Yaroslav D.

Sergeyev
Dmitri E. Kvasov (Eds.)
LNCS 11973

Numerical Computations:
Theory and Algorithms
Third International Conference, NUMTA 2019
Crotone, Italy, June 15–21, 2019
Revised Selected Papers, Part I
Lecture Notes in Computer Science 11973

Founding Editors
Gerhard Goos
Karlsruhe Institute of Technology, Karlsruhe, Germany
Juris Hartmanis
Cornell University, Ithaca, NY, USA

Editorial Board Members

Elisa Bertino
Purdue University, West Lafayette, IN, USA
Wen Gao
Peking University, Beijing, China
Bernhard Steffen
TU Dortmund University, Dortmund, Germany
Gerhard Woeginger
RWTH Aachen, Aachen, Germany
Moti Yung
Columbia University, New York, NY, USA
More information about this series at http://www.springer.com/series/7407
Yaroslav D. Sergeyev Dmitri E. Kvasov (Eds.)
•

Numerical Computations:
Theory and Algorithms
Third International Conference, NUMTA 2019
Crotone, Italy, June 15–21, 2019
Revised Selected Papers, Part I

123
Editors
Yaroslav D. Sergeyev Dmitri E. Kvasov
University of Calabria University of Calabria
Rende, Italy Rende, Italy
Lobachevsky University of Nizhny Lobachevsky University of Nizhny
Novgorod Novgorod
Nizhny Novgorod, Russia Nizhny Novgorod, Russia

ISSN 0302-9743 ISSN 1611-3349 (electronic)

Lecture Notes in Computer Science
ISBN 978-3-030-39080-8 ISBN 978-3-030-39081-5 (eBook)
https://doi.org/10.1007/978-3-030-39081-5
LNCS Sublibrary: SL1 – Theoretical Computer Science and General Issues

© Springer Nature Switzerland AG 2020

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of the
material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now
known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this book are
believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors
give a warranty, expressed or implied, with respect to the material contained herein or for any errors or
omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in
published maps and institutional affiliations.

This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
 Preface

This volume, edited by Yaroslav D. Sergeyev and Dmitri E. Kvasov, contains selected
peer-reviewed papers from the Third Triennial International Conference and Summer
School on Numerical Computations: Theory and Algorithms (NUMTA 2019) held in
Le Castella – Isola Capo Rizzuto (Crotone), Italy, during June 15–21, 2019.
The NUMTA 2019 conference has continued the previous successful editions of
NUMTA that took place in 2013 and 2016 in Italy in the beautiful Calabria region.
NUMTA 2019 was organized by the University of Calabria, Department of Com-
puter Engineering, Modeling, Electronics and Systems Science, Italy, in cooperation
with the Society for Industrial and Applied Mathematics (SIAM), USA. This edition
had the high patronage of the municipality of Crotone – the city of Pythagoras and his
followers, the Pythagoreans. In fact, Pythagoras established the first Pythagorean
community in this city in the 6th century B.C. It was a very special feeling for the
participants of NUMTA 2019 to visit these holy, for any mathematician, places with a
conference dedicated to numerical mathematics.
The goal of the NUMTA series of conferences is to create a multidisciplinary round
table for an open discussion on numerical modeling nature by using traditional and
emerging computational paradigms. Participants of the NUMTA 2019 conference
discussed multiple aspects of numerical computations and modeling starting from
foundations and philosophy of mathematics and computer science to advanced
numerical techniques. New technological challenges and fundamental ideas from
theoretical computer science, machine learning, linguistic, logic, set theory, and phi-
losophy met the requirements, as well as fresh, new applications from physics,
chemistry, biology, and economy.
Researchers from both theoretical and applied sciences were invited to use this
excellent opportunity to exchange ideas with leading scientists from different research
fields. Papers discussing new computational paradigms, relations with foundations of
mathematics, and their impact on natural sciences were particularly solicited. Special
attention during the conference was dedicated to numerical optimization techniques
and a variety of issues related to the theory and practice of the usage of infinities and
infinitesimals in numerical computations. In particular, there were a substantial number
of talks dedicated to a new promising methodology allowing one to execute numerical
computations with finite, infinite, and infinitesimal numbers on a new type of a
computational device – the Infinity Computer patented in the EU, Russia, and the USA.
This edition of the NUMTA conference was dedicated to the 80th birthday of
Professor Roman Strongin. For the past 50 years Roman Strongin has been a leader and
an innovator in Global Optimization, an important field of Numerical Analysis having
numerous real-life applications. His book on Global Optimization, published in 1978,
was one of the first in the world on this subject. Now it is a classic and has been used by
many as their first introduction and continued inspiration for Global Optimization.
Since that time, Roman has published numerous books and more than 400 papers in
vi Preface

several scientific fields and has been rewarded with many national and international
honors including the President of the Russian Federation Prize. For decades Roman
served as Dean, First Vice-Rector, and Rector of the famous Lobachevsky State
University of Nizhny Novgorod. Since 2008 he has been President of this university.
He is also Chairman of the Council of Presidents of Russian Universities,
Vice-President of the Union of the Rectors of Russian Universities, and Chairman
of the Public Chamber of the Nizhny Novgorod Region.
We are proud to inform you that 200 researchers from the following 30 countries
participated at the NUMTA 2019 conference: Argentina, Bulgaria, Canada, China,
Czech Republic, Estonia, Finland, France, Germany, Greece, India, Iran, Italy, Japan,
Kazakhstan, Latvia, Lithuania, the Netherlands, Philippines, Portugal, Romania,
Russia, Saudi Arabia, South Korea, Spain, Switzerland, Thailand, Ukraine, the UK,
and the USA.
The following plenary lecturers shared their achievements with the NUMTA 2019
participants:
• Louis D’Alotto, USA: “Infinite games on finite graphs using Grossone”
• Renato De Leone, Italy: “Recent advances on the use of Grossone in optimization
and regularization problems”
• Kalyanmoy Deb, USA: “Karush-Kuhn-Tucker proximity measure for convergence
of real-parameter single and multi-criterion optimization”
• Luca Formaggia, Italy: “Numerical modeling of flow in fractured porous media and
fault reactivation”
• Jan Hesthaven, Switzerland: “Precision algorithms”
• Francesca Mazzia, Italy: “Numerical differentiation on the Infinity Computer and
applications for solving ODEs and approximating functions”
• Michael Vrahatis, Greece: “Generalizations of the intermediate value theorem for
approximations of fixed points and zeroes of continuous functions”
• Anatoly Zhigljavsky, UK: “Uniformly distributed sequences and space-filling”
Moreover, the following tutorials were presented during the conference:
• Roberto Natalini, Italy: “Vector kinetic approximations to fluid-dynamics
equations”
• Yaroslav Sergeyev, Italy and Russia: “Grossone-based Infinity Computing with
numerical infinities and infinitesimals”
• Vassili Toropov, UK: “Design optimization techniques for industrial applications:
Challenges and progress”
These proceedings of NUMTA 2019 consist of two volumes: Part I and Part II. The
book you have in your hands is the first part containing peer-reviewed papers selected
from big special streams and sessions held during the conference. The second volume
contains peer-reviewed papers selected from the general stream, plenary lectures, and
small special sessions of NUMTA 2019. The special streams and sessions from which
the papers selected for this volume have been chosen are listed below (in the alpha-
betical order):
 Preface vii

(1) Approximation: methods, algorithms and applications; organized by Alessandra De

Rossi (University of Turin, Italy), Francesco Dell’Accio (University of Calabria,
Italy), Elisa Francomano (University of Palermo, Italy), and Donatella Occorsio
(University of Basilicata, Italy).
(2) Computational Methods for Data Analysis; organized by Rosanna Campagna,
Salvatore Cuomo, and Francesco Piccialli (all from the University of Naples
Federico II, Italy).
(3) First Order Methods in Optimization: Theory and Applications; organized by
Simone Rebegoldi (University of Modena and Reggio Emilia, Italy) and Marco
Viola (Sapienza University of Rome, Italy).
(4) High Performance Computing in Modelling and Simulation; organized by William
Spataro, Donato D’Ambrosio, Rocco Rongo (all from the University of Calabria,
Italy), and Andrea Giordano (ICAR–CNR, Italy).
(5) Numbers, Algorithms, and Applications; organized by Fabio Caldarola, Gianfranco
D’Atri, Mario Maiolo (all from the University of Calabria, Italy), and Giuseppe
Pirillo (University of Florence, Italy).
(6) Optimization and Management of Water Supply; organized by Fabio Caldarola,
Mario Maiolo, Giuseppe Mendicino, and Patrizia Piro (all from the University of
Calabria, Italy).
All the papers which were accepted for publication in this LNCS volume underwent
a thorough peer review process (required up to three review rounds for some manu-
scripts) by the members of the NUMTA 2019 Program Committee and independent
reviewers. The editors thank all the participants for their dedication to the success of
NUMTA 2019 and are grateful to the reviewers for their valuable work. The support
of the Springer LNCS editorial staff and the sponsorship of the Young Researcher Prize
by Springer are greatly appreciated.
The editors express their gratitude to the institutions that offered their generous
support to the international conference NUMTA 2019. This support was essential for
the success of this event:
– University of Calabria (Italy)
– Department of Computer Engineering, Modeling, Electronics and Systems Science
of the University of Calabria (Italy)
– Italian National Group for Scientific Computation of the National Institute for
Advanced Mathematics F. Severi (Italy)
– Institute of High Performance Computing and Networking of the National Research
Council (Italy)
– International Association for Mathematics and Computers in Simulation
– International Society of Global Optimization
Many thanks go to Maria Chiara Nasso from the University of Calabria, Italy, for
her valuable support in the technical editing of this volume.

October 2019 Yaroslav D. Sergeyev

Dmitri E. Kvasov
 Organization

General Chair
Yaroslav Sergeyev University of Calabria, Italy, and Lobachevsky
University of Nizhny Novgorod, Russia

Scientific Committee

Lidia Aceto, Italy Michael Emmerich, The Netherlands

Andy Adamatzky, UK Adil Erzin, Russia
Francesco Archetti, Italy Yury Evtushenko, Russia
Thomas Bäck, The Netherlands Giovanni Fasano, Italy
Roberto Battiti, Italy Şerife Faydaoğlu, Turkey
Roman Belavkin, UK Luca Formaggia, Italy
Giancarlo Bigi, Italy Elisa Francomano, Italy
Paola Bonizzoni, Italy Masao Fukushima, Japan
Luigi Brugnano, Italy David Gao, Australia
Sergiy Butenko, USA Manlio Gaudioso, Italy
Sonia Cafieri, France Victor Gergel, Russia
Tianxin Cai, China Jonathan Gillard, UK
Cristian Calude, New Zealand Daniele Gregori, Italy
Antonio Candelieri, Italy Vladimir Grishagin, Russia
Mario Cannataro, Italy Mario Guarracino, Italy
Giovanni Capobianco, Italy Nicola Guglielmi, Italy
Domingos Cardoso, Portugal Jan Hesthaven, Switzerland
Francesco Carrabs, Italy Felice Iavernaro, Italy
Emilio Carrizosa, Spain Mikhail Khachay, Russia
Leocadio Casado, Spain Oleg Khamisov, Russia
Carmine Cerrone, Italy Timos Kipouros, UK
Raffaele Cerulli, Italy Lefteris Kirousis, Greece
Marco Cococcioni, Italy Yury Kochetov, Russia
Salvatore Cuomo, Italy Olga Kostyukova, Belarus
Louis D’Alotto, USA Vladik Kreinovich, USA
Oleg Davydov, Germany Dmitri Kvasov, Italy
Renato De Leone, Italy Hoai An Le Thi, France
Alessandra De Rossi, Italy Wah June Leong, Malaysia
Kalyanmoy Deb, USA Øystein Linnebo, Norway
Francesco Dell’Accio, Italy Antonio Liotta, UK
Branko Dragovich, Serbia Marco Locatelli, Italy
Gintautas Dzemyda, Lithuania Stefano Lucidi, Italy
Yalchin Efendiev, USA Maurice Margenstern, France
x Organization

Vladimir Mazalov, Russia Khodr Shamseddine, Canada

Francesca Mazzia, Italy Sameer Shende, USA
Maria Mellone, Italy Vladimir Shylo, Ukraine
Kaisa Miettinen, Finland Theodore Simos, Greece
Edmondo Minisci, UK Vinai Singh, India
Nenad Mladenovic, Serbia Majid Soleimani-Damaneh, Iran
Ganesan Narayanasamy, USA William Spataro, Italy
Ivo Nowak, Germany Maria Grazia Speranza, Italy
Donatella Occorsio, Italy Giandomenico Spezzano, Italy
Marco Panza, USA Rosamaria Spitaleri, Italy
Massimo Pappalardo, Italy Alexander Strekalovskiy, Russia
Panos Pardalos, USA Roman Strongin, Russia
Remigijus Paulavičius, Lithuania Gopal Tadepalli, India
Hoang Xuan Phu, Vietnam Tatiana Tchemisova, Portugal
Stefan Pickl, Germany Gerardo Toraldo, Italy
Raffaele Pisano, France Vassili Toropov, UK
Yuri Popkov, Russia Hiroshi Umeo, Japan
Mikhail Posypkin, Russia Michael Vrahatis, Greece
Oleg Prokopyev, USA Song Wang, Australia
Davide Rizza, UK Gerhard-Wilhelm Weber, Poland
Massimo Roma, Italy Luca Zanni, Italy
Valeria Ruggiero, Italy Anatoly Zhigljavsky, UK
Maria Grazia Russo, Italy Antanas Žilinskas, Lithuania
Nick Sahinidis, USA Julius Žilinskas, Lithuania
Leonidas Sakalauskas, Lithuania Joseph Zyss, France
Yaroslav Sergeyev, Italy

Organizing Committee
Francesco Dall’Accio University of Calabria, Rende (CS), Italy
Alfredo Garro University of Calabria, Rende (CS), Italy
Vladimir Grishagin Lobachevsky University of Nizhny Novgorod,
Nizhny Novgorod, Russia
Dmitri Kvasov (Chair) University of Calabria, Rende (CS), Italy, and
Lobachevsky University of Nizhny Novgorod,
Nizhny Novgorod, Russia
Marat Mukhametzhanov University of Calabria, Rende (CS), Italy, and
Lobachevsky University of Nizhny Novgorod,
Nizhny Novgorod, Russia
Maria Chiara Nasso University of Calabria, Rende (CS), Italy
Clara Pizzuti National Research Council of Italy (CNR),
Institute for High Performance Computing and
Networking (ICAR), Rende (CS), Italy
Davide Rizza University of East Anglia, Norwich, UK
Yaroslav Sergeyev University of Calabria, Rende (CS), Italy, and
Lobachevsky University of Nizhny Novgorod,
Nizhny Novgorod, Russia
 Organization xi

Sponsors

In cooperation with

and under the high patronage of the

municipality of Crotone, Italy.
 Contents – Part I

Approximation: Methods, Algorithms, and Applications

Towards an Efficient Implementation of an Accurate SPH Method . . . . . . . . 3

Laura Antonelli, Daniela di Serafino, Elisa Francomano,
Francesco Gregoretti, and Marta Paliaga

A Procedure for Laplace Transform Inversion Based on Smoothing

Exponential-Polynomial Splines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
Rosanna Campagna, Costanza Conti, and Salvatore Cuomo

An Adaptive Refinement Scheme for Radial Basis Function Collocation . . . . 19

Roberto Cavoretto and Alessandra De Rossi

A 3D Efficient Procedure for Shepard Interpolants on Tetrahedra . . . . . . . . . 27

Roberto Cavoretto, Alessandra De Rossi, Francesco Dell’Accio,
and Filomena Di Tommaso

Interpolation by Bivariate Quadratic Polynomials and Applications

to the Scattered Data Interpolation Problem . . . . . . . . . . . . . . . . . . . . . . . . 35
Francesco Dell’Accio and Filomena Di Tommaso

Comparison of Shepard’s Like Methods with Different Basis Functions . . . . . 47

Francesco Dell’Accio, Filomena Di Tommaso, and Domenico Gonnelli

A New Remez-Type Algorithm for Best Polynomial Approximation . . . . . . . 56

Nadaniela Egidi, Lorella Fatone, and Luciano Misici

An SVE Approach for the Numerical Solution of Ordinary

Differential Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
Nadaniela Egidi and Pierluigi Maponi

Uniform Weighted Approximation by Multivariate Filtered Polynomials . . . . 86

Donatella Occorsio and Woula Themistoclakis

Computational Methods for Data Analysis

A Travelling Wave Solution for Nonlinear Colloid Facilitated Mass

Transport in Porous Media . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
Salvatore Cuomo, Fabio Giampaolo, and Gerardo Severino
xiv Contents – Part I

Performance Analysis of a Multicore Implementation for Solving

a Two-Dimensional Inverse Anomalous Diffusion Problem . . . . . . . . . . . . . 109
Pasquale De Luca, Ardelio Galletti, Giulio Giunta, Livia Marcellino,
and Marzie Raei

Adaptive RBF Interpolation for Estimating Missing Values

in Geographical Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
Kaifeng Gao, Gang Mei, Salvatore Cuomo, Francesco Piccialli,
and Nengxiong Xu

Stochastic Mechanisms of Information Flow in Phosphate Economy

of Escherichia coli . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
Ozan Kahramanoğulları, Cansu Uluşeker, and Martin M. Hancyzc

NMR Data Analysis of Water Mobility in Wheat Flour Dough:

A Computational Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
Annalisa Romano, Rosanna Campagna, Paolo Masi,
and Gerardo Toraldo

First Order Methods in Optimization: Theory and Applications

A Limited Memory Gradient Projection Method for Box-Constrained

Quadratic Optimization Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
Serena Crisci, Federica Porta, Valeria Ruggiero, and Luca Zanni

A Gradient-Based Globalization Strategy for the Newton Method . . . . . . . . . 177

Daniela di Serafino, Gerardo Toraldo, and Marco Viola

On the Steplength Selection in Stochastic Gradient Methods. . . . . . . . . . . . . 186

Giorgia Franchini, Valeria Ruggiero, and Luca Zanni

Efficient Block Coordinate Methods for Blind Cauchy Denoising . . . . . . . . . 198

Simone Rebegoldi, Silvia Bonettini, and Marco Prato

High Performance Computing in Modelling and Simulation

A Parallel Software Platform for Pathway Enrichment . . . . . . . . . . . . . . . . . 215

Giuseppe Agapito and Mario Cannataro

Hierarchical Clustering of Spatial Urban Data. . . . . . . . . . . . . . . . . . . . . . . 223

Eugenio Cesario, Andrea Vinci, and Xiaotian Zhu

Improving Efficiency in Parallel Computing Leveraging

Local Synchronization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232
Franco Cicirelli, Andrea Giordano, and Carlo Mastroianni
 Contents – Part I xv

A General Computational Formalism for Networks of Structured Grids . . . . . 243

Donato D’Ambrosio, Paola Arcuri, Mario D’Onghia, Marco Oliverio,
Rocco Rongo, William Spataro, Andrea Giordano, Davide Spataro,
Alessio De Rango, Giuseppe Mendicino, Salvatore Straface,
and Alfonso Senatore

Preliminary Model of Saturated Flow Using Cellular Automata. . . . . . . . . . . 256

Alessio De Rango, Luca Furnari, Andrea Giordano, Alfonso Senatore,
Donato D’Ambrosio, Salvatore Straface, and Giuseppe Mendicino

A Cybersecurity Framework for Classifying Non Stationary Data Streams

Exploiting Genetic Programming and Ensemble Learning. . . . . . . . . . . . . . . 269
Gianluigi Folino, Francesco Sergio Pisani, and Luigi Pontieri

A Dynamic Load Balancing Technique for Parallel Execution

of Structured Grid Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 278
Andrea Giordano, Alessio De Rango, Rocco Rongo,
Donato D’Ambrosio, and William Spataro

Final Sediment Outcome from Meteorological Flood Events:

A Multi-modelling Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 291
Valeria Lupiano, Claudia R. Calidonna, Elenio Avolio,
Salvatore La Rosa, Giuseppe Cianflone, Antonio Viscomi,
Rosanna De Rosa, Rocco Dominici, Ines Alberico, Nicola Pelosi,
Fabrizio Lirer, and Salvatore Di Gregorio

Parallel Algorithms for Multifractal Analysis of River Networks . . . . . . . . . . 307

Leonardo Primavera and Emilia Florio

A Methodology Approach to Compare Performance of Parallel

Programming Models for Shared-Memory Architectures. . . . . . . . . . . . . . . . 318
Gladys Utrera, Marisa Gil, and Xavier Martorell

Numbers, Algorithms, and Applications

New Approaches to Basic Calculus: An Experimentation

via Numerical Computation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 329
Luigi Antoniotti, Fabio Caldarola, Gianfranco d’Atri,
and Marco Pellegrini

A Multi-factor RSA-like Scheme with Fast Decryption Based on Rédei

Rational Functions over the Pell Hyperbola . . . . . . . . . . . . . . . . . . . . . . . . 343
Emanuele Bellini and Nadir Murru
xvi Contents – Part I

Paradoxes of the Infinite and Ontological Dilemmas Between Ancient

Philosophy and Modern Mathematical Solutions . . . . . . . . . . . . . . . . . . . . . 358
Fabio Caldarola, Domenico Cortese, Gianfranco d’Atri,
and Mario Maiolo

The Sequence of Carboncettus Octagons . . . . . . . . . . . . . . . . . . . . . . . . . . 373

Fabio Caldarola, Gianfranco d’Atri, Mario Maiolo,
and Giuseppe Pirillo

On the Arithmetic of Knuth’s Powers and Some Computational Results

About Their Density . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 381
Fabio Caldarola, Gianfranco d’Atri, Pietro Mercuri,
and Valerio Talamanca

Combinatorics on n-sets: Arithmetic Properties and Numerical Results. . . . . . 389

Fabio Caldarola, Gianfranco d’Atri, and Marco Pellegrini

Numerical Problems in XBRL Reports and the Use of Blockchain

as Trust Enabler . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 402
Gianfranco d’Atri, Van Thanh Le, Dino Garrì, and Stella d’Atri

Modelling on Human Intelligence a Machine Learning System . . . . . . . . . . . 410

Michela De Pietro, Francesca Bertacchini, Pietro Pantano,
and Eleonora Bilotta

Algorithms for Jewelry Industry 4.0 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 425

Francesco Demarco, Francesca Bertacchini, Carmelo Scuro,
Eleonora Bilotta, and Pietro Pantano

Clustering Analysis to Profile Customers’ Behaviour in POWER CLOUD

Energy Community . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 437
Lorella Gabriele, Francesca Bertacchini, Simona Giglio,
Daniele Menniti, Pietro Pantano, Anna Pinnarelli, Nicola Sorrentino,
and Eleonora Bilotta

A Grossone-Based Numerical Model for Computations with Infinity:

A Case Study in an Italian High School. . . . . . . . . . . . . . . . . . . . . . . . . . . 451
Francesco Ingarozza, Maria Teresa Adamo, Maria Martino,
and Aldo Piscitelli

A Computational Approach with MATLAB Software for Nonlinear

Equation Roots Finding in High School Maths . . . . . . . . . . . . . . . . . . . . . . 463
Annarosa Serpe

Task Mathematical Modelling Design in a Dynamic Geometry

Environment: Archimedean Spiral’s Algorithm . . . . . . . . . . . . . . . . . . . . . . 478
Annarosa Serpe and Maria Giovanna Frassia
 Contents – Part I xvii

Optimization and Management of Water Supply

Numerical Experimentations for a New Set of Local Indices

of a Water Network. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 495
Marco Amos Bonora, Fabio Caldarola, Joao Muranho, Joaquim Sousa,
and Mario Maiolo

Performance Management of Demand and Pressure Driven Analysis

in a Monitored Water Distribution Network . . . . . . . . . . . . . . . . . . . . . . . . 506
Marco Amos Bonora, Manuela Carini, Gilda Capano, Rocco Cotrona,
Daniela Pantusa, Joaquim Sousa, and Mario Maiolo

Algebraic Tools and New Local Indices for Water Networks:

Some Numerical Examples. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 517
Fabio Caldarola and Mario Maiolo

Identification of Contamination Potential Source (ICPS):

A Topological Approach for the Optimal Recognition of Sensitive
Nodes in a Water Distribution Network . . . . . . . . . . . . . . . . . . . . . . . . . . . 525
Gilda Capano, Marco Amos Bonora, Manuela Carini,
and Mario Maiolo

Seeking for a Trade-Off Between Accuracy and Timeliness

in Meteo-Hydrological Modeling Chains . . . . . . . . . . . . . . . . . . . . . . . . . . 537
Luca Furnari, Alfonso Senatore, and Giuseppe Mendicino

Optimization Model for Water Distribution Network Planning

in a Realistic Orographic Framework. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 545
Mario Maiolo, Joaquim Sousa, Manuela Carini,
Francesco Chiaravalloti, Marco Amos Bonora, Gilda Capano,
and Daniela Pantusa

Scenario Optimization of Complex Water Supply Systems for Energy

Saving and Drought-Risk Management . . . . . . . . . . . . . . . . . . . . . . . . . . . 557
Jacopo Napolitano and Giovanni M. Sechi

Optimizing Rainwater Harvesting Systems for Non-potable Water Uses

and Surface Runoff Mitigation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 570
Stefania Anna Palermo, Vito Cataldo Talarico, and Behrouz Pirouz

New Mathematical Optimization Approaches for LID Systems . . . . . . . . . . . 583

Behrouz Pirouz, Stefania Anna Palermo, Michele Turco,
and Patrizia Piro

Evaluation of an Integrated Seasonal Forecast System for Agricultural

Water Management in Mediterranean Regions . . . . . . . . . . . . . . . . . . . . . . 596
Alfonso Senatore, Domenico Fuoco, Antonella Sanna, Andrea Borrelli,
Giuseppe Mendicino, and Silvio Gualdi
xviii Contents – Part I

Optimization of Submarine Outfalls with a Multiport Diffuser Design . . . . . . 604

Salvatore Sinopoli, Marco Amos Bonora, Gilda Capano,
Manuela Carini, Daniela Pantusa, and Mario Maiolo

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 619

 Contents – Part II

Long Papers

Numerical Algorithms for the Parametric Continuation of Stiff

ODEs Deriving from the Modeling of Combustion with Detailed
Chemical Mechanisms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
Luigi Acampora and Francesco S. Marra

Stability Analysis of DESA Optimization Algorithm . . . . . . . . . . . . . . . . . . 17

Rizavel C. Addawe and Joselito C. Magadia

Are Humans Bayesian in the Optimization of Black-Box Functions? . . . . . . . 32

Antonio Candelieri, Riccardo Perego, Ilaria Giordani,
and Francesco Archetti

A New Syntax for Diagrammatic Logic: A Generic Figures Approach. . . . . . 43

Gianluca Caterina and Rocco Gangle

Objective and Violation Upper Bounds on a DIRECT-Filter Method

for Global Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
M. Fernanda P. Costa, Ana Maria A. C. Rocha,
and Edite M. G. P. Fernandes

The Approximate Synthesis of Optimal Control for Heterogeneous

Discrete Systems with Intermediate Criteria . . . . . . . . . . . . . . . . . . . . . . . . 72
Olga Danilenko and Irina Rasina

Modelling Climate Changes with Stationary Models: Is It Possible

or Is It a Paradox? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
Davide Luciano De Luca, Andrea Petroselli, and Luciano Galasso

Differential Equations and Uniqueness Theorems for the Generalized

Attenuated Ray Transforms of Tensor Fields . . . . . . . . . . . . . . . . . . . . . . . 97
Evgeny Yu. Derevtsov, Yuriy S. Volkov, and Thomas Schuster

Multiextremal Optimization in Feasible Regions with Computable

Boundaries on the Base of the Adaptive Nested Scheme . . . . . . . . . . . . . . . 112
Victor Gergel, Vladimir Grishagin, and Ruslan Israfilov

On Polyhedral Estimates of Reachable Sets of Discrete-Time Systems

with Uncertain Matrices and Integral Bounds on Additive Terms . . . . . . . . . 124
Elena K. Kostousova
xx Contents – Part II

Numerical Simulation of Hyperbolic Conservation Laws Using High

Resolution Schemes with the Indulgence of Fuzzy Logic . . . . . . . . . . . . . . . 139
Ruchika Lochab and Vivek Kumar

Methodology for Interval-Valued Matrix Games with 2-Tuple Fuzzy

Linguistic Information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
Tanya Malhotra, Anjana Gupta, and Anjali Singh

Set-Membership Computation of Integrals with Uncertain Endpoints . . . . . . . 169

Olivier Mullier and Julien Alexandre dit Sandretto

Epidemic Spreading Curing Strategy Over Directed Networks . . . . . . . . . . . 182

Clara Pizzuti and Annalisa Socievole

Learning Aerial Image Similarity Using Triplet Networks . . . . . . . . . . . . . . 195

Vytautas Valaitis, Virginijus Marcinkevicius, and Rokas Jurevicius

In the Quest for Invariant Structures Through Graph Theory, Groups

and Mechanics: Methodological Aspects in the History
of Applied Mathematics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208
Sandra Visokolskis and Carla Trillini

Generalizations of the Intermediate Value Theorem for Approximating

Fixed Points and Zeros of Continuous Functions. . . . . . . . . . . . . . . . . . . . . 223
Michael N. Vrahatis

Modelling Population Size Using Horvitz-Thompson Approach Based

on the Zero-Truncated Poisson Lindley Distribution . . . . . . . . . . . . . . . . . . 239
Ratchaneewan Wongprachan

Numerical Analysis of the Model of Optimal Consumption

and Borrowing with Random Time Scale . . . . . . . . . . . . . . . . . . . . . . . . . . 255
Aleksandra Zhukova and Igor Pospelov

Short Papers

Conditions of the Stability Preservation Under Discretization

of a Class of Nonlinear Time-Delay Systems . . . . . . . . . . . . . . . . . . . . . . . 271
Alexander Yu. Aleksandrov

Numerical Investigation of Natural Rough-Bed Flow . . . . . . . . . . . . . . . . . . 280

Giancarlo Alfonsi, Domenico Ferraro, Agostino Lauria,
and Roberto Gaudio

A Dynamic Precision Floating-Point Arithmetic Based on the Infinity

Computer Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 289
Pierluigi Amodio, Luigi Brugnano, Felice Iavernaro,
and Francesca Mazzia
 Contents – Part II xxi

Numerical Strategies for Solving Multiparameter Spectral Problems . . . . . . . 298

Pierluigi Amodio and Giuseppina Settanni

Algorithms of 3D Wind Field Reconstructing by Lidar Remote

Sensing Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 306
Nikolay Baranov

Stationarity Condition for Nonsmooth MPVCs with Constraint Set . . . . . . . . 314

David Barilla, Giuseppe Caristi, and Nader Kanzi

Molecular Dynamics Performance Evaluation with Modern

Computer Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 322
Emanuele Breuza, Giorgio Colombo, Daniele Gregori,
and Filippo Marchetti

Artificial Neural Networks Training Acceleration Through Network

Science Strategies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 330
Lucia Cavallaro, Ovidiu Bagdasar, Pasquale De Meo,
Giacomo Fiumara, and Antonio Liotta

Grossone Methodology for Lexicographic Mixed-Integer Linear

Programming Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 337
Marco Cococcioni, Alessandro Cudazzo, Massimo Pappalardo,
and Yaroslav D. Sergeyev

Infinite Games on Finite Graphs Using Grossone . . . . . . . . . . . . . . . . . . . . 346

Louis D’Alotto

A Novel Geometric Approach to the Problem

of Multidimensional Scaling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 354
Gintautas Dzemyda and Martynas Sabaliauskas

A Simulink-Based Infinity Computer Simulator and Some Applications . . . . . 362

Alberto Falcone, Alfredo Garro, Marat S. Mukhametzhanov,
and Yaroslav D. Sergeyev

Generalizing Pure and Impure Iterated Prisoner’s Dilemmas

to the Case of Infinite and Infinitesimal Quantities . . . . . . . . . . . . . . . . . . . 370
Lorenzo Fiaschi and Marco Cococcioni

Multidimensional Global Search Using Numerical Estimations

of Minimized Function Derivatives and Adaptive Nested
Optimization Scheme. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 378
Victor Gergel and Alexey Goryachikh
xxii Contents – Part II

2-Approximation Polynomial-Time Algorithm for a Cardinality-Weighted

2-Partitioning Problem of a Sequence . . . . . . . . . . . . . . . . . . . . . . . . . . . . 386
Alexander Kel’manov, Sergey Khamidullin, and Anna Panasenko

Exact Linear-Time Algorithm for Parameterized K-Means Problem

with Optimized Number of Clusters in the 1D Case . . . . . . . . . . . . . . . . . . 394
Alexander Kel’manov and Vladimir Khandeev

Polynomial-Time Approximation Scheme for a Problem of Searching

for the Largest Subset with the Constraint on Quadratic Variation . . . . . . . . . 400
Vladimir Khandeev

On a Comparison of Several Numerical Integration Methods

for Ordinary Systems of Differential Equations . . . . . . . . . . . . . . . . . . . . . . 406
Anatoliy G. Korotchenko and Valentina M. Smoryakova

On Acceleration of Derivative-Free Univariate Lipschitz Global

Optimization Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 413
Dmitri E. Kvasov, Marat S. Mukhametzhanov, Maria Chiara Nasso,
and Yaroslav D. Sergeyev

Numerical Simulation of Ski-Jump Hydraulic Behavior . . . . . . . . . . . . . . . . 422

Agostino Lauria and Giancarlo Alfonsi

On Linear Spline Wavelets with Shifted Supports . . . . . . . . . . . . . . . . . . . . 430

Svetlana Makarova and Anton Makarov

An Online Learning Approach to a Multi-player N-armed

Functional Bandit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 438
Sam O’Neill, Ovidiu Bagdasar, and Antonio Liotta

The Singular Value Decomposition of the Operators of the Dynamic

Ray Transforms Acting on 2D Vector Fields . . . . . . . . . . . . . . . . . . . . . . . 446
Anna P. Polyakova, Ivan E. Svetov, and Bernadette N. Hahn

Numerical Investigation into a Combined Ecologically Pure Plasma-Vortex

Generator of Hydrogen and Thermal Energy. . . . . . . . . . . . . . . . . . . . . . . . 454
Denis Porfiriev, Svetlana Kurushina, Nonna Molevich,
and Igor Zavershinsky

Computer Modeling of Electrochemical Impedance Spectra for Defected

Phospholipid Membranes: Finite Element Analysis . . . . . . . . . . . . . . . . . . . 462
Tomas Raila, Tadas Meškauskas, Gintaras Valinčius, Marija Jankunec,
and Tadas Penkauskas

A Compact Filter Regularization Method for Solving Side

ways Heat Equation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 470
Ankita Shukla and Mani Mehra
 Contents – Part II xxiii

Acceleration of Global Search by Implementing Dual Estimates

for Lipschitz Constant . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 478
Roman Strongin, Konstantin Barkalov, and Semen Bevzuk

The Method of Approximate Inverse in Slice-by-Slice Vector

Tomography Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 487
Ivan E. Svetov, Svetlana V. Maltseva, and Alfred K. Louis

Computational Fluid Dynamics Methods for Wind Resources Assessment . . . 495

Sabine Upnere, Valerijs Bezrukovs, Vladislavs Bezrukovs,
Normunds Jekabsons, and Linda Gulbe

On Stationary Points of Distance Depending Potentials . . . . . . . . . . . . . . . . 503

Alexei Uteshev and Marina Goncharova

A Partition Based Bayesian Multi-objective Optimization Algorithm . . . . . . . 511

Antanas Žilinskas and Linas Litvinas

Problem Statement for Preparing a Single Batch of End Product

Under Uncertainty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 519
Anna Zykina, Olga Kaneva, Vladimir Savkin, and Tatyana Fink

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 529

Approximation: Methods, Algorithms,
and Applications
 Towards an Efficient Implementation
of an Accurate SPH Method

Laura Antonelli1 , Daniela di Serafino2 , Elisa Francomano3(B) ,

Francesco Gregoretti1 , and Marta Paliaga3
1
Institute for High Performance Computing and Networking (ICAR),
CNR, Naples, Italy
{laura.antonelli,francesco.gregoretti}@icar.cnr.it
2
Department of Mathematics and Physics, University of Campania
“Luigi Vanvitelli”, Caserta, Italy
[email protected]
3
Department of Engineering, University of Palermo, Palermo, Italy
{elisa.francomano,marta.paliaga}@unipa.it

Abstract. A modified version of the Smoothed Particle Hydrodynamics

(SPH) method is considered in order to overcome the loss of accuracy of
the standard formulation. The summation of Gaussian kernel functions
is employed, using the Improved Fast Gauss Transform (IFGT) to reduce
the computational cost, while tuning the desired accuracy in the SPH
method. This technique, coupled with an algorithmic design for exploit-
ing the performance of Graphics Processing Units (GPUs), makes the
method promising, as shown by numerical experiments.

Keywords: Smoothed Particle Hydrodynamics · Improved Fast Gauss

Transform · Graphics Processing Units

1 Introduction
In the last years mesh-free methods have gained much attention in many appli-
cation fields [3,8]. The Smoothed Particle Hydrodynamics (SPH) is a mesh-free
technique originally introduced in astrophysics by Gingold and Monaghan [5]
and by Lucy [9], and widely used because of its ability of dealing with highly
complex geometries [1,8]. However, when irregular data distributions are con-
sidered, a loss of accuracy can occur. Many techniques have been developed to
overcome this problem. In this paper we discuss a strategy based on the Taylor
series expansion, which simultaneously improves the approximation of a func-
tion and its derivatives obtained with the SPH method. The improvement in
accuracy comes at the cost of an additional computational effort, which can be
alleviated by employing fast summations [4] in the computational scheme.
The availability of General-Purpose Graphics Processing Units (GPGPUs)
provides an opportunity to further speed up the SPH method. This work can be
also considered as a first step toward an efficient implementation of the improved
c Springer Nature Switzerland AG 2020
Y. D. Sergeyev and D. E. Kvasov (Eds.): NUMTA 2019, LNCS 11973, pp. 3–10, 2020.
https://doi.org/10.1007/978-3-030-39081-5_1
4 L. Antonelli et al.

SPH version on GPGPUs. To this aim, we develop a GPU implementation of

one of its most computationally intensive tasks.
The remainder of this paper is organized as follows. A brief description of
the standard SPH method and its improved formulation is given in Sect. 2. A
sequential implementation of the improved SPH method, to be used as a refer-
ence CPU version for the development of a GPU version, is discussed in Sect. 3
along with results of numerical experiments. A partial implementation on GPU
is presented in Sect. 3.1. Conclusions are provided in Sect. 4.

2 The Improved SPH Method

The standard SPH method computes an approximation to a function f : Ω ⊂

Rd → R starting from the kernel approximation

< fh (x) >= f (ξ)K(x, ξ; h)dΩ, (1)
Ω

where x, ξ ∈ Ω, h ∈ R+ is the smoothing length, and K(x, ξ; h) is the kernel

function, which is usually normalized to unity and is required to be symmetric
and sufficiently smooth. By assuming that f is at least continuously differen-
tiable, the kernel approximation has second-order accuracy [4,5]. By considering
N source points ξ j ∈ Ω, each associated with a subdomain Ωj ⊂ Ω, the so-called
particle approximation of (1) is defined as


N
fh (x) = f (ξ j )K(x, ξ j ; h)dΩj (2)
j=1

where dΩj is the measure of Ωj . Note that (2) generally does not satisfy second-
order accuracy, e.g., when irregular point distributions are considered.
If f is sufficiently smooth, the k -th order Taylor expansion of f (ξ) can be
used in order to increase the accuracy of the approximation:
 1
f (ξ) = (ξ − x)α Dα f (x) + O(hk+1 ), (3)
α!
|α|≤k

d
where α = (α(1) , α(2) , . . . , α(d) ) ∈ Nd is a multi-index, |α| = i=1 α(i) , α! =
d (1) (2) (d)
∂ |α|
(i)
i=1 (α )!, y
α
= y1α · y2α . . . · ydα , and Dα = (1) α(1) (d) α(d)
. By
(∂x ) ....(∂x )
multiplying (3) and its derivatives up to the k-th order by the kernel function,
and integrating over Ω, we get
 Towards an Efficient Implementation of an Accurate SPH Method 5
  1 
f (ξ)K(x, ξ; h)dΩ = (ξ − x)α Dα f (x)K(x, ξ; h)dΩ
Ω α! Ω

|α|≤k

+ O(hk+1 )K(x, ξ; h)dΩ

Ω
.. (4)
 .
 1 
f (ξ)D K(x, ξ; h)dΩ =
k
(ξ − x)α Dα f (x)Dk K(x, ξ; h)dΩ
Ω α! Ω

|α|≤k

+ O(hk+1 )Dk K(x, ξ; h)dΩ.

Ω

By adopting the particle approximation and neglecting the last terms in the
right-hand side of (4), we get an approximation to f and its derivatives at each
evaluation point x, which can be written as a linear system of size m = (d +
k)!/(d! k!):
A(k) (k) (k)
x cx = bx , (5)
where
⎛ ⎞

N  N
(d) (d) k
1
⎜ K(x, ξ j ; h)dΩj ... (ξj − x ) K(x, ξ j ; h)dΩj ⎟
⎜ k! ⎟
⎜ j=1 j=1 ⎟
⎜ ⎟
A(k)
x =⎜
⎜
..
.
..
.
..
.
⎟,
⎟ (6)
⎜ N ⎟
⎜ 
N ⎟
⎝ k
D K(x, ξ j ; h)dΩj ... 1
(ξ − x ) D K(x, ξ ; h)dΩ ⎠
(d) (d) k k
k! j j j
j=1 j=1

⎛ ⎞

N

⎛ ⎞ ⎜ f (ξ)K(x, ξ j ; h)dΩj ⎟
⎜ ⎟
f (x) ⎜ j=1 ⎟
⎜ ⎟ ⎜ ⎟
c(k)
x =⎝ ..
. ⎠, b(k)
x =⎜
⎜
..
.
⎟.
⎟ (7)
⎜ N ⎟
D f (x)
k
⎜ ⎟
⎝ f (ξ)Dk K(x, ξ ; h)dΩj ⎠
j
j=1

This procedure improves the accuracy of the standard method as discussed in [4],
and has been successfully used in electromagnetic simulations [1].
The improved SPH method is more expensive than the standard one, because
(k)
it requires, for each evaluation point x, the construction of the matrix Ax and
(k) 2
the right-hand side bx , i.e., the computation of m + m sums, and the solution
of the linear system (5). As outlined in [4], when K(x, ξ; h) is the Gaussian ker-
nel, the computational cost can be significantly reduced using the Fast Gauss
Transform (FGT) [6]. More precisely, the FGT lowers the cost of computing M
Gaussian sums using N source points from O(N M ) to O(N + M ), by approxi-
mating the sums with a required order of accuracy. Actually, the Improved FGT
(IFGT) [10,13] is used, because it achieves higher efficiency than the original
FGT, especially when the dimension d increases, by combining a suitable fac-
torization of the exponential function in the Gauss transform with an adaptive
space partitioning scheme where the N source points are grouped into clusters,
6 L. Antonelli et al.

and exploiting the fast decay of the Gaussian function. The description of the
IFGT is beyond the scope of this paper. Here we provide only some details
(k) (k)
related to the computation of Ax and bx , for d = 2 and k = 1, i.e., m = 3.
(k)
The computation of Ax via IFGT requires m(m + 1)/2 = 6 Gauss transforms
(k)
using N source points, while the computation of bx requires m Gauss trans-
forms and the evaluation of f at each source point. For any evaluation point x,
the transforms to be computed have the following form:


N
Gl (x) = wl (ξ j )K(x, ξ j ; h), l = 1, . . . , L,
j=1

where L = m(m + 3)/2 = 9, the weights wl (ξ j ) are defined as

2
(1)
w1 (ξ j ) = ρj , w4 (ξ j ) = ρj ξj , w7 (ξ j ) = ρj f (ξ j ),

2
(1) (2) (1) (8)
w2 (ξ j ) = ρj ξj , w5 (ξ j ) = ρj ξj , w8 (ξ j ) = ρj ξj f (ξ j ),
(2) (1) (2) (2)
w3 (ξ j ) = ρj ξj , w6 (ξ j ) = ρj ξj ξj , w9 (ξ j ) = ρj ξj f (ξ j ),
(i)
ξj , i = 1, 2, is the i-th coordinate of the point ξ j , and ρj = 1
πh2 dΩj . The entries
(k) (k)
of Ax and bx in (6) and (7) can be obtained as
(k) (k) (k) (k)
A11 = Ã1 , A12 = Ã2 , A13 = Ã3 , b1 = Ã7 ,
(k) (k) (k) (k)
A21 = −Ã2 , A22 = Ã4 , A23 = Ã5 , b2 = −Ã8 ,
(k) (k) (k) (k)
A31 = −Ã3 , A32 = Ã5 , A33 = −Ã6 , b3 = −Ã9 ,
where
Ã1 = G1 , Ã4 = G4 − 2x(1) G2 + (x(1) )2 G1 , Ã7 = G7 ,
Ã2 = G2 − x G1 , Ã5 = G5 − x(1) G2 − x(2) G3 + x(1) x(2) G1 , Ã8 = G8 − x(1) G7 , (9)
(1)

Ã3 = G3 − x(2) G1 , Ã6 = G6 − 2x(2) G3 + (x(2) )2 G1 , Ã9 = G9 − x(2) G7 ,

and the dependence of Gl on x has been removed to simplify the notations.

(k) (k)
Note that all the matrices Ax and the right-hand sides bx can be computed
simultaneously.

3 Implementation and Numerical Experiments

The improved SPH method can be structured into five computational tasks:
(a) generate the source points ξ j and the evaluation points xi ; (b) compute
the weights wl (ξ j ); (c) evaluate the Gauss transforms Gl (xi ); (d) compute the
(k) (k)
matrices Axi and the right-hand sides bxi ; (e) solve the linear systems (5).
All the tasks were implemented in C++, with the aim of developing an effi-
cient serial CPU version to be used as a baseline for a GPU implementation.
Concerning task (a), Ω was set equal to [0, 1]d and three distributions of source
points were considered: uniform d-dimensional mesh, Halton [7] and Sobol’ [14]
d-dimensional sequences. The Halton and Sobol’ points were generated by using
the C++ code available from [2]. The evaluation points xi were distributed on
 Towards an Efficient Implementation of an Accurate SPH Method 7

Table 1. Functions used in the numerical experiments.

fa (x(1) , x(2) ) = 16 x(1) x(2) (1 − x(1) )(1 − x(2) )

1
fb (x(1) , x(2) ) = tanh (9(x(2) − x(1) ) + 1)
9
1.25 + cos(5.4 x(2) )
fc (x(1) , x(2) ) =
6 + 6(3 x(1) − 1)2
  2  2 
1 81 1 1
fd (x(1) , x(2) ) = exp − x(1) − + x(2) −
3 16 2 2

a uniform mesh over Ω. Task (c) exploits the IFGT method implemented in
the figtree package [10], which tunes IFGT parameters to the source distri-
bution to get tighter error bounds, according to the desired accuracy and the
selected smoothing length. Task (e) was implemented by using the LAPACK
routines DGETRF and DGETRS from the auto-tuning ATLAS library [15];
(k)
DGETRF computes the LU factorization of Axi , while DGETRS performs the
corresponding triangular solves. Finally, the implementation of tasks (b) and (d)
was a straightforward application of (8) and (9), storing all the weights wl (ξ j )
(k) (k)
and all the entries of the matrices Axi and the right-hand sides bxi into three
arrays of lengths LN , LM and mM , respectively. Note that tasks (c) and (e)
account for most of the execution time.
We performed numerical experiments with d = 2 and k = 1, using the four
test functions reported in Table 1 and values of N and M much greater than the
ones considered in [4]. For each distribution of source√points, we set N = (2n +1)2
and h = 1/2n , with n = 7, 8, 9, 10, 11, and M = ( N + 1)2 . The accuracy for
IFGT computations was set to 10−6 . The experiments were run on an Intel
Xeon E5–2670 2.50 GHz CPU with 192 GB of RAM, with the Linux CentOS 6.8
operating system and the GNU 4.4.7 C++ compiler.
Figure 1 shows, for each test function f , the maximum error in approximating
f with the function fh computed by our SPH implementation,

max |fh (xi ) − f (xi )|,

i=1,...,M

as N varies. It also shows the maximum error for the derivative of the function
with respect to x(1) (the error in the approximation of the derivative with respect
to x(2) is comparable). We see that when N increases from (27 + 1)2 = 16641
to (211 + 1)2 = 4198401, the maximum error in approximating f and its deriva-
tive decreases by about two orders and one order of magnitude, respectively, for
each test function and each source point data set. These results confirm that
increasing the number of source points can be strongly beneficial in terms of
accuracy for the improved SPH method. On the other hand, the availability of
high-throughput many-core processors, such as GPGPUs, encourages the devel-
opment of SPH implementations able to deal with high-dimensional problems
and very large sets of source and evaluation points. A first step in this direction
is discussed in the next section.
8 L. Antonelli et al.

Test function fa Test function fb

1⋅100 1⋅10-1

1⋅10-1 1⋅10-2
Max error (log scale)

Max error (log scale)

1⋅10-2 1⋅10-3

Uniform f
-3 -4
1⋅10 Uniform Dx(1)f 1⋅10
Halton f
-4 Halton Dx(1)f -5 Uniform f
1⋅10 1⋅10 Uniform Dx(1)f
Sobol’ f
Sobol’ Dx(1)f Halton f
1⋅10-5 1⋅10-6 Halton Dx(1)f
Sobol’ f
Sobol’ Dx(1)f
-6 -7
1⋅10 4 5 6 6 1⋅10 4 5 6 6
1⋅10 1⋅10 1⋅10 5⋅10 1⋅10 1⋅10 1⋅10 5⋅10
N (log scale) N (log scale)

Test function fc Test function fd

1⋅10-1 1⋅10-1

-2 -2
1⋅10 1⋅10
Max error (log scale)

Max error (log scale)

-3
1⋅10 1⋅10-3

-4 -4
1⋅10 1⋅10

Uniform f Uniform f
1⋅10-5 Uniform Dx(1)f 1⋅10-5 Uniform Dx(1)f
Halton f Halton f
1⋅10-6 Halton Dx(1)f 1⋅10-6 Halton Dx(1)f
Sobol’ f Sobol’ f
Sobol’ Dx(1)f Sobol’ Dx(1)f
-7 -7
1⋅10 4 5 6 6 1⋅10 4 5 6 6
1⋅10 1⋅10 1⋅10 5⋅10 1⋅10 1⋅10 1⋅10 5⋅10
N (log scale) N (log scale)

Fig. 1. Maximum error in the approximation of the test functions and their first deriva-
tives with respect to x(1) .

3.1 A First Step Toward a GPU Implementation

Nowadays GPUs can be found in many computers, ranging from high-end

machines to laptops. In particular, an NVIDIA GPGPU is made of an array
of streaming multiprocessors, each consisting of a fixed number of scalar proces-
sors, one or more instruction fetch units and on-chip fast memory. On NVIDIA
GPUs, computations are performed by following the Single Instruction Multi-
ple Threads programming model, implemented in the CUDA software frame-
work [11], where multiple independent threads execute concurrently the same
instruction. A CUDA program consists of a host program that runs on a CPU
host and a kernel program that executes on a GPU device. Since thousands of
cores are available in a GPU, the kernel program must expose a high level of par-
allelism to keep the GPU busy; balanced workload among threads, minimization
of data transfers from the CPU to the GPU and vice versa, and suitable memory
access patterns are also fundamental to get high performance. Our final goal is
to efficiently map all the computational tasks of the improved SPH method onto
GPGPUs, in order to build a parallel computational tool for SPH simulations of
large-scale real-time applications. In this work we focus on task (e). The number
of linear systems (5) to be solved is equal to the number, M , of evaluation points,
while the size of the systems, m = (d + k)!/(d! k!), depends on the order of the
Taylor expansion, k, and the problem dimension, d. By setting d = 2 and k = 1
and letting M increase from the order of 104 to the order of 106 , we are faced
 Towards an Efficient Implementation of an Accurate SPH Method 9

GPU vs CPU - Solution of linear systems

2500 20

18
2000 16

14

Time (ms)
1500

Speedup
CPU 12
GPU
10
Speedup
1000
8

6
500
4

2
0
1⋅105 1⋅106 2⋅106 3⋅106 4⋅106
M (log scale)

Fig. 2. GPU vs CPU performance comparison for task (e).

with the challenge of solving from thousands to millions of very small systems.
In order to avoid the overhead of multiple API calls, we implemented a GPU
version of task (e) that makes use of the batched dense linear algebra kernels
available in the NVIDIA cuBLAS library [12]. These kernels, within a single API
call, are capable of solving all the linear systems concurrently, thus increasing the
exploitation of the underlying hardware. In particular, the factorization of the
matrices was performed by using the cuBLAS function cublasDgetrfBatched
and the solution of the triangular systems by using cublasDgetrsBatched. We
executed the parallel code on an NVIDIA Tesla K20 GPU with CUDA 9.0 and
compared the results with the serial version running on one core of the CPU
used in the previous experiments. Figure 2 shows the run-time in milliseconds
for the execution on GPU and√CPU as well as the speedup achieved. For each
number of evaluations, M = ( N + 1)2 , the time is the average over the runs
for all the test functions and source point sets. We see that the GPU implemen-
tation of task (e) reduces the execution time by a factor up to 20, thus yielding
a significant improvement in this phase of the method.

4 Conclusions
We discussed an improved SPH method based on the Taylor series expansion
with the Gaussian kernel function, focusing on its efficient implementation by
applying the IFGT. Numerical experiments confirmed that the accuracy of the
method can significantly benefit from the use of a very large number of points.
A GPU implementation of one of the SPH tasks was also developed, showing
that the method can take advantage from the computational power of modern
high-throughput architectures. The implementation on GPUs of the whole SPH
method will be the subject of future work.

Acknowledgments. This work has been partially supported by the INdAM-GNCS

research Project 2019 “Kernel-based approximation, multiresolution and subdivision
methods, and related applications”. This research has been also carried out within
RITA (Rete ITaliana di Approssimazione).
10 L. Antonelli et al.

References
1. Ala, G., Francomano, E.: A marching-on in time meshless kernel based solver for
full-wave electromagnetic simulation. Numer. Algorithms 62, 541–558 (2013)
2. Burkhardt, J.: https://github.com/cenit/jburkardt/tree/master/halton, https://
github.com/cenit/jburkardt/tree/master/sobol
3. Francomano, E., Hilker, F.M., Paliaga, M., Venturino, E.: An efficient method to
reconstruct invariant manifolds of saddle points. Dolomites Res. Notes Approx. 10,
25–30 (2017)
4. Francomano, E., Paliaga, M.: Highlighting numerical insights of an efficient SPH
method. Appl. Math. Comput. 339, 899–915 (2018)
5. Gingold, R.A., Monaghan, J.J.: Smoothed particle hydrodynamics: theory and
application to non-spherical stars. Mon. Not. R. Astron. Soc. 181, 375–89 (1977)
6. Greengard, L., Strain, J.: The fast Gauss transform. SIAM J. Sci. Stat. Comput.
12, 79–94 (1991)
7. Halton, J.H.: On the efficiency of certain quasi-random sequences of points in
evaluating multi-dimensional integrals. Numer. Math. 2, 84–90 (1960)
8. Liu, M.B., Liu, G.R.: Smoothed particle hydrodynamics (SPH): an overview and
recent developments. Arch. Comput. Methods Eng. 17, 25–76 (2010)
9. Lucy, L.B.: A numerical approach to the testing of the fission hypothesis. Astron.
J. 82, 1013–24 (1977)
10. Morariu, V.I., Srinivasan, B.V., Raykar V.C., Duraiswami, R., Davis, L.S.: Auto-
matic online tuning for fast Gaussian summation. In: Advances in Neural Infor-
mation Processing Systems, pp. 1113–1120 (2008). https://github.com/vmorariu/
figtree
11. NVIDIA: CUDA C Programming Guide (2019). https://docs.nvidia.com/cuda/
cuda-c-programming-guide/
12. NVIDIA: cuBLAS (2019). https://developer.nvidia.com/cublas
13. Raykar, V., Yang, C., Duraiswami, R., Gumerov, N.: Fast computation of sums of
Gaussians in high dimensions. Technical report CS-TR-4767/UMIACS-TR-2005-
69, University of Maryland, College Park, MD, USA (2005)
14. Sobol’, I.M.: On the distribution of points in a cube and the approximate evaluation
of integrals. USSR Comput. Math. Math. Phys. 7, 86–112 (1967)
15. Whaley, R.C., Petitet, A., Dongarra, J.J.: Automated empirical optimizations of
software and the ATLAS project. Parallel Comput. 27, 3–35 (2001)
 A Procedure for Laplace Transform
Inversion Based on Smoothing
Exponential-Polynomial Splines

Rosanna Campagna1(B) , Costanza Conti2 , and Salvatore Cuomo3

1
Department of Agricultural Sciences, University of Naples Federico II,
Portici, Naples, Italy
[email protected]
2
Department of Industrial Engineering, University of Florence, Florence, Italy
[email protected]
3
Department of Mathematics and Applications “R. Caccioppoli”,
University of Naples Federico II, Naples, Italy
[email protected]

Abstract. Multi-exponential decaying data are very frequent in appli-

cations and a continuous description of this type of data allows the use of
mathematical tools for data analysis such as the Laplace Transform (LT).
In this work a numerical procedure for the Laplace Transform Inversion
(LTI) of multi-exponential decaying data is proposed. It is based on
a new fitting model, that is a smoothing exponential-polynomial spline
with segments expressed in Bernstein-like bases. A numerical experiment
concerning the application of a LTI method applied to our spline model
highlights that it is very promising in the LTI of exponential decay data.

Keywords: Laplace Transform Inversion · Exponential-polynomial

spline · Multi-exponential data

1 Introduction
In several application fields inverse problems occur where, starting from observed
data, the function of interest is related to them through an integral operator as
follows:
Problem 1 (Discrete data inversion). Let be given {Fi }N
i=1 , s.t.:
 b
Fi = F (si ) + εi , i = 1, ..., N, with F (s) = K(s, t)f (t)dt
a

with K suitable kernel, εi , i = 1, ..., N unknown noises and a, b ∈ R. Compute

f , or some values of f .
The solution of the previous problem is related to the Laplace transform
inversion (LTI), bearing in mind the following definition:
c Springer Nature Switzerland AG 2020
Y. D. Sergeyev and D. E. Kvasov (Eds.): NUMTA 2019, LNCS 11973, pp. 11–18, 2020.
https://doi.org/10.1007/978-3-030-39081-5_2
12 R. Campagna et al.

Definition 1 (Laplace
∞ Transform). Let f be a given integrable function and
s0 ∈ C such that 0 e−s0 t f (t)dt < ∞. Then, for s ∈ C : Re(s) > Re(s0 ) we
∞ ∞
have 0 e−st f (t)dt < ∞. Defining C(f ) := {s ∈ C, 0 e−st f (t)dt < ∞} (the so
called region of convergence) the complex valued function F :
 ∞
F : s ∈ C(f ) → F (s) = e−st f (t)dt ∈ C
0

is the Laplace Transform (LT) of f , generally referred to as F (s) = L[f (t)].

The LT is an analytical function in its convergence region; the LTI is, in general,
expressed by complex formulas; here we focus on the so-called real inversion and
we refer to the LT main properties (see e.g. [10]):
F (s)
lim F (s) = 0, lim s · F (s) < ∞, lim = 0.
s→∞ s→∞ s→∞ s
Particularly, most of the LTs have rational or exponential behaviours. The LTI
consists in inverting the functional operator, computing f = L−1 [F ]. Starting
from the Problem 1, in this paper we suggest a procedure to solve in a stable way
the LTI from multi-exponential data. Data-driven approaches are very frequent
in several fields (see e.g. [15,16,18]) and it is common to adapt the classic fitting
models of which the analytical properties and the order of accuracy are known, to
the characteristics of the problem under examination, taking information from
the data. The LTI is a framework at the base of numerous applications, e.g.
Nuclear Magnetic Resonance (NMR) for the study of porous media [9]. There-
fore, the literature is wide and diversified so as the applications of the LT in
science and engineering problems [12]. In some works the numerical inversion of
the LT is based on collocation formulas of the LT inverse function [4], or on its
development in series of orthogonal functions, particularly polynomials [3,11,13]
or singular functions [1]. Moreover, the ill-posed nature of the problem and the
existence of general software packages for LTI (e.g. [6,7]), that are not specific for
applications, lead to use general purpose methods and related mathematical soft-
ware. Very often multi-exponential data are analysed by a Non-Negative Least
Squares (NNLS) procedure e.g. implemented in the software package CONTIN
[14]. From the above considerations, comes the idea of a fitting model that enjoys
properties similar to those of a LT function making the application of some LTI
methods meaningful. Our model is an exponential-polynomial smoothing spline
that follows the nature of the data and enjoys some of the main LT properties.
The procedure that we propose is made of two phases as synthesized below:

Procedure 1. LTI of multi-exponential data

1: input: multi-exponential data (si , Fi ), i = 1, . . . , n
2: Data-driven definition of the spline space parameters
definition of the GB-splines via the definition of Bernstein-like bases
derivation of the smoothing exponential-polynomial spline fitting the input data
3: LT inversion
4: output: Laplace transform inversion
 A Procedure for LTI Based on Smoothing Exponential-Polynomial Splines 13

The numerical experiment on LTI here presented confirms the correctness of

our ideas, even though more accurate numerical experiments on LTI are under
investigation. Open numerical issues require: (i) a detailed analysis of the sta-
bility, connected to the knots distribution; (ii) the boundary basis definition,
affecting the conditioning of the involved linear systems. These issues will be the
object of following investigations.
In Sect. 2 we report a description of the spline model used in our approach and
its derivation both from the known models in literature and from the properties
we impose to characterize it; in Sect. 3 we present a test that proves the reliability
of the model in a LTI procedure applied to exponential data. Last section refers
conclusions.

2 Smoothing Spline Approach

It is well known that a wide literature is dedicated to the splines and that the
polynomial splines are generalized in several different ways. A generalization of
polynomial splines is given by Chebyshev splines and, even further, by general-
ized splines (see [17], for example) whose local pieces belong to different spaces
with connecting conditions specified at the knots. GB-splines (i.e. basis splines
for the generalized spline spaces) are characterized by minimal support with
respect to “degree”, smoothness, and domain partition. They are largely used
in computer graphics, curve and surface reconstructions but also in image and
signal processing.
In this paper we present a special instance of generalized spline model dic-
tated by the multi-exponential decaying of the data also outside the knots and
we present a test of LTI based on this model. The results confirm that our spline
is a continuous model for data with exponential decay, that allows with success
using LTI methods to solve discrete data inversion problems.
In a previous paper [5] a cubic complete smoothing polynomial spline with an
exponential-polynomial decay only outside the knots and global C 1 -regularity
was considered. More recently (see [2]) we have defined a natural smoothing
exponential-polynomial spline enjoying C 2 regularity in a wider interval includ-
ing the knots. Here, setting a functional space reflecting the (exponential) nature
of the data, an exponential-polynomial smoothing spline (special instance of L-
splines) is defined minimizing a functional balancing fidelity to the data and
regularity request, according to the following:
Definition 2. Given a partition of the interval [a, b], Δ := {a < x1 < . . . xN <
b} and a vector (y1 , · · · , yN ) ∈ RN , a natural smoothing L-spline, related to a
fixed differential operator Ln of order n, on Δ, is the solution of the problem
N  b 
 2
min
n
(wi [u(xi ) − yi ] + λ
2
(Ln u(x)) dx , (1)
u∈H [a,b] a
i=1

with (w1 , . . . , wN ) non zero weights, λ ∈ R a regularization parameter and

H n [a, b] a suitable Hilbert space.
14 R. Campagna et al.

The first issue when dealing with a natural smoothing L-spline is given by
the selection of the differential operator while the second one is the construction
of the GB-splines for the corresponding spline space in order to derive a stable
solution of (1) expressed as
N

s(x) = cj ϕj (x). (2)
j=1

With the aim to solve the Problem 1, in our model we assume like knots
the xi ∈ Δ, i = 1, . . . , N , localized at the abscissae of the experimental data.
We work with generalized splines with segments in the spaces of exponential-
polynomial functions,

E2 := span{e−αx , x e−αx } and E4 := span{eαx , x eαx , e−αx , x e−αx }, α ∈ R+ ,

and construct the associated GB-splines using Bernstein-like basis. The corre-
sponding spline model is a Smoothing Generalized B-Spline on Bernestein basis,
SGBSB for short. A short algorithmic description of our model construction fol-
lows (see [2] for details).

Procedure 2. SGBSB
1: Definition of the differential operator and the related null-spaces E2 of L2 and E4
of L∗2 L2 ,
2: Definition of the GB-spline basis functions, {ϕ (x)}=1,...,N , in (2):

(a) the support of each ϕ is compact;

(b) ϕ ∈ C 2 [a, b];
(c) ϕ ∈ E4 for x ∈ (xi , xi+1 ), i = 1, . . . , N − 1;
(d) ϕ ∈ E2 , for x ∈ (a, x1 ) or x ∈ (xN , b);
(e) ϕ , j = 1, . . . , N , are positive and bell shaped functions.

3: Representation of each GB-spline in Bernstein-like basis:

3.1 : definition of regular basis functions:


3
ϕ (x)|[xj ,xj+1 ] = γ,j,i B̃i (x − xj ), j =  − 2, · · · ,  + 1,  = 3, · · · , N − 2
i=0

with B̃i Bernstein-like basis functions of E4 , i = 0, ..., 3, and C 2 regularity at

the internal points x−1 , x , x+1 ;
3.2 : definition of boundary left and right basis functions:
like the regular ones but with pieces in E2 , outside the knots interval, in (a, x1 ]
and in [xn , b) respectively.
 A Procedure for LTI Based on Smoothing Exponential-Polynomial Splines 15

Table 1. LTI results. The columns refer, from left to right: evaluation points (t),
computed values by applying GS to the real function F (fF ) and to the SGBSB s (fs );
pointwise errros between fF and fs (|fF − fs |), at the t points.

t fF fs |fF − fs |
2.0000e+00 3.9920e−01 3.9925e−01 4.6485e−05
2.5000e+00 2.5880e−01 2.5881e−01 5.7951e−06
3.0000e+00 1.6189e−01 1.6187e−01 2.1254e−05
3.5000e+00 9.9504e−02 9.9504e−02 1.0816e−07
4.0000e+00 6.0095e−02 6.0115e−02 2.0280e−05
4.5000e+00 3.5306e−02 3.5294e−02 1.2250e−05
5.0000e+00 1.9735e−02 1.9776e−02 4.0631e−05
5.5000e+00 9.9866e−03 1.0065e−02 7.8389e−05
6.0000e+00 3.9346e−03 3.9392e−03 4.5941e−06
6.5000e+00 2.4117e−04 2.4450e−04 3.3348e−06
7.0000e+00 −1.9429e−03 −1.9464e−03 3.5076e−06
7.5000e+00 −3.1609e−03 −3.1904e−03 2.9417e−05
8.0000e+00 −3.7633e−03 −3.8100e−03 4.6736e−05
8.5000e+00 −3.9771e−03 −3.9692e−03 7.9306e−06
9.0000e+00 −3.9505e−03 −3.4085e−03 5.4200e−04
9.5000e+00 −3.7799e−03 −2.4630e−03 1.3168e−03
1.0000e+01 −3.5281e−03 −1.4930e−03 2.0351e−03

Fig. 1. Curves of the computed LTI, by applying the GS algorithm to F (fF ) and to
SGBSB (fs ). The results are comparable and overlapping and they follow the behaviour
of the real inverse function f described by the red curve (‘*-’). The abscissae are referred
in the first column of the Table 1. (Color figure online)
16 R. Campagna et al.

3 Numerical Results
In this section we present some results about a test of LTI. Let be given N = 60
samples (xi , yi )N
i=1 , uniformly distributed in [0.05, 3]. The test was carried out
with MATLAB R2018a software on a Intel(R) Core(TM) i5, 1.8 GHz processor.
We assume that yi = F (xi ) with F the following LT:

F (x) = e−x /(1 + x),

whose LT inverse function is:

f (t) = e−(t−1) u(t − 1), t > 1, u(t) step function. (3)

The described SGBSB, s, requires the definition of some model parameters, that
is a, b, α (see [2] for details). In this test we set a = −1, b = 3.5, and we compute
α = 1.6 through a nonlinear least-squares regression of the data. The Gaver-
Stehfest (GS) algorithm for LTI [8,19] is used to compute an approximation of
f at the points

tj ∈ [2, 10], t1 = 2, tj+1 = tj + 0.5, j = 1, ..., 16.

The Matlab code gavsteh.m is available at the Mathworks File Exchange; the
algorithm for the SGBSB is implemented in a proprietary Matlab code, available
from the authors.
Numerical results are reported in Table 1 whose columns, from left to right,
contain:

1. t: the evaluation points, tj , j = 1, . . . , 17

2. fF : the computed values by applying GS to the real function F ;
3. fs : the computed values by applying GS to the SGBSB;
4. |fF − fs |: the computed pointwise erros between fF and fs , at the t points.

The corresponding curves of the computed values, together with the graph of f
in (3), are in Fig. 1.
We remark that the GS algorithm requires, as input parameters, the evalua-
tion point t, and the function to be inverted; then, it dynamically evaluates the
input function at suitable points, depending, among the other method parame-
ters, on t. In the test here presented, we evaluated also the approximation error
between F and the spline model, inverted in its place, and we observe that the
maximum absolute approximation error on the LT:

F − s∞ = 9.9464 × 10−4

is at most of the same order of the maximum error on the computed solution,
as confirmed by the values referred in the last column of Table 1.
 A Procedure for LTI Based on Smoothing Exponential-Polynomial Splines 17

4 Conclusion
In this work, we present the results of a LTI method, when applied to a natural
smoothing exponential-polynomial spline modelling exponential decay data. This
scenario is found in many applications. A detailed analysis of the parameters and
the sensitivity of the model are under investigation and will be object of future
studies.

Acknowledgements. The authors are members of the INdAM Research group GNCS
and of the Research ITalian network on Approximation (RITA).

References
1. Bertero, M., Brianzi, P., Pike, E.R.: On the recovery and resolution of exponential
relaxation rates from experimental data: Laplace transform inversion in weighted
spaces. Inverse Probl. 1, 1–15 (1985)
2. Campagna, R., Conti, C., Cuomo, S.: Smoothing exponential-polynomial splines
for multiexponential decay data. Dolomites Res. Notes Approx. 12, 86–100 (2019)
3. Campagna, R., D’Amore, L., Murli, A.: An efficient algorithm for regularization of
Laplace transform inversion in real case. J. Comput. Appl. Math. 210(1), 84–98
(2007)
4. Cuomo, S., D’Amore, L., Murli, A., Rizzardi, M.: Computation of the inverse
Laplace transform based on a collocation method which uses only real values. J.
Comput. Appl. Math. 198(1), 98–115 (2007)
5. D’Amore, L., Campagna, R., Galletti, A., Marcellino, L., Murli, A.: A smoothing
spline that approximates Laplace transform functions only known on measurements
on the real axis. Inverse Probl. 28(2), 025007 (2012)
6. D’Amore, L., Campagna, R., Mele, V., Murli, A.: ReLaTIve. An Ansi C90 software
package for the real Laplace transform inversion. Numer. Algorithms 63(1), 187–
211 (2013)
7. D’Amore, L., Campagna, R., Mele, V., Murli, A.: Algorithm 946: ReLIADiff-A
C++ software package for real Laplace transform inversion based on algorithmic
differentiation. ACM Trans. Math. Softw. 40(4), 31:1–31:20 (2014)
8. D’Amore, L., Mele, V., Campagna, R.: Quality assurance of Gaver’s formula for
multi-precision Laplace transform inversion in real case. Inverse Probl. Sci. Eng.
26(4), 553–580 (2018)
9. Galvosas, P., Callaghan, P.T.: Multi-dimensional inverse Laplace spectroscopy in
the NMR of porous media. C.R. Phys. 11(2), 172–180 (2010). Multiscale NMR
and relaxation
10. Henrici, P.: Applied and Computational Complex Analysis, Volume 1: Power Series
Integration Conformal Mapping Location of Zero (1988)
11. Miller, M.K., Guy, W.T.: Numerical inversion of the Laplace transform by use of
Jacobi polynomials. SIAM J. Numer. Anal. 3(4), 624–635 (1966)
12. Naeeni, M.R., Campagna, R., Eskandari-Ghadi, M., Ardalan, A.A.: Performance
comparison of numerical inversion methods for Laplace and Hankel integral trans-
forms in engineering problems. Appl. Math. Comput. 250, 759–775 (2015)
13. Papoulis, A.: A new method of inversion of the Laplace transform. Q. Appl. Math.
14(4), 405–414 (1957)
18 R. Campagna et al.

14. Provencher, S.W.: Contin: a general purpose constrained regularization program

for inverting noisy linear algebraic and integral equations. Comput. Phys. Com-
mun. 27(3), 229–242 (1982)
15. Romano, A., Campagna, R., Masi, P., Cuomo, S., Toraldo, G.: Data-driven
approaches to predict states in a food technology case study. In: 2018 IEEE 4th
International Forum on Research and Technology for Society and Industry (RTSI),
pp. 1–5 (2018)
16. Romano, A., Cavella, S., Toraldo, G., Masi, P.: 2D structural imaging study of
bubble evolution during leavening. Food Res. Int. 50(1), 324–329 (2013)
17. Schumaker, L.: Spline Functions: Basic Theory, 3rd edn. Cambridge Mathematical
Library, Cambridge University Press, Cambridge (2007)
18. Severino, G., D’Urso, G., Scarfato, M., Toraldo, G.: The IoT as a tool to combine
the scheduling of the irrigation with the geostatistics of the soils. Future Gener.
Comput. Syst. 82, 268–273 (2018)
19. Stehfest, H.: Algorithm 368: numerical inversion of Laplace transforms. Commun.
ACM 13(1), 47–49 (1970)
 An Adaptive Refinement Scheme
for Radial Basis Function Collocation

Roberto Cavoretto(B) and Alessandra De Rossi

Department of Mathematics “Giuseppe Peano”, University of Torino,

Via Carlo Alberto 10, 10123 Turin, Italy
{roberto.cavoretto,alessandra.derossi}@unito.it

Abstract. In this paper we present an adaptive refinement algorithm

for solving elliptic partial differential equations via a radial basis function
(RBF) collocation method. The adaptive scheme is based on the use of
an error indicator, which is characterized by the comparison of two RBF
collocation solutions evaluated on a coarser set and a finer one. This
estimate allows us to detect the domain parts that need to be refined by
adding points in the selected areas. Numerical results support our study
and point out the effectiveness of our algorithm.

Keywords: Meshfree methods · Adaptive algorithms · Refinement

techniques · Elliptic PDEs

1 Introduction

In this paper we present a new adaptive refinement scheme for solving elliptic
partial differential equations (PDEs). Our adaptive algorithm is applied to a non-
symmetric radial basis function (RBF) collocation method, which was originally
proposed by Kansa [5]. This approach has engendered a large number of works,
mainly by scientists from several different areas of science and engineering (see
e.g. [1–4,7] and references therein). Basically, the adaptive scheme we propose
is based on the use of an error indicator characterized by the comparison of two
approximate RBF collocation solutions, which are evaluated on a coarser set and
a finer one. This estimate allows us to identify the domain parts that need to be
refined by adding points in the selected areas. In our numerical experiments we
show the efficacy of our refinement algorithm, which is tested by modeling some
Poisson-type problems.
The paper is organized as follows. In Sect. 2 we review some basic informa-
tion on Kansa’s collocation method, which is applied to elliptic PDEs. Section 3
describes the adaptive refinement algorithm. In Sect. 4 we show some numerical
results carried out to illustrate the performance of the adaptive scheme. Section 5
contains conclusions.

c Springer Nature Switzerland AG 2020

Y. D. Sergeyev and D. E. Kvasov (Eds.): NUMTA 2019, LNCS 11973, pp. 19–26, 2020.
https://doi.org/10.1007/978-3-030-39081-5_3
20 R. Cavoretto and A. De Rossi

2 Nonsymmetric RBF Collocation

Given a domain Ω ⊂ Rd , we consider a (time independent) elliptic PDE along
with its boundary conditions

Lu(x) = f (x), x ∈ Ω,
(1)
Bu(x) = g(x), x ∈ ∂Ω,

where L is a linear elliptic partial differential operator and B is a linear boundary

operator.
For Kansa’s collocation method we choose to represent the approximate solu-
tion û by a RBF expansion analogous to that used in the field of RBF interpo-
lation [3], i.e. û is expressed as a linear combination of basis functions


N
û(x) = cj φε (||x − zj ||2 ), (2)
j=1

where cj is an unknown real coefficient, || · ||2 denotes the Euclidean norm, and
φε : R≥0 → R is some RBF depending on a shape parameter ε > 0 such that

φε (||x − z||2 ) = φ(ε||x − z||2 ), ∀ x, z ∈ Ω.

In Table 1 we list some examples of popular globally supported RBFs, which are
commonly used for solving PDEs (see [3] for details).

Table 1. Some examples of popular RBFs.

RBF φε (r)
2 2
Gaussian (GA) e−ε r

Inverse MultiQuadric (IMQ) (1 + ε2 r2 )−1/2

MultiQuadric (MQ) (1 + ε2 r2 )1/2

In (2) we can distinguish between the set X = {x1 , . . . , xN } of collocation

points and the set Z = {z1 , . . . , zN } of centers. Additionally, for the sake of
convenience we split the set X into a subset XI of interior points and a subset
XB of boundary points, so that X = XI ∪ XB .
Matching the PDE and the boundary conditions in (1) at the collocation
points X, we obtain a linear system of equations

Φc = v,

where c = (c1 , . . . , cN )T is the vector of coefficients, v = (v1 , . . . , vN )T is the

vector of entries

f (xi ), xi ∈ XI ,
vi =
g(xi ), xi ∈ XB .
 Adaptive Refinement Scheme for RBF Collocation 21

and Φ ∈ RN ×N is the collocation matrix

 
ΦL
Φ= . (3)
ΦB

The two blocks in (3) are defined as

(ΦL )ij = Lφε (||xi − zj ||2 ), xi ∈ XI , zj ∈ Z,

(ΦB )ij = Bφε (||xi − zj ||2 ), xi ∈ XB , zj ∈ Z,

Since the collocation matrix (3) may be singular for certain configurations of
the centers zj , it follows that the nonsymmetric collocation method cannot be
well-posed for arbitrary center locations. However, it is possible to find sufficient
conditions on the centers so that invertibility of Kansa’s matrix is ensured. For a
more detailed analysis of Kansa’s collocation method and some variations thereof
derived from applications, see e.g. [3,6] and references therein.

3 Adaptive Refinement Algorithm

In this section we present the adaptive algorithm proposed to solve time inde-
pendent PDE problems by Kansa’s approach.
Step 1. We define two sets, XN (0) and XN (0) , of collocation points and two
1 2
(0) (0)
sets, ZN (0) and ZN (0) of centers. Each couple of sets has size N1 and N2 ,
1 2
(0) (0)
respectively, with N1 < N2 and the symbol (0) identifying the initial iteration.
We then split the related sets as follows:

– XN1 (0) = XI,N1 (0) ∪ XB,N1 (0) and XN2 (0) = XI,N2 (0) ∪ XB,N2 (0) are sets of
interior and boundary collocation points, respectively;
– ZN1 (0) = ZI,N1 (0) ∪ ZB,N
A
(0) and ZN2 (0) = ZI,N2 (0) ∪ ZB,N (0) are sets of
A
1 2
interior and additional boundary centers, respectively.

Here we assume that XI,Ni (0) = ZI,Ni (0) , with i = 1, 2, while the set ZB,NA
(0) of
i
centers is taken outside the domain Ω as suggested in [3]. However, we note that
it is also possible to consider only a set of data as collocation points and centers.
Step 2. For k = 0, 1, . . ., we iteratively find two collocation solutions of the form
(k) (k)
(2), called ûN1 (k) and ûN2 (k) , which are respectively computed on N1 and N2
collocation points and centers.
Step 3. We compare the two approximate RBF solutions by evaluating error on
(k)
the (coarser) set containing N1 points, i.e.

|ûN2 (k) (xi ) − ûN1 (k) (xi )|, xi ∈ XN1 (k) .

(k)
Observe that here we assume that the solution computed on N2 discretization
(k)
points gives more accurate results than the ones obtained with only N1 points.
22 R. Cavoretto and A. De Rossi

Step 4. After fixing a tolerance tol, we determine all points xi ∈ XN1 (k) such
that

|ûN2 (k) (xi ) − ûN1 (k) (xi )| > tol. (4)

Step 5. In order to refine the distribution of discretization points, we compute

the separation distance
1
qXN (k)
= min ||xi − xj ||2 , xi ∈ XN1 (k) . (5)
1 2 i=j
Step 6. For k = 0, 1, . . . we update the two sets XN1 (k+1) and XN2 (k+1) of collo-
cation points (and accordingly the corresponding sets ZN1 (k+1) and ZN2 (k+1) of
centers) as follows. For each point xi ∈ XN1 (k) , such that the condition (4) is
satisfied, we add to xi :

– four points (the blue circles depicted in the left frame of Fig. 1), thus creating
the set XN1 (k+1) ;
– eight points (the red squares shown in the right frame of Fig. 1), thus gener-
ating the set XN2 (k+1) .

In both cases the new points are given by properly either adding or subtracting
the value of (5) to the components of xi . Furthermore, we remark that in the
illustrative example of Fig. 1 the point xi is marked by a black cross, while the
new sets are such that XN1 (k) ⊂ XN2 (k) , for k = 1, 2, . . ..
Step 7. The iterative process stops when having no points anymore which fulfill
the condition (4), giving the set XN2 (k∗ ) back. Note that k ∗ is here used to denote
the last algorithm iteration.

1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6

0.5 0.5
2

2
x

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x1 x1

Fig. 1. Illustrative example of refinement for sets XN1 (k) (left) and XN2 (k) (right) in
the adaptive algorithm. (Color figure online)
 Adaptive Refinement Scheme for RBF Collocation 23

4 Numerical Results
In this section we summarize the results derived from application of our adaptive
refinement algorithm, which is implemented in Matlab environment. All the
results are carried out on a laptop with an Intel(R) Core(TM) i7-6500U CPU
2.50 GHz processor and 8 GB RAM.
In the following we restrict our attention on solving some elliptic PDE prob-
lems via the nonsymmetric RBF collocation method. In particular, in (1) we
consider a few Poisson-type problems, taking the Laplace operator L = −Δ and
assuming Dirichlet boundary conditions. Hence, the PDE problem in (1) can be
defined as follows:
−Δu(x) = f (x), x ∈ Ω,
(6)
u(x) = g(x), x ∈ ∂Ω.

Then, we focus on two test problems of the form (6) defined on the domain
Ω = [0, 1]2 . The exact solutions of such Poisson problems are

P1 : u1 (x1 , x2 ) = sin(x1 + 2x22 ) − sin(2x21 + (x2 − 0.5)2 ),

1  4 1
P2 : u2 (x1 , x2 ) = x2 cos(4x21 + x22 − 1) + x1 .
2 4
A graphical representation of these analytic solutions is shown in Fig. 2.

Fig. 2. Graphs of exact solutions u1 (left) and u2 (right) of Poisson problems.

In our numerical tests we analyze the performance of the adaptive refine-

ment strategy applied to Kansa’s collocation method by using globally supported
RBFs such as MQ, IMQ and GA (see Table 1). We remark that the use of com-
pactly supported RBFs is also possible and effective but our tests showed that
the most accurate results were obtained with quite large supports. So the use of
compactly supported functions does not provide particular benefits w.r.t. glob-
ally supported RBFs. For this reason and for the sake of brevity, we do not
consider this case in the present paper.
The two starting sets defined in Sect. 3 consist of N1 (0) = 289 and N2 (0) =
1089 grid collocation points, while the tolerance in (4) is given by tol = 10−4 .
24 R. Cavoretto and A. De Rossi

In particular, in order to measure the quality of our results, we compute the

Maximum Absolute Error (MAE), i.e.,
MAE = max |u(yi ) − û(yi )|.
1≤i≤Neval

which is evaluated on a grid of Neval = 40 × 40 evaluation points. Moreover,

in regards to the efficiency of the adaptive scheme, we report the CPU times
computed in seconds.
In Tables 2, 3 and 4 we present the results obtained, also indicating the
final number Nf in of collocation points required to achieve the fixed tolerance.
Further, as an example, for brevity in only one case for each test problem,
we report the “refined grids” after applying iteratively the adaptive algorithm.
More precisely, in Fig. 3 we graphically represent the final distribution of points
obtained after the last algorithm iteration by applying: the MQ-RBF with ε = 4
for the test problem P1 (left), and the IMQ-RBF with ε = 3 for the test problem
P2 (right).

Table 2. MQ, ε = 4, tol = 10−4 .

Test problem Nf in MAE CPU time

P1 856 8.28 × 10−5 1.2
P2 798 2.66 × 10−4 1.7

Table 3. IMQ, ε = 3, tol = 10−4 .

Test problem Nf in MAE CPU time

P1 1001 3.10 × 10−5 1.7
P2 808 1.10 × 10−4 1.8

Table 4. GA, ε = 9, tol = 10−4 .

Test problem Nf in MAE CPU time

−4
P1 898 3.49 × 10 2.1
P2 843 2.08 × 10−4 1.9

Analyzing the numerical results, we can observe as the adaptive algorithm

allows us to increase the number of points in the regions where the solution is not
accurate enough. From the tables we note as MQ and IMQ give more accurate
results than GA. In fact, though the number of points required to satisfy the fixed
tolerance is quite similar for all used RBFs, we can remark a greater instability
of GA that needs a larger value of ε to work effectively. Finally, in terms of
computational efficiency the algorithm converges in few seconds in each of the
tests carried out.
 Adaptive Refinement Scheme for RBF Collocation 25

1.2 1.2

1 1

0.8 0.8

0.6 0.6
2

2
x

x
0.4 0.4

0.2 0.2

0 0

−0.2 −0.2
−0.2 0 0.2 0.4 0.6 0.8 1 1.2 −0.2 0 0.2 0.4 0.6 0.8 1 1.2
x x
1 1

Fig. 3. Final distribution of points obtained after applying the refinement process with
MQ, ε = 4, for problem P1 (left) and with IMQ, ε = 3, for problem P2.

5 Conclusions
In this work we presented an adaptive refinement algorithm to solve time inde-
pendent elliptic PDEs. This refinement strategy is tested on a nonsymmetric
RBF collocation scheme, known as Kansa’s method. More precisely, here we pro-
posed an adaptive approach based on a refinement technique, which consisted
in comparing two collocation solutions computed on a coarser set of collocation
points and a finer one. This process allowed us to detect the domain areas where
it is necessary to adaptively add points, thus enhancing accuracy of the method.
Numerical results supported this study by showing the algorithm performance
on some Poisson-type problems.
As future work we are interested in investigating and possibly extending
our adaptive schemes to hyperbolic and parabolic PDE problems. Moreover, we
are currently working on the optimal selection of the RBF shape parameter in
collocation schemes. However, this is out of the scopes of the present paper and
it will be dealt with in forthcoming works.

Acknowledgments. The authors sincerely thank the two anonymous referees for
carefully reviewing this paper and for their valuable comments and suggestions. They
acknowledge support from the Department of Mathematics “Giuseppe Peano” of the
University of Torino via Project 2018 “Algebra, geometry and numerical analysis”
and Project 2019 “Mathematics for applications”. Moreover, this work was partially
supported by INdAM – GNCS Project 2019 “Kernel-based approximation, multires-
olution and subdivision methods and related applications”. This research has been
accomplished within RITA (Research ITalian network on Approximation).

References
1. Cavoretto, R., De Rossi, A.: Adaptive meshless refinement schemes for RBF-PUM
collocation. Appl. Math. Lett. 90, 131–138 (2019)
2. Chen, W., Fu, Z.-J., Chen, C.S.: Recent Advances on Radial Basis Function Collo-
cation Methods. Springer Briefs in Applied Science and Technology. Springer, Berlin
(2014). https://doi.org/10.1007/978-3-642-39572-7
26 R. Cavoretto and A. De Rossi

3. Fasshauer, G.E.: Meshfree Approximation Methods with Matlab. Interdisciplinary

Mathematical Sciences, vol. 6. World Scientific Publishing Co., Singapore (2007)
4. Libre, N.A., Emdadi, A., Kansa, E.J., Shekarchi, M., Rahimian, M.: A fast adap-
tive wavelet scheme in RBF collocation for nearly singular potential PDEs. CMES
Comput. Model. Eng. Sci. 38, 263–284 (2008)
5. Kansa, E.J.: Multiquadrics - a scattered data approximation scheme with applica-
tions to computational fluid-dynamics - II solutions to parabolic, hyperbolic and
elliptic partial differential equations. Comput. Math. Appl. 19, 147–161 (1990)
6. Schaback, R.: Convergence of unsymmetric kernel-based meshless collocation meth-
ods. SIAM J. Numer. Anal. 45, 333–351 (2007)
7. Sarra, S.A., Kansa, E.J.: Multiquadric radial basis function approximation methods
for the numerical solution of partial differential equations. In: Advances in Compu-
tational Mechanics, vol. 2. Tech Science Press, Encino (2009)
 A 3D Efficient Procedure for Shepard
Interpolants on Tetrahedra

Roberto Cavoretto1 , Alessandra De Rossi1(B) ,

Francesco Dell’Accio2 , and Filomena Di Tommaso2
1
Department of Mathematics “Giuseppe Peano”, University of Turin,
Via Carlo Alberto 10, 10123 Turin, Italy
{roberto.cavoretto,alessandra.derossi}@unito.it
2
Department of Mathematics and Computer Science, University of Calabria,
via P. Bucci, Cubo 30A, 87036 Rende (CS), Italy
[email protected], [email protected]

Abstract. The need of scattered data interpolation methods in the mul-

tivariate framework and, in particular, in the trivariate case, motivates
the generalization of the fast algorithm for triangular Shepard method.
A block-based partitioning structure procedure was already applied to
make the method very fast in the bivariate setting. Here the searching
algorithm is extended, it allows to partition the domain and nodes in
cubic blocks and to find the nearest neighbor points that need to be
used in the tetrahedral Shepard interpolation.

Keywords: Scattered data interpolation · Tetrahedral Shepard

operator · Fast algorithms · Approximation algorithms

1 Introduction
Given a set of values of a function f at certain scattered nodes Xn = {x1 , . . . , xn }
in a compact convex domain Ω ⊂ R2 , the triangular Shepard method [8] can be
applied efficiently to interpolate the target function f : Ω → R. In [5] we pro-
posed a triangular Shepard method which combines triangle-based basis func-
tions with linear combinations of the values f (xi ) at the vertices of the trian-
gles. Moreover, the triangulation can be found in an efficient way by reducing the
number of triangles. The triangulation considered is called compact triangulation
and it allows the triangles to overlap or being disjoint. These triangulations are
determined by minimizing the bound of the error of the linear interpolant on the
vertices of the triangle, chosen in a set of nearby nodes. For these triangulations
a block-based partitioning structure procedure was presented in [3] to make the
method very fast, since the vertices of the triangles must be chosen in a set of
nearby nodes.
In recent years an increasing attention to the multivariate framework was
given. For this reason we propose in this paper a generalization to the 3D setting.
More precisely, we propose a fast searching procedure to apply to the tetrahedral
c Springer Nature Switzerland AG 2020
Y. D. Sergeyev and D. E. Kvasov (Eds.): NUMTA 2019, LNCS 11973, pp. 27–34, 2020.
https://doi.org/10.1007/978-3-030-39081-5_4
28 R. Cavoretto et al.

Shepard interpolation. It allows to partitioning the 3D domain and nodes in

cubic blocks and to find the nearest neighbor points to compute the Shepard
interpolant on tetrahedra. Similar algorithms were also analized in [2] in the
context of trivariate partition of unity methods combined with the use of local
radial kernels.
The paper is organized as follows. In Sect. 2 the tetrahedral Shepard method
for trivariate interpolation is recalled. In Sect. 3 we give a pseudo-code of the
complete interpolation algorithm, presenting the procedures used to identify and
search the nearest neighbor points in the 3D interpolation scheme. In Sect. 4 we
show some numerical experiments obtained to illustrate the performance of our
tetrahedral Shepard algorithm. Finally, Sect. 5 contains conclusions and future
work.

2 Tetrahedral Shepard Interpolant

Let be Xn = {x1 , . . . , xn } a set of data points or nodes of R3 with an associated
set of function data Fn = {f1 , . . . , fn } and H = {h1 , . . . , hm } a set of tetrahedra
with vertices in Xn . Let us denote by Wj = {xj1 , xj2 , xj3 , xj4 } the set of vertices
of hj , j = 1, . . . , m. Moreover, we assume that the set {Wj }j=1,...,m constitutes
a cover of Xn , that is
m
Wj = X n .
j=1

We can associate to each tetrahedra hj the set of barycentric coordinates of a

point x ∈ R3 , that is
W (x, xj2 , xj3 , xj4 ) W (xj1 , x, xj3 , xj4 )
μj,j1 (x) = , μj,j2 (x) = ,
W (xj1 , xj2 , xj3 , xj4 ) W (xj1 , xj2 , xj3 , xj4 )

W (xj1 , xj2 , x, xj4 ) W (xj1 , xj2 , xj3 , x)

μj,j3 (x) = , μj,j4 (x) = ,
W (xj1 , xj2 , xj3 , xj4 ) W (xj1 , xj2 , xj3 , xj4 )

where W (x, y, v, z) denotes 16 the signed volume of the tetrahedra hj . The linear
polynomial λj (x) which interpolates the data at the vertices of the tetrahedra
hj can be expressed in terms of barycentric coordinates in the following form


4
λj (x) = μj,jk (x) fjk , j = 1, . . . , m. (1)
k=1

The tetrahedral basis functions are a normalization of the product of the inverse
distances from the vertices of the tetrahedra hj


4
−ν
||x − xj ||
=1
βν,j (x) = , j = 1, . . . , m, ν > 0, (2)

m 
4
−ν
||x − xk ||
k=1 =1
 A 3D Efficient Procedure for Shepard Interpolants 29

where || · || is the Euclidean norm. The tetrahedral Shepard method is defined

by
m

Tν [f ] (x) = βν,j (x) λj (x) . (3)
j=1

Tetrahedral basis functions form a partition of unity, as the triangular Shepard

ones, and allow the interpolation of functional and derivative values. In fact,
the following results hold, the proofs can be easily obtained in analogy with [5,
Proposition 2.1].
Proposition 1. The tetrahedral basis function βν,j (x) and its gradient (that
exists for ν > 1) vanish at all nodes xi ∈ Xn that are not a vertex of the
corresponding tetrahedron hj . That is,
βν,j (xi ) = 0, (4)
∇βν,j (xi ) = 0, ν > 1, (5)
for any j = 1, . . . , m and i ∈
/ {j1 , j2 , j3 , j4 }. Moreover, they form a partition of
unity, that is
m
βν,j (x) = 1 (6)
j=1

and consequently, for each i = 1, . . . , n,


βν,j (xi ) = 1, (7)
j∈Ji

∇βν,j (xi ) = 0, ν > 1, (8)
j∈Ji
 
where Ji = k ∈ {1, . . . , m} : i ∈ {k1 , k2 , k3 , k4 } is the set of tetrahedra which
have xi as a vertex.
These properties imply that the operator Tν satisfies the following ones, see [4]
for details.
Proposition 2. The operator Tν is an interpolation operator, that is,
Tν [f ](xi ) = fi , i = 1, . . . , n,
and reproduces polynomials up to the degree 1.
The procedure to select the compact 3D-triangulation (by tetrahedra) of the
node set Xn strongly affects the results of the analysis of the convergence of the
operator Tν [f ] (x).
In order to determine the approximation order of the tetrahedral operator,
we denote by Ω ⊂ R3 a compact convex domain containing Xn and by C 1,1 (Ω)
the class of differentiable functions f : Ω → R whose partial derivative of order
1 are Lipschitz-continuous, equipped with the seminorm
 μ
D f (u) − Dμ f (v)
f 1,1 = sup : u, v ∈ Ω, u = v, μ = 1 . (9)
u − v
30 R. Cavoretto et al.

We also denote by ek, = xjk − xj , with k,  = 1, 2, 3, 4, the edge vectors of the
tetrahedron hj . Then, the following result holds (for the proof see [4]).
Proposition 3. Let f ∈ C 1,1 (Ω) and hj ∈ H a tetrahedron of vertices xj1 , xj2 ,
xj3 , xj4 . Then, for all x ∈ Ω we have


2 27
|f (x) − λj (x)| ≤ ||f ||1,1 3 ||x − xj1 ||2 + Cj kj ||x − xj1 ||2 , (10)
2

where kj = maxk,=1,2,3,4 ek,  and Cj is given by the ratio between the max-
imum edge and the volume, and then is a constant which depends only on the
shape of the tetrahedron hj . The error bound is valid for any vertex.

3 Trivariate Shepard Interpolation Algorithm on

Tetrahedra

In this section we present the interpolation algorithm, which performs the tetra-
hedral Shepard method (3) using the block-based partitioning structure and the
associated searching procedure. Here we consider Ω = [0, 1]3 .
INPUTS: n, number of data; Xn = {x1 , . . . , xn }, set of data points; Fn =
{f1 , . . . , fn }, set of data values; ne , number of evaluation points; nw , localiz-
ing parameter.
OUTPUTS: Ene = {Tν [f ](z1 ), . . . , Tν [f ](zne )}, set of approximated values.
Step 1: Generate a set Zne = {z1 , . . . , zne } ⊆ Ω of evaluation points.
Step 2: For each point xi , i = 1, . . . , n, construct a neighborhood of radius
√   
3 n 1/3
δ= , with d = .
d 8

where the value of d is suitably chosen extending the definition contained in [2].
This phase performs the localization.
Step 3: Compute the number b of blocks (along one side of the unit cube Ω)
defined by  
1
b= .
δ
In this way we get the side of each cubic block is equal to the neighborhood
radius. This choice enables us to examine in the searching procedure only a
small number of blocks, so to reduce the computational cost as compared to
the most advanced searching techniques, as for instance the kd-trees [10]. The
benefit is proved by the fact that this searching process is carried out in constant
time, i.e. O(1). Further, in this partitioning phase we number the cube-shaped
blocks from 1 to b3 .
 A 3D Efficient Procedure for Shepard Interpolants 31

Step 4: Build the partitioning structure on the domain Ω and split the set
Xn of interpolation nodes in b3 cubic blocks. Here we are able to obtain a fast
searching procedure to detect the interpolation points nearest to each of nodes.
Step 5: For each neighborhood or point (i.e., the neighborhood centre), solve
the containing query and the range search problems to detect all nodes Xnk ,
k = 1, . . . , b3 , belonging to the k-th block and its twenty-six neighboring blocks
(or less in case the block lies on the boundary). This is performed by repeatedly
using a quicksort routine.
Step 6: For each data point xi ∈ Xn , fix its nw nearest neighbors N (xi ) ⊂ Xn .
Among the
nw (nw − 1) (nw − 2)
6
tetrahedra with a vertex in xi , name it xj1 and other three vertices in N (xi ),
choose the one which locally reduces the bound for the error of the local linear
interpolant

2 27 kj3
3 ||x − xj1 ||2 + kj ||x − xj1 ||2 .
2 W (xj1 , xj2 , xj3 , xj4 )

Step 7: Compute the local basis function βν,j (z), j = 1, . . . , m, at each evalu-
ation point z ∈ Zne .
Step 8: Compute the linear interpolants λj (z), j = 1, . . . , m, at each evaluation
point z ∈ Zne .
Step 9: Apply the tetrahedral Shepard method (3) and evaluate the trivariate
interpolant at the evaluation points z ∈ Zne .

4 Numerical Results

We present here accuracy and efficiency results of the trivariate interpolation

algorithm proposed. The algorithm was implemented in Matlab. All the numer-
ical experiments have been carried out on a laptop with an Intel(R) Core i7
6500U CPU 2.50 GHz processor and 8.00 GB RAM.
In the following we analize the results obtained about several tests car-
ried out. We solved very large interpolation problems by means of the tetra-
hedral Shepard method (3). To do this we considered two different distribu-
tions of irregularly distributed (or scattered) nodes contained in the unit cube
Ω = [0, 1]3 ⊂ R3 , and taking a number n of interpolation nodes that varies
from 2 500 to 20 000. More precisely, as interpolation nodes we focus on a few
sets of uniformly random Halton points generated through the Matlab pro-
gram haltonseq.m [6], and pseudo-random points obtained by using the rand
Matlab command. In addition, the interpolation errors are computed on a grid
consisting of ne = 21 × 21 × 21 evaluation points, while as localizing parameter
we fix the value nw = 13 and ν = 2.
In the various experiments we discuss the performance of our interpolation
algorithm assuming the data values are given by the following two trivariate test
32 R. Cavoretto et al.

functions:
f1 (x1 , x2 , x3 ) = cos(6x3 )(1.25 + cos(5.4x2 ))/(6 + 6(3x1 − 1)2 ),
f2 (x1 , x2 , x3 ) = exp(−81/16((x1 − 0.5)2 + (x2 − 0.5)2 + (x3 − 0.5)2 ))/3.
These functions are usually used to test and validate new approximation
methods and algorithms (see e.g. [9]).
As a measure of the accuracy of our results, we compute the Maximum Abso-
lute Error (MAE) and the Root Mean Square Error (RMSE), whose formulas
are respectively given by
MAE = ||f − Tν [f ]||∞ = max |f (zi ) − Tν [f ](zi )|
1≤i≤ne

and

 ne
1 1 
RMSE = √ ||f − Tν [f ]||2 =  2
|f (zi ) − Tν [f ](zi )| ,
ne ne i=1

where zi ∈ Zne is an evaluation point belonging to the domain Ω.

In Tables 1 and 2 we report MAEs and RMSEs that decrease when the num-
ber n of interpolation points increases. Comparing then the errors obtained by
using the two data distributions, we can note that a (slightly) better accuracy is
achieved whenever we employ Halton nodes. This fact is basically due to greater
level of regularity of Halton points than pseudo-random Matlab nodes. Ana-
lyzing the error behavior with the test functions f1 and f2 , we get similar results
in terms of accuracy of the interpolation scheme.

Table 1. MAE and RMSE computed on Halton points.

n f1 f2
MAE RMSE MAE RMSE
2 500 4.29E−2 4.63E−3 1.37E−2 1.99E−3
5 000 3.75E−2 3.04E−3 1.03E−2 1.14E−3
10 000 2.39E−2 2.05E−3 5.96E−3 7.33E−4
20 000 1.72E−2 1.33E−3 3.41E−3 4.50E−4

Then we compare the performance of the optimized searching procedure

based on the partitioning of nodes in cubic blocks with a standard implemen-
tation of the algorithm where one computes all the distances between the inter-
polation nodes. With regard to the efficiency of the 3D Shepard interpolation
algorithm the CPU times computed in seconds are around 27 and 55 for the
sets of 10000 and 20000 nodes, respectively. Using a standard implementation
the seconds increase to about 41 and 318 for the two sets. From this study we
highlight a remarkable enhancement in terms of computational efficiency when
the new partitioning and searching techniques are applied.
 A 3D Efficient Procedure for Shepard Interpolants 33

Table 2. MAE and RMSE computed on pseudo-random Matlab points.

n f1 f2
MAE RMSE MAE RMSE
2 500 6.56E−2 5.49E−3 2.28E−2 2.49E−3
5 000 3.89E−2 3.89E−3 1.62E−2 1.63E−3
10 000 4.00E−2 2.71E−3 8.86E−3 1.03E−3
20 000 1.77E−2 1.65E−3 7.60E−3 6.38E−4

5 Conclusions and Future Work

In this paper we presented a new trivariate algorithm to efficiently interpolate
scattered data nodes using the tetrahedral Shepard method. Since this inter-
polation scheme needs to find suitable tetrahedra associated with the nodes,
we proposed a fast searching procedure based on the partitioning of domain in
cube blocks. Such a technique turned out to be computationally more efficient
than a standard one. Numerical experiments showed good performance of our
procedure, which enabled us to quickly deal with a large number of nodes.
Another possible extension is given by a spherical triangular or tetrahedral
Shepard method, which can be applied on the sphere S2 or other manifolds (see
e.g. [1,11]).

Acknowledgments. The authors acknowledge support from the Department of

Mathematics “Giuseppe Peano” of the University of Torino via Project 2019 “Math-
ematics for applications”. Moreover, this work was partially supported by INdAM
– GNCS Project 2019 “Kernel-based approximation, multiresolution and subdivision
methods and related applications”. This research has been accomplished within RITA
(Research ITalian network on Approximation).

References
1. Allasia, G., Cavoretto, R., De Rossi, A.: Hermite-Birkhoff interpolation on scat-
tered data on the sphere and other manifolds. Appl. Math. Comput. 318, 35–50
(2018)
2. Cavoretto, R., De Rossi, A., Perracchione, E.: Efficient computation of partition
of unity interpolants through a block-based searching technique. Comput. Math.
Appl. 71, 2568–2584 (2016)
3. Cavoretto, R., De Rossi, A., Dell’Accio, F., Di Tommaso, F.: Fast computation of
triangular Shepard interpolants. J. Comput. Appl. Math. 354, 457–470 (2019)
4. Cavoretto R., De Rossi A., Dell’Accio F., Di Tommaso F.: An efficient trivariate
algorithm for tetrahedral Shepard interpolation (2019, submitted)
5. Dell’Accio, F., Di Tommaso, F., Hormann, K.: On the approximation order of
triangular Shepard interpolation. IMA J. Numer. Anal. 36, 359–379 (2016)
6. Fasshauer, G.E.: Meshfree Approximation Methods with Matlab. World Scientific
Publishing Co., Singapore (2007)
34 R. Cavoretto et al.

7. Fasshauer, G.E., McCourt, M.J.: Kernel-Based Approximation Methods using

Matlab. Interdisciplinary Mathematical Sciences, vol. 19, World Scientific Pub-
lishing Co., Singapore (2015)
8. Little, F.F.: Convex combination surfaces. In: Barnhill R.E., Boehm, W. (eds.)
Surfaces in Computer Aided Geometric Design, Amsterdam, North-Holland, pp.
99–108 (1983)
9. Renka, R.J.: Multivariate interpolation of large sets of scattered data. ACM Trans.
Math. Softw. 14, 139–148 (1988)
10. Wendland, H.: Scattered Data Approximation. Cambridge Monographs on Applied
and Computational Mathematics, vol. 17. Cambridge University Press, Cambridge
(2005)
11. Zhang, M., Liang, X.-Z.: On a Hermite interpolation on the sphere. Appl. Numer.
Math. 61, 666–674 (2011)
 Interpolation by Bivariate Quadratic
Polynomials and Applications to the
Scattered Data Interpolation Problem

Francesco Dell’Accio(B) and Filomena Di Tommaso

Department of Mathematics and Computer Science, University of Calabria,

Rende, Italy
[email protected], [email protected]

Abstract. As specified by Little [7], the triangular Shepard method

can be generalized to higher dimensions and to set of more than three
points. In line with this idea, the hexagonal Shepard method has
been recently introduced by combining six-points basis functions with
quadratic Lagrange polynomials interpolating on these points and the
error of approximation has been carried out by adapting, to the case of
six points, the technique developed in [4]. As for the triangular Shepard
method, the use of appropriate set of six-points is crucial both for the
accuracy and the computational cost of the hexagonal Shepard method.
In this paper we discuss about some algorithm to find useful six-tuple
of points in a fast manner without the use of any triangulation of the
nodes.

Keywords: Multinode Shepard methods · Rate of convergence ·

Approximation order

1 Introduction

Let X = {x1 , ..., xn } be a set of n distinct points of a compact convex domain

Ω ⊂ R2 with associated function evaluation data fi = f (xi ), i = 1, . . . , n. We
assume that points in X are scattered, that is they do not obey any structure
or order between their relative locations. The problem of reconstruction of an
unknown function from these kind of data is well known and well studied in
approximation theory and several methods both which require a mesh (here we
mention that ones based on multivariate splines, finite elements, box splines) or
meshless methods, mainly based on radial basis functions, have been developed
with this goal and are successfully applied in different contexts. In this paper,
however, we focus on some methods which are variations of the classic Shep-
ard method (for a survey see [5]) and, in particular, to the triangular Shepard
method. Our goal is to extend the triangular Shepard method to set of more
than three points.

c Springer Nature Switzerland AG 2020

Y. D. Sergeyev and D. E. Kvasov (Eds.): NUMTA 2019, LNCS 11973, pp. 35–46, 2020.
https://doi.org/10.1007/978-3-030-39081-5_5
36 F. Dell’Accio and F. Di Tommaso

2 Triangular Shepard Method

Let T = {t1 , t2 , . . . , tm } be a triangulation of X, where tj = [xj1 , xj2 , xj3 ] denotes
the triangle with vertices xj1 , xj2 , xj3 ∈ X. The triangular Shepard interpolant
is defined by

m
Kμ [f ] (x) = Bμ,j (x) Lj [f ] (x) , x∈Ω
j=1

where
Lj [f ] (x) = λj,j1 (x)f (xj1 ) + λj,j2 (x)f (xj2 ) + λj,j3 (x)f (xj3 )
is the linear polynomial based on the vertices of tj in barycentric coordinates
and the triangular Shepard basis function is given by


3
−μ
|x − xjk |
k=1
Bμ,j (x) = , j = 1, . . . , m, μ > 0.

m 
3
−μ
|x − xkl |
k=1 l=1

The operator Kμ [f ] interpolates f (xi ) for each i = 1, . . . , n and reproduces

linear polynomials. Little noticed that it surpasses Shepard’s method greatly
in aesthetic behaviour but did not give indications neither on the choice of
the triangulation nor on the approximation order of the method. In a recent
paper [4] we have tackled these two problems and noticed that in order to reach
an adequate order of approximation, the triangulation must satisfy particular
conditions. For instance, the triangles can form a Delaunay triangulation of X
or a so called compact triangulation, that is they may overlap or being disjoint.
In the second case T is composed by a significantly smaller number of triangles
with respect to the Delaunay triangulation (see Figs. 1 and 2) and the accuracy
of approximation of the two cases are comparable [4]. The routine to detect a
compact triangulation can be organized in a fast algorithm using the localizing
searching technique [2] and the computational cost to implement the triangular
Shepard scheme is O(n log n) [1].

1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

Fig. 1. Triangular Shepard basis function Bµ,i (x) with respect to the triangle in bold
for the Delaunay triangulation of X
 Interpolation by Bivariate Quadratic Polynomials and Applications 37

1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

Fig. 2. Triangular Shepard basis function Bµ,i (x) with respect to the triangle in bold
for the a compact triangulation of X

The goal of the paper is to discuss on the extension of the triangular Shepard
method to a method based on six-point configurations of nodes in Ω. The interest
relies in the better accuracy of approximation, with respect to the triangular
Shepard method, provided that the local configurations of nodes are identified
in a proper manner. At a first glance, however, it is not clear at all which are the
configurations of six nodes that allow the solution of the Lagrange interpolation
problem in a set of scattered data (without the computation of the Vandermonde
determinant). The possible generalization of the triangular Shepard method to
set of more than three points has already been announced by Little [7] without
any suggestion on how to realize it.

3 Hexagonal Shepard Method

Let S = {s1 , s2 , . . . , sm } be a cover of X by means of six-point subsets of X:

each sj = {xjk }k=1,...,6 is a set of pairwise distinct nodes xj1 , xj2 , . . . , xj6 ∈ X
and
m
{j1 , j2 , j3 , j4 , j5 , j6 } = {1, 2, . . . , n}.
j=1

With S = {s1 , s2 , . . . , sm } we denote also the set of six-tuples sj = (xjk )k=1,...,6

that we identify with the hexagon (possibly  with self-intersections)
 bounded by
the finite chain of straight line segments xjk , xjk+1 , k = 1, . . . , 6, xj7 = xj1 ,
respectively. It will be clear from the context if we are dealing with subsets or
six-tuples, depending on the need of the order of the nodes in the subsets. We
assume that the Lagrange interpolation problem associated to each sj is poised
in the space Π22 of bivariate polynomials of degree less than or equal to 2, that
is the Lagrange interpolation polynomial exists and it is unique. For each μ > 0
the hexagonal Shepard operator is defined by

m
Hμ [f ] (x) = Eμ,j (x) Lj [f ] (x) , x∈Ω
j=1

where Lj [f ] (x) is the quadratic Lagrange interpolation polynomial on the six-

tuple sj , j = 1, . . . , m and the hexagonal Shepard basis functions with respect
38 F. Dell’Accio and F. Di Tommaso

to the set S are given by


6
−μ
|x − xjk |
k=1
Eμ,j (x) = , j = 1, . . . , m, μ > 0.

m 6
−μ
|x − xkl |
k=1 l=1

The six-points Shepard basis functions satisfy the following properties:

1. Eμ,j (x) ≥ 0,

m
2. Eμ,j (x) = 1,
j=1
3. Eμ,j (xi ) = 0, ∀j = 1, . . . , m, ∀i : xi ∈
/ sj ,
As a direct consequence of properties 2 and 3 we have
m

∀xi ∈ sj , Eμ,j (xi ) = 1.
j=1

It follows that
(i) Hμ [f ] (x) interpolates function evaluations at each node xi ;
(ii) Hμ [f ] (x) reproduces all polynomials up to degree 2
It is of interest to see the graphic behaviour of the basis functions: for a six-tuple
which has a point (in green) rounded by the others 5 (in red), the hexagonal
basis function has the shape displayed Fig. 3. Analogously, for a six-tuple which
has a point near to the boundary (in green) and the others 5 (in red) on its right
side, the hexagonal basis function has the shape displayed Fig. 4.

1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Fig. 3. Hexagonal basis function for a six-tuple which has a point (in green) rounded
by the others 5 (in red) (Color figure online)

3.1 Quadratic Lagrange Interpolation and Detection of Appropriate

Six-Tuple Set
The first problem to solve is that of the explicit representation of the quadratic
polynomial interpolant on the points of a six-tuple sj = {xji }i=1,...,6 ∈ S.
 Interpolation by Bivariate Quadratic Polynomials and Applications 39

1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Fig. 4. Hexagonal basis function for a six-tuple which has a point near to the boundary
(in green) and the others 5 (in red) on its right side (Color figure online)

A useful choice is the Lagrange form


6
Lj [f ](x) = λj,ji (x)f (xji ) , j = 1, . . . , m
i=1

which requires the computation of the Lagrange basis functions λj,ji (x), i =
1, . . . , 6 through the Kronecker’s delta property

1, i = k
λj,ji (xjk ) =
0, otherwise.

We then focus on the quadratic polynomial

1 (x)
λj,j1 (x) =
1 (xj1 )

which vanishes at all points in sj with the exception of xj1 : its analytical expres-
sion can be written such that the remaining polynomials λj,ji (x), i = 2, . . . , 6 are
obtained by means of the permutation i → i (mod 6) + 1 of the indices 1, . . . , 6.
We denote by
1 x1 x2

A(x, y, z) = 1 y1 y2
1 z1 z2

twice the signed area of triangle of vertices x, y, z ∈ R2 . Since 1 (x) is different

from zero on xj1 and vanishes on xj2 , . . . , xj6 , its analytical expression can be
obtained by linear combination

1 (x) = αQ0 (x) + βQ1 (x), (1)

of the quadratic polynomials Q0 (x) = A(x, xj2 , xj3 )A(x, xj5 , xj6 ) and Q1 (x) =
A(x, xj2 , xj6 )A(x, xj3 , xj5 ) both vanishing in xj2 , xj3 , xj5 , xj6 (Figs. 5 and 6).
The coefficients α and β which assure the vanishing of 1 (x) in xj4 are straight-
forward:

α = A(xj4 , xj3 , xj5 )A(xj4 , xj6 , xj2 ) and β = A(xj4 , xj2 , xj3 )A(xj4 , xj5 , xj6 )
40 F. Dell’Accio and F. Di Tommaso

Fig. 5. Triangles involved in the definition of polynomials Q0 (x) and Q1 (x)

Fig. 6. Triangles involved in the definition of constants α and β

For reasons of stability and accuracy, the crucial point in the definition of the
hexagonal Shepard method is the choice of the set of six-tuples S. We can take
into account an algorithm proposed by J. Dalik in the paper [3] on 2008. In
this paper he focuses on quadratic interpolation in vertices of unstructured tri-
angulations of a bounded closed domain Ω ⊂ R2 and gives conditions which
guarantee the existence, the uniqueness and the optimal-order of local interpo-
lation polynomials. More precisely, he proposes an algorithm to identify useful
six-tuples of points which are vertices of a triangulation T of Ω satisfying specific
properties: (i) the triangles of T must have no obtuse angles; (ii) the triangles
of T must have area greater than, or equal to, a fixed constant times h2 , where
h is the meshsize of T , i.e. the longest length of the sides of the triangles in T .
The core of the algorithm is the following procedure which allows to select a set
of 5 counterclockwise nodes around each interior vertex xj1 of T in a set of ν
nearby nodes as follows. Let us consider the ν nodes sharing with xj1 a common
side of the triangulation ordered counterclockwise.

– If ν = 5 we associate to xj1 the five nodes around it;

– if ν > 5 we select the 5 nodes around xj1 as follows: at each step k (1 ≤ k ≤
ν − 5) we eliminate the node (in red in the Fig. 7) such that

αi + αi+1 = min{αj + αj+1 : j = 1, . . . , ν − k}

– if ν < 5, according to the orientation, we add 5− ν nodes (in red in the Fig. 8)
which share an edge with one of the ν triangles around xj1 .

Nevertheless, to overcome the necessity of a preliminary triangulation, it is

possible to use fast algorithms to find six-tuple of points around a node. As an
example, if xj ∈ X and Nj = {xj1 , . . . , xjν } ⊂ X is the sequence of ν ≥ 5
 Interpolation by Bivariate Quadratic Polynomials and Applications 41

Fig. 7. Procedure to select a set of 5 counterclockwise nodes around xj1 if ν > 5

Fig. 8. Procedure to add 5 − n nodes which share an edge with one of the n triangles
around xj1 if ν < 5 (Color figure online)

nodes nearest to xj , enumerated with respect to the increasing distances from

xj , we can always re-enumerate the sequence in order that the nodes in Nj are
counterclockwise around xj . A simple algorithm can be organized as follows:

Algorithm 1. Re-enumerating neighbours of xj counterclockwise around it

INPUT xj , Nj
for all l = 1 : ν do
Nu (l, :) = [(xjl − xj )/||xjl − xj ||, l]
end for
sort Nu (Nu (:, 2) ≥ 0) in decreasing order with respect to the first column
sort Nu (Nu (:, 2) < 0) in increasing order with respect to the first column
concatenate the two vectors [Nj (Nu (:, 3)); Nj (Nu (:, 3))]

In order to select the set of 5 counterclockwise nodes around each interior ver-
tex xj we can adapt the Dalik’s procedure for maximizing the angles xjk xj xjk+1 ,
k = 1, . . . , ν, xjν+1 = xj1 or use some other technique which allows to reduce
the bound of the remainder term in the Lagrange interpolation on sj (see The-
orem 1).
In Table 1 we compare the approximation accuracies of H4 with those of the
2
triangular Shepard operator K2 using 1000 Halton interpolation nodes on [0, 1] .
The six-tuple set S is computed by means of the Dalik’s algorithm applied to
the Delaunay triangulation of the nodes. The numerical experiments are realized
42 F. Dell’Accio and F. Di Tommaso

by considering the set of test functions generally used in this field, defined in [8]
(see Fig. 9).
The numerical results show that the accuracy of approximation achieved by
H4 based on the Dalik’s algorithm is comparable but not better than those
of K2 . It is worth noting that it is not always possible to extract a strongly
regular triangulation from X; moreover, by applying the Dalik’s algorithm to
the Delaunay triangulation of X, we observe the presence of six-tuples of points
whose hexagons are not relatively small (see Fig. 10) and of a large number of
hexagons.
In order to improve the accuracy of approximation of H4 , in line with the
case of the triangular Shepard method, we need to (i) avoid six-tuples sj which
contain nodes not relatively near; (ii) be able to reduce the number of hexagons
of the cover S. Then it is necessary to find a procedure to compare two or more
six-tuples of points sharing a common vertex. We reach our goal by minimizing
the error bound for Lj [f ](x). In fact, the Lagrange interpolation polynomials at
the points {{xjk }6k=1 }m
j=1 is given by


6
Lj [f ](x) = λj,ji (x)f (xji ) , j = 1, . . . , m
i=1

so that by considering the second order Taylor expansion of f (xji ), i = 2, . . . , 6

centered at xj1 with integral remainder, we get
Lj [f ](x) = T2 [f, xj1 ](x) + δj (x)
where

6
T2 [f, xj1 ] (x) = λj,j1 (x) f (xj1 ) + λj,ji (x) (f (xj1 ) + (xji − xj1 ) · ∇f (xj1 )
i=2

+(xji − xj1 )Hf (xj1 ) (xji − xj1 )T

and

6 3 
|xji − xj1 | 1
∂ 3 f (xj1 + t(xji − xj1 )) 2
δj (x) = λj,ji (x) (1 − t) dt.
i=2
2 0 ∂νj3i

Therefore
|f (x) − Lj [f ] (x)| ≤ |f (x) − T2 [f, xj1 ] (x)| + |δj (x)| .
We can bound the remainder term δj (x) [6].
Theorem 1. For each x ∈ Ω we have
1 2
|δj (x)| ≤ ||f ||2,1 Kj hj |x − xjmax |
6
where |x − xjmax | = max |x − xji | and Kj is the sum of constants which con-
i=2,...,6
trol the shape of the triangles involved in the definition of the Lagrange basis
functions.
Interpolation by Bivariate Quadratic Polynomials and Applications 43

f1 f2

f3 f4

f5 f6

f7 f8

f9 f10

Fig. 9. Test functions

44 F. Dell’Accio and F. Di Tommaso

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Fig. 10. Six-tuples of points whose hexagons are not relatively small (in red) (Color
figure online)

The cubic approximation order of Hμ [f ] is guaranteed by the following the-

orem [6].
Theorem 2. Let Ω be a compact convex domain which contains X. If f ∈
C 2,1 (Ω) then for each μ > 5/6 we have

||f − Hμ [f ]|| ≤ KM ||f ||2,1 h3 ,

where

– h > 0 is as smaller as the nodes and the hexagons distributions are uniform
and the sizes of the hexagons are small;
– K > 0 is a as smaller as the constants Kj are
– M > 0 is as smaller as the maximum of the local number of hexagons is small.

After the application of the Dalik’s algorithm, we use the previous bounds to
associate to each node xi , which is vertex of two or more hexagons, the one with
ν
the smallest bound. This procedure certainly generates a cover {sj }j=1 of X
by means of six-point subsets and the number of hexagons is considerably lower
with respect to the number of nodes. For example, in the case of the 1000 Halton
points, we observe the presence of 355 hexagons versus the 976 hexagons used to
define the operator H4 based only on the Dalik’s algorithm. As a consequence
the constant M which appears in the previous Theorem is smaller since we avoid
clustered hexagons. The accuracy of approximation produced by this operator
are reported in the second column of Table 1. Finally, the results in the third
column are obtained by applying the fast Algorithm 1 for re-enumerating 9
neighbours of xj counterclockwise around it, by applying the Dalik’s procedure
to select a set of 5 counterclockwise nodes around xj and by using previous
bounds to associate to each node xj , which is vertex of two or more hexagons,
the one with the smallest bound.
 Interpolation by Bivariate Quadratic Polynomials and Applications 45

Table 1. Numerical comparison among the approximation accuracies of the operator

H4 on the six-tuples set generated by the Dalik’s algorithm, the operator H4 on the six-
tuples set generated by the refined Dalik’s algorithm, the operator H4 on the six-tuples
set generated by the fast algorithm and the operator K2 on the compact triangulation
of X

H4 H4 H4 K2
Dalik refined Dalik fast refined
f1 emax 8.93e−02 3.19e−03 7.86e−03 2.97e−02
emean 2.53e−04 2.27e−04 2.73e−04 1.22e−03
eMS 2.00e−03 3.85e−04 5.13e−04 2.21e−03
f2 emax 1.45e−02 1.59e−02 4.43e−03 6.04e−03
emean 9.18e−05 1.15e−04 1.09e−04 3.40e−04
eMS 4.47e−04 4.45e−04 2.89e−04 7.26e−04
f3 emax 4.42e−03 2.83e−04 4.66e−04 3.87e−03
emean 2.09e−05 1.86e−05 2.14e−05 2.43e−04
eMS 1.09e−04 2.80e−05 3.28e−05 4.05e−04
f4 emax 7.86e−03 8.48e−04 7.49e−04 9.05e−03
emean 2.10e−05 8.10e−06 8.42e−06 3.02e−04
eMS 2.24e−04 2.86e−05 2.36e−05 5.35e−04
f5 emax 1.81e−03 1.24e−03 2.57e−03 1.34e−02
emean 4.17e−05 5.92e−05 6.58e−05 3.90e−04
eMS 9.06e−05 1.11e−04 1.34e−04 7.98e−04
f6 emax 4.93e−03 7.19e−04 6.41e−04 8.66e−03
emean 3.46e−05 3.07e−05 3.29e−05 2.94e−04
eMS 1.81e−04 5.58e−05 5.64e−05 5.24e−04
f7 emax 1.19e+00 1.13e−01 1.11e−01 2.50e−01
emean 5.27e−03 3.86e−03 4.44e−03 1.89e−02
eMS 3.13e−02 6.27e−03 7.53e−03 2.92e−02
f8 emax 2.96e−01 2.38e−02 3.36e−02 1.35e−01
emean 1.07e−03 9.39e−04 1.10e−03 4.32e−03
eMS 6.82e−03 1.81e−03 2.12e−03 8.25e−03
f9 emax 1.20e+01 2.06e+00 1.48e+00 8.18e+00
emean 4.94e−02 6.28e−02 7.20e−02 3.08e−01
eMS 2.18e−01 1.28e−01 1.46e−01 6.61e−01
f10 emax 1.72e−01 3.35e−02 6.88e−02 1.54e−01
emean 8.49e−04 9.42e−04 1.09e−03 4.26e−03
eMS 4.15e−03 2.14e−03 2.65e−03 8.07e−03
46 F. Dell’Accio and F. Di Tommaso

4 Conclusion and Prospectives of Research

Using a procedure to select five nodes around a sixth one which guarantees
the existence and uniqueness of the Lagrange interpolation and a criterion to
compare the bounds of Lagrange interpolants on two or more six-tuples of points
sharing a common vertex, we have generalized the triangular Shepard method
to the hexagonal Shepard method, that is a fast method based on six points
with cubic approximation order and which compares well with the triangular
Shepard method in numerical accuracies. Further researches will be devoted to
the generalization of the hexagonal Shepard method to the case of scattered data
of R3 .

Acknowledgments. This work was partially supported by the INdAM-GNCS 2019

research project “Kernel-based approximation, multiresolution and subdivision meth-
ods and related applications”. This research has been accomplished within RITA
(Research ITalian network on Approximation).

References
1. Cavoretto, R., De Rossi, A., Dell’Accio, F., Di Tommaso, F.: Fast computation of
triangular Shepard interpolants. J. Comput. Appl. Math. 354, 457–470 (2019)
2. Cavoretto, R., De Rossi, A., Perracchione, E.: Efficient computation of partition of
unity interpolants through a block-based searching technique. Comput. Math. Appl.
71(12), 2568–2584 (2016)
3. Dalı́k, J.: Optimal-order quadratic interpolation in vertices of unstructured trian-
gulations. Appl. Math. 53(6), 547–560 (2008)
4. Dell’Accio, F., Di Tommaso, F., Hormann, K.: On the approximation order of tri-
angular Shepard interpolation. IMA J. Numer. Anal. 36, 359–379 (2016)
5. Dell’Accio, F., Di Tommaso, F.: Scattered data interpolation by Shepard’s like meth-
ods: classical results and recent advances. Dolomites Res. Notes Approx. 9, 32–44
(2016)
6. Dell’Accio, F., Di Tommaso, F.: On the hexagonal Shepard method. Appl. Numer.
Math. 150, 51–64 (2020). https://doi.org/10.1016/j.apnum.2019.09.005. ISSN 0168-
9274
7. Little, F.: Convex combination surfaces. In: Barnhill, R.E., Boehm, W. (eds.) Sur-
faces in Computer Aided Geometric Design, North-Holland, vol. 1479, pp. 99–108
(1983)
8. Renka, R.J.: Algorithm 790: CSHEP2D: cubic shepard method for bivariate inter-
polation of scattered data. ACM Trans. Math. Softw. 25(1), 70–73 (1999)
 Comparison of Shepard’s Like Methods
with Different Basis Functions

Francesco Dell’Accio1 , Filomena Di Tommaso1(B) ,

and Domenico Gonnelli2
1
Department of Mathematics and Computer Science, University of Calabria,
Via P. Bucci, Cubo 30A, Rende, Italy
{francesco.dellaccio,filomena.ditommaso}@unical.it
2
NTT DATA Italia, Contrada Cutura, Rende, Italy
[email protected]

Abstract. The problem of reconstruction of an unknown function from

a finite number of given scattered data is well known and well studied in
approximation theory. The methods developed with this goal are several
and are successfully applied in different contexts. Due to the need of
fast and accurate approximation methods, in this paper we numerically
compare some variation of the Shepard method obtained by considering
different basis functions.

Keywords: Shepard methods · Interpolation · Quasi-interpolation

1 Introduction
Let be X a convex domain of R2 , Xn = {x1 , x2 , . . . , xn } ⊂ X a set of nodes
and Fn = {f1 , . . . , fn } ⊂ R a set of associated function values. The problem of
reconstruction of a continuous function from the data (Xn , Fn ) is well known
and well studied in approximation theory. When the points of Xn bear no reg-
ular structure at all, we talk about scattered data approximation problem [17].
Several methods have been developed with this goal and are successfully applied
in different contexts. Some of these methods require a mesh, some others are
meshless, mainly based on radial basis functions (see for example [11,12]).
Recently, it has been pointed out the need of fast approximation methods
which overcome the high computational cost and the slowness of interpolation
schemes which entail the solution of large linear systems or the use of elaborated
mathematical procedures to find the values of parameters needed for setting
those schemes [15]. The Shepard method [16] and some of its variations [4,15]
belong to this class of methods.
The Shepard scheme consists in the construction of the function
n

φ(x − xj )fj
j=1
x → Sφ,n [f ](x) = n (1)

φ(x − xj )
j=1
c Springer Nature Switzerland AG 2020
Y. D. Sergeyev and D. E. Kvasov (Eds.): NUMTA 2019, LNCS 11973, pp. 47–55, 2020.
https://doi.org/10.1007/978-3-030-39081-5_6
48 F. Dell’Accio et al.

to approximate a target function f in X whose values fj = f (xj ) at the nodes are

known. In the original scheme [16] φ is the λ-power (λ > 0) of the inverse distance
−1
function x → |x| in the Euclidean space R2 . Since φ has a singularity in x = 0,
the Shepard approximant (1) interpolates the values of the target function at the
node xj , j = 1, . . . , n, but the reproduction quality of the interpolation operator
Sφ,n [·] is limited to constant functions. Moreover its approximation accuracy is
poor and the presence of flat spots in the neighborhood of all nodes, if λ ≥ 1,
or cusps at each xj , if λ < 1, is not visually pleasing. In the years several varia-
tions of the Shepard method have been proposed to improve its approximation
accuracy. They are based on the substitution of the function values fj with the
values at x of interpolation polynomials at xj and on the use of local support
basis functions. These schemes are called combined Shepard methods (for more
details see [10] and the references therein) and give the possibility to interpolate
additional derivative data if these are provided [1–3,6,7,9]. More recent methods
focus both on readily implementation and efficiency of the approximation. The
scaled Shepard methods [15] are based on the general scheme (1) but, instead of
−1
|x| , they make use of basis functions K(x) satisfying

• K : R2 → R+ is continuous on R2 ;
• min K(x) > 0;
|x|≤1
• for a fixed α > 0, there holds
 −α
2
K(x) ≤ κ 1 + |x| , x ∈ R2 , (2)

where κ is a constant independent on x.

For the optimality of the approximation order, they take into account the fill
distance [17]
hXn = sup inf |x − y|
x∈X y∈Xn

through a dilation factor βn , by setting φ(x) = K(βn x). The continuity of the
basis function at x = 0 causes the lost of the interpolation feature and the
resulting approximant is quasi-interpolant, since it preserves the reproduction
property of constant functions.
Another improvement of the Shepard method (1) is the triangular Shep-
ard operator [13], obtained by combining triangle-based basis functions with
local linear interpolants on the vertices of a triangulation of Xn . The triangular
Shepard is an interpolation operator which reproduces polynomials up to the
degree 1.
Being motivated by the need of fast and accurate approximation methods, as
specified above, in this paper we provide an explicit numerical comparison among
the accuracies and the CPU times (in seconds) of the Shepard, scaled Shepard
and triangular Shepard approximants. In Sect. 2, after recalling the definitions of
 Comparison of Shepard’s Like Methods with Different Basis Functions 49

the parameters necessary for rightly setting the scaled Shepard method, we make
a comparison among Shepard and scaled Shepard approximants on a set of 10 test
functions commonly used in this field [14]. In Sect. 3 we compare the triangular
Shepard method with its scaled versions. All tests have been carried out on a
laptop with an Intel(R) Core i7 5500U CPU 2.40 GHz processor and 8.00 GB
RAM. Finally, in Sect. 4 we formulate conclusions about our investigations.

2 Scaled Shepard Approximants

As pointed out in [15], the scaled Shepard approximants perform better if the
points of Xn are well-separated and Xn is quasi-uniformly distributed in X.
These conditions are fulfilled as soon as constants c and C, independent on n,
exist such that
inf |x − y| ≥ cn−1/2
x,y∈Xn
x=y

and
hXn ≤ Cn−1/2 .
√
Under these assumption, the dilation factor βn = C −1 n ≥ 1 can be computed
and the condition (2) is equivalent to
 
−2α
K(βn x) ≤ κ min 1, (βn |x|) , x ∈ R2 . (3)

The scaled Shepard approximant is

n

K(βn (x − xj ))f (xj )
j=1
x → SK,n [f ](x) = n . (4)

K(βn (x − xj ))
j=1

The following numerical experiments confirm the better accuracy of approxima-

tion achieved by the quasi-interpolant operator (4) with respect to those of the
original Shepard operator (1) according to the theoretical results proven in [15].
The experiments are realized by considering the following functions
2 2
K1 (x) = e−a |x| Gaussian function,
K2 (x) = e−a|x| C 0 -Matérn function, (5)
 −a
2
K3 (x) = 1 + |x| Generalized inverse multiquadric function
50 F. Dell’Accio et al.

1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2

Fig. 1. Behaviour of function K1 (x) (left) and K2 (x) (right) for a = 2, 3, 4 (in blue,
red and yellow) and of function (1 + |x|2 )−2 (in black). (Color figure online)

which are continuous and positive in R2 with positive minimum in the closed
ball |x| ≤ 1 for each a > 0. In the case of the Gaussian function K1 (x), by setting
|x| = r, condition (2) becomes
2 2
e−a r
≤ κ(1 + r2 )−α , r > 0
2
and for r = 1 we get e−a < κ 2−α , which is satisfied for each a2 ≥ α ≥ 1
and κ = 1. In Fig. 1 (left) we display the behaviour of the function K1 (x) for
the values of parameter a = 2, 3, 4 (in blue, red and yellow) and of the function
(1+|x| )−2 (in black). Analogously, in the case of the C 0 -Matérn function K2 (x),
2

condition (2) becomes

e−ar ≤ κ(1 + r2 )−α , r > 0

and for r = 1 we get e−a < κ 2−α which is satisfied for each a ≥ α ≥ 1 and κ = 1.
In Fig. 1 (right) we display the behaviour of function K2 (x) for a = 2, 3, 4 (in
blue, red and yellow) and of function (1 + |x| )−2 (in black). Finally, in the case
2

of the generalized inverse multiquadric function K3 (x), condition (2) is trivially

satisfied for each a ≥ α and κ = 1. In Table 1 we compare the approximation
accuracies of the original Shepard scheme S|·|−2 ,n [f ](x) with those of the quasi-
interpolant schemes SKj ,n [f ](x), j = 1, 2, 3 by setting a = 50. The numerical
experiments are performed by considering the 10 test functions defined in [14].
In Table 1 we report the absolute values of the maximum error emax , the average
error emean , and the mean square error eM S computed by using the n = 1000
Halton interpolation nodes [18] in the unit square [0, 1] × [0, 1], for which we
report the value βn ≈ 2.6087. The errors are computed at the ne = 101 × 101
points of a regular grid of the square [0, 1]×[0, 1]. The obtained numerical results
show that quasi-interpolant operators SKj ,n , j = 1, 2, 3 perform better than the
original Shepard method S|·|−2 ,n but similar to the operator S|·|−6 ,n . It is then
meaningful to compare above operators in terms of CPU times and we report
the results, computed in seconds, in Table 2.
 Comparison of Shepard’s Like Methods with Different Basis Functions 51

Table 1. Comparison among the original Shepard approximant S|·|−2 ,n , S|·|−6 ,n and
the quasi-interpolant Shepard schemes SKi ,n , i = 1, 2, 3 with parameter a = 50.

S|·|−2 ,n S|·|−6 ,n SK1 ,n SK2 ,n SK3 ,n

f1 emax 2.26e − 1 7.03e − 2 8.48e − 2 6.81e − 2 7.75e − 2
emean 3.71e − 2 7.42e − 3 9.13e − 3 5.13e − 3 8.74e − 3
eM S 5.02e − 2 1.13e − 2 1.42e − 2 8.16e − 3 1.33e − 2
f2 emax 4.66e − 2 3.36e − 2 3.87e − 2 3.25e − 2 3.13e − 2
emean 1.56e − 2 1.74e − 3 2.12e − 3 1.26e − 3 2.82e − 3
eM S 1.76e − 2 3.89e − 3 4.86e − 3 2.96e − 3 4.90e − 3
f3 emax 9.57e − 2 2.16e − 2 2.45e − 2 2.13e − 2 2.77e − 2
emean 1.48e − 2 2.94e − 3 3.63e − 3 1.98e − 3 2.69e − 3
eM S 1.98e − 2 4.00e − 3 5.03e − 3 2.83e − 3 4.15e − 3
f4 emax 6.67e − 2 1.45e − 2 1.65e − 2 1.42e − 2 1.70e − 2
emean 1.59e − 2 3.19e − 3 3.91e − 3 2.20e − 3 3.07e − 3
eM S 1.97e − 2 3.90e − 3 4.89e − 3 2.82e − 3 4.29e − 3
f5 emax 1.08e − 1 2.22e − 2 3.40e − 2 1.96e − 2 1.86e − 2
emean 1.46e − 2 2.42e − 3 3.03e − 3 1.62e − 3 2.41e − 3
eM S 2.12e − 2 4.23e − 3 5.43e − 3 2.90e − 3 3.99e − 3
f6 emax 1.52e − 1 3.74e − 2 3.58e − 2 3.78e − 2 5.86e − 2
emean 1.37e − 2 3.29e − 3 3.96e − 3 2.39e − 3 4.07e − 3
eM S 1.91e − 2 4.58e − 3 5.45e − 3 3.74e − 3 6.86e − 3
f7 emax 1.41e + 0 6.93e − 1 6.38e − 1 6.84e − 1 9.48e − 1
emean 2.94e − 1 7.99e − 2 9.65e − 2 5.79e − 2 1.24e − 1
eM S 3.80e − 1 1.03e − 1 1.27e − 1 7.81e − 2 1.57e − 1
f8 emax 8.43e − 1 1.86e − 1 2.62e − 1 1.52e − 1 2.10e − 1
emean 1.10e − 1 2.15e − 2 2.66e − 2 1.48e − 2 2.81e − 2
eM S 1.50e − 1 3.35e − 2 4.25e − 2 2.35e − 2 4.02e − 2
f9 emax 5.27e + 1 1.69e + 1 2.39e + 1 1.48e + 1 1.31e + 1
emean 7.65e + 0 1.53e + 0 1.88e + 0 1.06e + 0 1.97e + 0
eM S 1.10e + 1 2.60e + 0 3.28e + 0 1.86e + 0 3.18e + 0
f10 emax 5.82e − 1 1.45e − 1 2.14e − 1 1.16e − 1 3.14e − 1
emean 6.78e − 2 1.59e − 2 1.93e − 2 1.14e − 2 2.68e − 2
eM S 9.18e − 2 2.38e − 2 3.01e − 2 1.68e − 2 3.64e − 2

3 Scaled Triangular Shepard Approximants

The triangular Shepard scheme [13] requires the definition of a triangulation

T = {t1 , . . . , tm } of the node set Xn and, in line with previous notation, it
consists in the construction of the function
52 F. Dell’Accio et al.

Table 2. CPU times in seconds computed on the 1000 Halton interpolation points in
the case of function f1 .

Operator CPU time

S|·|2 ,n 0.2928
S|·|−6 ,n 1.0640
SK1 ,n 0.7624
SK2 ,n 0.8485
SK3 ,n 1.5148

m 
 3
φ (x − xj ) Lj [f ](x)
j=1 =1
x → Tφ,m [f ](x) = m 
(6)
 3
φ (x − xk )
k=1 =1

−1
where φ is the λ-power (λ > 0) of the inverse distance function x → |x| in
the Euclidean space R2 and Lj [f ](x) is the linear interpolation polynomial at
the vertices {xj1 , xj2 , xj3 } of the triangle tj ∈ T . The triangular Shepard scheme
reaches better accuracy of approximation with respect to the original Shepard
one both for a Delaunay triangulation or a compact triangulation of the node set
Xn . A compact triangulation of Xn , in particular, consists of a set of triangles
which may overlap or being disjoint. These triangles are determined in order to
reduce the error bound of the local linear interpolant and its number is about
1/3 the number of triangles in the Delaunay triangulation [8]. The routine to
detect a compact triangulation can be organized in a fast algorithm using the
localizing searching technique developed in [5] and the computational cost to
implement the scheme (6) is O(n log n) [4]. In line with Sect. 2, we can consider
the quasi-interpolants
m 
 3
K (βn (x − xj )) Lj [f ](x)
j=1 =1
x → TK,m [f ](x) = m 
(7)
 3
K (βn (x − xk ))
k=1 =1

which reproduce polynomials up to the degree 1 and numerically compare, in

accuracy of approximation and CPU time, these operators with the triangular
Shepard one. We use the same data and functions of previous experiments and
report the obtained numerical results in Tables 3 and 4.
 Comparison of Shepard’s Like Methods with Different Basis Functions 53

Table 3. Comparison among the triangular Shepard approximant T|·|−2 ,m and the
quasi-interpolant triangular Shepard schemes TKi ,n , i = 1, 2, 3 with parameter β = 50.

T|·|−2 ,m TK1 ,m TK2 ,m TK3 ,m

f1 emax 1.62e − 2 1.81e − 2 1.46e − 2 1.90e − 2
emean 1.18e − 3 1.63e − 3 1.36e − 3 1.57e − 3
eM S 1.99e − 3 2.55e − 3 2.11e − 3 2.53e − 3
f2 emax 6.78e − 3 9.20e − 3 8.01e − 3 6.49e − 3
emean 3.48e − 4 4.82e − 4 4.00e − 4 4.47e − 4
eM S 7.42e − 4 1.01e − 3 8.41e − 4 8.85e − 4
f3 emax 5.14e − 3 5.92e − 3 3.77e − 3 4.79e − 3
emean 2.93e − 4 3.99e − 4 3.34e − 4 3.71e − 4
eM S 4.95e − 4 5.94e − 4 4.95e − 4 5.87e − 4
f4 emax 3.22e − 3 2.96e − 3 2.54e − 3 2.93e − 3
emean 2.38e − 4 3.19e − 4 2.67e − 4 3.12e − 4
eM S 3.78e − 4 4.50e − 4 3.78e − 4 4.57e − 4
f5 emax 1.12e − 2 9.99e − 3 9.53e − 3 1.08e − 2
emean 3.91e − 4 5.59e − 4 4.61e − 4 4.98e − 4
eM S 7.91e − 4 1.01e − 3 8.27e − 4 9.72e − 4
f6 emax 8.37e − 3 5.95e − 3 6.91e − 3 7.10e − 3
emean 2.92e − 4 3.19e − 4 2.74e − 4 3.95e − 4
eM S 4.81e − 4 4.45e − 4 3.96e − 4 5.35e − 4
f7 emax 2.42e − 1 3.05e − 1 2.49e − 1 2.33e − 1
emean 1.86e − 2 2.58e − 2 2.17e − 2 2.38e − 2
eM S 2.75e − 2 3.65e − 2 3.05e − 2 3.29e − 2
f8 emax 1.16e − 1 1.15e − 1 1.08e − 1 1.17e − 1
emean 4.31e − 3 6.07e − 3 5.06e − 3 5.61e − 3
eM S 8.04e − 3 1.07e − 2 8.85e − 3 9.82e − 3
f9 emax 5.73e + 0 8.07e + 0 5.43e + 0 5.88e + 0
emean 2.91e − 1 4.27e − 1 3.55e − 1 3.71e − 1
eM S 5.65e − 1 7.82e − 1 6.42e − 1 6.97e − 1
f10 emax 1.36e − 1 1.71e − 1 1.35e − 1 1.56e − 1
emean 4.20e − 3 5.81e − 3 4.91e − 3 5.44e − 3
eM S 7.66e − 3 9.64e − 3 8.30e − 3 9.78e − 3
54 F. Dell’Accio et al.

Table 4. CPU times in seconds computed on the 1000 Halton interpolation points in
the case of function f1 .

Operator CPU time

T|·|2 ,n 0.5610
TK1 ,n 1.6901
TK2 ,n 1.6012
TK3 ,n 2.1759

4 Conclusion
In this paper we numerically compared classic Shepard method with some of its
variations obtained by considering different basis functions. The classic Shepard
method with exponent λ = 2 is the fastest but compares worse with respect to
all others. For λ = 6 its approximation accuracy is comparable with that one
of the quasi-interpolants recently introduced in [15] but the CPU time increases
notably. At the end of the day, the triangular Shepard method is a good com-
promise between computational efficiency and accuracy of approximation, as
demonstrated by comparing all numerical tables.

Acknowledgments. This work was partially supported by the INdAM-GNCS 2019

research project “Kernel-based approximation, multiresolution and subdivision meth-
ods and related applications”. This research has been accomplished within RITA
(Research ITalian network on Approximation).

References
1. Caira, R., Dell’Accio, F., Di Tommaso, F.: On the bivariate Shepard-Lidstone
operators. J. Comput. Appl. Math. 236(7), 1691–1707 (2012)
2. Cătinaş, T.: The combined Shepard-Lidstone bivariate operator. In: Mache, D.H.,
Szabados, J., de Bruin, M.G. (eds.) Trends and Applications in Constructive
Approximation, pp. 77–89. Birkhäuser Basel, Basel (2005)
3. Cătinaş, T.: The bivariate Shepard operator of Bernoulli type. Calcolo 44(4), 189–
202 (2007)
4. Cavoretto, R., De Rossi, A., Dell’Accio, F., Di Tommaso, F.: Fast computation of
triangular Shepard interpolants. J. Comput. Appl. Math. 354, 457–470 (2019)
5. Cavoretto, R., De Rossi, A., Perracchione, E.: Efficient computation of partition
of unity interpolants through a block-based searching technique. Comput. Math.
Appl. 71(12), 2568–2584 (2016)
6. Costabile, F.A., Dell’Accio, F., Di Tommaso, F.: Enhancing the approximation
order of local Shepard operators by Hermite polynomials. Comput. Math. Appl.
64(11), 3641–3655 (2012)
7. Costabile, F.A., Dell’Accio, F., Di Tommaso, F.: Complementary Lidstone inter-
polation on scattered data sets. Numer. Algorithms 64(1), 157–180 (2013)
8. Dell’Accio, F., Di Tommaso, F., Hormann, K.: On the approximation order of
triangular Shepard interpolation. IMA J. Numer. Anal. 36, 359–379 (2016)
 Comparison of Shepard’s Like Methods with Different Basis Functions 55

9. Dell’Accio, F., Di Tommaso, F.: Complete Hermite-Birkhoff interpolation on scat-

tered data by combined Shepard operators. J. Comput. Appl. Math. 300, 192–206
(2016)
10. Dell’Accio, F., Di Tommaso, F.: Scattered data interpolation by Shepard’s like
methods: classical results and recent advances. Dolomites Res. Notes Approx. 9,
32–44 (2016)
11. Fasshauer, G.E.: Meshfree Approximation Methods with Matlab. World Scientific
Publishing Co. Inc., Singapore (2007)
12. Franke, R.: Scattered data interpolation: tests of some methods. Math. Comput.
38(157), 181–200 (1982)
13. Little, F.: Convex combination surfaces. In: Barnhill, R.E., Boehm, W. (eds.) Sur-
faces in Computer Aided Geometric Design, North-Holland, vol. 1479, pp. 99–108
(1983)
14. Renka, R.J.: Algorithm 790: CSHEP2D: cubic Shepard method for bivariate inter-
polation of scattered data. ACM Trans. Math. Softw. 25(1), 70–73 (1999)
15. Senger, S., Sun, X., Wu, Z.: Optimal order Jackson type inequality for scaled
Shepard approximation. J. Approx. Theor. 227, 37–50 (2018)
16. Shepard, D.: A two-dimensional interpolation function for irregularly-spaced data.
In: Proceedings of the 1968 23rd ACM National Conference, ACM 1968, pp. 517–
524. ACM, New York (1968)
17. Wendland, H.: Scattered Data Approximation, vol. 17. Cambridge University
Press, Cambridge (2004)
18. Wong, T.T., Luk, W.S., Heng, P.A.: Sampling with Hammersley and Halton points.
J. Graph. Tools 2(2), 9–24 (1997)
 A New Remez-Type Algorithm for Best
Polynomial Approximation

Nadaniela Egidi , Lorella Fatone(B) , and Luciano Misici

Dipartimento di Matematica, Università degli Studi di Camerino,

Via Madonna delle Carceri 9, 62032 Camerino, MC, Italy
{nadaniela.egidi,lorella.fatone,luciano.misici}@unicam.it

Abstract. The best approximation problem is a classical topic of the

approximation theory and the Remez algorithm is one of the most famous
methods for computing minimax polynomial approximations. We present
a slight modification of the (second) Remez algorithm where a new app-
roach to update the trial reference is considered. In particular at each
step, given the local extrema of the error function of the trial polynomial,
the proposed algorithm replaces all the points of the trial reference con-
sidering some “ad hoc” oscillating local extrema and the global extremum
(with its adjacent) of the error function. Moreover at each step the new
trial reference is chosen trying to preserve a sort of equidistribution of the
nodes at the ends of the approximation interval. Experimentally we have
that this method is particularly appropriate when the number of the local
extrema of the error function is very large. Several numerical experiments
are performed to assess the real performance of the proposed method in
the approximation of continuous and Lipschitz continuous functions. In
particular, we compare the performance of the proposed method for the
computation of the best approximant with the algorithm proposed in [17]
where an update of the Remez ideas for best polynomial approximation
in the context of the chebfun software system is studied.

Keywords: Best polynomial approximation · Remez algorithm ·

Interpolation

1 Introduction
A classical problem of the approximation theory, going back to Chebyshev him-
self, is to look for a polynomial among those of a fixed degree that minimizes the
deviation in the supremum norm from a given continuous function on a given
interval. It is known that this polynomial exists and is unique, and is known as
the best, uniform, Chebyshev or minimax approximation to the function.
In 1934, Evgeny Yakovlevich Remez published in a series of three papers the
algorithm, that now bears his name, for the computation of the best polynomial
approximation [20–22]. With the help of three female students at the University
of Kiev, Remez obtained the coefficients and the errors of the best approxima-
tions to |x| by polynomials of degree 5, 7, 9, and 11 accurate to about 4 decimal
c Springer Nature Switzerland AG 2020
Y. D. Sergeyev and D. E. Kvasov (Eds.): NUMTA 2019, LNCS 11973, pp. 56–69, 2020.
https://doi.org/10.1007/978-3-030-39081-5_7
 A New Remez-Type Algorithm for Best Polynomial Approximation 57

places. In that period both theoretical properties and computational aspects of

best polynomial approximation were deeply investigated and the Remez algo-
rithm can be considered as one of the first nontrivial algorithms for a nonlinear
computational problem that was solved before the invention of computers.
With the advent of computers, in the mid-20th century, researches on the
computation of best approximations increased (some publications report exam-
ples of best approximations with degrees of a couple of dozens) but, after this
period, the interest diminished probably due to the lack of minimax approxi-
mation applications aside from digital signal processing and approximation of
special functions.
In recent years best polynomial approximation became a mandatory topic
in books of approximation theory but not always as useful as one might image
in practical applications. In fact, in practice, other types of polynomial inter-
polation (for example Chebyshev interpolants) are often as good or even bet-
ter minimax polynomial approximations (see, e.g. [1,15]). Nevertheless the best
approximation is a fundamental topic of approximation theory, going back more
than 150 years, and we believe that the study and development of robust algo-
rithms to compute this kind of approximation is an interesting line of research.
Moreover best approximation is a challenging problem and some new additional
ideas sometimes could indicate directions for its improvement.
In this paper we present a new version of the Remez algorithm that increases
its robustness (see [10]). Moreover this new version is reliable and able to com-
pute high accuracy best polynomial approximations with degrees in the hundreds
or thousands to different functions. In particular we propose a slight modification
of the Remez algorithm presented by the authors in [17] in the context of the
chebfun software system (see [2,17] and, for more details, https://www.chebfun.
org).
The paper is organized as follows. In Sect. 2 we review the classical Remez
algorithm specifying how the computation of a trial reference and a trial poly-
nomial is performed. In particular the Remez algorithm proposed in the chebfun
software system is presented. Section 3 shows a new strategy to adjust the trial
reference from the error of the trial polynomial. Finally in Sect. 4 we compare the
performance of the chebfun Remez algorithm and the slight variation proposed
here.

2 Classical Remez Algorithm

Let Pn be the set of polynomials of degree less than or equal to n ∈ N and having
real coefficients. Let f be a continuous function defined on a finite closed interval
I = [a, b], a, b ∈ R, a < b, that is f ∈ C (I), and let ·∞ be the supremum norm
of · on I, that is

f ∞ = max |f (x)| . (1)

x∈I
58 N. Egidi et al.

The classical problem of best polynomial approximation can be stated as

follow: given f ∈ C (I) find qn∗ ∈ Pn such that:

f − qn∗ ∞ ≤ f − pn ∞ , ∀pn ∈ Pn . (2)

It is well known that in C (I) there exists a unique best uniform approxima-
tion polynomial qn∗ ∈ Pn of f (see, for example, [1,6,14–16,19] and the references
therein).
Since the operator that assigns to each continuous function its best polyno-
mial approximation qn∗ , although continuous, is nonlinear (see, e.g., [14]), itera-
tive methods have been developed to compute qn∗ . The Remez algorithm is one
of these methods (see [20–22]). Given an initial grid, the Remez algorithm con-
sists of an iterative procedure that, modifying at every step the grid of trial
interpolants, converges quadratically to qn∗ , under some suitable assumptions
on f .
The Remez algorithm is essentially based on the following two theorems.

Theorem 1. (Chebyshev Equioscillation Theorem) Let f ∈ C (I). A polynomial

qn ∈ Pn is the best approximation to f (that is qn = qn∗ ) if and only if there
exists a set of n + 2 distinct points, a ≤ x0 < x1 < · · · < xn+1 ≤ b in I such
that

f (xi ) − qn (xi ) = λσi f − qn∗ ∞ , i = 0, 1, . . . , n + 1, (3)

where σi = (−1)i and λ = 1 or λ = −1 is a fixed constant.

For the proof of this fundamental theorem (in original or generalized form) see,
for example, [4–6,8,12,13,16,18].
The above theorem concerns the property of equioscillation of f − qn∗ . Next
theorem provides a lower estimate of the best approximation error and it is
very useful in numerical methods for finding the polynomial of best uniform
approximation.

Theorem 2. (De La Vallée Poussin) Let f ∈ C (I). Suppose we have a polyno-

mial qn ∈ Pn which satisfies

f (yi ) − qn (yi ) = (−1)i ei , i = 0, 1, . . . , n + 1, (4)

where a ≤ y0 < y1 < · · · < yn+1 ≤ b, and all ei , i = 0, 1, . . . , n + 1, are nonzero

and of the same sign. Then, for every pn ∈ Pn , we have:

min |f (yi ) − qn (yi )| ≤ max |f (yi ) − pn (yi )| , (5)

i i

and, in particular:

min |f (yi ) − qn (yi )| ≤ f − qn∗ ∞ ≤ f − qn ∞ . (6)

i
 A New Remez-Type Algorithm for Best Polynomial Approximation 59

For the proof see [7].

We define a set of n + 2 ordered points x = (x0 , x1 , . . . , xn+1 ) in I that
satisfies (3) a reference equioscillation set or simply a reference.
Estimates of the best uniform approximation error in terms of the smoothness
of the function f can be found in [1] (see also [15]).
Let
rn (x) = f (x) − qn (x), x ∈ I = [a, b],
be the remainder (i.e. the error function) when we approximate f with qn ∈ Pn .
If rn (a) = 0 and rn (b) = 0 and rn oscillates exactly n + 2 times then rn is called
standard remainder, otherwise it is called not standard remainder. We note that
when rn is standard then it has exactly n+2 local minima or maxima. Chebyshev
equioscillation theorem ensures that the residual rn∗ = f − qn∗ is standard.
From Theorems 1 and 2 we have that a polynomial qn ∈ Pn whose error
function oscillates n + 2 times (i.e. (4) holds), is a “near-best” approximation of
f , in the sense that
f − qn ∞
f − qn ∞ ≤ C f − qn∗ ∞ , C = ≥ 1. (7)
min |f (yi ) − qn (yi )|
i

On the basis of the above considerations it is possible to deduce methods that

often give a good estimate of the minimax approximation. Among all the near-
minimax approximations we mention the least-squares approximation, the inter-
polation at the Chebyshev nodes and, perhaps the most used, forced-oscillation
best approximation. For more details see, e.g., [1] and the references therein.
The Remez algorithm is one of the most famous methods for the com-
putation of minimax polynomial approximations [20–22]. This algorithm con-
structs a sequence of trial references x(k) = (x0 , x1 , . . . , xn+1 ) ∈ Rn+2 , with
(k) (k) (k)

(k) (k) (k)

a ≤ x0 < x1 < · · · < xn+1 ≤ b, and a sequence of trial polynomials
(k)
qn (x) ∈ Pn , k = 0, 1, . . . , that satisfy the alternation condition (4) in such
(k)
a way that C (k) , given by (7) when we replace qn with qn , approaches to 1
when k → +∞.
Remez Algorithm
Let {φi , i = 0, 1, . . . , n} be a basis of Pn , and let

x(0) = (x0 , x1 , . . . , xn+1 ) ∈ Rn+2 ,

(0) (0) (0)
(8)
(0) (0) (0)
such that a ≤ x0 < x1 < · · · < xn+1 ≤ b. Given x(k) at each kth step,
k = 0, 1, . . ., perform the following steps.
1. Computation of the trial polynomial.
n
 (k)
Construct qn(k) (x) = an,i φi (x) ∈ Pn such that
i=0
   
= (−1)j En(k) ,
(k) (k)
f xj − qn(k) xj j = 0, 1, . . . , n + 1, (9)
60 N. Egidi et al.

(k)
where En is called the levelled error (it may be positive or negative).
(k)
The polynomial qn (x) is obtained from the solution of linear system (9) that
(k) (k)
has n + 2 equations and n + 2 unknowns: the levelled error En plus an,i ,
 
 (k) 
i = 0, 1, . . . , n. Note that En  ≤ f − qn∗ ∞ .
 
 (k) 
2. If En  = f − qn∗ ∞ from the Theorem 1 we have that qn∗ = qn and the
(k)

algorithm stops.
3. Adjustment of the trial reference from the error of the trial polynomial.
 
 (k) 
If 0 < En  < f − qn∗ ∞ , the goal is to construct a new trial reference
x(k+1) such that the residual

rn(k) (x) = f (x) − qn(k) (x), (10)

(k+1)
oscillates at xj , j = 0, 1, . . . , n, that is
   
= (−1)j En,j ,
(k+1) (k+1) (k+1)
f xj − qn(k) xj j = 0, 1, . . . , n + 1, (11)

(k+1)
with En,j , j = 0, 1, . . . , n + 1, having all the same sign and such that
   
 (k+1)   (k) 
En,j  ≥ En , j = 0, 1, . . . , n + 1.

This last step will be described in more details in the next section both for
the classical Remez algorithm and for the proposed slight modification.
We note that theorem of De La Vallée Poussin guarantees the existence of
this new trial reference. In fact, to be sure of increasing the levelled error, the
replacement of the old trial reference with the new one must satisfies:
 
 
|En(k) | ≤ min |En,j | ≤ f − qn∗ ∞ ≤ f − qn(k+1)  .
(k+1)
(12)
j ∞

(k+1)
This implies that the new polynomial qn equioscillates
   with  a levelled error
 (k+1)   (k) 
greater in modulus than the previous one, i.e. En  > En . The monotonic
increase of the modulus of the levelled error is the key observation in order to
show that the algorithm converges to qn∗ , i.e. qn converges uniformly to qn∗ when
(k)

k → +∞ (for more details see, for example, [19]).

Remez proposed two strategies to perform Step 3. of the previous algorithm,
that is, he suggested two approaches to construct the new trial reference x(k+1)
(k)
knowing x(k) and qn . In the first Remez algorithm, x(k+1) is constructed by
(k)
moving one of the points of x(k) to the abscissa of the global extremum of rn
while keeping the sign alternation. In the second Remez algorithm x (k+1)
is con-
structed by replacing all the points of the trial reference with alternating abscissa
(k)
of the extrema of rn including the global extremum. The linear convergence
of the second Remez algorithm can be proved for each continuous function (see
[18]). Furthermore it is possible to prove that if f is twice differentiable, the error
 A New Remez-Type Algorithm for Best Polynomial Approximation 61

decays with a quadratic rate at every n + 2 steps in the case of the first Remez
algorithm [19] and at every step in the case of the second Remez algorithm [23].
In the next section the second Remez algorithm is explained in detail and
a slight modification of the second Remez algorithm is proposed, where a new
approach to update the trial reference is considered. We will see that this new
strategy is particularly effective when the trial polynomial has a very large num-
ber of extrema.
Obviously some stopping criteria for the Remez algorithm must be consid-
ered (see Step 2.). Among the others we mention the control of the difference
between the absolute value of the levelled error and the maximum error of the
trial polynomial (see [2]), the monitoring of the relative error between the trial
polynomial and the function (especially when the approximation is used to com-
pute a given function on a computer that uses floating point arithmetic) and,
obviously, the restriction on the maximum number of iterations.
It is worthwhile to note that the choice of basis {φi , i = 0, 1, . . . , n} of Pn
in Step 1. of the Remez algorithm is a critical point in the construction of best
approximants. In fact the basis used to represent polynomials determines the
numerical properties of the linear system (9) used for the computation of the
trial polynomials (see [9]). For example, the monomial basis, is a bad choice,
in fact the condition number of the resulting Vandermonde matrix, generally,
grows exponentially [11]. Usually the Chebyshev polynomial basis is used, but
also this choice can give an ill-conditioned system for arbitrary sets of points.
To overcome this difficulty instead of solving at each Remez iteration k, k =
0, 1, . . . , the linear system (9) we can use the following strategy. Let us construct
(k) (k) (k) (k)
two standard interpolating polynomials sn , tn of degree n, i.e. sn , tn ∈ Pn ,
such that:
     
= (−1)j ,
(k) (k) (k)
s(k)
n xj = f xj , t (k)
n xj j = 0, 1, . . . , n. (13)

That is, sn interpolates f and tn interpolates the ordinates (−1)j , j = 0, 1,

(k) (k)
(k)
. . . , n, at the first n + 1 nodes of the trial reference. Let En be given by the
following formula
     
xn+1 En(k) = f xn+1 − (−1)n+1 En(k) ,
(k) (k) (k)
s(k)
n xn+1 − t(k)
n (14)
(k)
then it is easy to prove that En is the levelled error at step k and the corre-
sponding trial polynomial is given by the following linear combination
qn(k) (x) = s(k)
n (x) − tn (x) En ,
(k) (k)
(15)
(k)
that is also a polynomial of degree n, i.e. qn ∈ Pn , satisfying Eq. (9).
(k) (k)
Note that the interpolants sn , tn ∈ Pn , can be constructed with any classi-
cal interpolation method. In [17] the authors proposed to use the Lagrange basis
and evaluate the trial polynomial with an ad hoc barycentric formula. This
barycentric Lagrange formulation is enough stable and effective for the evalua-
tion of high-degree polynomials. For more details about barycentric interpolation
formulas, see, for example, [3].
62 N. Egidi et al.

Going into details in [17] the authors present an update of the second Remez
algorithm in the context of the chebfun software system. This algorithm carries
out numerical computing with functions rather than numbers. The chebfun sys-
tem is a free/open-source software system written in MATLAB for numerical
computation with functions of one real variable. It was introduced in its original
form in [2]. The command remez, inside chebfun, allows a practical computation
of a minimax approximation by using the second Remez algorithm. In particular,
(k)
command remez, at each iteration step, computes the levelled error En of the
Remez algorithm thanks to an explicit formula resulting from the barycentric
Lagrange representation formula used in the manipulation of trial polynomi-
als (see Step 1. of the Remez algorithm) and uses chebfun global rootfinding
(k)
to compute all the local extrema of rn from which the new trial reference is
constructed (see Step 3. of the Remez algorithm).
Here and in [10] we also adopt the barycentric Lagrange representation for-
mula proposed in [17] for the construction of trial polynomials and we also use
(k)
chebfun global rootfinding to compute all the local extrema of rn from which the
new trial reference is constructed. In particular, in the next section we present
two algorithms for the construction of the new trial reference from the local
(k)
extrema of rn , that is, the one implemented in the command remez in chebfun
software system and the new version proposed in this paper.

3 From a Trial Polynomial to a Trial Reference: A New

Approach
This section provides more insights on Step 3. of the second Remez algorithm
outlined in the previous section. In particular at each Remez iteration, after
having computed a trial polynomial, a new trial reference must be constructed.
Thus, let us suppose that at each Remez iteration k, k = 0, 1, . . . , given a trial
(k) (k)
reference x(k) , there is a polynomial qn ∈ Pn defining the error function rn in
(k)
(10). Moreover let us suppose to correctly locate the oscillating extrema of rn .
In the classical second Remez algorithm the new trial reference is constructed
as follows.
Algorithm 1: the second Remez algorithm to compute a trial reference.
(k)
Let m be the number of the local maxima or minima of rn defined in (10)
and let z (k) ∈ Rm be the vector of their abscissas sorted in ascending order.
(k)
(i) If m = n+2, that is if rn has exactly n+2 local extrema, then x(k+1) = z (k)
.
(ii) If m > n + 2, let rk,n ∈ Rm be the vector defined as rik,n = rn zi ,
(k) (k)

i = 1, 2, . . . m. The new trial reference x(k+1) is obtained refining z (k) as

follows.    
   (k) 
(iia) Delete from z (k) the components where rik,n  < En . Then, for each
resulting set of consecutive points with the same sign, consider only those
where rk,n attains the largest value and delete the others. Let m ≥ n + 2
be the number of the remaining extrema abscissas and y (k) ∈ Rm be the
 A New Remez-Type Algorithm for Best Polynomial Approximation 63

corresponding vector obtained from z (k) after this refining. Note that by
construction y (k) contains the abscissas of oscillating extrema.
(iib) Given y (k) ∈ Rm obtained in (iia) choose n + 2 consecutive points among
the components of y (k) ∈ Rm including the global extremum z of rn ,
(k)

defined as
|rn(k) (z)| = ||rk,n ||∞ , (16)
by using the following rules:
– if m = n + 2 set x(k+1) = y (k) ,
– if m > n + 2 and on the left of z there are less than n + 2 points, then
x(k+1) is equal to the first n + 2 components of y (k) ,
– if m > n + 2 and on the left of z there are more than n + 2 points,
then x(k+1) is equal to the first n + 1 components of y (k) on the left
of z and z itself.
In this paper we propose to modify Step (iib) of Algorithm 1 as follows.
Algorithm 2: a new approach to compute a trial reference.
(iib) Choose n + 2 consecutive points among the components of y (k) ∈ Rm
that has been obtained after refining (iia) including the global extremum z
(k)
of rn defined in (16), by using the following rules.
– If m = n + 2 set x(k+1) = y (k) .
– If m = n + 3 and z is among the first
(n + 2)/2 components of y (k) ,
the new trial reference x(k+1) is equal to the first n + 2 components of
y (k) , otherwise x(k+1) is equal to the last n + 2 components of y (k) , by
indicating with
· the integer part of ·.
– If m > n+3 and m and n+2 are even numbers than consider the following
rules.
• If z belongs to the first (n + 2)/2 components of y (k) or to the
last (n + 2)/2 components of y (k) , then x(k+1) is the vector given by
the first (n + 2)/2 components of y (k) followed by the last (n + 2)/2
components of y (k) .
• If z belongs to the components of y (k) between the position
(n + 2)/2 + 1 and the position m − (n + 2)/2, then x(k+1) is the vector
having increasing components formed by the global extremum z, the
component of y (k) just on the left of z, plus n components chosen
between the other components of y (k) so that both the equidistribu-
tion of the nodes at the ends of y (k) and the alternation of signs of
the error function are maintained.
– Similar rules apply when m > n + 3 and m, n + 2 are odd or when
m > n + 3 and m is even and n + 2 is odd (or vice versa) (see [10]). In
this paper we only note that special attention must be paid when n + 2
is odd and m is even (or vice versa), since in these cases keeping the sign
alternation of the error function is not a trivial matter. For more details
see [10].
64 N. Egidi et al.

4 Numerical Results
Let us denote by R and Rnew the Remez algorithm within the chebfun software
system obtained by using, respectively, the Algorithm 1 and the Algorithm 2 of
Sect. 3 to compute the trial reference. Note that algorithm R is implemented in
the command remez in chebfun software system (see [2,17]), while Rnew is the
slight modification of the Remez algorithm proposed in this paper and further
examined in [10].
In this section we compare the performance of the two algorithms R and Rnew
through the computation of some best approximants. To do this comparison we
consider some of the functions used in [17] besides other functions obtained from
the previous ones by adding an oscillating noise.
In particular we use R and Rnew to compute minimax polynomial approx-
imation in the domain [a, b] = [−1, 1]. For the sake of brevity in this paper we
report only numerical results obtained when n = 30, but very similar results
have been achieved for different other choices of n (see [10]).
In addition to a maximum number of iteration Kmax , we use the follow-
ing stopping criterion for the two Remez algorithms: given a tolerance tol, the
algorithm stops at the first iteration k where

||rn(k) ||∞ − |En(k) | ≤ tol. (17)

Note that the quantity on the left of (17) is always positive. In the numerical
experiments we use Kmax = 100 and tol = 10−7 .
Both the algorithms R and Rnew start from the following trial reference:

x(0) = (x0 , x1 , . . . , xn+1 ) ∈ Rn+2 ,

(0) (0) (0)
(18)

defined as
 
(0) jπ
xj = cos , j = 0, 1, . . . , n + 1, (19)
n+1
(0)
and so they construct the same qn . That is, the two Remez algorithms R and
(0)
Rnew construct the same initial trial polynomial qn .
(0)
We observe that xj , j = 0, 1, . . . , n + 1, in (19) are the n + 2 maxima
of |Tn+1 (x)|, where Tn+1 is the Chebyshev polynomial of degree n + 1. It is
worthwhile to note that this trial reference is considered an excellent choice for
a near-minimax approximation, see [1].
Tables 1, 2 show the numerical results obtained looking for best polynomial
approximations by polynomials of degree n = 30 to some Hölder continuous
functions. In these tables, for each function f , we report the following quantities:

 (0) 
f − qn  , i.e. the infinite norm of the error obtained approximating f with
∞
the initial trial polynomial qn , f − qn∗ ∞ , that is the approximated minimax
(0)

error obtained applying Remez algorithms R and Rnew , and finally K, i.e. the
number of iterations necessary to obtain the corresponding minimax error. When
 A New Remez-Type Algorithm for Best Polynomial Approximation 65

the algorithm is unstable we simply type not stable in the table, avoiding to
report the minimax error.
Note that the functions listed in Table 2 are obtained from those reported in
Table 1 by adding an oscillating noise of the form  sin(Ax), where  represents
the magnitude of the error.
As we can see from Table 1, without the noise, the two Remez algorithms
R and Rnew have comparable performances. In fact, they converge in the same
iterations number K, reaching the same approximated minimax error. Similar
results are obtained also for different choices of n.

Table 1. Best polynomial approximation qn∗ of degree n = 30 to f obtained with the

two algorithms R and Rnew . The variable K denotes the number of iterations necessary
to obtain the corresponding minimax error.
 
 (0) 
f (x) f − qn  R f − qn∗ ∞ K Rnew f − qn∗ ∞ K
∞
|x| 0.0322581 0.0093325 5 0.0093325 5
√
x+1 0.0134517 0.0066025 4 0.0066025 4

|x − 0.1| 0.1908033 0.0636224 6 0.0636224 6
1 − sin(5|x − 0.5|) 0.1230295 0.0406541 7 0.0406541 7

Table 2. Best polynomial approximation qn∗ of degree n = 30 to f obtained with

the two algorithms R and Rnew . When the algorithm is unstable we write not stable,
while when it converges we report the number of iterations K necessary to obtain the
corresponding minimax error.
 
 (0)  ∗ ∗
f (x) f − qn  R f − qn ∞ Rnew f − qn ∞ K
∞
|x| +  sin(Ax) 0.1072307 not stable 0.0541669 29
√
x + 1 +  sin(Ax) 0.1071668 not stable 0.0500000 32

|x − 0.1| +  sin(Ax) 0.1916529 not stable 0.1008867 25
1 − sin(5|x − 0.5|) +  sin(Ax) 0.1239266 not stable 0.0659363 25

Problems arise when the functions exhibit a great number of oscillations and
n is sufficiently large. To simulate this situation we add to the functions listed
in Table 1 a very oscillating noise. The results obtained considering a particular
noise of the form  sin(Ax) with  = 0.05, A = 100 are shown in Table 2, but very
similar findings are achieved when other types of oscillating errors and/or other
choices of n are taken into account (see further details in [10]). From Table 2
we have that, unlike Rnew , the classical Remez algorithm is unstable, in fact
66 N. Egidi et al.

Fig. 1. Error obtained at the last Kmax = 100 iteration of the Remez algorithm R
when computing the best polynomial approximation of degree n = 30 to f (x) = 1 −
sin(5|x − 0.5|) + 0.05 sin(100x), x ∈ [−1, 1].

the algorithm R doesn’t converge when applied to this kind of functions. The
instability is mainly due to the fact that the noise added to the functions has
a great number of oscillations and, as a consequence, the number of the local
extrema of the error function is very large. In such a case, an ad hoc choice of
the trial reference (as the one proposed here) is essential for the stability of the
Remez algorithm.
To have an idea of what happens in these cases, we plot in Figs. 1, 2 the
(k)
error rn defined in (10) obtained applying the Remez algorithms R and Rnew
to compute the minimax approximation to function f (x) = 1 − sin(5|x − 0.5|) +
0.05 sin(100x), x ∈ [−1, 1] when n = 30.
In this case Rnew converges after K = 25 iterations (see Table 2) and the
graph of the corresponding error function at the 25th iteration is shown in Fig. 2.
As can be seen from Fig. 2, since the new algorithm Rnew converges, the error
curve equioscillates in n + 2 = 32 points.
Instead the classical Remez algorithm R is unstable and it doesn’t converge
after Kmax = 100 iterations (see Table 2). In Fig. 1 we can see the instability
of the algorithm R reflected on the magnitude of the error obtained at the last
iteration fixed in the computation.
 A New Remez-Type Algorithm for Best Polynomial Approximation 67

Fig. 2. Error of best polynomial approximation of degree n = 30 to f (x) = 1 −

sin(5|x − 0.5|) + 0.05 sin(100x), x ∈ [−1, 1], obtained by using the Remez algorithm
Rnew .

5 Conclusions
In this paper a slight modification of the (second) Remez algorithm, for the
computation of the best approximation to a function f , has been presented,
developed and implemented in the Remez algorithm Rnew . In this algorithm, a
new approach to update the trial reference has been considered. In particular
at each step of the algorithm, given the local extrema of the error function, the
new trial reference is chosen in such a way to contain the global extremum (with
its adjacent) of the error function together with some other ad hoc oscillating
local extrema that must be equidistributed in the approximation interval.
This slight modification has been compared with the classical (second) Remez
algorithm R implemented in the chebfun system (a free/open-source software
system written in MATLAB).
Numerical results show that the two algorithms R and Rnew perform sim-
ilarly well on classical examples used to test algorithms for minimax approx-
imation. However, when the degree of the approximant polynomial n is large
and the function f has many oscillations, the classical algorithm R is not stable
whereas the new slight modification Rnew easily computes the searched minimax
approximant.
In conclusion when the error curve for a best polynomial approximation by
polynomials of degree n has many more oscillating points than the n + 2 points
68 N. Egidi et al.

of equioscillations of the minimax approximant, then the new Remez algorithm

Rnew often outperforms the Remez algorithm R.

References
1. Atkinson, K.E.: An Introduction to Numerical Analysis, 2nd edn. Wiley, New York
(1989)
2. Battles, Z., Trefethen, L.N.: An extension of MATLAB to continuous functions
and operators. SIAM J. Sci. Comput. 25(5), 1743–1770 (2004)
3. Berrut, J.P., Trefethen, L.N.: Barycentric Lagrange interpolation. SIAM Rev. 46,
501–517 (2004)
4. Blichfeldt, H.F.: Note on the functions of the form f (x) ≡ ϕ(x)+a1 xn−1 +· · ·+an .
Trans. Am. Math. Soc. 2, 100–102 (1901)
5. Chebyshev, P.L.: Théorie des mécanismes connus sous le nom de parallélogrammes,
Mémoires de l’Académie impériale des sciences de St. Pétersbourg, vol. 7, pp. 539–
564 (1854)
6. Davis, P.: Interpolation and Approximation. Ginn (Blaisdell), Boston (1963)
7. De La Vallée Poussin, Ch.J.: Leçons sur l’approximation des fonctions d’une vari-
able réelle. Gautier-Villars, Paris (1919)
8. Devore, R.A., Lorentz, G.G.: Constructive Approximation. Grundlehren der math-
ematischen Wissenschaften, vol. 303. Springer, Berlin (1993)
9. Dunham, C.B.: Choice of basis for Chebyshev approximation. ACM Trans. Math.
Softw. 8, 21–25 (1982)
10. Egidi, N., Fatone, L., Misici, L.: A stable Remez algorithm for minimax approxi-
mation (in preparation)
11. Higham, N.J.: Accuracy and Stability of Numerical Algorithms, 2nd edn. SIAM,
Philadelphia (2002)
12. Kirchberger, P.: Über Tschebysheff’sche Annäherungsmethoden, dissertation,
Göttingen (1902)
13. Korneichuk, N.: Exact Constants in Approximation Theory, Encyclopedia of Math-
ematics and its Applications, vol. 38. Cambridge University Press, Cambridge
(1991)
14. Lorentz, G.G.: Approximation of Functions. Holt, Rinehart and Winston, New
York (1966)
15. Mastroianni, G., Milovanovic, G.V.: Interpolation Processes. Basic Theory and
Applications. Springer Monographs in Mathematics. Springer, Berlin (2008).
https://doi.org/10.1007/978-3-540-68349-0
16. Meinardus, G.: Approximation of Functions: Theory and Numerical Methods.
Springer, New York (1967). https://doi.org/10.1007/978-3-642-85643-3. (transl.
L. Schumaker)
17. Pachón, R., Trefethen, L.N.: Barycentric-Remez algorithms for best polynomial
approximation in the chebfun system. BIT Numer. Math. 49(4), 721–741 (2009)
18. Petrushev, P.P., Popov, V.A.: Rational Approximation of Real Function, Encyclo-
pedia of Mathematics and its Applications, vol. 28. Cambridge University Press,
Cambridge (1987)
19. Powell, M.J.D.: Approximation Theory and Methods. Cambridge University Press,
Cambridge (1981)
20. Remez, E.Y.: Sur la détermination des polynomes d’approximation de degré donné.
Commun. Kharkov Math. Soc. 10, 41–63 (1934)
 A New Remez-Type Algorithm for Best Polynomial Approximation 69

21. Remez, E.Y.: Sur le calcul effectif des polynomes d’approximation de Tchebychef.
Comptes rendus de l’Académie des Sciences 199, 337–340 (1934)
22. Remez, E.Y.: Sur un procédé convergent d’approximations successives pour
déterminer les polynomes d’approximation. Comptes rendus de l’Académie des
Sciences 198, 2063–2065 (1934)
23. Veidinger, L.: On the numerical determination of the best approximation in the
Chebyshev sense. Numer. Math. 2, 99–105 (1960)
 An SVE Approach for the Numerical
Solution of Ordinary Differential
Equations

Nadaniela Egidi(B) and Pierluigi Maponi

Scuola di Scienze e Tecnologie, Università di Camerino, 62032 Camerino, MC, Italy

{nadaniela.egidi,pierluigi.maponi}@unicam.it

Abstract. The derivative operator is reformulated as a Volterra integral

operator of the first kind. So, the singular value expansion (SVE) of the
kernel of such integral operator can be used to obtain new numerical
methods to solve differential equations. We present such ideas in the
solution of initial value problems for ordinary differential equations of
first order. In particular, we develop an iterative scheme where global
error in the solution of this problem is gradually reduced at each step.
The global error is approximated by using the system of the singular
functions in the aforementioned SVE.
Some experiments are used to show the performances of the proposed
numerical method.

Keywords: Approximation · Ordinary differential equation ·

Numerical differentiation · Singular value expansion · Volterra integral
equation

1 Introduction

Differential equations are one of the most important mathematical tools used
in many scientific fields such as physics, chemistry, biology, economics. So, each
advancement in the numerical solution of differential models has great influence
on applied sciences. In particular, numerical methods for solving initial value
problems for ordinary differential equations of first order have a central role
in numerical analysis. The most popular numerical methods for solving such
problems can be organised in two classes: multistep methods [1–3], and Runge-
Kutta methods [5–7]. Both these classes share the well known Euler’s method.
The approximation of derivatives is an important tool in the solution of
ordinary differential equations (ODEs) and allows the definition of algebraic
equations for the corresponding numerical solution [9]. In this paper, we propose
an iterative method to solve ODEs where the global error in the solution is
gradually reduced at each step, and the initial guess is obtained by the Euler’s
method. This iterative method is based on the singular value expansion of a
particular Volterra integral operator that gives a reformulation of the derivative
c Springer Nature Switzerland AG 2020
Y. D. Sergeyev and D. E. Kvasov (Eds.): NUMTA 2019, LNCS 11973, pp. 70–85, 2020.
https://doi.org/10.1007/978-3-030-39081-5_8
 An SVE Approach for the Numerical Solution of ODEs 71

operator. Hence, such a SVE can be a useful tool for the approximation of the
derivative operator and for the construction of new numerical methods to solve
differential equations.
In Sect. 2 we recall the formulation of the derivative operator as a Volterra
integral operator, and the SVE of the corresponding kernel. In Sect. 3 we present
the iterative method to solve initial value problems for ODEs of first order. In
Sect. 4 we present the numerical results obtained on some classical ODEs. In
Sect. 5 we give conclusions and future developments.

2 The Derivative Operator as a Volterra Integral

Operator

In this section we give the reformulation of the derivative operator as a Volterra

integral operator, and we recall the SVE of the obtained integral operator.

2.1 The Volterra Integral Operator

Let ν ≥ 1 be a given integer number, and let g : [0, 1] → R be a continuously

differentiable function up to order ν. Let g (j) be the jth derivative of g. We
suppose that g (j) (0), j = 0, 1, . . . , ν − 1, are known or already calculated with
lower values of ν and let
ν−1
 g (j) (0) j
h(t) = g(t) − t , t ∈ [0, 1]. (1)
j=0
j!

We note that h(j) (0) = 0, j = 0, 1, . . . , ν − 1 and h(ν) (t) = g (ν) (t), t ∈ [0, 1].
Hence, from standard arguments on Maclauren formula, we have that the deriva-
tion problem for h (and so for g) can be formulated as the following Volterra
integral equation of the first kind:
 t
(t − s)ν−1 (ν)
h(t) = h (s)ds, t ∈ [0, 1]. (2)
0 (ν − 1)!

Therefore v = h(ν) is the unique solution (see [11] for details) of the following
integral equation
 1
K(t, s)v(s)ds = h(t), t ∈ [0, 1], (3)
0

where h is a known function, and the integral kernel K : [0, 1] × [0, 1] → R is

defined as follows:

(t−s)ν−1
K(t, s) = (ν−1)! , 0 ≤ s < t ≤ 1, (4)
0, 0 ≤ t ≤ s ≤ 1.
72 N. Egidi and P. Maponi

Let K be the integral operator having kernel K defined by (4), then Eq. (3) can
be rewritten as

Kv = h. (5)

From standard arguments on integral operators with square integrable ker-

nels, we have that there exists a Singular Value Expansion (SVE) of the kernel
(4), that is
∞

K(t, s) = μl ul (t)vl (s), t, s ∈ [0, 1], (6)
l=1

where the non-null values μ1 ≥ μ2 ≥ . . . are the singular values of K, and for
l = 1, 2, . . . , the functions ul and vl are respectively the left-singular function
and right-singular function associated with μl , see [4,10] for details.
Let K∗ be the adjoint integral operator of K, then its kernel is

K ∗ (s, t) = K(t, s), t, s ∈ [0, 1], (7)

and, from (6) we easily obtain

Kvl = μl ul , K∗ ul = μl vl , l = 1, 2, . . . , (8)

moreover
dν dν
vl = μl ul , ul = (−1)ν μl vl . (9)
dtν dtν
and from the orthonormality properties of the singular functions, we obtain that
the solution of (5) is
∞
 h, ul 
h(ν) (t) = vl (t), (10)
μl
l=1

where ·, · denotes the inner product on the space of real square integrable
functions on [0, 1].

2.2 The Singular Value Expansion of the Integral Operator

In this section we describe the fundamental formulas for the computation of the
SVE of kernel (4), these formulas have been obtained in [8], where the reader
can find the necessary details.
Let ρ2 : Z → {0, 1} be the function such that, for k ∈ Z, ρ2 (k) gives the
reminder after the division of k by 2, and let ρ−
2 (k) = 1 − ρ2 (k). For k, q, j ∈ Z,
γ ∈ R, we define
 An SVE Approach for the Numerical Solution of ODEs 73
 qπ
2q + ρ2 (ν) ν , if ν is even,
θq = π= (2q+1)π
2ν , if ν is odd,
2ν
cq = cos θq , sq = sin θq , zq = eιθq = cq + ιsq
ck,q = cos ((k + 1)θq ) , sk,q = sin ((k + 1)θq ) ,
c(γ)
q = cos (γsq ) , s(γ)
q = sin (γsq ) ,
(γ)
ck,q = cos ((k + 1)θq − γsq ) = ck,q c(γ) (γ)
q + sk,q sq , (11)
(γ)
sk,q = sin ((k + 1)θq − γsq ) = sk,q c(γ)
q − ck,q sq ,
(γ)
(12)
q γcq
αq(γ) = (−1) e ,
ν − ρ2 (ν)
η= ,
2
 T
c·,j = c0,j , c1,j , . . . , cν−1,j ∈ Rν ,
(γ) (γ) (γ) (γ)

 T
s·,j = s0,j , s1,j , . . . , sν−1,j ∈ Rν ,
(γ) (γ) (γ) (γ)

c·,j = c·,j ∈ Rν , s·,j = s·,j ∈ Rν .

(0) (0)

We have the following results.

Theorem 1. Let μl > 0 be a singular value of the integral operator K defined by
√
its kernel (4), and let γl = 1/ ν μl . Then, the singular functions corresponding
to μl are
ν−ρ2 (ν)
  
ul (t) = eγl cp t Cp(u) cos(γl sp t) + Sp(u) sin(γl sp t) , t ∈ [0, 1], (13)
p=0
ν−ρ2 (ν)
  
vl (t) = eγl cp t Cp(v) cos(γl sp t) + Sp(v) sin(γl sp t) , t ∈ [0, 1], (14)
p=0

(·) (·)
where coefficients Cp , Sp ∈ R, p = 0, 1, . . . , ν − ρ2 (ν), are defined by the fol-
lowing relations:

– if ν is odd

Cp(u) = (−1)p+1 Sp(v) , Sp(u) = (−1)p Cp(v) , (15)


ν−1 
Cp(v) ck,p − Sp(v) sk,p = 0, k = 0, 1, . . . , ν − 1, (16)
p=0
ν−1
  
(γ ) (γ )
αp(γl ) Sp(v) ck,pl + Cp(v) sk,pl = 0, k = 0, 1, . . . , ν − 1; (17)
p=0
74 N. Egidi and P. Maponi

– if ν is even
(v) (u)
S0 = Sν(v) = S0 = Sν(v) = 0, (18)
p p
Cp(u) = (−1) Cp(v) , Sp(u)
= (−1) Sp(v) , (19)
ν  
Cp(v) ck,p − Sp(v) sk,p = 0, k = 0, 1, . . . , ν − 1, (20)
p=0
ν  
(γ ) (γ )
αp(γl ) Cp(v) ck,pl − Sp(v) sk,pl = 0, k = 0, 1, . . . , ν − 1. (21)
p=0

Proof. See [8] for the proof of this Theorem. 

Theorem 2. When μl , l = 1, 2, . . . , is a singular value of integral operator K

√
defined by its kernel (4), then γl = 1/ ν μl is a zero of the function

hν : R+ → R, hν (γ) = det (M (γ))

where M (γ) is the coefficients matrix of linear system (15)–(17) or (18)–(21)

when we substitute γl with γ.
When ν = 1 we have

h1 (γ) = − cos(γ), (22)

when ν = 2 we have

h2 (γ) = −4 (1 − cos(γ) cosh(γ)) , (23)

and when ν ≥ 3 we have that the function hν satisfies the following asymptotic
relation:

hν (γ) = (−1)η+1 d2 22η−1 (2ρ2 (ν) − ρ−

2 (ν)) cos(γ)e
γξ
+ gν (γ), (24)

where gν (γ) = O(eγξ0 ) when γ → +∞, d > 0 is given by

 
d= ρ2 (ν) + 2ρ−
2 (ν) (1 + cq ) sq (−cq ) ·
η+1≤q≤ν−1
  
2 2
|zp − z q | |zp − zq | , (25)
η+1≤p<q≤ν−1

and
η−1

ξ=2 ci − ρ−
2 (ν)c0 , (26)
i=0
η−2

ξ0 = cη−1 + 2 ci − ρ−
2 (ν)c0 . (27)
i=0
 An SVE Approach for the Numerical Solution of ODEs 75

Proof. See [8] for the proof of this Theorem. 

We note that, given γl a zero of the function hν , the singular functions
associated with μl respect to the base functions (13)–(14) are given by a non
trivial solution of the corresponding linear system (15)–(17) or (18)–(21) with
coefficients matrix M (γl ). So that Theorems (1) and (2) give the main tools for
the computation of the SVE of operator K, in fact they respectively provides
the formula for the computation of the singular functions and singular values.
So, this computation of the SVE can be used to approximate the derivative
of a given function but also to solve a differential equation as shown in the
next section, where, for the sake of simplicity, we consider only the case ν = 1.
In particular in this case we have that h1 : R+ → R is given in (22) and for
l = 1, 2, . . . and t ∈ [0, 1] the SVE is
π 1 √ √
γl = + (l − 1)π, μl = , ul (t) = 2 sin(γl t), vl (t) = 2 cos(γl t), (28)
2 γl

we note that ul (0) = vl (1) = 0, vl = μl ul , ul = −μl vl .

3 Initial Value Problem for First Order Ordinary

Differential Equation
We consider an initial value problem for first order ordinary differential equation


y (x) = f (x, y(x)), x ∈ (0, 1)
(29)
y(0) = 0,

where f : [0, 1] × R → R is a known continuous function, and the solution

y : [0, 1] → R is a continuously differentiable function.
Let ỹ be an approximation to y solution of (29) such that ỹ(0) = 0, let

e(x) = ỹ(x) − y(x) (30)

r(x) = ỹ  (x) − f (x, ỹ(x)) (31)

then

e (x) = ỹ  (x) − y  (x) = r(x) + f (x, ỹ(x)) − f (x, y(x)) (32)

≈ r(x) + fy (x, ỹ(x))e(x) (33)

3.1 Numerical Solution of ODEs

Let
∞

e(x) = el ul (x), 0 ≤ x ≤ 1, (34)
l=1

then from formula (10)

76 N. Egidi and P. Maponi

∞

e (x) = el γl vl (x), 0 ≤ x < 1. (35)
l=1

We note that this last relation cannot be valid for x = 1, since in general e (1) = 0
but vl (1) = 0 for all l (see formula (28)). On the other hand, this relation can
provide a good approximation of e when this function is small.
The coefficients el , l = 1, 2, . . . of (28) are computed as the minimizer of the
following problem
 2
min e (x) − fy (x, ỹ(x))e(x) − r(x)2 , (36)
e

where ·2 denotes the 2-norm. In particular, from the first-order conditions for
the minimum, for l = 1, 2, . . . , these coefficients must satisfy
 1
  
e (x) − fy (x, ỹ(x))e(x) − r(x) γl vl (x) − fy (x, ỹ(x))ul (x) dx = 0,
0

that is
∞
  1 
em γl γm vl vm − fy (x, ỹ(x)) (γm vm ul + γl vl um ) +
m=1 0
 
 2 1 
fy (x, ỹ(x)) um ul dx = γl vl − fy (x, ỹ(x))ul r(x)dx (37)
0

We propose the following method, based on a truncation of formulas (34) and

(37).
Given yk , k = 0, 1, . . . , an approximation of the solution y of (29) such that
yk (0) = 0, we define
rk (x) = yk (x) − f (x, yk (x))
Given L > 0, let ek = (ek,1 , ek,2 , . . . , ek,L )T ∈ RL be the solution of

M ek = b (38)

where b = (b1 , b2 , . . . , bL )T is given by

 1 
bl = γl vl (x) − fy (x, yk (x))ul (x) rk (x)dx, l = 1, 2, . . . , L, (39)
0

and M ∈ RL,L has entries for l, m = 1, 2, . . . , L given by

 1

Ml,m = γl γm vl (x)vm (x) − fy (x, yk (x)) (γm vm (x)ul (x)
0
 2 
+ γl vl (x)um (x)) + fy (x, yk (x)) um (x)ul (x) dx, (40)

where we can observe that M is a symmetric matrix.

 An SVE Approach for the Numerical Solution of ODEs 77

We note that the matrix M seems to be almost positive definite, in the sense
that if we neglect the first few rows, the other rows are diagonally dominant
with positive diagonal entries. This fact suggest that, under suitable conditions,
M is non singular and that system (38) can be solved efficiently by a proper
modification of the Cholesky method.
Given
L

ek (x) = ek,l ul (x), (41)
l=1

we note that by construction ek (x) yk (x) − y(x) so that we define

yk+1 (x) = yk (x) − ek (x). (42)

The above mentioned important properties of matrix M and the convergence

analysis of {yk }k deserve a future study of the proposed method that will allow
also the evaluation of the applicability for this new method.

3.2 Implementation Details

Let N > 0, h = 1/N , and xj = jh, j = 0, 1, . . . , N or j = 1/2, 3/2, . . . , N − 1/2,
we note that x0 = 0 and xN = 1. Let yk (x) be an approximation of y(x) then
the integrals in (39) and (40) are computed by using the midpoint rule with
quadrature nodes xj+1/2 ∈ (0, 1), j = 0, 1, . . . N − 1, that is x1/2 < x3/2 < · · · <
x(2N −1)/2 , note that 0 = x0 < x1/2 < x1 < x3/2 < · · · < x(2N −1)/2 < xN = 1.
Let
Ỹ k = (Yk,0 , Yk,1 , . . . , Yk,N )T ∈ RN +1 ,
Y k = (Yk,1/2 , Yk,3/2 , . . . , Yk,(2N −1)/2 )T ∈ RN ,
where Yk,j yk (xj ).
The initial guess y (0) (x) of the recursive procedure is obtained by the Euler
method, in particular we construct:

Y0,0 = 0, Y0,i = Y0,i−1 + hf (xi−1 , Y0,i−1 ), i = 1, . . . N,

Y0,i + Y0,i+1
Y0,i+1/2 = , i = 0, 1, . . . N − 1,
2
y0 (xi+1/2 ) = f (xi , Y0,i ), i = 0, 1, . . . N − 1.
We note that, at each step k, we need yk (xi+1/2 ), i = 0, 1, . . . N − 1, because
these values are necessary to compute the vector b by (39) and (31). Given Y k
and yk (xi+1/2 ), i = 0, 1, . . . N −1, we find ek ∈ RL as solution of system (38) and

we construct Y k+1 from (41) and (42). Moreover we construct yk+1 (xi+1/2 ), i =
0, 1, . . . N −1, by differentiating (41) and (42), we note that differentiating (41) we
obtain a truncation of (35). Also the first N entries of Ỹ k are obtained by using
(41) and (42), instead Ỹk+1,N = Yk+1,N −1/2 + (h/2)f (xk+1,N −1/2 , Yk+1,N −1/2 ),
because in this case we cannot use (35).
78 N. Egidi and P. Maponi

4 Numerical Results
We present the results of some numerical experiments to evaluate the perfor-
mance of the proposed method.
In these experiments, we have considered the implementation described in
the previous section. The system (38) has been solved by using two different
strategies: the Gaussian elimination method and Gauss-Seidel method. In the
Gauss-Seidel method a fixed number P of iterations has been performed. In
particular, at each iteration k = 0, 1, . . . of the proposed method, and each
(p)
iteration p = 0, 1, . . . , P of the Gauss-Seidel method, the vector ek has been
(0)
constructed, where the initial approximation ek is chosen equal to the null
(0) (P )
vector for k = 0, instead for k ≥ 1 we have ek = ek−1 , finally the solution of
(P )
system (38) is ek = ek .
Let Ỹ ∈ RN +1 be an approximation of the exact solution of problem (29) at
nodes xi , i = 0, 1, . . . , N , i.e. y = (y(x0 ), y(x1 ), . . . , y(xN ))T ∈ RN +1 .
We consider the following performance indices given by the infinite norm and
the quadratic mean of the error, that is:
 
   
Ỹ − y 
 
E∞ (Ỹ ) = Ỹ − y  , E2 (Ỹ ) = √ 2
. (43)
∞ N +1
In particular we use the following notation:
E
E∞ = E∞ (Ỹ 0 ), E2E = E2 (Ỹ 0 ), (44)

are the errors for the initial approximation Ỹ 0 obtained with the Euler method;
G G
G,K
E∞ = E∞ (Ỹ K ), E2G,K = E2 (Ỹ K ), (45)
G
are the errors for the approximation Ỹ K obtained with the proposed algorithm
after K iterations and solving the linear systems (38) by the Gaussian elimination
method;
GS GS
GS,K
E∞ = E∞ (Ỹ K ), E2GS,K = E2 (Ỹ K ), (46)
GS
are the errors for the approximation Ỹ K obtained with the proposed algo-
rithm after K iterations and solving the linear systems (38) by the Gauss-Seidel
method, with P = 3;
To test the proposed method we use the following examples.
Example 1. The following initial-value problem


y (x) = y + 1, x ∈ (0, 1),
(47)
y(0) = 0,

has solution y(x) = ex − 1.

 An SVE Approach for the Numerical Solution of ODEs 79

Example 2. The following initial-value problem for Bernoulli equation

  
y (x) = (y + 1) A − π2 tan π4 x , x ∈ (0, 1),
(48)
y(0) = 0,

where A ∈ R, has solution

π 
y(x) = eAx cos2 x − 1.
4
Example 3. The following initial-value problem for the Riccati equation


y (x) = f2 (x)y 2 + f1 (x)y + f0 (x), x ∈ (0, 1),
(49)
y(0) = 0,

where for A > 1, and

f2 (x) = 2x, (50)

f1 (x) = 2x(1 − 2A(x − 1)2 ), (51)
f0 (x) = 2A(x − 1) − 2A x(x − 1) − 2Ax(x − 1)(1 − 2A(x − 1) ), (52)
2 4 2

has solution 
1
y(x) = A (x − 1)2 − .
A + (1 − A)e−x2

Table 1. The errors for Example 1 having solution y(x), when the approximation
is computed by: the Euler’s method (E·E ); the proposed iterative method with K
iterations and Gaussian elimination method (E·G,K ); the proposed iterative method
with K iterations and Gauss-Seidel method with P = 3 (E·GS,K ). The notation x(y)
means x · 10y .

y(x) = ex − 1
E G,1 GS,1 G,3 GS,3
N L E∞ E∞ E∞ E∞ E∞
25 5 5.24(−2) 1.45(−2) 1.61(−2) 1.45(−2) 1.44(−2)
50 10 2.67(−2) 3.98(−3) 5.33(−3) 3.98(−3) 3.86(−3)
N L E2E E2G,1 E2GS,1 E2G,3 E2GS,3
25 5 2.51(−2) 6.25(−3) 7.16(−3) 6.25(−3) 6.20(−3)
50 10 1.26(−2) 1.67(−3) 2.46(−3) 1.67(−3) 1.61(−3)

In the numerical experiments we choose L = 5, 10 and N = 25, 50, the results

are reported in Tables 1, 2 and 3 and Figs. 1, 2 and 3. From these results, we
can observe that a few iterations of the proposed method is able to substantially
decrease the error in the initial approximation Ỹ 0 computed with the Euler’s
method. So, the good results obtained with the simple implementation described
80 N. Egidi and P. Maponi

Table 2. The errors for Example 2, with A = 0.5, having solution y(x), when the
approximation is computed by: the Euler’s method (E·E ); the proposed iterative
method with K iterations and Gaussian elimination method (E·G,K ); the proposed
iterative method with K iterations and Gauss-Seidel method with P = 3 (E·GS,K ).
The notation x(y) means x · 10y .
π 
y(x) = ex/2 cos2 4
x −1
E G,1 GS,1 G,3 GS,3
N L E∞ E∞ E∞ E∞ E∞
25 5 2.25(−2) 8.14(−5) 7.41(−5) 8.14(−5) 8.14(−5)
50 10 1.11(−2) 2.84(−5) 2.43(−5) 2.84(−5) 2.84(−5)
N L E2E E2G,1 E2GS,1 E2G,3 E2GS,3
25 5 1.48(−2) 4.24(−5) 3.76(−5) 4.24(−5) 4.24(−5)
50 10 7.35(−3) 1.37(−5) 1.08(−5) 1.37(−5) 1.37(−5)

Table 3. The errors for Example 3, with A = 10, having solution y(x), when the
approximation is computed by: the Euler’s method (E·E ); the proposed iterative
method with K iterations and Gaussian elimination method (E·G,K ); the proposed
iterative method with K iterations and Gauss-Seidel method with P = 3 (E·GS,K ).
The notation x(y) means x · 10y .
 
y(x) = 10 (x − 1)2 − 1
2
10−9e−x
E G,1 GS,1 G,3 GS,3
N L E∞ E∞ E∞ E∞ E∞
25 5 3.57(−1) 4.82(−2) 4.85(−2) 4.65(−2) 4.65(−2)
50 10 1.71(−1) 1.35(−3) 1.39(−3) 1.27(−3) 1.27(−3)
N L E2E E2G,1 E2GS,1 E2G,3 E2GS,3
25 5 1.48(−1) 1.96(−2) 1.96(−2) 1.93(−2) 1.93(−2)
50 10 7.10(−2) 4.65(−4) 5.47(−4) 6.12(−4) 6.12(−4)

in Sect. 3.2 deserve further study to evaluate the effective potential of the pro-
posed method and its eventual application in the solution of partial differential
equations.
There is no much difference between performing the iterative method with
one or three iterations, independently from the method used to solve the linear
systems (38), so that in Figs. 1, 2 and 3 we report the results only for one iteration
of the method K = 1.
Moreover, the results obtained by solving the linear systems (38) with the
Gaussian elimination method or with the Gauss-Seidel method with P = 3 are
very similar, as we can see from Tables 1, 2 and 3 and in Figs. 1, 2 and 3.
So a good strategy is to implement the proposed iterative method with a
single iteration K = 1, by solving systems (38) with P = 3 iterations of the
Gauss-Seidel method.
 An SVE Approach for the Numerical Solution of ODEs 81

Ỹ 0
y
2.

1.5

1.

0.5

x
0.5 1

G
Ỹ K
y
2.

1.5

1.

0.5

x
0.5 1

GS
Ỹ K
y
2.

1.5

1.

0.5

x
0.5 1

Fig. 1. The graphs of the solution y = ex − 1 of Example 1 when x ∈ [0, 1] (the lines),
G
of its initial approximation Ỹ 0 (the dots at the top), and of its approximations, Ỹ K
GS
and Ỹ K (the dots at the middle and at the bottom, respectively), obtained with the
proposed method for N = 25, L = 5, K = 1, the corresponding errors are given in
Table 1.
82 N. Egidi and P. Maponi

Ỹ 0
y
0.2

0.1

x
0.5 1

0.1

0.2

G
Ỹ K
y
0.2

0.1

x
0.5 1

0.1

0.2

GS
Ỹ K
y
0.2

0.1

x
0.5 1

0.1

0.2

 
Fig. 2. The graphs of the solution y = ex/2 cos2 π4 x − 1 of Example 2 when A = 0.5
and x ∈ [0, 1] (the lines), of its initial approximation Ỹ 0 (the dots at the top), and of its
G GS
approximations, Ỹ K and Ỹ K (the dots at the middle and at the bottom, respectively),
obtained with the proposed method for N = 25, L = 5, K = 1, the corresponding errors
are given in Table 2.
 An SVE Approach for the Numerical Solution of ODEs 83

Ỹ 0
y

x
0.5 1

0.45

0.9

1.35

1.8

G
Ỹ K
y

x
0.5 1

0.45

0.9

1.35

1.8

GS
Ỹ K
y

x
0.5 1

0.45

0.9

1.35

1.8

 
Fig. 3. The graphs of the solution y(x) = 10 (x − 1)2 − 1
2 of Example 3 when
10−9e−x
A = 10 and x ∈ [0, 1] (the lines), of its initial approximation Ỹ 0 (the dots at the top),
G GS
and of its approximations, Ỹ K and Ỹ K (the dots at the middle and at the bottom,
respectively), obtained with the proposed method for N = 50, L = 10, K = 1, the
corresponding errors are given in Table 3.
84 N. Egidi and P. Maponi

We note that, for the first two considered examples, where the ODEs are
linear, the errors do not change by performing one or more iterations of the
proposed method, when the Gaussian elimination method is used to solve the
linear systems (38). In the third example, the errors rapidly reach their limit
values.

5 Conclusions
The derivative operator is reformulated as a Volterra integral operator, and its
SVE is used to approximate the derivative operator and to obtain a new iterative
method for the solution of ODEs. This iterative method is constructed in order
to reduce, at each step, the global error of the approximating solution. As initial
approximation is used the one obtained by Euler’s method.
Numerical experiments show that the proposed iterative method gives satis-
factory results at the first iteration, with a small number of singular functions
and a small number of iterations in the Gauss-Seidel solution of the linear sys-
tems. In fact, the initial error obtained by the Euler’s method is decreased of
one order of magnitude or more.
These numerical results confirm that by using the SVE of the reformulated
derivative operator we can construct new numerical methods to solve differential
equations. In particular, this initial study has shown interesting results for the
proposed method, even if a number of questions must be addressed, such as
the refinement of the approximation techniques defined in Sect. 3.2 to improve
the efficiency of the proposed method, the application of similar ideas to partial
differential equations, the study of convergence and applicability of the proposed
method, finding a proper modification of the Cholesky method to direct solve
the envolved linear systems.

References
1. Dahlquist, G.: Convergence and stability in the numerical integration of ordi-
nary differential equations. Math. Scand. 4, 33–53 (1956). https://doi.org/10.7146/
math.scand.a-10454
2. Henrici, P.: Discrete Variable Methods in ODEs. Wiley, New York (1962)
3. Lambert, J.D.: Computational Methods in Ordinary Differential Equations. Wiley,
New York (1973)
4. Jörgens, K.: Linear Integral Operators. Pitman Advanced Pub. Program, Boston
(1982)
5. Atkinson, K.E.: An Introduction To Numerical Analysis. Wiley, New York (1978)
6. Butcher, J.C.: Numerical Methods for Ordinary Differential Equations. Wiley, New
York (2003)
7. Lapidus, L., Seinfeld, J.: Numerical Solution of Ordinary Differential Equations.
Academic Press, New York (1971)
8. Egidi, N., Maponi, P.: The singular value expansion of the Volterra integral equa-
tion associated to a numerical differentiation problem. J. Math. Anal. Appl. 460,
656–681 (2018)
 An SVE Approach for the Numerical Solution of ODEs 85

9. Fernández, D.C.D.R., Hicken, J.E., Zingg, D.W.: Review of summation-by-parts

operators with simultaneous approximation terms for the numerical solution of
partial differential equations. Comput. Fluids 95, 171–196 (2014)
10. Kress, R.: Linear Integral Equations. Springer, Berlin (1989). https://doi.org/10.
1007/978-3-642-97146-4
11. Zhang, H., Zhang, Q.: Sparse discretization matrices for Volterra integral operators
with applications to numerical differentiation. J. Integr. Equ. Appl. 23(1), 137–156
(2011)
 Uniform Weighted Approximation
by Multivariate Filtered Polynomials

Donatella Occorsio1,2(B) and Woula Themistoclakis2

1
Department of Mathematics, Computer Science and Economics,
University of Basilicata, v.le dell’Ateneo Lucano 10, 85100 Potenza, Italy
[email protected]
2
C.N.R. National Research Council of Italy, I.A.C. Istituto per le Applicazioni del
Calcolo “Mauro Picone”, via P. Castellino, 111, 80131 Napoli, Italy
[email protected]

Abstract. The paper concerns the weighted uniform approximation of

a real function on the d−cube [−1, 1]d , with d > 1, by means of some
multivariate filtered polynomials. These polynomials have been deduced,
via tensor product, from certain de la Vallée Poussin type means on
[−1, 1], which generalize classical delayed arithmetic means of Fourier
partial sums. They are based on arbitrary sequences of filter coefficients,
not necessarily connected with a smooth filter function. Moreover, in
the continuous case, they involve Jacobi–Fourier coefficients of the func-
tion, while in the discrete approximation, they use function’s values
at a grid of Jacobi zeros. In both the cases we state simple sufficient
conditions on the filter coefficients and the underlying Jacobi weights,
in order to get near–best approximation polynomials, having uniformly
bounded Lebesgue constants in suitable spaces of locally continuous func-
tions equipped with weighted uniform norm. The results can be useful
in the construction of projection methods for solving Fredholm integral
equations, whose solutions present singularities on the boundary. Some
numerical experiments on the behavior of the Lebesgue constants and
some trials on the attenuation of the Gibbs phenomenon are also shown.

Keywords: Weighted polynomial approximation · de la Vallée Poussin

means · Filtered approximation · Lebesgue constants · Projection
methods for singular integral equations · Gibbs phenomenon

1 Introduction
Many problems arising in the applied science can be modeled by integral equa-
tions on D = [−1, 1]d and in many cases, since the solution is unbounded on
∂D, they are uniquely solvable in some Banach space of locally continuous func-
tions, equipped with weighted uniform norm (see, e.g. [15,16,18] and the ref-
erences therein). In such cases, the classical projection methods fail in view of
the unboundedness of the corresponding Lebesgue constants (LC) associated
with the applied projection (typically, Lagrange and Fourier projections). In one
c Springer Nature Switzerland AG 2020
Y. D. Sergeyev and D. E. Kvasov (Eds.): NUMTA 2019, LNCS 11973, pp. 86–100, 2020.
https://doi.org/10.1007/978-3-030-39081-5_9
 Uniform Weighted Approximation by Multivariate Filtered Polynomials 87

dimension [17] a solution has been offered by certain quasi–polynomial projec-

tions, which generalize trigonometric de la Vallée Poussin means (briefly VP
means) and fall into the class of filtered polynomials (see e.g. [5,10,13,22–24]).
They have been successfully applied in [17] to construct a numerical method
which is uniformly convergent and stable in [−1, 1]. Here we provide the basic
theorems for the extension to higher dimensions.
We introduce and study multivariate filtered polynomial sequences deduced,
via tensor product, from the univariate generalized VP means recently stud-
ied in [24]. We state simple sufficient conditions in order that these polynomial
sequences converge (in a given weighted uniform norm) to the function they aim
to approximate with uniformly bounded related LC. This remarkable behavior is
especially desirable in proving the convergence and the stability of some numer-
ical methods for solving several functional equations (see e.g. [4,6,7,9,19,25]).
Moreover, from the general approximation view–point, our results offer many
different possibilities of factorizing the function in the product of a weight func-
tion to compensate the singular factor, and a smooth factor, which can be
approximated very well by polynomials.
The numerical experiments focus on the bidimensional case D = [−1, 1]2
where several optimal systems of nodes (e.g. Padua points, Xu points) have
been already constructed. Nevertheless, most of them cannot be used in the case
of functions unbounded at ∂D, since some nodes lie on the boundary of D (see
e.g. [1–3,26]). Comparisons with these kinds of polynomial approximation can
be found in the forthcoming paper by the authors [20], being them outside the
scope of this paper. Indeed, our focus mainly deals with the uniform weighted
approximation of functions which can be also unbounded at the border of the
square.
The outline of the paper is the following. Section 2 introduces some notation
and it deals with classical Fourier and Lagrange polynomial projections. Section 3
is devoted to filtered approximation in the continuous and the discrete case.
Finally, Sect. 4 contains some numerical experiments.

2 Basic Definitions and Properties

In order to approximate a function f on D, we consider multivariate polynomials

in the variable x := (x1 , . . . , xd ) ∈ D, which are of degree at most mk w.r.t. the
variable xk , for any k = 1, . . . , d (briefly ∀k ∈ N1d ). Setting m := (m1 , . . . , md ) ∈
INd , this polynomial space will be denoted by Pm .
Denoting by v γ,δ the Jacobi weight

v γ,δ (z) := (1 − z)γ (1 + z)δ , z ∈ (−1, 1), γ, δ > −1,

throughout the paper, for any γk , δk ≥ 0, k ∈ N1d , we set


d
u(x) = v γ1 ,δ1 (x1 )v γ2 ,δ2 (x2 ) · · · v γd ,δd (xd ) := uk (xk ), (1)
k=1
88 D. Occorsio and W. Themistoclakis

and we consider functions f that can be also unbounded on ∂D, but they are
such that
fu∞ = max |f (x)|u(x) < ∞.
x∈D

We denote by L∞
:= {f : fu∞ < ∞} the Banach space equipped with the
u
previous norm.
Moreover, we denote by Em (f )u the error of best weighted polynomial
approximation by means of polynomials in Pm , i.e.

Em (f )u = inf (f − P)u∞ .

P∈Pm

The extension of the Weierstrass theorem to D ensures that

lim Em (f )u = 0, ∀f ∈ Cu ,
m→∞

where Cu ⊂ L∞ u denotes the subspace of all functions that are continuous in the
interior of D and tend to zero as x tends to the edges of D (i.e. as xk → ±1, in
correspondence of some γk , δk > 0, k ∈ N1d ).
Along all the paper the constant C will be used several times, having different
meaning in different formulas. Moreover we write C = C(a, b, . . .) in order to say
that C is a positive constant independent of the parameters a, b, . . .
Finally, for r = (r1 , . . . , rd ) and s = (s1 , . . . , sd ) in INd , we set

IN[r, s] := {i = (i1 , . . . , id ) ∈ INd : rk ≤ ik ≤ sk , k ∈ N1d },

where for r = 0, 1 we agree that 0 = (0, . . . , 0) and 1 = (1, . . . , 1).

Along the paper, at the occurrence, other notations will be introduced. For
the convenience of the reader a list of notations is given at the end of the paper.

2.1 Optimal Polynomial Projectors

Setting
wk := v αk ,βk , αk , βk > −1, k = 1, . . . , d,
we denote by {pm (wk )}∞ m=0 the corresponding sequence of the univariate
orthonormal Jacobi polynomials with positive leading coefficients.
For any x = (x1 , . . . , xd ), we set


d 
d
w(x) := wk (xk ) = (1 − xk )αk (1 + xk )βk ,
k=1 k=1

and for any m = (m1 , . . . , md ) ∈ INd , we consider in Pm the following polynomial

basis
d
pm (w, x) := pmk (wk , xk ).
k=1
 Uniform Weighted Approximation by Multivariate Filtered Polynomials 89

Then the d–dimensional Fourier polynomial Sm (w, f , x) of f w.r.t. w is given

by

Sm (w, f , x) = ci (w, f )pi (w, x),
i∈IN[0,m]

ci (w, f ) = f (y)pi (w, y)w(y)dy. (2)
D

This polynomial can be deduced via tensor product from one dimensional Jacobi–
Fourier projections.
By using [14, Th. 4.34, p. 276], the following result can be easily proved
Theorem 1. Let w and u satisfy the following bounds
⎧
⎪ αk 1 αk 3
⎪
⎨ 2 + 4 < γk ≤ 2 + 4 and 0 ≤ γk < αk + 1,
∀k ∈ N1d . (3)
⎪
⎪
⎩ βk + 1 < δk ≤ βk + 3 and 0 ≤ δk < βk + 1,
2 4 2 4
Then for any f ∈ L∞
u and m ∈ IN it follows that
d

Sm (w, f )u∞ ≤ Cfu∞ log m, C = C(f , m), (4)

d
where log m := k=1 log mk .
This result and the invariance property

Sm (w, f ) ≡ f , ∀f ∈ Pm , (5)

yield the following

Corollary 1. Under the assumptions of Theorem 1, we have

Em (f )u ≤ [f − Sm (w, f )]u∞ ≤ CEm (f )u log m, C = C(f , m). (6)

m
Denoted by {xmk , (wk )}=1k
the zeros of the univariate polynomial pmk (wk ) and
by {λmk , (wk )}=1 the associated Christoffel numbers, for any i = (i1 , . . . , id ) ∈
mk

IN[1, m] we set
(m)
xi (w) = (xm1 ,i1 (w1 ), ..., xmk ,ik (wk ), ..., xmd ,id (wd )),

(m)

d
Λi (w) = λmk ,ik (wk ).
k=1

Then the Gauss-Jacobi cubature formula of order m obtained via tensor product
of univariate Gauss-Jacobi rules, is given by
  (m) (m)
f (x)w(x)dx = Λi (w)f (xi (w)) + Rm (w, f ), (7)
D
i∈IN[1,m]
90 D. Occorsio and W. Themistoclakis

where the remainder term Rm (w, f ) vanishes when f ∈ P2m−1 .

By discretizing the Fourier coefficients of Sm (w, f ) in (2) by the previous
rule (7) of order (m + 1), we get the following polynomial
 (m+1) (m+1) (m+1)
Lm (w, f , x) = Λi (w)Km (w, x, xi (w))f (xi (w)), (8)
i∈IN[1,m]

where

Km (w, x, y) := pr (w, y)pr (w, x).
r∈IN[0,m]

By definition it follows that Lm (w, f ) ∈ Pm and

Lm (w, f ) ≡ f , ∀f ∈ Pm .

Moreover, it is easy to check that the following interpolation property holds

(m+1) (m+1)
Lm (w, f , xi (w)) ≡ f (xi (w)), ∀i ∈ IN[1, (m + 1)],

i.e. Lm (w, f ) is the multivariate Lagrange polynomial in Pm interpolating f at

the zeros of pm+1 (w). The following approximation theorem generalizes [18,
Prop. 2.1] obtained in the case d = 2.
Theorem 2. Let w and u be such that
⎧

⎪ αk 1 αk 5
⎪
⎨ max 0, 2 + 4 ≤ γk ≤ 2 + 4 ,

 k ∈ N1d . (9)
⎪
⎪ β 1 βk 5
⎩ max 0, k
+ ≤ δk ≤ + ,
2 4 2 4

Then, for any f ∈ L∞

u and m ∈ IN , we have
d

Lm (w, f )u∞ ≤ Cfu∞ log m, (10)

and
[f − Lm (w, f )]u∞ ≤ CEm (f )u log m, (11)
where in both the estimates C = C(f , m).

3 Main Results
Setting N = (N1 , . . . , Nd ) ∈ INd and M = (M1 , . . . , Md ) ∈ INd , with Nk < Mk
for all k ∈ N1d , let us consider d uniformly bounded sequences of filter coefficients
{hN
k ,Mk
}l=1,2,.. , such that

Nk ,Mk 1, if ∈ [0, Nk ]
h = l = 1, 2, . . . , ∀k ∈ N1d . (12)
0, if ∈
/ [0, Mk ],
 Uniform Weighted Approximation by Multivariate Filtered Polynomials 91

As pointed out in [24] these coefficients can be either samples of filter func-
tions having different smoothness (see e.g. [12,21] and the references therein) or
they can be connected to no filter function. This last one is the case of some
generalized de la Vallée Poussin means, firstly introduced in [11] by using the
convolution structure of orthogonal polynomials.
In what follows we assume N ∼ M which means

Nk < Mk ≤ CNk , ∀k ∈ N1d , C = C(N, M).

Moreover, we set


d
hN,M
i := hN
ik
k ,Mk
, i = (i1 , . . . , id ) ∈ INd .
k=1

3.1 Continuous Filtered Approximation

By means of the previous filter coefficients, we define the following filtered Fourier
sum (or generalized de la Vallée Poussin mean)

VNM
(w, f , x) = hN,M
i ci (w, f )pi (w, x). (13)
i∈IN[0,M]

Note that this polynomial is a weighted delayed mean of the previous Fourier
sums. Indeed, the following summation by part formula


M 
M −1 
j
aj bj = aM sM + sj (aj − aj+1 ), sj := br , (14)
j=N j=N r=N

and the assumptions in (12), yield


M
VN (w, f , x) = dN,M
r Sr (w, f , x), (15)
r∈IN[N,M]

where we set

d
dN,M
r = (hN
rk
k ,Mk
− hN k ,Mk
rk +1 ), r = (r1 , . . . , rd ).
k=1

M
We observe that VN (w, f ) ∈ PM . Moreover, by (12) we easily get
M
VN (w, f ) = SN (w, f ) = f , ∀f ∈ PN . (16)

The following theorem supplies conditions under which the operator

M
VN (w) : f ∈ L∞ M ∞
u → VN (w, f ) ∈ PM ⊂ Lu is uniformly bounded w.r.t. N ∼ M.
It can be derived from the univariate case in [24, Theorem 3.1].
92 D. Occorsio and W. Themistoclakis

Theorem 3. Assume that the Jacobi weights u and w satisfy

γk − δk − αk − βk < 1, ∀k ∈ N1d (17)
2

and
⎧
⎪ αk 1 αk 5
⎪
⎨ 2 − 4 < γk ≤ 2 + 4 and 0 ≤ γk < αk + 1,
∀k ∈ N1d . (18)
⎪
⎪
⎩ βk − 1 < δk ≤ βk + 5 and 0 ≤ δk < βk + 1,
2 4 2 4
Moreover, assume that N ∼ M and that besides (12), the filter coefficients defin-
M
ing VN (w, f ) satisfy
Mk

2 Nk ,Mk C
Δ h ≤ , C = C(Nk , Mk ), k ∈ N1d , (19)
Nk
=Nk

where, as usual, Δ2 hN

k ,Mk
= hNk ,Mk
+2 − 2hNk ,Mk
+1 + hN

k ,Mk
, ∀k ∈ N1d .
∞
Then for any f ∈ Lu , we have
M
VN (w, f )u∞ ≤ Cfu∞ , C = C(N, M, f ). (20)
By the previous theorem, in view of the invariance property (16), the follow-
ing corollary comes down
Corollary 2. Under the assumptions of Theorem 3, for all f ∈ L∞
u , it is
M
EM (f )u ≤ [f − VN (w, f )]u∞ ≤ CEN (f )u , C = C(f , N, M). (21)

3.2 Discrete Filtered Approximation

Now we introduce the following discrete version of (13)

M
hN,M
(m)
VN (w, f , x) = i ci (w, f )pi (w, x) (22)
i∈IN[0,M]

(m)

ci (w, f ) = Λ(m)
r (w)f (x(m)
r )pi (w, x(m)
r ),
r∈IN[1,m]

obtained by approximating the coefficients ci (w, f ) in (13) by the m − th Gauss-

Jacobi rule in (7), with  
M+N+1
m= .
2
This choice of m assures that
M M
VN (w, f , x) = VN (w, f ) = f , ∀f ∈ PN .
Moreover, from [24, Theorem 4.1], the following near–best approximation result
can be deduced.
 Uniform Weighted Approximation by Multivariate Filtered Polynomials 93

Theorem 4. Assume that u and w satisfy (17) and

⎧
⎪ αk 1 αk 5
⎪
⎨ 2 − 4 < γk ≤ 2 + 4 and γk ≥ 0,
∀k ∈ N1d . (23)
⎪
⎪
⎩ βk − 1 < δk ≤ βk + 5 and δk ≥ 0,
2 4 2 4
Moreover, let be N ∼ M, and assume that the filter coefficients defining
M
VN (w, f ) satisfy (12) and (19).
Then for any f ∈ L∞ u , we have

M
VN (w, f )u∞ ≤ Cfu∞ , C = C(N, M, f ) (24)

and
M
EM (f )u ≤ [f − VN (w, f )]u∞ ≤ CEN (f )u , C = C(M, N, f ). (25)
In conclusion, under the assumptions of the previous theorems, we have that
M M
for any f ∈ Cu , both the polynomials VN (w, f ) and VN (w, f ) are near–best
approximation polynomials converging to f as N → ∞ and M ∼ N, with the
same order of the best polynomial approximation of f .

4 Numerical Experiments

Now we propose some tests exploiting the behaviors of the discrete filtered poly-
nomials previously introduced, in the case D = [−1, 1]2 .
We point out that in all the experiments we focus on the number m1 · m2 of
nodes determined by m = (m1 , m2 ). For simplicity, we fix m1 = m2 = m and,
for any given 0 < θ < 1, we take

N = (1 − θ)m, and M = (1 + θ)m,

so that the assumption N ∼ M is satisfied. Moreover, unless stated otherwise,

we set θ = 0.6.
As filter coefficients, we take


Nk ,Mk − Nk
h =h 1+ , ∈ IN, k ∈ {1, 2},
M k − Nk + 1

where the filter function h(x) vanishes outside of [0, 2], it is equal 1 on [0, 1] and
for x ∈ (1, 2] it is given by
⎧
⎪ 2 − x, classic de la Vallée Poussin filter (VP filter) ,
⎪
⎪
⎪
⎨
sin(π(x−1))
h(x) = π(x−1) , Lanczos filter, (26)
⎪
⎪
⎪
⎪
⎩ 1+cos(π(x−1))
2 , raised-cosine filter.
94 D. Occorsio and W. Themistoclakis

Note that all the previous filter choices satisfy the hypotheses (12) and (19),
required in our main results.
Moreover, all the computations have been performed in double-machine pre-
cision (eps ∼ 2.220446049250313e − 16) by MatLab version R2018a.
In the first test we want to compare the approximation provided by the
discrete filtered polynomial VN,M (w, f ) corresponding to the classical VP filter,
with that one of the Lagrange polynomial L  m−1 (w, f ) based on the same nodes.
To this aim, for increasing m, we compute the respective Lebesgue constant
(LC), namely the following operator norms
Lm−1 (w, f )u∞
Lm−1 (w)L∞ := sup ,
u
f =0 fu∞
VN,M (w, f )u∞
VN,M (w)L∞ := sup ,
u
f =0 fu∞

according to the formulas


(m) (m) u(x)
Lm−1 (w)L∞
u
= sup Λi Km (w, x, xi (w))
(m)
,
x∈D u(xi (w))
i∈IN[1,m]

M (m) u(x)
Λi v N,M (w, x, xi (w))
(m)
VN (w)L∞
u
= sup (m)
,
x∈D u(x (w))
i∈IN[1,m] i

where

Km (w, x, y) := pr (w, x)pr (w, y),
r∈IN[0,m]

v N,M (w, x, y) := hN,M
r pr (w, x)pr (w, y).
r∈IN[0,M]

Fig. 1. LC of Lagrange and filtered operators for αk = βk = −0.7, γk = δk = 0,

k = 1, 2.
 Uniform Weighted Approximation by Multivariate Filtered Polynomials 95

Fig. 2. LC of filtered operator for αk = βk = −0.7, γk = δk = 0, k = 1, 2.

The results are displayed in Fig. 1 for the unweighted case u(x) = 1 and,
for the sake of clarity, the plot of LC associated with filtered approximation is
repeated in Fig. 2.
The second experiment concerns the weighted approximation, namely the
case u(x) = 1. In order to test the goodness of the convergence ranges in the
assumptions (17) and (23), we computed, for increasing m, the LC of the filtered
operator associated with the classical VP filter for two different choices of w.
Figure 3 shows the resulting behaviors in a case where the previous assumptions
hold (left plot), and in another case where they do not (right plot).

Fig. 3. LC for parameters αk = βk = 0.5, γk = δk = 1 (k = 1, 2) on the left hand

side, and parameters (out of ranges) αk = βk = 0.6, γk = δk = 2 (k = 1, 2) on the
right–hand side

In the third experiment we are going to fix the couple of weights uk = 1, wk =

1 1
v − 2 ,− 2 , k = 1, 2, and we let to vary the filter coefficients or the parameter θ.
More precisely, in Fig. 4 we show the behavior of VN,M (w)L∞ u
for the different
filter functions in (26), while Fig. 5 displays the LC corresponding to the classic
VP filter and θ varying in {0.1 : 0.1 : 0.9}.
96 D. Occorsio and W. Themistoclakis

Fig. 4. LC for different filters and αk = βk = −0.5, γk = δk = 0, k ∈ N12 .

Fig. 5. LC for several theta and αk = βk = −0.5, γk = δk = 0, k = 1, 2.

Finally, the case of classical VP filter and variable θ is also shown in Fig. 6
for a different choice of the weight w defining the filtered operator VN,M (w).
The last experiment deals with the Gibbs phenomenon. We consider the
approximation of the following two bounded variation functions

f1 (x) = sign(x1 ) + sign(x2 ), x = (x1 , x2 ) ∈ [−1, 1]2 ,

and 
1, if x21 + x22 ≤ 0.62
f2 (x) = x = (x1 , x2 ) ∈ [−1, 1]2 ,
0, otherwise
the last function being also considered in [8].
In these cases it is well–known that Lagrange polynomials present overshoots
and oscillations close to the singularities, which are preserved also in the regular
 Uniform Weighted Approximation by Multivariate Filtered Polynomials 97

Fig. 6. LC for several theta and αk = βk = 0.2, γk = δk = 0, k = 1, 2.

Fig. 7. Lagrange and filtered (VP filter) polynomials of the function f1 (plotted at the
top) for m = 50 (2500 nodes) and parameters αk = βk = γk = δk = 0 (k = 1, 2).

part of the function. Figures 7 and 8 show that this phenomenon can be strongly
reduced if we take the discrete filtered (VP filter) polynomial based on the same
nodes set and a suitable value of the parameter θ ∈ (0, 1).
98 D. Occorsio and W. Themistoclakis

Fig. 8. Lagrange and filtered (VP filter) polynomial of the function f2 (plotted at the
top) for m = 300 and parameters αk = βk = −0.5, γk = δk = 0 (k = 1, 2).

List of Notations

– x = (x1 , x2 , . . . , xd ) and similarly for other bold letters (N, M, i etc.) denoting
vectors
– D = [−1, 1]d , d > 1 (d = 2 in Sect. 4)
– N1d = {1, . . . , d}
– IN[r, s] = {i = (i1 , . . . , id ) ∈ INd : ik ∈ [rk , sk ], ∀k ∈ N1d }
d
– u(x) = v γ1 ,δ1 (x1 ) · v γ2 ,δ2 (x2 ) · · · v γd ,δd (xd ) = k=1 uk (xk )
d
– w(x) = v α1 ,β1 (x1 ) · v α2 ,β2 · · · v αd ,βd (xd ) = k=1 wk (xk )
– pm (w, t) denotes the value at t of the univariate orthonormal Jacobi polyno-
mial of degree
d m associated with w and having positive leading coefficient
– pi (w, x) = k=1 pik (wk , xk )
– xm, (w) ( = 1, . . . , m) are the zeros of pm (w, t)
(m)
– xi (w) = (xm1 ,i1 (w1 ), xm2 ,i2 (w2 ), . . . , xmd ,id (wd )),
– λm, (w) ( = 1, . . . , m) are the univariate Christoffel numbers of order m
m−1
associated with w, i.e. λm, (w) = [ j=0 p2j (w, xm, (w))]−1
(m) d
– Λi (w) = k=1 λmk ,ik (wk )
d
– log m = k=1 log mk , for any m = (m1 , . . . , md ) ∈ INd
d
– hN,M
i = k=1 hN ik
k ,Mk

Acknowledgments. The authors are grateful to the anonymous referees for carefully
reviewing this paper and for their valuable comments and suggestions.
This research has been accomplished within Rete ITaliana di Approssimazione
(RITA) and partially supported by the GNCS-INdAM funds 2019, project “Discretiz-
zazione di misure, approssimazione di operatori integrali ed applicazioni”.
 Uniform Weighted Approximation by Multivariate Filtered Polynomials 99

References
1. Bos, L., Caliari, M., De Marchi, S., Vianello, M.: A numerical study of the Xu
polynomial interpolation formula in two variables. Computing 76, 311–324 (2006)
2. Bos, L., Caliari, M., De Marchi, S., Vianello, M., Xu, Y.: Bivariate Lagrange inter-
polation at the Padua points: the generating curve approach. J. Approx. Theory
143, 15–25 (2006)
3. Caliari, M., De Marchi, S., Vianello, M.: Bivariate Lagrange interpolation at the
Padua points: computational aspects. J. Comput. Appl. Math. 221, 284–292 (2008)
4. Capobianco, M.R., Criscuolo, G., Junghanns, P., Luther, U.: Uniform convergence
of the collocation method for Prandtl’s integro-differential equation. ANZIAM J.
42(1), 151–168 (2000)
5. Capobianco, M.R., Themistoclakis, W.: Interpolating polynomial wavelets on
[−1, 1]. Adv. Comput. Math. 23(4), 353–374 (2005)
6. De Bonis, M.C., Occorsio, D.: Quadrature methods for integro-differential equa-
tions of Prandtl’s type in weighted uniform norms. AMC (to appear)
7. De Bonis, M.C., Occorsio, D.: On the simultaneous approximation of a Hilbert
transform and its derivatives on the real semiaxis. Appl. Numer. Math. 114, 132–
153 (2017)
8. De Marchi, S., Erb, W., Marchetti, F.: Spectral filtering for the reduction of the
gibbs phenomenon for polynomial approximation methods on Lissajous curves with
applications in MPI. Dolom. Res. Notes Approx. 10, 128–137 (2017)
9. Fermo, L., Russo, M.G., Serafini, G.: Numerical methods for Cauchy bisingular
integral equations of the first kind on the square. J. Sci. Comput. 79(1), 103–127
(2019)
10. Filbir, F., Mhaskar, H.N., Prestin, J.: On a filter for exponentially localized kernels
based on Jacobi polynomials. J. Approx. Theory 160, 256–280 (2009)
11. Filbir, F., Themistoclakis, W.: On the construction of de la Vallée Poussin means
for orthogonal polynomials using convolution structures. J. Comput. Anal. Appl.
6, 297–312 (2004)
12. Gottlieb, D., Shu, C.-W.: On the Gibbs phenomenon and its resolution. SIAM Rev.
39(4), 644–668 (1997)
13. Mhaskar, H.N.: Localized summability kernels for Jacobi expansions. In: Rassias,
T.M., Gupta, V. (eds.) Mathematical Analysis, Approximation Theory and Their
Applications. SOIA, vol. 111, pp. 417–434. Springer, Cham (2016). https://doi.
org/10.1007/978-3-319-31281-1 18
14. Mastroianni, G., Milovanovic, G.: Interpolation Processes. Basic Theory and
Applications. Springer Monographs in Mathematics. Springer, Heidelberg (2008).
https://doi.org/10.1007/978-3-540-68349-0
15. Mastroianni G., Milovanović, G., Occorsio, D.: A Nyström method for two variables
Fredholm integral equations on triangles. Appl. Math. Comput. 219, 7653–7662
(2013)
16. Mastroianni, G., Russo, M.G., Themistoclakis, W.: The boundedness of the Cauchy
singular integral operator in weighted Besov type spaces with uniform norms.
Integr. Eqn. Oper. Theory 42, 57–89 (2002)
17. Mastroianni, G., Themistoclakis, W.: A numerical method for the generalized air-
foil equation based on the de la Vallée Poussin interpolation. J. Comput. Appl.
Math. 180, 71–105 (2005)
18. Occorsio, D., Russo, M.G.: Numerical methods for Fredholm integral equations on
the square. Appl. Math. Comput. 218(5), 2318–2333 (2011)
100 D. Occorsio and W. Themistoclakis

19. Occorsio, D., Russo, M.G.: Nyström methods for Fredholm integral equations using
equispaced points. Filomat 28(1), 49–63 (2014)
20. Occorsio, D., Themistoclakis, W.: Uniform weighted approximation on the square
by polynomial interpolation at Chebyshev nodes. Submitted to a Special Iusse of
NUMTA 2019
21. Sloan, I.H., Womersley, R.S.: Filtered hyperinterpolation: a constructive polyno-
mial approximation on the sphere. Int. J. Geomath. 3, 95–117 (2012)
22. Themistoclakis, W.: Weighted L1 approximation on [−1, 1] via discrete de la Vallée
Poussin means. Math. Comput. Simul. 147, 279–292 (2018)
23. Themistoclakis, W.: Uniform approximation on [−1, 1] via discrete de la Vallée
Poussin means. Numer. Algorithms 60, 593–612 (2012)
24. Themistoclakis, W., Van Barel, M.: Generalized de la Vallée Poussin approxima-
tions on [−1, 1]. Numer. Algorithms 75, 1–31 (2017)
25. Themistoclakis, W., Vecchio, A.: On the solution of a class of nonlinear systems
governed by an M -matrix. Discrete Dyn. Nat. Soc. 2012, 12 (2012)
26. Xu, Y.: Lagrange interpolation on Chebyshev: points of two variables. J. Approx.
Theory 87(2), 220–238 (1996)
Computational Methods for Data
Analysis
 A Travelling Wave Solution for Nonlinear
Colloid Facilitated Mass Transport in
Porous Media

Salvatore Cuomo1(B) , Fabio Giampaolo1 , and Gerardo Severino2

1
Department of Mathematics, University of Naples - Federico II, Naples, Italy
[email protected], [email protected]
2
Department of Agricultural Sciences, University of Naples - Federico II,
Naples, Italy
[email protected]

Abstract. Colloid facilitated solute transport through porous media is

investigated. Sorption on the matrix is modelled by the linear equilib-
rium isotherm whereas sorption on colloidal sites is regulated by non-
linearly equilibrium vs nonequilibrium. A travelling wave-type solution
is obtained to describe the evolution in both the liquid and colloidal
concentration.

Keywords: Nonlinear transport · Colloids · Travelling wave

1 Introduction

Aquifers’ contamination and adverse effects on the environment have become

a matter of considerable concern. Hence, it would be desirable having a pre-
dicting model to asses the effects of such a contamination risk [13]. A large
number of mathematical models have been developed in the past [11,12,15].
Current models often emphasize either physical or chemical aspects of trans-
port [2,8]. However, a general approach relies upon multicomponent transport
modelling, and it has been only recently formulated in [6]. In this case, both
adsorption and complexation in solution were taken into account along with
precipitation/dissolution phenomena. In particular, being concerned with com-
plex chemistry, such a model comes with non linear transport equations. The
main interest related to the solution for such a system of PDEs is that it allows
one to analyze the sensitivity of mass transport to the variation of different
parameters. Unfortunately, generally analytical (closed form) solutions are not
achievable, and concurrently one has to resort with numerical approximation.
However, such a stand point may arise two serious issues. One is about the
discretization in order to suppress instability and numerical dispersion [9]. The
other is that in some situations (typically at the very beginning of the transport
process) very steep concentration(s) gradients may develop. To accurately mon-
itor these gradients, very fine discretizations may be necessary: an undesirable
c Springer Nature Switzerland AG 2020
Y. D. Sergeyev and D. E. Kvasov (Eds.): NUMTA 2019, LNCS 11973, pp. 103–108, 2020.
https://doi.org/10.1007/978-3-030-39081-5_10
104 S. Cuomo et al.

situation in view of computational times [4]. Alternatively, one can introduce a

few approximations which enable one to gain analytical solution. In particular,
in the present paper a travelling wave solution is derived for the case of colloid
facilitaed mass transport with nonlinear sorption/desorption.

2 The Transport Model

We consider steady flow with constant velocity U in a one dimensional porous
medium that carries simultaneously solute and colloidal particles. In such a
system the mass flux vectors for the colloids and solute are

∂C (z, t)
Jc (z, t) = nCc (z, t) Uc (z, t) , J (z, t) = nC (z, t) U (z, t) − D (1)
∂z
respectively, where Cc and C are the concentrations (per unit volume of fluid) of
colloidal and solute particles, D is the pore scale dispersion and n the porosity.
Generally, advection colloidal velocity Uc is larger than U , due to “exclusion
phenomena”. We model this ‘exclusion’ process by assuming that Uc depends
linearly upon U through a constant coefficient Re ≥ 1, i.e. Uc = Re U . Colloidal
generation/removal is quantified by the mass balance equation for colloidal par-
ticles. However, in the context of the present paper we neglect both generation
and removal, such that one can assume that the concentration Cc of colloidal par-
ticles is uniform. The applicability of such an assumption is throughly discussed
in [5].
Let S denote the solute concentration on colloids, defined per unit colloidal
concentration; the actual solute concentration Ccs on colloidal particles is SCc .
Furthermore, we denote with N the solute concentration (mass of sorbed solute
per unit bulk volume) sorbed on the porous matrix. Thus, the total solute con-
centration writes as Ct = n (C + Ccs ) + N , and concurrently the (solute) mass
balance equation is
∂ ∂
[n (C + Ccs ) + N ] + (Jc + J) = 0. (2)
∂t ∂z
Substitution of (1) into (2) leads to the reaction diffusion equation

 
∂C ∂Ccs ∂ 1 ∂N ∂2C D
+ +U (Re Ccs + C) = − +D 2 D= (3)
∂t ∂t ∂z n ∂t ∂z n

where, for simplicity, we have regarded n and D as constant. Generally, both

Ccs and N depend upon C in a very complex fashion, however we shall assume
that sorption on the porous medium is governed only by the linear equilibrium
model, i.e. N = nKd C (Kd is the linear partitioning coefficient between the fluid
and sorbed phase), and we mainly focus on the effect of mass exchange between
the fluid and colloidal solute concentration. For this reason, we consider a quite
general dependence of Ccs upon C accounting for a nonlinear mass exchange, i.e.
 Nonlinearly Colloid Facilitated Solute Transport 105

∂ s
C = L [ϕ (C) − Ccs ] , (4)
∂t c
being L a given rate transfer coefficient. We assume that solute is continuously
injected at z = 0:

C (0, t) = C0 , (5)
whit zero initial C-concentration. The nonlinear reaction function ϕ describes
the equilibrium between the two phases, and its most used expressions are those
of Langmuir and Freudlich [8]. When kinetics is fast enough, i.e. L  1, the
left hand side of (4) may be neglected, up to a transitional boundary layer [10],
therefore leading to a non linear equilibrium sorption process.

3 Travelling Wave Solution

Overall, the system of Eqs. (3)–(4) can not be solved analytically. To obtain a
simple solution (which nevertheless keeps the main physical insights), we consider
a travelling wave type solution. Specifically, we assume that a travelling wave
solution, that generally occurs after a large time. This allows one to assume that
C at the inlet boundary is approximately equal to the feed concentration C0 , so
that the initial/boundary conditions may be approximate as follows

C0 for z = −∞
C(z, t)  (6)
0 for z = +∞,
for any t ≥ 0. We introduce the moving coordinate system η = z − αt, where α
represents the constant speed of the travelling wave (that will be determined later
on). We assume that each concentration, both in the liquid and in the colloidal
phase, moves with the same velocity α, which authorizes to write C (z, t) = C (η)
and Ccs (z, t) = Ccs (η). Of course this approximation works better and better as
the time increases. With these assumptions, the system (3)–(4) is reduced to
an ODE-problem
⎧
⎪
⎪ d
−α (RC + Ccs ) + U
d
(C + Ccs ) = 0
⎪
⎪
⎨ dη dη
(7)
⎪
⎪
⎪
⎪ d
⎩−α Ccs = L [ϕ (C) − Ccs ]
dη
with boundary conditions given by

C (−∞) = C0 , C (+∞) = 0 (8)

Ccs (−∞) = ϕ0 , Ccs (+∞) = ϕ (0) = 0, (9)

where ϕ0 ≡ ϕ (C0 ). We now identify the wave velocity α by integrating the first
of (7) over η
106 S. Cuomo et al.

− α (RC + Ccs ) + U (C + Ccs ) = A = const. (10)

From boundary conditions (8)–(9), one has
 
C0 + ϕ0
A=0 ⇒ α=U . (11)
RC0 + ϕ0
that inserted in (7) yields, after some algebra, the following boundary value
problem for C:
⎧

⎨ d C = τ ϕ0 C − ϕ (C)
0
dη C0 (12)
⎩
C (−∞) = C0 , C (+∞) = 0,
with the constant τ0 defined as:
 
L L ϕ0 + RC0
τ0 = = . (13)
α U ϕ0 + C0

Solution of (12) is possible if ϕ ≡ ϕ (x) is a “convex” isotherm (i.e. d2 ϕ/dx2 <

0). This is always the case for Freundlich/Langmuir type ϕ-function. Moreover,
the presence of the reaction rate L gives smooth concentration profiles that
become steeper and steeper as L increases, in particular for L → ∞ solution
of (12) asymptotically behaves as the one pertaining to the (nonlinear) equi-
librium [8]. For the application to the quantification of the pollution risk, it is
important to quantify the extent of the concentration front Sf . This latter exists
if and only if
 ε
dx
<∞ for any ε > 0. (14)
0 ϕ (x)
The differential equation in (12) is expressed in integral form upon integration
with respect to an arbitrary reference C (ηr ) = Cr . The final result is
 Cr
dx
τ0 (ηr − η) = ϕ0 = G (C, Cr ) , (15)
C x − ϕ (x)
C0
being the shape of the G-function depending upon the structure of the reaction
function ϕ. Since the solution (15) is given as function of an arbitrary reference
value Cr ≡ C (ηr ), it is important to identify the position ηr pertaining to such a
concentration value. The adopted approach relies on the method of moments [1,
3,7]. In particular, we focus on the zero-order moments [L], i.e.
 +∞  +∞
1 1
ηX = dη C (η) , ηY = dη ϕ (η) . (16)
C0 −∞ ϕ0 −∞
Before proceeding further, we wish noting here that when the number (say n)
of unknown parameters is greater than 1, one has to sort with moments of
 Nonlinearly Colloid Facilitated Solute Transport 107

order 0, 1, . . . , n − 1. The moment ηX can be regarded as the distance along

which C has increased from 0 to C0 (at a fixed time t). A similar physical insight
can be attached to the moment ηY pertaining to the sorbed concentration on
colloidal sites. While integrals (16) are not bounded, their difference it is. Indeed,
integration from η = −∞ to η = +∞ of the ODE in (12) by virtue of (16) leads
to ηY − ηX = τ0Cϕ0 0 . We can now focus on how to use moments (16) to identify
the reference point ηr . Thus, we integrate (15) from 0 to C0 , and division by

C0 results in ηr − ηX = τ0GC0 , with G  = C0 dx G (x, Cr ). Once G  has been
0
computed, the reference position is uniquely fixed.

4 Concluding Remarks
A travelling wave solution for colloid facilitated mass transport through porous
media has been obtained. A linear, reversible kinetic relation has been assumed
to account for mass transfer from/toward colloidal particles. This leads to a
simple BV-problem, that can be solved by means of a standard finite difference
numerical method. The present study is also the fundamental prerequisite to
investigate the dispersion mechanisms of pullatants under under (more complex)
flow configurations along the lines of [14]. Some of them are already part of
ongoing research projects.

References
1. Aris, R.: On the dispersion of linear kinematic waves. Proc. Roy. Soc. Lond. Ser.
A Math. Phys. Sci. 245(1241), 268–277 (1958)
2. Bellin, A., Severino, G., Fiori, A.: On the local concentration probability den-
sity function of solutes reacting upon mixing. Water Resour. Res. 47(1), W01514
(2011). https://doi.org/10.1029/2010WR009696
3. Bolt, G.: Movement of solutes in soil: principles of adsorption/exchange chromatog-
raphy. In: Developments in Soil Science, vol. 5, pp. 285–348 (1979)
4. Campagna, R., Cuomo, S., Giannino, F., Severino, G., Toraldo, G.: A semi-
automatic numerical algorithm for turing patterns formation in a reaction-diffusion
model. IEEE Access 6, 4720–4724 (2018)
5. Cvetkovic, V.: Colloid-facilitated tracer transport by steady random ground-water
flow. Phys. Fluids 12(9), 2279–2294 (2000)
6. Severino, G., Campagna, R., Tartakovsky, D.: An analytical model for carrier-
facilitated solute transport in weakly heterogeneous porous media. Appl. Math.
Model. 44, 261–273 (2017). https://doi.org/10.1016/j.apm.2016.10.064
7. Severino, G., Comegna, A., Coppola, A., Sommella, A., Santini, A.: Stochastic
analysis of a field-scale unsaturated transport experiment. Adv. Water Resour.
33(10), 1188–1198 (2010). https://doi.org/10.1016/j.advwatres.2010.09.004
8. Severino, G., Dagan, G., van Duijn, C.: A note on transport of a pulse of nonlin-
early reactive solute in a heterogeneous formation. Comput. Geosci. 4(3), 275–286
(2000). https://doi.org/10.1023/A:1011568118126
9. Severino, G., Leveque, S., Toraldo, G.: Uncertainty quantification of unsteady
source flows in heterogeneous porous media. J. Fluid. Mech. 870, 5–26 (2019).
https://doi.org/10.1017/jfm.2019.203
108 S. Cuomo et al.

10. Severino, G., Monetti, V., Santini, A., Toraldo, G.: Unsaturated transport with
linear kinetic sorption under unsteady vertical flow. Transp. Porous Media 63(1),
147–174 (2006). https://doi.org/10.1007/s11242-005-4424-0
11. Severino, G., Tartakovsky, D., Srinivasan, G., Viswanathan, H.: Lagrangian models
of reactive transport in heterogeneous porous media with uncertain properties.
Proc. Roy. Soc. A Math. Phys. Eng. Sci. 468(2140), 1154–1174 (2012)
12. Severino, G., Cvetkovic, V., Coppola, A.: On the velocity covariance for steady
flows in heterogeneous porous formations and its application to contaminants trans-
port. Comput. Geosci. 9(4), 155–177 (2005). https://doi.org/10.1007/s10596-005-
9005-3
13. Severino, G., Cvetkovic, V., Coppola, A.: Spatial moments for colloid-enhanced
radionuclide transport in heterogeneous aquifers. Adv. Water Resour. 30(1), 101–
112 (2007). https://doi.org/10.1016/j.advwatres.2006.03.001
14. Severino, G., De Bartolo, S., Toraldo, G., Srinivasan, G., Viswanathan, H.: Travel
time approach to kinetically sorbing solute by diverging radial flows through het-
erogeneous porous formations. Water Resour. Res. 48(12) (2012). https://doi.org/
10.1029/2012WR012608
15. Severino, G., Santini, A., Monetti, V.M.: Modelling water flow water flow and
solute transport in heterogeneous unsaturated porous media. In: Pardalos, P.M.,
Papajorgji, P.J. (eds.) Advances in Modeling Agricultural Systems, vol. 25, pp.
361–383. Springer, Boston (2009). https://doi.org/10.1007/978-0-387-75181-8 17
 Performance Analysis of a Multicore
Implementation for Solving a
Two-Dimensional Inverse Anomalous
Diffusion Problem

Pasquale De Luca1(B) , Ardelio Galletti2 , Giulio Giunta2 ,

Livia Marcellino2 , and Marzie Raei3
1
Department of Computer Science, University of Salerno, Fisciano, Italy
[email protected]
2
Department of Science and Technology, University of Naples Parthenope,
Naples, Italy
{ardelio.galletti,giulio.giunta,livia.marcellino}@uniparthenope.it
3
Department of Mathematics, Malek Ashtar University of Technology, Isfahan, Iran
[email protected]

Abstract. In this work we deal with the solution of a two-dimensional

inverse time fractional diffusion equation, involving a Caputo fractional
derivative in his expression. Since we deal with a huge practical prob-
lem with a large domain, by starting from an accurate meshless local-
ized collocation method using RBFs, here we propose a fast algorithm,
implemented in a multicore architecture, which exploits suitable parallel
computational kernels. More in detail, we firstly developed, a C code
based on the numerical library LAPACK to perform the basic linear
algebra operations and to solve linear systems, then, due to the high
computational complexity and the large size of the problem, we propose
a parallel algorithm specifically designed for multicore architectures and
based on the Pthreads library. Performance analysis will show accuracy
and reliability of our parallel implementation.

Keywords: Fractional models · Multicore architecture · Parallel

algorithms

1 Introduction

In recent decades, fractional calculus has become highly attractive due to wide
applications in science and engineering. Indeed, fractional models are beneficial
and powerful mathematical tools to describe the inherent properties of processes
in mechanics, chemistry, physics, and other sciences. Meshless methods represent
a good technique for solving these models in high-dimensional and complicated
computational domains. In particular, in the current work we deal with the

c Springer Nature Switzerland AG 2020

Y. D. Sergeyev and D. E. Kvasov (Eds.): NUMTA 2019, LNCS 11973, pp. 109–121, 2020.
https://doi.org/10.1007/978-3-030-39081-5_11
110 P. De Luca et al.

solution of a two-dimensional inverse time fractional diffusion equation [1,2],

defined as follows:
c α
0 Dt v(x, t) = κΔv(x, t) + f (x, t), x = (x, y) ∈ Ω ⊆ R2 , t ∈ ]0, T ], (1)

with the following initial and Dirichlet boundary conditions:

v(x, 0) = ϕ(x), x ∈ Ω,
v(x, t) = ψ1 (x, t), x ∈ Γ1 , t ∈ ]0, T ], (2)
v(x, t) = ψ2 (x)ρ(t), x ∈ Γ2 , t ∈ ]0, T ],

and the non-local boundary condition


v(x, t)dx = h(t), t ∈ ]0, T ], (3)
Ω

∂ α
where v(x, t) and ρ(t) are unknown functions and c0 Dtα = ∂t α denotes the Caputo

fractional derivative [12,13] of order α ∈ ]0, 1]. It is obvious and well-known that
in such problems, collocation methods based on global radial basis functions
lead to ill-conditioned coefficient matrices. Also, if we deal with a huge practical
problem with a large domain, the computational cost and ill-conditioning will
grow dramatically. Therefore, the use of fast algorithms and parallel computa-
tional kernels become unavoidable and necessary. So, in this work, by starting
from an accurate meshless localized collocation method for approximating the
solution of (1) based on the local radial point interpolation (LRPI) technique,
a parallel procedure exploiting the multicore environment capabilities is pro-
posed. The parallel algorithm is based on a functional decomposition approach
in order to perform an asynchronous kind of parallelism, by solving in parallel
different tasks of the overall work, in order to obtain a meaningful gain in terms
of performance.
The rest of the paper is organized as follows: in Sect. 2 the numerical proce-
dure to discretize the inverse time fractional diffusion equation is shown; Sect. 3
deals with the description of both sequential and parallel implementation details
of the algorithm; in Sect. 4 we provide tests and experiments that prove accuracy
and efficiency, in terms of performance, of the parallel implementation; finally,
in Sect. 5, we draw conclusions.

2 Numerical Background: The Time Discretization and

Meshless Localized Collocation

In this section, the numerical approach to discretize the problem (1) is sum-
marized. Following [1,2], we firstly consider an implicit time stepping proce-
dure discretize the fractional model (1) in time direction, then we make use of
a meshless localized collocation method to evaluate the unknown functions in
some collocation points.
 Performance Analysis of a Multicore Implementation 111

2.1 The Time Discretization Approximation

In order to discretize the fractional model (1) in time direction an implicit time
stepping procedure is employed. More in particular, by choosing a time step
τ > 0 and setting tn = nτ , for n = 0, . . . , T /τ (assume that T /τ is an integer),
and by substituting t = tn+1 in the Eq. (1), the following relation is obtained:
c α n+1
0 Dt v(x, t ) = κΔv(x, tn+1 ) + f (x, tn+1 ), (x, tn+1 ) ∈ Ω × (0, T ], (4)

Therefore we make use of the following second-order time discretization for the
Caputo derivative of v(x, t) at point t = tn+1 [3,4]:
  n+1
 ω α (j) t−α
c α n+1−j
0 Dt v(x, t) = v(x, t ) − v(x, 0) + O(τ 2 ), (5)
t=tn+1
j=0
τα Γ (1 − α)

where
⎧ ⎧
⎪
⎨ α + 2 pα ⎨ 1, j = 0,
α 0, j = 0,

ω (j) = α +2 and pα
j = α+1
⎪
⎩
2 α α α
pj − pj−1 , j > 0, ⎩ 1− pα
j−1 , j ≥ 1.
2 2 j

By substituting the Eqs. (5) in (4) the following relation is obtained:

n
ω α (0) n+1 n+1 ω α (j) n+1−j t−α
α
v − κΔv = − α
v + v 0 + f n+1 , (6)
τ j=1
τ Γ (1 − α)

with f n+1 = f (x, tn+1 ) and v n+1−j = v(x, tn+1−j ) (j = 0, . . . , n + 1).

2.2 The Meshless Localized Collocation Method

In last decades, meshless methods have been employed in many different fields
of science and engineering [8–11]. This is because these methods can compute
numerical solutions without using any given mesh of the problem domain. In the
meshless localized collocation method, the global domain Ω is partitioned into
local sub-domains Ωi (i = 1, . . . , N ) corresponding to every point. These sub-
domains ordinarily are circles or squares and cover the entire global domain Ω.
Then the radial point interpolation shape functions, φi , are constructed locally
over each Ωi by combining radial basis functions and the monomial basis function
[5] corresponding to each local field point xi . In the current work, one of the most
popular RBFs, i.e., the generalized multiquadric radial basis function (GMQ-
RBF) is used as follows:
φ(r) = (r2 + c2 )q , (7)
where c is the shape parameter. We highlight that other RBF basis function
could be also used (for example Gaussian or Wendland RBFs) without changing
the description structure below. However, as shown in the experimental section,
GMQ-RBFs guarantee good results in terms of accuracy. The local radial point
112 P. De Luca et al.

interpolation shape function generates the N × N sparse matrix Φ. Therefore v

can be approximated by
N
v(x) = φi (x)vi (8)
i=1

where φi (x) = φ(x − xi 2 ) (the norm x − xi 2 denotes the Euclidean distance

between x and field point xi ). Substituting approximation formula (8) in Eqs. (6),
(2) and (3) yields:
N  α
  n N
ω (0) ∂ 2 φi ∂ 2 φi ω α (j) 
φi (xj ) −κ + (x j ) v n+1
i = − φi (xj )vin+1−j
i=1
τα ∂x2 ∂y 2 j=1
τ α
i=1

 N
t−α
+ φi (xj )v 0 + f n+1 , j = 1, . . . , NΩ (9)
Γ (1 − α) i=1

N

φi (xj )vin+1 = ψ1n+1 (xj ), j = NΩ + 1, . . . , NΩ + NΓ1 , (10)
i=1
N

φi (xj )vin+1 = ψ2n+1 (xj )ρn+1 , j = NΩ +NΓ1 +1, . . . , NΩ +NΓ1 +NΓ2 , (11)
i=1
N

 
φi (xj )dΩ vin+1 = hn+1 . (12)
i=1 Ω

The collocation equations (9) are referred to the NΩ interior points in Ω, while
the NΓ1 equations (10) and the NΓ2 equations (11) (involving also the unknown
ρn+1 = ρ(tn+1 )) arise from the initial and Dirichlet boundary conditions. Finally,
a further equation is obtained by applying 2D Gaussian-Legendre quadrature
rules of order 15. Therefore, the time discretization approximation and the local
collocation strategy construct a linear system of N + 1 linear equations with
N + 1 unknown coefficients (N = NΩ + NΓ1 + NΓ2 ). The unknown coefficients

v(n+1) = (v1n+1 , . . . , vN
n+1 n+1
,ρ )

are obtained by solving the sparse linear system:

Av(n+1) = B (n+1) , (13)

where A is a (N + 1) × (N + 1) coefficient matrix and B (n+1) is the (N + 1)

vector. In this regard it should be noted that, unlike B (n+1) , the coefficient
matrix A does not change its entries along the time steps. Moreover, due to the
local approach, where only nearby points of each field point xi are considered,
each related equation of the linear systems involves few unknown values and,
consequently the coefficient matrix A is sparse.
Previous discussion allows us to introduce the following scheme, Algorithm
1, which summarizes the main steps needed to solve the numerical problem.
 Performance Analysis of a Multicore Implementation 113

Algorithm 1. Pseudo-code for problem (1)

Input: κ, α, T, τ ,ϕ, Ψ1 , Ψ2 , h,
{xi }N Ω
i=1 , % interior points
NΓ
{xi+NΩ }i=11 % Γ1 boundary points
NΓ
{xi+NΩ +NΓ1 }i=12 % Γ2 boundary points
 T /τ −1
Output: {vin+1 }N
i=1 ,
n=0
T /τ −1
{ρn+1 }n=0

1: build A % by following (9,10,11,12)

2: for n = 0, 1, . . . , T /τ − 1 % loop on time slices
3: build B(n+1) % by following (9,10,11,12)
(n+1)
4: compute v : % solution of Av(n+1) = B(n+1) in (13)
5: endfor

3 Sequential and Parallel Implementation

To perform an effective implementation of Algorithm 1, we firstly developed a

sequential version of the code, then a multicore parallel version. More precisely,
in the following we present a detailed description of the sequential algorithm
implementation:

(ds1) generate input points, called CenterPoints, in a 2D regular grid;

(ds2) partition CenterPoints in interior points center I and boundary points
center b1 and center b2;
(ds3) find, for each fixed interior point (i = 1, . . . ,size(center I)) its local neigh-
bors and, by evaluating the Laplacian of the local RBF interpolating func-
tion, we build the i-th row of A total (i.e. A). We highlight that this step
requires to solve multiple linear systems of small size (number of neigh-
bors) for each point in center I);
(ds4) build next size(center b1) + size(center b2) rows of A total (by using
(10) and (11));
(ds5) build last row of A total by evaluating the integral in (3) by means of
2D Gaussian-Legendre quadrature rules of order 15;
(ds6) generate a discrete 1D time interval tt= [0 : τ : T ], with step τ ;
(ds7) for each time in tt, build the right-hand side vector B (i.e. B(n+1) ) and
solve the sparse linear system A total · sol = B, where sol is the com-
puted value of v(n+1) ;
(ds8) finally, the condition number of A total is also estimated.

Previous steps are carried over the following Algorithm 2, where: the multiple
linear systems, presented in (ds3), are built at lines 9, 10, 11 and then solved
114 P. De Luca et al.

by using the routine dgesv of the LAPACK library based on the LU factor-
ization method, while to solve the sparse linear system, at line 16, a specific
routine of the CSPARSE library is employed [15], i.e. the cs lusol routine,
typical for linear systems characterized by sparse coefficient matrices. Finally,
to evaluate the condition number of the sparse matrix at line 21, we used the
condition simple1 routine of the CONDITION library [14].

Algorithm 2. Sequential algorithm

1: STEP 0: input phase
2: generate CenterPoints
3: find Center I % interior points
4: find Center b1 % boundary points
5: find Center b2 % boundary points
6: STEP 1: construction of the coefficient matrix
7: for each point of CenterPoints
8: find its local neighbors
9: solve multiple linear systems % one for each point of center I
10: solve multiple linear systems % one for each point of center b1
11: solve multiple linear systems % one for each point of center b2
12: endfor
13: STEP 2: loop on time
14: for n = 0; n < T /τ ; step = 1 do
15: build B % (by using results of lines 8,9,10,11)
16: solve A total · sol = B
17: set sol M[n+1]:=sol
18: end for
19: STEP 3: output phase and condition number evaluation
20: reshape matrix sol M
21: compute cond % condition number of matrix A total

Despite the good accuracy achieved in our experiments, we observed that

executing the software on a latest-generation CPU requires very large execution
times. In order to improve the performance, a parallel approach is introduced. To
be specific, we have chosen to parallelize main kernels in Algorithm 2 by using
the well-known powerful of multicore processors, that are widely used across
many application domains, including general-purpose, embedded, network and
digital signal processing [6,7].
Our parallel algorithm is based on a functional decomposition combined with
a classical domain decomposition approach. In other words, the pool of threads
 Performance Analysis of a Multicore Implementation 115

works exploiting an asynchronous parallelism by solving different tasks of the

overall work in a parallel way. In the following, the parallel strategy implemented
is described in detail.
The input phase in STEP 0 uses a parallel definition of the local neighbors
by considering the classical domain decomposition approach of the CenterPoint
set and using the information related to the id thread, i.e. the index of each
thread. In this way, we can define a local size n loc, by dividing the global size
of the data structure by the thread number n thread. Hence, we deduce the
start and the final positions of each local structure as:

start = id thread × n loc, end = (id thread + 1) × n loc.

The local block, for each subset of point which each thread deals with, are found
and collected by a suitable synchronization barrier. In fact, since this phase uses
several system calls to copy the local chunk into the global neighbors, which are
stored in the shared memory, we manage this critical region by protecting the
shared data to a race condition, using semaphores (see libraries in [17,18]).
The STEP 1, is related to the construction of the coefficient matrix A total.
In order to parallelize this task, we firstly use the domain decomposition app-
roach in which each sub-set of the CenterPoint finds its local neighbors, as in
the STEP 0; then, for the first time an asynchronous approach to menage the
work of threads is used. Every threads works in parallel to solve the multiple
linear systems shown at lines 9, 10 and 11 of Algorithm 2, by using the dgesv
routine of the LAPACK library. After this task, still a critical region occurs.
Then, to ensure the correct data writing a similar synchronization mechanism
to the one used for the STEP 0, has been implemented. To complete this phase
a check at the barrier happens, and then a semaphore to unlock STEP 1 is
activated.
The STEPS 2 and 3, which provide the sparse linear systems solution and the
condition number computation are executed in a similar way to what happens

Algorithm 3. parallel sparse linear systems resolution and condition number

evaluation - asynchronous approach
1: shared variables: NG, A total
2: one thread execution
3: compute the condition number of the A total matrix in serial way into cond
4:
5: other threads execution
6: for n = 0; n < T /τ ; step = 1 do
7: build B
8: solve A total · sol = B
9: set sol M[n+1]:=sol
10: end for
11: check barrier for all threads
116 P. De Luca et al.

in STEP 1. This phase is shown in Algorithm 3 and it is the main computational

kernel of the parallel algorithm.
Again, a combination of domain decomposition and functional decomposition
approach is used to perform in parallel these final STEPS. In fact, while a single
thread compute the condition number of the sparse matrix, the other threads
build in parallel the B vectors, for each time step, breaking down the work by
means of a domain decomposition scheme. Moreover, they solve the sparse linear
system, at each time, by using the multicore parallel version of the cs lusol
routine of CSPARSE library.
To be specific, for the second time we exploit the power of such environ-
ment by using the possibility to operate in an asynchronous way. Well-suited
semaphores help the execution in order to manage the critical regions and to
protect the routines which are not thread-safe, i.e. those routines that cannot be
called from multiple threads without unwanted interaction between them. This
ensures, always, that the parallel software has the correct data consistency.

4 Implementation Details and Numerical Tests

Our parallel algorithm is developed and tested on two CPU Intel Xeon with 6
cores, E5-2609v3, 1.9 Ghz, 32 GB of RAM memory, 4 channels 51Gb/s memory
bandwidth. We use the Pthreads library [16], for UNIX systems, as specific inter-
face for the IEEE POSIX 1003.1c standard: a set of C language programming
types and procedures for managing the synchronization and concurrency in a
multicore environment.

4.1 Accuracy

Here we are interested in measuring the error due to the numerical approximation
introduced by using both time discretization and the local collocation approach.
Following results are obtained by using the parallel algorithm with 4 threads. It
is also useful to point out that the accuracy does not depend on the number of
threads. As measures of the approximation error we use both the relative root
mean square v and maximum absolute error ερ , defined as:

N

 ṽ(xi , tn ) − v(xi , tn ) 2
v = max  n)
, ρ = max |ρ̃(tn )−ρ(tn )|,
n=1,...,T /τ
i=1
v(x i , t n=1,...,T /τ

where ṽ and ρ̃ denote the computed values of the true solutions v and ρ, respec-
tively. All results in this section are referred to the following case study:

– as set of points we use a uniform two-dimensional grid in Ω = [0, 1]2 so that

N = (1 + h−1 )2 , being h the space discretization step along both x and y
directions;
– the boundaries are Γ2 = {(x, y)|x = 1, 0 < y < 1} and Γ1 = ∂Ω \ Γ2 ;
 Performance Analysis of a Multicore Implementation 117

– the source function is:


t2−α
f (x, t) = 2 − t exp(x + y),
2
Γ (3 − α)
while the initial and Dirichlet boundary conditions are set by taking:

ϕ(x) = 0, ψ1 (x, t) = t2 exp(x + y) and ψ2 (x) = exp(x + y)

– the true solutions of the problem are:

v(x, t) = t2 exp(x + y) and ρ(t) = t2 .

Table 1 shows the error behaviour in terms of time step size (τ ) by letting the
number of points N = 400 and two different values of the fractional order α.
This table illustrates the convergence and the accuracy of the proposed method
while decreasing τ . In Table 2 the error estimates and the condition number of
the coefficient matrix are reported by increasing the number of points N and
taking τ = 0.01 for two different values of the fractional order α. The results
show that the accuracy of the method is improved by increasing N and the con-
dition number indicates that the coefficient matrix A has an acceptable condi-
tion number. Figure 1 demonstrates the approximate solution and the point-wise
absolute error for v(x, t) by letting N = 625, τ = 0.01 and α = 0.75. Also, local
sub-domain for a center point and sparsity pattern of A are shown in Fig. 2.

Table 1. Behaviour of the condition number μ and of the approximation error for
α = 0.25, α = 0.85, for N = 400 and for several values of τ .

α τ v ρ α τ v ρ

0.25 1
10
9.45 · 10−5 9.96 · 10−5 0.85 1
10
9.29 · 10−5 9.79 · 10−5
−5 −5
1
20
5.11 · 10 5.39 · 10 1
20
4.92 · 10−5 5.19 · 10−5
1
40
3.82 · 10−5 4.02 · 10−5 1
40
3.74 · 10−5 3.91 · 10−5
−5 −5
1
80
3.61 · 10 3.66 · 10 1
80
3.55 · 10−5 3.58 · 10−5
1
160
3.56 · 10−5 3.56 · 10−5 1
160
3.51 · 10−5 3.49 · 10−5
−5 −5
1
320
3.54 · 10 3.54 · 10 1
320
3.49 · 10−5 3.47 · 10−5

4.2 Performance Analysis

This section deals with the efficiency analysis of our parallel algorithm in terms
of execution times, in order to highlight the gain achieved.
In Table 3 several executions of our software are showed, by varying the input
size and the threads number t. We can see a significant gain with respect to the
serial version, already starting from the run with 2 cores.
118 P. De Luca et al.

Table 2. Behaviour of the condition number μ and of the approximation error for
α = 0.35, α = 0.75, for τ = 0.01 and for several values of N .

α N v ρ μ α N v ρ μ

0.35 36 7.07 · 10−4 8.64 · 10−4 1.98 · 104 0.75 36 7.06 · 10−4 8.63 · 10−4 1.88 · 104
−4 −4 4
81 2.16 · 10 2.44 · 10 8.32 · 10 81 2.15 · 10−4 2.44 · 10−4 8.15 · 104
144 6.56 · 10−5 7.18 · 10−5 3.50 · 105 144 6.58 · 10−5 7.21 · 10−5 3.22 · 105
−5 −5 5
196 2.35 · 10 2.30 · 10 6.69 · 10 196 2.34 · 10−5 2.33 · 10−5 6.15 · 105
289 2.91 · 10−5 1.57 · 10−5 1.49 · 106 289 2.88 · 10−5 1.53 · 10−5 1.38 · 106
−5 −5 6
400 3.59 · 10 3.63 · 10 2.93 · 10 400 3.56 · 10−5 3.58 · 10−5 2.86 · 106
576 4.96 · 10−5 5.18 · 10−5 6.23 · 106 576 4.92 · 10−5 5.14 · 10−5 6.05 · 106

Fig. 1. (a) Exact solution v(x, t); (b) absolute error (N = 625, τ = 0.01, α = 0.75)

1 0

0.9
100
0.8

0.7 200

0.6

300
0.5
y

0.4
400
0.3

0.2 500

0.1
600
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 100 200 300 400 500 600
x nz = 37212

(a) (b)

Fig. 2. (a) Local sub-domain for a given center point; (b) sparsity pattern of A
 Performance Analysis of a Multicore Implementation 119

Table 3. Execution times in seconds (s) achieved by varying both number of threads
t (t = 1, 2, 4, 6, 9, 12) and size of the problem N .

N Serial time (s) Parallel time (s)

1 2 4 6 9 12
3 4
8.1 × 10 1.2 × 10 30.02 15.02 10.97 9.03 8.94 7.98
1.0 × 104 3.19 × 104 39.98 20.01 14.97 12.08 11.99 11.01
1.69 × 104 1.66 × 105 97.06 56.92 41.95 35.05 33.99 29.01
1.98 × 104 2.46 × 105 114.98 85.02 56.97 45.64 40.81 38.80
2.25 × 104 3.9 × 105 195.96 116.93 83.94 77.05 74.96 68.01

We observe that, the execution time of the parallel version is improved espe-
cially for the use of an “ad hoc” memory allocation and a suitable scheduling
policy. More precisely, the gain is so large, by using two cores or more, for the
massive use of shared-memory combined with a suitable employment of the local
stack memory level of each thread. The strong performance, compared to the
sequential version, is confirmed also by increasing the number of cores until six,
while, for a larger number of cores, the performance degrades. This can also be
observed through Tables 4 and 5, which show the speed-up and the efficiency,
respectively. This is due to the hardware characteristics of our supercomputer
(two CPUs, with 6 cores) and, precisely, to the fact that while dual CPU systems
setups pack many core counts and outshine single processor servers by a large
margin, our tests show a marginal performance increase over single CPU config-
urations, caused by the fact that the CPUs worked on the same data, at the same
time, in the shared memory. In other words, the synchronization, which needs
the access to the shared memory by the CPUs, slows down the performance and
decreases the earnings expectations inevitably.

Table 4. Speed-up achieved by varying both number of threads t (t = 2, 4, 6, 9, 12)

and size of the problem N .

N Parallel Speed-up
2 4 6 9 12
3
8.1 × 10 1.99 2.73 3.32 3.35 3.76
1.0 × 104 1.99 2.67 3.30 3.34 3.63
1.69 × 104 1.70 2.31 2.76 2.93 3.34
1.98 × 104 1.35 2.01 2.51 2.81 2.96
2.25 × 104 1.67 2.33 2.54 2.61 2.88
120 P. De Luca et al.

Table 5. Efficiency achieved by varying both number of threads t (t = 2, 4, 6, 9, 12)

and size of the problem N .

N Parallel efficiency
2 4 6 9 12
3
8.1 × 10 0.99 0.68 0.55 0.37 0.31
1.0 × 104 0.99 0.66 0.55 0.37 0.30
1.69 × 104 0.85 0.57 0.43 0.32 0.27
1.98 × 104 0.67 0.50 0.41 0.31 0.24
2.25 × 104 0.83 0.58 0.42 0.29 0.24

5 Conclusion and Future Work

In this paper, we proposed a parallel code in order to solve a two-dimensional

inverse time fractional diffusion equation. The algorithm implements a numerical
procedure which is based on the discretization of the Caputo fractional derivative
and on the use of a meshless localized collocation method exploiting the radial
basis functions properties. Because the high computational complexity and the
large size of the problem, the parallel algorithm is designed for multicore archi-
tectures and, as seen in the experiments, this code has proved to be very efficient
and accurate.

References
1. Yan, L., Yang, F.: The method of approximate particular solutions for the time-
fractional diffusion equation with a non-local boundary condition. Comput. Math.
Appl. 70, 254–264 (2015)
2. Abbasbandy, S., Ghehsareh, H.R., Alhuthali, M.S., Alsulami, H.H.: Comparison
of meshless local weak and strong forms based on particular solutions for a non-
classical 2-D diffusion model. Eng. Anal. Boundary Elem. 39, 121–128 (2014)
3. Tian, W.Y., Zhou, H., Deng, W.H.: A class of second order difference approxima-
tions for solving space fractional diffusion equations. Math. Comput. 84, 1703–1727
(2015)
4. Kilbas, A., Srivastava, M.H., Trujillo, J.J.: Theory and Application of Fractional
Differential Equations, North Holland Mathematics Studies, vol. 204 (2006)
5. Liu, Q., Gu, Y.T., Zhuang, P., Liu, F., Nie, Y.F.: An implicit RBF meshless app-
roach for time fractional diffusion equations. Comput. Mech. 48, 1–12 (2011)
6. Cuomo, S., De Michele, P., Galletti, A., Marcellino, L.: A GPU parallel imple-
mentation of the local principal component analysis overcomplete method for DW
image denoising. In: Proceedings - IEEE Symposium on Computers and Commu-
nications, vol. 2016, pp. 26–31, August 2016
7. D’Amore, L., Arcucci, R., Marcellino, L., Murli, A.: A parallel three-dimensional
variational data assimilation scheme. In: AIP Conference Proceedings, vol. 1389,
pp. 1829–1831 (2011)
 Performance Analysis of a Multicore Implementation 121

8. Cuomo, S., Galletti, A., Giunta, G., Marcellino, L.: Reconstruction of implicit
curves and surfaces via RBF interpolation. Appl. Numer. Math. 116, 157–171
(2017). https://doi.org/10.1016/j.apnum.2016.10.016
9. Ala, G., Ganci, S., Francomano, E.: Unconditionally stable meshless integration of
time domain Maxwell’s curl equations. Appl. Math. Comput. 255, 157–164 (2015)
10. Cuomo, S., Galletti, A., Giunta, G., Starace, A.: Surface reconstruction from scat-
tered point via RBF interpolation on GPU. In: 2013 Federated Conference on
Computer Science and Information Systems, FedCSIS 2013, art. no. 6644037, pp.
433–440 (2013)
11. Francomano, E., Paliaga, M., Galletti, A., Marcellino, L.: First experiences on an
accurate SPH method on GPUs. In: Proceedings - 13th International Conference on
Signal-Image Technology and Internet-Based Systems, SITIS 2017, 2018-January,
pp. 445–449 (2018). https://doi.org/10.1109/SITIS.2017.79
12. Ala, G., Di Paola, M., Francomano, E., Pinnola, F.: Electrical analogous in vis-
coelasticity. Commun. Nonlinear Sci. Numer. Simul. 19(7), 2513–2527 (2014)
13. Podlubny, I.: Fractional Differential Equation. Academic Press, San Diego (1999)
14. https://people.sc.fsu.edu/jburkardt/c src/condition/condition.html
15. http://faculty.cse.tamu.edu/davis/suitesparse.html
16. https://computing.llnl.gov/tutorials/pthreads/
17. https://pubs.opengroup.org/onlinepubs/7908799/xsh/semaphore.h.html
18. https://pubs.opengroup.org/onlinepubs/009695399/functions/pthread mutex
lock.html
 Adaptive RBF Interpolation for
Estimating Missing Values
in Geographical Data

Kaifeng Gao1 , Gang Mei1(B) , Salvatore Cuomo2 , Francesco Piccialli2 ,

and Nengxiong Xu1
1
China University of Geosciences (Beijing), Beijing, China
2
University of Naples Federico II, Naples, Italy
[email protected]

Abstract. The quality of datasets is a critical issue in big data mining.

More interesting things could be found for datasets with higher qual-
ity. The existence of missing values in geographical data would worsen
the quality of big datasets. To improve the data quality, the missing
values are generally needed to be estimated using various machine learn-
ing algorithms or mathematical methods such as approximations and
interpolations. In this paper, we propose an adaptive Radial Basis Func-
tion (RBF) interpolation algorithm for estimating missing values in geo-
graphical data. In the proposed method, the samples with known val-
ues are considered as the data points, while the samples with missing
values are considered as the interpolated points. For each interpolated
point, first, a local set of data points are adaptively determined. Then,
the missing value of the interpolated point is imputed via interpolating
using the RBF interpolation based on the local set of data points. More-
over, the shape factors of the RBF are also adaptively determined by
considering the distribution of the local set of data points. To evaluate
the performance of the proposed method, we compare our method with
the commonly used k -Nearest Neighbor (k NN) interpolation and Adap-
tive Inverse Distance Weighted (AIDW) interpolation, and conduct three
groups of benchmark experiments. Experimental results indicate that the
proposed method outperforms the k NN interpolation and AIDW inter-
polation in terms of accuracy, but worse than the k NN interpolation and
AIDW interpolation in terms of efficiency.

Keywords: Data mining · Data quality · Data imputation · RBF

interpolation · k NN

1 Introduction
Datasets are the key elements in big data mining, and the quality of datasets
has an important impact on the results of big data analysis. For a higher quality
dataset, some hidden rules can often be mined from it, and through these rules we
c Springer Nature Switzerland AG 2020
Y. D. Sergeyev and D. E. Kvasov (Eds.): NUMTA 2019, LNCS 11973, pp. 122–130, 2020.
https://doi.org/10.1007/978-3-030-39081-5_12
 Adaptive RBF Interpolation for Estimating Missing Values 123

can find some interesting things. At present, big data mining technology is widely
used in various fields, such as geographic analysis [10,16], financial analysis,
smart city and biotechnology. It usually needs a better dataset to support the
research, but in fact there is always noise data or missing value data in the
datasets [9,14,17]. In order to improve data quality, various machine learning
algorithms [7,15] are often required to estimate the missing value.
RBF approximation techniques combined with machine learning algorithms
such as neural networks can be used to optimize numerical algorithms. Besides,
RBF interpolation algorithm is a popular method for estimating missing values
[4–6]. In large-scale computing, the cost can be minimized by using adaptive
scheduling method [1]. RBF is a distance-based function, which is meshless and
dimensionless, thus it is inherently suitable for processing multidimensional scat-
tered data. Many scholars have done a lot of work on RBF research. Skala [13]
used CSRBF to analyze big datasets, Cuomo et al. [2,3] studied the reconstruc-
tion of implicit curves and surfaces by RBF interpolation. Kedward et al. [8] used
multiscale RBF interpolation to study mesh deformation. In RBF, the shape fac-
tor is an important factor affecting the accuracy of interpolation. Some empirical
formulas for optimum shape factor have been proposed by scholars.
In this paper, our objective is to estimate missing values in geographical
data. We proposed an adaptive RBF interpolation algorithm, which adaptively
determines the shape factor by the density of the local dataset. To evaluate
the performance of adaptive RBF interpolation algorithm in estimating miss-
ing values, we used three datasets for verification experiments, and compared
the accuracy and efficiency of adaptive RBF interpolation with that of k NN
interpolation and AIDW.
The rest of the paper is organized as follows. Section 2 mainly introduces the
implementation process of the adaptive RBF interpolation algorithm, and briefly
introduces the method to evaluate the performance of adaptive RBF interpola-
tion. Section 3 introduces the experimental materials, and presents the estimated
results of missing values, then discusses the experimental results. Section 4 draws
some conclusions.

2 Methods

In this paper, our objective is to develop an adaptive RBF interpolation algo-

rithm to estimate missing values in geospatial data, and compare the results with
that of k NN and AIDW. In this section, we firstly introduce the specific imple-
mentation process of the adaptive RBF interpolation algorithm, then briefly
introduces the method to evaluate the performance of adaptive RBF interpola-
tion.
The basic ideas behind the RBF interpolation are as follows. Constructing
a high-dimensional function f (x) , x ∈ Rn £, suppose there is a set of discrete
points xi ∈ Rn , i = 1, 2, · · · N$ with associated data values f (xi ) ∈ R, i =
1, 2, · · · N . Thus, the function f (x) can be expressed as a linear combination of
124 K. Gao et al.

RBF in the form (Eq. (1)):

N  
f (x) = aj φ x − xj 2 (1)
j=1

where N is the number of interpolation points, {aj } is the undetermined coeffi-

cient, the function φ is a type of RBF.
The kernel function selected in this paper is Multi-Quadric RBF(MQ-RBF),
which is formed as (Eq. (2)):

φ (r) = (r2 + c2 ) (2)

where r is the distance between the interpolated point and the data point, c is the
shape factor. Submit the data points (xi , yi ) into Eq. (1), then the interpolation
conditions become (Eq. (3)):
N  
yi = f (xi ) = aj φ xi − xj 2 , i = 1, 2, · · · N (3)
j=1

When using the RBF interpolation algorithm in a big dataset, it is not prac-
tical to calculate an interpolated point with all data points. Obviously, the closer
the data point is to the interpolated point, the greater the influence on the inter-
polation result and the data point far from the interpolated point to a certain
distance, its impact on the interpolated point is almost negligible. Therefore, we
calculate the distances from an interpolated point to all data points, and select
20 points with the smallest distances as a local dataset for the interpolated point.
In Eq. (2), the value of the shape factor c in MQ-RBF has a significant influ-
ence on the calculation result of interpolation. We consult the method proposed
by Lu and Wang [11,12], adaptively determining the value c of the interpo-
lated points by the density of the local dataset. The expected density Dexp is
calculated by the function (Eq. (4)):

Ndp
Dexp = (4)
(Xmax − Xmin ) (Ymax − Ymin )

where Ndp is the number of data points in the dataset, Xmax is the maximum
value of xi for the data points in the dataset, Xmin is the minimum value of xi
in dataset, Ymax is the maximum value of yi in dataset, Ymin is the minimum
value of yi in dataset.
And the local density Dloc is calculated by (Eq. (5)):

Nloc
Dloc = (5)
(xmax − xmin ) (ymax − ymin )

where Nloc is the number of data points in the local dataset, in this paper, we
set Nloc as 20. xmax is the maximum value of xi for the data points in local
dataset, xmin is the minimum value of xi in local dataset, ymax is the maximum
value of yi in local dataset, ymin is the minimum value of yi in local dataset.
 Adaptive RBF Interpolation for Estimating Missing Values 125

With both the local density and the expected density, the local density statis-
tic D can be expressed as (Eq. (6)):

Dloc
D (s0 ) = (6)
Dexp

where s0 is the location of an interpolated point. Then normalize the D (s0 )

measure to μD by a fuzzy membership function (Eq. (7)):
⎧ (7)
⎪
⎪
⎨ 0  D (s0 ) ≤ 0
μD = 0.5 − 0.5 cos π D (s0 ) 0 ≤ D (s0 ) ≤ 2
⎪
⎪ 2
⎩
1 D (s0 ) ≥ 2

Finally, determine the shape factor c by a triangular membership func-

tion. The calculation process of triangular membership functions with differ-
ent adaptive shape factors is shown in Fig. 1. For example, if μD is 0.45,
according to the proportional relationship, the value of shape factor c will be
(0.25 × c2 + 0.75 × c3 ); see Eq. (8).

⎧ (8)
⎪
⎪
⎪
c1 0.0 ≤ μD ≤ 0.1
⎪
⎪
⎪ c1
⎪
⎪ [1 − 5 (μD − 0.1)] + 5c2 (μD − 0.1) 0.1 ≤ μD ≤ 0.3
⎪
⎪
⎨ 5c3 (μD − 0.3) + c2 [1 − 5 (μD − 0.3)] 0.3 ≤ μD ≤ 0.5
c=
⎪
⎪
⎪
⎪
c3 [1 − 5 (μD − 0.5)] + 5c4 (μD − 0.5) 0.5 ≤ μD ≤ 0.7
⎪
⎪
⎪
⎪
⎪
⎪
5c5 (μD − 0.7) + c4 [1 − 5 (μD − 0.7)] 0.7 ≤ μD ≤ 0.9
⎩
c5 0.9 ≤ μD ≤ 1.0
where c1 , c2 , c3 , c4 , c5 are five levels of shape factor.

Fig. 1. Triangular membership function for different degrees of the adaptive shape
factor [11]

After determining the shape factor c, the next steps are the same as the gen-
eral RBF calculation method. The specific process of the adaptive RBF inter-
polation algorithm is illustrated in Fig. 2.
In order to evaluate the computational accuracy of the adaptive RBF inter-
polation algorithm, we use the metric, Root Mean Square Error (RMSE) to
measure the accuracy. The RMSE evaluates the error accuracy by comparing
126 K. Gao et al.

Start Determine the local data point

set by the kNN method

Input Data Calculate the density of points in

the local data point set

Calculate expected density

Adaptively determine the shape
factor c of RBF by the density of
the local data point set

Output Data
Estimate the missing values
using the determined shape
factor RBF
End

Fig. 2. The flow chart of the adaptive RBF interpolation algorithm

the deviation between the estimated value and the true value. In addition, we
record each calculation time as a basis for evaluating the efficiency of interpola-
tion calculation. Then, we compare the accuracy and efficiency of adaptive RBF
estimator with the results of k NN and AIDW estimators.

3 Results And Discussion

3.1 Experimental Materials

To evaluate the performance of the presented adaptive RBF interpolation algo-

rithm, we use three datasets to test it. The details of the experimental environ-
ment are listed in Table 1.

Table 1. Basic information of experimental environment

Specification Details
OS Windows 7. Professional
CPU Intel (R) i5-4210U
CPU Frequency 1.70 GHz
CPU RAM 8 GB
CPU Core 4

In our experiments, we use three datasets from three cities’ Digital Elevation
Model (DEM) images; see Fig. 3. The range of three DEM images are the same.
 Adaptive RBF Interpolation for Estimating Missing Values 127

Figure 3(a) is the DEM image of Beijing working area. The landform of this
area is mountainous in a small part of Northwest and plains in other areas.
Figure 3(b) is the DEM image of Chongqing city. There are several mountains
in the northeast-southwest direction, and the southeast area is a mountainous
area. Figure 3(c) is the DEM image of Longyan city, which is hilly and high in
the east and low in the west.
We randomly select 10% observed samples from each dataset as the samples
with missing values, and the rest as the samples with known values. It should be
noted that the samples with missing values have really elevation values in fact,
but for testing, we assume the elevations are missing. Basic information of the
datasets is listed in Table 2.

Table 2. Experimental data

Dataset Number of known values Number of missing values Illustration

Beijing 1,111,369 123,592 Fig. 3(a)
Chongqing 1,074,379 97,525 Fig. 3(b)
Longyan 1,040,670 119,050 Fig. 3(c)

(a) The DEM map of Beijing City, China (b) The DEM map of Chongqing City, China

(c) The DEM map of Longyan City, China

Fig. 3. The DEM maps of three cities for the experimental tests
128 K. Gao et al.

3.2 Experimental Results and Discussion

We compare the accuracy and efficiency of adaptive RBF estimator with that of
k NN and AIDW estimators.

(a) Computational accuracy

(b) Computational efficiency

Fig. 4. The comparisons of computational accuracy and efficiency

In the Fig. 4, we find that the accuracy of the adaptive RBF estimator is
the best performing, and the k NN estimator with the lowest accuracy. With
the number of known data points in the datasets decreases, the accuracy of
three estimators decreases significantly. Moreover, the computational efficiency
of adaptive RBF estimator is worse than that of k NN estimator and AIDW esti-
mator. Among the three methods, k NN has the best computational efficiency.
With the increase of data quantity, the disadvantage of the computational effi-
ciency of k NN estimator becomes more and more obvious.
 Adaptive RBF Interpolation for Estimating Missing Values 129

The data points selected from DEM images are evenly distributed, and the
shape factor c of the adaptive RBF interpolation algorithm is adapted according
to the density of the points in the local dataset. Therefore, when the missing
data is estimated in a dataset with a more uniform data point, the advantages
of the adaptive RBF interpolation algorithm may not be realized. We need to
do further research in datasets with uneven datasets.

4 Conclusions
In this paper, we have proposed an adaptive RBF interpolation algorithm for
estimating missing values in geographical data and evaluated its computational
accuracy and efficiency. We have conducted three groups of experiments. The
results show that the accuracy of the adaptive RBF interpolation performs better
than k NN interpolation and AIDW in regularly distributed datasets. Therefore,
we consider that the proposed adaptive method can effectively improve the accu-
racy of estimating missing data in geographical data by adaptively determining
the shape factor.

Acknowledgments. This research was jointly supported by the National Natural

Science Foundation of China (11602235), and the Fundamental Research Funds for
China Central Universities (2652018091, 2652018107, and 2652018109). The authors
would like to thank the editor and the reviewers for their contribution.

References
1. Barone, G.B., et al.: An approach to forecast queue time in adaptive scheduling:
how to mediate system efficiency and users satisfaction. Int. J. Parallel Program.
45(5), 1–30 (2016)
2. Cuomo, S., Galletti, A., Giunta, G., Marcellino, L.: Reconstruction of implicit
curves and surfaces via rbf interpolation. Appl. Numer. Math. 116, 60–63 (2016)
3. Cuomo, S., Gallettiy, A., Giuntay, G., Staracey, A.: Surface reconstruction from
scattered point via RBF interpolation on GPU (2013)
4. Ding, Z., Mei, G., Cuomo, S., Tian, H., Xu, N.: Accelerating multi-dimensional
interpolation using moving least-squares on the GPU. Concurr. Comput. 30(24),
e4904 (2018)
5. Ding, Z., Gang, M., Cuomo, S., Li, Y., Xu, N.: Comparison of estimating miss-
ing values in IoT time series data using different interpolation algorithms. Int. J.
Parallel Programm. 1–15 (2018)
6. Ding, Z., Gang, M., Cuomo, S., Xu, N., Hong, T.: Performance evaluation of gpu-
accelerated spatial interpolation using radial basis functions for building explicit
surfaces. Int. J. Parallel Programm. 157, 1–29 (2017)
7. Gao, D., Liu, Y., Meng, J., Jia, Y., Fan, C.: Estimating significant wave height
from SAR imagery based on an SVM regression model. Acta Oceanol. Sinica 37(3),
103–110 (2018)
8. Kedward, L., Allen, C.B., Rendall, T.C.S.: Efficient and exact mesh deformation
using multiscale rbf interpolation. J. Computat. Phys. 345, 732–751 (2017)
130 K. Gao et al.

9. Keogh, R.H., Seaman, S.R., Bartlett, J.W., Wood, A.M.: Multiple imputation
of missing data in nested case-control and case-cohort studies. Biometrics 74(4),
1438–1449 (2018)
10. Liang, Z., Na, Z., Wei, H., Feng, Z., Qiao, Q., Luo, M.: From big data to big
analysis: a perspective of geographical conditions monitoring. Int. J. Image Data
Fusion 9(3), 194–208 (2018)
11. Lu, G.Y., Wong, D.W.: An adaptive inverse-distance weighting spatial interpola-
tion technique. Comput. Geosci. 34(9), 1044–1055 (2008)
12. Mei, G., Xu, N., Xu, L.: Improving GPU-accelerated adaptive IDW interpolation
algorithm using fast KNN search. Springerplus 5(1), 1389 (2016)
13. Skala, V.: RBF interpolation with CSRBF of large data sets. Proc. Comput. Sci.
108, 2433–2437 (2017)
14. Sovilj, D., Eirola, E., Miche, Y., Bjork, K.M., Rui, N., Akusok, A., Lendasse, A.:
Extreme learning machine for missing data using multiple imputations. Neurocom-
puting 174(PA), 220–231 (2016)
15. Tang, T., Chen, S., Meng, Z., Wei, H., Luo, J.: Very large-scale data classification
based on k-means clustering and multi-kernel SVM. Soft Comput. 1, 1–9 (2018)
16. Thakuriah, P.V., Tilahun, N.Y., Zellner, M.: Big data and urban informatics: inno-
vations and challenges to urban planning and knowledge discovery. In: Thakuriah,
P.V., Tilahun, N., Zellner, M. (eds.) Seeing Cities Through Big Data. SG, pp.
11–45. Springer, Cham (2017). https://doi.org/10.1007/978-3-319-40902-3 2
17. Tomita, H., Fujisawa, H., Henmi, M.: A bias-corrected estimator in multiple impu-
tation for missing data. Stat. Med. 47(1), 1–16 (2018)
 Stochastic Mechanisms of Information
Flow in Phosphate Economy
of Escherichia coli

Ozan Kahramanoğulları1,3(B) , Cansu Uluşeker2 ,

and Martin M. Hancyzc3,4
1
Department of Mathematics,
University of Trento, Trento, Italy
[email protected]
2
Department of Chemistry, Bioscience and Environmental Engineering,
University of Stavanger, Stavanger, Norway
3
Department of Cellular, Computational and Integrative Biology,
University of Trento, Trento, Italy
4
Chemical and Biological Engineering, University of New Mexico, Albuquerque, USA

Abstract. In previous work, we have presented a computational model

and experimental results that quantify the dynamic mechanisms of auto-
regulation in E. coli in response to varying external phosphate levels. In
a cycle of deterministic ODE simulations and experimental verification,
our model predicts and explores phenotypes with various modifications
at the genetic level that can optimise inorganic phosphate intake. Here,
we extend our analysis with extensive stochastic simulations at a single-
cell level so that noise due to small numbers of certain molecules, e.g.,
genetic material, can be better observed. For the simulations, we resort
to a conservative extension of Gillespie’s stochastic simulation algorithm
that can be used to quantify the information flow in the biochemical
system. Besides the common time series analysis, we present a dynamic
visualisation of the time evolution of the model mechanisms in the form
of a video, which is of independent interest. We argue that our stochastic
analysis of information flow provides insights for designing more stable
synthetic applications that are not affected by noise.

Keywords: Synthetic biology · E. coli · Modelling · Stochasticity ·

Noise

1 Introduction
The rapidly growing field of synthetic biology, at the crossroads of molecular
biology, genetics and quantitative sciences, aims at developing living technolo-
gies by re-engineering the makeup of organisms. The applications in this field
are designed by channeling the quantitative understanding of the molecular pro-
cesses to a methodological workflow that can be compared to the use of mechan-
ics in civil engineering. The aim, in these applications, is to modify the organisms
c Springer Nature Switzerland AG 2020
Y. D. Sergeyev and D. E. Kvasov (Eds.): NUMTA 2019, LNCS 11973, pp. 131–145, 2020.
https://doi.org/10.1007/978-3-030-39081-5_13
132 O. Kahramanoğulları et al.

to enhance and benefit from their natural capacity for certain tasks. For example,
in enhanced biological phosphorus removal (EBPR) micro-organisms such as E.
coli are used to profit from their inherent regulatory mechanisms that efficiently
respond to phosphate starvation. Achieving a quantitative understanding of the
molecular mechanisms involved in such processes from signal transduction to
gene regulation has implications in biotechnology applications.
In previous work [1], we have presented a computational model of the
dynamic mechanisms in E. coli phosphate economy. Our model, based on a chem-
ical reaction representation, explores the biochemical auto-regulation machinery
that relays the information on extracellular inorganic phosphate (Pi ) concentra-
tion to the genetic components. The ordinary differential equation simulations
with our model quantify the dynamic response to varying external Pi levels of E.
coli with which it optimises the expression of the proteins that are involved in the
Pi intake. The analysis of the simulations with our model and their experimental
verification showed that our model captures the variations in phenotype result-
ing from modifications on the genetic components. This allowed us to explore a
spectrum of synthetic applications that respond to various external Pi concen-
trations with varying levels of gene expression.
Besides the deterministic processes that are faithfully captured by ordinary
differential equations, many aspects of gene expression employ stochastic pro-
cesses [2]. In particular, the randomness in transcription and translation due to
small molecule numbers of genetic material can result in significant fluctuations
in mRNA and protein numbers in individual cells. This, in return, can lead to
cell-to-cell variations in phenotype with consequences for function [3].
For the case of synthetic applications that involve modifications in the genetic
makeup of the cells, the stochasticity in the biochemical processes introduces
yet another parameter that needs to be monitored and can even be exploited
beyond the development of more reliable synthetic devices [4]. In this regard, the
stochastic effects in gene expression have been the topic of extensive research.
In particular, the fluctuations that depend on gene network structure and the
biochemical affinity of the interacting biochemical components have been inves-
tigated both experimentally and theoretically, e.g., [5,6].
Here we extend our analysis in [1] with extensive stochastic simulations with
Gillespie’s SSA algorithm [6,7]. Our simulations, at a single-cell level allow us
to monitor the noise due to small numbers of molecules. This way, we quantify
the effect of the model parameters corresponding to various synthetic promoter
designs on signal robustness in conditions of different regimes of external Pi
concentrations. For the simulations, we resort to a conservative extension of SSA
that can be used to quantify the information flow [8,9]. Our analysis reveals
the distribution of the system resources and the resulting information flow in
terms of species fluxes between system components in response to external Pi
signal at different time intervals. Based on this, we provide a quantification of
the noise in the system due to stochastic processes in different conditions. We
argue that our analysis provides insights that can guide the design of synthetic
applications, where the effect of stochasticity can be predicted and controlled.
 Stochasticity in Phosphate Economy of E. coli 133

We present a visualisation of the dynamic evolution of the system fluxes in the

form of a video, which is of independent interest. In particular, we show that
such visualisations can provide insights in the analysis of stochastic simulations
with chemical reaction networks beyond the time series they produce.

2 Phosphate Economy of E. coli

The regulatory mechanisms in E. coli that control the inorganic phosphate (Pi )
uptake involve the interplay between two complementary mechanisms. When the
external Pi concentration is above the millimolar range, Pi is transported into
the cell mainly by the low-affinity Pi transporter (Pit) system, which is constitu-
tively expressed and dependent on the proton motive force [10]. However, when
the external Pi concentration falls below the 0.2 mM range, the high-affinity
Phosphate specific transport (Pst) system is induced. This triggers the expres-
sion of an operon that includes an ABC transporter, which actively transports
Pi by ATP-consumption. The Pst system involves a positive feedback loop, and
it induces its own expression via a two-component system (TCS) consisting of
the histidine kinase PhoR and the transcription factor PhoB. Both Pit and Pst
are highly specific for Pi .
Pi intake by Pst system is a negative process, whereby a high external Pi con-
centration turns the system off; the activation is the default state. The current
evidence suggests that the TCS mechanism is turned off by the PhoU protein
that monitors the ABC transporter activity. In mechanistic terms, when there is
sufficient Pi flux, PhoU stabilises the PhoR and this prevents the TCS mecha-
nism from relaying the signal to the transcription factor PhoB. Contrarily, when
the external Pi concentration is limited, PhoU does not inhibit the TCS. As
a result of the decrease in the external Pi concentration, the concentration of
PhoR molecules that are not inhibited by PhoU increases. Thus, the auto-cross-
phosphorylation activity of PhoR dimers provides a proxy for the external Pi
concentration signal. This is because the Pst signal is relayed by auto-cross-
phosphorylation of PhoR dimers that are not inhibited by PhoU.
The Chemical Reaction Network (CRN) model in [1] is displayed in the
Appendix section. The model describes the signal transduction processes down-
stream of PhoU to the genetic components, and the feedback of the gene expres-
sion to the Pst system. Our model makes use of the interaction mechanism
between PhoU and PhoR by employing a scalar factor for the PhoR auto-cross-
phosphorylation activity: the reactions r01, r02, r03, and r04 model the signal
transduction from PhoR, where fc is this factor describing the PhoR activity
resulting from the external Pi concentration. The fc = 1.0 models the starvation
response to the external Pi concentration of 0 μM. An increase in the external
Pi concentration and the resulting inhibition of PhoR by PhoU is modelled by
a decrease in the fc. Thus, fc = 0 models a Pi concentration over 0.2 mM.
Following this process, phosphorylated PhoR activates PhoB by phospho-
transferase (r05, r06, r07, r08,r09, r10). Phosphorylated PhoB dimerises to
constitute an active transcription factor (r11, r12) and binds the promoter
134 O. Kahramanoğulları et al.

region of PhoA and PhoB genes to activate their transcription (r16, r17, r18,
r19). The factors bf and uf in reactions r16, r18, r17, and r19 model the affin-
ity of the active transcription factor to the promoter region. The default value
of 1.0 for these factors results in the control model, whereas variations in bf and
uf model synthetic promoters that can be stronger or weaker.
The histidine kinase PhoR is a bifunctional enzyme that performs two oppos-
ing tasks: on one hand, it activates the PhoB dimers as described above. On the
other hand, it dephosphorylates the phosphorylated PhoB (r13, r14, r15). The
activated promoters transcribe the mRNA molecules for the expression of PhoA,
PhoB, and PhoR (r20, r21, r22, r23, r24), which can be subject to degradation
or dilution (r25, r26, r27, r28, r29).
The control model in [1] is parameterised within the biologically feasible
range and the parameter values are narrowed down by random restart least-
squares optimisation by fitting the model dynamics to experimental data. The
deterministic simulation plots in Fig. 1 display the concentration dynamics of
the active transcription factor dimers DiPhoBpp, the active promoter pPhoAa,
and the protein PhoA, which is the yield of the system. As described above, the
external Pi concentration is modelled by the fold change fc applied to the auto-
cross-phosphorylation rate of the reactions r01 and r03 as this rate is a function
of the ABC transporter activity. These simulations show that our model captures
the mean behaviour of the system components in agreement with fluorescence
readings in experiments. The plots also show that the active transcription factor
DiPhoBpp concentration and the active promoter pPhoAa concentration are as
expected functions of the external Pi concentration. More importantly, these
signals are not affected from the changes in other protein concentrations.

3 Stochastic Analysis and Quantifying Information Flow

Stochastic simulations with Chemical Reaction Networks (CRNs) are commonly
performed by using one of the various versions of Gillespie’s stochastic simulation
Concentration (µM )

DiPhoBpp pPhoAa PhoA

Fig. 1. The time-series plots with ordinary differential equation simulations display the
response to the variations in external Pi concentration. The externalPi concentration
is given with the fold-change fc. A higher external Pi concentration is modelled with
a smaller factor and vice versa. The different fc values are color coded in the legend.
(Color figure online)
 Stochasticity in Phosphate Economy of E. coli 135

algorithm (SSA) [7]. Given an initial state as a vector of species quantities, the
algorithm constructs a trajectory of the network with respect to the underlying
continuous time Markov chain semantics. At each simulation step, the algorithm
performs a Monte Carlo procedure to sample from the probability distribution of
the possible reaction instances at that state to pick a reaction and its time. The
algorithm updates the state and continues in the same way until the end-time
is reached. The simulation terminates after logging the trajectory to a file.
Deterministic and stochastic simulations reflect the two facets of the CRNs
with respect to the underlying mass action dynamics. Because a stochastic simu-
lation trajectory represents one of the many possible “realisations” of the system,
it can capture the fluctuations in species numbers and possible extinctions that
may arise due to low species numbers. The deterministic simulations, on the
other hand, reflect the mean behaviour of the network, thus they do not cap-
ture noise or extinction events. Consequently, the stochastic simulations, at their
limit of large numbers, overlap with the deterministic differential equation sim-
ulations. The stochastic simulation plots depicted in Fig. 2 exemplify this idea
in comparison with the deterministic simulation plots in Fig. 1.
As described above, the SSA generates a stochastic simulation trajectory by
sequentially sampling a reaction instance one after another from the distribution
of available reactions. The time between two reaction instances is obtained by
sampling from an exponential distribution, which is a function of the reaction
propensities available at that state. Each reaction instance modifies the system
state. The algorithm then continues to pick a reaction instance until it reaches
the end-time. The algorithm logs the reaction instances, which provides the com-
mon time series representation of the simulations. As a result of this sequential
procedure, the timestamps of the reaction instances follow a total order.
However, when we inspect the dependencies of each reaction instance on
the available resources at that state, a different point of view arises. This is
because each reaction instance consumes reactants as resources that were pro-
duced by another reaction at some previous time point. The reaction instance
No. of molecules

DiPhoBpp pPhoAa PhoA

Fig. 2. Stochastic time series with different external Pi concentrations, where the hor-
izontal and vertical axes are the time in seconds and the number of molecules. As in
the simulations in Fig. 1, a higher external Pi concentration is given with a smaller
factor fc, color coded in the legend. The number of promoters, given by 10 plasmids,
gives rise to a greater noise in the number of active promoters pPhoAa in comparison
to those in active transcription factor DiPhoBpp and PhoA. PhoA quantifies the yield.
136 O. Kahramanoğulları et al.

produces other resources that become available for consumption at any time
later on. Thus, the product of a reaction remains available for consumption,
but it is not necessarily consumed by its immediate successor. Conesequently,
the production and consumption relationships between the reaction instances
in a simulation follow a partial order instead of a total order. The standard
SSA loses these relationships, which provide the information on the resource
dependencies between the reaction instances. However, these dependencies can
be used to quantitatively describe the causality as well as the information flow
in the system.

Fig. 3. Flux graphs obtained from a simulation with the CRN in the Appendix section.
The graph on the left displays the fluxes in the first 1000 simulated seconds, whereas
the second displays the fluxes in the following 1000 simulated seconds. The graphs are
rendered with Cytoscape [11] using yFiles. The thickness of the arrows is proportional
with the weight of the fluxes. The graphs show that, within the first 1000 s, the fast
times-scale dimerisation of the transcription factor PhoB, given by reaction 11, func-
tions as an attractor for the majority of the fluxes. Within the next 1000 s, as the
system approaches the steady state, the fluxes to reactions 13, 14 and 15 increase.
The concomitant feedback from the genetic circuit to the TCS results in an increased
activity in the reactions 03, 04, 05, 06, 07, 08, 09, and 10, performing TCS activation.
(Color figure online)

In previous work, we have introduced a conservative extension of Gillespie’s

stochastic simulation algorithm [7], called fSSA, that keeps track of the resource
dependencies between reaction instances [8,9]. In addition to the time series,
fSSA logs the dependency information as a partial order by introducing a con-
stant cost to each simulation step. The algorithm introduces two data structures,
one for logging the source reaction of the available resources at the current state
 Stochasticity in Phosphate Economy of E. coli 137

as a look-up table and another for logging the quantity of the simulated depen-
dencies between reaction instances. The latter structure results in a dag, called
flux graph, that provides a quantification of the flow of resources between reac-
tion instances for any user-specified time interval. The flux graph is updated at
every simulation step by sampling from the table of available resources.
The flux graphs are edge-coloured directed graphs that consist of the flux
edges from a time point t ≥ 0 to another t > t. Each edge of the graph is of the
x,n
form p −→ q, where p and q are nodes representing the reactions of the CRN,
x is a network species, and n is the weight. The edge colour x, n on the arrow
denotes that between time points t and t , species x flowed from p to q with
a weight of n. The weight n denotes the multiplicity of the species x that are
logged to have flowed from p to q within the chosen time interval.
The flux graphs in Fig. 3 are obtained from a simulation with the CRN in
the Appendix section with an fc value of 1.0. The two graphs display the fluxes
within the first 1000 simulated seconds and within the following 1000 s.
Inc. Pi
Dec. Pi

DiPhoBpp pPhoAa PhoA

Fig. 4. Stochastic time series of simulations. At 7000 simulated seconds, the phospho-
rylation fold change fc is decreased from 1.0 to 0.3 (top row) or increased from 0.3
to 1.0 (bottom row) in relation to a change in Pi concentration. For comparison, the
deterministic trajectories are plotted with dashed lines. The stochastic simulations are
scaled down to one tenth of the E. coli volume such that there is a single promoter on a
plasmid per simulation, and the binding and unbinding effects on the promoter become
observable in the plots of the active promoter pPhoAa. An increase in the unbinding
events results in the fully painted area in the pPhoAa plot. A decrease introduces gaps
to the painted area. The right-most column displays the adjustment of the system’s
yield, given by PhoA, in response to the change in external Pi levels.
138 O. Kahramanoğulları et al.

4 Coherence in Response to Phosphate Concentration

In E. coli, the TCS mechanism relays the information on external Pi concen-
tration to the genetic level. Thus, the activity level of the transcription factor
DiPhoBpp provides an internal representation of the external Pi concentration. In
agreement with experimental observations, the simulations with the differential
equations in [1] show that the mean behaviour of our model of the TCS mech-
anism in response to external Pi concentration remains robust to perturbations
in many of the system parameters. As shown in the simulations in Figs. 1 and
2, the system maintains a certain steady state in accordance with the external
Pi levels also with the feedback provided by the expression of the TCS compo-
nents that are regulated by this transcription factor: the increased activity of
the transcription factor results in the expression of the transcription factor itself
as well as the histidine kinase PhoR. Still, the steady state level of PhoB dimers
remains in equilibrium as a function of only the input signal. This phenomenon
is a consequence of the bi-functional role of PhoR, which participates in both
phosphorylation and dephosphorylation of its cognate response regulator PhoB
[12–14]. This dual role of PhoR is a mechanism that enhances signal robustness.
In experiments, it has been shown that the phosphatase activity in the TCS
provides a rapid dephosphorylation mechanism that tunes the system when it
becomes subject to changes, and thereby restores it to the original state [15].
To verify this with our model, we have performed simulations, whereby the sys-
tem is exposed to a change in external Pi after the active transcription factor
DiPhoBpp has reached deterministic equilibrium: the model is first instantiated
with an auto-cross-phosphorylation value (fc). In two sets of stochastic simu-
lations, the fc value is then decreased or increased at 7000 simulated seconds,
corresponding to a sudden change in the external Pi concentration. The stochas-
tic simulation results, depicted in Fig. 4, show that after these perturbations the
system tunes its transcription factor levels accordingly. As a result of this, the
promoter activity and the yield of the system (PhoA) adjusts to the modified
activity.
Figure 5 displays, at 100 s intervals, the system fluxes just before the pertur-
bation, i.e., F[6900, 7000], just after the perturbation, i.e., F[7000, 7100], and
the equilibrium interval F[16100, 16200]. Here, F[t, t ] denotes the flux graph
between time t and t . The arrows are scaled proportional with the flux weights.
The left-most flux graph in Fig. 5 displays the flow of the system resources
within the 100-s intervals from 6900 to 7000 s. This is the time interval immedi-
ately before the perturbation in the top row of Fig. 4, where the system is initi-
ated at the starvation condition. In this interval, just before the increase in the
external Pi concentration, the transcription factor activity, given by DiPhoBpp,
is at steady state. The flux graph shows that the system is performing cycles of
phosphorylation and dephosphorylation of PhoB as well as its complexation to
form the active dimer DiPhoBpp and the decomplexation of the dimer back to
the monomer PhoBp. The balance between these events maintains the steady
state levels of the active transcription factor DiPhoBpp.
 Stochasticity in Phosphate Economy of E. coli 139

The middle flux graph in Fig. 5 displays the fluxes within the 100 s intervals
from 7000 to 7100 s. This is the time interval immediately after the perturbation
in the top row of Fig. 4, where the external Pi concentration suddenly increases.
The flux graph indicates a shift of resources to feed flux a cycle between the
reactions 2, 4, 6, 8, and 9. As a consequence of this increase, the transcription
factor activity shifts to a new lower steady state. This, in return, accommodates
the reduction in phosphorylation of PhoB and the consequent reduction of the
transcription factor activity.
The right-most flux graph in Fig. 5 displays the fluxes from 16100 to 16200 s,
which is the period at the end of the simulation in the top row of Fig. 4. Here, the
system has already adapted to the increase in the external Pi concentration and
the transcription factor activity, given by DiPhoBpp, has reached a new steady
state. Thus, the situation in the flux graph is similar to the one in the left-
most flux graph, whereby the activation and inactivation events are balanced.
However, we observe a reduction in the fluxes to the dimerisation reaction 11,
which explains the reduced transcription factor activity.1

5 Stochasticity in Promoter Design

As demonstrated by the simulations above, the TCS transcription factor activity,

that is, the concentration of the phosphorylated PhoB dimers, serves as a proxy

[6900, 7000] [7000, 7100] [16100, 16200]

Fig. 5. Stochastic simulation fluxes with a chemical reaction network that models the
two-component system response in E. coli to a change in external phosphate concen-
tration. The graphs display the fluxes before and after the perturbation and at the end
of the simulation. The notation F [t, t ] denotes the flux graph between the time points
t and t . The numbers are the reactions of the CRN in the Appendix section. For visual
clarity, flux species are omitted. For a one-minute-long video of the complete simula-
tion fluxes, see: https://youtu.be/PiKRCYyR57k. The network simulates the first 4,5
h after the starvation signal. At 7000 s the phosphate concentration increases and the
network responds by lowering DiPhoBpp activity as in Fig. 4, top row.

1
For an exposure to the changes in system fluxes throughout the simulation, we refer
to the online video of the complete simulation: https://youtu.be/PiKRCYyR57k.
140 O. Kahramanoğulları et al.

for the external Pi concentration, given by the fc value. The resulting active
transcription factor signal activates the promoter and this feeds back as the
expression of the TCS components as well as other proteins, e.g., PhoA. This
process thus provides the specific adaptation of gene expression dependent on
the external Pi response stimuli by providing the appropriate promoter activity.
In this setting, the promoter activity, pPhoAa and pPhoBa, is proportional to the
affinity of the promoter to the active transcription factor DiPhoBpp as well as
its concentration, as described by mass action law.
The binding rate of the active transcription factor to the promoter is deter-
mined by the specific nucleotide sequence of the promoter, which also determines
how long the promoter remains bound, thus activated, after binding. A muta-
tion in a single nucleotide can result in a drastic modification of the binding and
unbinding rates [16–18]. Many applications in synthetic biology are based on
exploiting such mechanisms by introducing random mutations to the promoter
sequence and, this way, generating libraries of promoters with desired strengths.
In [1], to explore the effect of variations in promoter strength on protein
expression, we have performed a class of deterministic simulations. In these sim-
ulations, we have measured the PhoA protein yield of the system in conditions
of different external Pi concentrations. For each external Pi concentration, we
have scanned 100 different promoter designs by varying the promoter binding
factors, given by bf in the reactions r16 and r18, and the promoter unbinding
rates, given by uf in the reactions r17 and r19, in a spectrum of 10 different
values for each. A representative heat-map for these simulations that displays

fc = 1.0 fc = 0.5 fc = 0.2 fc = 0.1

uf in r17 and r19
pPhoAa (µM )

bf in r16 and r18

Fig. 6. Heatmaps for the activity of various promoter designs as in [1]. The heatmaps
are ordered according to the external Pi concentration given by the fold changes fc
applied to the PhoR autphosphorylation reactions. The left most column with 1.0 as
the fc value is the starvation condition with 0µM external Pi . Each heatmap scans
100 simulations by applying 10 different fold change values to the promoter binding
rates, given with bf in r16 and r18, as well as 10 different fold change values to the
promoter unbinding rates, given with uf in r17 and r18. The heatmaps display the
resulting steady state levels of the active promoter pPhoAa in deterministic ordinary
differential equation simulations. The intersection of the dashed lines in the left column
delivers the experimentally observed regime reported in [1]. The levels of this regime
that display the starvation response are highlighted in all the heatmaps.
 Stochasticity in Phosphate Economy of E. coli 141

the mean promoter activity pPhoAa as in [1] is depicted in Fig. 6. These simula-
tions show that in order to obtain the starvation response in the conditions with
higher external Pi concentration, promoter binding rates need to be increased
and unbinding decreased via the appropriate nucleotide sequence.
Cells with the same genetic makeup can exhibit phenotypic variation in the
expression of their different proteins. Some of this variation is attributed to
noise that is extrinsic to the protein expression machinery, characterised as the
fluctuations in other cellular components. On the other hand, the biochemical
process of gene expression can be a source of significant intrinsic noise that
results in loss of coherence in the output signal, especially in the context of low
molecule numbers [2,19]. The differential equation simulations capture the mean
deterministic behaviour that would emerge within a population that employs
such mechanisms. However, they do not capture the extent of fluctuations in
individuals and the possible variation within the population.
To detect the intrinsic noise in gene expression in the system described by our
model, we have run a set of repeated stochastic simulations under three external
Pi concentration conditions, modelled as fc values 1.0, 0.3, and 0.1. We have
then varied the binding and unbinding rates of the transcription factor and the
promoter by applying the factors bf ∈ {0.5, 1.0, 1.5} for the reactions r16, r18
and uf ∈ {0.5, 1.0, 1.5} for the reactions r17, r19. The time series of 27 sets of
simulations for 5 repeats each are depicted in Fig. 7.
In accordance with [1], in these simulations we observe that a concomitant
increase in binding rates and decrease in unbinding rates provide higher mean
levels of active promoter, given with pPhoAa. However, a fine-tuned balance of
these rates is required for the system in order not to overshoot the mean gene
expression levels in lower external Pi concentration conditions, given with fc
values closer to 1.0. From a biological point of view, such an overshoot can have
implications on function and introduce a selective pressure.
Stochastic simulations with our model demonstrate an appreciable increase
in fluctuations with an increase in unbinding rates (uf) and a moderate decrease
in the fluctuations with an increase in binding rates (bf). For a quantification of
noise in the system, we have computed the coefficient of variation (CV) for the
active promoter pPhoAa and mRNAa. We observed that the CV value for pPhoAa
increased with a decrease in fc. However, within all the external Pi concentration
regimes, the noise given with CV value for pPhoAa increased together with an
increase in the unbinding factor uf. Similarly, an increase in the binding factor
bf consistently reduced the noise in promoter activity in all the regimes. The
highest fidelity in the promoter activity signal was obtained with bf = 1.5 and
uf = 0.5. For the mRNAa signal, however, a significant consistent change in CV
value as a result of a change in unbinding factor uf is observable only with
fc values 0.3 and 0.1, corresponding to higher Pi concentrations. Similarly, an
increase in binding factor bf resulted in a decrease in noise in terms of CV for
the mRNAa signal only at higher Pi concentrations.
The CV value provides a quantification of noise for individual species. How-
ever, they may not be representative of the noise in the complete machinery. For
142 O. Kahramanoğulları et al.

a more general quantification of noise in the complete system, we have measured

the distance between the complete simulation fluxes of the simulations within
the same regime. For this, we have computed the mean distance between the
normalised flux graphs of the simulations within the same regime. We have then
compared these values in relation to variations in the external Pi concentration
as well as variations in binding and unbinding rates. We have computed the dis-
tance between the two flux graphs F and F  as the sum of squared differences

fc = 1.0 fc = 0.3 fc = 0.1

bf | uf 0.5 1.0 1.5 0.5 1.0 1.5 0.5 1.0 1.5
0.5
pPhoAa
1.0
1.5
0.5
mRNAa
1.0
1.5
0.5
PhoA
1.0
1.5

Fig. 7. The time series of stochastic simulations with promoter designs, given by the
factors bf and uf. The top panel displays the promoter activity, i.e., pPhoAa, the mid
panel displays the levels of PhoA mRNA, i.e., mRNAa, and bottom panel displays the
yield of the system as PhoA expression, i.e., PhoA. The plots are ordered according to
the external Pi concentration given by the fold changes fc applied to the PhoR aut-
phosphorylation rate. The three left most columns with fc value of 1.0 is the starvation
condition with 0 µM external Pi concentrations, followed by those for increasing levels
with fc values 0.3 and 0.1, respectively. For each external Pi concentration, the bf and
uf are scanned for values 0.5, 1.0, and 1.5 for the promoter binding rates r16 and r18
and the unbinding rates r17 and r18. The instantiation of the factors as fc = 1.0,
bf = 1.0, and uf = 1.0 delivers the control regime.
 Stochasticity in Phosphate Economy of E. coli 143

of the edge weights [20]. If an edge does not exist, its weight is assigned a value
of 0. The distance function δ, which is a metric [20], is thus defined as follows:
 x,ni
  ni , p −→ q ∈ Fi 
δ(F1 , F2 ) = sqrt (w1 −w2 ) for i = 1, 2, wi =
2

x
0, otherwise
p−→q ∈ F1 ∪F2

The mean distance δ of the simulations are displayed in the lower panel
of Fig. 7 together with the time series of PhoA. The distances follow a trend
that is similar to CV of mRNAa signal: a consistent change in the mean distance
δ between two simulations within the same regime as a result of a change in
unbinding factor uf and binding factor bf is observable only with fc values 0.3
and 0.1. Again, the highest fidelity in terms of noise is obtained with bf = 1.5 and
uf = 0.5. However, in accordance with the observations in Fig. 6, a more optimal
design that prevents an overshoot in promoter activity is obtained with uf > 0.5.
This indicates that there is a trade-off between the design of promoters that are
capable of exhibiting a starvation response in higher external Pi concentration
conditions and the noise in the synthetic system.

6 Discussion

We have presented an analysis of the noise in the biochemical machinery in the

phosphate economy of E. coli based on stochastic simulations with the SSA
algorithm. Besides the common time evolution of the biochemical species quan-
tities, our simulations capture the mechanisms with which information flows
between different system components, thereby permitting for a quantification of
the fluctuations, not only in the quantities of individual species, but also for the
complete system. Because the bacterial two-component system in our model is a
central component of many bacterial auto-regulatory processes, our model and
its deterministic and stochastic analysis should serve as a template for studying
other similar systems involved in the uptake of extracellular molecules. This, in
return, should pave the way for the systematic development of a class of syn-
thetic biology technologies that exploit such bacterial processes. In this regard,
the visualisation of the time evolution of the system fluxes that we have pre-
sented should be relevant for such applications as well as for other models that
use a chemical reactions network representation.

Acknowledgements. This work has been partially funded by the European Union’s
Horizon 2020 research and innovation programme under the grant agreement No 686585
– LIAR, Living Architecture.

Appendix

The CRN in [1] that models the auto-regulation mechanism of E. coli in response
to varying external phosphate concentrations. The time unit of the reactions is in
144 O. Kahramanoğulları et al.

seconds. The fold-change fc factor in reactions r01 and r03 model the variations
in external Pi concentration. The fc = 1.0 value corresponds to the starvation
condition and a lower fc value corresponds to a higher external Pi concentration.
The binding factor bf in reactions r16, r18 and unbinding factor uf in reactions
r17, r19 are scalar factors. They represent the affinity of the active transcription
factor to the promoter region. In the control model, the default values of bf = 1.0
and uf = 1.0 are used.

reactions
r01 : DiPhoR -> DiPhoRp , 25.3658*fc;
r02 : DiPhoRp -> DiPhoR , 8.1165;
r03 : DiPhoRp -> DiPhoRpp , 25.3658*fc;
r04 : DiPhoRpp -> DiPhoRp , 8.1165;
r05 : DiPhoRpp + PhoB -> DiPhoRpp_PhoB , 100;
r06 : DiPhoRpp_PhoB -> DiPhoRpp + PhoB , 44.9411;
r07 : DiPhoRpp_PhoB -> DiPhoRp + PhoBp , 21.3718;
r08 : DiPhoRp + PhoB -> DiPhoRp_PhoB , 100;
r09 : DiPhoRp_PhoB -> DiPhoRp + PhoB , 94.9411;
r10 : DiPhoRp_PhoB -> DiPhoR + PhoBp , 21.3718;
r11 : PhoBp + PhoBp -> DiPhoBpp , 100;
r12 : DiPhoBpp -> PhoBp + PhoBp , 24.9411;
r13 : DiPhoR + PhoBp -> DiPhoR_PhoBp , 100;
r14 : DiPhoR_PhoBp -> DiPhoR + PhoBp , 34.9411;
r15 : DiPhoR_PhoBp -> DiPhoR + PhoB , 12.95;
r16 : DiPhoBpp + pPhoA -> pPhoAa , 10000*bf;
r17 : pPhoAa -> DiPhoBpp + pPhoA , 1000*uf;
r18 : DiPhoBpp + pPhoB -> pPhoBa , 10000*bf;
r19 : pPhoBa -> DiPhoBpp + pPhoB , 1000*uf;
r20 : pPhoAa -> pPhoAa + mRNAa , 0.0540;
r21 : mRNAa -> mRNAa + PhoA , 0.0302;
r22 : pPhoBa -> pPhoBa + mRNAb , 0.130;
r23 : mRNAb -> mRNAb + PhoB , 0.036;
r24 : mRNAb -> mRNAb + DiPhoR , 0.0302;
r25 : PhoA -> , 0.0001;
r26 : PhoB -> , 0.0001;
r27 : DiPhoR -> , 0.0001;
r28 : mRNAa -> , 0.0055;
r29 : mRNAb -> , 0.0055;

initial state
0.22 DiPhoR; 0.22 PhoB; 0.0166 pPhoA; 0.0166 pPhoB;
 Stochasticity in Phosphate Economy of E. coli 145

References
1. Uluşeker, C., Torres-Bacete, J., Garcia, J.L., Hancyzc, M.M., Nogales, J.,
Kahramanoğulları, O.: Quantifying dynamic mechanisms of auto-regulation in
Escherichia coli with synthetic promoter in response to varying external phos-
phate levels. Sci. Rep. 9, 2076 (2019)
2. Elowitz, M.B., Levine, A.J., Siggia, E.D., Swain, P.S.: Stochastic gene expression
in a single cell. Science 297(5584), 1183–1186 (2002)
3. Raj, A., van Oudenaarden, A.: Nature, nurture, or chance: stochastic gene expres-
sion and its consequences. Cell 135(2), 216–226 (2008)
4. Bandiera, L., Furini, S., Giordano, E.: Phenotypic variability in synthetic biology
applications: dealing with noise in microbial gene expression. Front Microbiol. 7,
479 (2016)
5. Yu, J., Xiao, J., Ren, X., Lao, K., Xie, X.S.: Probing gene expression in live cells,
one protein molecule at a time. Science 311(5767), 1600–1603 (2006)
6. McAdams, H.H., Arkin, A.: Stochastic mechanisms in gene expression. PNAS
94(3), 814–819 (1997)
7. Gillespie, D.T.: Exact stochastic simulation of coupled chemical reactions. J. Phys.
Chem. 81(25), 2340–2361 (1977)
8. Kahramanoğulları, O., Lynch, J.: Stochastic flux analysis of chemical reaction net-
works. BMC Syst. Biol. 7, 133 (2013)
9. Kahramanoğulları, O.: Quantifying information flow in chemical reaction net-
works. In: Figueiredo, D., Martı́n-Vide, C., Pratas, D., Vega-Rodrı́guez, M.A. (eds.)
AlCoB 2017. LNCS, vol. 10252, pp. 155–166. Springer, Cham (2017). https://doi.
org/10.1007/978-3-319-58163-7 11
10. Harris, R.M., Webb, D.C., Howitt, S.M., Cox, G.B.: Characterization of PitA and
PitB from Escherichia coli. J Bacteriol. 183, 5008–5014 (2001)
11. Shannon, P., et al.: Cytoscape: a software environment for integrated models of
biomolecular interaction networks. Genome Res. 13(11), 2498–504 (2003)
12. Miyashiro, T., Goulian, M.: High stimulus unmasks positive feedback in an autoreg-
ulated bacterial signaling circuit. PNAS 105, 17457–17462 (2008)
13. Tiwari, A., et al.: Bistable responses in bacterial genetic networks: designs and
dynamical consequences. Math. Biosci. 231(1), 76–89 (2011)
14. Shinar, G., Milo, R., Matinez, M.R., Alon, U.: Input output robustness in simple
bacterial signaling systems. PNAS 104(50), 19931–19935 (2007)
15. Mukhopadhyay, A., Gao, R., Lynn, D.G.: Integrating input from multiple sig-
nals: the VirA/VirG two-component system of Agrobacterium tumefaciens. Chem-
biochem 5, 1535–1542 (2004)
16. Hawley, D.K., McClure, W.R.: Compilation and analysis of Escherichia coli pro-
moter DNA sequences. Nucleic Acids Res. 11(8), 2237–2255 (1983)
17. Jensen, P.R., Hammer, K.: The sequence of spacers between the consensus
sequences modulates the strength of prokaryotic promoters. Appl. Environ. Micro-
biol. 64(1), 82–87 (1998)
18. Jensen, P.R., Hammer, K.: Artificial promoters for metabolic optimization.
Biotechnology Bioeng. 58(2–3), 191–195 (1998)
19. Leveau, J.H., Lindow, S.E.: Predictive and interpretive simulation of green fluo-
rescent protein expression in reporter bacteria. J. Bacteriol 183(23), 6752–6762
(2001)
20. Kahramanoğulları, O.: On quantitative comparison of chemical reaction network
models. In: Proceedings of PERR 2019 3rd Workshop on Program Equivalence and
Relational Reasoning, volume 296 of EPTC, pp. 14–27 (2019)
 NMR Data Analysis of Water Mobility
in Wheat Flour Dough: A Computational
Approach

Annalisa Romano1,2 , Rosanna Campagna1(B) , Paolo Masi1,2 ,

and Gerardo Toraldo2,3
1
Department of Agricultural Sciences, University of Naples Federico II,
Portici, Naples, Italy
{anromano,rosanna.campagna,paolo.masi}@unina.it
2
Centre for Food Innovation and Development in the Food Industry (CAISIAL),
University of Naples Federico II, Portici, Naples, Italy
[email protected]
3
Department of Mathematics and Applications “R. Caccioppoli”,
University of Naples Federico II, Naples, Italy

Abstract. The understanding of the breadmaking process requires to

understand the changes in water mobility of dough. The dough ingre-
dients as well as the processing conditions determine the structure of
baked products which in turn is responsible for their apparence, tex-
ture, taste and stability. The transition from wheat flour to dough is a
complex process in which several transformations take place, including
those associated with changes in water distribution [13]. The molecular
mobility of water in foods can be studied with proton Nuclear Magnetic
Resonance (1H NMR). In this study, the measured transverse relaxation
times (T2) were considered to investigate the wheat dough development
during mixing. The interactions of the flour polymers with water during
mixing reduce water mobility and result in different molecular mobil-
ities in dough. The molecular dynamics in heterogeneous systems are
very complex. From a mathematical point of view the NMR relaxation
decay is generally modelled by the linear superposition of a few exponen-
tial functions of the relaxation times. This could be a too rough model
and the classical fitting approaches could fail to describe physical reality.
A more appealing procedure consists in describing the NMR relaxation
decay in integral form by the Laplace transform [2]. In this work a dis-
crete Inverse Laplace Transform procedure is considered to obtain the
relaxation times distribution of a dataset provided as case study.

Keywords: Bread making · Dough · Laplace transform

1 Introduction
The water molecular mobility in flour dough system is of paramount importance
because it has direct influence on the rheological properties of the dough and
c Springer Nature Switzerland AG 2020
Y. D. Sergeyev and D. E. Kvasov (Eds.): NUMTA 2019, LNCS 11973, pp. 146–157, 2020.
https://doi.org/10.1007/978-3-030-39081-5_14
 NMR Data Analysis of Water Mobility in Wheat Flour Dough 147

its baking performance. During mixing, the ingredients are transformed into a
dough through the formation of gluten, a continuous cohesive viscoelastic pro-
tein network in which are dispersed starch granules. At the dough stage, flour
particles are hydrated, and the transition from wheat flour to dough is a complex
process in which several transformations take place, including those associated
with changes in water distribution [13,21]. Generally these transformations take
place during the first minutes of mixing and can be reasonably monitored with
a Brabender farinograph that also allows an empirical characterization of the
flour on the basis of few dough mixing parameters, including water absorption
capacity, dough development time, dough stability time and mixing tolerance
index [19,22]. Although very practical for routine and industrial applications,
this approach does not provide any direct information on the physical state
of water imbibed by the flour at the molecular and supra-molecular level. The
Nuclear Magnetic Resonance (NMR) can provide information about phenomena
that involve single molecules or relatively small clusters of molecules in dough
and bread. The NMR relaxation times highlight the material composition and
are used for quality control. In particular, the molecular mobility of water and
biopolymers in food products can be studied with proton nuclear magnetic res-
onance (1H NMR). Low-resolution (LR) 1H NMR has been mainly used to
measure transverse relaxation time T2 in dough and bread (see [1,3,12]). When
water is bound tightly to the substrate (e.g. flour), it is highly immobilized
and shows reduced T2; whereas free water is mobile and has relatively long T2
[14]. Thus, useful information on the strength or degree of water binding can be
obtained. In this study, the relaxation time T2, measured by LR 1H NMR, was
used to investigate wheat dough development during mixing.
A NMR-based analysis of water mobility in wheat flour dough is carried out
and a Laplace Transform Inversion (LTI) recovery test is presented. The LTI
functional formulation is indeed useful for the analysis of the NMR relaxation
times, resulting in a relaxation time spectra which may provide information on
the different water molecules population involved in a process. The drawback of
the direct application of LTI numerical methods is the needed of a continuous
model describing the data [11]. In order to overcome this issue, the behaviour
of the true signal generating the data, is deduced by piecewise functions, fit-
ting sequences of discrete data, also of long duration, that cannot be fitted by a
single function, such as a polynomial. Spline functions are preferred, since they
do not require equally time intervals and therefore may be used to fit gaps in
data files. In literature smoothing spline models reflecting the exponential decay
taking into account main Laplace Transform (LT) properties can be found in
[4]. This approach allows to overcome the limits of the analysis of discrete data
[20], since the problem is converted in a continuous form so that general software
packages for LTI (see [9,10]) can be used. However, in the applications, the most
used approach in relaxation studies is a Regularized Inverse Laplace Transform
(RILT) algorithm based on weighted least squares solution (see [18]), that allows
to solve discrete inverse problems described by integral equations of first species.
The integral form of the decaying magnetization signal gives the LT functional
148 A. Romano et al.

relationship between the relaxation signal s and the corresponding distribution

function g of the relaxation times. A LTI of s reveals the frequency distribution
of relaxation times; in this relaxation time spectrum, the integral of each peak
reveals the contribution of each species exhibiting the corresponding specific spin
relaxation.
Regarding materials and methods, the sample data were prepared by using:
soft wheat flour (Caputo “00”, Italy), deionised water, salt and compressed
yeast (Mastro Fornaio, Paneangeli).R Dough formulation was: soft wheat flour
(60.35%), salt (1.2%), yeast (1.1%), deionised water (37.35%) according to
farinographic absorption. In fact, the amount of water (absorption) required
to obtain a stability period at 500 Brabender Units (BU) was used. All doughs
were prepared in a Brabender Farinograph (O. H. Duisburg, Germany), equipped
with a 300 g bowl. Mixing time and temperature were kept constant and equal
to 1, 2, 3 and 13 min and 25 ◦ C, respectively. Proton transverse (T2) relaxation
times were measured with a low resolution (20 MHz) 1H NMR spectrometer (the
minispec mq-20, Bruker, Milano, Italy) operating at 25±1 ◦ C. 3 g of sample were
placed into a 10 mm external diameter NMR tube (filled with sample to about
3 cm height). NMR tubes were immediately sealed with Parafilm R to prevent
moisture loss during the NMR experiment. The transverse relaxation times, 1H
T2, were determined with a CPMG pulse sequence (Carr, Purcell, Meiboom
and Gill) using a recycle delay of 0.4 s and an interpulse spacing of 0.04 ms.
The number of data points acquired was 1500 for the CPMG sequence and 20
for the Inversion Recovery. Three NMR tubes were analyzed for each sample of
dough. Each average value represents the mean of 3 independent measurements.
In Sect. 2 the mathematical modelling and the numerical scheme are presented;
Sect. 3 gives the results for a dataset of real samples. In the last section there
are some conclusions.

2 Mathematical Modelling

The numerical scheme that we use to analyse the effects of the NMR spec-
troscopy consists in: (a) a continuous description of the magnetization decay, as
a function of the time, like the sum of exponential decay functions; (b) a LTI of
the relaxation function, resulting in a relaxation time distribution, characteristic
for the samples under investigation. For sake of completness some definitions are
reported.

Definition 1 (Laplace Transform). Let f be a given integrable function and

s0 ∈ C such that  ∞
e−s0 t f (t)dt < ∞.
0

Then, for s ∈ C : Re(s) > Re(s0 ) we have

 ∞
e−st f (t)dt < ∞.
0
 NMR Data Analysis of Water Mobility in Wheat Flour Dough 149

Defining   
∞
−st
C(f ) := s ∈ C, e f (t)dt < ∞
0
(the so called region of convergence) the complex valued function F :
 ∞
F : s ∈ C(f ) → F (s) = e−st f (t)dt ∈ C
0

is the Laplace Transform of f , generally referred to as F (s) = L[f (t)].

Definition 2 (Laplace Transform inversion problem). Given F , compute
f such that F (s) = L[f (t)].
Definition 3 (Discrete data inversion). Given
Fi = F (si ) + εi , i = 1, ..., N
with εi , i = 1, ..., N unknown noises and a, b ∈ R. Compute f , or some values
of f , where
 b
F (s) = K(s, t)f (t)dt,
a
with K suitable kernel.
The real LTI is an ill-posed problem, according to Hadamard definition, since
the solution doesn’t depend continuously on the data. Due to the ill conditioning
of the discrete problem, strong amplifications of the errors occur on the final
result, thus regularization techniques are needed. When K(s, t) = e−st , the
Definition 3 is referred to as Discrete Laplace transform inversion (DLTI).
To solve the NMR data analysis by DLTI, first of all the magnetization is
modeled as a decreasing function of time, formulated as the sum of exponential
terms, depending on the amplitudes gj and relaxation times, Tj of the compo-
nents, as follows:
M

s(t) = gj e−t/Tj (1)
j=1

where: t is the experimental time, M is the number of micro-domains having

the same spin density, gj are the amplitudes and the relaxation time Tj of the
different components; Tj stands for the transverse relaxation time (T2 )j . Usually,
the time constants vary from a few microseconds up to several seconds, depend-
ing on the material under investigation. The sum is over the i different water
environments. The number of exponential components, M , can range from 1 to
a large number, especially in highly heterogeneous systems. In the applications,
most relaxation curves can be well described with 2 to 3 exponential terms, even
if this number strongly affects the solution. In complex multiphase systems, the
decay constants have values close enough to each other, so a continuous distri-
bution of the different water sites better represents the signal in integral form
as  ∞
s(t) = K(t, τ )g(τ )dτ.
0
150 A. Romano et al.

Under these assumptions, LTI methods are mandatory to analyse the NMR
relaxation decay. Nevertheless LTI methods require the LT function analytically
known almost everywhere in its domain [5–7,17], so a preprocessing of the data,
is needed to bridge the discrete data with the continuous model. For our problem
we describe the finite dataset of samples like:
 b
si = K(ti , τ )g(τ )dτ + εi , a, b ∈ R+ , with si = s(ti ), i = 1, . . . , N,
a
(2)
with i noise components. Once assumed the integral form (2), the solution g
is computed by solving a constrained quadratic programming problem that is
defined by regularizing the following problem:

min s − Ag2 (3)

g
s.t. g(τ ) ≥ 0 (4)

where · is the Euclidean norm, s is the data vector and g the discrete solution,
following the next procedure:

Procedure . DLTI for NMR data analysis

1: Approximate the integrals (2) by numerical integration:

Ng

si = am K(ti , τm )g(τm ) + εi , i = 1, . . . , N (5)
m=1

with am weights of the quadrature formula, Ng number of the τm grid points. The
computation of the si in (5) requires the solution of the linear system:
Ng

si = Ai,m gm + εi , i = 1, . . . , N (6)
m=1

with gm := g(τm ) and Ai,m = am K(ti , τm ), m = 1, . . . , Ng , i = 1, . . . , N .

2: Constraints on the solution; in order to reduce the degree of freedoms on the solu-
tion, the non-negative inverse function is required, equal to zero at the extremes:

gm ≥ 0, m = 1, . . . , Ng (7)
g1 = 0, gNg = 0 (8)

3: Regularization; moving from (2) to (4) the ill-posedness becomes ill-conditioning,

so a regularized solution of (6) is computed by solving the weighted least-square
problem

min M−1/2 (s − Ag)2 + α2 Dg2 subject to (7) and (8), (9)

g
 NMR Data Analysis of Water Mobility in Wheat Flour Dough 151

Procedure . DLTI for NMR data analysis (cont.)

where M is the (positive definite) covariance matrix of the error distribution εi

and D is the matrix depending on the discretization scheme adopted for the second
order derivatives of g. Statistical prior knowledge is needed in order to consider the
randomness on the solution due to the noise. The penalty term Dg2 in (9), is
imposed, forcing the smoothness, selecting the solution with minimum number of
peaks. The regularization parameter α is critical to the quality of the inversion.

Fig. 1. Transverse relaxation time (T2) for water protons in flour doughs at mixing
time 13 min long. N = 200 samples.

As concern the choice of α, several methods have been proposed in literature:

Discrepancy Principle, Generalized Cross Validation, L-Curve Method. In our
numerical tests we choose it empirically, and a more rigorous estimation proce-
dure will be considered in future studies. The numerical solution of the problem
(9) by ad hoc algorithms for quadratic programming (see [24,25] and references
therein), is under investigation.

3 Numerical Experiments
In this section we present the data set and the corresponding inversion results.
A typical plot of data for determining T2 for water protons in flour doughs with
dough consistency of 500 BU is shown in Fig. 1. The relaxation curve should
152 A. Romano et al.

Fig. 2. Relaxation time distribution at different mixing time equal to 1 and 2 min (from
up to down).

reveal two components characterized by a short and a long relaxation time:

the first component represents the less mobile, or more tightly bound water
fraction, while the second one represents the more mobile water fraction [16]. The
most mobile water fraction is responsible for the large value of relative humidity
 NMR Data Analysis of Water Mobility in Wheat Flour Dough 153

Fig. 3. Relaxation time distribution at different mixing time equal to 3 min.

(RH) and water activity of the dough (above 95%) [8] allowing optimization of
storage condition and development of process condition which may result in an
extension of shelf life of this product family. The other water fraction behaves
like a structure component of the starch and the gluten: this fraction is hard
to remove and does not contribute to the overall RH of the system (see [23]).
Each of these water fractions is indeed split in several families of molecules that
specifically interact with the various components of the dough. Therefore the
information on water molecules with large or short relaxation times may be
correlated to bread staling processes.
The numerical experiments about the LTI of the mean values of three dif-
ferent acquisitions, each of N = 200 samples, after a mixing time of 1, 2, 3
and 13 min long, were carried out with MATLAB R2018a software on a Intel(R)
Core(TM) i5, 1.8 GHz processor. The RILT is computed by the Matlab code
rilt.m available at the Mathworks File Exchange. We remark that the LTI is an
ill-posed problem so only a possible solution is given, satisfying some constraints
imposed by the user, according to the data. We fixed the default constraints (7)–
(8). The relaxation times distribution {gj }j is the DLTI of the sampled signal
s, calculated by the described procedure. The minimization is computed by the
Matlab function fminsearch that uses the Nelder-Mead simplex algorithm as
described in [15]. In the Figs. 2 and 3 we describe the computed relaxation time
spectrum of the magnetization decay, corresponding to different mixing times,
equal to 1, 2 and 3 minutes, respectively. Due to the material dependent variety
of the NMR relaxation times the time scale of the relaxation time spectrum has
to be optimized for the samples under investigation. The starting relaxation time
154 A. Romano et al.

Fig. 4. Relaxation time distribution (top) and corresponding normalized residuals |si −
(Ag̃)i |/N , i = 1, . . . , N , between the measured si and the computed magnetization
decay values (Ag̃)i (bottom), after a time mixing 13 min long.
 NMR Data Analysis of Water Mobility in Wheat Flour Dough 155

should be in the order of the time resolution of the experiment, but larger than
zero, the end of the time scale in the order of twice the largest expected relax-
ation time component. In our experiments the maximum mixing time is 13 min,
so the end of the time scale is fixed to 106 milliseconds (ms). In combination
with the time scale, the resolution can be specified by the number of points,
which determines the calculation time of the ILT in a roughly quadratic way.
The number of points of the relaxation time spectrum is set to 20. The relax-
ation spectrum presents some peaks, related to the relaxation time: the larger
the relaxation time, the wider the peak. The integral of a peak in the relaxation
spectrum corresponds therefore to the contribution of the species exhibiting this
particular relaxation behaviour. In Fig. 4 (top) we report the relaxation time dis-
tribution after a time mixing 13 min long. The curve proves that one component
is clearly resolved in 13 min.
In order to get an estimate of the goodness of the ILT, a plot of the residuals as
function of experiment time, between the measured and the calculated magneti-
zation decay corresponding to a mixing time of 13 min, is also displayed in Fig.
4 (bottom).

4 Conclusions
The interactions of the flour polymers with water during mixing reduce water
mobility and result in different molecular mobilities in dough. The results showed
that 1H NMR is an effective method to have a deeper view of wheat dough
development during mixing. A NMR-inversion recovery experiment based on
a RILT algorithm furnishes the relaxation time distribution and reveals the
dominance of one component responsible in the changes in water mobility of
dough. In the LTI algorithm the choice of the regularization parameter α and
the more suitable algorithm for the research of the optimal solution, are still
open issues. The calibration on the data tailoring the model on the exponential
sum defining the relaxation decay function will be object of future studies.

Acknowledgements. This work was partially supported by Gruppo Nazionale per il

Calcolo Scientifico - Istituto Nazionale di Alta Matematica (GNCS-INdAM) and by the
Centre for Food Innovation and Development in the Food Industry (CAISIAL), Uni-
versity of Naples Federico II, Portici (Naples), Italy. Rosanna Campagna was also sup-
ported by the MOD CELL DEV Project - Programma di finanziamento della Ricerca
di Ateneo, University of Naples Federico II.

References
1. Balestra, F., Laghi, L., Saa, D.T., Gianotti, A., Rocculi, P., Pinnavaia, G.: Physico-
chemical and metabolomic characterization of kamutkhorasan
R and durum wheat
fermented dough. Food Chem. 187, 451–459 (2015)
2. Barone, P., Ramponi, A., Sebastiani, G.: On the numerical inversion of the Laplace
transform for nuclear magnetic resonance relaxometry. Inverse Prob. 17(1), 77–94
(2001)
156 A. Romano et al.

3. Bosmans, G.M., Lagrain, B., Deleu, L.J., Fierens, E., Hills, B.P., Delcour, J.A.:
Assignments of proton populations in dough and bread using NMR relaxometry of
starch, gluten, and flour model systems. J. Agric. Food Chem. 60(21), 5461–5470
(2012)
4. Campagna, R., Conti, C., Cuomo, S.: Smoothing exponential-polynomial splines
for multiexponential decay data. Dolomites Res. Notes Approx. 12, 86–100 (2019)
5. Cuomo, S., D’Amore, L., Murli, A.: Error analysis of a collocation method for
numerically inverting a laplace transform in case of real samples. J. Comput. Appl.
Math. 210(1), 149–158 (2007). Proceedings of the Numerical Analysis Conference
2005
6. Cuomo, S., D’Amore, L., Murli, A., Rizzardi, M.: Computation of the inverse
Laplace Transform based on a collocation method which uses only real values. J.
Comput. Appl. Math. 198(1), 98–115 (2007)
7. Cuomo, S., D’Amore, L., Rizzardi, M., Murli, A.: A modification of weeks’ method
for numerical inversion of the Laplace transform in the real case based on automatic
differentiation. In: Bischof, C.H., Bücker, H.M., Hovland, P., Naumann, U., Utke,
J. (eds.) Advances in Automatic Differentiation. LNCSE, pp. 45–54. Springer, Hei-
delberg (2008). https://doi.org/10.1007/978-3-540-68942-3 5
8. Czuchajowska, Z., Pomeranz, Y.: Differential scanning calorimetry, water activ-
ity, and moisture contents in crumb center and near-crust zones of bread during
storage. Cereal Chem. 66, 305–309 (1989)
9. D’Amore, L., Campagna, R., Mele, V., Murli, A.: ReLaTIve. An Ansi C90 software
package for the real Laplace transform inversion. Numer. Algorithms 63(1), 187–
211 (2013)
10. D’Amore, L., Campagna, R., Mele, V., Murli, A.: Algorithm 946: ReLIADiff-A
C++ software package for real laplace transform inversion based on algorithmic
differentiation. ACM Trans. Math. Softw. 40(4), 31:1–31:20 (2014)
11. D’Amore, L., Mele, V., Campagna, R.: Quality assurance of Gaver’s formula for
multi-precision Laplace transform inversion in real case. Inverse Probl. Sci. Eng.
26(4), 553–580 (2018)
12. Doona, C., Baik, M.Y.: Molecular mobility in model dough systems studied by
time-domain nuclear magnetic resonance spectroscopy. J. Cereal Sci. 45(3), 257–
262 (2007)
13. Goesaert, H., Slade, L., Levine, H., Delcour, J.A.: Amylases and bread firming: an
integrated review. J. Cereal Sci. 50(3), 345–352 (2009). Special Section: Enzymes
in Grain Processing
14. Hills, B.: NMR studies of water mobility in foods. In: Roos, Y., Leslie, R., Lillford,
P. (eds.) Water Management in the Design and Distribution of Quality Foods, pp.
107–131. CRC Press, USA (1999)
15. Lagarias, J.C., Reeds, J.A., Wright, M.H., Wright, P.E.: Convergence properties of
the Nelder-Mead simplex method in low dimensions. SIAM J. Optim. 9, 112–147
(1998)
16. Leung, H.K., Magnuson, J.A., Bruinsma, B.L.: Pulsed nuclear magnetic resonance
study of water mobility in flour doughs. J. Food Sci. 44(5), 1408–1411 (1979)
17. Murli, A., Cuomo, S., D’Amore, L., Galletti, A.: Numerical regularization of a real
inversion formula based on the Laplace transform’s eigenfunction expansion of the
inverse function. Inverse Probl. 23(2), 713–731 (2007)
18. Provencher, S.W.: A constrained regularization method for inverting data repre-
sented by linear algebraic or integral equations. Comput. Phys. Commun. 27(3),
213–227 (1982)
 NMR Data Analysis of Water Mobility in Wheat Flour Dough 157

19. Raiola, A., Romano, A., Shanakhat, H., Masi, P., Cavella, S.: Impact of heat treat-
ments on technological performance of re-milled semolina dough and bread. LWT
117, 108607 (2020)
20. Romano, A., Campagna, R., Masi, P., Cuomo, S., Toraldo, G.: Data-driven
approaches to predict states in a food technology case study. In: 2018 IEEE 4th
International Forum on Research and Technology for Society and Industry (RTSI),
pp. 1–5, September 2018
21. Romano, A., Cavella, S., Toraldo, G., Masi, P.: 2D structural imaging study of
bubble evolution during leavening. Food Res. Int. 50(1), 324–329 (2013)
22. Romano, A., Cesarano, L., Sarghini, F., Masi, P.: The influence of arabinogalactan
on wheat dough development during mixing and leavening process. Chem. Eng.
Trans. 32, 1765–1770 (2013)
23. Schiraldi, A., Fessas, D.: Classical and knudsen thermogravimetry to check states
and displacements of water in food systems. J. Therm. Anal. Calorim. 71(1), 225–
235 (2003)
24. di Serafino, D., Toraldo, G., Viola, M., Barlow, J.: A two-phase gradient method
for quadratic programming problems with a single linear constraint and bounds
on the variables. SIAM J. Optim. 28(4), 2809–2838 (2018)
25. di Serafino, D., Ruggiero, V., Toraldo, G., Zanni, L.: On the steplength selection in
gradient methods for unconstrained optimization. Appl. Math. Comput. 318, 176–
195 (2018). Recent Trends in Numerical Computations: Theory and Algorithms
First Order Methods in Optimization:
Theory and Applications
 A Limited Memory Gradient Projection
Method for Box-Constrained Quadratic
Optimization Problems

Serena Crisci1,3(B) , Federica Porta2,3 , Valeria Ruggiero1,3 ,

and Luca Zanni2,3
1
Department of Mathematics and Computer Science,
University of Ferrara, via Machiavelli 30, 44121 Ferrara, Italy
{serena.crisci,valeria.ruggiero}@unife.it
2
Department of Physics, Informatics and Mathematics,
University of Modena and Reggio Emilia, via Campi 213/B, 41125 Modena, Italy
{federica.porta,luca.zanni}@unimore.it
3
Member of the INdAM-GNCS Research group, Rome, Italy

Abstract. Gradient Projection (GP) methods are a very popular tool

to address box-constrained quadratic problems thanks to their simple
implementation and low computational cost per iteration with respect,
for example, to Newton approaches. It is however possible to include,
in GP schemes, some second order information about the problem by
means of a clever choice of the steplength parameter which controls
the decrease along the anti-gradient direction. Borrowing the analysis
developed by Barzilai and Borwein (BB) for an unconstrained quadratic
programming problem, in 2012 Roger Fletcher proposed a limited mem-
ory steepest descent (LMSD) method able to effectively sweep the spec-
trum of the Hessian matrix of the quadratic function to optimize. In this
work we analyze how to extend the Fletcher’s steplength selection rule
to GP methods employed to solve box-constrained quadratic problems.
Particularly, we suggest a way to take into account the lower and the
upper bounds in the steplength definition, providing also a theoretical
and numerical evaluation of our approach.

Keywords: Quadratic programming · Gradient projection methods ·

Steplength selection rule · Ritz-like values

1 Introduction
Let consider the following box-constrained quadratic problem (QP)
1 T
min f (x) ≡ x Ax − bT x + c, (1)
≤x≤u 2
where A ∈ Rn×n is a symmetric and positive definite matrix, b, , u are vectors
of Rn , with  ≤ u, and c is a scalar. Hereafter, we denote the feasible region by
Ω = {x ∈ Rn :  ≤ x ≤ u}.
c Springer Nature Switzerland AG 2020
Y. D. Sergeyev and D. E. Kvasov (Eds.): NUMTA 2019, LNCS 11973, pp. 161–176, 2020.
https://doi.org/10.1007/978-3-030-39081-5_15
162 S. Crisci et al.

We are interested in solving problem (1) by means of the well-known Gradient

Projection (GP) method combined with a linesearch strategy along the feasible
direction [2, Chapter 2]. The standard GP iteration can be written as
x(k+1) = x(k) + νk dk =
  (2)
= x(k) + νk PΩ (x(k) − αk g (k) ) − x(k) ,

where PΩ (·) denotes the projection operator onto the constraints set, g (k) stands
for ∇f (x(k) ), αk is the steplength parameter that controls the decrease along the
anti-gradient and νk ∈ (0, 1] is computed by means of a backtracking procedure,
as for example the monotone or non-monotone Armijo linesearch [2,8].
Convergence results on the GP method have been obtained also in the case
of general continuously differentiable objective functions, as summarized in the
following theorem. For further details, see [3,10].
Theorem 1. [3, Th 2.2] Let consider the problem
min F (x),
x∈Ω
n
where F : R → R is a continuously differentiable function and Ω is a closed
convex subset of Rn . Let assume αk ∈ [αmin , αmax ], with 0 < αmin < αmax ,
and νk obtained by either a monotone or a non-monotone linesearch. Then,
each limit point of the sequence {xk }k∈N generated by GP method is a stationary
point for the considered problem. If, in addition, F is a convex function with a
Lipschitz-continuous gradient, the set of solutions is not empty and the initial
level set is bounded, we have that
 
∗ 1
F (x ) − F = O
(k)
,
k
where F ∗ is the minimum value [10, Th. 3.3].
Thanks to the very general hypothesis on αk , a clever choice of such param-
eter can be exploited in order to accelerate the practical performance of the
GP method. In order to understand the key principle to properly select the
steplength αk , we firstly recall some useful considerations relative to the easier
case of the unconstrained minimization of the quadratic function in (1).
In this case, the GP approach (2) reduces to a standard gradient method whose
iteration is given by
x(k+1) = x(k) − αk g (k) , (3)
and the corresponding gradient recurs according to the rule
g (k+1) = g (k) − αk Ag (k) . (4)
By denoting with {λ1 , λ2 , . . . , λn } the eigenvalues of A and with {v1 , v2 , . . . , vn }
a set of associated orthonormal eigenvectors, the gradient g (k+1) can be expressed
as
n
(k+1)
g (k+1) = μi vi , (5)
i=1
 A Limited Memory Gradient Projection Method 163

(k+1)
where μi ∈ R is called the i-th eigencomponent of g (k+1) and satisfies the
following recurrence formula:
k

(k+1) (0) (k)
μi = μi (1 − αj λi ) = μi (1 − αk λi ). (6)
j=0

Since, in absence of constraints, the gradient components at the solution are

equal to zero, equality (6) highlights that if the steplength αk is able to approxi-
(k+1)
mate the inverse of any eigenvalue λi , the corresponding eigencomponent μi
is damped. Indeed, the most popular steplengths strategies developed for the
unconstrained quadratic case (in particular those based on the Barzilai-Borwein
rules [1]) are usually aimed at capturing this property, providing suitable approx-
imations of the inverse of the eigenvalues of the Hessian matrix A and trying to
reduce the gradient eigencomponents by means of an effective sweeping of the
spectrum of A−1 (see, for example, [1,5,6,12,13]).
Among them, a promising steplength updating strategy has been suggested
in [5] leading to the development of the Limited Memory Steepest Descent
(LMSD) algorithm. Several numerical experiments [5] show the effectiveness of
this steplength selection rule with respect to other state-of-the art strategies for
choosing the steplength.
As a consequence of these encouraging results and taking into account that
the convergence for the GP method (2) is guaranteed for every choice of the
steplength in a closed and bounded interval, the aim of this paper is to under-
stand how to apply the limited memory steplength approach to the GP algo-
rithms for box-constrained quadratic problems. Particularly, we investigate if
it may be still convenient to approximate the eigenvalues of the whole Hessian
matrix or the spectrum of a more suitable matrix should be considered.
The paper is organized as follows. In Sect. 2 we recall the standard LMSD
method and we develop its possible generalization to quadratic problems with
upper and lower bounds. In Sect. 3 we present the results of different numerical
experiments to evaluate the validity of the suggested generalized limited memory
steplength strategy. Our conclusions are reported in Sect. 4.

Notation. In the following we denote by AC,D the submatrix of A of order

C × D given by the intersection of the rows and the columns with indices in
the sets C and D respectively. Moreover, xC ∈ RC stands for the subvector of x
with entries indexed in C.

2 A Limited Memory Steplength Selection Rule for

Box-Constrained Quadratic Problems

We start this section by recollecting the main features of the original LMSD
algorithm for unconstrained quadratic optimization problems and then we move
to the more general case of box-constrained ones.
164 S. Crisci et al.

2.1 The State-of-the Art About LMSD

The basic idea of this method is to divide up the sequence of iterations of a
steepest descent scheme into groups of m iterations, referred to as sweeps, where
m is a small positive integer, with the aim of providing, at each new sweep, m
approximations of the eigenvalues of the Hessian matrix.
The starting point of such idea is that m steps of the Lanczos iterative pro-
cess applied to the matrix A, with starting vector q1 = g (k) /g (k) , provide a
tridiagonal m × m matrix T whose eigenvalues, called Ritz values, are approxi-
mations of m eigenvalues of A [7]. Indeed, given an integer m ≥ 1, the Lanczos
process generates orthonormal n-vectors {q1 , q2 , . . . , qm } that define a basis for
the Krylov sequence

g (k) , Ag (k) , A2 g (k) , . . . , Am−1 g (k) (7)

and such that the matrix

T = QT AQ (8)
T
is tridiagonal, where Q = [q1 , q2 , . . . , qm ], Q Q = I. The inverses of the m eigen-
values of the matrix T are employed as steplengths for the successive m itera-
tions. The matrix T can be computed without involving the matrix A explicitly.
To reach this goal we need the two following remarks.
1. If a limited number m of back values of the gradient vectors
G = [g (k) g (k+1) · · · g (k+m−1) ]
is stored and the (m + 1) × m matrix J containing the inverses of the corre-
sponding last m steplengths is considered
⎡ 1 ⎤
αk
⎢ .. ⎥
⎢− 1 . ⎥
⎢ ⎥
J = ⎢ αk ⎥
⎢ .. 1 ⎥
⎣ . αk+m−1 ⎦
− αk+m−1
1

then the equations arising from (4) can be rewritten in matrix form as
AG = [G, g (k+m) ]J. (9)
2. Taking into account (4) and that the columns of G are in the space generated
by the Krylov sequence (7), we have G = QR, where R is m × m upper
triangular and nonsingular, assuming G is full-rank.
We remark that R can be obtained from the Cholesky factorization of GT G and
the computation of Q is not required. Then from both (9) and GT G = RT R, it
follows that the tridiagonal matrix T can be written as
T = QT AQ = R−T GT [G g (k+m) ]JR−1 = [R r]JR−1 , (10)
where the vector r is the solution of the linear system RT r = GT g (k+m) . Proper
techniques can be adopted to address the case of rank-deficient G.
 A Limited Memory Gradient Projection Method 165

2.2 From the LMSD Method to the Limited Memory GP method

When we solve a quadratic box-constrained problem by means of the GP method
(2), the procedure to build the matrix G using the whole gradients seems not
convenient. Indeed, in this case, the relation (9) does not hold and the matrix
obtained by the procedure (10) does not exhibit a tridigonal structure but it
is an upper Hessenberg matrix. A first attempt at suggesting a possible way to
employ the limited memory steplength selection rule in the framework of box-
constrained optimization has been done in [11]. Here, the authors propose to
consider a generalized matrix T , called T̃ , defined as
T̃ = R̃−T G̃T [G̃ g̃ (k+m) ]J˜R̃−1 ,
where the general vector g̃ (k) is given by
 (k) (k)
(k) gj if j < xj < uj
g̃j = , (11)
0 otherwise

G̃ is the n × m matrix [g̃ (k) g̃ (k+1) · · · g̃ (k+m−1) ], R̃ is such that R̃T R̃ = G̃T G̃
and J˜ is the (m + 1) × m lower bidiagonal matrix
⎡ 1 ⎤
αk νk
⎢ .. ⎥
⎢− 1 . ⎥
⎢ ⎥
J˜ = ⎢ αk νk ⎥. (12)
⎢ .. 1 ⎥
⎣ . αk+m−1 νk+m−1 ⎦
− αk+m−11νk+m−1

However, in [11] no investigation related to both the spectral properties of the

steplengths generated by the suggested approach and possible extended recur-
rence formulas of kind (9) has been addressed. In the following, starting from
an analysis aimed at clarifying the relation between subsequent gradients in a
sweep, a different idea to generalize the standard limited memory procedure to
GP algorithms for box-constrained quadratic problems will be developed.
At any iteration, let denote the following set of indices
(k) (k)
Fk = {i : i ≤ xi − αk gi ≤ ui }
Bk = N \ Fk , with N = {1, ..., n}
in order to distinguish the cases where the projection onto the feasible set is
computed from the cases where it has no effect on the components of the current
iterate. The entries of the iterate x(k+1) generated by the GP method (2) are
⎧  
⎨ x(k) + νk x(k) − αk g (k) − x(k) i ∈ Fk ,
i i i i
(k+1)
xi =   (13)
⎩x +(k) (k)
νk γi − xi
(k)
i ∈ Bk ,
i

where 
(k) (k)
(k) i if xi − αk gi < i ,
γi = (k) (k)
ui if xi − αk gi > ui .
166 S. Crisci et al.

As a consequence, for any i = 1, . . . , n, the new gradient components are given by

n

(k+1) (k+1)
gi = aij xj − bi =
j=1
     
(k) (k) (k) (k) (k)
= aij xj − νk αk gj + aij xj − νk (xj − γj ) − bi
j∈Fk j∈Bk

(k)
 (k)
 (k)
xj
(k)
− γj
= gi − νk αk aij gj − νk αk aij .
αk
j∈Fk j∈Bk

From the previous equation we can write

  1 
(k) αk νk
(k) (k+1)
AFk ,Fk gFk = gF gFk − AFk ,N p(k) (14)
k − αk1νk

where p(k) is a vector with n entries defined as


(k)
0 i ∈ Fk ,
pi = x(k) −γ
(k) (15)
i
αk
i
i ∈ Bk .

At the next iteration, using the same argument employed to obtain (14), we get
 
  1
(k+1) (k+1) (k+2) αk+1 νk+1
AFk+1 ,Fk+1 gFk+1 = gFk+1 gFk+1 − AFk+1 ,N p(k+1) (16)
− αk+11νk+1

with the obvious definitions for Fk+1 and Bk+1 . At this point, under the assump-
tion Fk ∩ Fk+1
= ∅, we consider the following subsets of indices by taking into
account all the possible cases that may occur at the (k + 1)-th iteration

F(k,k+1) := Fk ∩ Fk+1
k
F (k,k+1) := Fk \ (Fk ∩ Fk+1 )
k+1
F (k,k+1) := Fk+1 \ (Fk ∩ Fk+1 ) .

From (14) and (16), we may write

 
(k) (k+1)
AF(k,k+1) ,F(k,k+1) gF(k,k+1) gF(k,k+1) =
⎡ 1 ⎤
  αk νk 0
(k) (k+1) (k+2) ⎢ 1 1 ⎥

= gF(k,k+1) gF(k,k+1) gF(k,k+1) ⎣− αk νk αk+1 νk+1 ⎦ − AF(k,k+1) ,N p(k) p(k+1) +
0 − αk+11νk+1

(k) (k+1)
− AF (k,k+1),F
k g k
(k,k+1)
AF
F (k,k+1) ,F
k+1 g k+1
(k,k+1)
.
(k,k+1) F (k,k+1)
 A Limited Memory Gradient Projection Method 167

The argument can be generalized to a sweep of length m starting from the

iteration k. By defining F(k,k+m−1) := ∩k+m−1 s=k Fs
= ∅ we have
 
(k) (k+m−1)
AF(k,k+m−1) ,F(k,k+m−1) gF(k,k+m−1) · · · gF(k,k+m−1) (17)
 
(k) (k+m−1) (k+m)
= gF(k,k+m−1) · · · gF(k,k+m−1) gF(k,k+m−1) J˜ +
 
−AF(k,k+m−1) ,N p(k) · · · p(k+m−1) +
 (k) (k+m−1)

− AF ,F
k g k · · · A F ,F
k+m−1 g k+m−1 ,
(k,k+m−1) (k,k+m−1) F (k,k+m−1) (k,k+m−1) (k,k+m−1) F (k,k+m−1)

where J˜ is the (m + 1) × m lower bidiagonal matrix given in (12).

If Fk+j ⊆ F(k,k+m−1) , for j = 0, . . . , m − 1, the term
(k+j)
AF k+j g k+j
(k,k+m−1) ,F k,k+m−1 F(k,k+m−1) ,F k,k+m−1

does not contribute to the relation (17), which can be rewritten as

 
(k) (k+m−1)
AF(k,k+m−1) ,F(k,k+m−1) gF(k,k+m−1) · · · gF(k,k+m−1) (18)
 
(k) (k+m−1) (k+m)
= gF(k,k+m−1) · · · gF(k,k+m−1) gF(k,k+m−1) J˜ +
 
−AF(k,k+m−1) ,N p(k) · · · p(k+m−1) .

In order to preserve the validity of (18) and to correctly neglect the term
(k+j)
AF ,F
k+j g k+j , for j = 0, . . . , m − 1, we propose to
(k,k+m−1) k,k+m−1 F(k,k+m−1) ,F k,k+m−1
interrupt a sweep and to restart the collection of new restricted gradient vectors
when the condition Fk+j ⊆ F(k,k+m−1) is not satisfied, by developing a tech-
nique which adaptively controls the length of the sweep, up to the given value m.
At the beginning of the iterative process, this condition does not typically hold
and the sweeps have a length at most equal to 1; however, as the number of
iterations increases, the components that are going to be projected onto the fea-
sible set tend to stabilize and, as a consequence, Fk+j ⊆ F(k,k+m−1) occurs for
a growing number of iterations. Hereafter, we suppose that Fk+j ⊆ F(k,k+m−1) ,
j = 0, ..., m − 1, holds. We can state that the equality (18) can be considered a
possible extension of the Eq. (9) which holds in the unconstrained framework.
As a consequence, in presence of box-constraints, we suggest to not store m back
whole gradients vectors (for which no recurrence formula holds) but to consider
m back gradients restricted to the set of indices F(k,k+m−1) . Driven by these
considerations, our implementation of the limited memory steplength rule for
the constrained case is based on the following generalization of the matrix G:
 
(k) (k+m−1)
G(k,k+m−1) = gF(k,k+m−1) · · · gF(k,k+m−1) . (19)
168 S. Crisci et al.

T
Given m ≥ 1 and the m × m matrix R(k,k+m−1) such that R(k,k+m−1)
T
R(k,k+m−1) = G(k,k+m−1) G(k,k+m−1) , we propose to compute, at each new
sweep, m steplengths as inverses of the eigenvalues of the symmetric matrix

−T
T̃(k,k+m−1) = R(k,k+m−1) GT(k,k+m−1) AF(k,k+m−1) ,F(k,k+m−1) G(k,k+m−1) R(k,k+m−1)
−1
,

with the aim of approximating the inverses of the eigenvalues of the matrix
AF(k,k+m−1) ,F(k,k+m−1) . This idea mimics the approach proposed in [4] where, in
the case of box constraints, novel versions of the BB rules sweeping the spectrum
of a proper submatrix of A turned out to be convenient with respect to the
standard ones. Indeed, in this case, under the special assumptions m = 1, Fk−1 =
(k) (k)
Fk and νk−1 = 1, in view of γBk = xBk , the recurrence (14) can be simplified as
(k+1) (k) (k)
gFk = gFk − νk αk AFk ,Fk gFk .
By denoting with {δ1 , . . . , δr } and {w1 , . . . , wr } the eigenvalues and the associ-
ated orthonormal eigenvectors of AFk ,Fk , respectively, where r = Fk , and by
(k+1) r (k+1) (k) r (k)
writing gFk = i=1 μ̄i wi and gFk = i=1 μ̄i wi , we obtain the following
recurrence formula for the eigencomponents:
(k+1) (k)
μ̄i = μ̄i (1 − νk αk δi ).
1
This means that if the selection rule provides a good approximation of δi , a
(k+1)
useful reduction of |μ̄i |
can be achieved. We underline that, if m = 1, αk
is computed in order to estimate the inverse of an eigenvalue of AFk−1 ,Fk−1 ;
obviously, if Fk−1 = Fk , αk can also provide a good approximation of one of the
(k+1)
values δ1i and thus reduce the corresponding desired component |μ̄i |.
We remark that, in view of (18), the matrix T̃(k,k+m−1) has the following
form
 
−T ˜ −1
GT(k,k+m−1) G(k,k+m−1) gF(k,k+m−1) JR
(k+m)
T̃(k,k+m−1) = R(k,k+m−1) (k,k+m−1) +
−T
  −1
−R(k,k+m−1) GT(k,k+m−1) AF(k,k+m−1) ,N p(k) · · · p(k+m−1) R(k,k+m−1) =
  −1
= R(k,k+m−1) , r(k,k+m−1) JR ˜
(k,k+m−1) + (20)
−T
  −1
+R(k,k+m−1) GT(k,k+m−1) AF(k,k+m−1) ,N p(k) · · · p(k+m−1) R(k,k+m−1) ,
T
where the vector r(k,k+m−1) is the solution of the system R(k,k+m−1) r(k,k+m−1) =
GT(k,k+m−1) gF(k,k+m−1) .
(k+m)

Despite the carried out analysis, from the practical point of view, we want
to avoid to explicitly make use of the matrix AF(k,k+m−1) ,N and, hence, we do
not consider the
 exact relation (18), but its inexact version where the term
AF(k,k+m−1) ,N p(k) · · · p(k+m−1) is neglected. For this reason, we do not com-
pute the eigenvalues of T̃(k,k+m−1) but the eigenvalues of the symmetric part of
the matrix
 
Z(k,k+m−1) = R(k,k+m−1) , r(k,k+m−1) JR ˜ −1
(k,k+m−1) .
 A Limited Memory Gradient Projection Method 169

To explain the relation between the eigenvalues of T̃(k,k+m−1) and the ones of
the symmetric part of Z(k,k+m−1) , we start to clarify the details of our approach
from the more understandable case of m = 1 where T̃(k,k+m−1) reduces to a
(k)
scalar. In this case, at iteration k + 1 only one gradient is available Gk = gFk .
We are interested in computing
T̃k = Rk−T GTk AFk ,Fk Gk Rk−1 ,
where
 
GTk Gk = (gFk )T gFk = (gFk )T gFk = RkT Rk .
(k) (k) (k) (k) (k) (k)
(gFk )T gFk

Then, using (14), the matrix T̃k is given by

  1  
−T T
Rk Gk (k) (k+1)
gFk gFk αk νk
− AFk ,N p (k)
Rk−1 (21)
− αk1νk
 (k) (k+1)
1 (k) T gF k − g F k 1
= (gFk ) − AFk ,N p(k) 
(k) T (k) αk νk (k) (k)
(gFk ) gFk (gFk )T gFk
(gFk )T AFk ,Fk gFk
(k) (k)
= (k) (k)
.
(gFk )T gFk
Hence, for the special case m = 1, if we would consider the exact expression
of AFk ,Fk Gk given by the right hand side of (14), we would obtain as unique
(k) (k)
(gF )T AFk ,Fk gF
k k
eigenvalue of T̃k the value (k) (k) which is the inverse of the Rayleigh
(gF )T gF
k k
quotient of the matrix AFk ,Fk and, hence, belongs to its spectrum.
However, in practice we compute the scalar
(gFk )T AFk ,Fk gFk (gFk )T AFk ,N p(k)
(k) (k) (k)
Zk = = (k) (k)
+ (k) (k)
(22)
(gFk )T gFk (gFk )T gFk
that is a value in the spectrum of AFk ,Fk affected by en error due to the presence
of the second term at the right-hand side of Eq (22). An estimation of this error,
at iteration k + 1, is given by
AFk ,N p(k) 
ρk ≤ (k)
.
gFk 
From Eqs. (13) and (15), the following results hold
(k+1) (k) (k+1) (k)
(k) xBk − xBk  (k) xFk − xFk 
p(k)  = pBk  = and gFk  = .
αk νk αk νk
As a consequence,
(k) (k) (k+1) (k)
AFk ,Bk pBk  AFk ,Bk pBk  AFk ,Bk xBk − xBk 
ρk ≤ (k)
≤ (k)
= (k+1) (k)
.
gFk  gFk  xFk − xFk 
170 S. Crisci et al.

From this bound on ρk , we can state that, when k is sufficiently large so as

the components to project onto the feasible region are almost settled, the error
ρk is not significant and the steplength αk+1 is approximating the inverse of
an eigenvalue of AFk ,Fk . In the more general case of m > 1, we compute the
eigenvalues of the symmetric part of Z(k,k+m−1) given by

1 T

Z̃(k,k+m−1) = Z(k,k+m−1) + Z(k,k+m−1) . (23)
2
Then, from Eq. (20) we have

Z̃(k,k+m−1) = T̃(k,k+m−1) + (24)

1    
−T −1
+ R(k,k+m−1) GT(k,k+m−1) AF(k,k+m−1) ,N p(k) · · · p(k+m−1) R(k,k+m−1) +
2
1   (k) T 
−T −1
+ R(k,k+m−1) p · · · p(k+m−1) ATF(k,k+m−1) ,N G(k,k+m−1) R(k,k+m−1) .
2
A result of perturbation matrix theory (see Corollary 6.3.4 [9]) ensures that
!    ! " "
! ! " "
!λj Z̃(k,k+m−1) − λj T̃(k,k+m−1) ! ≤ "Z̃(k,k+m−1) − T̃(k,k+m−1) " , (25)

where λj (C) is the j−th eigenvalue of C. By denoting with DF the Frobenius
norm of a matrix D, the right-hand side of (25) can be bounded from above as
" "
" "
"Z̃(k,k+m−1) − T̃(k,k+m−1) " ≤
"   −1 "
" −T "
≤ "R(k,k+m−1) GT(k,k+m−1) AF(k,k+m−1) ,N p(k) · · · p(k+m−1) R(k,k+m−1) "≤
"  (k) "
≤ "AF(k,k+m−1) ,N p · · · p(k+m−1) " R(k,k+m−1)  ≤
−1
" " " " −1
≤ "AF(k,k+m−1) ,N " " p(k) · · · p(k+m−1) " R(k,k+m−1) ≤
" " " "
≤ "AF(k,k+m−1) ,N " " p(k) · · · p(k+m−1) "F R(k,k+m−1)  ≤
−1

#
$m−1 "
" "$  " (k+i) " "2 −1
≤ "AF(k,k+m−1) ,N " % "pBk+i " R(k,k+m−1) ,
i=0

where in the second inequality we use


−T
R(k,k+m−1) GT(k,k+m−1)  = −T
R(k,k+m−1) −1
GT(k,k+m−1) G(k,k+m−1) R(k,k+m−1)  = 1.

We can conclude that, if m is relatively small, as the number of iterations

k increases, the matrix Z̃(k,k+m−1) (of which we compute the eigenvalues)
approaches the matrix T̃(k,k+m−1) (of which we would compute the eigenval-
ues).
 A Limited Memory Gradient Projection Method 171

3 Numerical Experiments
In this section we analyse, on some box-constrained quadratic problems, the
practical behaviour of the GP method combined with

– the original limited memory (LM) steplength selection rule (which collects
the whole back gradients),
– the modified LM steplength selection rule suggested in [11] (which considers
the modified gradients given in (11)),
– the modified LM steplength selection rule suggested in Sect. 2.2 (which
exploits the matrix G(k,k+m−1) defined in (19)).

Our main aim is to investigate the distribution of the inverses of the steplengths
generated by the three approaches with respect to the eigenvalues of a proper
submatrix of the Hessian matrix.
Since in the practical implementation of the modified LM updating strategy
proposed in Sect. 2.2, we compute the steplengths as the inverses of the eigenval-
ues of the matrix (23), by analogy we generate the steplengths provided by both
the standard LM procedure and the one proposed in [11] as the inverses of the
eigenvalues of the symmetric part of the Hessenberg matrices T and T̃ , respec-
tively, instead of reducing them to a tridiagonal form, as made in [5] and [11].
Nevertheless, numerical experiments show that this modification does not signif-
icantly affect the results. Hereafter we denote by LMGP, Box-LMGP1 and Box-
LMGP2 the GP algorithm equipped with the standard LM steplength selection
rule and the modified versions developed in [11] and Sect. 2.2, respectively.
In our tests, the LMGP, Box-LMGP1 and Box-LMGP2 methods share a
monotone Armijo-type linesearch procedure to select νk and the same stopping
criterion:
ϕ(x(k) ) ≤ tolg(x(0) ), (26)
where ϕ(x(k) ) is the projected gradient at x(k) , i.e., the vector with entries
(k)
ϕi , i = 1 . . . , n, defined as
⎧
⎪
⎪
(k) (k)
⎨ gi for i < xi < ui ,
(k) (k) (k)
ϕi = max{0, gi } for xi = ui , (27)
⎪
⎪
⎩ (k) (k)
min{0, gi } for xi = i .

The following parameter setting is used: tol = 10−8 , αmin = 10−10 , αmax = 106 ,
T T
α0 = (g (0) g (0) )/(g (0) Ag (0) ); furthermore, different values for the parameter m
are used, i.e. m = 3, 5, 7. The feasible initial point x(0) is randomly generated
with inactive entries.
We start to analyse the effect of the considered steplength selection rules
within the GP method on a toy problem of size n = 20 with ten active con-
straints at the solution; the eigenvalues of the Hessian matrix are logarithmi-
cally distributed and the condition number is equal to 500. In Fig. 1 we report
the behaviour of α1k (red crosses) with respect to the eigenvalues of the Hessian
172 S. Crisci et al.

Table 1. Main features of two quadratic test problems subject to lower bounds

n = 1000 na = 400
λmin (A) λmax (A) Distribution of the λmin (AF ∗ ,F ∗ ) λmax (AF ∗ ,F ∗ )
eigenvalues of A
TP1 1 1000 log-spaced 3.08 753.26
TP2 9.40 1013.95 log-spaced 10 1000

matrix (green dotted lines) and the restricted Hessian submatrix (black dotted
line) at each iteration k, for the case m = 5. In each panel of the figure, the blue
lines show the maximum and the minimum eigenvalues of the whole Hessian
matrix and the blue symbols “o” are used to denote the eigenvalues of the sub-
matrix AF ∗ ,F ∗ . We observe that the inverses of the steplengths produced by the
LMGP method may sometimes fall outside the spectrum of the restricted Hes-
sian or even the spectrum of the whole Hessian, while the other two approaches
1
are able to restrain this effect. In particular, the sequence αk generated by
the Box-LMGP2 scheme, belongs to the spectra of the current restricted Hessian
matrices, providing also a reduction of the iterations needed to satisfy the stop-
ping criterion. Indeed, the effectiveness of the Box-LMGP2 procedure allows an
earlier stabilization of the active set, with respect to the other two approaches.
The numerical behaviour of the considered methods was also verified on test
problems of larger size. In particular, we randomly generated quadratic test
problems subject to lower bounds, in which the solution, the number of active
constraints at the solution, and the distribution of the eigenvalues of the dense
symmetric positive definite Hessian matrix of the objective function are prefixed.
For the sake of brevity, we only report here the results obtained on two test prob-
lems described in Table 1, where na is the number of active lower constraints at
the solution and F ∗ denotes the set of the indices of the inactive constraints at
the solution, so that AF ∗ ,F ∗ is the submatrix of the Hessian matrix defined by
the intersection of the rows and the columns with indices in F ∗ .

102 102 102

101 101 101

100 100 100

0 50 100 150 0 20 40 60 80 100 0 20 40 60 80

Fig. 1. Distribution of α1k with respect to iterations for LMGP (left panel), Box-
LMGP1 (middle panel) and Box-LMGP2 (right panel) on a toy problem of size n = 20,
for a sweep of length m = 5.
 A Limited Memory Gradient Projection Method 173

In Table 2 we report, for each method, the number of iterations (iterations)

necessary to satisfy the stopping criterion, the total number of backtracking
reductions (backtr.) on the parameter νk , and the final error on the objective
function values, i.e. Fk = |f (x∗ ) − f (x(iterations) )|.

Table 2. Numerical results for problems TP1 and TP2

method iterations backtr. Fk iterations backtr. Fk

TP1 TP2
LMGP (m = 3) 384 131 3.96e-09 292 81 1.12e-06
Box-LMGP1 (m = 3) 233 25 4.19e-09 116 13 2.72e-07
Box-LMGP2 (m = 3) 222 23 9.31e-10 109 16 1.96e-07
LMGP (m = 5) 355 112 1.51e-08 172 25 1.04e-06
Box-LMGP1 (m = 5) 252 18 1.00e-08 114 8 9.76e-08
Box-LMGP2 (m = 5) 221 25 7.22e-09 101 8 4.82e-07
LMGP (m = 7) 375 89 1.40e-08 152 16 1.11e-06
Box-LMGP1 (m = 7) 255 16 2.03e-08 112 5 5.49e-07
Box-LMGP2 (m = 7) 233 25 7.92e-09 104 4 1.02e-07

Figure 2 enables to show the behaviour of the rules on test problem TP1. Similar
results can be observed for TP2. In each panels of Fig. 2, at the k-th iteration,
the black dots denote 20 eigenvalues (with linearly spaced indices, included the
maximum and the minimum eigenvalues) of the submatrix of the Hessian matrix,
defined by neglecting the rows and columns with indices corresponding to the
active variables at the current iterate, and the red cross corresponds to the
inverse of the steplength αk . The blue lines show the maximum and the minimum
eigenvalues of the whole Hessian matrix and the blue symbols “o” are used to
denote 20 eigenvalues of the submatrix AF ∗ ,F ∗ , with linearly spaced indices
(included the maximum and the minimum eigenvalues). The results shown in
Figs. 1 and 2 and Table 2 allow us to make the following considerations:

– in general the steplengths generated by the Box-LMGP1 and the Box-LMGP2

methods seem to better approximate the eigenvalues of the submatrices of A
restricted to the rows and columns corresponding to the inactive components
of the current iterate with respect to the LMGP algorithm;
– the adaptive strategy implemented by the Box-LMGP2 approach turns out
to be more effective for m = 5 and m = 7, thanks to its previously described
ability in adaptively controlling the length of the sweep. Indeed such ability
allows to consider shorter sweeps at the beginning of the iterative process
when the sequence of the restricted Hessian matrices is not yet stabilized
towards AF ∗,F ∗ and, hence, the inverses of the steplengths generated by the
limited memory strategy in a particular sweep could not provide suitable
approximations of the eigenvalues of the restricted Hessian matrices involved
174 S. Crisci et al.

250

200
300

200

150
150
200

100
100
100

50
50
0

0
0
103

102

101

100

103

102

101

100
103

102

101

100
250

200
300

200

150
150
200

100
100
100

50
50
0

0
0
103

102

101

100

103

102

101

100
103

102

101

100

200
200
300

150
150
200

100
100
100

50
50
0

0
103

102

101

100

103

102

101

100

103

102

101

100

Fig. 2. Distribution of α1k with respect to the iterations for LMGP (first row), Box-
LMGP1 (second row) and Box-LMGP2 (third row) on TP1, for different values of the
length of the sweep.

in the next sweep. Longer sweeps are instead promoted with the stabilization
of the final active set; furthermore, from Table 2, we may observe that the Box-
LMGP2 method, compared with LMGP and Box-LMGP1, generally requires
lower numbers of backtracking steps;
 A Limited Memory Gradient Projection Method 175

– within the same or higher accuracy, the effectiveness of the Box-LMGP2

method is confirmed by the lower number of iterations with respect to the
LMGP and Box-LMGP1 methods.

4 Conclusions
In this paper we developed an updating strategy to select the steplength in
gradient projection methods for the minimization of box-constrained quadratic
problems. In particular, we generalized a steplength selection rule proposed in
the unconstrained optimization framework and based on the storage of a limited
number of consecutive objective function gradients. By preserving the same basic
idea of exploiting stored gradient vectors, we detailed how to possibly modify the
original updating strategy in order to take into account the lower and the upper
bounds. Numerical experiments carried out on box-constrained quadratic test
problems showed that the modified procedure allowed the gradient projection
method to reach better practical performance with respect to the standard one.

Acknowledgments. We thank the anonymous reviewers for their careful reading of

our manuscript and their many insightful comments and suggestions. This work has
been partially supported by the INDAM research group GNCS.

References
1. Barzilai, J., Borwein, J.M.: Two-point step size gradient methods. IMA J. Numer.
Anal. 8, 141–148 (1988)
2. Bertsekas, D.P.: Nonlinear Programming, 2nd edn. Athena Scientific, Boston
(1999)
3. Birgin, E.G., Martinez, J.M., Raydan, M.: Non-monotone spectral projected gra-
dient methods on convex sets. SIAM J. Optim. 10(4), 1196–1211 (2000)
4. Crisci, S., Ruggiero, V., Zanni, L.: Steplength selection in gradient projection meth-
ods for box-constrained quadratic programs. Appl. Math. Comput. 356, 312–327
(2019)
5. Fletcher, R.: A limited memory steepest descent method. Math. Program. Ser. A
135, 413–436 (2012)
6. Frassoldati, G., Zanghirati, G., Zanni, L.: New adaptive stepsize selections in gra-
dient methods. J. Ind. Manage. Optim. 4(2), 299–312 (2008)
7. Golub, G.H., van Loan, C.F.: Matrix Computations, 3rd edn. John Hopkins Uni-
versity Press, Baltimore (1996)
8. Grippo, L., Lampariello, F., Lucidi, S.: A nonmonotone line-search technique for
Newton’s method. SIAM J. Numer. Anal. 23(4), 707–716 (1986)
9. Horn, R.A., Johnson, C.R.: Matrix Analysis. Cambridge University Press, Cam-
bridge (2012)
10. Huang, Y., Liu, H.: On the rate of convergence of projected Barzilai-Borwein meth-
ods. Optim. Meth. Softw. 30(4), 880–892 (2015)
11. Porta, F., Prato, M., Zanni, L.: A new steplength selection for scaled gradient
methods with application to image deblurring. J. Sci. Comput. 65, 895–919 (2015)
176 S. Crisci et al.

12. di Serafino, D., Ruggiero, V., Toraldo, G., Zanni, L.: On the steplength selection
in gradient methods for unconstrained optimization. Appl. Math. Comput. 318,
176–195 (2018)
13. Zhou, B., Gao, L., Dai, Y.H.: Gradient methods with adaptive step-sizes. Comput.
Optim. Appl. 35(1), 69–86 (2006)
 A Gradient-Based Globalization Strategy
for the Newton Method

Daniela di Serafino1 , Gerardo Toraldo2 , and Marco Viola1(B)

1
Department of Mathematics and Physics, University of Campania “L. Vanvitelli”,
Caserta, Italy
{daniela.diserafino,marco.viola}@unicampania.it
2
Department of Mathematics and Applications, University of Naples Federico II,
Naples, Italy
[email protected]

Abstract. The Newton method is one of the most powerful methods

for the solution of smooth unconstrained optimization problems. It has
local quadratic convergence in a neighborhood of a local minimum where
the Hessian is positive definite and Lipschitz continuous. Several strate-
gies have been proposed in order to achieve global convergence. They are
mainly based either on the modification of the Hessian together with a
line search or on the adoption of a restricted-step strategy. We propose
a globalization technique that combines the Newton and gradient direc-
tions, producing a descent direction on which a backtracking Armijo line
search is performed. Our work is motivated by the effectiveness of gradi-
ent methods using suitable spectral step-length selection rules. We prove
global convergence of the resulting algorithm, and quadratic rate of con-
vergence under suitable second-order optimality conditions. A numerical
comparison with a modified Newton method exploiting Hessian modifi-
cations shows the effectiveness of our approach.

Keywords: Newton method · Gradient method · Global convergence

1 Introduction
We consider the following unconstrained minimization problem
min f (x), x ∈ Rn , (1)
where f is a twice continuously differentiable function. In the following, the
gradient and the Hessian of f are denoted by g(x) and H(x), respectively.
The Newton method for problem (1) generates a sequence {xk } by using the
iterative scheme
xk+1 = xk + dN k , k = 0, 1, 2, ..., (2)
This work was partially supported by Gruppo Nazionale per il Calcolo Scientifico -
Istituto Nazionale di Alta Matematica (GNCS-INdAM). Marco Viola was also sup-
ported by the MOD CELL DEV Project - Programma di finanziamento della Ricerca
di Ateneo, University of Naples Federico II.
c Springer Nature Switzerland AG 2020
Y. D. Sergeyev and D. E. Kvasov (Eds.): NUMTA 2019, LNCS 11973, pp. 177–185, 2020.
https://doi.org/10.1007/978-3-030-39081-5_16
178 D. Serafino et al.

where dN
k solves the linear system

Hk d = −gk , (3)
with Hk = H(xk ) and gk = g(xk ). If x∗ is a solution of (1), with H(x∗ )  0 and
H(x) Lipschitz continuous in a neighborhood of x∗ , then the Newton method is
locally quadratically convergent to x∗ [11]. Global convergence holds for convex
problems, provided some form of reduction on the pure Newton step dN k is
allowed (see, e.g., the damped Newton method by Nesterov [16]). For nonconvex
optimization the Newton direction computed through (3) may not be a descent
direction for f at xk . In order to deal with nonconvex problems, Modified Newton
(MN) methods have been proposed. The basic idea is to replace Hk in (3) by
a matrix H  k = Hk + Ek , where Ek is such that H  k is positive definite [14], so
MN 
that the solution dk of the linear system Hk d = −gk is a descent direction at
xk , and xk+1 can be computed by performing a line search along dM k
N
. Global
convergence can be proved for MN methods, provided that the matrices H  k have
uniformly bounded condition numbers, i.e.,
k ) ≤ ζ
κ(H

for some ζ > 0 independent of k. Effective ways to compute H  k have been

proposed, such as those in [10,12], based on modified Cholesky factorizations
which automatically compute the Cholesky factorization of a positive definite
matrix Hk + Ek , where Ek is a nonnegative diagonal matrix which is zero if Hk
is “sufficiently” positive definite. Another popular choice is to take Ek = λI,
where I is the identity matrix and λ is a suitable positive constant, so that
dM
k
N
= −(Hk + λI)−1 gk . (4)
Let us recall that the Newton method computes xk+1 by using the second-
order Taylor approximation to f at xk :
1
ψk (x) = f (xk ) + gkT (x − xk ) + (x − xk )T Hk (x − xk ). (5)
2
In MN methods, the quadratic model at xk is forced to be convex, in order to
provide a descent direction. Because of that, if f is nonconvex in a neighborhood
of xk , the quadratic model may be inadequate. Starting from this consideration,
we propose a globalization strategy for the Newton method in which the search
direction is a combination of the Newton and Steepest Descent (SD) directions,
such as
−1
dk = dNk − λgk = −(Hk + λI)gk , (6)
where λ is chosen so that dk satisfies some descent condition at xk . The direc-
tions (4) and (6) are somehow related to the Trust Region (TR) approach, also
known as restricted-step approach, which represents the most popular globaliza-
tion strategy for the Newton method. In a TR approach, given xk , the iterate
xk+1 is computed as the (possibly approximate) solution of the problem
min {ψk (x) : x − xk  ≤ Δk } , (7)
 A Gradient-Based Globalization Strategy for the Newton Method 179

where Δk is the radius of the ball where the quadratic model is trusted to
be a reliable approximation of f (x). This radius is updated at each iteration,
according to some rule. Moré and Sorensen (see [14] and references therein) show
k solves (7) if and only if dk = x
that x k − xk satisfies

(Hk + λI) d = −gk ,

(8)
λ (Δk − d) = 0,

for some λ > 0 such that Hk +λI is positive semidefinite; therefore, the MN direc-
tion (4) can be seen as a TR direction. A search direction of the form (6) reminds
us of the so-called dogleg method, and more generally of the two-dimensional
subspace minimization, which are designed to give an approximate solution to
the TR subproblem (7). These strategies make a search in the two-dimensional
space spanned by the steepest descent direction and the Newton one.
A search direction of the form (6) is a way to combine a locally superlinearly
convergent strategy (Newton) with a globally convergent strategy (SD) in such
a way that the latter brings the iterates close to a solution, namely in a so-called
basin of attraction of the Newton method. The effectiveness of the approach
obviously relies on a suitable choice of λ in (6). Although appearing a quite
natural strategy, we found the use of the SD direction in globalizing Newton
or quasi-Newton methods (see, e.g., [13] and [3, Section 1.4.4]) to be much less
popular than we expected. This is probably due to lack of confidence in the SD
methods, which have long been considered rather ineffective because of their
slow convergence rate and their oscillatory behavior. However, starting from the
publication of the Barzilai and Borwein (BB) method [2], it has become more and
more evident that suitable choices of the step length in gradient methods may
lead to effective algorithms [4,6,7], which have also shown good performance in
solving problems arising in several application fields [1,5,9,19].
The remainder of this paper is organized as follows. In Sect. 2 we describe our
globalization strategy, proving global convergence of the corresponding Newton
method. In Sect. 3 we report results of numerical experiments carried out with
our algorithm, including a comparison with an MN method. We provide some
conclusions in Sect. 4.

2 A Globalized Newton Method

A line-search method for solving (1) uses the iterative scheme

xk+1 = xk + αk dk , k = 0, 1, 2, ...,

where dk is the search direction and αk > 0 is the step length. Decreasing line-
search methods, i.e., such that {f (xk )} is a strictly decreasing sequence, require
dk to be a direction of strict descent, i.e., a direction such that

dTk gk < 0. (9)

180 D. Serafino et al.

Algorithm 1. NSD method

1: Let x0 ∈ Rn , ε ∈ (0, 1), σ ∈ (0, 1/2) and ρ ∈ (0, 1).
2: for k = 0, 1, 2, . . . do
3: k = −gk /gk ;
g
4: if Hk is nonsingular then
−1
5: dNk = −Hk gk ;
6: 
dk = dk /dN
N N
k ;  
7: βk = ρh , where h = min s ∈ N0 : cosρs d Nk + (1 − ρ )
s
gk , −gk  > ε ;
8: else
9: βk = 0;
10: end if
11: N
d k = βk d k + (1 − βk )gk ;
12: αk = βk dN k  + (1 − βk )ξk gk  for a suitably chosen ξk > 0;
13: select αk satisfying (11) by a backtracking procedure starting from αk ;
14: xk+1 = xk + αk dk ;
15: end for

The convergence of the line-search method depends on the choice of dk and

αk . The descent condition (9) is not sufficient to guarantee convergence, and a
key requirement for dk is to satisfy the following angle criterion:

−gkT dk
cos−gk , dk  = >ε (10)
gk dk 
for some ε > 0 independent of k. About the step length, the Armijo condition

f (xk + αk dk ) − f (xk ) ≤ σαk dTk gk , σ ∈ (0, 1), (11)

is a common choice. We note that (11) does not prevent the method from taking
too small steps. Such a drawback can be overcome if a backtracking line-search
procedure is adopted to choose αk (see [17, page 37]).
In our globalization strategy for the Newton method, the search direction is
defined as a convex combination of the normalized Newton and steepest descent
directions. A general description of our approach, referred to as NSD method,
is given in Algorithm 1. We note that if Hk is singular βk is set to 0 and dk to
k , i.e., the iteration becomes a gradient step. The parameter ξk is chosen using
g
a spectral step-length selection rule for the gradient method [6].
The following theorem shows that the limit points of the sequence generated
by the NSD method are stationary.
Theorem 1. Let f ∈ C 2 (Rn ). Then the NSD method is well defined and, for
any choice of x0 , every limit point of the sequence {xk } is a stationary point.

Proof. We note that the strategy used to choose βk in Algorithm 1 ensures

that a direction dk satisfying condition (10) can be found. Indeed, if Hk is
singular, then (10) trivially holds. If Hk is nonsingular and d  N satisfies (10),
k
then βk = ρ0 = 1 and dk = d  N ; otherwise, since cos
gk , −gk  = 1 and dk → g k
k
 A Gradient-Based Globalization Strategy for the Newton Method 181

as s → ∞, a value of βk such that (10) holds will be computed in a finite

number of steps. Since αk is obtained through a backtracking strategy to satisfy
the Armijo rule (11), Proposition 1.2.1 in [3] can be applied to conclude the
proof.

It is well known that Newton-type methods are superlinearly convergent

when started from an initial point sufficiently close to a local minimum. One
would hopefully expect a globalization strategy to be able to preserve the con-
vergence rate of the pure Newton method. This is the case of the NSD method,
as shown by the following theorem.

Theorem 2. Let f ∈ C 2 (Rn ) and let {xk } be the sequence generated by the
NSD method. Suppose that there exists a limit point x̂ of {xk } where H(x̂)  0
and that H(x) is Lipschitz continuous in a neighborhood of x̂. Then {xk } con-
verges to x̂ provided that the value of ε used in Algorithm 1 is sufficiently small.
Furthermore, the rate of convergence is quadratic.

Proof. Let K ⊂ N be the index set of the subsequence of {xk } converging to x̂

and let λ1 and λn be the smallest and the largest eigenvalue of H(x̂), respec-
tively. If we choose μ1 , μ2 > 0 such that 0 < μ1 < λ1 ≤ λn < μ2 , the continuity
of H(x) implies that there exists a scalar δ > 0 such that H(x)  0 and its
eigenvalues lie in (μ1 , μ2 ) for each x ∈ Bδ (x̂) = {x ∈ Rn : x − x̂ < δ}. With-
out loss of generality, we can choose δ such that, starting from a point in Bδ (x̂),
the iterates of the Newton method remain in Bδ (x̂) (see [14, Theorem 2.3]). We
can also assume that H(x) is Lipschitz continuous Bδ (x̂). Let kδ ∈ K be such
that xkδ ∈ Bδ (x̂). We have that

gkTδ Hk−1 gkδ μ1

cosdN
kδ , −gkδ  =
δ
> .
Hk−1
δ
gkδ gkδ  μ2
Therefore, if ε < μμ12 , the backtracking strategy at line 7 of Algorithm 1 selects
βkδ = 1 and the (kδ + 1)-st iterate is obtained by a line search over the Newton
direction. By reasoning as in the proof of part (b) in [18, Theorem 1.4.9], we have
that if kδ is large enough, then gkδ  is small and the step length αk = 1 satisfies
the Armijo condition (11), because H(x) is Lipschitz continuous in Bδ (x̂) and
σ < 12 . By the choice of δ we have that xkδ +1 ∈ Bδ (x̂). Then Algorithm 1 becomes
the Newton method and the sequence {xk } converges to x̂ with quadratic rate.

3 Numerical Experiments
We developed a MATLAB implementation of the NSD method and compared
it with a MATLAB implementation of the MN method available from https://
github.com/hrfang/mchol, which exploits the modified Cholesky factorizations
described in [10]. In the NSD method, we set ξk = max{ξkBB2 , ν}, where ξkBB2
is computed by using the Barzilai-Borwein step-length selection rule defined
182 D. Serafino et al.

Table 1. Performance of the NSD and MN methods in the solution of the test problems
(the mark “—” indicates that the required accuracy has not been satisfied within 1000
iterations). Note that we set to 0 all the values below the machine epsilon.

prob NSD MN
fval # its # evals fval # its # evals
1 0.0000 e+00 20 49 0.0000 e+00 18 96
3 1.1279 e−08 3 5 1.1279 e−08 3 5
4 3.4891 e−10 438 3859 3.7800 e−05 — 2007
6 0.0000 e+00 11 21 0.0000 e+00 11 21
7 6.9588 e−02 9 17 6.9588 e−02 9 17
8 2.2500 e−05 37 83 2.2500 e−05 37 83
9 9.3763 e−06 140 354 9.3763 e−06 138 344
10 0.0000 e+00 6 12 2.4571 e−13 967 1933
11 8.5822 e+04 10 19 8.5822 e+04 10 19
12 0.0000 e+00 8 16 0.0000 e+00 23 58
13 3.0282 e−04 16 35 3.0282 e−04 9 37
14 0.0000 e+00 25 57 0.0000 e+00 25 57
15 1.7085 e−10 18 35 1.7085 e−10 18 35
16 0.0000 e+00 10 22 4.5201 e−01 — 2137
17 0.0000 e+00 456 1022 0.0000 e+00 59 169
18 0.0000 e+00 7 14 0.0000 e+00 7 14
19 0.0000 e+00 33 78 0.0000 e+00 33 78
20 -4.0000 e+00 — 21981 −4.0000 e+00 — 21981
21 0.0000 e+00 2 3 0.0000 e+00 2 3
22 0.0000 e+00 2 3 0.0000 e+00 2 3
23 6.9492 e−15 29 67 6.9492 e−15 29 67
25 3.0227 e−10 18 35 3.0227 e−10 18 35
27 4.8254 e−01 8 15 5.0000 e−01 2 20
28 3.4102 e+00 4 8 8.8077 e+00 9 54
30 3.9789 e−01 5 9 3.9789 e−01 5 29
31 −1.0153 e+01 10 20 −1.0153 e+01 13 26
32 −1.0402 e+01 9 28 −4.9728 e−06 10 39
33 −3.8351 e+00 9 20 −1.9733 e+00 8 20
34 −2.1546 e−01 7 14 −2.1546 e−01 7 14
35 −1.3803 e+01 6 12 −1.4427 e+01 5 9
37 −1.0000 e+00 11 23 0.0000 e+00 2 23
38 2.2875 e+00 6 13 2.2875 e+00 6 13
39 2.1831 e−01 7 16 2.1831 e−01 6 21
40 5.1001 e−01 6 13 −4.6516 e−01 7 33
41 0.0000 e+00 40 101 0.0000 e+00 28 78
43 3.7532 e−16 6 13 0.0000 e+00 20 79
 A Gradient-Based Globalization Strategy for the Newton Method 183

# iterations # obj fun evaluations

1 1

0.8 0.8

0.6 0.6
( )

( )
0.4 0.4

0.2 0.2
NSD NSD
MN MN
0 0
1 1.5 2 2.5 3 1 1.5 2 2.5 3

Fig. 1. Performance profiles of the NSD and MN methods in the solution of the test
problems with 10 starting points. We consider the number of iterations (left) and the
number of objective function evaluations (right).

in [2, equation (5)] and ν = 10−5 . Moreover, we set ρ = 0.9 and ε = 10−6 . In
the MN method, we chose the modified Cholesky factorization named GMW-
II, which minimizes the 2-norm of the matrix Ek in the modified Hessian. Both
methods used an Armijo backtracking line search with quadratic and cubic inter-
polation (see [17, Section 3.5]) and σ = 10−4 . The methods were stopped as soon
as gk  < 10−6 or a maximum number of 1000 iterations was achieved.
The two algorithms were run on 36 problems from the collection
available at https://people.sc.fsu.edu/∼jburkardt/m src/test opt/test opt.html,
which includes MATLAB implementations of the well-known Moré-Garbow-
Hillstrom benchmark problems [15] and other unconstrained optimization prob-
lems. All the experiments were carried out using MATLAB R2018b on a 64-bit
Intel i7-6500 (3.10 GHz) processor, with 12 GB of RAM and 4 MB of cache
memory and the Windows 10 operating system.
We first tested the two methods using the starting points provided with
the test problems. For each method and each problem, we report in Table 1
the objective function value at the computed solution (fval), the number of
iterations (# its) and the number of objective function evaluations (# evals).
These results show that the NSD method was generally able to obtain objective
function values at the computed solutions smaller than or equal to those reached
by the MN method. Furthermore, we verified that in most of the cases where
the two algorithms computed the same solution, NSD performed a number of
iterations (each requiring the solution of a linear system) that is comparable with
the number of iterations of the MN method. However, NSD generally required
a smaller number of objective function evaluations, indicating a smaller number
of line-search steps per iteration.
In order to further assess the performance of NSD, we ran tests using mul-
tiple starting points. Besides the original starting point x0 , for each problem
we considered 9 more points obtainedby adding to each entry (x0 )i a value γi
randomly chosen in −10−2 ai , 10−2 ai , where ai = |(x0 )i |. This allowed us to
184 D. Serafino et al.

compare the two algorithms on a total of 360 problem instances. Again NSD was
generally able to achieve objective function values smaller than or comparable
with those computed by MN. The results of these tests are summarized by the
performance profiles [8] in Fig. 1, which compare the two algorithms in terms
of iterations and objective function evaluations on the 273 instances (76% of
the problems) where they computed the same optimal solution up to the third
significant digit. We see that NSD was more efficient in terms of both iterations
(i.e., number of linear system solved) and objective function evaluations (i.e.,
line-search steps), and appeared to be more robust than MN.

4 Conclusions and Future Work

We proposed a globally convergent algorithm for unconstrained minimization, in
which the Newton and the steepest descent directions are combined to produce a
descent direction. Although simple and straightforward, this idea has been little
explored to globalize the Newton method. Preliminary computational exper-
iments show that the proposed strategy achieves promising results compared
with a well-established modified Newton approach. Future work will include
extensive testing of the proposed globalization strategy and its application to
quasi-Newton methods.

References
1. Antonelli, L., De Simone, V., di Serafino, D.: On the application of the spectral pro-
jected gradient method in image segmentation. J. Math. Imaging and Vis. 54(1),
106–116 (2016)
2. Barzilai, J., Borwein, J.M.: Two-point step size gradient methods. IMA J. Numer.
Anal. 8(1), 141–148 (1988)
3. Bertsekas, D.P.: Nonlinear Programming, 3rd edn. Athena Scientific, Belmont
(2016)
4. Crisci, S., Ruggiero, V., Zanni, L.: Steplength selection in gradient projection meth-
ods for box-constrained quadratic programs. Appl. Math. Comput. 356, 312–327
(2019)
5. De Asmundis, R., di Serafino, D., Landi, G.: On the regularizing behavior of the
SDA and SDC gradient methods in the solution of linear ill-posed problems. J.
Comput. Appl. Math. 302, 81–93 (2016)
6. di Serafino, D., Ruggiero, V., Toraldo, G., Zanni, L.: On the steplength selection
in gradient methods for unconstrained optimization. Appl. Math. Comput. 318,
176–195 (2018)
7. di Serafino, D., Toraldo, G., Viola, M., Barlow, J.: A two-phase gradient method
for quadratic programming problems with a single linear constraint and bounds
on the variables. SIAM J. Optim. 28(4), 2809–2838 (2018)
8. Dolan, E.D., Moré, J.J.: Benchmarking optimization software with performance
profiles. Math. Program. Ser. B 91(2), 201–213 (2002)
9. Dostál, Z., Toraldo, G., Viola, M., Vlach, O.: Proportionality-based gradient meth-
ods with applications in contact mechanics. In: Kozubek, T., et al. (eds.) HPCSE
2017. LNCS, vol. 11087, pp. 47–58. Springer, Cham (2018). https://doi.org/10.
1007/978-3-319-97136-0 4
 A Gradient-Based Globalization Strategy for the Newton Method 185

10. Fang, H., O’Leary, D.P.: Modified Cholesky algorithms: a catalog with new
approaches. Math. Program. 115(2), 319–349 (2008)
11. Fletcher, R.: Practical Methods of Optimization. Wiley, New York (2013)
12. Gill, P., Murray, W.: Newton-type methods for unconstrained and linearly con-
strained optimization. Math. Program. 7(1), 311–350 (1974)
13. Han, L., Neumann, M.: Combining quasi-Newton and Cauchy directions. Int. J.
Appl. Math. 22(2), 167–191 (2003)
14. Moré, J.J., Sorensen, D.C.: Newton’s method. In: Golub, G. (ed.) Studies in Numer-
ical Analysis, pp. 29–82. The Mathematical Association of America, Providence
(1984)
15. Moré, J.J., Garbow, B.S., Hillstrom, K.E.: Algorithm 566: FORTRAN subrou-
tines for testing unconstrained optimization software [C5], [E4]. ACM Trans. Math.
Softw. 7(1), 136–140 (1981)
16. Nesterov, Y.: Introductory Lectures on Convex Optimization: A Basic Course, vol.
87. Springer, Boston (2013). https://doi.org/10.1007/978-1-4419-8853-9
17. Nocedal, J., Wright, S.: Numerical Optimization. Springer Series in Operations
Research, 2nd edn. Springer, New York (2006). https://doi.org/10.1007/978-0-387-
40065-5
18. Polak, E.: Optimization: Algorithms and Consistent Approximations, vol. 124.
Springer, New York (2012). https://doi.org/10.1007/978-1-4612-0663-7
19. Zanella, R., Boccacci, P., Zanni, L., Bertero, M.: Efficient gradient projection
methods for edge-preserving removal of Poisson noise. Inverse Prob. 25(4), 045010
(2009)
 On the Steplength Selection in Stochastic
Gradient Methods

Giorgia Franchini1,2,3(B) , Valeria Ruggiero2,3 , and Luca Zanni1,3

1
Department of Physics, Informatics and Mathematics, University of Modena
and Reggio Emilia, Modena, Italy
{giorgia.franchini,luca.zanni}@unimore.it
2
Department of Mathematics and Computer Science, University of Ferrara,
Ferrara, Italy
[email protected]
3
INdAM Research Group GNCS, Rome, Italy

Abstract. This paper deals with the steplength selection in stochas-

tic gradient methods for large scale optimization problems arising in
machine learning. We introduce an adaptive steplength selection derived
by tailoring a limited memory steplength rule, recently developed in the
deterministic context, to the stochastic gradient approach. The proposed
steplength rule provides values within an interval, whose bounds need to
be prefixed by the user. A suitable choice of the interval bounds allows to
perform similarly to the standard stochastic gradient method equipped
with the best-tuned steplength. Since the setting of the bounds slightly
affects the performance, the new rule makes the tuning of the param-
eters less expensive with respect to the choice of the optimal prefixed
steplength in the standard stochastic gradient method. We evaluate the
behaviour of the proposed steplength selection in training binary classi-
fiers on well known data sets and by using different loss functions.

Keywords: Stochastic gradient methods · Steplength selection rule ·

Ritz-like values · Machine learning

1 Introduction
One of the pillars of machine learning is the development of optimization meth-
ods for the numerical computation of parameters of a system designed to make
decisions based on yet unseen data. Supported on currently available data or
examples, these parameters are chosen to be optimal with respect to a loss func-
tion [3], measuring some cost associated with the prediction of an event. The
problem we consider is the unconstrained minimization of the form

min F (x) ≡ E[f (x, ξ)] (1)

xRd

where ξ is a multi-value random variable and f represents the cost function.

c Springer Nature Switzerland AG 2020

Y. D. Sergeyev and D. E. Kvasov (Eds.): NUMTA 2019, LNCS 11973, pp. 186–197, 2020.
https://doi.org/10.1007/978-3-030-39081-5_17
 On the Steplength Selection in Stochastic Gradient Methods 187

While it may be desirable to minimize (1), such a goal is untenable when

one does not have complete information about the probability distribution of ξ.
In practice, we seek the solution of a problem that involves an estimate of the
objective function F (x). In particular, we minimize the sum of cost functions
depending on a finite training set, composed by sample data ξi , i ∈ {1 . . . n}:
1 1
n n
min Fn (x) ≡ f (x, ξi ) ≡ fi (x), (2)
x∈Rd n i=1 n i=1

where n is the number of samples and each fi (x) ≡ f (x, ξi ) denotes the cost
function related to the instance ξi of the training set elements. For very large
training set, the computation of ∇Fn (x) is too expansive and it results inappo-
site for an online training with a growing amount of samples; then, exploiting
the redundancy of the data, Stochastic Gradient (SG) method and its variants,
requiring only the gradient of one or few terms of Fn (x) at each iteration, have
been chosen as the main approaches for addressing the problem (2). In Algo-
rithm 1 we resume the main steps of a generalized SG method [3]. Each k-th
iteration of SG requires a realization ξk of the random variable ξ, a strategy to
devise a stochastic gradient vector g(xk , ξk ) ∈ Rd at the current iterate xk and
an updating rule for the steplength or scalar learning rate ηk > 0.

Algorithm 1. Stochastic Gradient (SG) method

1: Choose an initial iterate x1 .
2: for k = 1, 2, . . . do
3: Generate a realization of the random variable ξk .
4: Compute a stochastic vector g(xk , ξk ).
5: Choose a learning rate ηk > 0.
6: Set the new iterate as xk+1 ← xk − ηk g(xk , ξk ).
7: end for

In particular, we point out two different strategies for the choices of ξk and
g(xk , ξk ):
– a realization of ξk may be given by the choice of a single training sample,
or, in other words, a random index ik is chosen from {1, 2, . . . , n} and the
stochastic gradient is defined as
g(xk , ξk ) = ∇fik (xk ), (3)
where ∇fik (xk ) denotes the gradient of the ik -th component function of (2)
at xk ;
– the random variable ξk may represent a small subset Sk ⊂ {1, ..., n} of sam-
ples, randomly chosen at each iteration, so that the stochastic gradient is
defined as
1 
g(xk , ξk ) = ∇fi (xk ) (4)
|Sk |
i∈Sk

where |Sk | denotes the number of elements of the set Sk .

188 G. Franchini et al.

The described approaches give rise to the standard simple SG and its mini-batch
version, respectively.
The convergence results of Algorithm 1 equipped with the rule (3) or (4)
apply to both the objective functions F (x) and Fn (x). In [3] the theoretical
properties of generalized SG schemes are proved for strongly convex and non-
convex loss functions. In particular, we refer to the case of a strongly convex loss
function, such as the ones used in the numerical results section. We recall that a
function F is said to be strongly convex with parameter c when for all x, y and
a ∈ [0, 1] we have
1
F (ax + (1 − a)y) ≤ aF (x) + (1 − a)F (y) − ca(1 − a)x − y2 ,
2
where c is a positive scalar.
Standard assumptions for the analysis of SG methods are that the gradient of
F is L-Lipschitz continuous and the first and second moments of the stochastic
directions {g(xk , ξk )} satisfy the following inequalities:

μG  ∇F (xk ) 22 ≥ ∇F (xk )T Eξk [g(xk , ξk )] ≥ μ  ∇F (ωk ) 22

Vξk [g(xk , ξk )] ≤ M + MV  ∇F (xk ) 22 ,

where μ, μG , M, MV are suitable positive scalars. Thus, when the variable

steplength ηk is in the interval [η min , η max ], with η min > 0 and η max ≤ LM
μ
G
,
the expected optimality gap related to the SG iteration satisfies the following
asymptotic condition [3, Th. 4.6]
k→∞
E[F (xk ) − F∗ ] ≤−−−−→ η max LM
2cμ ,

where F∗ is the required minimum value. This result shows that if the steplength
is sufficiently small, then the expected objective values will converge to a neigh-
borhood of the optimal value. In practice, since the constants related to the
assumptions, such as the Lipschitz parameter of ∇F , or the parameters involved
in the bounds of the moments of the stochastic directions, are unknown and
not easy to approximate, the steplength is selected as a fixed small value η. We
observe that there is no guidance on the specific choice of this value, which,
however, plays a crucial role in the effectiveness of the method. Indeed, a too
small steplength can give rise to a very slow learning process. For this reason,
in some recent papers (see, for example, [8,9]), rules for an adaptive selection of
the steplength have been proposed. In this work, we tailor the limited memory
steplength selection rule proposed in [4] to the SG framework.

2 Selections Based on the Ritz-like Values

In the deterministic framework, a very effective approach for the steplength
selection in the gradient methods is proposed in [4] for unconstrained quadratic
programming problems and then extended to general nonlinear problems. In
 On the Steplength Selection in Stochastic Gradient Methods 189

order to capture some second order information of the considered problem, the
steplengths are defined as the inverse of suitable approximations of the eigenval-
ues of the Hessian matrix, given by its Ritz values. The key point is to obtain
the Ritz values in an inexpensive way.
Let assume we have to solve the quadratic programming problem minx φ(x) ≡
x
1 T
2 Ax − bT x by means of the gradient method. The basic idea in [4] is to divide
the sequence of iterations into groups of mR iterations referred to as sweeps,
where mR is a small positive integer, and, for each sweep, to set the steplengths as
the inverse of some Ritz values of the Hessian matrix A, computed by exploiting
the gradients of the previous sweep. In particular, at the iteration k ≥ mR ,
we denote by G and J the matrices obtained collecting mR gradient vectors
computed at previous iterates and the related steplengths:
⎛ −1 ⎞
ηk−mR
⎜ −1 .. ⎟
⎜−η . ⎟
⎜
G = [gk−mR , . . . , gk−1 ] , J = ⎜ k−m R ⎟,
.. ⎟
⎝ −1 ⎠
. ηk−1
−1
−ηk−1

where we use the notation gi ≡ ∇φ(xi ). In view of the recurrence formula linking
the gradients of φ(x) at two successive iterates
gi = gi−1 − ηi−1 Agi−1 , i > 0,
we can write
AG = [G, gk ]J. (5)
This equation is useful to compute the tridiagonal matrix T obtained from the
application of mR iterations of the Lanczos process to the matrix A, with start-
ing vector q1 = gk−mR / gk−mR ; this procedure generates an orthogonal
matrix Q = [q1 , . . . , qmR ], whose columns are a basis for the Krylov subspace
{gk−mR , Agk−mR , A2 gk−mR , ..., AmR −1 gk−mR }, such that
T = QT AQ. (6)
The steplengths for the next mR gradient iterations are defined as the inverse
of the eigenvalues θi of T , that are the so-called Ritz values:
1
ηk−1+i = , i = 1, . . . , mR . (7)
θi
The explicit computation of the matrix Q can be avoided, by observing that
G = QR, where R is upper triangular, and R can be obtained from the Cholesky
factorization of GT G, that is GT G = RT R. Then the matrix T can be computed
from equation (6) as follows:
T = R−T GT AGR−1 = R−T GT [G, gk ]JR−1 = [R, r]JR−1 , (8)
where the vector r is the solution of the linear system RT r = GT gk .
190 G. Franchini et al.

For a non-quadratic objective function, the recurrence (5) does not hold and
the Eq. (8) provides an Hessenberg matrix; nevertheless, we can compute the
symmetric tridiagonal matrix T by replacing the strictly upper triangle of T by
the transpose of its strictly lower triangle (in Matlab notation T̄ = tril(T ) +
tril(T, −1) ); the eigenvalues θi of T tend to approximate mR eigenvalues of the
Hessian matrix [4,7].
In the stochastic context, we propose to introduce in the SG methods a selec-
tion rule for the steplength based on the just described Ritz-like approach; in this
case, the implementation of this technique involves some important differences.
The main difference is the use of stochastic gradients instead of full gradients in
the construction of the matrix G:

G = [gk−mR (xk−mR , ξk−mR ), . . . , gk−1 (xk−1 , ξk−1 )];

we observe that the stochastic gradients are related to different samples of the
data. By means of this matrix G, the matrix T of (8) can be computed; we
propose to approximate second order information with the eigenvalues of its
symmetric part T̃ = (T + T T )/2. Another key point is that, among the mR
eigenvalues θi of T̃ , only the Ritz values belonging to an appropriate range
[ηmin , ηmax ] have to be considered. As a consequence, the steplengths in a new
sweep are defined in the following way:

1
ηk−i+1 = max min ηmax , , ηmin , i = 1, ..., mR . (9)
θi

Moreover, a further value ηini ∈ [ηmin , ηmax ] is introduced for setting the starting
sweep: ηi = ηini , i = 0, ..., mR − 1. This reference value is also used as steplength
in a recovery procedure, when all the eigenvalues of T̃ are negative and they
have to be discarded.
The proposed steplength approach depends on the chosen interval
[ηmin , ηmax ] and on ηini . However, the effectiveness of the corresponding SG
methods is weakly affected by variations of these parameters. This behaviour
introduces greater flexibility with respect to the choice of a fixed small scalar,
that must be carefully best-tuned. In particular, the numerical results of the
next section highlight that the version of SG equipped with the Ritz-like selec-
tion rule for the steplengths appears to be more robust than that with a constant
steplength and it provides numerical results with a comparable accuracy.

3 Numerical Experiments

In order to evaluate the effectiveness of the proposed steplength rule for SG

methods, we consider the optimization problems arising in training binary clas-
sifiers for two well known data-sets:

– the MNIST data-set of handwritten digits (http://yann.lecun.com/exdb/

mnist), commonly used for testing different systems that process images; the
 On the Steplength Selection in Stochastic Gradient Methods 191

images are in gray-scale [0, 255], in our case normalized in the interval [0, 1],
centered in a box of 28 × 28 pixels; from the whole data-set of 60, 000 images,
11, 800 images were extracted exclusively relating to digits 8 and 9;
– the web data-set w8a downloadable from https://www.csie.ntu.edu.tw/
∼cjlin/libsvmtools/datasets/binary.html, containing 49,749 examples; each
example is described by 300 binary features.
We built linear classifiers corresponding to three different loss functions; in
all cases, a regularization term was added to avoid overfitting. Thus the mini-
mization problem has the form
λ
min Fn (x) + x22 , (10)
x∈Rd 2

where λ > 0 is a regularization parameter. By denoting as ai ∈ Rd and bi ∈

{1, −1} the feature vector and the class label of the i-th sample, respectively,
the loss function Fn (x) assumes one of the following form:
– logistic regression loss:

1
n  T

Fn (x) = log 1 + e(−bi ai x) ;
n i=1

– square loss:
1
n
Fn (x) = (1 − bi aTi x)2 ;
n i=1
– smooth hinge loss:
⎧
n ⎪⎨ 2 − bi ai x, if bi aTi x ≤ 0
1 T
1 
Fn (x) = 2 (1 − bi ai x) , if 0 < bi ai x < 1
1 T 2 T
n i=1 ⎪
⎩
0, if bi aTi x ≥ 1.

We compare the effectiveness of the following schemes:

– simple SG with fixed steplength and SG equipped with the Ritz-like
steplength rule (9), named SG FR;
– SG with a fixed mini-batch size in the version with fixed steplength, denoted
by SG mini-batch, and the one equipped with the Ritz-like steplength rule
(9), named SG FR mini-batch.

3.1 Numerical Results

In all the numerical experiments we use the following setting:
– the regularization parameter λ is equal to 10−8 ;
– the size of the mini-batch is |S| = 20;
– in the FR methods, the length of the sweep is mR = 3;
192 G. Franchini et al.

– each method is stopped after 15 epochs, i.e., after a time interval equivalent to
15 evaluations of a full gradient of Fn ; in this way we compare the behaviour
of the methods in a time equivalent to 15 iterations of a full gradient method
applied to Fn (x).
In the following tables we report the results obtained by the considered meth-
ods on the MNIST and w8a data-sets, by using the three loss functions (logistic
regression, square and smooth hinge functions). For any test problem, we per-
form 10 runs of each method and we report the following results:
– the average value of the optimality gap Fn (x) − F∗ , where x is the iterate
obtained at the end of the 15 epochs and F∗ is an estimate of the optimal
objective value obtained by a full gradient method with a large number of
iterations;
– the related average accuracy A(x) with respect to the training set employed
for training the binary classifier, that is the percentage of well-classified exam-
ples.
We carried out the 10 different simulations with the same parameters, but leaving
the possibility to the random number generator to vary. Indeed, due to the
stochastic nature of the methods, the average values in different simulations
provide more reliable results.
First of all, we describe the numerical results obtained by the different ver-
sions of SG related to the best-tuned setting of the parameters. In Table 1, we
report the value of the fixed steplength ηOP T corresponding to the best perfor-
mance of SG in 15 epochs. The steplength of SG mini-batch is set as |S| · ηOP T .

Table 1. Values of the best-tuned steplength ηOP T in 15 epochs for the SG method
in the case of the two data-sets and the three loss functions.

MNIST w8a
Loss Logistic Square Smooth Logistic Square Smooth
function regression hinge regression hinge
ηOP T 10−2 10−4 10−2 10−1 10−3 5 10−2

In the FR case, the following setting provides the best results:

– in SG FR, for both MNIST and w8a, we set [ηmin , ηmax ] = [10−4 ηOP T ,
5ηOP T ] and ηini = 0.1 ηOP T ;
– in SG FR mini-batch, for both MNIST and w8a, we set [ηmin , ηmax ] =
[10−8 ηOP T , 50ηOP T ] and ηini = 0.1 ηOP T . We can observe that, in the mini-
batch version, the method allows to choose the steplengths within a greater
interval, showing more robustness.

In Tables 2, 3 and 4, we show the results obtained for the logistic regression,
square and smooth hinge loss functions, respectively.
 On the Steplength Selection in Stochastic Gradient Methods 193

Table 2. Numerical results of the considered methods with Fn (x) given by the logistic
regression after 15 epochs.

MNIST w8a
Method Fn (x) − F∗ A(x) Fn (x) − F∗ A(x)
SG 0.0107 0.987 0.00419 0.903
SG FR 0.0210 0.984 0.0251 0.899
SG mini-batch 0.0103 0.987 0.0141 0.901
SG FR mini-batch 0.0129 0.986 0.00441 0.903

Table 3. Numerical results of the considered methods with Fn (x) given by the square
loss after 15 epochs.

MNIST w8a
Method Fn (x) − F∗ A(x) Fn (x) − F∗ A(x)
SG 0.00557 0.978 0.00229 0.889
SG FR 0.00729 0.978 0.00417 0.888
SG mini-batch 0.00585 0.977 0.00145 0.890
SG FR mini-batch 0.00647 0.977 0.00871 0.887

We observe that the results obtained with the FR adaptive steplength rule are
well comparable with the ones obtained with the standard SG method equipped
with the best-tuned steplength.
However, the considerable numerical experimentation carried out to obtain
the above tables allows to remark that the optimal steplength search process for
the SG method is often computationally long and expensive, since it requires
a trial-and-error approach, while the FR adaptive rule appears to be weakly
affected by the values ηmin and ηmax defining its working interval. In the follow-
ing figures, we highlight the differences between the two approaches with respect
to different settings of parameters in terms of the behavior of the optimality gap
Fn (x) − F∗ in 15 epochs. In particular, in Figs. 1 and 2, we show the behaviour

Table 4. Numerical results of the considered methods with Fn (x) given by the smooth
hinge loss after 15 epochs.

MNIST w8a
Method Fn (x) − F∗ A(x) Fn (x) − F∗ A(x)
SG 0.00607 0.989 0.00136 0.904
SG FR 0.00847 0.987 0.0127 0.898
SG mini-batch 0.00754 0.987 0.000656 0.904
SG FR mini-batch 0.00793 0.987 0.00361 0.902
194 G. Franchini et al.

Fig. 1. Behaviour of SG and SG FR in 15 epochs on the MNIST data-set; test problem

with smooth hinge loss function.

Fig. 2. Behaviour of SG mini-batch and SG FR mini-batch in 15 epochs on the MNIST

data-set; test problem with smooth hinge loss function.

of the considered methods when the smooth hinge is used as loss function on
the MNIST data-set, while the results obtained with the logistic regression loss
function on the w8a data-set are reported in Figs. 3 and 4. On the left panels
of the figures, the behaviour of SG and SG mini-batch methods equipped with
different values of the steplength η in 15 epochs is shown; the right panels report
the results obtained when the methods SG FR and SG FR mini-batch are exe-
cuted with the same values for ηini and ηmin and different values of ηmax . We
observe that the adaptive steplength rule FR seems to be slightly dependent
on the value of ηmax , making the choice of a suitable value of this parameter a
less difficult task with respect to the setting of η in the SG and SG mini-batch
schemes.
 On the Steplength Selection in Stochastic Gradient Methods 195

Fig. 3. Behaviour of SG and SG FR in 15 epochs on the w8a data-set; test problem

with logistic regression loss function.

Fig. 4. Behaviour of SG mini-batch and SG FR mini-batch in 15 epochs on the w8a

data-set; test problem with logistic regression loss function.

Finally, the last two figures compare the accuracy obtained in 15 epochs
by the two approaches, SG with constant steplength and SG equipped by FR
adaptive steplength rule, when the parameters η and ηmax are not set at the best-
tuned values, as in the experiments related to the previous tables. In particular,
in Figs. 5 and 6, we show a comparison between SG or SG mini-batch with
a prefixed non-optimal steplength η or |S|η respectively, and the corresponding
versions equipped with the adaptive steplength rule based on the Ritz-like values.
In these figures, for the FR case we set ηmin = 10−4 η, ηmax = 5η, ηini = 10−1 η
and for FR mini-batch case we set ηmin = 10−8 η, ηmax = 50η, ηini = 10−1 η. In
practice, in order to obtain the numerical results shown in Fig. 5 (w8a data-set
and logistic regression loss function), the best-tuned values have been multiplied
by a factor 10−3 , while in the case of Fig. 6 (MNIST data-set and smooth hinge
loss function), the parameters are set equal to 10−2 times the best-tuned values.
196 G. Franchini et al.

Fig. 5. Comparison between SG with respect to SG FR (on the left) and between SG
mini-batch and SG FR mini-batch (on the right) in 15 epochs on the w8a data-set;
test problem with logistic regression loss function.

Fig. 6. Comparison between SG with respect to SG FR (on the left) and between SG
mini-batch and SG FR mini-batch (on the right) in 15 epochs on the MNIST data-set;
test problem with smooth hinge loss function.

As can be seen in the two figures, the selected steplength values guarantee
the convergence of SG and SG mini-batch, but they are too small and produce a
slow descent of the optimality gap; on the other hand, the FR approach appears
less dependent on an optimal setting of the parameters and it enables us to
obtain smaller optimality gap values after the same number of epochs exploited
by SG.

4 Conclusions
In this work we propose to tailor the steplength selection rule based on the Ritz-
like values, used successfully in the deterministic gradient schemes, to the SG
 On the Steplength Selection in Stochastic Gradient Methods 197

methods. The numerical experimentation highlights that this strategy enables

to achieve the performance of SG with fixed best-tuned steplength. Although,
also in this case, it is necessary to carefully select a thresholding range for the
steplengths, the proposed strategy appears slightly dependent on the param-
eters defining the bounds for the steplengths, making the parameters setting
less expensive with respect to the SG framework. Furthermore, we remark that
the mini-batch version of the proposed approach, allows more flexibility, since
it needs of less restricted bounds for the steplengths range. In conclusion, the
proposed technique provides a guidance on the learning rate selection and it
allows to perform similarly to the SG approach equipped with the best-tuned
steplength.
Future works will involve the introduction of the proposed technique in the
variance reductions methods and its validation on other loss functions. Following
the suggestions in [1,2,5,6], a very interesting analysis will concern the possibility
of combining the proposed steplength selection rule with inexact line search
techniques used in SG methods.

References
1. Bellavia, S., Krejic, N., Krklec Jerinkic, N.: Subsampled inexact Newton methods
for minimizing large sums of convex functions. arXiv:1811.05730 (2018)
2. Bollapragada, R., Byrd, R., Nocedal, J.: Adaptive sampling strategies for stochastic
optimization. SIAM J. Optim. 28(4), 3312–3343 (2018)
3. Bottou, L., Curtis, F.E., Nocedal, J.: Optimization methods for large-scale machine
learning. SIAM Rev. 60(2), 223–311 (2018)
4. Fletcher, R.: A limited memory steepest descent method. Math. Program. Ser. A
135, 413–436 (2012)
5. Krejic, N., Krklec Jerinki, N.: Nonmonotone line search methods with variable sam-
ple size. Numer. Algorithms 68, 711–739 (2015)
6. Paquette, C., Scheinberg, K.: A stochastic line search method with convergence rate
analysis. arXiv:1807.07994v1 (2018)
7. di Serafino, D., Ruggiero, V., Toraldo, G., Zanni, L.: On the steplength selection
in gradient methods for unconstrained optimization. Appl. Math. Comput. 318,
176–195 (2018)
8. Sopyla, K., Drozda, P.: SGD with BB update step for SVM. Inf. Sci. Inform. Com-
put. Sci. Intell. Syst. Appl. 316(C), 218–233 (2015)
9. Tan, C., Ma, S., Dai, Y., Qian, Y.: BB step size for SGD. In: Lee, D., Sugiyama,
M., Luxburg, U., Guyon, I., Garnett, R. (eds.) Advances in Neural Information
Processing Systems (NIPS 2016), vol. 29 (2016)
 Efficient Block Coordinate Methods
for Blind Cauchy Denoising

Simone Rebegoldi1(B) , Silvia Bonettini2 , and Marco Prato2

1
Dipartimento di Scienze Biomediche, Metaboliche e Neuroscienze,
Università di Modena e Reggio Emilia, Via Campi 287, 41125 Modena, Italy
[email protected]
2
Dipartimento di Scienze Fisiche, Informatiche e Matematiche,
Università di Modena e Reggio Emilia, Via Campi 213/b, 41125 Modena, Italy
{silvia.bonettini,marco.prato}@unimore.it

Abstract. This paper deals with the problem of image blind decon-
volution in presence of Cauchy noise, a type of non-Gaussian, impul-
sive degradation which frequently appears in engineering and biomedical
applications. We consider a regularized version of the corresponding data
fidelity function, by adding the total variation regularizer on the image
and a Tikhonov term on the point spread function (PSF). The result-
ing objective function is nonconvex with respect to both the image and
PSF block, which leads to the presence of several uninteresting local
minima. We propose to tackle such challenging problem by means of a
block coordinate linesearch based forward backward algorithm suited for
nonsmooth nonconvex optimization. The proposed method allows per-
forming multiple forward-backward steps on each block of variables, as
well as adopting variable steplengths and scaling matrices to accelerate
the progress towards a stationary point. The convergence of the scheme is
guaranteed by imposing a linesearch procedure at each inner step of the
algorithm. We provide some practical sound rules to adaptively choose
both the variable metric parameters and the number of inner iterations
on each block. Numerical experiments show how the proposed approach
delivers better performances in terms of efficiency and accuracy if com-
pared to a more standard block coordinate strategy.

Keywords: Blind deconvolution · Cauchy noise · Nonconvex

optimization

1 Problem Formulation
In image restoration, the vast majority of the literature relies on the assump-
tion that the observed data is corrupted by additive white Gaussian noise [9,24].
Such assumption, combined with a maximum a posteriori approach, leads to
the minimization of a penalized least squares functional, which is usually con-
vex and hence easy to minimize. However, in several real applications, the most
adequate noise model may not be the Gaussian one. Typical examples include
c Springer Nature Switzerland AG 2020
Y. D. Sergeyev and D. E. Kvasov (Eds.): NUMTA 2019, LNCS 11973, pp. 198–211, 2020.
https://doi.org/10.1007/978-3-030-39081-5_18
 Efficient Block Coordinate Methods for Blind Cauchy Denoising 199

Poisson noise in photon counting imaging [4], impulse noise due to analogue-to-
digital conversion errors [19] or multiplicative noise in radar imagery [1].
Recently, much attention has been dedicated to Cauchy noise [14,16,18,25], a
kind of impulsive noise which corrupts synthetic aperture radar (SAR) images,
underwater acoustic signals, low-frequency atmospheric signals and multiple-
access in wireless communication systems [3,20,25]. The probability density
function associated to the “zero-centered” Cauchy distribution is given by
1 γ
p(v) = (1)
π γ + v2
2

where γ > 0 is a scale parameter. We observe that (1) is even, bell-shaped and
exhibits a thicker tail than the Gaussian bell curve, which tells us that an image
corrupted by Cauchy noise is more likely to contain major noise peaks than the
same image corrupted by white Gaussian noise.
Assume that a digital image g ∈ Rn has been acquired according to the
model g = ω ⊗ x + v, where x ∈ Rn is the true image, ω ∈ Rn is the point spread
function (PSF) describing the blurring process, v ∈ Rn is the random noise vector
following a Cauchy distribution and ⊗ denotes the convolution operation under
periodic boundary conditions. Then, according to the maximum a posteriori
approach [25], it is possible to derive a variational model to restore the corrupted
image, where the corresponding fit-to-data term is given by
n
1  2 
JC (x; ω) = log γ + ((ω ⊗ x)i − gi )2 . (2)
2 i=1

The image restoration is then performed by minimizing, with respect to the

unknown x, the sum of JC plus some appropriate convex regularizers, such as
the total variation functional [24]. Note that the data fidelity term (2) is non-
convex, thus several local minima may exist and any optimization method may
get stuck in one of them. The authors in [25] address the nonconvexity of the
problem by adding a quadratic penalty term to the objective function, whereas
the works [7,18] directly address the problem by means of first order splitting
algorithms, for which convergence guarantees hold even in the nonconvex case.
In this paper, we extend the previously described model to blind deconvo-
lution, namely to the restoration of images corrupted by Cauchy noise when
the blur is unknown. Blind deconvolution is a severely ill-posed problem which
has been widely treated in the literature, especially in the case of Gaussian and
Poisson noise [2,17,21–23]. However, to the best of our knowledge, Cauchy noise
has not yet been considered in the blind deconvolution framework. In this case,
the problem amounts to minimizing the data fidelity (2) with respect to the
coupled unknowns (x, ω), possibly adding regularizers and constraints on both
x and ω in order to prevent the model from favouring trivial solutions [17,21].
Here we choose to add the total variation function on the image and a zero-order
Tikhonov term on the PSF, while imposing nonnegativity on both unknowns and
200 S. Rebegoldi et al.

a normalization constraint on the PSF only. Therefore, we obtain the regularized

problem

argmin J(x; ω) ≡ JC (x; ω) + ρx T V (x) + ιΩx (x) + ρω ω2 + ιΩω (ω) (3)
x∈Rn , ω∈Rn

n
where T V (x) = i=1 ∇i x is the total variation regularizer, being ∇i ∈ R2×n
the discrete gradient operator at pixel i, ρx , ρω > 0 are the regularization
m param-
eters, Ωx = {x ∈ Rn : x ≥ 0} and Ωω = {ω ∈ Rm : ω ≥ 0, i=1 ωi = 1}
are the constraints sets, and ιΩx (resp. ιΩω ) denotes the indicator function of
Ωx (resp. Ωω ). Problem (3) is nonsmooth and nonconvex, not only globally, but
also with respect to the variables blocks x and ω; therefore it is crucial to devise
efficient optimization tools able to avoid uninteresting local minima and speed
up the convergence to a meaningful stationary point.
The aim of this paper is to propose an efficient alternating minimization
strategy to solve problem (3). The proposed method applies a specific variable
metric forward–backward algorithm to each block of variables, and ensures global
convergence towards a stationary point by performing a backtracking procedure
along the descent direction. Unlike other standard approaches, it allows selecting
the metric parameters in a variable manner; this could be helpful in order to get
better reconstructions of both the image x and the PSF ω with reduced com-
putational times. We support this remark by reporting a numerical experience
on some test images corrupted by Cauchy noise, where we show the validity of
the proposed approach in terms of both the quality of the reconstructed objects
and the convergence speed towards the limit point.

2 The Proposed Algorithm

In its standard version, the alternating minimization strategy [13] consists in

cyclically minimizing the objective function J with respect to a single block
of variables, while keeping the other one fixed. Starting from an initial guess
(x(0) , ω (0) ), the resulting iterative sequence {(x(k+1) , ω (k+1) )}k∈N is given by
⎧
⎪
⎨x(k+1) = argmin J(x; ω (k) )
x∈Rn
(4)
⎪
⎩ ω (k+1)
= argmin J(x(k+1) ; ω)
ω∈R n

It is known that solving exactly the subproblems in (4) could be too costly or
generate oscillating sequences in the absence of strict convexity assumptions [13].
Since the objective function in (3) can be decomposed into the sum

J(x; ω) = J0 (x; ω) + J1 (x) + J2 (ω)

 Efficient Block Coordinate Methods for Blind Cauchy Denoising 201

where J0 (x; ω) = JC (x; ω) + ρω ω2 is continuously differentiable, J1 (x) =

ρx T V (x) + ιΩx (x) and J2 (ω) = ιΩω (ω) are continuous on their domains and
convex, it is much more convenient to adopt a proximal–linearized version of the
previous alternating scheme. To this aim, we recall that the proximal operator
of a convex function g is defined as [9, Sect. 3.4]
1
proxα,D
g (x) = argmin g(u) + u − x2D (5)
u∈Rn 2α

where α > 0, D ∈ Rn×n is a symmetric positive definite matrix and yD =

y T Dy denotes the norm induced by D of a vector y ∈ Rn . Then one can

consider the following inexact scheme
⎧ (k) (k)
 
⎨x(k+1) = proxαx ,Dx x (k)
− αx
(k)
(Dx
(k) −1
) ∇x J (x(k)
; ω (k)
)
ρx T V +ιΩx 0
(k) (k)
  (6)
⎩ω (k+1) = proxα ω ,Dω
ιΩω ω (k) − αω (Dω )−1 ∇ω J0 (x(k+1) ; ω (k) )
(k) (k)

where αx , αω > 0 are the steplength parameters, Dx , Dω ∈ Rn×n are the

(k) (k) (k) (k)

scaling matrices and ∇x J0 (x; ω), ∇ω J0 (x; ω) denote the partial gradients of J0
at point (x, ω) with respect to the two blocks of variables.
The alternating scheme (6) is also denominated block coordinate forward–
backward algorithm [5,8,10], since it alternates, on each block of variables, a
forward (gradient) step on the differentiable part with a backward (proximal)
step on the convex part. A standard strategy for the parameters selection in
(k) (k)
(6) consists in choosing Dx (resp. Dω ) as the identity matrix, while taking
(k) (k)
αx (resp. αω ) as the inverse of the Lipschitz constant of the partial gradient
∇x J0 (x; ω ) (resp. ∇ω J0 (x(k+1) ; ω)), see for instance [5]. However, this choice
(k)

is not applicable whenever the partial gradients fail to be Lipschitz continuous

and it might generate sequences of extremely small steplengths. Furthermore,
computing the proximal operators appearing in (6) could still be computation-
α(k) ,D (k)
ally expensive. Indeed, even though proxιΩωω ω reduces to the projection onto
the set Ωω , which can be computed by means of linear-time algorithms [6], the
α(k) ,D (k)
evaluation of proxρxxT V +ι
x
Ωx
requires an inner routine which could heavily slow
down the iterations of the outer loop, if an excessive precision in the evaluation
is required.
In order to overcome such limitations, the authors in [8] propose a modifica-
tion of the inexact scheme (6) which can be seen as a block coordinate extension
of the variable metric linesearch based forward–backward algorithm (VMILA)
[7]. At each outer iteration k ∈ N, the proposed block-VMILA algorithm applied
to problem (3) generates two sequences of inner iterates {x(k,) } and {ω (k,) },
202 S. Rebegoldi et al.

(k) (k)
by performing Nx VMILA steps on the image block and Nω VMILA steps
on the PSF block. The simplified outline of the algorithm is reported below.
⎧ (k,0)
⎪
⎪x = x(k)
⎪
⎪ (k)
⎪
⎪FOR  = 0, 1, . . . , Nx − 1
⎪
⎪ (k,)  
⎪
⎪ (k,)
(k,)
(k,) (k,)
⎪
⎪ ux ≈η proxα x ,Dx
x(k,) − αx (Dx )−1 ∇x J0 (x(k,) ; ω (k) )
⎪
⎪
ρx T V +ιΩx
⎪
⎪ (k,) (k,)
⎪
⎪ x(k,+1) = x(k,) + λx (ux − x(k,) )
⎪
⎪ (k)
⎪
⎪ (k+1)
= x(k,Nx )
⎨x
(7)
⎪
⎪
⎪
⎪ω (k,0) = ω (k)
⎪
⎪
⎪
⎪ (k)
⎪
⎪FOR  = 0, 1, . . . , Nω − 1
⎪
⎪  
⎪
⎪ (k,)
= proxα
(k,)
ω
(k,)
,Dω (k,) (k,)
ω (k,) − αω (Dω )−1 ∇ω J0 (x(k+1) ; ω (k,) )
⎪
⎪ uω ιΩω
⎪
⎪
⎪
⎪
(k,) (k,)
ω (k,+1) = ω (k,) + λω (uω − ω (k,) )
⎪
⎪
⎩ (k+1) (k)
ω = ω (k,Nω )

We now clarify the meaning of the iterates and parameters involved in (7).

∈ Rn is a suitable approximation of the forward–backward

(k,)
– The point ux
(k,)
step performed on the image block, where αx > 0 is a steplength param-
(k,)
eter belonging to the interval [αmin , αmax ] and Dx ∈ Dμ is a symmet-
ric positive definite scaling matrix with eigenvalues in [ μ1 , μ], μ > 1. In
(k,)
order to compute ux , we consider the convex minimum problem asso-
(k,)
ciated to the proximal evaluation at the gradient step zx = x(k,) −
(k,) (k,) −1
αx (Dx ) ∇x J0 (x (k,) (k)
; ω ), namely
(k,) 2
u − zx  (k,)
Dx
min h(k,)
x (u) ≡ (k,)
+ ρx T V (u) + ιΩx (u) + Cx(k,)
u∈Rn 2αx
(k,)
where Cx ∈ R is a suitable constant, and then we reformulate it as the
equivalent dual problem [7, Sect. 3]
(k,) −1
AT v − zx
(k,) (k,) 2
αx (Dx )  (k,)
Dx
max Ψx(k,) (v) ≡− (k,)
− ιΩx∗ (v) + Cx(k,)
v∈R3n 2αx
where A = (∇T1 · · · ∇Tn I)T ∈ R3n×n and Ωx∗ = {v ∈ R3n : (v2i−1 , v2i )2 ≤
ρx , i = 1, . . . , n, vi ≤ 0, i = 2n + 1, . . . , 3n}. Chosen a parameter η ∈
∈ Ωx and a dual variable v (k,) ∈ Ωx∗
(k,)
(0, 1), we look for a primal point ux
satisfying [7, Eq. 15]

h(k,)
x (u(k,)
x ) ≤ ηΨx(k,) (vx(k,) ). (8)
(k,) (k,)
The primal-dual pair (ux , vx ) can be computed by generating a sequence
of dual variables converging to the solution of the dual problem and then
stopping the dual iterates when condition (8) is achieved.
 Efficient Block Coordinate Methods for Blind Cauchy Denoising 203

(k,)
– The point uω is the projection of the gradient step onto the nonnegativity
orthant plus a normalization constraint. As described in [6], this problem is
equivalent to solving a one-dimensional root finding problem for a piecewise
linear monotonically nondecreasing function, which can be tackled by spe-
cialized linear time algorithms. In the numerical experiments, we will adopt
the same secant-based method used in [6,11].
(k,) (k,)
– The linesearch parameters λx , λω ∈ (0, 1] are chosen in such a way
that the sufficient decrease conditions (9)-(10) are imposed on the image
and PSF block, respectively. This is done by performing a backtracking pro-
(k,) (k,)
cedure along the descent directions defined by dx = ux − x(k,) and
(k,) (k,)
dω = uω − ω (k,) . Note that the stopping criterion (8) and the projection
(k,) (k,) (k,) (k,)
operation onto Ωω imply that hx (ux ) < 0 and hω (uω ) < 0, so that
the two inequalities (9)-(10) can be actually considered as descent conditions.
The block-VMILA scheme is entirely reported in Algorithm 1. Under mild
assumptions on the objective function and provided that the proximal operators
are computed with increasing accuracy, it is possible to prove that each limit
point (if any) of the sequence generated by the block-VMILA scheme is station-
ary for the objective function [8, Theorem 1]. In particular, Algorithm 1 applied
to problem (3) satisfies the required assumptions for ensuring the stationarity of
the limit points. We report the corresponding theoretical result below. For the
sake of clarity, we recall that the subdifferential of a convex function g at point
y is given by ∂g(y) = {v ∈ Rn : g(z) ≥ g(y) + v T (z − y), ∀ z ∈ Rn }, whereas
the normal cone to a convex set Ω at point y is defined as NΩ (y) = {v ∈ Rn :
v T (z − y) ≤ 0, ∀ z ∈ Rn }.

Theorem 1. Let J(x; ω) be defined as in problem (3). Any limit point (x̄, ω̄) of
the sequence {(x(k) , ω (k) )}k∈N generated by Algorithm 1 is stationary, that is

0 ∈ ∇J0 (x̄; ω̄) + (ρx ∂T V (x̄) + NΩx (x̄)) × NΩω (ω̄).

The convergence of the iterates {(x(k) , ω (k) )}k∈N to a stationary point can
be proved when the proximal operators are computed exactly and by assuming
that the objective function satisfies the so-called Kurdyka–Lojasiewicz property
at each of its stationary points [5, Definition 3]. The proof of this stronger con-
vergence result for the block-VMILA scheme can be found in [8, Theorem 2].

3 Numerical Experience
For our numerical experiments, we consider a dataset of five 256 × 256 grayscale
images and assume that the true PSF ω is associated to a Gaussian kernel with
window size 9 × 9 and standard deviation equal to 2. The scale parameter γ
has been set equal to 0.02 as in [25], whereas the regularization parameters
ρx and ρω have been manually tuned in order to provide the most visually
satisfactory reconstructions for both the image and the PSF. The blurred and
noisy images have been obtained by convolving the true objects with ω and then
204 S. Rebegoldi et al.

Algorithm 1. Block-VMILA for blind Cauchy denoising

Choose the initial guesses x(0) ∈ Ωx , ω (0) ∈ Ωω , the real numbers 0 < αmin ≤ αmax ,
μ ≥ 1, δ, β, η ∈ (0, 1) and the nonnegative integers N̄x , N̄ω .

FOR k = 0, 1, 2, . . .

STEP 1. 1. Set x(k,0) = x(k) .

(k)
2. Choose the number of inner iterations Nx ≤ N̄x .
(k)
3. FOR  = 0, 1, . . . , Nx − 1
(k,) (k,)
1. Choose αx ∈ [αmin , αmax ], Dx ∈ Dμ .
(k,) (k,) (k,)
2. Set zx = x(k,) − αx (Dx )−1 ∇x J0 (x(k,) ; ω (k) ).
(k,) (k,)
(k,) (k,)
3. Compute ux ≈η proxαx ,Dx
ρx T V +ιΩ (zx ) according to (8).
x
(k,) (k,) (k,)
4. Set dx
= −x ux .
5. Compute the smallest nonnegative integer m such that

J(x(k,) + δ m d(k,)
x ; ω (k) ) ≤ J(x(k,) ; ω (k) ) + βδ m h(k,)
x (u(k,)
x ). (9)
(k,)
6. Compute x(k,+1) = x(k,) + δ m dx .

4. Set x(k+1) = x(k,Nx ) .

STEP 2. 1. Set ω (k,0) = ω (k) .

(k)
2. Choose the number of inner iterations Nω ≤ N̄ω .
(k)
3. FOR  = 0, 1, . . . , Nω − 1
(k,) (k,)
1. Choose αω ∈ [αmin , αmax ], Dω ∈ Dμ .
(k,) (k,) (k,)
2. Set zω = ω (k,) − αω (Dω )−1 ∇ω J0 (x(k+1) ; ω (k,) ).
(k,) (k,)
(k,) (k,)
3. Compute uω = proxαω
ιΩ
,Dω
(zω ).
ω
(k,) (k,) (k,)
4. Set =
dω −ω uω .
5. Compute the smallest nonnegative integer m such that

J(x(k+1) ; ω (k,) + δ m d(k,)

ω ) ≤ J(x(k+1) ; ω (k,) ) + βδ m h(k,)
ω (u(k,)
ω ).
(10)
(k,)
6. Compute ω (k,+1) = ω (k,) + δ m dω .

4. Set ω (k+1) = ω (k,Nω ) .

adding Cauchy noise, which has been generated by dividing two independent
realizations of a normal random variable and then multiplying the result by the
parameter γ (see [25, Sect. 5.1] for more details).
We consider the following two alternative choices for the scaling matrices
(k,) (k,)
Dx and Dω appearing in Algorithm 1.

– Split Gradient (SG): according to the strategy proposed in [15], we first

decompose the partial gradients ∇x J0 (x(k,) ; ω (k) ) and ∇ω J0 (x(k+1) ; ω (k,) )
 Efficient Block Coordinate Methods for Blind Cauchy Denoising 205

into the difference of a positive and a nonnegative part, and then we choose
(k,) (k,)
Dx and Dω as diagonal matrices whose diagonal elements are given by
 
 −1 xj
(k,)
1
(k,)
Dx = max min (k)
, μk ,
jj Vx (x(k,) ) μk
 
 −1 ωj
(k,)
1
(k,)
Dω = max min (k)
, μk ,
jj Vω (ω (k,) ) μk

(k) T (k)
ω ⊗x−g (k) T
ω⊗x −g (k+1)
where Vx (x) = Ω (k) γ 2 +(ω (k) ⊗x−g)2 and Vω (ω) = X (k) γ 2 +(ω⊗x(k+1) −g)2

are the positive parts coming from the decompositions of the partial gradients,
being Ω (k) , X (k) ∈ Rn×n the two block circulant matrices with circulant
blocks such that ω (k) ⊗ x = Ω (k) x and ω ⊗ x(k+1) = X
ω, whereas {μk }k∈N
(k)

is a sequence of positive thresholds chosen as μk = 1 + P/k p with p = 2,

P = 1010 , so that the scaling matrices are gradually converging to the identity
matrix.
(k,) (k,)
– Identity Matrix (I): in this case, we let Dx = Dω = In for all k, .
(k,)
For both choices of the scaling matrices, the steplength parameters αx and
(k,)
αω are computed by appropriately alternating the two scaled Barzilai-Borwein
(BB) rules [6]. If we focus on the image block (the reasoning follows identically
for the PSF block), each scaled BB steplength is obtained by imposing a quasi-
Newton property on the matrix B(α) = α−1 Dx , thus obtaining
(k,)

 −1
(k,) T
(k,) T (k,) (k,) (k,)
(k,) (k,)
sx Dx Dx sx sx Dx yx
BB1 BB2
αx = , αx =  −1  −1
(k,) T (k,) (k,) (k,) T (k,) (k,) (k,)
sx Dx yx yx Dx Dx yx

(k,)
where sx = x(k,) − x(k,−1) is the difference between two consecutive inner
(k,)
iterates and yx = ∇x J0 (x(k,) ; ω (k) ) − ∇x J0 (x(k,−1) ; ω (k) ) is the difference
between two consecutive partial gradients. At each inner iterate, one of the two
scaled BB steplengths is chosen according to the alternation strategy described in
[6, Sect. 3.3]. The chosen value is then constrained within the interval [αmin , αmax ]
with αmin = 10−10 and αmax = 102 .
It is well known that, in the quadratic case, the BB rules well approximate
the reciprocals of some eigenvalues of the Hessian matrix, which is a desirable
property in order to ensure fast convergence of gradient methods [12]. There
is some evidence that such good behaviour still holds for non-quadratic prob-
lems and it can be further enhanced when the BB rules are combined with the
split gradient strategy, as confirmed by several numerical experiments in previ-
ous works [6–8]. Therefore it seems reasonable to expect the same accelerated
behaviour for the block-VMILA algorithm.
(k) (k)
According to Algorithm 1, the number of inner steps Nx and Nω may
vary at each outer iteration k, provided that they do not exceed the a priori
fixed upper bounds N̄x and N̄ω . Then following the strategy adopted in [8], we
206 S. Rebegoldi et al.

stop the inner iterates x(k,) (resp. ω (k,) ) when either the maximum number of
(k,) (k,) (k,) (k,)
inner iterations is achieved or when the quantity hx (ux ) (resp. hω (uω ))
(k,) (k,) (k,) (k,)
is sufficiently small. This is reasonable, since hx (ux ) (resp. hω (uω )) is
zero if and only the inner iterate is stationary for the objective function restricted
to the single block of variables (see [7,8] and references therein). In conclusion,
the inner iterations numbers are set as

Nx(k) = min{min{ ∈ N : |h(k,)

x (u(k,)
x )| ≤ (k)
x }, N̄x }

Nω(k) = min{min{ ∈ N : |h(k,)

ω (u(k,)
ω )| ≤ (k)
ω }, N̄ω }

(k) (k)
where N̄x = N̄ω = 5 and the tolerance parameters x and ω are halved when-
ever the inner routine does not perform more than one inner iteration.
In addition to the two variants of Algorithm 1 described above, we implement
also the more standard block coordinate forward–backward scheme (6). For this
(k) (k)
method, the scaling matrices Dx and Dω are chosen equal to the identity
(k) (k)
matrix, whereas the steplength parameters are selected as αx = 1/Lx and
αω = 1/Lω , being Lx = γ −2 Ω (k) 1 Ω (k) ∞ , Lω = γ −2 X (k) 1 X (k) ∞
(k) (k) (k) (k)

two upper estimates for the Lipschitz constants of the partial gradients
∇x J0 (x(k,) ; ω (k) ) and ∇ω J0 (x(k+1) ; ω (k,) ). The proximal operator of the total
variation term plus the indicator function of Ωx is computed inexactly, by using
the same stopping criterion (8) adopted for Algorithm 1. The resulting method
can be considered as an inexact version of the proximal alternating linearized
minimization (PALM) algorithm devised in [5], which does not originally include
the possibility of computing inexactly the proximal operators of the convex terms
involved in the minimization problem. For that reason, we will refer to this imple-
mentation of the block scheme (6) as the inexact PALM (inePALM in short).
For all algorithms, the initial guess x(0) has been chosen equal to the observed
image g, whereas the initial PSF ω (0) has been set as either the constant image
satisfying the normalization constraint or a Gaussian function with standard
deviation equal to 1. In Fig. 1, we observe that the block-VMILA scheme is able
to provide accurate reconstructions of the ground truth, whereas the inePALM
estimated images still look blurry, as if no information on the PSF had been
retrieved by the algorithm. This is confirmed by looking at Figs. 4-5, where we
report the plots of the relative root-mean-square error (RMSE) with respect to
time on both the reconstructed image x(k) and PSF ω (k) , i.e.

x(k) − x̄ ω (k) − ω̄

RM SE(x(k) ) = , RM SE(ω (k) ) =
x̄ ω̄

being x̄, ω̄ the true image and PSF, respectively. We see that the inePALM
algorithm does not move away from the initial PSF, whatever the test problem
considered. Consequently, the algorithm fails in deblurring the observed images
and gets stuck in stationary points which provide bigger values than the ones
identified by block-VMILA, as we can deduce from the function plots reported in
(k)
Figs. 2-3. The inePALM failure is due to the excessively small steplengths αω ,
 Efficient Block Coordinate Methods for Blind Cauchy Denoising 207

Fig. 1. From left to right: blurred and noisy images, VMILA-SG and inexact PALM
reconstructions obtained by initializing the point spread function with a Gaussian blur.
208 S. Rebegoldi et al.

105 105 105

-1.2 -1.2 -1
VMILA I VMILA I VMILA I
VMILA SG VMILA SG VMILA SG
-1.2
-1.4 inePALM -1.4 inePALM inePALM

-1.4
-1.6 -1.6
f(x)

f(x)

f(x)
-1.6

-1.8 -1.8
-1.8

-2 -2
-2

-2.2 -2.2 -2.2

0 10 20 30 40 50 60 0 10 20 30 40 50 60 0 10 20 30 40 50 60
Time (s) Time (s) Time (s)

105 105
-0.8 -0.8
VMILA I VMILA I
-1 VMILA SG -1 VMILA SG
inePALM inePALM
-1.2 -1.2

-1.4 -1.4
f(x)

f(x)

-1.6 -1.6

-1.8 -1.8

-2 -2

-2.2 -2.2
0 10 20 30 40 50 60 0 10 20 30 40 50 60
Time (s) Time (s)

Fig. 2. Decrease of the objective function values vs time obtained by initializing the
point spread function with a constant image. From left to right: baboon, boat and
cameraman (top row), parrot and peppers (bottom row).
105 105 105
-1.7 -1.6 -1.7
VMILA I VMILA I VMILA I
-1.75 VMILA SG VMILA SG -1.75 VMILA SG
inePALM -1.7 inePALM inePALM
-1.8 -1.8

-1.85 -1.85
-1.8
f(x)

f(x)

f(x)

-1.9 -1.9

-1.9
-1.95 -1.95

-2 -2
-2
-2.05 -2.05

-2.1 -2.1 -2.1

0 10 20 30 40 50 60 0 10 20 30 40 50 60 0 10 20 30 40 50 60
Time (s) Time (s) Time (s)

105 105
-1.7 -1.6
VMILA I VMILA I
-1.75 VMILA SG VMILA SG
inePALM -1.7 inePALM
-1.8

-1.85
-1.8
f(x)

f(x)

-1.9

-1.9
-1.95

-2
-2
-2.05

-2.1 -2.1
0 10 20 30 40 50 60 0 10 20 30 40 50 60
Time (s) Time (s)

Fig. 3. Decrease of the objective function values vs time obtained by initializing the
point spread function with a Gaussian function. From left to right: baboon, boat and
cameraman (top row), parrot and peppers (bottom row).
 Efficient Block Coordinate Methods for Blind Cauchy Denoising 209

0.24 0.25 0.28

VMILA I VMILA I VMILA I
VMILA SG VMILA SG 0.26 VMILA SG
0.22
inePALM inePALM inePALM
0.24
0.2 0.2
0.22
RMSE(x(k))

RMSE(x(k))

RMSE(x(k))
0.18 0.2

0.18
0.16 0.15
0.16
0.14
0.14

0.12 0.1 0.12

0 10 20 30 40 50 60 0 10 20 30 40 50 60 0 10 20 30 40 50 60
Time (s) Time (s) Time (s)
0.3 0.24
VMILA I VMILA I
VMILA SG 0.22 VMILA SG
inePALM inePALM
0.25 0.2

0.18
RMSE(x(k))

RMSE(x(k))
0.2 0.16

0.14

0.15 0.12

0.1

0.1 0.08
0 10 20 30 40 50 60 0 10 20 30 40 50 60
Time (s) Time (s)

Fig. 4. Relative root-mean-square error (RMSE) on the image vs time obtained by

initializing the point spread function with a Gaussian function. From left to right:
baboon, boat and cameraman (top row), parrot and peppers (bottom row).

1.4 1.4 1.4

VMILA I VMILA I VMILA I
VMILA SG VMILA SG VMILA SG
1.2 1.2 1.2
inePALM inePALM inePALM

1
1 1
RMSE( (k))

RMSE( (k))

RMSE( (k))

0.8
0.8 0.8

0.6
0.6 0.6

0.4
0.4 0.4
0.2
0.2 0.2
0 10 20 30 40 50 60 0 10 20 30 40 50 60 0 10 20 30 40 50 60
Time (s) Time (s) Time (s)
1.4 1.4
VMILA I VMILA I
VMILA SG 1.2 VMILA SG
1.2
inePALM inePALM
1
1
RMSE( (k))

RMSE( (k))

0.8
0.8
0.6
0.6
0.4

0.4 0.2

0.2 0
0 10 20 30 40 50 60 0 10 20 30 40 50 60
Time (s) Time (s)

Fig. 5. Relative root-mean-square error (RMSE) on the PSF vs time obtained by

initializing the point spread function with a Gaussian function. From left to right:
baboon, boat and cameraman (top row), parrot and peppers (bottom row).
210 S. Rebegoldi et al.

(k)
which in turn depend on the extremely large Lipschitz constants Lω . This issue
is completely avoided by the block-VMILA scheme, thanks to the use of larger
steplengths in the computation of the forward–backward step.
From the plots in Figs. 4-5, it is also evident that the block-VMILA algo-
rithm performs better when equipped with the split gradient strategy for the
scaling matrix. Indeed, for each test problem, the VMILA-SG variant provides a
reconstructed PSF with relative error smaller than 50% in less than 10 seconds
on a laptop equipped with a 2.60 GHz Intel(R) Core(TM) i7-4510U and 8 GB
of RAM, whereas the non-scaled VMILA-I variant takes in average sextuple the
time to get to the same accuracy level.

4 Conclusions
In this paper, we have addressed the problem of blind Cauchy denoising by means
of a block coordinate linesearch based forward backward method. The proposed
approach allows the user to freely select the metric parameters, while imposing
a backtracking procedure along the feasible direction in order to guarantee the
convergence towards a stationary point. The scheme is accelerated in practice
by making use of very well-known adaptive strategies for the selection of the
parameters, such as the Barzilai-Borwein rules or the split-gradient strategy,
which capture some second order local information of the objective function. The
numerical experiments here presented demonstrate that the proposed approach is
highly competitive with respect to standard block coordinate methods, avoiding
uninteresting stationary points and providing more accurate reconstructions of
both the image and the point-spread-function.

Acknowledgements. The authors are members of the INDAM research group GNCS.

References
1. Aubert, G., Ajoul, J.-F.: A variational approach to removing multiplicative noise.
SIAM J. Appl. Math. 68(4), 925–946 (2008)
2. Ayers, G.R., Dainty, J.C.: Iterative blind deconvolution method and its applica-
tions. Opt. Lett. 13(7), 547–549 (1988)
3. Banerjee, S., Agrawal, M.: Underwater acoustic communication in the presence of
heavy-tailed impulsive noise with bi-parameter Cauchy-Gaussian mixture model.
In: Ocean Electronics (SYMPOL), pp. 1–7 (2013)
4. Bertero, M., Boccacci, P., Ruggiero, V.: Inverse Imaging with Poisson Data, pp.
2053–2563. IOP Publishing, Bristol (2018)
5. Bolte, J., Sabach, S., Teboulle, M.: Proximal alternating linearized minimization for
nonconvex and nonsmooth problems. Math. Program. 146(1–2), 459–494 (2014)
6. Bonettini, S., Zanella, R., Zanni, L.: A scaled gradient projection method for con-
strained image deblurring. Inverse Problems 25(1), 015002 (2009)
7. Bonettini, S., Loris, I., Porta, F., Prato, M., Rebegoldi, S.: On the convergence of
a linesearch based proximal-gradient method for nonconvex optimization. Inverse
Problems 33(5), 055005 (2017)
 Efficient Block Coordinate Methods for Blind Cauchy Denoising 211

8. Bonettini, S., Prato, M., Rebegoldi, S.: A block coordinate variable metric line-
search based proximal gradient method. Comput. Optim. Appl. 71(1), 5–52 (2018)
9. Chambolle, A., Pock, T.: An introduction to continuous optimization for imaging.
Acta Numerica 25, 161–319 (2016)
10. Chouzenoux, E., Pesquet, J.C., Repetti, A.: A block coordinate variable metric
forward-backward algorithm. J. Glob. Optim. 66(3), 457–485 (2016)
11. Dai, Y.H., Fletcher, R.: New algorithms for singly linearly constrained quadratic
programming problems subject to lower and upper bounds. Math. Program. 106,
403–421 (2006)
12. Frassoldati, G., Zanghirati, G., Zanni, L.: New adaptive stepsize selections in gra-
dient methods. J. Ind. Manag. Optim. 4(2), 299–312 (2008)
13. Grippo, L., Sciandrone, M.: On the convergence of the block nonlinear Gauss–
Seidel method under convex constraints. Oper. Res. Lett. 26(3), 127–136 (2000)
14. Idan, M., Speyer, J.: Cauchy estimation for linear scalar systems. IEEE Trans.
Autom. Control 55(6), 1329–1342 (2010)
15. Lantéri, H., Roche, M., Cuevas, O., Aime, C.: A general method to devise max-
imum likelihood signal restoration multiplicative algorithms with non-negativity
constraints. Signal Process. 81(5), 945–974 (2001)
16. Laus, F., Pierre, F., Steidl, G.: Nonlocal myriad filters for Cauchy noise removal.
J. Math. Imaging Vision 60(8), 1324–1354 (2018)
17. Levin, A., Weiss, Y., Durand, F., Freeman, W.T.: Understanding blind decon-
volution algorithms. IEEE Trans. Pattern Anal. Mach. Intell. 33(12), 2354–2367
(2011)
18. Mei, J.-J., Dong, Y., Huang, T.-Z.: Cauchy noise removal by nonconvex ADMM
with convergence guarantees. J. Sci. Comput. 74(2), 743–766 (2018)
19. Nikolova, M.: A variational approach to remove outliers and impulse noise. J. Math.
Imaging Vision 20(1–2), 99–120 (2004)
20. Peng, Y., Chen, J., Xu, X., Pu, F.: SAR images statistical modeling and classifica-
tion based on the mixture of alpha-stable distributions. Remote Sens. 5, 2145–2163
(2013)
21. Perrone, D., Favaro, P.: A clearer picture of total variation blind deconvolution.
IEEE Trans. Pattern Anal. Mach. Intell. 38(6), 1041–1055 (2016)
22. Prato, M., La Camera, A., Bonettini, S., Bertero, M.: A convergent blind decon-
volution method for post-adaptive-optics astronomical imaging. Inverse Problems
29, 065017 (2013)
23. Prato, M., La Camera, A., Bonettini, S., Rebegoldi, S., Bertero, M., Boccacci, P.:
A blind deconvolution method for ground based telescopes and Fizeau interferom-
eters. New Astron. 40, 1–13 (2015)
24. Rudin, L.I., Osher, S., Fatemi, E.: Nonlinear total variation based noise removal
algorithms. J. Phys. D. 60(1–4), 259–268 (1992)
25. Sciacchitano, F., Dong, Y., Zeng, T.: Variational approach for restoring blurred
images with Cauchy noise. SIAM J. Imaging Sci. 8(3), 1894–1922 (2015)
High Performance Computing in
Modelling and Simulation
 A Parallel Software Platform for Pathway
Enrichment

Giuseppe Agapito(B) and Mario Cannataro

University of Catanzaro, 88100 Catanzaro, Italy

{agapito,cannataro}@unicz.it

Abstract. Biological pathways are complex networks able to provide a

view on the interactions among bio-molecules inside the cell. They are
represented as a network, where the nodes are the bio-molecules, and the
edges represent the interactions between two biomolecules. Main online
repositories of pathways information include KEGG that is a reposi-
tory of metabolic pathways, SIGNOR that comprises primarily signaling
pathways, and Reactome that contains information about metabolic and
signal transduction pathways. Pathways enrichment analysis is employed
to help the researchers to discriminate relevant proteins involved in the
development of both simple and complex diseases, and is performed with
several software tools. The main limitation of the current enrichment
tools are: (i) each tool can use only a single pathway source to compute
the enrichment; (ii) researchers have to repeat the enrichment analysis
several times with different tools (able to get pathway data from different
data sources); (iii) enrichment results have to be manually merged by
the user, a tedious and error-prone task even for a computer scientist.
To face this issues, we propose a parallel enrichment tool named Parallel
Enrichment Analysis (PEA) ables to retrieve at the same time path-
ways information from KEGG, Reactome, and SIGNOR databases, with
which to automatically perform pathway enrichment analysis, allowing
to reduce the computational time of some order of magnitude, as well as
the automatic merging of the results.

Keywords: Pathways · Pathway database · Parallel computing ·

Pathway enrichment analysis

1 Introduction
Biological pathways are human representations of the existent interactions
among biomolecules, that regulate how cellular functions are carried out both
in healthy and diseased state and how cells can interact with the external envi-
ronment. Biological Pathways can be classified into three categories: Signalling
Pathways, Metabolic Pathways, and Regulatory Pathways. Several online
databases store, represent, and share different types of pathways. For exam-
ple, Reactome and KEGG store all three types of pathways while SIGNOR

c Springer Nature Switzerland AG 2020

Y. D. Sergeyev and D. E. Kvasov (Eds.): NUMTA 2019, LNCS 11973, pp. 215–222, 2020.
https://doi.org/10.1007/978-3-030-39081-5_19
216 G. Agapito and M. Cannataro

includes only signaling pathways and Metacyc comprises only metabolic path-
ways. Although, databases containing the same kinds of pathways (like KEGG
and Reactome) show minimal overlap on the number of pathways and gene cov-
erage as described in [9]. Each database has its own representation conventions
and data access methods, making data integration from multiple databases a
considerable challenge. This calls for a need to define a unique file format that
makes it possible to standardize data coming from various data sources.
Pathways are represented using structured file formats (i.e., XML, OWL,
RDF, XML-based, and psimi-xml) or unstructured text files. The most used
structured file format is the Biological Pathway eXchange format (BIOPAX),
categorized in - (LEVELS 1,2,3) [2]. Thus, the search, comparison, and iden-
tification of similar data types from different sources is often difficult. To fill
this gap, in this work we present a parallel software algorithm named Paral-
lel Enrichment Analysis (PEA) with which to simply and effectively manage
information contained in several online databases based on BioPAX and XML
formats. Pathway enrichment tools such as CePa [5], SPIA [10], and TPEA
[7] can perform pathway enrichment analysis, exploiting data coming from a
single specific data source. PEA instead, can perform pathway enrichment anal-
ysis, exploiting data coming from several data source into the same analysis (in
this version, PEA can retrive information from Reactome, KEGG, and SIGNOR
databases, as well as each data compliant with the BioPAX Level 3 format). PEA
is implemented using the cross platform language Java 8, using a multi-threads
solution, where threads are mapped on the available physical cores. Retrieving
data from multiple databases is an easily parallelizable task, because there is
no need to share information among threads to retrieve information from inde-
pendent data sources. As a result, the enrichment results coming from different
data source are automatically merged together, allowing thus to obtain more
informative results without be necessary to use multiple software tools. Path-
ways enrichment analysis in PEA is implemented as a customized version of the
Hypergeometric distribution function.
The remaining part of the manuscript is organized as follows: Sect. 2 presents
the state of the art of pathway databases, along with some well-known enrich-
ment tools. Section 3 introduces the PEA software platform and its capability,
whereas in Sect. 4 the PEA’s performance are evaluated. Finally, Sect. 5 con-
cludes the paper and delineate some possible future works and extensions.

2 Related Works
The number of pathway databases is growing quickly in recent years. This is
advantageous because biologists often need to use information from many sources
to support their research. Here we report a short list of well-known pathway
databases.
– Reactome is an open source pathway database [3,4]. Currently Reactome con-
tains the whole known pathways coming from 22 different organisms includ-
ing human. Pathway can be download in different formats comprising SBML,
BioPAX and other graphical formats.
 A Parallel Software Platform for Pathway Enrichment 217

– KEGG is a collection of 19 interconnected databases, including genomic,

chemical, pathway and phenotypic information [6]. KEGG stores pathways
from several organisms, including human. KEGG data can be accessed using
the KEGG API or KEGG FTP, allowing users to download each pathway in
the KGML format (KEGG XML format).
– SIGnaling Network Open Resource (SIGNOR) is a collection of approximately
12,000 manually-annotated causal relationships and about 2800 human pro-
teins participating in signal pathways [8]. Pathways are manually curated,
and available for the download from the Download page.
– PathwayCommons [1] is a collection of public pathway databases, providing
an access point for a collection of public databases. PathwayCommons pro-
vides a web interface to browse pathways, as well as a web service API for
automatic access to the data. Also, PSI-MI and BioPAX, formats are sup-
ported for the data download.
On the other hand, each database has its own representation conventions, path-
way are encoded using different type of file such as: BIOPAX -(LEVEL 1,2,3) -,
SIF, SBML, XML, and flat text files. BIOPAX -Biological Pathway Exchange-
, is a meta language defined in OWL and is represented in the RDF/XML
format. Simple Interaction Format (SIF) is a data exchange format used to
represent molecular interactions and extended to represent biological models.
Systems Biology Markup Language (SBML) is focused on describing biological
networks. Here we introduce some available pathway enrichment tools and their
functionalities.

– Topology-based Pathway Enrichment Analysis (TPEA) [7] is a pathway

enrichment tool based on graph topology. Pathways enrichment analysis in
TPEA is obtained by using the KEGG database. TPEA is available as a R
package at https://cran.r-project.org/web/packages/TPEA/.
– CePa [5] performs enrichment pathways analysis based on topological infor-
mation in addition to gene-set information. Enrichment analysis in CePa is
obtained exploiting the pathways contained in KEGG database. CePa is avail-
able as R package, at https://cran.r-project.org/web/packages/CePa/.
– Signaling Pathway Impact Analysis (SPIA) [10] combines the evidence
obtained from the classical enrichment analysis with the measure of the
perturbation on pathways under a given condition. Pathways are obtained
from KEGG. SPIA is an R package available at https://bioconductor.org/
packages/SPIA/.

All the listed pathway enrichment tools can perform enrichment analysis
exploiting the information available into a single data source. Due to this limita-
tion, researchers have to repeat several time the enrichment analysis by using the
opportune software tool compatibles with the chosen data source. In many cases,
software tools are available only for well-known data repository such as KEGG or
Reactome, limiting the number of data source usable from the researchers. This
scenario, calls for the need to define software tools can dig with many databases
at the same time. To overcome the limitation of the existent software tools, we
218 G. Agapito and M. Cannataro

propose Parallel Enrichment Analysis (PEA) software tools. PEA preforms par-
allel pathway enrichment analysis from different data sources, providing more
relevant results, allowing researchers to foster a greater understanding of the
underlying biology. PEA can analyze data coming from KEGG, Reactome, and
SIGNOR in parallel as well as, to perform pathways enrichment using any path-
way data coded using the BioPAX-Level 3 format.

3 Parallel Enrichment Analysis Platform

Parallel Enrichment Analysis (PEA) is a software platform developed to perform
pathway enrichment analysis in parallel, using data coming from different data
sources. PEA is developed by using the Java 8 cross-platform programming lan-
guage, making it compatible with all the operating systems that support Java.
To efficiently manage concurrency we employed the built-in support for multi-
threaded programming. PEA is based on a modular architecture that provides
a two-fold advantage, first each module can handle a specific type of data, and
second it is possible to add new modules making PEA easily compatible with
new data sources. The PEA’s architecture is depicted in Fig. 1.

Fig. 1. PEA architecture.

The key modules of PEA are: Importer-Container, Enrichment-

Computational-Engine, Enrichment-Results-Collector, and Statistical-Library. In
the current version of PEA the Importer-Container comprises KEGG, SIG-
NOR, Reactome, and the General BioPAX and XML - Importer. The Importer-
Container provides built-in functionalities, allowing to include additional mod-
ules with a minimum programming effort, simulating the Java plug-in importer.
 A Parallel Software Platform for Pathway Enrichment 219

The available importer-modules are independent, property that would enable

PEA to retrieve, and preprocess data in parallel making the import process scal-
able. Preprocessing is mandatory to convert data in a suitable format for the
subsequent enrichment analysis. Conversely, from the other enrichment tools,
PEA provides an outlet with which to get data from multiple data sources i.e.,
Reactome, KEGG, and SIGNOR. In addition, it provides a generic BioPAX
and XML importer. Traditional pathways enrichment tools cannot use a generic
pathway, i.e., a new pathway coded in BioPAX or XML by the user, with which
to perform pathway enrichment analysis. PEA, introducing a general BioPAX-
XML importer, can also use generic pathways to perform pathway enrichment
analysis. The Enrichment-Computational-Engine module, allows to carry out the
enrichment analysis quickly, efficiently, and automatically from all the available
data source in the same experiment. The enrichment analysis is done in parallel
for each data source. When an importer finishes its job, it sends the prepro-
cessed data to the Enrichment-Computational-Engine that in parallel starts to
perform the enrichment per each single data source. The enrichment results are
handled from the Enrichment-Results-Collector that merges together the results
obtained from the multiple enrichment in an single result file. Pathway enrich-
ment results are sorted in a increasing probability value (p-value) of statistical
relevance, and only the results for which the computed p-value is greater than
a threshold defined by the user (default p-value is 0.05) are stored and provided
to the users. In addition, in PEA the FDR (false discovery rate) and Bonfer-
roni methods of multiple statistical test adjustment are available. Finally, the
Statistical-Library module is a collection of statistical measures for the automatic
analysis of pathways coming from multiple data sources.

4 Performance Evaluation

This Section presents the performance evaluation of PEA currently implemented

as a multi-threaded Java application. The experiments were run on a single
cluster’s node made available by the ICAR-CNR Institute, Italy, and composed
by 24 computing nodes (CN) each one containing 132 GBytes of RAM, 1 Tbytes
hard disk, and 16 Intel(R) Xeon(R) CPU E5–2670 at (2.60GHz) with 16 cores
each. Since Intel uses the hyper-threading technology, each CN has 16 CPUs
and 32 virtual cores. The operating system is Linux Red Hat version 4.4.6-4
(x86 64), whereas the versions of JRE and JDK used are, OpenJDK Runtime
Environment, and OpenJDK 64-Bit Server VM both at 64 bit.
For our experiments, we used Reactome, KEGG, and SIGNOR pathway
databases, the data sources from which to retrieve the information and per-
form the multiple pathway enrichment analysis. To compute pathway enrich-
ment analysis, we simulated an up-regulate genes dataset, including about 50,000
distinct genes. PEA parallel engine presents a multi-level of parallelism: the
first one concerns the capability to load in parallel the pathways data from
the three databases. In PEA, pathway data are loaded through ad-hoc Pipes.
Pipes are independent, making it possible to retrieve data in parallel from all
220 G. Agapito and M. Cannataro

the available databases. In the current implementation, pipes are implemented

as single-thread, to avoid that performance decreases significantly due to the
delay introduced by I/O operations performed from additional threads per sin-
gle database. In the current version of PEA, the downloaded information from
the three databases have been stored into three separate files. The second level of
parallelism concerns with the enrichment. In PEA, to speed up the enrichment
analysis, the enrichment-engine splits the enrichment genes dataset among the
available physical cores, to balance the workload, speeding-up the time necessary
to compute the enrichment. In the experiments, PEA has run on one computa-
tion node of the cluster varying the number of active threads in the enrichment
analysis. Mapping each thread on a physical core, i.e., changing the number of
parallel worker between 2, 4, 8, and 12. Each thread is executed on a core for a
maximum of 12 cores for single data source (to avoid that PEA-Threads saturate
the available phisical cores). The sequential PEA has run on just one core, run-
ning each data importer and enrichment analysis sequentially, employing only
a single thread for the whole analysis process. We measured the response time
of PEA (i.e., the execution times of the preprocess and enrichment). For each
experiment configuration, we made 10 runs and measured the mean of execution
times. The execution times are reported in Table 1.

Table 1. Table shows the execution times obtained by PEA using 2, 4, 6, 8, 10, 12 cores,
computing pathway enrichment analysis from all the available databases. Loading is
done in parallel, using one thread per database. EA refers to Enrichment Analysis.

#Cores Prep times (ms) EA Times (ms)

1 236970 36313
2 192739 7850
4 192228 7590
6 191624 6600
8 191624 6240
10 192228 5760
12 192225 5180

Figure 2 conveys the PEA’s speed-up for loading the input data and for the
pathway enrichment analysis, respectively.
The PEA’s enrichment analysis speed-up (black line in Fig. 2) is super-linear
until 4 cores, and start to decrease when passing to 6, 8, 10 and 12 cores for each
databases. This behavior can be explained because, until 4 threads per single
data-source the available physical cores are enough to directly handle 12 threads
(the CPU has 16 physical cores). The speed-up starts to decrease when, to map
the threads per single data-source, in addition to the physical cores it is neces-
sary to use the Hyper-Threading technology available in the Intel’s CPUs. This
knowledge can be used, to automatically set the maximum number of threads to
 A Parallel Software Platform for Pathway Enrichment 221

12

Linear Speedup
Preprocessing Speed-up
Enrichment Speed-up
10

8
Speed-Up

0
0 2 4 6 8 10 12
Number of Cores

Fig. 2. PEA speed up.

use in the enrichment, avoiding to decrease the performance of the system due
to bad resource management. The PEA’s pre-processing speed-up (dashed blue
line in Fig. 2) is sub-linear due to disk readings. On the other hand, reading in
parallel the three dataset allow to save time, respect the sequential read (let see
Table 1). Using more threads to read a single file, it results in a decreasing of the
system performance. In fact, in the current version of PEA data from the three
datasets are stored as single huge files, making multi-thread reading ineffective.
Thus, using a redundant array of independent disks (RAID), and splitting data
on more small files, could contribute to increase the overall performance using
more threads to load pathway data. To alleviate the computational impact of
pre-processing on the performance of PEA, we are planning to load the databases
only at the startup, keeping them in the main memory.

5 Conclusion

In this article, we have proposed PEA a parallel framework that can man-
age information from multiple pathway databases, to produce more informative
pathway enrichment analysis. Most of the existing pathway enrichment analysis
tools are designed only to use data coming from a single database. PEA instead,
is the first attempt to perform pathway enrichment analysis exploiting infor-
mation coming from different pathway databases in a single experiment. In the
current version of PEA, we incorporated three pathway databases: KEGG, Reac-
tome, SIGNOR, and generic pathways coded in BioPAX (Level 3), with which
to perform enrichment analysis. In our simulation studies, we demonstrated that
222 G. Agapito and M. Cannataro

the proposed parallel approach outperforms the sequential pathway enrichment

analysis methods.
Future work will regard the possibility to make PEA compatible with other
pathway databases. In addition, we will provide an engine able to store the
pathway data on multiple files, contributing to increase the performance of the
loading task as well as the overall performance of PEA. Finally, PEA can be
implemented using the distributed computing model, based on message passing
to better exploit the performance of distributed architecture as well as multi
CPUs or multi-cores architectures.

References
1. Cerami, E.G., et al.: Pathway commons, a web resource for biological pathway
data. Nucleic Acids Res. 39, D685–D690 (2010)
2. Demir, E., et al.: The BioPAX community standard for pathway data sharing. Nat.
Biotechnol. 28(9), 935 (2010)
3. Fabregat, A., et al.: The reactome pathway knowledgebase. Nucleic Acids Res.
46(D1), D649–D655 (2017)
4. Fabregat, A., et al.: Reactome pathwayanalysis: a high-performance in-memory
approach. BMC Bioinformatics 18(1), 142 (2017). https://doi.org/10.1186/s12859-
017-1559-2
5. Gu, Z., Wang, J.: CePa: an R package for finding significant pathways weighted
by multiple network centralities. Bioinformatics 29(5), 658–660 (2013)
6. Kanehisa, M., Goto, S.: KEGG: kyoto encyclopedia of genes and genomes. Nucleic
Acids Res. 28(1), 27–30 (2000)
7. Liu, D., et al.: Pathway enrichment analysis approach based on topological struc-
ture and updated annotation of pathway. Briefings in Bioinf. 20(1), 168–177 (2017)
8. Perfetto, L., et al.: SIGNOR: a database of causal relationships between biological
entities. Nucleic Acids Res. 44(D1), D548–D554 (2015)
9. Rahmati, S., Abovsky, M., Pastrello, C., Jurisica, I.: pathDIP: an annotated
resource for known and predicted human gene-pathway associations and pathway
enrichment analysis. Nucleic Acids Res. 45(D1), D419–D426 (2016)
10. Tarca, A.L., et al.: A novel signaling pathway impact analysis. Bioin-
formatics 25(1), 75–82 (2009). https://doi.org/10.1093/bioinformatics/btn577.
https://www.ncbi.nlm.nih.gov/pubmed/18990722
 Hierarchical Clustering of Spatial
Urban Data

Eugenio Cesario1,2 , Andrea Vinci2(B) , and Xiaotian Zhu1

1
Monmouth University, West Long Branch, NJ, USA
{ecesario,s1201934}@monmouth.edu
2
ICAR-CNR, Rende, Italy
[email protected]

Abstract. The growth of data volume collected in urban contexts opens

up to their exploitation for improving citizens’ quality-of-life and city
management issues, like resource planning (water, electricity), traffic, air
and water quality, public policy and public safety services. Moreover,
due to the large-scale diffusion of GPS and scanning devices, most of
the available data are geo-referenced. Considering such an abundance
of data, a very desirable and common task is to identify homogeneous
regions in spatial data by partitioning a city into uniform regions based
on pollution density, mobility spikes, crimes, or on other characteristics.
Density-based clustering algorithms have been shown to be very suitable
to detect density-based regions, i.e. areas in which urban events occur
with higher density than the remainder of the dataset. Nevertheless, an
important issue of such algorithms is that, due to the adoption of global
parameters, they fail to identify clusters with varied densities, unless the
clusters are clearly separated by sparse regions. In this paper we provide
a preliminary analysis about how hierarchical clustering can be used to
discover spatial clusters of different densities, in spatial urban data. The
algorithm can automatically estimate the area of data having different
densities, it can automatically estimate parameters for each cluster so as
to reduce the requirement for human intervention or domain knowledge.

1 Introduction

In our days, we are experiencing the most rapid growth in urbanization in history.
The population which lives in cities is growing from 2.86 billion in 2000 to 4.98
billion in 2030, according to a United Nations report. Thus, the 60% of people in
the world is going to live in cities by 2030. Such rapid urbanization is bringing
significative environmental, economic and social changes, and also raises new
issues in city management, related to public policy, safety services, resource
management (electricity, water) and air pollution.
Moreover, a large-scale diffusion of scanning devices, gps, and image process-
ing leads to an abundance of geo-referenced data. Furthermore, more and more
Point of Interest (POI) databases are created which annotate spatial objects
with categories, e.g. buildings are identified as restaurants, and systems, such as
c Springer Nature Switzerland AG 2020
Y. D. Sergeyev and D. E. Kvasov (Eds.): NUMTA 2019, LNCS 11973, pp. 223–231, 2020.
https://doi.org/10.1007/978-3-030-39081-5_20
224 E. Cesario et al.

Google Earth, already fully support the visualization of POI objects on maps.
Considering such an abundance of data, it is becoming crucial their acquisi-
tion, integration, and analysis of big and heterogeneous urban information to
tackle the major issues that cities face today, including air pollution, energy
consumption, traffic flows, human mobility, environmental preservation, com-
mercial activities and savings in public spending [6].
As more and more data become available for a spatial area, it is desirable
to identify different functions and roles which different parts of this spatial area
play; in particular, a very desirable and common task is to identify homogeneous
regions in spatial data and to describe their characteristics, creating high-level
summaries for spatial datasets which are valuable for planners, scientists, and
policy makers. For example, environmental scientists are interested in partition-
ing a city into uniform regions based on pollution density and on other environ-
mental characteristics. Similarly, city planners might be interested in identifying
uniform regions of a city with respect to the functions they serve for the people
who live in or visit this part of a city. Furthermore, policy officers are interested
in detecting high crime density areas (or crime hotspots), to better control the
city territory in terms of public safety.
Among several spatial analysis approaches, density-based clustering algo-
rithms have been shown to be very suitable to detect density-based regions, i.e.
areas in which urban events (i.e., pollution peaks, traffic spikes, crimes) occur
with higher density than the remainder of the dataset. In fact, they can detect
dense regions within a given geographical area, where shapes of the detected
regions are automatically traced by the algorithm without any pre-fixed division
in areas. Also, they can find arbitrarily shaped and differently sized clusters,
which are considered to be the dense regions separated by low-density regions.
Moreover, density-based clustering requires no prior information regarding the
number of clusters, which is another positive benefit that makes such method-
ology suitable for these cases.
An important issue of such density-based algorithms is that, due to the adop-
tion of global parameters, they fail to identify clusters with varied densities,
unless the clusters are clearly separated by sparse regions. In fact, they can
result in the discovery of several small non significant clusters that actually do
no represent dense regions, or they can discover a few large regions that actually
are no longer dense as well. In this paper we provide a preliminary analysis about
how hierarchical clustering can be used to discover spatial clusters of different
densities, in spatial urban data. The algorithm can automatically estimate the
area of the data having different densities, it can automatically estimate param-
eters for each cluster so as to reduce the requirement for human intervention or
domain knowledge.
The rest of the paper is organized as follows. Section 2 reports the most
important approaches in spatial clustering literature exploiting different densities
and the most representative projects in that field of research. Section 3 presents
the proposed algorithm by describing its main steps. Section 4 describes a pre-
 Hierarchical Clustering of Spatial Urban Data 225

liminary experimental evaluation, performed on a real-world case study. Finally,

Sect. 5 concludes the paper and plans future research works.

2 Related Works

In urban datasets the detection of areas in which events occur with higher density
than the remainder of the dataset is becoming a more and more desirable and
common task [2,3,5,9,10]. To this purpose, several density-based approaches
have been used and we briefly describe here some of the most representative
ones.
DBSCAN [7] is the classic density-based algorithm proposed in literature,
which builds clusters by grouping data points that are sufficiently dense, where
the density associated with a point is obtained by counting the number of points
in a region of specified radius  around this point. The main issue of this algo-
rithm is that it does not perform well under multi-density circumstance and
requires much subjective intervention in the parameter estimation.
OPTICS [1] (Ordering Points to Identify the Clustering Structure) is an
enhanced method upon DBSCAN, which creates an ordering of the objects
augmented by reachability distance and makes a reachability-plot out of this
ordering. The reachability-plot, which contains information about the intrinsic
clustering structure, is the basis for interactive cluster analysis, but it does not
produce the clustering result explicitly.
DBSCAN-DLP (Multi-density DBSCAN based on Density Levels Partition-
ing) [11] partitions a dataset into different density level sets by analyzing the
statistical characteristics of its density variation, and then estimates  for each
density level set. Finally, DBSCAN clustering is performed on each density level
set with corresponding  to get clustering results.
GADAC (A new density-based scheme for clustering based on genetic algo-
rithm) has been proposed in [8] to determine appropriate parameters for
DBSCAN. It exploits a genetic algorithm to find clusters of varied densities,
by selecting several radius values.
The VDBSCAN [9] (Varied Density Based Spatial Clustering of Applications
with Noise) is an algorithm which detects clustering models at different values of
densities. Specifically, the approach computes k-dist value for each object (i.e., th
minimum distance such that k points are included in the object’s neighborhood)
and sorts them in ascending order, then make a visualization of the sorted values.
The sharp changes in the k-dist plot correspond to a list of radiuses for different
density varied clusters.
KDDClus [10] is another density-based clustering approach for an automatic
estimation of parameters, to detect clusters of varied densities implementing
a bottom-up approach. It estimates the density of a pattern by averaging the
distances of all its k-nearest neighbors, and uses 1-dimension clustering on these
density values to get a partition of different levels. Then, the algorithm detects
radiuses for different densities, and finally DBSCAN is run to find out clusters
of varied densities.
226 E. Cesario et al.

3 A Framework for Hierarchical Clustering of Urban

Spatial Data

This section describes the main steps of the algorithm we exploit to perform
hierarchical density-based clustering on spatial urban data. Let be D a dataset
collecting spatial urban data instances, D =< x1 , x2 , . . . , xN >, where each xi
is a data tuple described by < latitude, longitude >, i.e. coordinates of the
place event occurs. The meta-code of the approach, based on the algorithm
proposed in [11], is reported in Fig. 1. The algorithm receives in input the dataset
D, and returns the discovered knowledge models, i.e., a set of spatial clusters
DR = {DR1 , . . . , DRK } of different densities.
The algorithm begins by computing, for each point xi , the k-nearest neigh-
bor distance of xi , given a certain k. This is performed by the Compute-K-
Dist(K, D) method, which computes the distance between each xi ∈ D and
its k th -nearest neighbor (line L1). It is worth noting that the k-nearest neigh-
bor distance value of a certain point xi can indicate its density appropriately
(for more details, see [11]): higher such a distance, lower is the density of
points around xi . As soon as this step is completed, the Compute-Density-
Variation(K-Dist-List) method computes the density variation of each point
pi with respect to pj (with i = j) and returns the density variation list (line
L2). On the basis of the computed density variation values, the Partition-
Density-Variation(Density − V ariation − List, D) method creates a list of
density level sets (line L3): a density level set consists of points whose densi-
ties are approximately the same ([11]), that is, density variations of data points
within the same density level set are lower than τ , where τ is a density varia-
tion threshold. Doing this, a multi-density dataset can be divided into several
density level sets, each of which stands for a density distribution. At this point,
the Compute-Eps-Values() method computes coefficient of variability values
(which are used to compare the dispersion of two sample sets) and scatterness
values, which are suitable to compute the level-turning line for the  values (line
L4). Such values are stored and returned in the  − list, i.e., a list of  values
that are estimated as the best values with respect to the different densities in the
data ([11]). Finally, for each  in the -list, the clustering algorithm is executed
and the discovered clusters are added to the final cluster set (lines L5–L8). All
non-marked points are recognized as noise. The final result consists in a set of
spatial clusters, each one representing a event-dense urban region, detected by
different -value settings (i.e., by different densities).

4 Experimental Results

This section describes a preliminary experimental evaluation of the approach

described above. To do that, we executed different tests on a real-world dataset
collecting geo-referenced crime events occurred in an urban area of Chicago. The
goal of our analysis is to perform a comparative analysis of spatial clustering to
 Hierarchical Clustering of Spatial Urban Data 227

Fig. 1. Multi-density clustering algorithm

detect dense regions of geo-localized urban events, when a unique global density
parameter or multi-density parameters are used.
The data that we used to train the models and perform the experimental
evaluation is housed on Plenario [4], a publicly available data search and explo-
ration platform that was developed (and currently managed) by the University
of Chicago’s Urban Center for Computational and Data. The platform hosts sev-
eral data sets regarding various city events, i.e., traffic crashes, food inspections,
crime events, etc. For the sake of our experimental evaluation, the analysis has
been performed on the ‘Crimes - 2001 to present’ dataset, a collection of crime
events occurred in a large area of Chicago on 2012. The selected area includes
different zones of the city, some growing in terms of population, others in terms of
business activities, with different crime-densities over their territory (so making
it interesting for a comparison between single-density and multi-density spatial
analysis). Its perimeter is about 52 Km and its area is approximately 135 Km2 .
The total number of collected crimes is 100 K, while the average number of
crimes per week is 2,275. The total size of this data set is 123 MB.
Figure 2 shows the number of data points having at least min pts neighbors
within the reachability radius , varying the specific density (pts/m2 ) (in log
scale) and the min pts parameter. Given the min pts and density values, the 
parameter is computed as density = min pts/(π ∗ 2 ). In other words, the chart
shows how many core points are counted considering the chosen  and min pts
parameters. The chart shows how the points density is sensitive to the min pts
parameter, and how min pts and  parameter values impact on the results of
the density-based clustering algorithm. In particular two phenomena can be
observed: (i) the number of data points having a specific density decreases as
min pts increases and (ii) for each min pts value, there is a density limit beyond
which the number of core points does not decrease. The first phenomenon shows
how the min pts value has to be carefully chosen to make valuable the results
of the clustering algorithm. The second phenomenon is due to the nature of
the exploited dataset, where the granularity of spatial coordinates is of 1 foot
(0.3048 m), thus having data points which share the same coordinates, and
228 E. Cesario et al.

having neighbors at a zero distance. The same phenomena can be observed also
in Fig. 3, which depicts the number of core points with respect to  and for
several values of min pts.

Fig. 2. Number of core points, w.r.t. the specific density (pts/m2 ) and min pts.

Fig. 3. Number of core points, w.r.t. the  and min pts parameters.

Figure 4 shows a preliminary result of our analysis, where dense regions

are discovered exploiting global or local density parameter values. In partic-
ular, Fig. 4(a) shows the results achieved by fixing a global  = 150 m and
min pts = 64, where each region is represented by a different colour. Inter-
estingly, the algorithm detects a set of significant regions clearly recognizable
through different colours: a large region (in red) in the central part of the area
along with seven smaller areas (in green, blue and light-blue) on the left and
right side, corresponding to zones with the highest concentration of geo-localized
 Hierarchical Clustering of Spatial Urban Data 229

Fig. 4. Dense regions discovered in an area of Chicago, detected exploiting global or

variable density parameter values. (Color figure online)

events. Many other smaller regions representing very local high-density zones
are distributed in the whole area. When a unique global parameter is used, the
density-based algorithm can fail to identify clusters with varied densities, unless
the clusters are clearly separated by sparse regions. In fact, they can result in
the discovery of several small non significant clusters that actually do no repre-
sent dense regions, or they can discover a few large regions that actually are no
longer dense as well. On the other side, Fig. 4(b) shows the result achieved when
different density parameter values are used. Interestingly, the algorithm detects
a set of eight significant regions clearly recognizable through different colors.
Comparing the two images, we can observe a large red region of the central part
of Fig. 4(a); on the other side, in Fig. 4(b) such a region is split in three clusters,
each one with different characteristics and different densities.
Figure 4(c) shows a density map of the same dataset. There is a set of sev-
eral significant regions clearly recognizable through different color intensities,
230 E. Cesario et al.

where darker (lighter) intensities of blue represent lower (higher) densities of geo-
localized events in the selected area. Moreover, in Fig. 4(b) and (c) we can observe
several high-density areas (in red, blue, cyan, yellow and green in Fig. 4(b); in
the central area, in low-density areas at the top and bottom in Fig. 4(c)), that
group points that can be analyzed as an homogeneous set, and that are not
detected by the global density algorithm.
We can conclude that, when a unique global parameter is used, the density-
based algorithm can fail to identify clusters with varied densities, unless the
clusters are clearly separated by sparse regions. In fact, they can result in the
discovery of several small (and non significant) clusters that actually do no rep-
resent dense regions, or they can discover a few large regions that actually are
no longer dense as well. A multi-density approach seems to be more suitable for
clustering and analyzing urban spatial data, as it can detect regions that are
more meaningful w.r.t. the global density approach.

5 Conclusions and Future Work

Spatial analysis of urban data is becoming a very desirable and common task,
aimed at describing building high-level summaries for spatial datasets which
are valuable for planners, scientists, and policy makers. Density-based clustering
algorithms have been shown to be very suitable to detect density-based regions,
but due to the adoption of global parameters, they fail to identify clusters with
varied densities. This paper has provided a preliminary analysis about a hierar-
chical multi-density clustering algorithm that can be used to discover clusters in
spatial urban data. A preliminary experimental evaluation of the approach, per-
formed on a real-world dataset, has shown a comparative analysis of the results
achieved when a unique global density parameter and multi-density parameters
are used. The initial results reported in the paper are encouraging and show
several benefits when a multi-density approach is used. As future work, we will
investigate this issue more in detail and we will perform a more extensive exper-
imental evaluation.

References
1. Ankerst, M., Breunig, M.M., Kriegel, H.P., Sander, J.: OPTICS: ordering points
to identify the clustering structure. ACM Sigmod Rec. 28, 49–60 (1999)
2. Catlett, C., Cesario, E., Talia, D., Vinci, A.: A data-driven approach for spatio-
temporal crime predictions in smart cities. In: 2018 IEEE International Conference
on Smart Computing (SMARTCOMP), pp. 17–24. IEEE (2018)
3. Catlett, C., Cesario, E., Talia, D., Vinci, A.: Spatio-temporal crime predictions in
smart cities: a data-driven approach and experiments. Pervasive Mob. Comput,
53, 62–74 (2019)
4. Catlett, C., et al.: Plenario: an open data discovery and exploration platform for
urban science. IEEE Data Eng. Bull. 37(4), 27–42 (2014)
 Hierarchical Clustering of Spatial Urban Data 231

5. Cesario, E., Talia, D.: Distributed data mining patterns and services: an architec-
ture and experiments. Concurrency Comput. Pract. Experience 24(15), 1751–1774
(2012)
6. Cicirelli, F., Guerrieri, A., Mastroianni, C., Spezzano, G., Vinci, A. (eds.): The
Internet of Things for Smart Urban Ecosystems. IT. Springer, Cham (2019).
https://doi.org/10.1007/978-3-319-96550-5
7. Ester, M., Kriegel, H.P., Sander, J., Xu, X., et al.: A density-based algorithm for
discovering clusters in large spatial databases with noise. In: KDD, vol. 96, pp.
226–231 (1996)
8. Lin, C.Y., Chang, C.C., Lin, C.C.: A new density-based scheme for clustering based
on genetic algorithm. Fundamenta Informaticae 68(4), 315–331 (2005)
9. Liu, P., Zhou, D., Wu, N.: VDBSCAN: varied density based spatial clustering of
applications with noise. In: 2007 International Conference on Service Systems and
Service Management, pp. 1–4. IEEE (2007)
10. Mitra, S., Nandy, J.: KDDClus: a simple method for multi-density clustering.
In: Proceedings of International Workshop on Soft Computing Applications and
Knowledge Discovery (SCAKD 2011), Moscow, Russia, pp. 72–76. Citeseer (2011)
11. Xiong, Z., Chen, R., Zhang, Y., Zhang, X.: Multi-density dbscan algorithm based
on density levels partitioning. J. Inf. Comput. Sci. 9(10), 2739–2749 (2012)
 Improving Efficiency in Parallel
Computing Leveraging Local
Synchronization

Franco Cicirelli , Andrea Giordano(B) , and Carlo Mastroianni

ICAR-CNR, via P. Bucci, cubo 8/9 C, Rende, CS, Italy

{franco.cicirelli,andrea.giordano,carlo.mastroianni}@icar.cnr.it

Abstract. In a parallel computing scenario, a complex task is typically

split among many computing nodes, which are engaged to perform por-
tions of the task in a parallel fashion. Except for a very limited class
of application, computing nodes need to coordinate with each other
in order to carry out the parallel execution in a consistent way. As a
consequence, a synchronization overhead arises, which can significantly
impair the overall execution performance. Typically, synchronization is
achieved by adopting a centralized synchronization barrier involving all
the computing nodes. In many application domains, though, such kind of
global synchronization can be relaxed and a lean synchronization schema,
namely local synchronization, can be exploited. By using local synchro-
nization, each computing node needs to synchronize only with a subset
of the other computing nodes. In this work, we evaluate the performance
of the local synchronization mechanism when compared to the global
synchronization approach. As a key performance indicator, the efficiency
index is considered, which is defined as the ratio between useful compu-
tation time and total computation time, including the synchronization
overhead. The efficiency trend is evaluated both analytically and through
numerical simulation.

Keywords: Parallel computing · Efficiency · Synchronization ·

Max-Plus Algebra

1 Introduction
In order to parallelize the computation needed to solve a problem, different
portions of the problem are assigned to different computing nodes which process
data in parallel. Important application fields in which parallel computing is of
outmost important to achieve significant improvements in terms of execution
and efficiency are: biology, geology, hydrology, logistics and transportation, social
sciences, smart electrical grids (see [1–4]). An interesting application field where
parallel computing is gaining importance is the urban-computing one: in this
context, it is necessary to analyze as different aspects as the mobility of people
or vehicles, the air quality, the consumption of water and electricity, and so on.
c Springer Nature Switzerland AG 2020
Y. D. Sergeyev and D. E. Kvasov (Eds.): NUMTA 2019, LNCS 11973, pp. 232–242, 2020.
https://doi.org/10.1007/978-3-030-39081-5_21
 Improving Efficiency in Parallel Computing 233

The objective is to improve the quality of the services offered to the citizens
(see [5–7]). Two further emerging fields are the “Internet of Things” (IoT) [8,9]
and some alternatives to the classic paradigm of Cloud Computing, i.e., the Fog
Computing and Edge Computing [10,11], where the computation is naturally
parallel and is brought closer to the user and to the data.
A common classification of parallel computing includes the notion of “embar-
rassingly parallel” [12]. In this case, computation at the single nodes can be per-
formed in isolation. A more common case is when parallel tasks exchange data
during computation, and therefore the advancement of execution at the different
nodes must be coordinated, or synchronized. Synchronization [13] means that, at
certain time instants, one node must wait for the data coming from other nodes
before proceeding to the next piece of computation. In this paper, we consider
the very common case of step-based computation: the computation is split in
work units called “steps”, and synchronization occurs at the end of each step.
It is important to define how and when the computing nodes synchronize with
each other. A useful distinction is between global and local synchronization [14].
Synchronization is global, or all–to–all, when each node can start the execution at
a given step only after the node itself and all the other nodes have completed their
execution of the step before. In many application domains, though, such kind
of global synchronization can be relaxed and another synchronization schema,
namely local synchronization, can be exploited. With local synchronization, each
computing node needs to synchronize only with a subset of the other computing
nodes, which are referred to as “adjacent” or “neighbor” nodes.

Node 1 Node 2 Node 3 Node 4 Node 5 Node 6 Node 7 Node 1 Node 2 Node 3 Node 4 Node 5 Node 6 Node 7

l3(1), l3(1),
T3(1) T3(1)

STEP
1

T3(2) T3(2)
l3(2)

l3(2)
STEP
2

Fig. 1. Dynamics of seven nodes for two steps using global synchronization (left) and
local synchronization (right). The solid vertical lines represent the execution times, the
dashed vertical lines are the waiting times and the horizontal dashed lines represent
the synchronization points.

Figure 1 shows an example of global and local synchronization for a system

composed of seven nodes, for two consecutive steps. In the case of global syn-
chronization, at each step, all the nodes must wait for the slowest one before
advancing to the next step (see figure on the left). More in detail, node 5 is the
slowest at step 1 while node 3 is the slowest at step 2. The scenario of local syn-
chronization is depicted in the right part of Fig. 1, in the case that each node has
234 F. Cicirelli et al.

Local Computation Local Computation Local Computation

for each step

for each step

for each step

Send Updates Send Updates Send Updates
to neighbors to neighbors to neighbors

Wait For Data Wait For Data Wait For Data

from neighbors from neighbors from neighbors

Fig. 2. Execution loop under local synchronization.

two neighbors, named left and right neighbors. Each node advances to the next
step of computation when the node itself and its two neighbors have completed
the execution of the current step.
Figure 2 shows the loop executed by each computing node, at each step,
when adopting the local synchronization pattern. The loop in composed of three
phases: (i) the node executes the local computation, related to the specific region
for the current step; (ii) the node sends data to its neighbour nodes; (iii) the
node waits for the analogous data coming from its neighbours.
The rest of the paper is organized as follows. Section 2 presents a mathe-
matical model for global and local synchronization, which allows to define the
computation time and the overhead time spent for synchronization. For the sake
of simplicity, the model of local synchronization considers the case in which each
node synchronizes only with two neighbors. Anyway, the achieved results are
general and hold also with a greater number of neighbors provided it is a con-
stant value. Section 3 presents an asymptotic analysis that shows that the mean
computation time per step is bounded with local synchronization and unbounded
with global synchronization. Section 4 reports an analysis of efficiency, defined as
the ratio between useful computation time and total computation time, includ-
ing the overhead time due to synchronization. Section 5 reports some numerical
results obtained with simulation, which confirm the provided analytical assess-
ment. Finally, Sect. 6 concludes the paper.

2 A Model for Local and Global Synchronziation

Let us denote by N the number of nodes, by li (k) the time needed by node i,
1 ≤ i ≤ N , to execute the local computation at the step k, and by Ti (k) the
time elapsed at node i from the beginning of the computation (i.e., start of the
step 1) until the end of the step k. As an example, in Fig. 1, assuming that the
values of li (k) are the same for the cases of global and local synchronization, we
show the corresponding execution advancements. It can be seen that the times
Ti (k) tend to be shorter in the case of local synchronization.
For the sake of ease of exposition, we consider a simple case of the model for
which communication times are negligible and each node i needs to synchronize
with two neighbors, indexed as i − 1 and i + 1, except for nodes 1 and N that
have only one neighbor. In addition, we assume that the computation time li (k)
 Improving Efficiency in Parallel Computing 235

depends only on the node i but does not depend on the step k, and that li (k) = li ,
1 ≤ i ≤ N , are i.i.d. random variables with average ¯li .
In the case of global synchronization, each step k begins contemporarily at
each node when the slowest node at the step k−1 has completed its computation.
So we have:
k−1

Ti (k) = max(l1 , . . . , lN ) + li , k ≥ 1, (1)
j=1
In the case of local synchronization we have the recursive equation:
Ti (k) = max(Ti (k − 1), Ti−1 (k − 1), Ti+1 (k − 1)) + li , k ≥ 1, 1 ≤ i ≤ N. (2)

in which we assume Ti (0) = 0, 1 ≤ i ≤ N and T0 (k) = TN +1 (k) = 0, k ≥ 0.

In both cases, the random sequence {T (k)/k, k ≥ 1}, where {T (k) =
(T1 (k), . . . TN (k))} falls into the framework of stochastic equations as described
in [15]. Using the results from [15] it can be shown that when N is finite and
{l(k) = (l1 (k), . . . , lN (k)), k ≥ 1} are independent, the sequence {T (k)/k, k ≥
1} is ergodic and stable.

2.1 Computation and Overhead Times

In the following analysis, it is appropriate to distinguish between the computa-
tion time (the time spent by nodes for useful computation) and the overhead
time (the waiting time needed for synchronization).
The computation time at step k, defined as the sum of computation times of
all the nodes at step k, and denoted as tS (k), is equal to:
tS (k) = l1 (k) + l2 (k) + · · · + lN (k) (3)
The average value of the computation step is:
t̄S (k) = ¯l1 + ¯l2 + · · · + ¯lN = N · ¯l (4)
where ¯l is the average of the average computation times ¯li of the N nodes.
With the assumptions specified above, t̄S (k) is the same at each step.
The overhead time at step k is defined as the sum of the times oi (k) spent by
each node for synchronization, after completing the computation at step k − 1
and before starting the computation at step k. It is equal to:
tO (k) = o1 (k) + o2 (k) + · · · + oN (k) (5)
In the case of global synchronization, the expression for the overhead time of
node i at step k, oi (k), must take into account the fact that node i needs to wait
until all the other nodes have finished their computation at step k − 1. Since
all the nodes start their computation precisely at the same time, the overhead
experienced by node i at step k is equal to:
oi (k) = max (lj ) − li (6)
j=1..N

and the expression for the overall overhead time is:

236 F. Cicirelli et al.

N
  
tO (k) = oi (k) = N · max lj − tS (k) (7)
j=1..N
i=1

In the case of local synchronization, oi (k) is the time that node i needs to
wait until its two neighbors have completed their computation of step k − 1, and
is equal to:
 
oi (k) = max Ti−1 (k − 1), Ti (k − 1), Ti+1 (k − 1) − Ti (k − 1) (8)

Now that we have defined the computation and overhead time for both global
and local synchronization, we define as tL (k) = tS (k) + tO (k) the full length of
the step k, including both the actual computation and the overhead. Recalling
the definition of Ti (k), it follows that the sum of the full lengths of computation
tL (k), from step 1 up to step k, is equal to the sum of Ti (k) for all the nodes i:
N
 k
 k−1
 N

Ti (k) = tL (j) = tL (j) + tL (k) = Ti (k − 1) + tL (k) (9)
i=1 j=1 j=1 i=1

From the above expression we find another expression for tL (k):

N
 N

tL (k) = Ti (k) − Ti (k − 1) (10)
i=1 i=1

3 Asymptotic Analysis of Global and Local

Synchronization
In the following, analytical results are given showing that with global synchro-
nization the mean computation time per step is unbounded when the number of
nodes increases. Conversely, with local synchronization the mean computation
time per step is bounded irrespective of the number of nodes. In the follow-
ing, the mean computation time per step is also referred to as the cycle time,
Tcycle = limk→∞ k1 E(Ti (k)).

3.1 Analysis of Global Synchronization

With global synchronization, represented in (1), the analysis can be performed

by exploiting the well-known results coming from order statistics and extreme
value theory [16–18]. In the case that the random variable li has a continuous
distribution with the support1 on a semi-infinite interval, we can state that, as

1
In the case that li are independent (not necessarily identically distributed) random
variables with a continuous distribution having support on a bounded interval, the
mean computation time limk→∞ E(Ti (k))/k is always a constant, irrespective of the
number of nodes N .
 Improving Efficiency in Parallel Computing 237

the number of nodes N grows, the cycle time limk→∞ k1 E(Ti (k)) grows as well,
and in the limit as N → ∞ we have that:
1
lim E(Ti (k)) = E(max(l1 , . . . , lN )) → ∞.
k→∞ k

For example, in the case that li are i.i.d random variables distributed exponen-
tially, having parameter λ and average 1/λ, for sufficiently large N we have2 :
1
lim E(Ti (k)) ≈ (1/λ) (ln N + 0.5772)
k→∞ k
Similar expressions are not available in the general case, but hold for a large
class of distributions. Such a class includes the distributions [19] having pure-
exponential and non-pure exponential tails, like the gamma distribution.
With global synchronization, if the computation times are random with
unbounded support, then as the number of nodes N increases, on average, the
time to complete the next computation step also becomes unbounded.

3.2 Analysis of Local Synchronization

With local synchronization, a node can start the next computation step after a
finite number of neighbors finish their computations of the current step. In this
scenario, a different conclusion is obtained: the cycle time limk→∞ k1 E(Ti (k))
becomes bounded irrespective of the number of nodes N . In what follow, for
simplicity we consider the case of synchronization with only two neighbors, i.e.,
the scenario described in (2). It is convenient to describe the model (2) by adopt-
ing the framework of discrete event dynamic systems, and the evolution can be
captured in terms of the Max-Plus algebra. Subsequently, the well-known results
of the Max-Plus theory can be used to study the cycle time limk→∞ k1 E(Ti (k)).
The Max-Plus theory defines an algebra over the vector space R ∪ {−∞}
with the operation x ⊕ y defined as max(x, y) and the operation x ⊗ y defined
as x + y.
The vector T (k) = (T1 (k), . . . , TN (k)) collects the values of Ti (k) for each
node i at the step k, where T (0) is the column-vector of zeros. The N ×N matrix
M (k) = [M (k)]ij , 1 ≤ i, j ≤ N , is defined as:
⎛ ⎞
l1 l1 −∞ −∞ . . . −∞ −∞
⎜ l2 l2 l2 −∞ . . . −∞ −∞ ⎟
⎜ ⎟
⎜−∞ l3 l3 l3 . . . −∞ −∞ ⎟
⎜ ⎟
⎜ ⎟
M(k) = ⎜−∞ −∞ l4 l4 . . . −∞ −∞ ⎟ .
⎜ .. .. .. .. . . . .. ⎟
⎜ . . . . . ... ⎟
⎜ ⎟
⎝−∞ −∞ −∞ −∞ . . . lN −1 lN −1 ⎠
−∞ −∞ −∞ −∞ . . . lN lN
2
It is well-known that the right part is the approximation of the expected
 maximum
−1
of N i.i.d. random variables with exponential distribution equal to µ Ni=1 i , with
N −1
i=1 i being the N harmonic number.
th
238 F. Cicirelli et al.

With this notation, equation (2) can be rewritten as

T (k + 1) = M(k) ⊗ T (k), k ≥ 0, (11)

where the matrix-vector product is defined by:

[M(k) ⊗ T (k)]i = max ([M(k)]ij + Tj (k))

1≤j≤N

Indeed, this expression equals to:

[T (k + 1)]i = [M(k) ⊗ T (k)]i = li ⊗ Ti−1 (k) ⊕ li ⊗ Ti (k) ⊕ li ⊗ Ti+1 (k) =

max(Ti−1 (k) + li , Ti (k) + li , Ti+1 (k) + li ) = max(Ti−1 (k), Ti (k), Ti+1 (k)) + li

where T0 (k) = TN +1 (k) = 0.

Now we can make use of the well-known asymptotic results from the Max-
Plus theory (see, for example, [20]). The matrix T(k) has at least one finite
entry on each row, which is the necessary and sufficient condition for Tj (k)
to be finite. From [20, Lemma 6.1] we find that there exists γ > 0 such that
limk→∞ k1 E(Ti (k)) = γ, 1 ≤ i ≤ N . In the case when li are stochastically
bounded by a single variable (say L), having moment generating function (say
L(s)), the upper bound for the value of γ is (see [20, Proposition 6.2]):

γ ≤ inf{x > E(L) such that M (x) > ln 3}, (12)

where M (x) = supθ∈R (θx − ln L(θ)) and E(L) = L (0).

This result tells us that in the case of local synchronization the mean com-
putation time remains finite for any number of nodes N .

4 Estimation of Efficiency
The aim of this section is to show that global and local synchronizations exhibit
very different behaviors in terms of efficiency, where the efficiency is the ratio
of the useful computation and the total time (useful computation time plus
overhead time).
To show this let us start defining the overall computation and overall over-
head times from the beginning of execution up to step k:
k
 k

TS (k) = tS (j) TO (k) = tO (j) (13)
j=1 j=1

where tS (k) and tS (k) are defined in (3) and (5). The efficiency is defined as:

T̄S (k) 1
Ef (k) = = (14)
T̄S (k) + T̄O (k) 1 + T̄O (k)/T̄S (k)
We are interested in the efficiency when both k and N grow indefinitely.
In the following subsections we see that:
 Improving Efficiency in Parallel Computing 239

– with global synchronization, the efficiency tends to zero as the number of

nodes grows;
– with local synchronization, the efficiency has a non-zero lower bound.

The consequences of these different trends are remarkable not only with very
large values of N, but also for ordinary values of N, which means that the use of
local synchronization leads to significant improvements in practical scenarios.

Efficiency with Global Synchronization. With global synchronization, the

ratio T̄O (k)/T̄S (k) can be derived from (4), (7) and (13) and is equal to:
 
T̄O (k) k · N · E maxj=1..N ¯lj − T̄S (k)
= (15)
T̄S (k) T̄S (k)

From (4) we see that T̄S (k) = k · N · ¯l, and the above expression becomes:
 
T̄O (k) E maxj=1..N ¯lj
= ¯l −1 (16)
T̄S (k)

The value of the denominator is clearly a bounded value, for each value of
N, since it is the average of N finite values. On the other hand, the value of the
numerator is unbounded, as N increases, for a wide set of probability distribu-
tions that can be assumed to characterize the computation time li . It follows that
the ratio (16) is unbounded and that the efficiency – see expression (14) – tends
to zero when increasing the number of nodes. In other words, when N increases,
the overhead becomes prevalent with respect to the actual computation time.
For example, if the li are i.i.d. and have an exponential distribution, with
parameter λ and average 1/λ, the maximum of N such distributions is known
to be approximated by the expression 1/λ · H(N )
1/λ · ln(N ), where H(N ) is
the harmonic number i=1..N (1/i).
In this case, the efficiency can be expressed as:
1
Ef (k)
(17)
ln(N )

This expression means that the efficiency decreases and tends to zero as
the number of nodes increases. When the local computation time is distributed
with any other distribution of the exponential family, the expression for the
maximum of a set of i.i.d random variables will be different but the trend will
still be logarithmic.

Efficiency with Local Synchronization. In the case of local synchronization,

the value of efficiency for k → ∞ can be computed as:
240 F. Cicirelli et al.

T̄S (k)
Ef = lim Ef (k) = lim = (18)
k→∞ k→∞ T̄O (k) + T̄S (k)
k
1/k · j=1 t̄S (j)
lim N = (19)
k→∞ 1/k ·
i=1 E(Ti (k))
1/k · k · N · ¯l
lim N = (20)
k→∞
i=1 E(Ti (k))/k
N · ¯l
N = (21)
i=1 limk→∞ E(Ti (k))/k
N · ¯l ¯l
≥ (22)
N · Tcycle γ

The expression (19) comes by applying expression (9) at the denominator,

and then multiplying the numerator and the denominator by 1/k. To obtain
expression (20) we apply (4), and in (21) we exchange the limit and the summa-
tion at the denominator. To obtain the left side of (22) we rely on Max-Plus that
ensures that the value of the limit, the cycle time, is the same at each node, see
Sect. 3.2. The final inequality applies because we know by the Max-Plus theory
that the cycle time Tcycle is finite and has un upper bound γ, i.e., Tcycle ≤ γ.
The conclusion is that, in the case of local synchronization, the efficiency has
a non-zero lower bound, irrespective of the number of nodes.

5 Numerical Results
In the following we report some numerical results both for global and local
synchronizations under negligible communication times and i.i.d. computation
times li . We used Matlab to simulate the computational behavior modeled by (1)
and (2) with different values of the number of nodes N . For li we considered many
types of distributions that are well-suited workload models in parallel computing
systems [21] (exponential, hyper-gamma, lognormal, etc.). We report here the
results obtained with the exponential distribution. However, we found that the
behavior described in the following applies also with the other distributions.
The performance is assessed by computing the efficiency using a single sim-
ulation run with k = 10000 and the batch-means method. We consider the case
in which li has an exponential distribution with the mean equal to 1.0. Figure 3
shows the values of the achieved efficiency versus the number of nodes N in the
case of global and local synchronization. The figure also reports the bound on
the efficiency obtained with the Max-Plus algebra (see Expression (22)) and the
value of efficiency obtained for N = 1000 nodes.
The experimental values are consistent with the theoretical bounds discussed
in the previous sections, and it can be seen that the use of local synchronization
allows the efficiency to be notably increased with respect to global synchroniza-
tion. In particular, it is confirmed that with global synchronization the value
of efficiency decreases towards zero, while with local synchronization the effi-
ciency is always higher than the theoretical bound, which in this case is 0.3.
 Improving Efficiency in Parallel Computing 241

0.8
local synch
global synch
local synch max-plus
0.6
Efficiency
local synch N=1000

0.4

0.2

0
0 10 20 30 40 50 60 70 80 90 100
No. of nodes, N

Fig. 3. Values of efficiency as function of N under local and global synchronization.

The local computation times have exponential distribution. The plot shows the bound
obtained with Max-Plus algebra and the numerical value with N = 1000.

From the figure it can be seen that a “practical” bound equal to about 0.41 is
already reached when the number of nodes is 50 or larger. It is also interesting
to notice that the bound on efficiency derived with Max-Plus algebra, although
underestimated, is still larger than the value of efficiency obtained with global
synchronization with 20 or more nodes.
From the unreported experiments performed by using different distributions
of the computation time we have seen that the advantage of local synchronization
increases with the variance of the adopted distribution.

6 Conclusions
In this paper, we analyzed and compared the performance of two different
approaches for synchronizing the execution of parallel computation, namely local
synchronization and global synchronization. The efficiency, which is given by the
ratio of useful computation time and the total computation time, including the
overhead, is used as a key performance indicator. We found that, as the number
of nodes increases, the efficiency tends to zero in the case of global synchroniza-
tion, while it has a non-zero lower bound in the case of local synchronization.
This result was achieved by using the Max-Plus algebra, and it was confirmed
experimentally though numerical simulation.

References
1. Cicirelli, F., Forestiero, A., Giordano, A., Mastroianni, C.: Transparent and efficient
parallelization of swarm algorithms. ACM Trans. Auton. Adapt. Syst. 11(2), 14
(2016)
2. Cicirelli, F., Forestiero, A., Giordano, A., Mastroianni, C.: Parallelization of space-
aware applications: modeling and performance analysis. J. Netw. Comput. Appl.
122, 115–127 (2018)
242 F. Cicirelli et al.

3. Gong, Z., Tang, W., Bennett, D.A., Thill, J.C.: Parallel agent-based simulation
of individual-level spatial interactions within a multicore computing environment.
Int. J. Geogr. Inf. Sci. 27(6), 1152–1170 (2013)
4. Tang, W., Bennett, D.A., Wang, S.: A parallel agent-based model of land use
opinions. J. Land Use Sci. 6(2–3), 121–135 (2011)
5. Mastroianni, C., Cesario, E., Giordano, A.: Efficient and scalable execution of
smart city parallel applications. Concurrency Comput. Pract. Experience 30(20),
e4258 (2018)
6. Zheng, Y., Capra, L., Wolfson, O., Yang, H.: Urban computing: concepts, method-
ologies, and applications. ACM Trans. Intell. Syst. Technol. 5(3), 1–55 (2014)
7. Blecic, I., Cecchini, A., Trunfio, G.A., Verigos, E.: Urban cellular automata with
irregular space of proximities. J. Cell. Automata 9(2–3), 241–256 (2014)
8. Atzori, L., Iera, A., Morabito, G.: The internet of things: a survey. Comput. Netw.
54(15), 2787–2805 (2010)
9. Lee, I., Lee, K.: The Internet of Things (IoT): applications, investments, and chal-
lenges for enterprises. Bus. Horiz. 58(4), 431–440 (2015)
10. Krishnan, Y.N., Bhagwat, C.N., Utpat, A.P.: Fog computing–network based cloud
computing. In: 2nd IEEE International Conference on Electronics and Communi-
cation Systems (ICECS), pp. 250–251 (2015)
11. Hu, P., Dhelim, S., Ning, H., Qiu, T.: Survey on fog computing: architecture,
key technologies, applications and open issues. J. Netw. Comput. Appl. 98, 27–42
(2017)
12. Ekanayake, J., Fox, G.: High performance parallel computing with clouds and cloud
technologies. In: Avresky, D.R., Diaz, M., Bode, A., Ciciani, B., Dekel, E. (eds.)
CloudComp 2009. LNICSSTE, vol. 34, pp. 20–38. Springer, Heidelberg (2010).
https://doi.org/10.1007/978-3-642-12636-9 2
13. Fujimoto, R.: Parallel and Distributed Simulation Systems. Wiley, New York (2000)
14. Cicirelli, F., Forestiero, A., Giordano, A., Mastroianni, C., Razumchik, R.: Global
and local synchronization in parallel space-aware applications. In: 32nd European
Conference on Modelling and Simulation, ECMS 2018, Wilhelmshaven, Germany,
May 2018
15. Borovkov, A.A.: Ergodicity and stability theorems for a class of stochastic equa-
tions and their applications. Theory Probab. Appl. 23(2), 227–247 (1979)
16. David, H.A., Nagaraja, H.N.: Order Statistics, 3rd edn. Wiley, New York (2003)
17. Ang, A.S., Tang, W.: Probability Concepts in Engineering Planning and Design,
vol. II. Rainbow Bridge (1984)
18. Madala, S., Sinclair, J.B.: Performance of synchronous parallel algorithms with
regular structures. IEEE Trans. Parallel Distrib. Syst. 2(1), 105–116 (1991)
19. Whitt, W., Crow, C., Goldberg, D.: Two-moment approximations for maxima.
Oper. Res. 55(3), 532–548 (2007)
20. Baccelli, F., Konstantopoulos, P.: Estimates of cycle times in stochastic petri nets.
In: Karatzas, I., Ocone, D. (eds.) Applied Stochastic Analysis, pp. 1–20. Springer,
Heidelberg (1992). https://doi.org/10.1007/BFb0007044
21. Lublin, U., Feitelson, D.G.: The workload on parallel supercomputers: modeling
the characteristics of rigid jobs. J. Parallel Distrib. Comput. 63(11), 1105–1122
(2003)
 A General Computational Formalism
for Networks of Structured Grids

Donato D’Ambrosio1(B) , Paola Arcuri1 , Mario D’Onghia1 ,

Marco Oliverio1 , Rocco Rongo1 , William Spataro1 , Andrea Giordano2 ,
Davide Spataro3 , Alessio De Rango4 , Giuseppe Mendicino4 ,
Salvatore Straface4 , and Alfonso Senatore4
1
DEMACS, University of Calabria, Rende, Italy
{donato.dambrosio,rocco.rongo,william.spataro}@unical.it
[email protected], [email protected], [email protected]
2
ICAR-CNR, Rende, Italy
[email protected]
3
ASML, Eindhoven, The Netherlands
[email protected]
4
DIATIC, University of Calabria, Rende, Italy
{alessio.derango,giuseppe.mendicino,
salvatore.straface,alfonso.senatore}@unical.it

Abstract. Extended Cellular Automata (XCA) represent one of the

most known parallel computational paradigm for the modeling and sim-
ulation of complex systems on stenciled structured grids. However, the
formalism does not perfectly lend itself to the modeling of multiple
automata were two or more models co-evolve by interchanging informa-
tion and by synchronizing during the dynamic evolution of the system.
Here we propose the Extended Cellular Automata Network (XCAN) for-
malism, an extension of the original XCA paradigm in which different
automata are described by means of a graph, with vertices representing
automata and inter-relations modeled by a set of edges. The formalism
is applied to the modeling of a theoretical 2D/3D coupled system, where
space/time variance and synchronization aspects are pointed out.

Keywords: Extended Cellular Automata Network · Modeling ·

Stenciled structured grid · Direct acyclic graph · Space/time
granularity · Synchronization

1 Introduction

Mesh algorithms are widely used in Science and Engineering for modeling and
simulating a wide variety of complex phenomena, e.g. in Computational Fluid
Dynamics [2,17]. In this field, even if unstructured grids are gaining popularity
(see e.g. [38,43]), structured grids are still the most utilized since they generally
require less computing power and memory, often allow for better convergence,

c Springer Nature Switzerland AG 2020

Y. D. Sergeyev and D. E. Kvasov (Eds.): NUMTA 2019, LNCS 11973, pp. 243–255, 2020.
https://doi.org/10.1007/978-3-030-39081-5_22
244 D. D’Ambrosio et al.

besides being more suitable for the execution on modern many-core accelerators
like GPUs due to data locality [3,7–9,28,39].
Cellular Automata probably represent one of the most known example of
parallel computational paradigms for stenciled structured grids (i.e., for grids on
which a spatial neighborhood relation is defined). Originally proposed by John
von Neumann to study self-reproduction in biological systems [34], the model
is based on an infinite grid of cells with integer coordinates in which each cell
changes state based on a local law of evolution (transition function). Cellular
Automata are universal computational models [4,10,42], widely adopted in both
theoretical studies [29,35,44], and applications [1,21,27,31]. Nevertheless, they
do not straightforwardly lend themselves to the modeling of some complex sys-
tems like, for instance, lava flows [11]. As a matter of fact, in this case it is not
straightforward to simulate lava feeding to vents in terms of local interactions, or
to perform evaluations regarding the global state of the system, for instance for
evaluating if the phenomenon evolution is terminated (i.e., if lava has solidified
everywhere). For this reason, Di Gregorio and Serra proposed an extension of
the original von Neumann’s cellular model, known as Multi-component, Com-
plex or even Extended Cellular Automata (XCA) [26]. Even if computationally
equivalent to the original CA, XCA permit to relax the locality constraint when
it is advantageous. In addition, it permits to split both the cell state in substates,
and the transition function in elementary processes or kernels. For this reason,
they were widely adopted for the simulation of different complex natural phe-
nomena [13,14,16,18,19,25,30,36,40,40]. Furthermore, the formalism is general
and other numerical methods, like for instance Explicit Finite Differences, can
be seen as a particular case of XCA (see e.g. [6,12,20,23,33]). Different software
systems were also developed from the XCA formalism. Among them, well known
examples are Camelot [15], libAuToti [41], OpenCAL [12,37], and vinoAC [24]1 .
Cellular Automata Networks were also proposed [5], albeit they essentially rep-
resent an alternative formulation of XCA, where the network defines the order
of application of the considered local/global laws.
However, though powerful and highly flexible, the computational paradigms
cited above are not straightforwardly applicable for modeling the co-evolution
of automata of different dimensions and space/time granularity, as pointed out
in [32], where automaton-automaton interactions were explicitly defined and
implemented. Specifically, they considered a coupled system composed by a two-
dimensional model simulating water runoff on a topographic surface, and a three-
dimensional automaton modeling water infiltration and groundwater propaga-
tion, respectively. In such a case, besides the different dimensions, the automata
could have different spacial granularity due to the data sampling methods
adopted. For instance, a LIDAR technique could be used for the high-resolution
sampling of the upper topographic surface elevations, while core drilling could be
adopted to (roughly) characterize the underground, resulting in a finer descrip-
tion (and therefore in a smaller grid cell) for the two-dimensional automaton.
In a similar context, a base time step variance could also be observed (with
1
The vinoAC acronym does not explicitly appear in the text.
 A General Computational Formalism for Networks of Structured Grids 245

the finer-grained spacial model generally needing a smaller time step), and a
higher level structure required for coordinating the dynamical co-evolution of
the automata [22].
In order to address the above issues, here we propose the Extended Cellular
Automata Network (XCAN) formalism for modeling a network of co-evolving
automata. XCAN defines the interfaces between interconnected automata, man-
ages the dynamical evolution of each model belonging to the network and their
synchronization, by taking into account the space/time granularity variance
within the network. The paper is organized as follows: Sect. 2 provides an alter-
native formulation of the XCA computational paradigm, by introducing some
features that are considered useful in the modeling of different complex systems;
Sect. 3 introduces XCAN by providing the formal definition; Sect. 4 describes a
simplified, hypothetical, example of application; Sect. 5 concludes the paper with
a general discussion and future developments.

2 Extended Cellular Automata

An Extended Cellular Automata (XCA) is defined as a tuple:

 
A = (D, S, I), X, P, Q, (Σ, Φ), R, Γ, (T, t), ω
where the round parentheses are used to group related objects, and the D, X,
Q, and (Σ, Φ) objects are mandatory. The meaning of each model object is
described below.

– D = [ι1min , ι1max ] × . . . × [ιdmin , ιdmax ] ⊂ Zd is the d-dimensional discrete

(integer
 coordinates) computational domain. 
– S = x = (x1 , x2 , . . . , xd ) ∈ Rd | μ : D → Rd , μ(ι) = x is the d-dimensional
(real coordinates) domain corresponding to the discrete domain D under the
function μ, this latter associating cells in D to points in Rd .
– I = {Ii = [ζi1min , ζi1max ] × . . . × [ζidmin , ζidmax ] ⊆ D | ιijmin ≤ ζijmin , ζijmax ≤
ιijmax | ∀j = 1, 2, . . . , d} is the set of domain interfaces. It can be used to
provide input to the system (e.g., to set boundary conditions), or for reduc-
tion/generalization
 (or aggregation/disaggregation) operations.
– X = Xi = {ξi1 , ξi2 , . . . , ξi|Xi | } | i = 1, 2, . . . , |X|, ξij ∈ Zd ∀j =

1, 2, . . . , |Xi | is the set of neighborhood stencils. The neighborhood of a cell
ι ∈ D under the stencil Xi is defined as:

V (ι, Xi ) = {ι + ξi1 , ι + ξi2 , . . . , ι + ξi|Xi | }

The neighborhood radius along the j th direction is defined as:

 |X| |X| 
rj = (rj− , rj+ ) = min(0, min ξij ), max(0, max ξij ) (1)
i=1 i=1

while the neighborhood radius is defined as r = maxdj=1 {rj− , rj+ }.

246 D. D’Ambrosio et al.

– P = Pi = {pi1 , pi2 , . . . , pi|Pi | } | i = 1, 2, . . . , |P |, pij ∈ QP ∀j =

1, 2, . . . , |Pi | is the set of parameters. It can be used to define space-invariant
values, as well as to store information resulting from non-local operations (see
below).
– Q = Q1 ×. . .×Q|Q| is the set of cell states, expressed as the Cartesian product
of the Qi substate sets. Accordingly, a state q ∈ Q is expressed as:

q = (q1 , q2 , . . . , q|Q| ) | qi ∈ Qi ∀i = 1, 2, . . . , |Q|

and qi ∈ Qi is referred as the ith substate. The cell state function defines cell
state assignment:
c : D → Q | ι ∈ D → q ∈ Q
Eventually, an automaton configuration is defined as the state assignment to
each domain cell:

C = {(ι, q) | q = c(ι), ι ∈ D, q ∈ Q}
 |P | 
– Σ = σ1 , σ2 , . . . , σ|Σ| | σi : QP j × Q|Xk | → Q, j ∈ {1, 2, . . . , |P |} is
the set of local transition functions. The input is given by the states of the
neighboring cells involved into the local transition rule, besides the values
of zero or more parameters. A local transition is classified as an internal
transition if the states of the neighboring cells, except the state of cell itself,
|P |
are irrelevant, i.e., if σi : QP j × Q → Q.
 |P | +
– Φ = φ , φ , . . . , φ|Σ| |φi : QP j × Q|Ik | → Q|Ik | , φi |ι∈Ik =
  1 2  
σi Pj , c V (ι, X) , j ∈ {1, 2, . . . , |P |}, k ∈ {1, 2, . . . , |I|} . Ik+ represents the
set resulting from the union of Ik and its adjacent borders along each direc-
tion, where the border width along each direction is defined by Eq. 1:

Ik+ = [ζk1min − r1− , ζk1max + r1+ ] × . . . × [ζkdmin − rd− , ζkdmax + rd+ ]

φi applies the corresponding local function σi to the non-local domain Ik .

 |P |
– R = ρ1 , ρ2 ,. . . , ρ|R| | ρi : QP j × Q|Ik | → QP , j ∈ {1, 2, . . . , |P |}, k ∈
{1, 2, . . . , |I|} is the set of reduction (or aggregation) functions.
 |P |
– Γ = γ1 , γ2 ,. . . , γ|Γ | | γi : QP j × Q → Q|Ik | , j ∈ {1, 2, . . . , |P |}, k ∈
{1, 2, . . . , |I|} is the set of generalization (or disaggregation) functions.
|Φ|+|R|+|Γ |
– T = ◦i=1 τi | τi ∈ Φ ∪ R ∪ Γ is the function determining the automaton
global transition.
– t ∈ R+ is the time corresponding to state transition of the system (i.e., to the
application of the T function).
– ω : {C | C is a configuration of A} → {0, 1} is the quiescent function. If ω is
true, A is in quiescent state.
 A General Computational Formalism for Networks of Structured Grids 247

Fig. 1. Example of a network composed by four XCA models, namely A, B, C, and

D, each of them characterized by a dimension and a space/time granularity. One-, two
and three-dimensional models are here considered, with different spacial cardinality
and time clock. The edges CA, AB, BA, and DB define dependence relationships of
A from C, of B from A, of A from B, and of b from D, respectively. Please, note that
the edges AB and BA define a two-sided relationship between the automata A and B.

3 Extended Cellular Automata Network

An Extended Cellular Automata Network (XCAN) is a direct acyclic graph
(DAG), where nodes represent XCA models, while edges connections between
couples of automata. An edge AB connecting the automaton A to the automa-
ton B specifies that the state of B (also) depends on the state of A. In the case
a dependence of A from B have to be modeled, a further edge BA can be con-
sidered. Additional components are considered to define topological/functional
relations between connected vertices and to characterize the dynamical evolution
of the network. An example is shown in Fig. 1. An XCAN is formally defined as:

N = <V, E, (I, π), τ, τN , ω>

where
– V = {Ai | i = 1, 2, . . . , |V |, Ai is a XCA} is the set of XCA models belonging
to the network.
– E = {AB | A, B ∈ V } is the set of edges. The notation AB denotes an edge
from A to B, meaning that the state of B is affected by the state of A.
 (A) (B) (A) (B) 
– I = {Ii , Ij } | AB ∈ E, Ii ∈ I (A) , Ij ∈ I (B) is the set of interfaces
of connected automata. The notation I (A) denotes that the set I belongs to
the automaton A.
248 D. D’Ambrosio et al.

(A) (B)
– π = {πAB : P(Ii ) → P(Ij ) | AB ∈ E} is the set of partition functions
that define a bijection between interface cells of one automaton and interface
subsets of the other. Specifically:

⎧ (A) (B) (A) (B) (A) (B)

⎪
⎨ πAB,γ : Ii → P(Ij ) | ι ∈ Ii → H ⊆ Ij if |Ii | < |Ij |
πAB = πAB,σ : Ii(A) → Ij(B) | ι ∈ Ii(A) → ι ∈ Ij(B) (A) (B)
if |Ii | = |Ij |
⎪
⎩ (B) (A) (A) (B)
πAB,ρ = πBA,γ | ι ∈ Ii → H ⊆ Ij if |Ii | > |Ij |

where H is the generic macrocell (i.e., a set of one or more cells) of the
macropartitioned interface.
|H (A) | |H (B) | |H (B) |
– τ = {τAB : Q(A) × Q(B) → Q(B) | AB ∈ E} is the set of
induction (or interaction) functions that, depending on the interface parti-
tioning, are defined as:
⎧ (B)
⎪ (B) |H | |H (B) | (A) (B)
⎪
⎨ τAB,γ : Q × Q
(A)
→ Q(B) if |Ii | < |Ij |
(A) (B)
τAB = τAB,σ : Q(A) × Q(B) → Q(B) if |Ii | = |Ij |
⎪
⎪
⎩ |H (A)
| (A) (B)
τAB,ρ : Q(A) × Q(B) → Q(B) if |Ii | > |Ij |
– τN : N × C (N ) → C (N ) is the network control unit. It determines the network
configuration update by executing the base steps reported in the ordered list
below at discrete iterations:
1. Evaluates the elapsed time that each automaton A ∈ V would have if
evolved for one more step;
2. Evolves the automata having the minimum elapsed time, as evaluated in
the previous base-step, by applying the corresponding global transition
function T ;
3. Applies the induction functions τAB , being B evolved in the current step,
A any other automaton s.t. AB ∈ E.
– ω : V → {true, f alse} is the termination function. When ω is true the net-
work computation halts.
(N )
At step n = 0, the network is in the initial state C0 . τN is then applied
to allow the network evolve, by producing a sequence of configurations
(N ) (N )
C1 , C2 , · · · . After each step, the ω function is applied. Possible halt cri-
teria could be the following:
• A predefined number of iterations have been computed;
• The analyzed process/phenomenon is completed;
• All the automata are in the quiescent state.

4 An Example of Extended Cellular Automata Network

As an example of application, here we consider a conceptual, oversimplified net-
work composed by two XCA, namely a two-dimensional model A and three-
dimensional model B, interconnected by a single edge AB, defining a depen-
dence of B from A (Fig. 2). Interaction between A and B occurs at domain
 A General Computational Formalism for Networks of Structured Grids 249

Fig. 2. Example of a simplified network composed by two XCA models, namely A and
B. The single edge AB connects the two automata by defining a dependence of B from
A. The time corresponding to a discrete step of the two automata is shown, being 3
physical time units for A, and 9 time units for B.

interfaces. Different spacial granularity here induces a macrocell partitioning on

the finer-grained interface, which is that of A (Fig. 3). Moreover, the different
temporal granularity causes that the finer-grained automaton, that is A again,
is updated more frequently. Automata interaction exclusively occurs when B
evolves. In the following, the above outlined conceptual models and the inter-
connection network are defined, by ignoring components and specifications that
are not assumed necessary to characterize the network dynamical behavior. The
symbol is used to mark missing or unspecified components.

Fig. 3. Example of a coupled 2D/3D system. (a) The domain of the 2D automaton A
is represented by a 8 × 4 structured grid of square cells, while the domain of the 3D
automaton B by a 4×2×3 structured grid of cubic cells. (b) A couple of interfaces, IA =
A and IB = {e3 = 0} on A and B, respectively, are defined to permit synchronization
(i.e., information exchange during the dynamical evolution of the system) between the
two automata. In particular, the finer-grained interface IA is partitioned in macrocells
(Ijk , with j, k = 1, 2, 3, 4) to math the coarse grained interface IB , where the original
partitioning in cells (ιjk , with j, k = 1, 2, 3, 4) is maintained.
250 D. D’Ambrosio et al.

The automaton A is defined as

 
A = (D, , I), , , Q, ( , ), ρ, , ( , t),
where:

– D = [1, 8] × [1, 4] ⊂ Z2 is the two-dimensional discrete domain.

– I = [1, 8] × [1, 4] = D is the interface with the B automaton.
– Q = R is the set of states for the cell.
– ρ : Q4 → R is a reduction function, which takes the states of four cells and
provides a reduced value, e.g., average. In such a case, we would have:

(q1 , q2 , q3 , q4 ) → (q1 + q2 + q3 + q4 )/4

– t = 3 is the time corresponding to the state transition (i.e., to the application

of the τ function - here omitted).

Similarly, the automaton B is defined as

 
B = (D, , I), , P, Q, (σ, ), , , ( , t),
where:

– D = [1, 4] × [1, 2] × [1, 3] ⊂ Z3 is the three-dimensional discrete domain.

– I = [1, 4] × [1, 2] × {1} ⊂ D(B) is the interface with the automaton A.
– P = {p} is the set of parameters, here composed by the single parameter p.
– Q = R is the set of states for the cell.
– σ : P × Q → Q is the internal transition that takes a real value and a cell
state and provides a new state for the cell. For instance, the function could
simply add the value of the parameter to the cell state. In such a case, we
would have:
(p, q) → q  = p + q
– t = 9 is the time corresponding to the state transition (i.e., to the application
of the τ function - here omitted).

The related network is defined as

N = <V, E, (I, π), τ, τN , >

– V = {A, B} is the set of XCA belonging to the network.

– E = {AB} is the set of edges, composed by the only edge AB defining the
dependence of B from A.
– I = {I (A) , I (B) } is the couple of interfaces of connected automata.
– π = πAB : I (B) → P(I (A) ) is the partition function defining the bijection
between each cell of I (B) and a square regions (macrocells) of 2 × 2 cells of
I (A) due to the different interface granularity (see Fig. 3b). In the specific
case, the following function could be adopted:

(i1 , i2 , 1) → [2 · (i1 − 1) + 1, 2 · (i1 − 1) + 2] × [2 · (i2 − 1) + 1, 2 · (i2 − 1) + 2]

 A General Computational Formalism for Networks of Structured Grids 251

Fig. 4. Sequence of the first nine steps of evolution for the network. Both the step
(n) and the time (t) axis are considered for reference. Note that, since the edge AB
is defined, the τAB function is executed each time the automaton B executes a state
transition.

4
– τ = τAB,ρ : Q(A) × Q(B) → Q(B) is the induction function from A to B that
takes in input the states of a macrocell on I (A) and the state of a cell on I (B)
and gives the new state for the cell on I (B)
 
(q1 , q2 , q3 , q4 )(A) , q (B) → q (B)

By considering the definitions of the automata A and B, it can be defined as

the composition of a reduction on A and an internal transition on B:
    
τAB,ρ (q1 , q2 , q3 , q4 )(A) , q (B) = σ (B) ρ(A) (q1 , q2 , q3 , q4 )(A) , q (B)
= (q1 + q2 + q3 + q4 )(A) /4 + q (B)

– τN : N × C (N ) → C (N ) is the network control unit, as defined in Sect. 3.

At step n = 0, the network is in the initial state. τN is then applied to let the
network evolve, by producing a sequence of configurations, as depicted in Fig. 4.
Note that, while A evolves independently from B, each time B evolves, the τAB
function is applied to take into consideration the dependence of B from A.

5 Conclusions

In this paper we have introduced the Extended Cellular Automata Network

formal computational paradigm. With respect to XCA, the network provides a
higher level of abstraction, by allowing for the co-evolution of different extended
cellular automata models, each one characterized by its own dimension, and
space/time granularity. Indeed, the network permits to define the hierarchical
relationship among automata by adopting a direct acyclic graph in which the
vertices represent the automata, and the edges the inter-relations between them.
The network also defines automaton-automaton communication interfaces and
external functions, which model effects that occur between connected automata.
Preliminarily, we have provided an alternative formulation of Extended Cel-
lular Automata that is more suitable for defining the new XCAN model. More-
over, we have provided a first theoretical example of a network composed by one
two-dimensional and one three-dimensional automaton, each one whit different
252 D. D’Ambrosio et al.

space/time granularity. In particular, we have pointed out aspects related to

the definition of communication interfaces, and have defined a simple external
function to account for the influence that the two-dimensional automaton has
on the three-dimensional one. Eventually, we have shown a possible co-evolution
process in which the automata evolution is coordinated by the network.
As future development of this work, we intend to define an application pro-
gram interface (API) based on XCAN and to provide an implementation of such
an API. The software will be developed with the purpose of both test differ-
ent parallelization scenarios, such as shared and distributed memory systems,
and strategies concerning, for instance, load balancing and halos exchanging
optimizations. In addition, different co-evolution strategies will be explored to
ensure the network evolve consistently. Eventually, the system will be applied to
the modeling and simulation of a real system, in order to evaluate correctness
and efficiency.

References
1. Aidun, C., Clausen, J.: Lattice-Boltzmann method for complex flows. Annu. Rev.
Fluid Mech. 42, 439–472 (2010)
2. Andersson, B., et al.: Computational Fluid Dynamics for Engineers. Cambridge
University Press, Cambridge (2011). https://doi.org/10.1017/CBO9781139093590
3. Belli, G., et al.: A unified model for the optimal management of electrical and
thermal equipment of a prosumer in a DR environment. IEEE Trans. Smart Grid
10(2), 1791–1800 (2019). https://doi.org/10.1109/TSG.2017.2778021
4. Burks, A.W.: Programming and the theory of automata. In: Burks, A.W. (ed.)
Essays on Cellular Automata, chap. 2, pp. 65–83. University of Illinois Press,
Urbana (1970)
5. Calidonna, C.R., Naddeo, A., Trunfio, G.A., Di Gregorio, S.: CANv2: a hybrid
CA model by micro and macro-dynamics examples. In: Bandini, S., Manzoni, S.,
Umeo, H., Vizzari, G. (eds.) ACRI 2010. LNCS, vol. 6350, pp. 128–137. Springer,
Heidelberg (2010). https://doi.org/10.1007/978-3-642-15979-4 13
6. Cervarolo, G., Mendicino, G., Senatore, A.: A coupled ecohydrological-three-
dimensional unsaturated flow model describing energy, H2O and CO2 fluxes. Eco-
hydrology 3(2), 205–225 (2010)
7. Cicirelli, F., Furfaro, A., Giordano, A., Nigro, L.: An agent infrastructure for dis-
tributed simulations over HLA and a case study using unmanned aerial vehicles.
In: 40th Annual Simulation Symposium (ANSS 2007), pp. 231–238, March 2007.
https://doi.org/10.1109/ANSS.2007.10
8. Cicirelli, F., Giordano, A., Nigro, L.: Distributed simulation of situated multi-
agent systems. In: 2011 IEEE/ACM 15th International Symposium on Distributed
Simulation and Real Time Applications, pp. 28–35, September 2011. https://doi.
org/10.1109/DS-RT.2011.11
9. Cicirelli, F., Forestiero, A., Giordano, A., Mastroianni, C.: Transparent and efficient
parallelization of swarm algorithms. ACM Trans. Auton. Adapt. Syst. (TAAS)
11(2), 14 (2016)
10. Cook, M.: Universality in elementary cellular automata. Complex Syst. 15(1), 1–40
(2004)
 A General Computational Formalism for Networks of Structured Grids 253

11. Crisci, G.M., et al.: Predicting the impact of lava flows at Mount Etna. Italy. J.
Geophys. Res. Solid Earth 115(B4), 1–14 (2010)
12. D’Ambrosio, D., et al.: The open computing abstraction layer for parallel complex
systems modeling on many-core systems. J. Parallel Distrib. Comput. 121, 53–70
(2018). https://doi.org/10.1016/j.jpdc.2018.07.005
13. D’Ambrosio, D., Rongo, R., Spataro, W., Trunfio, G.A.: Meta-model assisted evo-
lutionary optimization of cellular automata: an application to the SCIARA model.
In: Wyrzykowski, R., Dongarra, J., Karczewski, K., Waśniewski, J. (eds.) PPAM
2011. LNCS, vol. 7204, pp. 533–542. Springer, Heidelberg (2012). https://doi.org/
10.1007/978-3-642-31500-8 55
14. D’Ambrosio, D., Rongo, R., Spataro, W., Trunfio, G.A.: Optimizing cellular
automata through a meta-model assisted memetic algorithm. In: Coello, C.A.C.,
Cutello, V., Deb, K., Forrest, S., Nicosia, G., Pavone, M. (eds.) PPSN 2012. LNCS,
vol. 7492, pp. 317–326. Springer, Heidelberg (2012). https://doi.org/10.1007/978-
3-642-32964-7 32
15. Dattilo, G., Spezzano, G.: Simulation of a cellular landslide model with CAMELOT
on high performance computers. Parallel Comput. 29(10), 1403–1418 (2003)
16. De Rango, A., Napoli, P., D’Ambrosio, D., Spataro, W., Di Renzo, A., Di Maio, F.:
Structured grid-based parallel simulation of a simple DEM model on heterogeneous
systems. In: 2018 26th Euromicro International Conference on Parallel, Distributed
and Network-based Processing (PDP), pp. 588–595, March 2018. https://doi.org/
10.1109/PDP2018.2018.00099
17. Deng, X., Min, Y., Mao, M., Liu, H., Tu, G., Zhang, H.: Further studies on geo-
metric conservation law and applications to high-order finite difference schemes
with stationary grids. J. Comput. Phys. 239, 90–111 (2013)
18. Filippone, G., Spataro, W., D’Ambrosio, D., Spataro, D., Marocco, D., Trunfio, G.:
CUDA dynamic active thread list strategy to accelerate debris flow simulations. In:
Proceedings - 23rd Euromicro International Conference on Parallel, Distributed,
and Network-Based Processing, PDP 2015, pp. 316–320 (2015)
19. Filippone, G., D’Ambrosio, D., Marocco, D., Spataro, W.: Morphological coevo-
lution for fluid dynamical-related risk mitigation. ACM Trans. Model. Comput.
Simul. (TOMACS) 26(3), 18 (2016)
20. Folino, G., Mendicino, G., Senatore, A., Spezzano, G., Straface, S.: A model based
on cellular automata for the parallel simulation of 3D unsaturated flow. Parallel
Comput. 32(5), 357–376 (2006)
21. Frish, U., Hasslacher, B., Pomeau, Y.: Lattice gas automata for the Navier–Stokes
equation. Phys. Rev. Lett. 56(14), 1505–1508 (1986)
22. Fujimoto, R.M.: Parallel and Distribution Simulation Systems, 1st edn. Wiley, New
York (1999)
23. Cervarolo, G., Mendicino, G., Senatore, A.: Coupled vegetation and soil moisture
dynamics modeling in heterogeneous and sloping terrains. Vadose Zone J. 10, 206–
225 (2011)
24. Giordano, A., De Rango, A., D’Ambrosio, D., Rongo, R., Spataro, W.: Strategies
for parallel execution of cellular automata in distributed memory architectures.
In: 2019 27th Euromicro International Conference on Parallel, Distributed and
Network-Based Processing (PDP), pp. 406–413, February 2019. https://doi.org/
10.1109/EMPDP.2019.8671639
25. Giordano, A., et al.: Parallel execution of cellular automata through space parti-
tioning: the landslide simulation Sciddicas3-Hex case study. In: 2017 25th Euromi-
cro International Conference on Parallel, Distributed and Network-based Process-
ing (PDP), pp. 505–510, March 2017. https://doi.org/10.1109/PDP.2017.84
254 D. D’Ambrosio et al.

26. Di Gregorio, S., Serra, R.: An empirical method for modelling and simulating some
complex macroscopic phenomena by cellular automata. Future Gen. Comput. Syst.
16, 259–271 (1999)
27. Higuera, F., Jimenez, J.: Boltzmann approach to lattice gas simulations. Europhys.
Lett. 9(7), 663–668 (1989)
28. Jammy, S., Mudalige, G., Reguly, I., Sandham, N., Giles, M.: Block-structured
compressible Navier–Stokes solution using the OPS high-level abstraction. Int. J.
Comput. Fluid Dyn. 30(6), 450–454 (2016)
29. Langton, C.: Computation at the edge of chaos: phase transition and emergent
computation. Physica D 42, 12–37 (1990)
30. Lucà, F., D’Ambrosio, D., Robustelli, G., Rongo, R., Spataro, W.: Integrating
geomorphology, statistic and numerical simulations for landslide invasion hazard
scenarios mapping: an example in the Sorrento Peninsula (Italy). Comput. Geosci.
67(1811), 163–172 (2014)
31. McNamara, G., Zanetti, G.: Use of the Boltzmann equation to simulate lattice-gas
automata. Phys. Rev. Lett. 61, 2332–2335 (1988)
32. Mendicino, G., Pedace, J., Senatore, A.: Stability of an overland flow scheme in
the framework of a fully coupled eco-hydrological model based on the macroscopic
cellular automata approach. Commun. Nonlinear Sci. Numer. Simul. 21(1–3), 128–
146 (2015)
33. Mendicino, G., Senatore, A., Spezzano, G., Straface, S.: Three-dimensional unsat-
urated flow modeling using cellular automata. Water Resour. Res. 42(11), W11419
(2006)
34. von Neumann, J.: Theory of Self-Reproducing Automata. University of Illinois
Press, Champaign (1966)
35. Ninagawa, S.: Dynamics of universal computation and 1/f noise in elementary
cellular automata. Chaos Solitons Fractals 70(1), 42–48 (2015)
36. Oliverio, M., Spataro, W., D’Ambrosio, D., Rongo, R., Spingola, G., Trunfio, G.:
OpenMP parallelization of the SCIARA cellular automata lava flow model: perfor-
mance analysis on shared-memory computers. Procedia Comput. Sci. 4, 271–280
(2011)
37. De Rango, A., Spataro, D., Spataro, W., D’Ambrosio, D.: A first multi-GPU/multi-
node implementation of the open computing abstraction layer. J. Comput. Sci. 32,
115–124 (2019). https://doi.org/10.1016/j.jocs.2018.09.012
38. Reguly, I., et al.: Acceleration of a full-scale industrial CFD application with OP2.
IEEE Trans. Parallel Distrib. Syst. 27(5), 1265–1278 (2016). https://doi.org/10.
1109/TPDS.2015.2453972
39. Reguly, I., Mudalige, G., Giles, M., Curran, D., McIntosh-Smith, S.: The OPS
domain specific abstraction for multi-block structured grid computations. In: Pro-
ceedings of WOLFHPC 2014: 4th International Workshop on Domain-Specific Lan-
guages and High-Level Frameworks for High Performance Computing - Held in
Conjunction with SC 2014: The International Conference for High Performance
Computing, Networking, Storage and Analysis, pp. 58–67 (2014). https://doi.org/
10.1109/WOLFHPC.2014.7
40. Spataro, D., D’Ambrosio, D., Filippone, G., Rongo, R., Spataro, W., Marocco, D.:
The new SCIARA-fv3 numerical model and acceleration by GPGPU strategies.
Int. J. High Perform. Comput. Appl. 31(2), 163–176 (2017). https://doi.org/10.
1177/1094342015584520
41. Spingola, G., D’Ambrosio, D., Spataro, W., Rongo, R., Zito, G.: Modeling complex
natural phenomena with the libAuToti cellular automata library: an example of
 A General Computational Formalism for Networks of Structured Grids 255

application to lava flows simulation. In: PDPTA - International Conference on Par-

allel and Distributed Processing Techniques and Applications, pp. 277–283 (2008)
42. Thatcher, J.W.: Universality in the von Neumann cellular model. In: Burks, A.W.
(ed.) Essays on Cellular Automata, chap. 5, pp. 132–186. University of Illinois
Press, Urbana (1970)
43. Tie, B.: Some comparisons and analyses of time or space discontinuous Galerkin
methods applied to elastic wave propagation in anisotropic and heterogeneous
media. Adv. Model. Simul. Eng. Sci. 6(1) (2019). https://doi.org/10.1186/s40323-
019-0127-x
44. Wolfram, S.: A New Kind of Science. Wolfram Media Inc., Champaign (2002)
 Preliminary Model of Saturated Flow
Using Cellular Automata

Alessio De Rango1(B) , Luca Furnari1 , Andrea Giordano3 ,

Alfonso Senatore1 , Donato D’Ambrosio2 , Salvatore Straface1 ,
and Giuseppe Mendicino1
1
DIATIC, University of Calabria, Rende, Italy
[email protected]
2
DeMACS, University of Calabria, Rende, Italy
3
ICAR-CNR, Rende, Italy

Abstract. A fully-coupled from surface to groundwater hydrological

model is being developed based on the Extended Cellular Automata
formalism (XCA), which proves to be very suitable for high perfor-
mance computing. In this paper, a preliminary module related to three-
dimensional saturated flow in porous media is presented and imple-
mented by using the OpenCAL parallel software library. This allows to
exploit distributed systems with heterogeneous computational devices.
The proposed model is evaluated in terms of both accuracy and pre-
cision of modeling results and computational performance, using single
layered three-dimensional test cases at different resolutions (from local
to regional scale), simulating pumping from one or more wells, river-
groundwater interactions and varying soil hydraulic properties. Model
accuracy is compared with analytic, when available, or numerical (MOD-
FLOW 2005) solution, while the computational performance is evaluated
using an Intel Xeon CPU socket. Overall, the XCA-based model proves
to be accurate and, mainly, computationally very efficient thanks to the
many options and tools available with the OpenCAL library.

Keywords: High-performance computing · Extended Cellular

Automata · Computational fluid dynamics

1 Introduction

In a context of climate change, water resources management will become a key

factor to research sustainable development [14]. The development of more and
more efficient hydrological models is an essential element to study superficial
and subsurface water dynamics. Many models have been recently proposed to
predict these phenomena (cf. [20]). Most of them use PDEs (Partial Difference
Equation) to study and analyze the real event, approximating the solution by
adopting a specific numerical solver as, for instance, Finite Differences, Finite
Elements, Finite Volume Methods, etc. (cf. [4,9]). In most cases, an explicit
c Springer Nature Switzerland AG 2020
Y. D. Sergeyev and D. E. Kvasov (Eds.): NUMTA 2019, LNCS 11973, pp. 256–268, 2020.
https://doi.org/10.1007/978-3-030-39081-5_23
 Preliminary Model of Saturated Flow Using Cellular Automata 257

recurrence relation is obtained, which expresses the next state of the generic
(discrete) system element as a function of the current states of a limited number
of neighboring elements. For this reason, most explicit schemes can be formalized
in terms of Cellular Automata (CA) [18], which is one of the most known and uti-
lized decentralized discrete parallel computational models (cf. [2,3,11,15,16,21]).
Nevertheless, a kind of global control on the system to be simulated is often use-
ful to make the modeling of certain processes more straightforward. Therefore,
as pointed out in [6], both decentralized transitions and steering operations can
be adopted to model complex systems.
Different parallel integrated development environments (IDEs) and software
libraries were proposed to efficiently model complex systems formalized in terms
of Cellular Automata. Among them, Camelot represents a (commercial) exam-
ple of CA IDE, accounting for both local interactions and global steering, able
to accelerate the computation on clusters of distributed nodes interconnected
via network [1]. The libAuToti library [22] is essentially feature-equivalent to
Camelot, even released as an Open Source software library. Unfortunately, both
Camelot and libAutoti are no longer developed. However, the open source Open-
CAL [5] software library has been recently proposed as a modern alternative to
the aforementioned simulation software. Specifically, OpenCAL allows to exploit
heterogeneous computational devices like classical CPU sockets and modern
accelerators like GPUs on clusters of interconnected workstations. It was suc-
cessfully applied to the simulation of different systems, including a debris flow
evolving on real topographic surface, graphics convolutional filters, fractal gen-
eration, as well as particle systems based on the Discrete Element Method [7,8].
In this paper we developed a new, preliminary, groundwater model based on
the discretization of the Darcy law, by adopting an explicit Finite Difference
scheme to obtain a discrete formulation [12,17]. We then implemented it by
using OpenCAL, and therefore applied the model to simulate two different cases
of study, namely a constant pumping rate and an idealized case of interaction
between aquifer and river. We evaluated the model in terms of both accuracy and
computational efficiency. As regards accuracy, we compared the outcomes with
the analytical solutions, when available, and the MODFLOW 2005 model [13]
while, regarding computational performances, they were preliminary evaluated
by considering an Intel Xeon-based workstation.
The paper is organized as follows. Section 2 formalizes the groundwater flow
model, while Sect. 3 briefly presents the OpenCAL parallel software library.
Section 4 illustrates the considered case studies, together outcomes and accuracy
evaluation, while Sect. 5 describes the obtained computational performances.
Finally, Sect. 6 concludes the paper with a general discussion envisaging possible
future developments.

2 Preliminary Model of Saturated Flow

Groundwater phenomena is governed by the continuity equation that expresses
the mass balance for each cell. In particular, during a fixed time interval, the
258 A. De Rango et al.

sum of all flows into and out of the cell must be equal to the rate of change in
storage within the cell. The discrete governing equation is defined as follows:
 Δh
Qi = · Δx · Δy · Sy (1)
Δt
Where Q is the flow rate into the cell [m3 s−1 ], Δh is the head change [m] over
a time interval Δt [s], Δx and Δy identify the cell dimensions [m] and Sy is the
specific yield [.].
In the considered preliminary test cases, a two-dimensional space, composed
of regular cells, is adopted. In order to estimate the hydraulic head change within
the cell, the four cells at the cardinal points are considered. In particular, the
flows from the central cell are represented as positive inflow or negative outflow.
As shown in Fig. 1.

Fig. 1. Interaction between neighbor cells.

Therefore, the summation of flows is calculated as:


Qi = Qn + Qs + Qe + Qo + Qc (2)

Applying the Darcy law,

hi − hc
Qi = Ksi · Atransv · = Ti · (hi − hc ) (3)
Δx
Where Ksi is the saturated conductivity [m s−1 ], Atransv is the area of the
cell [m2 ], h is the hydraulic head [m] and Ti is hydraulic transmissivity [m2 s−1 ],
calculated as Ti = Ksi · b where b is the aquifer thickness [m].
Eventually, the explicit formula is obtained as follows:
 Preliminary Model of Saturated Flow Using Cellular Automata 259

Δt 
ht+1 = ht + · Qi (4)
Sy · Δx 2

The Cellular Automata formalism is used in order to solve the saturated flow
equation. Cellular Automata (CA) are computational models whose evolution is
governed by laws which are purely local. In their essential definition, a CA can
be described as a d–dimensional space composed of regular cells. At time t = 0,
cells are in an arbitrary state and the CA evolves by changing the states of the
cells in discrete steps of time and by applying simultaneously to each of them the
same law of evolution, or transition function. Input for each cell is given by the
states of neighboring cells and the neighborhood conditions are determined by a
geometrical pattern, which is invariant in time and space. Despite their simple
definition, CA may give rise to extremely complex behavior at a macroscopic
level. In fact, even if the local laws that regulate the dynamics of the system are
known, the global behavior of the system can be very hard to be predicted.
The Cellular Automata formulation of the aforementioned preliminary model
is defined as:

SFCA = <E d , X, S, P, σ>

• E d = {(x, y)|x, y ∈ N, 0 ≤ x ≤ lx , 0 ≤ y ≤ ly } is the set of square cells
covering the bi-dimensional finite region where the phenomenon evolves;
• X = {(0, 0), (−1, 0), (0, 1), (0, −1), (0, 1)} identifies the pattern of cells (Von
Neumann neighbourhood) that influence the cell state change;
• S = Sh × SKs × Sy × Sw is the finite set of states considered as Cartesian
product of substates where Sh is the head state, Sk identifies the saturated
conductivity, Sy identifies the specific yield and Sw identifies the source term;
• P = {Dimx , Dimy , b} is the finite set of parameters (invariant in time and
space) which affect the transition function used for the first case study. In
addition, for the second case study, further parameters are considered such
as: {Ksb , M, Hw , W, Rb };
• σ : identifies the transition function applied to each cell at every time step,
which describes water dynamics inside the aquifer by applying Eq. 4.

3 The OpenCAL Software System and Implementation

Details
OpenCAL (Open Computation Abstraction Layer) is an open source paral-
lel computing abstraction layer for scientific computing and is based on the
Extended Cellular Automata general formalism [10] as a Domain Specific Lan-
guage. Moreover, Cellular Automata, Finite Difference Method and other struc-
tured grid-based methods can be straightforwardly supported.
The library permits to abstract the implementation phase by providing the
building blocks of the XCA computational paradigm through its API structures
260 A. De Rango et al.

and functions. In particular, the computational domain, substates, and neigh-

borhood are formalized at the higher level of abstraction, as well as local inter-
actions and global operations (e.g., reductions operations), which are commonly
required by structured grid-based applications. Moreover, the initial conditions
of the system, and a termination criterion to halt the system evolution can be
defined. In addition, OpenCAL provides embedded optimization algorithms and
allows for a fine grained control over the simulation [5]. Furthermore, starting
from an OpenCAL-based serial implementation of a model, different parallel ver-
sions can be obtained with the minimum effort, including those for multi- and
many-core shared memory devices, as well as for distributed memory systems.
For this purpose, the OpenCAL implementation level is transparently exploited,
whose components are based on the OpenMP, OpenCL, and MPI APIs.
As regards the implementation of the proposed underground model, we con-
sidered two different versions in this preliminary work, namely a straightfor-
ward basic implementation, that does not consider any OpenCAL specific opti-
mizations, and an explicit implementation, that in most cases permits to avoid
unneeded memory access operations [5]. More in detail, in the first version, the
implicit update policy transparently updates the whole set of model substates at
the end of each execution of each elementary process composing the transition
function. Conversely, the explicit scheme permits to selectively update substates,
depending if they are actually modified.
Note that only the serial and OpenMP-based components of the library were
considered in this work, namely OpenCAL (which actually refers to both the
serial implementation component, and the name of the whole software library)
and OpenCAL-OMP, respectively.

4 Case Studies
In this preliminary development phase of the model, two different transient case
studies where conducted, considering an unconfined aquifer, following [19]. The
first one is a standard hydrology problem, the water table drawdown caused by
a well with a constant pumping rate. The second one is the aquifer response to
a stream-stage variation. In particular, in these case studies the Qc term from
the Eq. (2) is referred to the source term, which is equal to zero when well or
river cells are not considered. A square aquifer of 1 km2 is considered, in order
to test the accuracy and computational performance of the model. For both case
studies, two different meshes, with different sizes, are adopted. For the first test
Δx is fixed to 10 m, for a total of 100 × 100 = 104 cells, in the second test
Δx = 1 m, for a total of 1000 × 1000 = 106 cells.
Hydraulic saturated conductivity Ks is set equal to 1.25·10−5 m s−1 , and the
specific yield Sy is set 0.1. The aquifer thickness is 50 m, temporal step size Δt
Δx2 ·S
is calculated using the Courant-Friedrichs-Lewy (CFL) condition Δt = 4T y
and is set to 4000 s and 40 s, for the first and second test cases, respectively.
Both cases studies are compared with the widely used groundwater model
MODFLOW 2005 [13]. Its solutions are achieved, only on the less dense mesh,
 Preliminary Model of Saturated Flow Using Cellular Automata 261

using the Preconditioned Conjugate-Gradient(PCG) solver set with the same Δt

used in the SFCA model (4000 s).

4.1 Drawdown Made by a Well with Constant Pumping Rate

This problem could be resolved using the analytic solution (5) obtained by Theis
[23,24], which is valid for confined aquifer with full penetrating well.
Q
s= W (u) (5)
4πT
 ∞ −u
e
W (u) = du (6)
u u

r 2 · Sy
u= (7)
4tT
Where s is drawdown [m] (Fig. 2b), Q is the pumping rate [m3 s−1 ], t is the
time since pumping starts [s], T is the hydraulic transmissivity [m2 s−1 ], W (u)
is called well function and could be approximated using a numerical solution,
r is the radial distance from the well [m] and Sy is specific yield. This solution
could be used also for unconfined aquifer, if the drawdown is relatively small
compared to the saturated thickness.

Fig. 2. (a) Scheme of the position of the monitoring wells and the pumping well, (b)
vertical and horizontal section of the pumping well, s is the drawdown made by the
pumping in respect to the piezometric line which represents the undisturbed condition.

In this specific case study, a well is placed in the center of the domain with
a constant pumping rate of 0.001 m3 s−1 . The initial head is equal to 50 m all
262 A. De Rango et al.

over the domain and the Dirichlet condition is used for the boundaries, fixing
the hydraulic head to a constant value.
Monitoring wells are placed, according to [19], at distances equal to 150 m,
200 m and 300 m from the pumping well on one cardinal direction. Moreover,
two further monitoring wells are placed at distances of 127 m and 170 m on the
45◦ direction (Fig. 2a), to verify the quality of the Von Neumann neighborhood.
The results of a simulation with 10 m mesh size (100 × 100 cells) over 12 days
are shown in Fig. 3.
The lines represent the analytical solutions obtained by Theis, the red squares
are the numerical solutions obtained by MODFLOW 2005 and the blue crosses
are the numerical solutions obtained by SFCA . This visual comparison shows a
very good fit by the SFCA model and confirm that Von Neumann neighborhood
does not generate numerical distortions even in diagonal directions.

Fig. 3. Drawdown obtained by analytical solution, MODFLOW 2005 and SFCA at

each monitoring well (please refer to Table 1 for labels).

Table 1. RMSE [m] referred to analytical solution.

Well Distance [m] MODFLOW SFCA

MW1 150 2,10E−04 1,49E−04
MW2 200 8,00E−05 3,37E−05
MW3 300 4,71E−05 3,35E−06
MW4 127 1,30E−04 3,24E−04
MW5 170 3,13E−04 6,74E−05
 Preliminary Model of Saturated Flow Using Cellular Automata 263

Also Table 1 confirms a very got fit by SFCA compared to the analytical

solution, even better than MODFLOW 2005 results. In particular, the Root
Mean Square Error (RMSE) calculated using SFCA is always lower than the
MODFLOW RMSE, unless for MW4, which points out a better performance by
MODFLOW 2005.
Further tests are also executed with the denser grid (1000 × 1000 cells) with
similar results.

4.2 Interaction Between Aquifer and River

One of the most relevant problems in alluvial hydrology is the interaction

between the aquifer system and the river stage. Modeling this phenomenon
requires that two different behaviors are considered, being in one case the aquifer
a source for the river (therefore, aquifer head will decrease in the proximity of
the river) and, in the opposite case, the river a source for the aquifer (therefore,
the aquifer head will increase in the proximity of the river).
To discriminate these two different behaviors, the river package in MOD-
FLOW 2005 uses the bottom of streambed (Rb ) as a control variable. Specifically,
if the aquifer hydraulic head (h) is above the bottom of streambed (h > Rb ), it
acts as a source. Otherwise (h ≤ Rb ), the river acts as a source:
 Ksb ·Δx·W
M · (hw − h) if h > Rb
QC = Ksb ·Δx·W (8)
M · (H w + M ) if h ≤ Rb
Equation 8 describes how the river source is modeled both in MODFLOW
2005 and in the SFCA model, where Ksb is the streambed hydraulic conductivity,
set to 1.0 · 10−5 m s−1 , W is the river width fixed to 5 m, M is the riverbed
thickness set to 0.5 m, the bottom of streambed Rb is set to 46.5 m and Hw is the
water level in the stream above the surface of the streambed. hw is the hydraulic
head in the stream and in this case study it is assumed to change accordingly
to a triangular law during the simulations (Fig. 5). The relative schematization
is illustrated in Fig. 4b.
According to [19], river is located at a distance of 250 m from the west border
of the domain and flows from north to south. Neumann boundary conditions of
no flows are set in the northern and southern borders, while Dirichlet conditions
are imposed to the western and eastern borders. The initial head is fixed equal
to 50 m all over the domain.
Simulations lasted 30 days. During the first 10 days, Hw is set to 48 m,
between 10 and 15 days it linearly increases up to 50 m, during the next 5 days
it linearly decreases again to 48 m keeping this constant value until the end of
the experiment (Fig. 5). Therefore, three different behaviors can be identified:
initially, the aquifer recharges the river, then the river recharges the aquifer and,
finally, the aquifer again recharges the river.
Monitoring wells are placed at increasing distance from the west border.
Specifically, MW100 is placed at 100 m distance, MW350 at 350 m, MW450 at
450 m, MW 550 at 550 m and MW650 at 650 m (Fig. 4a). The closest well to
264 A. De Rango et al.

Fig. 4. (a) Overview of the position of the monitoring wells and the river, (b) Cross-
section of the case study identifying the interaction between the river and the aquifer.

the river is MW350, it is expected that in this monitoring well the influence
of the river stage variation is mostly amplified. On the other hand, MW650 is
the farthest well, where the head variation during the simulation should be the
lowest.
The results of the simulations, with 10 m mesh size, are represented in Fig. 5.
The red line represents the Hw variation, black lines the MODFLOW 2005 solu-
tions and blue crosses the SFCA model solutions. Also in this second case study,
the SFCA model shows an excellent fit and confirms the trend expected by phys-
ical considerations.

5 Computational Performance

In order to evaluate the computational performance of the two different Open-

CAL implementations of the SFCA groundwater flow model here proposed (cf.
Sect. 3), the simulation of the test studies reported in Sects. 4.1 and 4.2 were
performed over a computational domain of 1000 × 1000 cells of 1 m side. Both
OpenCAL and SFCA were compiled by considering the gcc GNU C compiler
version 8.3.0 with the -O3 optimization flag. All serial and multi-thread tests
were performed on a hyper-threading 8-core Intel Xeon 2.0 GHz E5-2650 CPU
based workstation running Arch Linux by considering 1, 2, 4, 8 and 16 threads.
Here, the OpenMP static scheduler for parallel loop work-sharing was considered
for the OpenCAL-OMP benchmarks.
The speed-up analysis results of the first test case related to OpenCAL and
OpenCAL-OMP versions are shown in Fig. 6, whereas the results of the second
test case are shown in Fig. 7. A speed-up value of about 12 was obtained by
considering the explicit update scheme for both the considered test cases, while
 Preliminary Model of Saturated Flow Using Cellular Automata 265

Fig. 5. Hydraulic head in the stream Hw and hydraulic head at different monitoring
wells obtained by MODFLOW 2005 and SFCA models. (Color figure online)

Fig. 6. Speed-up achieved by the OpenCAL-OMP simulation of the test study

described in Sect. 4.1. Results of both the implicit and explicit implementations are
shown. Speed-up values were evaluated by considering the corresponding serial execu-
tion time that was 1587 and 1304 s for the implicit and explicit OpenCAL schemes,
respectively.

speed-up of about 7.4 and 4.8 were registered for the first and second test cases,
respectively, when the implicit scheme was adopted.
The result obtained in the case of the implicit version is explained by consid-
ering that the whole (global) state of the automaton is transparently updated,
even if only the Sh substate is actually modified by the transition function (cf.
266 A. De Rango et al.

Sect. 2). Nevertheless, the implicit scheme implementation represents the first
fundamental step in the OpenCAL development process that permits to rapidly
obtain a non optimized working version of the model in order to test its cor-
rectness. It is worth to note that no in-depth knowledge of OpenCAL details is
required in this phase.

Fig. 7. Speed-up achieved by the OpenCAL-OMP simulation of the test study

described in Sect. 4.2. Results of both the implicit and explicit implementations are
shown. Speed-up values were evaluated by considering the corresponding serial execu-
tion time that was 4476 and 3227 s for the implicit and explicit OpenCAL schemes,
respectively.

6 Conclusions
This paper reports the first implementation of the SFCA model for the simulation
of groundwater phenomena, using the OpenCAL parallel computational library.
Two case studies are considered, one related to a drawdown made by a well
with constant pumping rate and the second regarding the interaction between
an aquifer and a river. The parallel implementation of the model is carried out
by considering the OpenMP component of OpenCAL. Several tests are carried
out to validate the accuracy and computational performance of the model. The
MODFLOW 2005 model and the analytical solution, obtained by Theis, are used
to compare the results of the developed model. The results show good accuracy of
the model for both case studies. In particular, for the first case study, the model
achieved a better agreement with the analytical solution than MODFLOW 2005
for most of the monitoring wells. Regarding computational performances, the
tests pointed out overall good performance and scalability, achieving a speed-up
of about 12 by considering 16 threads for both test cases.
 Preliminary Model of Saturated Flow Using Cellular Automata 267

Future work will regard the extension of the model to the non-saturated
case and performance improvements. In particular, regarding the performance,
other compilers will be tested in spite of gcc, e.g. clang. Moreover, OpenCAL
will be better exploited to improve performance. For instance, the embedded
quantization optimization could be considered that, based on a set of one or more
thresholds, is able to exclude cells classified as stationary from the computation.
Eventually, other OpenCAL components will be taken into account in order to
exploit modern accelerators like GPUs.

Acknowledgements. L. Furnari acknowledges for the program “POR Calabria

FSE/FESR 2014/2020 - Mobilitá internazionale di Dottorandi a Assegni di
ricerca/Ricercatori di Tipo A” Actions 10.5.6 and 10.5.12.

References
1. Cannataro, M., Di Gregorio, S., Rongo, R., Spataro, W., Spezzano, G., Talia, D.:
A parallel cellular automata environment on multicomputers for computational
science. Parallel Comput. 21(5), 803–823 (1995)
2. Cervarolo, G., Mendicino, G., Senatore, A.: A coupled ecohydrological-three-
dimensional unsaturated flow model describing energy, H2 O and CO2 fluxes. Eco-
hydrology 3(2), 205–225 (2010)
3. Cervarolo, G., Mendicino, G., Senatore, A.: Coupled vegetation and soil moisture
dynamics modeling in heterogeneous and sloping terrains. Vadose Zone J. 10, 206–
225 (2011)
4. Chang, K.S., Song, C.J.: Interactive vortex shedding from a pair of circular cylin-
ders in a transverse arrangement. Int. J. Numer. Methods Fluids 11(3), 317–329
(1990)
5. D’Ambrosio, D., et al.: The open computing abstraction layer for parallel complex
systems modeling on many-core systems. J. Parallel Distrib. Comput. 121, 53–70
(2018)
6. Dattilo, G., Spezzano, G.: Simulation of a cellular landslide model with camelot
on high performance computers. Parallel Comput. 29(10), 1403–1418 (2003)
7. De Rango, A., Napoli, P., D’Ambrosio, D., Spataro, W., Di Renzo, A., Di Maio, F.:
Structured grid-based parallel simulation of a simple DEM model on heterogeneous
systems, pp. 588–595 (2018)
8. De Rango, A., Spataro, D., Spataro, W., D’Ambrosio, D.: A first multi-GPU/multi-
node implementation of the open computing abstraction layer. J. Comput. Sci. 32,
115–124 (2019)
9. Deng, X., Min, Y., Mao, M., Liu, H., Tu, G., Zhang, H.: Further studies on geo-
metric conservation law and applications to high-order finite difference schemes
with stationary grids. J. Comput. Phys. 239, 90–111 (2013)
10. Di Gregorio, S., Serra, R.: An empirical method for modelling and simulating some
complex macroscopic phenomena by cellular automata. Futur. Gener. Comput.
Syst. 16, 259–271 (1999)
11. Filippone, G., D’Ambrosio, D., Marocco, D., Spataro, W.: Morphological coevo-
lution for fluid dynamical-related risk mitigation. ACM Trans. Model. Comput.
Simul. (TOMACS) 26(3), 18 (2016)
268 A. De Rango et al.

12. Folino, G., Mendicino, G., Senatore, A., Spezzano, G., Straface, S.: A model based
on cellular automata for the parallel simulation of 3D unsaturated flow. Parallel
Comput. 32(5), 357–376 (2006)
13. Harbaugh, A.: MODFLOW-2005, the U.S. geological survey modular ground-water
model-the ground-water flow process. U.S. Geological Survey (2005)
14. Kundzewicz, Z.W., et al.: The implications of projected climate change for fresh-
water resources and their management. Hydrol. Sci. J. 53(1), 3–10 (2008)
15. Lucá, F., D’Ambrosio, D., Robustelli, G., Rongo, R., Spataro, W.: Integrating
geomorphology, statistic and numerical simulations for landslide invasion hazard
scenarios mapping: an example in the Sorrento Peninsula (Italy). Comput. Geosci.
67(1811), 163–172 (2014)
16. Mendicino, G., Pedace, J., Senatore, A.: Stability of an overland flow scheme in
the framework of a fully coupled eco-hydrological model based on the Macroscopic
Cellular Automata approach. Commun. Nonlinear Sci. Numer. Simul. 21(1–3),
128–146 (2015)
17. Mendicino, G., Senatore, A., Spezzano, G., Straface, S.: Three-dimensional unsatu-
rated flow modeling using cellular automata. Water Resour. Res. 42(11), 2332–2335
(2006)
18. von Neumann, J.: Theory of Self-Reproducing Automata. University of Illinois
Press, Champaign (1966)
19. Ravazzani, G., Rametta, D., Mancini, M.: Macroscopic cellular automata for
groundwater modelling: a first approach. Environ. Model Softw. 26(5), 634–643
(2011)
20. Senatore, A., Mendicino, G., Smiatek, G., Kunstmann, H.: Regional climate change
projections and hydrological impact analysis for a Mediterranean basin in Southern
Italy. J. Hydrol. 399(1), 70–92 (2011)
21. Spataro, D., D’Ambrosio, D., Filippone, G., Rongo, R., Spataro, W., Marocco, D.:
The new SCIARA-fv3 numerical model and acceleration by GPGPU strategies.
Int. J. High Perform. Comput. Appl. 31(2), 163–176 (2017)
22. Spingola, G., D’Ambrosio, D., Spataro, W., Rongo, R., Zito, G.: Modeling complex
natural phenomena with the libAuToti cellular automata library: an example of
application to lava flows simulation. In: PDPTA - International Conference on Par-
allel and Distributed Processing Techniques and Applications, pp. 277–283 (2008)
23. Straface, S.: Estimation of transmissivity and storage coefficient by means of a
derivative method using the early-time drawdown. Hydrol. J. 17(7), 1679 (2009)
24. Theis, C.V.: The relation between the lowering of the piezometric surface and the
rate and duration of discharge of a well using ground-water storage. Trans. Am.
Geophys. Union 16, 519–524 (1935)
 A Cybersecurity Framework
for Classifying Non Stationary Data
Streams Exploiting Genetic Programming
and Ensemble Learning

Gianluigi Folino(B) , Francesco Sergio Pisani, and Luigi Pontieri

ICAR-CNR, Rende, CS, Italy

{gianluigi.folino,fspisani,luigi.pontieri}@icar.cnr.it

Abstract. Intrusion detection systems have to cope with many chal-

lenging problems, such as unbalanced datasets, fast data streams and
frequent changes in the nature of the attacks (concept drift). To this aim,
here, a distributed genetic programming (GP) tool is used to generate
the combiner function of an ensemble; this tool does not need a heavy
additional training phase, once the classifiers composing the ensemble
have been trained, and it can hence answer quickly to concept drifts,
also in the case of fast-changing data streams. The above-described app-
roach is integrated into a novel cybersecurity framework for classifying
non stationary and unbalanced data streams. The framework provides
mechanisms for detecting drifts and for replacing classifiers, which per-
mits to build the ensemble in an incremental way. Tests conducted on
real data have shown that the framework is effective in both detecting
attacks and reacting quickly to concept drifts.

Keywords: Cybersecurity · Intrusion detection · Genetic

programming

1 Introduction
The problem of classifying attacks in the cybersecurity field involves many issues,
such as the need of dealing with fast data streams, the non-stationary nature
of attacks (concept drift) and the uneven distribution of the classes. Classical
data mining algorithms usually are not able to handle all these issues. Ensemble-
based algorithms [5] fit well this challenging scenario, as they are incremental,
robust to noise, scalable and operate well on unbalanced datasets. Data-driven
strategies for combining the classifiers composing the ensemble have proven to
be more effective than classic combination schemes relying on non-trainable
aggregation functions (such as weighted voting of the base models’ predictions).
However, since it may be necessary to re-train (part of) the ensemble when
new data become available, this phase should not be computationally expensive.
In this work, we describe an approach to the detection of attacks in fast data
c Springer Nature Switzerland AG 2020
Y. D. Sergeyev and D. E. Kvasov (Eds.): NUMTA 2019, LNCS 11973, pp. 269–277, 2020.
https://doi.org/10.1007/978-3-030-39081-5_24
270 G. Folino et al.

streams, which can incrementally update an ensemble of classifiers when changes

in the data are detected. This approach, named CAGE-MetaCombiner (CM C
for short), is a distributed intrusion detection framework, which exploits the
distributed GP tool CAGE [4] to generate the combiner function of the ensem-
ble by mixing up a collection of non-trainable functions. A major advantage
of using non-trainable functions as building blocks is that they can be evolved
without any extra phase of training and, thus, they are particularly apt to han-
dle concept drifts, also in the case of stringent processing-time constraints. More
details on the ensemble-based algorithm and an experimentation on artificial
datasets can be found in [7], while here we describe the general architecture and
the usage of a real intrusion detection dataset. The rest of the paper is struc-
tured as follows. Section 2 gives some background information about Genetic
Programming. Section 3 describes the proposed approach and system architec-
ture. Section 4 illustrates experiments conducted to validate the approach and
to compare it with related ones. Finally, Sect. 5 draws some conclusions and
discusses a number of open issues and relevant lines of future work.

2 Background: Genetic Programming

Genetic Programming (GP), inspired by Darwin’s evolutionary theories, evolves

a population of solutions (individuals) to a problem for a number of genera-
tions. It can be used to learn both the structure and parameters of the model;
the individuals (chromosomes) of typical GP approaches are trees. The internal
nodes of the tree are functions and the leaves are typically the problem variables,
constants or random numbers. The initial population of GP is a set of trees gen-
erated randomly. During the evolutionary process, the individuals are evolved
until they reach the optimal solution (or a good approximation of it) of the
problem, or until a maximum number of generations is reached. The evolution is
driven by a function of fitness, which is chosen for the particular problem to be
solved and represents the goodness of a solution of the problem. Similar to other
evolutionary algorithms, for each generation, two genetic operators (crossover
and mutation) are performed on some individuals, chosen randomly on the basis
of their fitness: individuals with better fitness have more chance to be chosen.
The crossover operator swaps two random subtrees of two individuals (par-
ents) and generates two new individuals (children). Moreover, the mutation oper-
ator is performed on a single individual and mutates a random subtree and gen-
erate a new individual. Figure 1, show an example of the crossover and mutation
operator. By these two operators GP can search the problem landscape in order
to find the optimal solution. Then, the new generated individuals are added to
the populations and compete with other individuals based on their fitness, i.e.,
the better individuals have more chance to survive. This process leads to find
better solutions during the evolution of the process.
 A Cybersecurity Framework for Classifying Non Stationary Data Streams 271

Fig. 1. An example of GP crossover (left) and mutation (right). In GP crossover two

random subtrees of the parents are selected and swapped with each other and generates
two new individuals. Here, the function set contains Avg, +, and ×, and the terminal
set contains problem variables and some random numbers. In GP mutation, a random
subtree of the parent is selected and substituted with a new random subtree.

3 Proposed Framework: Approach and System

Architecture

This section describes our Intrusion Detection (ID) approach, and an IDS archi-
tecture supporting the main tasks (e.g., drift detection, models’ induction, etc.)
involved in it.

3.1 Stream-Oriented Ensemble Discovery and Prediction Approach

The proposed approach relies on analysing a stream D = d0 , d1 , . . . of log data,

containing both labelled and unlabelled data tuples, with the help of an ensemble
model E, which consists of two components: (i) a list base(E) of base classifiers,
say h1 , . . . , hk for some k in N, such that each hj encodes a function mapping
any data tuple d of D to an anomaly score hj (d) ∈ [0, 1] (the higher the score,
the more likely d is estimated to be an intrusion attack); (ii) a combiner function
φE , which maps any data tuple d of D to an overall anomaly score φE (d) ∈ [0, 1],
derived from the predictions h1 (d), . . . , hk (d).
Both the base classifiers and the ensemble E are built through a continu-
ous learning and prediction scheme, where incoming labelled tuples are used
as training examples, while unlabelled one are classified, as either malicious
or normal, using E, as they arrive. The approach relies on a window-based
processing strategy, where the input data stream D is split into a sequence
D0 , D1 , . . . , Di , Di+1 , . . . of non-overlapping fixed-length windows, consisting
each of n temporally contiguous tuples.
The main computation steps of our continuous learning and prediction app-
roach are summarised below.
272 G. Folino et al.

As soon as a new window Di has been gathered, the approach tries to assess
whether some kind of concept drift has occurred, by suitably comparing the data
distribution of Di with that of the previous window(s).

Fig. 2. A general architecture for classifying large streams of IDS data.

In case a drift is detected, the labelled tuples of Di are split into two sets:
a training subset Ti and a validation set Vi . The former is used to train α
different base classifiers, by resorting to different induction algorithms, like those
described in the experimental section. Among the discovered classifiers, the l ones
with the highest accuracy scores are chosen and added to base(E). For efficiency
reasons, at most m base classifiers are kept in E. Hence, whenever |base(E)|
overcomes m, |base(E)| − m classifiers must be removed from base(E), using
some suitable selection strategy. At this point, a novel combiner function ψ E is
derived for E by running a GP procedure that tries to maximise the accuracy
of ψ E over Vi . More details on this respect are given in the next subsection.
As a last processing step for the current window Di , the ensemble E is applied
to the tuples of Di , in order to associate the unlabelled ones with a class label
and to update the statistics needed for concept-drift detection.

3.2 Conceptual Architecture and Implementation Details

Figure 2 shows the architecture of the intrusion-detection system that was imple-
mented to validate the approach described above. The architecture features sev-
eral functional modules, which correspond to the main kinds of data-stream pro-
cessing tasks that are involved in the approach. We next briefly describe these
modules and some details on how their current implementation in the prototype
system.
 A Cybersecurity Framework for Classifying Non Stationary Data Streams 273

Change Detector. This module analyses the data stream in search of possible
changes in the distribution of the target classes (i.e. normal vs attacks), repre-
senting evidence for a concept drift. For the sake of efficiency (and suitability
for an online ID setting), such analysis relies on the incremental computation of
statistics for every new incoming data window Di .
The current prototype system resorts to a classic change-detection method
of this kind, namely ADaptative WINDdowing (ADWIN) [1] (using a confidence
level δ of 0.002). However, notice that our approach can adopt alternative drift
detection methods—see [9] for a survey on this topic.

Model Generator. This module is responsible for inducing a collection of base

models from given training tuples, and it is called when building an initial ver-
sion of the ensemble and whenever a concept drift is detected on a window Di
(to extract a collection of new base models that capture well the emerging con-
cept). Currently, this module implements a number of classic classifier-induction
methods, including decision tree learning, Bayesian methods, logistic regression
and k-NN (see Sect. 4 for details).

Update Ensemble. This module is meant to produce an updated version of the

ensemble E, starting from an updated collection of base models. To this end, it
must face two main tasks: (i) discard some of the base models, in case they are
more than the maximum threshold m, and (ii) generate a combiner function for
E via evolutionary computation.
As to the first task, many strategies were defined in the literature [8] for
discarding inaccurate/redundant/obsolete base models, like the one used in the
experimentation and described in the next section, which could be integrated in
our framework.
In order to define an optimal combiner function for the selected base models,
a GP strategy is adopted in the framework, briefly described below. Basically,
any candidate combiner (function) is encoded as a GP tree where each non-leaf
node can be associated with a non-trainable aggregation function (i.e. average,
weighted average, multiplication, maximum and median, replicated with a differ-
ent arity, from 2 to 5), while each leaf refers to one of the base models. The fitness
score for evaluating the quality of such a tree w.r.t. the current data window Di
is simply computed as the prediction error made by the resulting ensemble on the
validation set Vi – needing no extra computation on other training data. Further
details on how the combiner function of the ensemble is evolved are reported in
[6]. The tool used to evolve the combiner functions is a distributed GP imple-
mentation of a fine-grained cellular model, named CAGE [4], which can run on
both distributed-memory parallel computers and distributed environments.

Other Modules. In principle, the stream of input log data may come from differ-
ent kinds of sources, such as network-traffic logs, system/application logs, etc.
Thus, suitable data preprocessing and feature extraction methods must be imple-
mented, in the Data Preprocessing module, to turn these data into a homogenous
collection of data tuples, in order to make them undergo our analysis approach.
274 G. Folino et al.

Finally, the Alert Generation module is devoted to classify the incoming unla-
belled tuples, using the ensemble model E, and to generate and register suitable
alerts for each detected intrusion attack.

4 Experimental Section
In this section, we verified the quality of the proposed approach on the ISCX IDS
dataset [12]. This dataset was created by capturing seven days of network traffic
in a controlled testbed made of a subnetwork placed behind a firewall. Normal
traffic was generated with the aid of agents that simulated normal requests
of human users following some probability distributions extrapolated from real
traffic. Attack were generated with the aid of human operators. The result is a
fully labelled dataset containing realistic traffic scenarios (see Table 1). Different
days contain different attack scenarios, ranging from HTTP Denial of Service,
DDos, Brute Force SSH and attempts of infiltrating the subnetwork from the
inside.
A Linux cluster with 16 Itanium2 1.4 GHz nodes (2 GBytes of RAM), con-
nected by a Myrinet high performance network, was employed to conduct the
experiments. The GP framework was run by using the same parameters as in the
original paper, since in this work we are no interested in tuning the parameters
to improve the performance. In particular, the probability of crossover was set
to 0.7, the probability of mutation was set to 0.1, a population of 120 individuals
(each one with a maximum depth set to 7) was used for 500 generations. All the
results were obtained by averaging 30 runs.
Among the many metrics for evaluating classifier systems, in this paper we
choose recall and precision, because they give an idea of the capability of the
system to individuate the attacks and to reduce the number of false alarms.
We remind that the recall indicates the ratio between the correctly predicted
attacks to the total number of the attacks (a value of 100% means that all the
attacks were detected). The precision indicates the ratio between the number of
correctly predicted attacks and the total number of predicted attacks (a value
of 100% means that no false alarms were signaled).
The AUC metric quantifies the area under the ROC curve. The ROC curve
is computed comparing the false positive rate (i.e., recall) and the true positive
rate (i.e., the ratio between the false alarm signaled above all normal connections
processed). It is evident that an AUC close to 1 means an optimal recognition
rate.
In order to train the classifiers of the ensemble, each window of the stream
is divided into two equal parts (each on of 50%): the first part is used to train
the base classifiers, and the other part is used as a validation set for evolving the
combination function through the evolutionary algorithm. The base ensemble is
composed of 10 classifiers, while the maximum number of classifiers is 20.
CAGE-MetaCombiner (CMC) adopts many different learners as base classi-
fiers, all taken from the well-known WEKA tool1 . In more detail, the classifiers
1
http://www.cs.waikato.ac.nz/ml/weka.
 A Cybersecurity Framework for Classifying Non Stationary Data Streams 275

are the following ones: J48 (decision trees), JRIP rule learner (Ripper rule learn-
ing algorithm), NBTree (Naive Bayes tree), Naive Bayes, 1R classifier, logistic
model trees, logistic regression, decision stumps and 1BK (k-nearest neighbor
algorithm).

Table 1. Main characteristics of the ISCX IDS dataset.

Day Description Pcap file’ No. of Perc. of

size (GB) flows attacks
Day 1 Normal traffic without malicious 16.1 359,673 0.000%
activities
Day 2 Normal traffic with some malicious 4.22 134,752 1.545%
activities
Day 3 Infiltrating the network from the inside & 3.95 153,409 6.395%
Normal traffic
Day 4 HTTP Denial of Service & Normal traffic 6.85 178,825 1.855%
Day 5 Distributed Denial of Service using an 23.4 554,659 6.686%
IRC Botnet
Day 6 Normal traffic without malicious 17.6 505,057 0.000%
activities
Day 7 Brute Force SSH + Normal activities 12.3 344,245 1.435%

Table 1 illustrates the different distributions of the attacks, which usually are
grouped in a small range of windows; in addition, for different days, different
types of attack are detected. This characteristic is very useful in testing the
capability of our algorithm to handle drifts, as, when a new type of attack is
present in the data, usually a drift can be observed.
We compared our approach with the HoeffdingTree algorithm (classic and
boosted version) for different width of the windows (1,000, 2,000 and 5,000 of
tuples). The results are reported in Table 2.
Differently from our approach, the HoeffdingTree algorithm updates more
frequently its model; therefore, for small windows (1k) it preforms better in
terms of precision and AUC. However, as the size of the window increases (2k
and 5k), our approach performs sensibly better. Furthermore, while for both the
versions of the HoeffdingTree, a performance degradation of the ensemble-based
algorithm can be observed when the width is increased, this behavior does not
affect our approach; probably, it is due to the fact that, in our case, when a
drift is detected, the algorithm updates/replaces the models and re-weights the
classifiers (by recomputing the combination function).
276 G. Folino et al.

Table 2. Precision, rEcall and AUC for the comparison among our approach, the
Hoeffding tree (classical and boosted version) on the ISCX dataset.

Precision Recall AUC

1k 2k 5k 1k 2k 5k 1k 2k 5k
CMC 83.46 ± 0.28 87.59 ± 0.03 92.42 ± 2.20 88.39 ± 0.69 88.25 ± 1.08 82.44 ± 0.69 0.84 ± .022 0.87 ± .003 0.89 ± .001
HT boosted 84.72 ± 2.67 81.79 ± 2.82 79.50 ± 0.14 85.20 ± 1.04 81.35 ± 2.14 70.62 ± 1.15 0.82 ± .021 0.79 ± .007 0.75 ± .009
HoeffdingTree 89.22 ± 1.67 87.51 ± 2.23 87.23 ± 1.80 92.66 ± 0.92 87.76 ± 2.86 79.69 ± 0.53 0.89 ± .014 0.88 ± .001 0.85 ± .012

5 Conclusion, Discussion and Future Work

An ensemble-based framework for detecting intrusions in streaming logs has been

presented. The framework features a drift detection mechanism, allowing it to
cope with the non-stationary nature of intrusion attacks, and a wheel strategy
for removing inaccurate base models when such drift occur. The ensemble’s
combiner function is updated by using a GP method, meant to maximise the
ensemble’s accuracy on a small validation set. Preliminary test results on public
data confirmed that the framework is a viable and effective solution for detecting
intrusions in real-life application scenarios. Several issues major lines of extension
are open for this work, which are summarised below.

Integrating Outlier-Oriented Models. The base models currently used in our

ensemble approach are learnt by only reusing classifier-induction algorithms,
which need to be trained with labelled example tuples covering two classes (at
least): normal behaviour vs. intrusion attacks. To allow the framework to react
more promptly to novel attack types, we plan to also incorporate (unsuper-
vised) outlier detection models, and evaluate the ability of our ensemble learn-
ing scheme to effectively combine the anomaly scores returned by such base
models—along the recent research line of outlier ensembles [11].

Using More Expressive Combiners. In order to allow our framework to benefit

from the discovery of expressive combiner functions, one possible solution is to
enlarge the set of non-trainable functions that can be used as building block in
the GP search. To this end, in addition to the aggregation functions considered
in this work, we will investigate on using fuzzy-logics functions (e.g., t-norms,
t-corms), as well as generalised mixture functions [2], which compute a weighted
combination of their input variables where the weights are determined dynami-
cally, based on the actual values of the variables.
An even more flexible “context-aware” ensemble scheme could be obtained
when allowing the combiner function to possibly change the logics for combining
the predictions made by the base models on a given test instance, according to
the characteristics of the instance itself. This is in line with the classic area of
Mixture of Experts models [10] and the more recent one of Dynamic Ensemble
Selection/Weighting [3] approaches, where the final overall prediction returned
for a test instance x is (mainly) grounded on the predictions made by the base
 A Cybersecurity Framework for Classifying Non Stationary Data Streams 277

classifiers that look more competent for (the region of the instance space that
contain) x. How to efficiently and effectively extend our approach is a challenging
matter of study.

Other Work. A further direction of future work concerns adapting the fitness
score (used to guide the evolutionary search of the ensemble’s combiner function)
in a way that the degree of diversity among the base models is taken into account.
Indeed, if having a high level of diversity is a desideratum for an ensemble
classifier in general, it may become a key feature for making it robust enough
towards the dynamically changing nature of intrusion detection scenarios.

References
1. Bifet, A., Gavalda, R.: Learning from time-changing data with adaptive windowing.
In: SDM, vol. 7, pp. 443–448. SIAM (2007)
2. Costa, V.S., Farias, A.D.S., Bedregal, B., Santiago, R.H., Canuto, A.M.P.: Combin-
ing multiple algorithms in classifier ensembles using generalized mixture functions.
Neurocomputing 313, 402–414 (2018)
3. Cruz, R.M., Sabourin, R., Cavalcanti, G.D.: Dynamic classifier selection: recent
advances and perspectives. Inf. Fusion 41, 195–216 (2018)
4. Folino, G., Pizzuti, C., Spezzano, G.: A scalable cellular implementation of parallel
genetic programming. IEEE Trans. Evol. Comput. 7(1), 37–53 (2003)
5. Folino, G., Pisani, F.S.: Combining ensemble of classifiers by using genetic program-
ming for cyber security applications. In: Mora, A.M., Squillero, G. (eds.) EvoAp-
plications 2015. LNCS, vol. 9028, pp. 54–66. Springer, Cham (2015). https://doi.
org/10.1007/978-3-319-16549-3 5
6. Folino, G., Pisani, F.S., Sabatino, P.: A distributed intrusion detection framework
based on evolved specialized ensembles of classifiers. In: Squillero, G., Burelli, P.
(eds.) EvoApplications 2016. LNCS, vol. 9597, pp. 315–331. Springer, Cham (2016).
https://doi.org/10.1007/978-3-319-31204-0 21
7. Folino, G., Pisani, F.S., Sabatino, P.: An incremental ensemble evolved by using
genetic programming to efficiently detect drifts in cyber security datasets. In:
Genetic and Evolutionary Computation Conference, Companion Material Proceed-
ings, GECCO 2016, Denver, CO, USA, 20–24 July 2016, pp. 1103–1110 (2016)
8. Gama, J., Žliobaitė, I., Bifet, A., Pechenizkiy, M., Bouchachia, A.: A survey on
concept drift adaptation. ACM Comput. Surv. (CSUR) 46(4), 44 (2014)
9. Gonçalves Jr., P.M., de Carvalho Santos, S.G., Barros, R.S., Vieira, D.C.: A com-
parative study on concept drift detectors. Expert Syst. Appl. 41(18), 8144–8156
(2014)
10. Masoudnia, S., Ebrahimpour, R.: Mixture of experts: a literature survey. Artif.
Intell. Rev. 42(2), 275–293 (2014)
11. Micenková, B., McWilliams, B., Assent, I.: Learning outlier ensembles: the best
of both worlds-supervised and unsupervised. In: ACM SIGKDD 2014 Workshop
ODD2 (2014)
12. Shiravi, A., Shiravi, H., Tavallaee, M., Ghorbani, A.A.: Toward developing a sys-
tematic approach to generate benchmark datasets for intrusion detection. Comput.
Secur. 31(3), 357–374 (2012)
 A Dynamic Load Balancing Technique
for Parallel Execution of Structured
Grid Models

Andrea Giordano1 , Alessio De Rango2 , Rocco Rongo3 ,

Donato D’Ambrosio3 , and William Spataro3(B)
1
ICAR-CNR, Rende, Italy
[email protected]
2
DIATIC, University of Calabria, Rende, Italy
[email protected]
3
DEMACS, University of Calabria, Rende, Italy
{rocco.rongo,donato.dambrosio,william.spataro}@unical.it

Abstract. Partitioning computational load over different processing

elements is a crucial issue in parallel computing. This is particularly rel-
evant in the case of parallel execution of structured grid computational
models, such as Cellular Automata (CA), where the domain space is par-
titioned in regions assigned to the parallel computing nodes. In this work,
we present a dynamic load balancing technique that provides for perfor-
mance improvements in structured grid model execution on distributed
memory architectures. First tests implemented using the MPI technology
have shown the goodness of the proposed technique in sensibly reducing
execution times with respect to not-balanced parallel versions.

Keywords: Parallel computing · Parallel software tools · Load

balancing · Cellular Automata

1 Introduction
The computational demands of complex systems simulation, such as in Compu-
tational Fluid Dynamic (CFD), are in general very compute intensive and can be
satisfied only thanks to the support of advanced parallel computer systems. In
the field of Modeling and Simulation (M&S), approximate numerical solutions
of differential equations which rule a physical system (e.g., Navier-Stokes equa-
tions) are obtained by using parallel computers [4]. Classical approaches based
on calculus (e.g., Partial Differential Equations - PDEs) often fail to solve these
kinds of equations analytically, making a numerical computer-based methodol-
ogy mandatory in case of solutions for real situations. Discretization methods,
such as the Finite Element Method (FEM) or Finite Difference Method (FDM)
(c.f., [7,9,14,16]), which estimate values at points over the considered domain,
are often adopted to obtain approximate numerical solutions of the partial differ-
ential equations describing the system. Among these discrete numerical method-
ologies, Cellular Automata (CA) have proven to be particularly adequate for
c Springer Nature Switzerland AG 2020
Y. D. Sergeyev and D. E. Kvasov (Eds.): NUMTA 2019, LNCS 11973, pp. 278–290, 2020.
https://doi.org/10.1007/978-3-030-39081-5_25
 A Dynamic Load Balancing Technique for CA Execution 279

systems whose behavior can be described in terms of local interactions. Orig-

inally studied by John von Neumann for studying self-reproducing issues [27],
CA models have been developed by numerous researchers and applied in both
theoretical and scientific fields (c.f., [2,10,13,15,21,24–26,29,30,38]).
Thanks to their rule locality nature and independently from the adopted
formal paradigm, complex systems simulation based on the above numerical
methodologies can be straightforwardly implemented on parallel machines. Typ-
ically, parallelizations based on OpenMP, MPI and the more recent GPGPU app-
roach (e.g., with CUDA or OpenCL) have proven to be valuable solutions for effi-
cient implementations of computational models (e.g., [1,3,18,19,31,32,36,37]).
Nevertheless, computational layers based on higher level abstractions (i.e, by
using a Domain Specific Language - DSL) of the adopted paradigm have been
applied with success for simulating complex systems (e.g., [9,22,34]).
Parallel Computing applications require the best distribution of computa-
tional load over processing elements, in order to better exploit resources [23].
For instance, this is the case of CA which model the dynamics of topologically
connected spatial systems (e.g., lava or debris flows), in which the evolution
initially develops within a confined sub-region of the cellular space, and fur-
ther expanding on the basis of the topographic conditions and source location(s)
[6]. Indeed, when a data-parallel spatial decomposition is utilized, most com-
putation can take place where the so-called “active” cells are located, with the
risk of overloading the involved processing element(s). In this case, an effective
load-balancing technique can mitigate this issue.
Load Balancing (LB) consists in partitioning the computation between pro-
cessing elements of a parallel computer, to obtain optimal resource utilization,
with the aim of reducing the overall execution time. In general, the particular
parallel programming paradigm which is adopted and the interactions that occur
among the concurrent task determine the suitability of a static or dynamic LB
technique. Static LB occurs when tasks are distributed to the processing ele-
ments before execution, while dynamic LB refers to the case when the workload
is dynamically distributed during the execution of the algorithm. More specifi-
cally, if the computation requirements of a task are known a priori and do not
change during computation, a static mapping could represent the most appropri-
ate choice. Nevertheless, if these are unknown before execution, a static mapping
can lead to a critical imbalance, resulting in dynamic load balancing being more
efficient. However, if the amount of data to be exchanged by processors produces
elevated communication times (e.g., due to overhead introduced system moni-
toring, process migration and/or extra communication), a static strategy may
outperform the advantages of the dynamic mapping, providing for the necessity
of a static load balancing.
The simulation of complex physical phenomena implies the handling of an
important amount of data. Several examples of both static and dynamic LB
approaches can be find in literature for structured grid based models parallel
implementations. In CAMEL [11] authors adopt a static load balancing strat-
egy based on the scattered decomposition technique [28] which effectively can
280 A. Giordano et al.

reduce the number of non-active cells per processor. A dynamic load-balancing

algorithm is adopted in P-CAM [35]. P-CAM is a simulation environment based
on a computational framework of interconnected cells which are arranged in
graphs defining the cell interconnections and interactions. When a simulation
starts, a decomposition of these graphs on a parallel machine is generated and
a migration of cells occur thanks to a load balancing strategy. Another notable
example based on a dynamic LB technique is presented in [39], and applied to
an open-source parallel library for implementing Cellular Automata models, and
provides for meaningful performance improvements for the simulation of topolog-
ically connected phenomena. Other recent examples can be found in [12] referred
to multi-physics simulations centered around the lattice Boltzmann method, and
in [17] where authors study LB issues in decentralized multi-agent systems by
adopting a sandpile cellular automaton approach.
Accordingly, we here present an automatic domain detection feature that
dynamically (and optimally) balances computational load among processing ele-
ments of a distributed memory parallel computer during a simulation. The paper
is organized as follows. In Sect. 2 parallel decomposition techniques for CA mod-
els in distributed memory environments is discussed; subsequently, in Sect. 3, the
proposed dynamic load balancing technique for CA is presented while prelimi-
nary experimental results are reported in Sect. 4. Eventually, a brief discussion
and future developments section concludes the paper.

2 Parallel Execution of Cellular Automata in Distributed

Memory Architectures
As anticipated, the Cellular Automata (CA) computational paradigm is partic-
ularly suitable for describing some complex systems, whose dynamics may be
expressed in terms of local laws of evolution. In particular, CA can be fruitfully
adopted to model and simulate complex systems that are characterized by an
elevated number of interacting elementary components. Nevertheless, thanks to
their implicit parallel nature, CA can be fruitfully parallelized on different par-
allel machines to scale and speed up their execution. A CA can be considered
as a d-dimensional space (i.e., the cellular space), partitioned into elementary
uniform units called cells, representing a finite automaton (fa ). Input for each
fa is given by the cell’s state and the state of a geometrical pattern, invariant
in time and space over the cellular space, called the cell’s neighbourhood. Two
well-known examples of two-dimensional neighbourhoods are shown Fig. 1. At
time t = 0, the CA initial configuration is defined by the fa ’s states and the CA
subsequently evolves by applying the fa ’s transition function, which is simulta-
neously applied to each fa at each computational step. Despite their simplicity,
CA can produce complex global behavior and are equivalent to Turing Machines
from a computational viewpoint.
The CA execution on both sequential or parallel computers consists evalu-
ating the transition functions over the cellular space. In this case, at step t, the
 A Dynamic Load Balancing Technique for CA Execution 281

von Neuman Moore radius 2 Moore

Neighbourhood Neighbourhood Neighbourhood

Fig. 1. CA neighbourhoods: von Neumann and Moore neighbourhoods are the left
ones, respectively, both with visibility radius equal to 1, while a Moore neighbourhood
with visibility radius equal to 2 is shown on the right.

evaluation of a cell transition function takes as input the state of neighbour-

ing cells at step t − 1. This requires that the cell states at step t − 1 have to be
stored in memory when executing step t, obtaining the so-called “maximum par-
allelism” behaviour. This issue can be straightforwardly implemented by using
two matrices for the space state: the read and the write matrix. When executing
of a generic step, the evaluation of the transition function is obtained by reading
states from the read matrix and by then writing results to the write matrix.
When all of the transition functions of cells have been evaluated, the matrices
are swapped and the following CA step can take place.

Cellular Automata

Region 1 Region 2 Region 3 Region 4

(Node1) (Node 2) (Node 3) (Node 4)

Fig. 2. Cellular space partitioning in regions that are assigned to different parallel
computing nodes.

Being parallel computational models, CA execution can be efficiently paral-

lelized by adopting a typical data-parallel approach, consisting in partitioning
the cellular space in regions (or territories) and assigning each of then to a spe-
cific computing element [8,20] as shown in Fig. 2. Each region is assigned to a
different computing element (node), which is in responsible of executing the tran-
sition function of all the cells belonging to that region. Since the computation of
282 A. Giordano et al.

Node 1 Node 2

Cell
Neighbourhood

Fig. 3. The neighbourhood of a edge cell overlapping more regions

a transition function of a cell is based on the states of the cell’s neighbourhood,

this can overlap two regions for cells located at the edge part, as seen in Fig. 3.
Therefore, the transition function execution of cells in this area needs infor-
mation that belongs to a adjacent computing node. As a consequence, in order
to keep the parallel execution consistent, the states of these border cells (or halo
cells) need to be exchanged among neighboring nodes at each computing step.
In addition, the border area of a region (the halo cells) is divided in two different
sub-areas: the local border and the mirror border (see Fig. 4). The local border
is handled by the local node and its content replicated in the mirror border of
the adjacent node.

Node 1

Local Border
Node 1 Node 2
Mirror Border
{

Borders
{

Mirror Border

Local Border

Node 2

Fig. 4. Border areas of two adjacent nodes

 A Dynamic Load Balancing Technique for CA Execution 283

The parallel execution of a CA in a distributed memory architecture consists

in each node executing the loop illustrated in Algorithm 1. At each iteration,
the node sends its halo borders to neighbour nodes, receiving in turn the corre-
sponding halo cells (lines 3–4). Afterwards, the transition function is applied by
reading input values of cells from the read matrix and updating the new result
to the write matrix (i.e., recall the “maximum parallelism” behaviour described
above). After all cells have been have updated, a swap between the read and
write matrices occurs and the next CA step can be taken into account (line 7).

Algorithm 1. CA execution
1 while !StopCriterion() // Loop until CA stop criterion met
2 do
3 SendBorderToNeighbours() // Send halo borders to left and right
neighbour nodes
4 ReceiveBorderFromNeighbours() // Receive halo borders from left
and right neighbour nodes
5 ComputeTransitionFunction() // Read from read matrix and write
to write matrix
6 SwapReadWriteMatrices() // Swap read and write matrices
7 step ← step + 1 // Next CA step

3 A Dynamic Load Balancing Technique for Partitioned

Cellular Automata
The proposed LB algorithm consists in dynamically computing the optimal work-
load that each processing element has to take into account for achieving uniform
execution times over a simulation. The approach presented in this paper relies
on exchanging computation load among computing nodes at given steps on the
basis of a specific criterion. For example, the balancing of the load can occur at
a predefined rate of CA time steps, or when the execution times among nodes is
particularly unbalanced, and so on. The load balancing is actually achieved by
means of CA columns exchange among nodes. In particular, each node exchanges
columns with its neighbour nodes in a parallel fashion by exploiting the same
communication channel already in use for the halo exchange1 . During a simula-
tion, the execution times experienced by the nodes for executing the current step
are retrieved and stored in each node. When the LB phase has to be executed,
each node sends its step time to a specific “master” node, which is in charge of
establishing a suitable columns exchanges that nodes must perform in order to
achieve a balanced workload.
1
For the sake of simplicity, in this first implementation of the LB procedure the
exchange of columns is not toroidal, i.e. it is not possible to exchange columns
between the rightmost node and the leftmost node.
284 A. Giordano et al.

Algorithm 2. CA execution with Dynamic Load Balancing

1 while !StopCriterion() // Loop until CA stop criterion met
2 do
3 if LoadBalancingCriterion() // Load Balancing step reached
4 then
5 SendMyExecInfo() // send to Master my size and timings
6 SequenceOfFLows = ReceiveSeqFlows() // Receive sequence of
flows from Master
7 if IamLBMaster then
8 LBInfo=ReceiveExecInfo() // Receive sizes and timings
from all nodes
9 newRegionSizes = LoadBalance(LBInfo) // Determine new node
sizes
10 allSequenceFlows = ComputeFlows(NewRegionSizes)
// Determine sequence of flows for all nodes
11 SendFlows(allSequenceFlows) // Send sequence of flows to
nodes
12 forall f lows ∈ SequenceOf F Lows do
13 ExchangeLBFlows(flows)

// Back to normal CA execution

14 SendBorderToNeighbours()
15 ReceiveBorderFromNeighbours()
16 ComputeTransitionFunction()
17 SwapReadWriteMatrices()
18 step ← step + 1 // Next CA step

The new CA loop containing the code implementing the load balancing is
described in Algorithm 2. Please note that the normal CA execution takes place
at every step as reported at lines 14–18. As mentioned before, the LB is exe-
cuted when some conditions are met, i.e., when LoadBalancingCriterion() is
true (line 3). At line 5, each node sends its CA space size (i.e., the number
of columns), along with the elapsed time required for computing the last CA
step, to the master node (SendMyExecInfo()), and waits for receiving informa-
tion from this latter on columns to be exchanged to achieve load balancing (line
6). In particular, this information (SequenceOfFlows) consists in a sequence of
columns to be exchanged with the left and the right neighbours. The reason
behind why a sequence of columns exchange is necessary, rather than just an
simple exchange of a given number of columns to left and right nodes, will be
clarified in the following. At line 12–13 the actual columns exchange takes place.
In particular, each flows of the sequence, i.e, the columns to be exchanged with
the left and right neighbour, is in practice applied through ExchangeLBFlows().
Let us now summarize the master behaviour (lines 8–11). At line 8, the master
receives information about the nodes state (i.e., space size and elapsed times)
and determines, at line 9 (LoadBalance()), the new region sizes for the nodes,
 A Dynamic Load Balancing Technique for CA Execution 285

which minimize the unbalancing of the workload. On the basis of the new region
sizes, the determination of the flows of columns that must be exchanged among
nodes can be straightforwardly computed. However, it can be possible that some
of the determined flows may exceed the columns availability of a given node. For
instance, let us assume there are 3 nodes N1, N2 and N3, each having a 100
column CA space size right before the LB phase evaluation. Let us also assume
that the LoadBalance() function computes new region sizes as 70, 20 and 210.
In this case, in order to achieve 210 columns for N3, a flow of 110 columns should
be sent from N2 to N3 (recall that it is not possible to exchange columns between
N3 and N1, as reported in Sect. 2), though N2 hosts only 100 columns. In this
example, this issue can be addressed by simply considering 2 exchange phases.
In the first step, all the 100 columns between N2 to N3 are exchanged, while the
remaining 10 columns are exchanged in a second phase. Note that in the second
phase N2 hosts 30 columns, having received them from N1 in the first phase,
and so is now able to send the 10 columns to N3. The aforementioned sequence
of flows are computed by the master in the ComputeFlows() function (line 10)
and thus sent to all the nodes by the SendFlows() function (line 11).
It is worth to note that Algorithm 2 represents a general framework for achiev-
ing a dynamic load balancing during a simulation. Most of the functions seen in
the above pseudo-code do not require further specifications, except for the two
methods: LoadBalancingCriterion() and LoadBalance(). The implementa-
tion of these two functions determines when the LB should take place and how
to resize regions so as to achieve the “optimal” load balancing. In our preliminary
implementation, the load balancing occurs at a predefined rate of CA steps while
the LoadBalance() function follows an heuristic based on resizing the regions
taking into account the region time differences normalized with respect to their
old sizes. However, other strategies can be considered and linked to Algorithm 2
by implementing specific versions of the two methods just described.

4 Experimental Results

Preliminary experiments were carried out for testing the performance of the pro-
posed LB algorithm on the SciddicaT CA debris flow model [5]. The testbed is
composed by a grid of 296 columns × 420 rows, representing the DEM (Digital
Elevation Model) of the Tessina landslide, occurred in Northern Italy in 1992.
In order to create an initial unbalanced condition among processing nodes, land-
slide sources were located in the lower rows of the morphology, corresponding to
higher topographic elevations (see Fig. 5). As the simulation develops, the land-
slide expands to lower topographic altitudes, thus progressively interesting other
processing elements. The simulation was run for 4000 computational steps, cor-
responding to the full termination of the landslide event. Other parallelizations
(e.g., multi-node and multi-GPGPU implementations) performed on SciddicaT
on the same data set can be found in [9] and [33].
286 A. Giordano et al.

Fig. 5. Initial node partitioning for the Tessina landslide simulation. The initial land-
slide source in indicated in red. The upper numbering indicates the core id, the middle
indicates the node partitioning as number of columns. (Color figure online)

Fig. 6. Final load balanced configuration referred to the last step of the simulation,
with the landslide that has evolved towards the upper-right side of the CA space. Please
note the new node partitioning corresponding to a balanced configuration.
 A Dynamic Load Balancing Technique for CA Execution 287

4.1 Performances
The preliminary tests were performed on a 4-core i7 Linux-based PC with 8 GB
RAM. The opensource C++ OpenMPI 2.0 version of MPI was used for mes-
sage passing among processes. Besides a normal not-balanced execution, three
different load balancing tests were considered by considering different LB steps
applications, i.e., the LoadBalancingCriterion() is TRUE each 250, 400 and 800
steps, respectively, for each of the three LB experiments. Table 1 summarizes the
obtained results.

Table 1. Execution times of preliminary tests executed for assessing the performance of
the Load Balancing algorithm on the SCIDDICA CA debris-flow model. Four different
tests are reported, referred to Normal (i.e., not load balanced) execution, LB executed
each 250 steps, LB each 400 steps and LB each 800 steps, respectively.

Test Times (s) Improvement (%)

No LB 675 -
LB each 250 steps 503 25%
LB each 400 steps 501 26%
LB each 800 steps 484 28%

As seen, the application of the LB algorithm permitted an improvement up to

28% of the overall execution time (about 484 s versus 675 s of the not-balanced
algorithm). Furthermore, for these preliminary tests, improvements are noted
gradually as the number of the LB application step increases, proving that in
these experiments the LB algorithm indeed provides benefits in reducing execu-
tion times, though introducing, as expected, some degree of overhead which is
however counterbalanced by the aforementioned absolute time decrease. Even-
tually, Fig. 6 shows the last step of the simulation referred to the 800-step based
LB experiments. As noted, starting from an initial uniform node partitioning,
the simulation ends with a new node partitioning corresponding to a balanced
execution time node configuration.

5 Conclusions
We here present a dynamic load balancing feature that exploits computational
resources to reduce overall execution times in parallel executions of CA models
on distributed memory architectures. Specifically, the algorithm executes load
balancing among processors to reduce processor timings at regular intervals,
based to an algorithm which computes the optimal distribution load exchange
among adjacent nodes. Preliminary experiments, considering the SciddicaT CA
landslide model and executed for assessing the advantage of the dynamically load
balanced version with respect the non-balanced one, resulted in good improve-
ments on a standard 4-core i7 based PC. In particular, improvements up to 28%
288 A. Giordano et al.

were obtained when the LB algorithm is applied for the simulation of the 1992
Tessina landslide (Italy).
Further experiments will be carried out in order to compute most favorable
LB parameters (e.g., other LB steps), besides testing the algorithm with other LB
strategies as, for instance, considering a LB criterion only when elapsed times
between nodes are significant, and other heuristics to compute the new node
workload. For instance, automated optimization techniques such as evolutionary
algorithms or other heuristics, will be considered for calibrating LB parameters
referred to the particular parallel system and the adopted simulation model.
Future developments will regard the application of the LB algorithm on other
CA models, besides the extension on two-dimensional node partitioning, thus
permitting the application of the LB technique also on more complex network
topologies (i.e., meshes, hypercubes, etc.).

Acknowledgments. Authors thank BS student Rodolfo Calabrò from University of

Calabria for helping in code implementation and testing phases.

References
1. Abraham, M.J., et al.: GROMACS: high performance molecular simulations
through multi-level parallelism from laptops to supercomputers. SoftwareX 1, 19–
25 (2015)
2. Aidun, C., Clausen, J.: Lattice-Boltzmann method for complex flows. Annu. Rev.
Fluid Mech. 42, 439–472 (2010)
3. Amritkar, A., Deb, S., Tafti, D.: Efficient parallel CFD-DEM simulations using
OpenMP. J. Comput. Phys. 256, 501–519 (2014)
4. Andersson, B., Andersson, R., Håkansson, L., Mortensen, M., Sudiyo, R., van
Wachem, B.: Computational Fluid Dynamics for Engineers. Cambridge Univer-
sity Press (2011). https://doi.org/10.1017/CBO9781139093590
5. Avolio, M., et al.: Simulation of the 1992 Tessina landslide by a cellular automata
model and future hazard scenarios. Int. J. Appl. Earth Obs. Geoinf. 2(1), 41–50
(2000)
6. Cannataro, M., Di Gregorio, S., Rongo, R., Spataro, W., Spezzano, G., Talia, D.:
A parallel cellular automata environment on multicomputers for computational
science. Parallel Comput. 21(5), 803–823 (1995)
7. Cervarolo, G., Mendicino, G., Senatore, A.: A coupled ecohydrological-three-
dimensional unsaturated flow model describing energy, H2 O and CO2 fluxes. Eco-
hydrology 3(2), 205–225 (2010)
8. Cicirelli, F., Forestiero, A., Giordano, A., Mastroianni, C.: Parallelization of space-
aware applications: modeling and performance analysis. J. Netw. Comput. Appl.
122, 115–127 (2018)
9. D’Ambrosio, D., et al.: The open computing abstraction layer for parallel complex
systems modeling on many-core systems. J. Parallel Distrib. Comput. 121, 53–70
(2018). https://doi.org/10.1016/j.jpdc.2018.07.005
10. Di Gregorio, S., Filippone, G., Spataro, W., Trunfio, G.: Accelerating wildfire sus-
ceptibility mapping through GPGPU. J. Parallel Distrib. Comput. 73(8), 1183–
1194 (2013)
 A Dynamic Load Balancing Technique for CA Execution 289

11. Di Gregorio, S., Rongo, R., Spataro, W., Spezzano, G., Talia, D.: High performance
scientific computing by a parallel cellular environment. Future Gener. Comput.
Syst. 12(5), 357–369 (1997)
12. Duchateau, J., Rousselle, F., Maquignon, N., Roussel, G., Renaud, C.: An out-
of-core method for physical simulations on a multi-GPU architecture using lattice
Boltzmann method. In: 2016 International IEEE Conferences on Ubiquitous Intel-
ligence & Computing, Advanced and Trusted Computing, Scalable Computing and
Communications, Cloud and Big Data Computing, Internet of People, and Smart
World Congress, pp. 581–588. IEEE (2016)
13. Filippone, G., D’ambrosio, D., Marocco, D., Spataro, W.: Morphological coevo-
lution for fluid dynamical-related risk mitigation. ACM Trans. Model. Comput.
Simul. (ToMACS) 26(3), 18 (2016)
14. Folino, G., Mendicino, G., Senatore, A., Spezzano, G., Straface, S.: A model based
on cellular automata for the parallel simulation of 3D unsaturated flow. Parallel
Comput. 32(5), 357–376 (2006)
15. Frish, U., Hasslacher, B., Pomeau, Y.: Lattice gas automata for the Navier-Stokes
equation. Phys. Rev. Lett. 56(14), 1505–1508 (1986)
16. Cervarolo, G., Mendicino, G., Senatore, A.: Coupled vegetation and soil moisture
dynamics modeling in heterogeneous and sloping terrains. Vadose Zone J. 10, 206–
225 (2011)
17. Gasior, J., Seredynski, F.: A cellular automata-like scheduler and load balancer.
In: El Yacoubi, S., Was, J., Bandini, S. (eds.) ACRI 2016. LNCS, vol. 9863, pp.
238–247. Springer, Cham (2016). https://doi.org/10.1007/978-3-319-44365-2 24
18. Gerakakis, I., Gavriilidis, P., Dourvas, N.I., Georgoudas, I.G., Trunfio, G.A., Sir-
akoulis, G.C.: Accelerating fuzzy cellular automata for modeling crowd dynamics.
J. Comput. Sci. 32, 125–140 (2019)
19. Giordano, A., De Rango, A., D’Ambrosio, D., Rongo, R., Spataro, W.: Strategies
for parallel execution of cellular automata in distributed memory architectures,
pp. 406–413, February 2019. https://doi.org/10.1109/EMPDP.2019.8671639
20. Giordano, A., et al.: Parallel execution of cellular automata through space parti-
tioning: the landslide simulation Sciddicas3-Hex case study, pp. 505–510, February
2017
21. Higuera, F., Jimenez, J.: Boltzmann approach to lattice gas simulations. Europhys.
Lett. 9(7), 663–668 (1989)
22. Jammy, S., Mudalige, G., Reguly, I., Sandham, N., Giles, M.: Block-structured
compressible Navier-Stokes solution using the ops high-level abstraction. Int. J.
Comput. Fluid Dyn. 30(6), 450–454 (2016)
23. Kumar, V.: Introduction to Parallel Computing, 2nd edn. Addison-Wesley Long-
man Publishing Co., Inc., Boston (2002)
24. Langton, C.: Computation at the edge of chaos: phase transition and emergent
computation. Physica D 42, 12–37 (1990)
25. Lucà, F., D’Ambrosio, D., Robustelli, G., Rongo, R., Spataro, W.: Integrating
geomorphology, statistic and numerical simulations for landslide invasion hazard
scenarios mapping: an example in the Sorrento Peninsula (Italy). Comput. Geosci.
67(1811), 163–172 (2014)
26. McNamara, G., Zanetti, G.: Use of the Boltzmann equation to simulate lattice-gas
automata. Phys. Rev. Lett. 61, 2332–2335 (1988)
27. von Neumann, J.: Theory of Self-Reproducing Automata. University of Illinois
Press, Champaign (1966)
28. Nicol, D.M., Saltz, J.H.: An analysis of scatter decomposition. IEEE Trans. Com-
put. 39(11), 1337–1345 (1990)
290 A. Giordano et al.

29. Ninagawa, S.: Dynamics of universal computation and 1/f noise in elementary
cellular automata. Chaos, Solitons Fractals 70(1), 42–48 (2015)
30. Ntinas, V., Moutafis, B., Trunfio, G., Sirakoulis, G.: Parallel fuzzy cellular
automata for data-driven simulation of wildfire spreading. J. Comput. Sci. 21,
469–485 (2016)
31. Oliverio, M., Spataro, W., D’Ambrosio, D., Rongo, R., Spingola, G., Trunfio, G.:
OpenMP parallelization of the SCIARA Cellular Automata lava flow model: per-
formance analysis on shared-memory computers. Procedia Comput. Sci. 4, 271–280
(2011)
32. Procacci, P.: Hybrid MPI/OpenMP implementation of the ORAC molecular
dynamics program for generalized ensemble and fast switching alchemical simu-
lations (2016)
33. Rango, A.D., Spataro, D., Spataro, W., D’Ambrosio, D.: A first multi-GPU/multi-
node implementation of the open computing abstraction layer. J. Comput.
Sci. 32, 115–124 (2019). https://doi.org/10.1016/j.jocs.2018.09.012, http://www.
sciencedirect.com/science/article/pii/S1877750318303922
34. Reguly, I., et al.: Acceleration of a full-scale industrial CFD application with OP2.
IEEE Trans. Parallel Distrib. Syst. 27(5), 1265–1278 (2016). https://doi.org/10.
1109/TPDS.2015.2453972
35. Schoneveld, A., de Ronde, J.F.: P-CAM: a framework for parallel complex systems
simulations. Future Gener. Comput. Syst. 16(2–3), 217–234 (1999)
36. Spataro, D., D’Ambrosio, D., Filippone, G., Rongo, R., Spataro, W., Marocco, D.:
The new SCIARA-fv3 numerical model and acceleration by GPGPU strategies.
Int. J. High Perform. Comput. Appl. 31(2), 163–176 (2017). https://doi.org/10.
1177/1094342015584520
37. Was, J., Mróz, H., Topa, P.: Gpgpu computing for microscopic simulations of crowd
dynamics. Comput. Inform. 34(6), 1418–1434 (2016)
38. Wolfram, S.: A New Kind of Science. Wolfram Media Inc., Champaign (2002)
39. Zito, G., D’Ambrosio, D., Spataro, W., Spingola, G., Rongo, R., Avolio, M.V.: A
dynamically load balanced cellular automata library for scientific computing. In:
CSC, pp. 322–328 (2009)
 Final Sediment Outcome
from Meteorological Flood
Events: A Multi-modelling Approach

Valeria Lupiano1 , Claudia R. Calidonna2(B) , Elenio Avolio2 ,

Salvatore La Rosa4 , Giuseppe Cianflone4,6 , Antonio Viscomi4 ,
Rosanna De Rosa4 , Rocco Dominici4,6 , Ines Alberico3 , Nicola Pelosi3 ,
Fabrizio Lirer3 , and Salvatore Di Gregorio2,5
1
IRPI UOS, National Research Council, Rende, CS, Italy
2
ISAC UOS, National Research Council, Lamezia Terme, CZ, Italy
[email protected]
3
ISMAR UOS, National Research Council, Napoli, Italy
4
DIBEST, University of Calabria, 87036 Rende, CS, Italy
5
DeMaCS, University of Calabria, 87036 Rende, CS, Italy
6
EalCUBO, 87036 Arcavacata di Rende, CS, Italy

Abstract. Coastal areas are more and more exposed to the effects of cli-
matic change. Intense local rainfalls increases the frequency of flash floods
and/or flow-like subaerial and afterward submarine landslides. The over-
all phenomenon of flash flood is complex and involves different phases
strongly connected: heavy precipitations in a short period of time, soil
erosion, fan deltas forming at mouth and hyperpycnal flows and/or land-
slides occurrence. Such interrelated phases were separately modelled for
simulation purposes by different computational models: Partial Differen-
tial Equations methods for weather forecasts and sediment production
estimation and Cellular Automata for soil erosion by rainfall and sub-
aerial sediment transport and deposit. Our aim is to complete the model
for the last phase of final sediment outcome. This research starts from
the results of the previous models and introduces the processes con-
cerning the demolition of fan deltas by sea waves during a sea-storm
and the subsequent transport of and sediments in suspension by current
at the sea-storm end and their deposition and eventual flowing on the
sea bed. A first reduced implementation of the new model SCIDDICA-
ss2/w&c1 was applied on the partial reconstruction of the 2016 Bag-
nara case regarding the meteorological conditions and the flattening of
Sfalassà’s fan delta.

Keywords: Extreme event · Flood · Sediment transport and

deposition · Subaerial and subaqueous flow-like landslide · Modelling
and simulation

c Springer Nature Switzerland AG 2020

Y. D. Sergeyev and D. E. Kvasov (Eds.): NUMTA 2019, LNCS 11973, pp. 291–306, 2020.
https://doi.org/10.1007/978-3-030-39081-5_26
292 V. Lupiano et al.

1 Introduction
Climatic change in coastal areas increases dramatically the frequency of extreme
meteorological occurrences, which induce a chain of disastrous events: extremely
intense local rainfalls determine heavy soil erosion and can trigger flash floods
or debris/mud/granular flows, which, on the shore, can continue as subaqueous
debris flows or hyperpycnal flows at mouth of watercourses with the possibility
to evolve offshore in density currents. This study is interested in the last part
of this composite phenomenon, its evolution from the shoreline toward the open
sea.
Different scenarios may be considered: the flood (or the debris flow) that
reaches the sea as a hyperpycnal stream that flows on the sea bed and originates
a density current; a delta fan, which was produced by the flood, is demolished
by one or more sea-storm events of diverse intensity forming detrital submarine
flows, particular meteorological conditions could give rise to the suspension of
detrital matter and subsequent deposition, when the waves energy decreases.
The overall phenomenon is very complex, its modelling and computer simulation
(M&S) can be important in order to forecast the natural hazard and manage
risk situations.
Different computational approach were adopted for M&S of the diverse inter-
related phenomenological components, e.g. classic PDE approximation methods
for weather forecasting [1–3] and sediment production estimation [4]; an alter-
native computational paradigm, the Cellular Automata, was utilized for M&S of
“surface flows” [5], a CA methodology for M&S of complex systems was devel-
oped [6–8]. The most interesting models related to the phenomenological aspects
of final sediment outcome are treated in several CA studies of M&S: SCIDDICA-
ss2 and SCIDDICA-ss3 (Simulations through Computational Innovative meth-
ods for the Detection of Debris flow path using Interactive Cellular Automata
for subaerial, subaqueous and mixed subaerial and subsequent subaqueous land-
slides) [9–13], M&S of density currents of Salles et al. [14] which is partially
derived by a previous SCIDDICA version, SCAVATU (Simulation by Cellular
Automata for the erosion of VAst Territorial Units by rainfall) [15], RUSICA
(RUdimental SImulation of Coastal erosion by cellular Automata) [16], M&S of
hot mudflows [17]; M&S of long-term soil redistribution by tillage [18]; M&S of
soil surface degradation by rainfall [19]. The CA for surface flows [6] may be
regarded as a two-dimensions space, partitioned in hexagonal cells of uniform
size, the cell corresponds usually to a portion of surface; each characteristic,
relevant to the evolution of the system and relative to the surface portion cor-
responding to the cell, is individuated as a sub-state, the third dimension (the
height) features, e.g., the altitude, may be included among the sub-states of the
cell; each cells embeds an identical computing device, the elementary automa-
ton (ea), whose input is given by the sub-states of the six adjacent cells, the
CA evolves changing the state at discrete times simultaneously, according to the
transition function of the ea. The transition function accounts for the dynamics
of the system and is compound by a sequence of “elementary” processes.
 Final Sediment Outcome from Flood Events 293

SCIDDICA-ss2 was selected as a significant base for developing in incremen-

tal way a model able to account for the physical processes regarding the final
sediment outcome that are not present in SCIDDICA-ss2, but may opportunely
be imported by the other models SCAVATU and RUSICA. SCIDDICA-ss2 was
validated on the well-known 1997 Albano lake debris flow [10,11], this model
accounted for the following processes both in subaerial zones and in subaque-
ous ones, summing up: determination of debris outflows towards the adjacent
cells, altitude, kinetic head, debris thickness variation by detrital cover erosion,
kinetic head variation by turbulence dissipation, air-water interface effects on
the outflows. SCIDDICA-ss2 is able to reproduce the phenomenology of the
debris flows, which overcome the coastline and continue as submarine detrital
flows, while the following processes, concerning sea-storm events and their con-
sequences have to be introduced: energy transmission from subaqueous currents
and waves to granular matter in suspension, on the sea bed, in the flooded area
of the shoreline, processes of transport, suspension and sedimentation of gran-
ular matter. SCIDDICA, RUSICA and SCAVATU can consider different type
of flowing matter: debris, mud, sand, particles; in order to avoid confusion, the
general term granular matter will be adopted, granulometries will be specified if
necessary. A partial implementation of this extended model SCIDDICA-ss2/w1
was performed, the 2016 Bagnara flood event was considered for simulation in
its very final parts, when the Sfalassà’s fan delta [20,21] produced by the flood
was demolished during successive sea-storms. A partial reconstruction of this
last part, starting from the fan delta positioning and structure together with the
meteorological conditions during the sea-storms, permitted to “play” with many
possible scenarios in order to examine some possible final sediment outcomes.
The paper is organised as follows: next section introduces to the new version
of SCIDDICA, Sect. 3 presents the SCIDDICA-ss2/w&c1 partial implementa-
tion and the geological setting of the study area, the most interesting simulation
results are reported, and finally conclusions are discussed.

2 The Model SCIDDICAss2/w&c1

An outline of the model SCIDDICA-ss2/w&c1 is reported in this section, start-
ing from the original model SCIDDICA-ss2 new sub-states and new procedures
are introduced in order to account also for the phenomenology of dispersion of
sediments.

2.1 An Outline of the Original Model SCIDDICA-ss2

SCIDDICA-ss2 wants to capture the phenomenology of “surface flow” of granular

type, debris, mud and the like in terms of complexity emergence both in subaerial
areas and in subaqueous ones, it is able to model flow parting and confluence of
flows according to the morphology, soil erosion and so on, it is a two-dimensions
hexagonal CA model specified by the quintuple: <R, X, S, P, τ > where:
294 V. Lupiano et al.

– R = {(x, y)|(x, y ∈ Z) ∧ (0 ≤ x ≤ lx ) ∧ (0 ≤ y ≤ ly )} is the finite sur-

face with the regular hexagonal tessellation, that cover the region, where the
phenomenon evolves.
– X = (0, 0), (0, 1), (0, −1), (1, 0), (−1, 0), (1, 1), (−1, −1) identifies the geomet-
rical pattern of cells, which influence any state change of the generic cell,
identified as the central cell: X includes the central cell (index 0) and the six
adjacent cells (indexes 1, . . . , 6).
– S is the set of the ea states, they are specified in terms of sub-states of type
ground and flow sub-states.
• Ground sub-states: A is the cell altitude, D is the depth of soil erodable
stratum that could be transformed by erosion in granular matter.
• Granular matter sub-states: T is the average thickness of granular matter
of the cell, X and Y are the coordinates of its barycenter with reference
to the cell center, and K is its kinetic head.
• Flow sub-states: i E is the part of outflow, the so called “external flow”
(normalised to a thickness), that penetrates the adjacent cell i, 1 ≤ i ≤ 6,
from central cell, i XE and i YE are the coordinates of its barycenter with
reference to the adjacent cell center, i KE is its kinetic head, (six compo-
nents for each sub-state); i I is the part of outflow toward the adjacent
cell, the so called “internal flow”, (normalized to a thickness) that remains
inside the central cell, i XI and i YI are the coordinates of its barycenter
with reference to the central cell center, i KI is its kinetic head, (six com-
ponents for all the sub-states).
– P is the set of the global physical and empirical parameters of the phe-
nomenon, they are enumerated in the following list and are better explicated
in next section: pa is the cell apothem; pt is the temporal correspondence of a
CA step; padhw , padha are the air/water adhesion values, i.e. the landslide mat-
ter thickness, that may not be removed; pf cw , pf ca are the air/water friction
coefficient for the granular matter outflows; ptdw , ptda , pf cw , pedw , peda are
air/water parameters for energy dissipation by turbulence, air/water param-
eters for energy dissipation by erosion; pml is the matter loss in percentage
when the landslide matter enters into water; pmtw , pmta are the air/water
activation thresholds of the mobilization; perw , pera are the air/water pro-
gressive erosion parameters; pwr is the water resistance parameter.
– τ : S 7 → S is the deterministic state transition for the cells in R, basic
processes of the transition function are here sketched, note that ΔQ means
variation of the ground sub-state Q, ground sub-state Q of the adjacent cell i,
1 ≤ i ≤ 6, is specified as Qi ; the subscripts “w” and “a” in parameter names
are omitted, when the formula is considered valid both in water and air.
– Mobilization Effects. When the kinetic head value overcomes an opportune
threshold (K > pmt ), depending on the soil features and its saturation state,
a mobilization of the detrital cover occurs proportionally to the quantity
overcoming the threshold: per (K − pmt ) = ΔT = −ΔD = −ΔA (the detrital
cover depth diminishes as the granular matter thickness increases), −ΔK =
ped (K − pmt ) calculates the kinetic head loss.
 Final Sediment Outcome from Flood Events 295

– Turbulence Effect. The effect of the turbulence is modelled by a proportional

kinetic head loss at each SCIDDICA step: −ΔK = ptd K.
– Granular Matter Outflows. The computation of the outflows fi , 0 ≤ i ≤ 6
from the central cell toward the cell i (f0 , is the quantity that does not
flow) is performed in two steps: determination of the outflows minimizing the
differences of “heights” hi , 0 ≤ i ≤ 6 in the neighborhood by the “algorithm of
minimization of differences” [5,6], and computation of the shift of the outflows
with subsequent determination of external and internal outflow sub-states.
First step: The quantity d to be distributed from the central cell: d =
T − padh = Σ0≤i≤6 fi . “Heights” are specified as: h0 = A + K + padh ; hi =
Ai + Ti , 0 ≤ t ≤ 6; the values of flows are obtained in order to minimize
Σ0≤i<j≤6 (|(hi + fi ) − (hj + fj )|).
Second step: The si shift of fi with slope θi related to heights hi , h0 0 ≤
i ≤ 6 is modelled as the body barycentre’s movement on the slope:
si = vi + pt + g(sin√θi − pf ca cos θi )p2t , with g the gravity acceleration and
initial velocity vi = 2gH in the subaerial case;
si = (1 − exp(−pwr · pt ))vi · pt + g(sin θi − pf cw · cos θi pwr )) + g(sin θi − pf ca ·
cos θi )pt /pwr , the water resistance is considered for the subaqueous case, a
modified Stokes equations, it is adopted with a form factor proportional to
mass and g < g, accounting for buoyancy.
– Flows Composition. The new values of the ground and granular matter sub-
states are calculated according to the new values of flow sub-states.
– Air-Water Interface. An external flow from an air cell (altitude higher than
water level) to a water cell (altitude lower than water level) can imply a loss
of matter (water inside debris and fine grains) proportional to debris mass,
specified by pml ; it implies a correspondent loss of kinetic energy, determined
by kinetic head decrease.

At the beginning of the simulation, we specify the states of the cells in R,

defining the initial CA configuration. The initial values of the sub-states are
accordingly initialized. In particular, A assumes the morphology values except
for the detachment area, where the thickness of the landslide mass is subtracted
from the morphology value; T is zero everywhere except for the detachment
area, where the thickness of landslide mass is specified; D assumes initial values
corresponding to the maximum depth of the mantle of soil cover, which can be
eroded. All the values related to the remaining sub-states are zero everywhere.
At each next step, the function τ is applied to all the cells in R, so that the
configuration changes in time and the CA evolution is obtained.

2.2 Outline of the Preliminary Version of SCIDDICA-ss2/w&c1

SCIDDICA-ss2/w&c1 is a two-dimension hexagonal CA model, extension of

SCIDDICA-ss2, which is specified by the septuplet:
<Rw&c1 , Gw&c1 , Xw&c1 , Sw&c1 , Pw&c1 , τw&c1 , γw&c1 > where:
296 V. Lupiano et al.

– Rw&c1 = R
– Gw&c1 is the set of cells, which undergo to the influences of the “external
world”; in this case, they are the “underwater” cells, which are exposed to
the effect of the waves.
– Xw&c1 = X
– Sw&c1 is the set of the ea states, they are specified in terms of sub-states of
type ground, granular matter and flow sub-states. This model has to account
layers of matter of different granularity (n layers) and sub-states, related to
waves and currents.
• Ground sub-states are the same as in SCIDDICA-ss2.
• Granular matter sub-states for each layer j, 1 ≤ j ≤ n: Tj is the average
thickness of granular matter of the cell, Xj and Yj are the coordinates of
its barycenter with reference to the cell center, Kj is its kinetic head; Sj
is the granular matter of layer j in suspension in the cell, normalized to
a thickness.
• Flow sub-states for each layer j, 1 ≤ j ≤ n : i Ej is the part of outflow,
the so called “external flow” (normalized to a thickness), that penetrates
the adjacent cell i, 1 ≤ i ≤ n, from central cell, i XEj and i YEj are the
coordinates of its barycenter with reference to the adjacent cell center,
i
KEj is its kinetic head, (six components for each sub-state); i Ij is the
part of outflow toward the adjacent cell, the so called “internal flow”,
(normalized to a thickness) that remains inside the central cell, i XIj and
i
YEj are the coordinates of its barycenter with reference to the central cell
center, i KIj is its kinetic head, (six components for all the sub-states);
i
Sj is the part of suspended matter outflow (normalized to a thickness),
that penetrates the adjacent cell i, 1 ≤ i ≤ 6, from central cell, i XSj and
i
YSj are the coordinates of its barycenter with reference to the adjacent
cell center.
• Wave and current sub-states: Aw , the wave amplitude; Lw , the wave
length; Xw , Yw , component x − y of wave direction; the Cx , Cy , x − y
speed components of the surface current.
– Pw&c1 is the set of the global physical and empirical parameters of the phe-
nomenon, there are the same parameter of P except padhw , padha the air/water
adhesion values and padhw , padha the air/water friction coefficient for the
granular matter outflows; they are multiplied because take a different value
for each layer j, 1 ≤ j ≤ n : j padhw , j padha . Parameters, regarding the wave
demolition of the layers, are the activation thresholds of the mobilization j pmt
(j pmts for the suspension dynamics at the shoreline) and the progressive ero-
sion parameters j per (j pers for the suspension dynamics at the shoreline) for
each layer j, 1 ≤ j ≤ n.
– τw&c1 , contains all the elementary processes of τ , they are applied to each layer
according to proper sub-states and parameters; furthermore the following
processes are considered with only a type of granulometry for simplicity sake
in the exposition:
• Suspension by erosion for cells at the shoreline. When the wave amplitude
overcomes an opportune threshold (Aw > 1 pmts ), depending on the layer
 Final Sediment Outcome from Flood Events 297

features, a mobilization of the layer occurs proportionally to the quantity

overcoming the threshold: 1 pers (Aw − 1 pmts ) = ΔS1 = −ΔT1 (the layer’s
thickness diminishes as the matter in suspension increases). Note that
deposit process can follow immediately to the suspension process at the
shoreline.
• Suspension, deposit and deposit mobilization processes. These processes
derive from the model RUSICA [16,21], but they are adapted to the
phenomenological conditions of the sea storm: matter in suspension can
be increased by deposit erosion depending on the energy on the bottom
sea: 1 pers (eb − −1 pmts ) = ΔS1 = −ΔT1 , where eb is the energy at the
bottom sea; the wave energy inside a cell ec can maintain granular mat-
ter in suspension until a maximum density dmx = pmt (ec − pmne )/(−A)
depending on granulometry of the matter, otherwise precipitation of
a part of suspended matter occurs until the equilibrium is reached:
ΔT1 = −ΔS1 = S1 − Admx .
• Diffusion of the granular matter in suspension during a sea storm. A
diffusion mechanism, using the “algorithm of minimization of differences”
[5,6], is adopted, it is similar to that of the previous section, a value of K1
depending on the wave sub-states is introduced and A (value of central
cell) is considered constant in the neighborhood, because there is no slope
for matter in suspension, but only differences in density.
• Transport of the granular  matter in suspension by sea currents. The shift
s is specified as s = pt (Cx2 + Cy2 ), originating for the granular matter
in suspension, the outflows i S1 toward the adjacent cells according the Θ
i
angle of the current such that Σ1≤i≤6 S1 = S1 · s/pa . The slow deposition
process of granular matter transported by current is not considered.
– γw&c1 is the external influence function, that returns step by step (Z) the
significant values regarding waves and currents, it is split in the following
functions:
– γw : Z × Gw&c1 → Aw × Lw × Xw × Yw for generation of wave values cell by
cell of Gw&c1 according to the observations or meteorological forecasting.
– γc : Z × Gw&c1 → Cx × Cy for generation of current values cell by cell of
Gw&c1 according to the observations or meteorological forecasting.

SCIDDICA-ss2 and RUSICA coalesce with strong simplifications in this pre-

liminary version w&c1; layers of different granulometry are here considered sep-
arately, but phenomenological conditions of their possible innermost mixing are
tackled superficially and need simulations of real events in order to improve the
model by an accurate parametrization or introducing new sub-states, parame-
ters and elementary processes according to the methodology of incremental CA
modelling, furthermore the model is “cut” for reduced applications.
298 V. Lupiano et al.

3 Examples of Simulation of Demolition Process of Fan

Delta by Sea-Storm and Sediment Outcome: The
2015–16 Sfalassà Case

A summary of observation of demolition of fan delta by sea storms precedes in

this chapter the section specifying the partial implementation of SCIDDICA-
ss2/w&c1; the simulations of the demolition of the 2015 fan delta follow, at the
end, an hypothetical case of transport in suspension was simulated.

3.1 Summary of Observations Related to the Sfalassà’s Area and

Meteorological Events

The interested area is characterised as follows: the wide of continental shelf

varies from 120 m north of Marturano Promontory to 50 m at Cacili promon-
tory. It reaches the maximum extension, about 300 m, at south of Marturano
Promontory and narrows to about 100 m in the area close to Sfalassà’s mouth
to become again wider (about 180 m) southward (Calabria, Italy). The conti-
nental shelf has a slope ≤7◦ , it continues with a continental slope that between
20 m and 50 m depth has a slope range of 20◦ –30◦ , that becomes of 10◦ –20◦ at
higher depth. The submarine retrogressive scars confined at canyon head-walls
coincide with both Sfalassà and Favazzina fiumara mouth. The emerged beach
has a height varying between 0 and 3.5 m a.s.l. and a mean slope of 6◦ in accor-
dance with the slope of continental shelf. The volume of fan delta, grown in
consequence of flood event occurred on 2 November 2015 is of about 25.000 m3
this value is in agreement with the volume of fan delta formed in consequence of
flood events and the evaluation of yield sediment derived from EPM model [24].
From a meteorological point of view, the flood event was analyzed in detail
in a recent work [23], where different configurations of a mesoscale atmospheric
model were tested and several sensitivity tests were carried out. The precipita-
tion affected principally the eastern side of the Calabria, although huge quan-
tities of rain have been recorded on the southerly Tyrrhenian areas. An upper
level trough over Sicily, synergistically to a low-pressure area at the surface,
favored the advection of moist and warm air masses from the Ionian Sea towards
the eastern side of the Calabria. Considering the observed precipitation col-
lected by regional network of the “Centro Funzionale Multirischi” (http://www.
cfd.calabria.it), it can be seen how several rain gauges recorded precipitation
>500 mm for the whole event, some of which just located on the western slopes
of the Aspromonte Mountain, upstream to the Sfalassà’s area. The rain gauge
of Bagnara Calabra recorded 96 mm on 31 October and 112 mm on 1 Novem-
ber, demonstrating the extraordinariness of the event. Consequently to flood a
significant fan delta was formed at the mouth of Sfalassà watercourse by detri-
tal matter consisting of badly classified sediments from coarse sand 0.75 mm to
blocks of 1.0 m, an average value of 150 mm may be considered. The thickness
of this fan delta is approximately 5 cm in average, surface of the underwater
detrital cover, until the original coastline, was estimated 9250 m2 . Meteo-marine
 Final Sediment Outcome from Flood Events 299

Fig. 1. Interested area of Bagnara site and sequence of photographs showing effect of
flash-flood and sea storms

events (at least three strong sea storms) occurred during the period 2015.11.25
to 2016.01.28, they brought to destroy almost completely the fun delta. The first
sea storm (general information in Table 1st event) took away a thin strip of the
delta fan parallel to the beach, the stronger second sea storm (general infor-
mation in Table 2nd event) took away a larger strip of the delta fan parallel to
the beach, the last strongest sea storm (general information in Table 3rd event)
demolished the delta fan.
Regarding the storms, a marine-waves analysis was carried out. For the case
we considered high-resolution (0.125◦ × 0.125◦ ) simulations carried out with
the model WAM (WAve Model), developed by ECMWF (European Centre for
Medium-Range Weather Forecasts). The sea storms are classified reporting in
the tables the most significant marine parameters:
SW H (Significant wave height [m]); M W P (Mean wave period [s]); M W D
(Mean wave direction [degrees]); P P 1D (Peak period of 1D spectra [s]); HM AX
(Maximum individual wave height [m]); T M AX (Period corresponding to max-
imum individual wave height [s]).
300 V. Lupiano et al.

Table 1. Extreme events after flash flood occurred ranging November 2015 January
2016 here statistic values regarding model output WAM for the three entire period

ESWH (m) MWP [s] MWD [degrees] PP1D [s] HMAX [m] TMAX [s]
25–28/11/2015
MAX 1,85 6,52 297,92 7,43 3,59 5,90
MIN 0,98 5,80 266,58 6,19 1,88 4,80
AVE 1,32 6,24 283,30 6,62 2,55 5,62
std 0,26 0,18 10,88 0,40 0,53 0,22
03–08/01/2016
MAX 2,11 7,43 294,29 8,73 4,01 6,74
MIX 1,00 4,75 265,76 6,13 1,93 4,33
AVE 1,48 6,64 283,30 7,41 2,85 5,94
std 0,40 0,33 4,53 0,99 0,76 0,31
12–18/01/2016
MAX 2,43 7,62 317,07 9,06 4,63 6,82
MIX 0,89 5,32 275,50 6,17 1,69 5,07
AVE 1,53 6,60 294,97 7,71 2,92 5,93
std 0,43 0,56 12,17 0,87 0,81 0,43

The values are extrapolated by the full gridded output of WAM, in particular
at an offshore position located about 3 km away from the coastline. Such values
opportunely simulated with near-shore model amplify the effect especially for
the waves height. These parameters, for the different sea storms, are taken into
account for the SCIDDICA simulations.
The Table 1 shows considered events, according periods, reporting statistical
values for the entire storm.
The effect of sea storms may be deduced roughly by a series of photographs
of the area after the sea storms (see Fig. 1).
The most realistic scenario of the event considers that the strength of the waves
did not allow for suspension the matter, that constituted the fan delta because
of its granulometry, so granular matter, that was eroded by the strength of
the waves, flowed on the seabed without be significantly influenced by the cur-
rents [22].

3.2 Partial Implementation of SCIDDICA-ss2/w&c1 and First

Simulations
Implementation of SCIDDICA-ss2/w&c1 was performed on the previous one of
SCIDDICA-ss2 in language C++, introducing partially and adapting sub-states,
parameters and processes of RUSICA. An important limit regards the granu-
lometry of granular matter, there is only a layer, corresponding to the fan delta,
whose granulometry is an “equivalent” granulometry of a mixture of deposits.
 Final Sediment Outcome from Flood Events 301

This could be a solution in some cases, but it is not always satisfying, partic-
ularly when a very heterogeneous granulometry could involve both suspension
and flowing on the sea bed. The aim of this first version is to test the model
according phenomenological view-point, i.e. if the elementary processes account
for the main mechanisms of the overall phenomenon. Two cases are considered,
the real case of demolition of the fan delta by successive sea storms of different
intensity and the very hypothetical case of a sea storm in the same area, but
with suspension: if the granular matter would be very fine, the sea storm could
cause diffusion in suspension, then it could be transported by the currents at
the end of the sea storm. Note that the model provides for an adherence effect,
i.e. a smallest part of matter remains always in the cell (less than 1 mm) and
cannot be transported outside the cell; it permits to mark the flows.

3.3 Simulation of the Demolition of the 2015 Fan Delta of

Watercourse Sfalassà

Three sea storms are considered after the 2015 flash flood in the Bagnara area
that formed the fan delta Fig. 1. The first sea storm in the Table 1 is the shortest
lasting and lesser sea storm. It does not weight much, only a thin part facing the
sea of the fan delta is eroded and a small flow of the eroded part is channelized
toward the offshore depression; only a minimum quantity reaches it (Fig. 2b).
Simulation of the second sea storm starts from the final conditions imme-
diately after the first sea storm. The considered second storm 1 is longer and
stronger than the first one, it effects the central area of the fan delta; a subaque-
ous flow of granular matter begins to reach the depression (Fig. 2c).
Simulation of the third sea storm in Table 1 starts from the final condi-
tions immediately after the second sea storm. This sea storm is the longest and
strongest, the fan delta is destroyed except two small parts on the right and
on the left (Fig. 2d). Almost all the matter, that formed initially the fan delta,
reached the sea depression. The global evolution of the system in the simulation
reproduces significantly the real event: the erosion propagates progressively from
the water line to the internal fringe area.

3.4 A Very Hypothetical Case of Transport in Suspension

In Fig. 2 simulation steps of erosion of the Sfalassà’s fan delta and resulting
subaqueous flows of the eroded granular matter, thickness of granular matter is
reported: (a) initial erosion of the fan delta (initial conditions), (b) effect at the
end of the first sea storm corresponding approximately to 3 days and 6 h, (c)
effect at end of the second sea storm approximately with duration of 5 days and
12 h, (d) third storm (duration of 6 days and 21 h) dismantling delta fun i.e. area
until original coastline.
Regarding the calibration phase of parameters, the main effort regarded this
new part of model, i.e., the elementary process (suspension, deposit and deposit
mobilization) concerning the demolition of the fan delta; a simple trial and error
302 V. Lupiano et al.

Fig. 2. Simulation steps of erosion of the Sfalassà’s fan delta and resulting subaqueous
flows thickness: (a) step 1, initial erosion, (b) step 1350: end of the first sea storm,
(c) step 3550: end of the second sea storm approximately, (d) step 6300: the third sea
storm dismantling of the delta fun.

method was possible to be applied with satisfying results in comparison with the
partial reconstruction of the real case.
The second case consider a hypothetical sea storm on a not-realistic fan delta
as far composition in comparison with real case, its granulometry permits the
suspension after the delta demolition, after the sea storm, a constant current
intercepts the granular matter and transported it.
Initial position of the fan delta is the same as in the previous case, of course
with a different thin granulometry. The effect of the sea storm is just a diffusion of
the eroded granular matter in suspension (step 1000 of the simulation, (Fig. 3a).
The sea storm ceases at the step 2500, the further erosion at the step 2000 is
reported in the Fig. 3b).
After the sea storm, a hypothetical (and not realistic) current effected the
suspended granular matter and channelizes it in its direction, toward NE. The
transport path is reported clearly in (Fig. 3a) (step 3000, 500 steps of the trans-
 Final Sediment Outcome from Flood Events 303

Fig. 3. Simulation steps of hypothetical diffusion in suspension of the eroded granular

matter of the fan delta: (a) step 1000, (b) step 2000, (c) step 3000, (d) step 4000.

port in suspension). The last (Fig. 3d) step 4000 shows as the path continues to
be channelized in NE direction. Such a current does not exist and its distance
from the coast is improbable, but bathymetry data at disposal don’t permit to
see the evolution of the simulation for a sufficient space in order to evaluate the
model in the case of another direction of current; that is the reason so unnatural
distance from the coast has been hypothesized.
Regarding the calibration phase of parameters, a simple trial and error me-
thod was satisfying for the elementary processes concerning diffusion, suspension
and transport of granular matter for a hypothetical (not real) case.
The main effort regarded this new part of model, i.e., the elementary pro-
cess (suspension, deposit and deposit mobilization) concerning the demolition of
the fan delta; a simple trial and error method was possible to be applied with
satisfying results.
304 V. Lupiano et al.

4 Conclusions
SCIDDICA-ss2/w&c1 was defined on the base of two models CA model SCID-
DICA-ss2, that is a well-founded model, very successful for a large range of
applications and RUSICA still in a phase of development. Such a model is very
complex and our effort was the inclusion and refinement of elementary processes
for harmonizing the two models and introducing new features according the
incremental modelling method for macroscopic CA. Therefore, the implementa-
tion of the model was partial in order to focus itself on the phenomenology in a
first rough way and understand the factors that permit a correct emergence of
the overall phenomenon. This partial and rough implementation of SCIDDICA-
ss2/w&c1 was applied on a real case, whose data for simulation are incomplete
and approximate and on a hypothesized case, which isn’t realistic, considering
the geological characteristics of the same area of the first case, but interesting for
a comparison of the different behaviours in the same initial context. The devel-
opment of the simulated event in the first case may be considered successful in
the limits of the implementation if we consider the demolition of the fan delta by
the succession of the sea storms thanks to a good knowledge of meteorological
data. The final outcome of the sediments is correct but times for a complete
deposit in the depression could be different. About, the second case, the initial
diffusion in suspension of eroded matter of fan delta and the transport by current
was simulated satisfactory obviously only from a phenomenological viewpoint, a
real case with enough precise data is necessary for correcting and improving the
model. This is just a preliminary step. Further work and effort has to be pursued
in order to better outperform the model and its results due to presence of sev-
eral parameters necessary to describe a macro complex system such as this and
with a huge and long time involved area. In these case it is necessary to obtain
detailed and long term data to define better the parameters values interval and
consequently obtain a more precise pattern of sediments distribution.

References
1. Pielke Sr., R.A.: Mesoscale Meteorological Modeling, 3rd edn. Academic Press,
Boston (2013)
2. Kalnay, E.: Atmospheric Modeling, Data Assimilation and Predictability. Cam-
bridge University Press, Cambridge (2003)
3. Richardson, L.F.: Weather Prediction by Numerical Process. Cambridge University
Press, Cambridge (1922)
4. Larosa, S., et al.: A PyQGIS plug-in for the sediments production calculation based
on the Erosion Potential Method. In: Gao, W. (eds.) Frontiers of Earth Science
(2019, Subm.). ISSN 2095-0195 (print version), ISSN 2095-0209 (el. version)
5. Avolio, M.V., Di Gregorio, S., Spataro, W., Trunfio, G.A.: A theorem about the
algorithm of minimization of differences for multicomponent cellular automata.
In: Sirakoulis, G.C., Bandini, S. (eds.) ACRI 2012. LNCS, vol. 7495, pp. 289–298.
Springer, Heidelberg (2012). https://doi.org/10.1007/978-3-642-33350-7 30
 Final Sediment Outcome from Flood Events 305

6. Di Gregorio, S., Serra, R.: An empirical method for modelling and simulating some
complex macroscopic phenomena by cellular automata. Futur. Gener. Comput.
Syst. 16, 259–271 (1999)
7. Calidonna, C.R., Di Gregorio, S., Mango Furnari, M.: Mapping applications of
cellular automata into applications of cellular automata networks. Comput. Phys.
Commun. 147, 724–728 (2002)
8. Calidonna, C.R., Naddeo, A., Trunfio, G.A., Di Gregorio, S.: From classical infinite
space-time CA to a hybrid CA model for natural sciences modeling. Appl. Math.
Comput. 218, 8137–8150 (2012)
9. Calidonna, C.R., Di Napoli, C., Giordano, M., Mango Furnari, M., Di Gregorio,
S.: A network of cellular automata for a landslide simulation. In: ICS 2001, pp.
419–426 (2001)
10. Mazzanti, P., Bozzano, F., Avolio, M.V., Lupiano, V., Di Gregorio, S.: 3D numer-
ical modelling of submerged and coastal landslides propagation. In: Mosher, D.C.,
et al. (eds.) Submarine Mass Movements and Their Consequences. NTHR, vol.
28, pp. 127–139. Springer, Dordrecht (2010). https://doi.org/10.1007/978-90-481-
3071-9 11
11. Avolio, M.V., Lupiano, V., Mazzanti, P., Di Gregorio, S.: An advanced cellular
model for flow-type landslide with simulations of subaerial and subaqueous cases.
In: Proceedings of the EnviroInfo 2009, vol. 501, pp. 131–140 (2009)
12. Maria, V.A., Francesca, B., Di Gregorio, S., Valeria, L., Paolo, M.: Simulation of
submarine landslides by cellular automata methodology. In: Margottini, C., Canuti,
P., Sassa, K. (eds.) Landslide Science and Practice, pp. 65–72. Springer, Heidelberg
(2013). https://doi.org/10.1007/978-3-642-31427-8 8
13. Avolio, M.V., Di Gregorio, S., Lupiano, V., Mazzanti, P.: SCIDDICA-SS3 : a new
version of cellular automata model for simulating fast moving landslides. J. Super-
comput. 65(2), 682–696 (2013). ISSN 0920-8542
14. Salles, T., Lopez, S., Cacas, M.C., Mulder, T.: Cellular automata model of density
currents. Geomorphology 88(1–2), 1–20 (2007)
15. D’Ambrosio, D., Di Gregorio, S., Gabriele, S., Gaudio, R.: A cellular automata
model for soil erosion by water. Phys. Chem. Earth Part B 26(1), 33–39 (2001)
16. Calidonna, C.R., De Pino, M., Di Gregorio, S., Gullace, F., Gullı̀, D., Lupiano, D.:
A CA model for beach morphodynamics. In: AIP Conference Proceedings. Numer-
ical Computations: Theory and Algorithms (NUMTA 2016), vol. 1776 (2016)
17. Arai, K., Basuki, A.: Simulation of hot mudflow disaster with cell automaton and
verification with satellite imagery data. Int. Arch. Sci. Photogramm. Remote Sens.
Spat. Inf. 38 Part 8, 237–242 (2010)
18. Vanwalleghem, T., Jiménez-Hornero, F., Giráldez, J.V., Laguna, A.M.: Simulation
of long-term soil redistribution by tillage using a cellular automata model. Earth
Surf. Process. Landf. 35, 761–770 (2010)
19. Valette, G., Prévost, S., Laurent, L., Léonard, J.: SoDA project: a simulation of
soil surface degradation by rainfall. Comput. Graph. 30, 494–506 (2006)
20. Punzo, M., et al.: Remocean X-band wave radar for wave field analysis: case study
of Bagnara Calabra (South Tyrrhenian Sea, Italy). J. Sens. 1, 131–140. Applica-
tion of X-BandWave Radar for Coastal Dynamic Analysis: Case Test of Bagnara
Calabra (South Tyrrhenian Sea, IT) 501, 9 (2016). Article ID: 6236925
21. Gullace, F.: Un tentativo di Modellizzazione della Morfodinamica degli Arenili
con Automi Cellulari. Master theses in Physics, Department Physics, University
of Calabria, aa 2015/16 (2016)
22. Dominici R.: Personal communication (2019)
306 V. Lupiano et al.

23. Avolio, E., Federico, S.: WRF simulations for a heavy rainfall event in southern
Italy: verification and sensitivity tests. Atmos. Res. 209, 14–35 (2018)
24. Auddino, M., Dominici, R., Viscomi, A.: Evaluation of yield sediment in the
Sfalassà Fiumara (south western, Calabria) by using Gavrilovi method in GIS
enviroment. Rendiconti online della Soc. Geol. Ita. 33, 3–7 (2015)
 Parallel Algorithms for Multifractal
Analysis of River Networks

Leonardo Primavera1(B) and Emilia Florio2

1
Dipartimento di Fisica, Università della Calabria,
Cubo 31/C, Ponte P. Bucci, 87036 Rende, CS, Italy
[email protected]
2
Dipartimento di Matematica e Informatica, Università della Calabria,
Cubo 30/B, Ponte P. Bucci, 87036 Rende, CS, Italy
[email protected]

Abstract. The dynamical properties of many natural phenomena can

be related to their support fractal dimension. A relevant example is the
connection between flood peaks produced in a river basin, as observed
in flood hydrographs, and the multi-fractal spectrum of the river itself,
according to the Multifractal Instantaneous Unit Hydrograph (MIUH)
theory. Typically, the multifractal analysis of river networks is carried
out by sampling large collections of points belonging to the river basin
and analyzing the fractal dimensions and the Lipschitz-Hölder expo-
nents of singularities through numerical procedures which involve dif-
ferent degrees of accuracy in the assessment of such quantities through
different methods (box-counting techniques, the generalized correlation
integral method by Pawelzik and Schuster (1987), the fixed-mass algo-
rithms by Badii and Politi (1985), being some relevant examples). How-
ever, the higher accuracy in the determination of the fractal dimensions
requires considerably higher computational times. For this reason, we
recently developed a parallel version of some of the cited multifractal
methods described above by using the MPI parallel library, by reaching
almost optimal speed-ups in the computations. This will supply a tool
for the assessment of the fractal dimensions of river networks (as well as
of several other natural phenomena whose embedding dimension is 2 or
3) on massively parallel clusters or multi-core workstations.

Keywords: Multifractal dimension · River networks · Parallel

algorithms

1 Introduction
The multifractal analysis is a powerful tool in many fields of science to relate
the geometric characteristics of objects with their physical properties. Since the
pioneering works by Mandelbrot [10], Grassberger [8], Grassberger and Procaccia
[9], many contributions of the multifractal analysis have been given in physics,
chemistry, biology and engineering.
c Springer Nature Switzerland AG 2020
Y. D. Sergeyev and D. E. Kvasov (Eds.): NUMTA 2019, LNCS 11973, pp. 307–317, 2020.
https://doi.org/10.1007/978-3-030-39081-5_27
308 L. Primavera and E. Florio

In particular, in hydrology, the multifractal analysis has given considerable

contributions. De Bartolo et al. [4], showed that river networks are multifractal
objects and [3] that they are non plane-filling structures (their fractal support
dimension is lesser than two). Moreover, it was shown that there is a strict
correlation between the hydrological response of a basin and its multifractal
spectrum through the Multifractal Instantaneous Unit Hydrograph (MIUH) [2]1 .
The assessment of the multifractal spectrum can be realized through the
following procedure. First, a 2D projection of the river basin on the ground plane
is realized and the different channels belonging to the river basin are represented
through the so-called “blue-lines”. Second, a regular sample of the blue-lines is
extracted by picking up a set of points, the so-called “net-points”, which are
representative of the river basin. Finally, a multifractal analysis is performed on
this set of points through several possible techniques. The most used methods
are: fixed-size algorithms (FSA), the correlation integral (CI) by Pawelzik and
Schuster [11], and fixed-mass algorithms (FMA) by Badii and Politi [1].
However, some methods can be more relevant than others for hydrological
studies. The FSA is by far the easiest to be implemented. It is based on the idea
of partitioning the multifractal set in squares of decreasing size and counting
the number of points (the so called “mass”) falling into each element of the
partition. The fractal dimension of the object can then be obtained through a
suitable scaling procedure of the mass as a function of the size of the partition.
The FMA uses the opposite approach, namely one fixes the mass in each partition
and then computes the size of the circle containing the given mass inside it.
As pointed out by De Bartolo et al. [7], the important parameter for the
hydrological response of the basin, directly used in the MIUH, is the D−∞ frac-
tal dimension, which corresponds to the right part of the multifractal spectrum.
However, FSAs are not well suitable to give a correct assessment of this parame-
ter, since they present strong oscillations in the scalings of τq for negative values
of q, that corresponds just to the right part of the multifractal spectrum. The
correlation integral method by Pawelzik and Schuster, as suggested by De Bar-
tolo et al. [7], seems to be able to improve this situation, although the most
suitable method seems to be the FMA. This happens because the oscillations
in the scalings observed in the FSAs are mainly due to the poor statistics of
net-points in the partitions where there is a less dense distribution of the points.
The FMA, on the converse, does not suffer for this problem, since the number
of net-points in each subset of the partition is fixed a priori and the size of the
subsets is computed accordingly, which is the opposite of what happens in FSAs.
However, FMA can be very expensive, from the computational point of view,
due to the procedure needed to compute the size of the partition once the “mass”
is chosen. This problem was solved in the past by extracting randomly chosen
net-points and applying the algorithm only on such selected sub-sets. On the
one hand, this allows one to obtain the results of the multifractal analysis in
reasonable CPU times, on the other hand, this sampling of the net-points lowers

1
The problem of the relation between the floods in rivers and meteorologic forecasting
has interested scientists in different epochs. See, for instance [6].
 Parallel Algorithms for Multifractal Analysis of River Networks 309

the precision of the calculation, as well as making useless to select large sets of
net-points to improve the representation of the river network.
However, nowadays, with the development of multi-cores CPUs, which can be
found even on moderately expensive personal computers, workstations or even
small clusters, one can make an effort to parallelize the code in order to exploit
the power of parallel CPUs or clusters to improve the computational times. At
the best of our knowledge, this attempt has never been done for this algorithm.
This contribution will try to fill this gap.
The organization of the contribution is the following: in next section, we
briefly recall the fixed-size and fixed-mass techniques; then, we explain the par-
allelization strategy we adopted to speed-up the numerical computation; after
that, to be sure of the results of our numerical procedure, we make a comparison
between the numerical and theoretical results obtained by analyzing a random
multifractal for which the fractal dimensions and the multifractal spectra are
known; finally we show the results of the parallelization in terms of the scaling
of the computational time with the number of cores, computed on a relatively
small multi-node cluster. Finally, in the last section, we draw some conclusions
and future outlooks.

2 Fixed-Size and Fixed-Mass Multifractal Analysis

Both the FSA and FMA multifractal rely on the study of the scaling properties
of the set of points representing the multifractal object (let us call N the number
of such points). The assessment of the fractal dimension is then carried out by
partitioning the set in a number nc of non-overlapping cells, with a size i ,
i = 1, . . . , nc . Let us call “measure” or “mass” pi (i ), the number of points
falling inside each cell. For a fractal set, it is possible to show that:
nc q−1 τ
i=1 pi i
lim =k (1)
i →0 nc
where k is a constant for a suitable choice of the real numbers q and τ , which
are related to the fractal dimension of the object.
For a multifractal, the values of q and τ are not unique, but a ensemble of
solutions q, τq exists, that satisfies the relation (1). These two quantities are
related to the generalized fractal dimensions Dq of the set through the relation:
τq
Dq = , ∀q = 1 (2)
q−1
Finally, the Lipschitz-Hölder exponents αq , are given by the relation:
dτq
αq = (3)
dq
while the multifractal spectrum f (αq ) is defined as:

f (αq ) = qαq − τq (4)

310 L. Primavera and E. Florio

In practical terms, the determination of the solutions (q, τq ) is performed

numerically in the following way. For FSA, for instance the famous “box-
counting” algorithm, the dimension i =  is chosen the same for all the subsets
of the partition, in such a way that Eq. (1) gives:

lim log M (, q) = τq lim log  + log k (5)

→0 →0

where: nc
i=1 pq−1
i
M (, q) =
nc
is the moment of order q − 1 of the “mass”. Numerically, it is impossible to take
the limit for vanishing values of , therefore the assessment of the solution τq
of Eq. (5) is carried out by fixing a value of q, by choosing a sequence of expo-
nentially decreasing values of the size  of the partitioning cells and “counting”
the “mass” pi falling into each of the intervals and finally computing log M (, q)
for each value of log . Then τq will result as the slope of the linear best-fit
approximation of this curve, for each fixed value of q. The drawback of this
method is easily understood: for negative values of q, in the subsets of the par-
tition where few points are present (namely for small values of pi ), the function
M (, q) exhibits strong, nonphysical, oscillations for decreasing values of , due
to the poor statistics in those cells. Negative values of q correspond to the right
part of the multifractal spectrum f (αq ), whose fractal dimension D−∞ enters in
the MIUH forecasting model. Therefore, the assessment of D−∞ accomplished
through FSA can be affected by strong errors and should be adopted with par-
ticular care.
The approach followed in the FMA is opposite. In this case, the “mass” pi = p
is chosen the same for each subset of the partition and the corresponding “size”
of the partitioning cells is evaluated by finding, for each point, the p nearest
neighbors. The radius of the circle containing that “mass” is then taken as the
size i of the partitioning cell. For a fixed value of τq , then, the mass p can be
taken out from the summation, by yielding, for Eq. (1) a relation of the form:

lim log M (i , τq ) = (1 − q) lim log p + log k (6)

i →0 p→0

where we used the fact that p → 0 for i → 0, and:

nc −τq
i=1 i
M (i , τq ) = (7)
nc
is the moment of order −τq of the size i of the partitioning cells. Finally, the solu-
tion 1 − q, and therefore q, will result from computing the scaling of log M (i , τq )
with log p, for exponentially decreasing values of p, as in Eq. (6).
In the latter case, since the “mass” content in each cell of the partition is fixed
a-priori, the problems present with the FSA are completely overcome. However,
the computational times required to complete the analysis can be really very
long, even for a total number of net-points N relatively small (the analysis of
 Parallel Algorithms for Multifractal Analysis of River Networks 311

104 net-points may require up to three days of computation on a workstation

equipped with a Xeon E3-1225 V2 CPU, at 3.20 GHz, on a single core). The true
bottleneck of the computation lies in the search of the nearest neighbors points
for each single net-point, which is a N 2 algorithm and, therefore, the CPU time
increases very quickly with the number of points.
A way to mitigate this problem and make the FMAs attractive for realis-
tic cases found in hydraulic science (where the number of net-points can easily
exceed N = 105 ÷106 ) is the parallelization of the algorithm. In the past, another
approach widely adopted was to limit the search of the nearest neighbors points
to a randomly extracted sub-set of the net-points, so decreasing the complex-
ity of the computation, but also the precision of the assessment of the fractal
dimensions. Although the parallelization of the algorithm is absolutely feasible,
at the best of our knowledge this was never attempted before. In next section we
describe how we parallelized the numerical code and, afterwards, we show how
the scaling with the number of processors is not far away from an almost linear
speed-up, which indicates a promising way to obtain the results of the analysis
in reasonable times in realistic cases.

3 Parallelization of the Code

In order to parallelize the code in such a way that it can be run on the largest
possible configurations (either for single- and multi-processor machines, or on
massively parallel clusters) we used the Message-Passing-Interface (MPI) parallel
library.
The serial FMA works in the following way:
1. the coordinates of the points of the fractal object are read from a file and put
into a vector, each element being a structure containing two records, one for
the x and the other for the y coordinates;
2. according to the input data, the values for τq are selected and an exponentially
increasing partitioning scheme for the values of nearest neighbors points is
set-up, that corresponds to a partitioning in the mass pi ;
3. in a main loop, for each point, a vector is built with all the distances between
the point under consideration and the other points of the fractal set;
4. from this vector, all the distances including the first pi nearest neighbors
points are selected, which corresponds to consider the radius of the circle
containing pi nearest neighbors (or, in other words, the sizes i of the parti-
tion);
5. the logarithms of the quantities M (i , τq ) for all values selected for p are
computed;
6. these quantities are finally written in a file, along with the logarithm of the
mass p;
At the end of the simulation, a file containing the scalings of log M (i , τq ) as
a function of log pi , for each fixed value of τq , is available. The corresponding
values of q are obtained by plotting these quantities for the different values of
312 L. Primavera and E. Florio

τq , so that the corresponding values of 1 − q (and, therefore, of q) are obtained

through a best-fit procedure, as the angular coefficient of the linear relation (6).
This gives us the curve τq as a function of q. The Lipschitz-Hölder exponents
αq can then be computed from the values of τq through Eq. (3). This can be
accomplished in several ways: in some cases (when the number of values of
(q, τq ) is not very high), it may be convenient to interpolate the curve τq with
a polynomial function and then compute the values of αq as the derivatives of
this interpolating function. Another way to do this, more suitable when a higher
number of values for (q, τq ) is available, is to numerically compute the derivative
by using, for instance, a finite difference formula. A little problem with this
approach comes from the fact that the values of τq are equally spaced, while the
values of q, being obtained as the results of the linear interpolation procedure
of Eq. (6), are not. This can be overcome, for instance, by choosing a centered
scheme for the first derivative with a different step-size on the left and right side
of the value of q under consideration.
We chose this second approach and computed the derivatives of τq by using
the formula:

dτq  h2 h2 − h1 h1
=− τq (qi−1 ) + τq (qi ) + τq (qi+1 ) (8)
dq q=qi h1 (h1 + h2 ) h1 h2 h2 (h1 + h2 )
where: h1 = qi − qi−1 , h2 = qi+1 − qi , and the numerical error is proportional to
the product h1 h2 . In the case when h1 = h2 this scheme reduces to the simple
second order central difference scheme.
It is immediately understandable that the most computationally expensive
point in the algorithm is the third one, namely the evaluation of the distances
of each point from the others, in order to obtain the nearest neighbors. This is a
N 2 algorithm. Our parallelization strategy has focused mainly on this point. We
start all MPI processes at the beginning of the code and let all the computational
cores read the points of the fractal set and store them in a vector. We adopted
this approach in order to avoid subsequent communications among the processes
in order to exchange parts of this vector, which would result in higher latency
times and a worse scaling of the computational time. This may appear odd, but
it is actually not a serious problem in the majority of situations. For instance,
the memory requirement per core to store this vector would be 16 × N bytes for
a double precision input file made of N net-points. For instance, for N = 106
this would require a memory allocation of less than 16 MB per core, which is
fairly easy to find on the majority of modern machines. This is also the main
contribution to memory allocation, all the other vectors being of very small sizes.
Once each core has its own private copy of the net-points, it is assigned a
specific interval of points on which it has to operate. However, thanks to the fact
that the coordinates of the points are known to each core, the distances between
each point in this interval from all the other ones can be easily calculated. There-
fore, also the evaluation of the nearest neighbors for each of such points can be
done in parallel (point 4) and the evaluation of the moments (point 5) can be
performed “locally” on each core. However, this last point requires a final aver-
age of the moments for all the points, according to Eq. (7). This is accomplished
 Parallel Algorithms for Multifractal Analysis of River Networks 313

with an MPI REDUCE call, which adds up all the contributions to the moments on
each core and sends the average to the master process. The file with the results
(point 6 above) is then written only on the master process. This introduces a
serial part in the code, whose weight is anyway considerably small, compared to
the rest of the computation.

Fig. 1. A random multifractal set obtained according to the multiplicative process

described by (9) with N = 106 points.

4 Numerical Results
4.1 Results of the Analysis for a Random Multifractal
In order to check the validity of the parallel algorithm, we ran several analy-
ses either with the serial or the parallel version of the code on a deterministic
multifractal. The multifractal set is obtained according to a recursive procedure
by starting from a random seed (see, for instance, Falconer [5]), in the following
way: let us start from the unit square Q = [0, 1] × [0, 1] and let us consider the
following transformations:
1 1
S1 : (x, y) → (x , y  ) = ( x, y)
2 2
1 1
S2 : (x, y) → (x , y  ) = ( (x + 1), y)
2 2
1 1
S3 : (x, y) → (x , y  ) = ( x, (y + 1))
2 2
1 1
S4 : (x, y) → (x , y  ) = ( (x + 1), (y + 1))
2 2
314 L. Primavera and E. Florio

Fig. 2. Comparison between theoretical and numerical results for τq . Theoretical

curves, obtained through Eq. (9) are plotted as red lines, numerical values are plot-
ted as diamond marks. (Color figure online)

These transformations map a point of the original unit square Q into one of
the four squares Qi (i = 1, . . . , 4) of side 1/2 in which Q can be partitioned. We
assign to each of the four squares Qi , the following probabilities: p1 = 1/10; p2 =
2/10; p3 = 3/10; p4 = 4/10. We then generate a set of N random numbers in
the interval [0, 1[. If the number falls in the interval [0, 0.1[ (that happens with
a probability p1 ), or [0.1, 0.3[ (happens with a probability p2 ), or [0.3, 0.6[ (with
a probability p3 ) or, finally, [0.6, 1[ (with a probability p4 ), the transformation
corresponding to that probability is applied and by iterating this procedure, one
obtains the multifractal set visualized in Fig. 1.
The theoretical prediction for τq is given by the relation:
−τq q −τq q −τq q −τq q
r1 p1 + r2 p2 + r3 p3 + r4 p4 =1

where ri are the “contraction rates” used to construct the multifractal set. In
our case: ri = 1/2, therefore, the value of τq is given by:
4
log( i=1 pq1 )
τq = − (9)
log 2

Finally, the theoretical values for αq and f (αq ) can be computed through Eqs. (3)
and (4).
In Figs. 2 and 3, the theoretical (red lines) and numerical curves (black dia-
monds) are shown for τq and f (αq ), respectively. As visible, the agreement of
 Parallel Algorithms for Multifractal Analysis of River Networks 315

the theoretical and numerical curves for τq is almost perfect. A similar situation
holds for the f (αq ) plot, although some small differences are present in the right
part of the spectrum, close to the maximum. However, this is probably due to
the approximations introduced by the finite difference formula (8) we used for
the evaluation of αq .

Fig. 3. Comparison between theoretical and numerical results for f (αq ). Theoretical
curves are plotted as red lines, numerical values are plotted as diamond marks. (Color
figure online)

4.2 Results of the Parallelization

In order to evaluate the effectiveness of the parallelization procedure, we com-
puted the speed-up of the code. We ran the FMA analysis on the random theo-
retical multifractal shown in Fig. 1, with N = 106 points.
It is worth mentioning that the numerical results cannot be exactly the same
for all the runs. This is due to the fact that the computation of the moments
M (i , τq ) requires an MPI REDUCE operation among the processors and the sums
are done in a different order, therefore the results can be slightly different in the
various cases. We checked anyway that the maximum relative difference in all
the cases was lesser than 10−8 .
As usual, we define: Tn as the CPU time necessary to run the code on n
cores, and the speed-up as:
T1
S(n) =
Tn
316 L. Primavera and E. Florio

that is, the ratio of the serial over the parallel CPU times. The “ideal” value,
S(n) = n, represents the case of perfect parallelization (although it is not
unusual, in some special cases, to find even superlinear speed-ups because of
insufficient memory storage or excessive cache occupation in the serial case).

Fig. 4. Speed-up curve for the execution times vs. the number of processors (black-
diamonds marks) along with the theoretical scaling S(n) = n (red line). (Color figure
online)

In Fig. 4, we show the speed-up S(n) for the tests we did on a relatively
small cluster, made of 32 nodes, 2 CPUs/node, 10 cores/CPU, for a total of
640 computational cores. The CPUs are Intel Xeon processors ES-2680 with a
clock frequency of 2.8 GHz and the node interconnection is realized through an
Infiniband switch with a bandwidth of 40 Gb/s. The results show a fairly good
speed-up of the code (black-diamonds marks) with respect to the theoretical
curve S(n) = n (red line). The difference for increasing values of n are likely due
to the latency of the communications among different nodes.

5 Conclusions
We used the MPI library to parallelize a code to perform the fixed-mass multi-
fractal analysis. Such a code was widely used in the past, in its serial version, for
applications to the study of river networks to get useful parameters to be used
to study the hydrological response of a basin through the Multifractal Instanta-
neous Unit Hydrograph (MIUH).
 Parallel Algorithms for Multifractal Analysis of River Networks 317

For large numbers of net-points extracted from the blue-lines of the river
network, the numerical complexity of the calculation requires very long com-
putational times that are drastically reduced in the parallel version. This will
allow the code to run on multi-core workstations and/or multi-node clusters by
exploiting the whole potential of the CPUs.
Future possible improvements could consist in: (1) realizing an hybrid paral-
lelization with the Open-Message-Passing (OPEN/MP) paradigm, which would
avoid the inter-node communications, and (2) the porting of the code on GPUs,
that would allow a very efficient massively parallel execution of the code without
the need to buy expensive clusters or extremely powerful workstations.

References
1. Badii, R., Politi, A.: Hausdorff dimension and uniformity factor of strange attrac-
tors. Phys. Rev. Lett. 52, 1661–1664 (1984)
2. De Bartolo, S.G., Ambrosio, L., Primavera, L., Veltri, M.: Descrittori frattali
e caratteri morfometrici nella risposta idrologica. In: Caroni, E., Fiorotto, V.,
Mancinelli, A., Salandin, P. (eds.) La Difesa Idraulica del Territorio, pp. 47–60.
Tergeste, Trieste (2003)
3. De Bartolo, S.G., Gabriele, S., Gaudio, R.: Multifractal behaviour of river networks.
Hydrol. Earth Syst. Sci. 4, 105–112 (2000)
4. De Bartolo, S.G., Veltri, M., Primavera, L.: Estimated generalized dimensions of
river networks. J. Hydrol. 322, 181–191 (2006)
5. Falconer, K.: Fractal Geometry: Mathematical Foundations and Applications, 2nd
edn. Wiley, Chichester (2003)
6. Florio, E., Maierù, L.: The scientific knowledge of book XVI De subtilitate by
G. Cardano used in the Trattato sulla divinatione naturale cosmologica by P.A.
Foscarini. Open J. Human. 1, 385–422 (2019)
7. Gaudio, R., De Bartolo, S.G., Primavera, L., Veltri, M., Gabriele, S.: Procedures
in multifractal analysis of river networks, vol. 286, pp. 228–237. IAHS Publication
(2004)
8. Grassberger, P.: Generalized dimensions of strange attractors. Phys. Lett. A 97,
227–230 (1983)
9. Grassberger, P., Procaccia, I.: Characterization of strange attractors. Phys. Rev.
Lett. 50, 346–349 (1983)
10. Mandelbrot, B.B.: Possible refinement of the lognormal hypothesis concerning the
distribution of energy dissipation in intermittent turbulence. In: Rosenblatt, M.,
Van Atta, C. (eds.) Statistical Models and Turbulence. LNP, vol. 12, pp. 333–351.
Springer, Berlin (1972). https://doi.org/10.1007/3-540-05716-1 20
11. Pawelzik, K., Schuster, H.G.: Generalized dimensions and entropies from a mea-
sured time series. Phys. Rev. A 35, 481–484 (1987)
 A Methodology Approach to Compare
Performance of Parallel Programming
Models for Shared-Memory Architectures

Gladys Utrera(B) , Marisa Gil , and Xavier Martorell

Computer Architecture Department, Universitat Politècnica de Catalunya,

c/ Jordi Girona, 1-3, 08034 Barcelona, Catalonia, Spain
{gutrera,marisa,xavim}@ac.upc.edu

Abstract. The majority of current HPC applications are composed of

complex and irregular data structures that involve techniques such as
linear algebra, graph algorithms, and resource management, for which
new platforms with varying computation-unit capacity and features are
required. Platforms using several cores with different performance char-
acteristics make a challenge the selection of the best programming model,
based on the corresponding executing algorithm. To make this study,
there are approaches in the literature, that go from comparing in isola-
tion the corresponding programming models’ primitives to the evalua-
tion of a complete set of benchmarks. Our study shows that none of them
may provide enough information for a HPC application to make a pro-
gramming model selection. In addition, modern platforms are modifying
the memory hierarchy, evolving to larger shared and private caches or
NUMA regions making the memory wall an issue to consider depending
on the memory access patterns of applications. In this work, we propose
a methodology based on Parallel Programming Patterns to consider intra
and inter socket communication. In this sense, we analyze MPI, OpenMP
and the hybrid solution MPI/OpenMP in shared-memory environments.
We demonstrate that the proposed comparison methodology may give
more accurate predictions in performance for given HPC applications and
consequently a useful tool to select the appropriate parallel programming
model.

Keywords: MPI · OpenMP · NUMA · HPC · Parallel programming

patterns

1 Introduction
Current HPC platforms are composed of varying computation units capacities
and features connected by diverse, increasingly powerful and complex networks
to provide better performance not only for large size messages but also for mas-
sive receive/send from multiple nodes. These are characteristics foreseeable for
the Exascale era.
c Springer Nature Switzerland AG 2020
Y. D. Sergeyev and D. E. Kvasov (Eds.): NUMTA 2019, LNCS 11973, pp. 318–325, 2020.
https://doi.org/10.1007/978-3-030-39081-5_28
 A Methodology Approach for Performance Comparison of PPL 319

From the point of view of the software, applications also tend to be composed
of different data structures and the corresponding algorithms to read and modify
this data. In addition, based on the data size and the order in which the data is
processed, the performance in terms of scalability or reliability, can be affected
depending on the programming model in use.
This scheme has led to several approaches considering the use of pure Message
Passing library Interface (MPI) [7] versus OpenMP [9] primitives inside a node or
exploring several levels of hybrid message-passing and shared-memory proposals
to take advantage of the different cores’ characteristics. Furthermore, modern
platforms are also modifying the memory hierarchy differences, evolving to larger
shared and private caches or NUMA (Non-Uniform Access Memory) regions.
UMA (Uniform Memory Access) architectures, commonly referred to as SMP
(Symmetric Multiprocessing), have equal memory access latencies from any pro-
cessor. On the contrary, NUMA architectures are organized as interconnected
SMPs and the memory access latencies may differ between different SMPs. In
this situation, the memory wall is an issue to consider depending on the memory
access patterns the executing application exhibits: data message size, varied size
of synchronization or mutex areas; and an inter-socket evaluation is necessary.
This increasing complexity at both low- and high-level makes a challenge
the selection of the best programming model, to achieve the best performance
on a specific platform. In this work, we take a pattern-based approach to ana-
lyze application performance, based on the scalability achieved, including data
locality.
Contributions of this work:
– Review a methodology currently used to enhance parallel programming lan-
guages with the objective of comparing them.
– Performance comparison between MPI and OpenMP under the stencil and
reduce parallel patterns.
The rest of the paper is organized in the following way: Sect. 2 introduces
background, related work and our motivation; Sect. 3 presents the experimental
results. Finally, in Sect. 4 are the conclusions and future work.

2 Background and Motivation

The choice of the parallel programming model can be determinant in perfor-
mance for a given application.
Standard programming models may ensure portability. However, when com-
bined with the platform architecture there is not a general best approach.
In this work we focus on two standards: Message Passing Model (MPI) and
OpenMP. OpenMP [9] is the de facto standard model for shared memory systems
and MPI [7] is the de facto standard for distributed memory systems.
In the following subsections we introduce briefly MPI and OpenMP program-
ming models. After that we make an overview of existing performance compari-
son studies, which conduct us to the motivation of our proposal.
320 G. Utrera et al.

2.1 OpenMP
OpenMP is a shared-memory multiprocessing Application Program Inference
(API) for easy development of shared memory parallel programs. It provides a
set of compiler directives to create and synchronize threads, and easily paral-
lelize commonly used parallel-patterns. To that end, it uses a block-structured
approach to switch between sequential and parallel regions, which follows the
fork/join model. When entering a parallel region, a single thread splits into some
number of threads, then when finishing that region, only a sequential thread con-
tinuous execution.

2.2 MPI
MPI is a message passing library specification for parallel programming on a dis-
tributed environment. In a message passing model, the application is composed
of a set of tasks which exchange the data, local or distributed among a certain
number of machines, by message passing. There exist several implementations
like Intel MPI, and also open source like OpenMPI, MPICH.
Each task in the MPI model has its own address space and the access to
others’ tasks address space has to be done explicitly with message passing. Data
partitioning and distribution to the tasks of the application is required to be
programmed explicitly.
MPI provides point-to-point operations, which enable communication
between two tasks, and collective operations like broadcast or reduction, which
implement communication among several tasks. In addition, communication can
be synchronous where tasks are blocked until the message passing operation is
completed, or asynchronous where tasks can defer the waiting for the completion
of the operation until some predefined point in the program. The size of the data
exchanged can be from bytes to gigabytes.

2.3 Related Work and Motivation

There is an interesting study by Krawezik et al. [5], which involved the compar-
ison of some communication primitives from OpenMP and MPI, as well work-
sharing constructs from OpenMP and evaluations of the NAS Benchmarks in
shared-memory platforms. They recommend OpenMP for computing loops and
MPI for the communication part. On the other hand, Kang et al. [4] suggest
that for problems with small data sizes OpenMP can be a good choice, while if
the data is moderate and the problem computation-intensive then MPI can be
considered a framework. Another work by Piotrowski [10] which evaluates square
Jacobi relaxation under pure MPI the hybrid version with OpenMP, showed that
in a cluster of shared-memory nodes, none of the hybrid versions outperformed
pure MPI due to longer MPI point-to-point communication times. In the work
by Qi et al. [11] they also use the NAS benchmarks for the evaluation and show
that with simple OpenMP directives the performance obtained is as good a with
shared-memory MPI on IA64 SMP. They also claim that even OpenMP has easy
 A Methodology Approach for Performance Comparison of PPL 321

programming, improper programming is likely to happen leading to inferior per-

formance. In this sense, in the work by [2] they conclude that parallelization with
OpenMP can be obtained with little effort but the programmer has to be aware
of data management issues in order to obtain a reasonable performance. The per-
formance comparison by Yan et al. in their work [12] on the 3D Discrete Element
Method of ellipsoid-like complex-shaped particles, examine memory aspects, task
allocation as well as variations in the combined approach MPI/OpenMP. They
conclude that Pure MPI achieves both lower computational granularity (thus
higher spatial locality) and lower communication granularity (thus faster MPI
transmission) than hybrid MPI-OpenMP in 3D DEM, where computation dom-
inates communication, and it works more efficiently than hybrid MPI-OpenMP
in 3D DEM simulations of ellipsoidal and poly-ellipsoidal particles.
The reviewed works above perform interesting analysis on selected applica-
tions and/or algorithms but none of them provide enough information for a given
HPC application other than the ones specifically analyzed, to make a proper pro-
gramming model selection. Even more, there are contradictory recommendations
derived from the fact that they were evaluated with different platform charac-
teristics which is a relevant factor.

2.4 Selected Patterns

Pattern-based parallel programming consists on a set of customizable paral-

lel patterns used to instantiate blocks of applications. This approach is being
empowered in the last years with the objective to provide parallel programming
languages with additional features that bring more flexibility and reduce the
effort of parallelization. In the work by De Sensi et al. [3], the authors demon-
strate the feasibility of the approach.
In this work, we exploit the idea of pattern-based parallel programming lan-
guages to compare them in terms of performance.
We selected the loop-of-stencil-reduce pattern for being representative of
many HPC applications. This parallel pattern is general enough to subsume
other patterns like map, map-reduce, stencil, stencil-reduction and their usage
in a loop [1].
Below we can see a pseudocode of our MPI and OpenMP implementation
versions:
{ MPI version } { OpenMP version }
data-distribution
loop loop
start-border-recv #pragma omp parallel for
stencil-reduce (in, out) stencil-reduce (in, out)
start-border-send swap-matrix-in-out
complete-border-exchange end loop
swap-matrix-in-out
end loop
322 G. Utrera et al.

In order to make a fair comparison, we do not take into account any converge
condition; the main loop has a fixed number of iterations. There are no depen-
dencies within the blocks inside the stencil algorithm. At every loop there is an
input matrix and an output matrix. Before next iteration, we perform a swap
between both matrix. The parallelization is block-based. This means that a task
perform the stencil algorithm on a block matrix like the one shown in Fig. 1.
The memory access pattern for our implementation is a 5-point Stencil shown
in Fig. 1. In dark grey is shown the data shared in the border of each block. Each
task share data with top, bottom, left and right tasks.

Fig. 1. Memory access pattern for a 5-point stencil

Fig. 2. OpenMP and MPI loop-stencil-reduce parallel pattern performance comparison

when datasize does not fit in LLC. Bigger is better.

3 Experimental Results
In this section we show the performance results from the evaluation of the imple-
mentation in MPI, OpenMP and MPI/OpenMP of the selected parallel pattern.
The executions were performed on NordIII [8], a supercomputer based on
Intel SandyBridge processors, with Linux Operating System. It has 3.056 homo-
geneous compute nodes (2x Intel SandyBridge-EP E5-2670/1600 20M 8-core at
2.6 GHz) with 2 GB per core. We use the Intel MPI library 4.1.3.049 and C
compiler Intel/2017.4.
 A Methodology Approach for Performance Comparison of PPL 323

The evaluations are performed varying the number of tasks (1–16), task allo-
cation (within a NUMA-node or inter NUMA-nodes) and data size. For the data size
we consider two cases: (1) fits in the last-level cache (LLC); (2) does not fit in
the LLC but fits in main memory.
The work distribution among tasks in our test is well-balanced, so load bal-
ancing issues are not tackle in this analysis. There are no dependencies between
tasks during a given step, so tasks are embarrasingly parallel. The communica-
tion is performed between adjacent tasks (exchange of borders), but source data
is from the previous step. Notice that, the input matrix is only read and the
output matrix is only written (see the pseudocode at Sect. 2.4).
The initialization of data structures is done in parallel taking care of the
first touch Operating system data allocation policy to minimize remote accesses
during calculation. Tasks are binded to cores in order to ensure allocation poli-
cies: (1) compact, that is in the same NUMA node; (2) scatter, that is equally
distributed across NUMA nodes. The results are showed in Figs. 2 and 3.

Fig. 3. OpenMP and MPI loop-stencil-reduce parallel pattern performance comparison

when datasize does not fit in LLC. Bigger is better.

We can observe in Fig. 2 that the data size does not fit in the shared NUMA-
node LLC, when the NUMA node is full (8 tasks) or both NUMA nodes are full
(16 tasks) MPI performance degrades dramatically with respect to OpenMP.
However, if data fits in LLC, as shown in Fig. 3, then MPI has better performance
when having 8 tasks allocated in different NUMA nodes or when the NUMA node
is full (16 tasks).
The stencil parallel-pattern is characterized for having memory accesses
which are updated by other tasks in previous steps. This means that such data
has to be exchanged before doing the current calculation. There are memory
accesses not only within the block of data processed by a given task, but also for
data managed by neighbouring tasks (adjacent blocks). For shared-memory pro-
gramming models, collaborating tasks allocated to different NUMA-nodes have a
well-documented effect on memory access performance (e.g. OpenMP) [6]). This
is not the case for distributed memory programming models (e.g. MPI). In paral-
lel programming languages where memory is not shared among tasks (i.e. MPI),
this is exchanged explicitely between steps (i.e. MPI Isend and MPI Irecv for our
324 G. Utrera et al.

current implementation shown in Sect. 2.4. However, once the data is brought
becomes local (task allocation is transparent).
Taking all this into account we can appreciate the NUMA effect when data
fits in LLC. As memory is not an issue for this data size, MPI obtains better
performance than OpenMP. The data managed by each MPI task is local. On the
other hand, when data does not fit in LLC, then as MPI duplicates data faces
memory problems with the consequent increment in cache misses, degrading
performance (Fig. 4). Notice that if allocating tasks in different NUMA nodes,
this effect can be alleviated. Despite LLC misses in MPI for small data sizes are
larger than LLC misses in OpenMP, the remote accesses generated by adjacent
tasks penalize bringing better performance for MPI.
The hybrid approach MPI/OpenMPI performed worse than MPI and
OpenMP in isolation when allocated one MPI task per NUMA-node, which
means in our experimental platform, 2 task per node. The added memory over-
head plus the thread creation and other overheads (e.g. remote memory access)
do not compensate in performance. We believe that for larger NUMA-nodes this
results may be different. We are currently undergoing this study.
In conclusion, for small data sizes and small number of tasks, both parallel
programming languages can achieve the same performance no matter where tasks
are allocated (lesser tasks, lesser interaction between them). When incrementing
the number of tasks, there is more interaction between them penalizing remote
accesses for OpenMP, but duplicating data at the same time for MPI.

Fig. 4. Last Level Cache misses when datasize does not fit

4 Conclusions and Future Work

Based on the idea of pattern-based parallel programming we compared two stan-
dard like OpenMP, MPI and the hybrid approach MPI/OpenMP on a mul-
ticore platform with different memory accesses characteristics. We selected a
scheme pattern representative of many HPC applications like a loop of stencil
and reduce. We showed that both parallel programming languages can show
different performance characteristics in applications depending on the data size
being processed and the data locality which requires extra and conscious pro-
gramming effort. Considering the hybrid approach to solve gaps in the context
of a shared-platform showed to not work as good as separately.
 A Methodology Approach for Performance Comparison of PPL 325

In this work we only focused on data-parallel algorithms. We are planning

to extend our study to other parallel patterns like pipeline and unstructured. In
addition, the platform as already shown in terms of memory accesses has a sig-
nificant role so we plan to enrich the work by studying heterogeneous platforms.

Acknowledgements. This research was supported by the following grants Span-

ish Ministry of Science and Innovation (contract TIN2015-65316), the Generalitat de
Catalunya (2014-SGR-1051) and the European Commission through the HiPEAC-3
Network of Excellence (FP7/ICT-217068).

References
1. Aldinucci, M., et al.: A parallel pattern for iterative stencil + reduce. J. Super-
comput. 74(11), 5690–5705 (2018). https://doi.org/10.1007/s11227-016-1871-z
2. Bane, M.K., Keller, R., Pettipher, M., Computing, M., Smith, I.M.D.: A com-
parison of MPI and OpenMP implementations of a finite element analysis code
(2000)
3. Danelutto, M., De Matteis, T., De Sensi, D., Mencagli, G., Torquati, M.: P 3
ARSEC: towards parallel patterns benchmarking. In: Proceedings of the Sympo-
sium on Applied Computing, SAC 2017, pp. 1582–1589. ACM, New York (2017).
https://doi.org/10.1145/3019612.3019745
4. Kang, S.J., Lee, S.Y., Lee, K.M.: Performance comparison of OpenMP, MPI, and
MapReduce in practical problems. Adv. Multimed. 2015, 361–763 (2015). https://
doi.org/10.1155/2015/575687
5. Krawezik, G.: Performance comparison of MPI and three OpenMP programming
styles on shared memory multiprocessors. In: Proceedings of the Fifteenth Annual
ACM Symposium on Parallel Algorithms and Architectures, SPAA 2003, pp. 118–
127. ACM, New York (2003). https://doi.org/10.1145/777412.777433
6. Metzger, P., Cole, M., Fensch, C.: NUMA optimizations for algorithmic skeletons.
In: Aldinucci, M., Padovani, L., Torquati, M. (eds.) Euro-Par 2018. LNCS, vol.
11014, pp. 590–602. Springer, Cham (2018). https://doi.org/10.1007/978-3-319-
96983-1 42
7. MPI: Message Passing Interface (MPI) Forum. http://www.mpi-forum.org/
8. NORD3: Nord3 machine. https://www.bsc.es/support/Nord3-ug.pdf
9. OpenMP: The OpenMP API specification for parallel programming. http://www.
openmp.org/
10. Piotrowski, M.: Mixed mode programming on HPCx. In: Piotrowski, M. (ed.) Par-
allel Scientific Computing and Optimization. SOIA, vol. 27, pp. 133–143. Springer,
New York (2009). https://doi.org/10.1007/978-0-387-09707-7 12
11. Qi, L., Shen, M., Chen, Y., Li, J.: Performance comparison between OpenMP
and MPI on IA64 architecture. In: Bubak, M., van Albada, G.D., Sloot, P.M.A.,
Dongarra, J. (eds.) ICCS 2004. LNCS, vol. 3038, pp. 388–397. Springer, Heidelberg
(2004). https://doi.org/10.1007/978-3-540-24688-6 52
12. Yan, B., Regueiro, R.A.: Comparison between pure MPI and hybrid MPI-OpenMP
parallelism for discrete element method (DEM) of ellipsoidal and poly-ellipsoidal
particles. Comput. Part. Mech. 6(2), 271–295 (2019). https://doi.org/10.1007/
s40571-018-0213-8
Numbers, Algorithms, and Applications
 New Approaches to Basic Calculus:
An Experimentation via Numerical
Computation

Luigi Antoniotti, Fabio Caldarola(B) , Gianfranco d’Atri,

and Marco Pellegrini

Deparment of Mathematics and Computer Science,

Università della Calabria, Cubo 31/A, 87036 Arcavacata di Rende, CS, Italy
[email protected], {caldarola,datri}@mat.unical.it,
[email protected]

Abstract. The introduction of the first elements of calculus both in the

first university year and in the last class of high schools, presents many
problems both in Italy and abroad. Emblematic are the (numerous) cases
in which students decide to change their course of study or give it up com-
pletely cause the difficulties with the first exam of mathematics, which
usually deals with basic calculus. This work concerns an educational
experimentation involving (with differentiated methods) about 170 stu-
dents, part at the IPS “F. Besta” in Treviso (IT) with main focus on two
5th classes where the students’ age is about 19 years old, and part at
the Liceo Classico Scientifico “XXV Aprile” in Pontedera, prov. of Pisa
(IT). The experimental project aims to explore the teaching potential
offered by non-classical approaches to calculus jointly with the so-called
“unimaginable numbers”. In particular, we employed the computational
method recently proposed by Y.D. Sergeyev and widely used both in
mathematics, in applied sciences and, recently, also for educational pur-
poses. In the paper will be illustrated tools, investigation methodologies,
collected data (before and after the teaching unit), and the results of
various class tests.

Keywords: Mathematical education · Learning/teaching models and

strategies · Interactive learning environments · Infinite · Infinity
Computer · Grossone · Unimaginable numbers

1 Introduction
The difficulties that high school students encounter in their approach to uni-
versity mathematics courses are fairly well known, both because of the direct
experience of those who teach them, and because of the extensive literature that
deals with various aspects of the problem. For example, see [19] for the intrinsic
difficulties related to the transition to the third level of education, or [24,26–29]
for detailed analyses of students’ problems and approaches to calculus courses.
c Springer Nature Switzerland AG 2020
Y. D. Sergeyev and D. E. Kvasov (Eds.): NUMTA 2019, LNCS 11973, pp. 329–342, 2020.
https://doi.org/10.1007/978-3-030-39081-5_29
330 L. Antoniotti et al.

A research topic, very lively as well, which often overlaps with the previous ones,
deals with the role assumed by software, simulations and computer technologies
in general, as a support to academic teaching, and to the learning and personal
elaboration of concepts and methods by the students (see, e.g., [23,43]). In this
view, very interesting results were obtained in the paper [4], which in part served
as a model to build the working environments described here.1
A whole line of research, within the theme concerning the students’ approach
to calculus, focuses on methods and problems related to teaching and learning
the concept of limit and the processes with infinitely many steps. For example,
[14,15,18,29,42] investigate student’s difficulties, common misconceptions and
obstacles for the learning, [46–48] elaborate on conceptual images and students’
models of limit, and the very interesting paper [45] deals with the mathematical
intuition in limiting processes.2
We are interested, instead, in the interactions of the students, in particular
of the last year of high schools, with the first rudiments of non-classical analysis
systems and similar mathematics. The teaching of non-classical mathematics, in
many countries of the world, has become a consolidated fact from many years; for
example, the broader phenomenon is the teaching of Robinson’s non-standard
analysis both in university courses3 and in high schools. In Italy, in general,
traditional methods are taught almost everywhere, but the focus on non-classical
mathematics is strongly growing and many people are convinced of the usefulness
of teaching alternative theories and methods also in high schools.4
The research exposed in this paper wants to investigates the response of stu-
dents with respect the easy computational system proposed by Y. Sergeyev and,
in smaller measure, the notational system introduced by D.E. Knuth to write
the so-called unimaginable numbers (see Sect. 2 for a brief theoretical overview).
The experimentations described here involve 8 high school classes in two different
1
The paper [4] concerns a rather complex two-year experimental project conducted
at the University of Calabria within the Master’s Degree Program in Electronic
Engineering. There, once two groups were formed each year, an experimental one
equipped with computer-based tools and a control group associated with more tra-
ditional teaching methods, the aim was to analyze a series of student performances.
We inform the reader that in the experimentations described in the present paper,
we will find some slight traces of part of the methods used in [4]; the most visible is
the employ of the computational system called Infinity Computer which will be used
by a (very small) part of the sample, creating a hint of parallelism with the role of
the software used in [4].
2
The mathematical intuition will be important also for us, when our students will
work with the infinite and, in particular, when they will approach the new concept
of grossone (see also [44]).
3
See, for instance, the manual [22] adopted in many academic courses and now in
its third edition. See also [17] for an interesting comparison between Leibniz’ and
Robinson’s systems and “Sarah”s conceptions”.
4
See, for instance, [16, page 2] and the proceedings [7] of the national Italian con-
ference “Analisi nonstandard per le scuole superiori. VII Giornata di studio”. This
conference takes place every year and has now reached its ninth edition, Verona,
October 5, 2019.
 New Approaches to Basic Calculus 331

Italian institutes in Treviso and Pontedera (Pi), see Sect. 3 for details. The aim
of the research are multiple, and can be summarized in three main groups/points
to investigate:
(a) The students’ mathematical intuition and first approaches with respect the
grossone system without preliminary class lectures;
(b) The students’ responses, empathy and performances during and after a brief
cycle of class lectures;
(c) The students’ individual home working with also the support of the Infinity
Computer.
In Sect. 4 we will give the results and some brief conclusions.
One last important piece of information for the reader: the paper [21] con-
cerns a twin but independent experimentation, concerning the same themes and
carried out approximately simultaneously.

2 A Brief Overview on the Grossone-Based System and

on Unimaginable Numbers
Sergeyev began to develop the grossone-based methodology in the early 2000’s
and, in brief, we can say that his numerical system is based on two fundamental
units: the familiar 1 that generates natural numbers and a new infinite unit ①,
called grossone, which allows to write infinite numbers and to execute computa-
tions with them in a intuitive and easy way, similar at all to what people daily
do with finite numbers and ordinary integers. Moreover, infinitesimal numbers
then appear by taking the inverses of infinite, grossone-based numbers. We also
recall that, in Sergeyev’s system, ① is the number of elements of IN and n ≤ ①
for all n ∈ IN.
The extended set of natural numbers, denoted by IN,  can be written as

 = 1, 2, . . . , ① − 1, ①, ① + 1, . . . , 2①, . . . , 3①, . . . ,
IN
  
IN
(1)
①2 , ①2 + 1, . . . , ①2 + ①, . . ., 2①2 , . . . , ①3 , . . . , 2① , . . . ,
2① + ①, . . . . . . , ①① , . . . . . . ,

while the extended set of integers as Z  ∪ {0} ∪ {−N : N ∈ IN}.

Z = IN  Hence, the

quotients of elements of ZZ obviously yield the extended set of rational numbers
denoted with Q.
Introductory books for the grossone-based numerical system are [30,32], writ-
ten in popular or didactic way, while the interested reader can see [31,36,38,40]
for more detailed surveys and [1,8–10,13,20,33,34,37–41] (and the references
therein) for some applications. Sergeyev’s Infinity Computer [35] is, moreover,
an on-line computing software developed for the grossone-based system which
gives an extra level to our experimentation and connects it to previous works
like [2,4] and to other similar ones in progress using the frameworks of [3,5].
332 L. Antoniotti et al.

“Unimaginable numbers” are instead finite but very large natural numbers
with a completely different and ancient origin: the first unimaginable number
comes back in fact to Archimedes of Syracuse (see [6,11]) Usually an integer
n ∈ IN is said unimaginable if it is greater than 1 googol which is equal to
10100 . Writing unimaginable numbers in common scientific notation is almost
always impossible and we need notations developed ad hoc like Knuth’s up-arrow
notation that at its lower kevels gives the usual addition, multiplication and
exponentiation, then tetration, pentation, hexation, and so on (see [6,11,12,25]
and the references therein).

3 Description of the Samples and the Methods

of the Research
The participants to this research are divided in more groups:
(P ) A group P of students at the Liceo Classico Scientifico “XXV Aprile” in
Pontedera, prov. of Pisa, Italy. This group consists of 45 students coming
from a second and a third class named, by convenience, P 2 and P 3, respec-
tively. More details on this sample will be given in Table 1.
(T ) A group T of students at the Istituto Superiore Commerciale Professionale
“F. Besta” in Treviso, Italy. This group comes from 2 third and 2 fourth
classes conventionally indicated by T 3, T 3 , T 4 and T 4 , respectively.
(T ) A second group of students at the same institute in Treviso as group T , is
now indicated by the uppercase calligraphic letter T : such a group comes
from 2 fifth classes conventionally indicated by T 5 and T 5 . This group
is the more interesting both for being the last class before the university
career (for those who will choose to continue their studies) and because
the organization, started before for this group, allowed to carry out a more
extensive and articulated research.
A more accurate description of the groups listed above can be found in Table 1
where some data are shown for each class of the samples P, T and T . The second
column of Table 1 enumerates the students in each class and the third one gives
the subdivision male-female. The 4th column lists the “mean age” computed at
April 1, 2019 (in some case there is a certain margin of approximation caused
by incomplete data). The last three columns regard the final votes obtained in
mathematics at the end of the first semester (January 31, 2019): the 5th column
computes the “mean vote” of all the students, the 6th one gives separated means
male-female, and the last provides the maximum final vote in the class, separately
again male-female.

3.1 The Group P : Experimental Activities and Methods

For the group P the experimental activity was limited to a class test at the
end of May 2019, in agreement with the aim (a) described in the Introduction.
The class test was administered without prior knowledge of the grossone-based
 New Approaches to Basic Calculus 333

Table 1. The composition of the 8 classes constituting the samples, and some other
data.

Class Students M - F Mean age Mean vote Mean vote (M - F) Max vote (M - F)
P2 25 11 - 14 15.2 6.7/10 6.5/10 - 6.8/10 9/10 - 9/10
P3 20 10 - 10 16.2 7.6/10 7.9/10 - 7.3/10 9/10 - 8/10
T3 26 8 - 18 16.6 5.6/10 5.7/10 - 5.6/10 8/10 - 9/10
T 3 17 6 - 11 16.6 5.8/10 5.9/10 - 5.7/10 9/10 - 9/10
T4 15 4 - 11 17.5 5.8/10 5.9/10 - 5.7/10 8/10 - 9/10
T 4 27 8 - 19 17.7 5.4/10 5.6/10 - 5.0/10 8/10 - 9/10
T5 23 7 - 16 18.7 5.5/10 6.1/10 - 5.2/10 9/10 - 9/10
T 5 15 4 - 11 19.1 5.6/10 5.8/10 - 5.5/10 9/10 - 8/10

system and without students have seen before the symbol ①, (in brief, a “zero-
knowledge test”). The only information given to them was ① ∈ IN and ① ≥ n
for all n ∈ IN. The test had 15 multiple-choice questions and other 6 open ones.
The contents were elementary operations with ①, and about the order relation
 i.e., the extended set of rational numbers in Sergeyev’s framework (see
in Q,
[31,32,36,38,40]). Below we report in English some examples of the proposed
questions.

Question 1.5 Consider the writing ①+① and make a mark on the option that seems
most correct to you among the following:

(a) ① + ① has no sense;

(b) ① + ① = ①;
(c) ① + ① is impossible to execute;
(d) ① + ① = 2①;
(e) ① + ① = 0.

Question 2. Consider the writing ① − ① and make a mark on the option that seems
most correct to you among the following:

(a) ① − ① = −①;
(b) ① − ① is indeterminate;
(c) ① − ① = 0;
(d) ① − ① = ①;
(e) ① − ① has no sense.

Question 3. Consider the expression −3① + ① and make a mark on the option that
seems most correct to you among the following:
5
The numbers used here to enumerate questions are different from those in the stu-
dents’ test (cf. [21, Sect. 3]). We moreover precise that in some classes in Trento we
prepared two or four test versions changing for the order of questions and answers,
to prevent, together with other appropriate measures, any kind of influence among
students.
334 L. Antoniotti et al.

(a) −3① + ① = ①;
(b) −3① + ① is a writing without sense;
(c) −3① + ① = −3①;
(d) −3① + ① = −2①;
(e) −3① + ① = ①;
(f) −3① + ① is an indeterminate expression;
(g) −3① + ① = 0.

Question 4. Consider ①, ① + ① and ① · ①: mark the option (or the options) that
seem correct to you among those below.
(a) ① < ① + ①, but ① · ① = ①;
(b) ① < ① + ① < ① · ①;
(c) ① + ① and ① · ① are both equal to ①;
(d) ① ≤ ① + ① ≤ ① · ①;
(e) It is not possible to establish any order relation between ①, ① + ① and ① · ①;
(f) ① < ① + ① and ① + ① ≤ ① · ①;
(g) ① ≤ ① · ① < ① + ①;
(h) ① ≤ ① · ① ≤ ① + ①;
(i) The writings ① · ① and ① + ① have no sense;
(j) None of the previous is correct;
(k) Other:

Question 5. Consider the expression − 53 ① + 12 ① and choose the option that seems
most correct to you among the following:
(a) Both − 53 ① and 12 ① are writings without sense;
(b) − 53 ① + 12 ① is an indeterminate expression;
(c) − 53 ① + 12 ① = −①;
(d) − 53 ① + 12 ① = − 76 ①;
(e) − 53 ① + 12 ① = 0;
(f) − 53 ① + 12 ① = ①.


Question 6. Consider the expression − 23 ① + 2 ·(4①−3)+4 and choose the option
that seems most correct to you among the following:
(a) The first factor is
equal to−①, the second to +① and the addition of 4 is
irrelevant, hence − 23 ① + 2 · (4① − 3) + 4 = −①;


(b) − 23 ① + 2 · (4① − 3) + 4 = −①2 ;


(c) − 23 ① + 2 · (4① − 3) + 4 is an indeterminate expression;


(d) − 23 ① + 2 · (4① − 3) + 4 = − 83 ①2 + 10① − 2;
(e) It

is2 not possible
 to sum 23 ① with 2 and 4① with −3, hence the expression
− 3 ① + 2 · (4① − 3) + 4 = −① has no sense.

Question 7. Consider 12
8 ① and 3 ①: mark the option (or the options) that seem
5

correct to you among those below.

 New Approaches to Basic Calculus 335

8 ① and 3 ① have no sense;

(a) The writings 12 5

8 ① = 3 ①;
(b) 12 5

8 ① and 3 ① are both equal to ①;

(c) 12 5

8 ① > 53 ①;
12
(d)
8 ① < 53 ①;
12
(e)
8 ① ≥ 53 ①;
12
(f)
8 ① ≤ 53 ①;
12
(g)
8 ① and 53 ①.
12
(h) It is not possible to establish an order relation between

It is important to notice that the questions in the test (and also for the
ones of the samples T and T ) were grouped in small groups on separate sheets,
and the students were asked to read and answer the questions in the order of
presentation, without the possibility of changing an answer already given.
After the day of the test, and therefore outside our experimentation, there
were some discussions in classroom, often to answer the questions and curiosities
of the students themselves, about the meaning and use of the symbol ①. We
specify, however, that the experimentation in question did not in any way affect
the regular progress of the established school program.
The results of the test will be given and analyzed in Sect. 4.

3.2 The Group T : Experimental Activities and Methods

The experimental activities and the research methodologies used for the 4 classes
of group T were almost identical to those described in Subsect. 3.1. The only
exceptions concern a greater ease and a simplification of some test questions,
especially of some more complex ones not reported in the previous subsection,
dictated by the different type of school and, unlike the group P , to classes T 3 and
T 4 were administered, in the days after the test, an informative questionnaire
that aimed to capture the (eventual) interest and any curiosity aroused in the
students. We also precise that the test questions and their number were not the
same for all the classes that made up the group T , for didactic and organizational
reasons.

3.3 The Group T : Experimental Activities and Methods

The most complex and structured experimentation concerned the group T . In

addition to the zero-knowledge test as for groups P and T , but carried out about
one or two months in advance, a series of short lessons and class discussions was
proposed to both classes T 5 and T 5 . The organization relative to the class T 5
allowed also the possibility of a greater number of lessons, which in any case
covered a very limited amount of time (about 4 or 5 h split into packages of
half an hour each time, over two or three months). After the cycle of lectures,
a test very similar, especially for the class T 5 , to the one administered at the
beginning was proposed for both classes: for convenience and future references
we call it the final test.
336 L. Antoniotti et al.

Our experimental activity with group T included also some rudiments of cal-
culus. In particular, it has been discussed the concept of limit for x → ±∞ in
classical analysis compared with the evaluation of a function at a grossone-based
infinity in Sergeyev’s framework (mainly we took x = ±① than other infi-
nite numbers). Moreover, we showed the relation between some asymptotic
behaviours of a function and its derivative both in the traditional and in the
new context, and we also talked about the meaning and the way to perform
computations with infinitesimal quantities written in the new system. Examples
of closed and open questions proposed to the students of group T in the final
test are reported below translated in English.

Question 8. Consider the writing 3 − 1 and mark the right option/options among
2① ①
the following:

(a) 3
− 1
= 12 ①;
2① ①
(b) 3
− 1
is indeterminate;
2① ①
(c) The writings 3 and 1 have no sense because it is not possible to divide by
2① ①
infinity;
(d) 3
− 1 is zero because both 3 and 1 are equal to zero;
2① ① 2① ①
(e) 3
− 1 = 1 ;
2① ① 2①
(f) 3
− 1
= 2;
2① ① ①
(g) 3
− 1 = 3−1 = 1 ;
2① ① 2① ①
−1
(h) 3
− 1
= 0.5① .
2① ①
Question 9. Consider the writings 1
, − 1
and −3 1
+ 12 . Indicate the true
2① ① 2①
expressions among the following:

(a) − 3
+ 1
1
<−
< 1 ;
2① ①
2 2①
(b) − < − 3 + 12 < 1 ;
1
① 2① 2①
(c) − 1 < 1 < − 3 + 12 ;
① 2① 2①
(d) There are no order relations between 1 , − 1 and −3 1 + 12 because they are
2① ① 2①
not numbers;
(e) − 1
= 1 = 0 < − 3 + 12 = 12 ;
① 2① 2①
(f) The expressions 1
, − 1 and −3 1 + 12 have no sense because it is not possible
2① ① 2①
to divide by infinity.

Other questions of the test had the aim to solicit a comparison with the
symbol ∞ as used in traditional mathematics.

Question 10. Consider the writing ∞ + ∞ and ∞ · ∞. Choose the true options
among the following:
 New Approaches to Basic Calculus 337

(a) ∞ + ∞ and ∞ · ∞ have no sense in mathematics;

(b) ∞ + ∞ = ∞ and ∞ · ∞ = ∞;
(c) ∞ + ∞ = ∞ and ∞ + ∞ = ∞;
(d) ∞ + ∞ = 2∞ < ∞ · ∞ = ∞2 ;
(e) ∞ + ∞ and ∞ · ∞ are not comparable;
(f) ∞ < ∞ + ∞ < ∞ · ∞.

Question 11. Consider the writing ∞ − ∞ and mark the right options among the
following:
(a) ∞ − ∞ = ∞;
(b) −∞ < ∞ − ∞ < ∞;
(c) ∞ − ∞ is an indeterminate expression;
(d) ∞ − ∞ = 0;
(e) −∞ < ∞ − ∞ < ∞;
(f) ∞ − ∞ < ∞ + ∞;
(g) ∞ − ∞ and −∞ are not comparable.

Question 12. For each of the following items make a mark on “T” or “F”
if you believe the corresponding statement to be true or false, respectively.
(a) ① < +∞ T-F
(b) ① = +∞ T-F
(c) ① and ∞ are not comparable T-F
(d) ① ≤ +∞ T-F
(e) ① ≥ +∞ T-F
(f) ① and ∞ cannot be used together because the belong to
different settings T-F
(g) ①2 > +∞ T - F
(h) ①+1=① T - F
(i) ∞+1=∞ T - F
(j) ∞+1>∞ T - F

Question 13. In the classical setting, consider the function given by the analytical
x2
expression f (x) = x+1 .
(a) Compute the domain of the function and the limits at the extremal points of
the domain.
(b) Compute the asymptotes of f (vertical, horizontal and oblique).
x2
Question 14. In the grossone setting, consider the function f (x) = x+1 .

(a) Compute the values of f at −① and ①.

(b) Compute the values of f at −① + 2, −① − 2 and ① + 1.
(c) Let y = a(x) be the right oblique asymptote of f (if it exists).
• Compute the value a(①) and a(① + 1).
• Compute the difference f (①) − a(①) and f (① + 1) − a(① + 1).
338 L. Antoniotti et al.

Question 15 (excluded from the evaluation). In your opinion there are some advan-
tages in using ① in the place of +∞ (i.e., the grossone setting in the place of the
classical one)? Justify the answer and, if yes, list some of them.
The last extra question, although proposed together the final test, was
excluded from any attribution of a score and this was clearly written. In any
case, most of the students did not answer this question, or gave hasty and little
significant answers: a different outcome would probably have been recorded if it
had been proposed on a different day rather than at the end of a test with many
questions.
The students of class T 5 have also been spoken in classroom of the Infin-
ity Computer, trying to motivate them in a deepening and individual work at
home on it. From the compilation of the informative questionnaire and from a
continuous dialogue with the students, however, it emerged that only 4 of them
actually used the Infinity Computer at least once at home, independently. For
convenience, we will denote this group of students by T 5.1.
As regards unimaginable numbers we inform the reader that, for the class T 5,
a soft approach to Knuth’s notation and very large numbers was also planned,
but just in part developed with the students. In particular we presented tetra-
tions and pentations to them, and in a first moment it seemed very successful,
in particular the way to iterate the ordinary exponentiation and to write it
compactly under the form of a tetration. Many problems and much confusion
emerged instead in a successive lesson two weeks later, and we decided to ask
no questions about these topics in the final test (cf. the conclusions in Sect. 5).

4 Results and Their Analysis

In this section we will give the results of the tests, will discuss the situation
picture emerging from the questionnaires and the dialogues with the students,
and will finally draw up a balance of the experimentations giving some brief
conclusions jointly with the next section.
The zero-knowledge test has been proposed to all the 8 classes, with major
differences between the groups P and T ∪ T , but with several minor ones inside
the sample T ∪ T . In Table 2, the columns 2–5 are devoted to this test: the 2nd
and 3rd columns give the mean student score [Mean] obtained in each class and
the maximum score [Max], respectively, both over the total score of the test.
The 4th column gives the normalized mean (students’) score [NMS] and the 5th
column the standard deviation [SD]. The column from 6 to 9 are the twin ones
of 2–5, but relative to the final test (notations are hence obtained by adding “F”
to the previous ones).
The results of the zero-knowledge test are, in general, positive or very posi-
tive, in dependence of the cases, as Table 2 shows. In particular, the maximum
scores in the 3rd column unequivocally give the measure of how intuitive it is to
perform basic calculations with the grossone system.
The scores of the final test are higher and with more gap from the corre-
sponding one of the initial test, in particular for the classes with less teaching
 New Approaches to Basic Calculus 339

Table 2. The results of the tests for each class.

Class Mean Max NMS SD Mean F Max F NMS F SD F

P2 15.9/21 21/21 0.76 2.7/21 – – – –
P3 16.7/21 20/21 0.79 2.1/21 – – – –
T3 10.5/24 18/24 0.43 2.9/24 15.1/24 21/24 0.63 2.7/24
T 3 8.1/24 15/24 0.34 3.9/24 – – – –
T4 10.8/24 15/24 0.45 4.1/24 – – – –
T 4 8.4/24 15/24 0.35 3.7/24 13.2/24 20/24 0.55 3.1/24
T5 16.2/24 24/24 0.67 5.5/24 20.7/24 24/24 0.86 2.7/24
T 5 8.1/24 18/24 0.34 3.9/24 14.2/24 22/24 0.59 3.6/24
T 5.1 21.8/24 24/24 0.91 1.3/24 23.3/24 24/24 0.97 1.3/24

time. This result is rather unexpected and seems to be due to the fact that the
class T 5, the one with the highest number of lessons, starts from a very high
score in the initial test. But it could also mean that a few quick explanations
are enough to significantly improve some good initial performances (considering
that it is a text with zero knowledge) especially if not very high.
As regards the small T 5.1 group (i.e., the one consisting of the 4 students
of the class T 5 that used the Infinity Computer at least once at home), we can
observe very high performances in both columns 2 and 6, with a small increase.
Probably, a more difficult extra test for this group would have been interesting
both at the beginning and at the end of the experimentation. Finally, from
the questionnaire and, in particular, from a continuous conversation with the
students, we think that their approach with the Infinity Computer has been
fruitful to arouse attraction and to give further motivation and interest, probably
because it is seen as a form of concrete application (recall that we are dealing
with a technical-commercial school) of the grossone-based system.

5 Conlusion

The good results obtained in the various levels of the experimentation have
shown a remarkable usability for the students of the new concept of infinity rep-
resented by grossone. It should in fact be emphasized that in all the 8 classes that
took part in the experimentation, most of the students succeeded in assimilat-
ing, or better, effectively understanding, the distinctive properties of ①, and the
correct way of performing calculations in the associate numerical-computational
system in a few minutes, already at the initial zero-knowledge test. Very inter-
esting and useful to the students, it was also the comparison, made several times
during the lessons given to the classes T 5 and T 5 , between the classical con-
ception of infinity (dating back to Cantor and Weierstrass) and that related to
grossone: in fact the students, stimulated by the possibility of working computa-
tionally with ①, in an “unconsciously familiar” way, showed a marked interest,
340 L. Antoniotti et al.

very difficult to be aroused in general, also for the more theoretical aspects
concerning the two models of infinity. The possibility of carrying out a wider
experimentation and proposing Sergeyev’s model on a larger scale could there-
fore have relevant educational implications.
We also believe that the Infinity Computer can also have a good educational
value, which however we have not had the opportunity to investigate or test in
depth in a very short experimentation, and with many new features for students
like ours.
Similarly, with regard to unimaginable numbers, we conclude that they could
have very interesting didactic applications (especially from the point of view of
generalizing the usual operations of addition, multiplication and exponentiation
via tetration, pentation, hexation, etc.), but they require experimental activities
completely dedicated to them because a not so easy assimilation of such topics
has emerged, at least among the students of a technical school, more used to
calculations than to algebraic formalisms.

Aknowledgments. This work is partially supported by the research projects

“IoT&B, Internet of Things and Blockchain”, CUP J48C17000230006, POR Calabria
FESR-FSE 2014-2020.

References
1. Antoniotti, L., Caldarola, F., Maiolo, M.: Infinite numerical computing applied to
Peano’s, Hilbert’s and Moore’s curves. Mediterr. J. Math. to appear
2. Bertacchini, F., Bilotta, E., Pantano, P., Tavernise, A.: Motivating the learning
of science topics in secondary schools: a constructivistic edutainment setting for
studing Chaos. Comput. Educ. 59(4), 1377–1386 (2012)
3. Bertacchini, F., Bilotta, E., Caldarola, F., Pantano, P.: Complex interactions in
one-dimensional cellular automata and linguistic constructions. Appl. Math. Sci.
12(15), 691–721 (2018). https://doi.org/10.12988/ams.2018.8353
4. Bertacchini, F., Bilotta, E., Caldarola, F., Pantano, P.: The role of computer sim-
ulations in learning analytic mechanics towards chaos theory: a course experimen-
tation. Int. J. Math. Educ. Sci. Technol. 50(1), 100–120 (2019). https://doi.org/
10.1080/0020739X.2018.1478134
5. Bertacchini, F., Bilotta, E., Caldarola, F., Pantano, P., Renteria Bustamante, L.:
Emergence of linguistic-like structures in one-dimensional cellular automata. In:
Sergeyev, Ya.D., Kvasov, D.E., Dell’Accio, F., Mukhametzhanov, M.S. (eds.) 2nd
International Conference on NUMTA 2016 - Numerical Computations: Theory and
Algorithms, AIP Conference Proceedings, vol. 1776, p. 090044. AIP Publ., New
York (2016). https://doi.org/10.1063/1.4965408
6. Blakley, G.R., Borosh, I.: Knuth’s iterated powers. Adv. Math. 34(2), 109–136
(1979). https://doi.org/10.1016/0001-8708(79)90052-5
7. Bonavoglia, P. (ed.): Atti del convegno della VII Giornata Nazionale di Analisi
non Standard per le scuole superiori, Venezia, 30 settembre 2017. Edizioni Aracne,
Roma (2017). (in Italian)
8. Caldarola, F.: The Sierpiński curve viewed by numerical computations with infini-
ties and infinitesimals. Appl. Math. Comput. 318, 321–328 (2018). https://doi.
org/10.1016/j.amc.2017.06.024
 New Approaches to Basic Calculus 341

9. Caldarola, F.: The exact measures of the Sierpiński d-dimensional tetrahedron in

connection with a Diophantine nonlinear system. Commun. Nonlinear Sci. Numer.
Simul. 63, 228–238 (2018). https://doi.org/10.1016/j.cnsns.2018.02.026
10. Caldarola, F., Cortese, D., d’Atri, G., Maiolo, M.: Paradoxes of the infinite and
ontological dilemmas between ancient philosophy and modern mathematical solu-
tions. In: Sergeyev, Y.D., Kvasov, D.E. (eds.) NUMTA 2019. LNCS, vol. 11973,
pp. 358–372. Springer, Cham (2020)
11. Caldarola, F., d’Atri, G., Maiolo, M.: What are the unimaginable numbers? Sub-
mitted for publication
12. Caldarola, F., d’Atri, G., Mercuri, P., Talamanca, V.: On the arithmetic of Knuth’s
powers and some computational results about their density. In: Sergeyev, Y.D.,
Kvasov, D.E. (eds.) NUMTA 2019. LNCS, vol. 11973, pp. 381–388. Springer, Cham
(2020)
13. Caldarola, F., Maiolo, M., Solferino, V.: A new approach to the Z-transform
through infinite computation. Commun. Nonlinear Sci. Numer. Simul. 82, 105019
(2020). https://doi.org/10.1016/j.cnsns.2019.105019
14. Cornu, B.: Quelques obstacles à l’apprentissage de la notion de limite. Recherches
en Didactique des Mathematiques 4, 236–268 (1983)
15. Davis, R.B., Vinner, S.: The notion of limit: some seemingly unavoidable miscon-
ception stages. J. Math. Behav. 5, 281–303 (1986)
16. Di Nasso, M.: Un’introduzione all’analisi con infinitesimi. In: Analisi nonstandard
per le scuole superiori - V Giornata di studio, Atti del convegno, Verona, 10 Ottobre
2015, Matematicamente.it (2016). (in Italian)
17. Ely, R.: Nonstandard student conceptions about infinitesimals. J. Res. Math. Educ.
41(2), 117–146 (2010)
18. Ervynck, G.: Conceptual difficulties for first year university students in the acqui-
sition of the notion of limit of a function. In: Proceedings of the Fifth Conference
of the International Group for the Psychology of Mathematics Education, pp. 330–
333 (1981)
19. Guzman, M., Hodgson, B., Robert, A., Villani, V.: Difficulties in the passage from
secondary to tertiary education. In: Fischer, G., Rehmann, U. (eds.) Proceedings
of the International Congress of Mathematicians, Berlin 1998, vol. 3, pp. 747–762
(1998)
20. Iannone, P., Rizza, D., Thoma, A.: Investigating secondary school students’ episte-
mologies through a class activity concerning infinity. In: Bergqvist, E., et al. (eds.)
Proceedings of the 42nd Conference of the International Group for the Psychology
of Mathematics Education, Umeå(Sweden), vol. 3, pp. 131–138. PME (2018)
21. Ingarozza, F., Adamo, M.T., Martino, M., Piscitelli, A.: A Grossone-Based Numer-
ical Model for Computations with Infinity: A Case Study in an Italian High School.
In: Sergeyev, Y.D., Kvasov, D.E. (eds.) NUMTA 2019. LNCS, vol. 11973, pp. 451–
462. Springer, Cham (2020). https://doi.org/10.1007/978-3-030-39081-5 39
22. Keisler, H.J.: Elementary Calculus: An Approach Using Innitesimals, 3rd edn.
Dover Publications (2012)
23. Khan, S.: New pedagogies on teaching science with computer simulations. J. Sci.
Educ. Technol. 20, 215–232 (2011)
24. Klymchuk, S., Zverkova, T., Gruenwald, N., Sauerbier, G.: University students’
difficulties in solving application problems in calculus: student perspectives. Math.
Educ. Res. J. 22(2), 81–91 (2010)
25. Knuth, D.E.: Mathematics and computer science: coping with finiteness. Science
194(4271), 1235–1242 (1976). https://doi.org/10.1126/science.194.4271.1235
342 L. Antoniotti et al.

26. Lithner, J.: Students’ mathematical reasoning in university textbook exercises.

Educ. Stud. Math. 52, 29–55 (2003)
27. Lithner, J.: Mathematical reasoning in calculus textbook exercises. J. Math. Behav.
23, 405–427 (2004)
28. Lithner, J.: University students’ learning difficulties. Educ. Inquiry 2(2), 289–303
(2011)
29. Muzangwa, J., Chifamba, P.: Analysis of errors and misconceptions in the learning
of calculus by undergraduate students. Acta Didactica Napocensia 5(2) (2012)
30. Rizza, D.: Primi passi nell’aritmetica dell’infinito. Preprint (2019)
31. Sergeyev, Y.D.: A new applied approach for executing computations with infinite
and infinitesimal quantities. Informatica 19(4), 567–96 (2008)
32. Sergeyev, Y.D.: Arithmetic of Infinity. Edizioni Orizzonti Meridionali, Cosenza
(2003)
33. Sergeyev, Y.D.: Evaluating the exact infinitesimal values of area of Sierpinskinski’s
carpet and volume of Menger’s sponge. Chaos Solitons Fractals 42(5), 3042–6
(2009)
34. Sergeyev, Y.D.: Higher order numerical differentiation on the Infinity Computer.
Optim. Lett. 5(4), 575–85 (2011)
35. Sergeyev, Y.D.: (2004). http://www.theinfinitycomputer.com
36. Sergeyev, Y.D.: Lagrange Lecture: methodology of numerical computations with
infinities and infinitesimals. Rendiconti del Seminario Matematico Università e
Politecnico di Torino 68(2), 95–113 (2010)
37. Sergeyev, Y.D.: Solving ordinary differential equations by working with infinitesi-
mals numerically on the Infinity Computer. Appl. Math. Comput. 219(22), 10668–
81 (2013)
38. Sergeyev, Y.D.: Un semplice modo per trattare le grandezze infinite ed infinitesime.
Matematica, Cultura e Società: Rivista dell’Unione Matematica Italiana 8(1), 111–
47 (2015). (in Italian)
39. Sergeyev, Y.D.: Using blinking fractals for mathematical modelling of processes of
growth in biological systems. Informatica 22(4), 559–76 (2011)
40. Sergeyev, Y.D.: Numerical infinities and infinitesimals: methodology, applications,
and repercussions on two Hilbert problems. EMS Surv. Math. Sci. 4(2), 219–320
(2017)
41. Sergeyev, Y.D., Mukhametzhanov, M.S., Mazzia, F., Iavernaro, F., Amodio, P.:
Numerical methods for solving initial value problems on the Infinity Computer.
Int. J. Unconv. Comput. 12(1), 55–66 (2016)
42. Sierpinska, A.: Humanities students and epistemological obstacles related to limits.
Educ. Stud. Math. 18, 371–397 (1987)
43. Steinberg, R.N.: Computer in teaching science: to simulate or not simulate? Phys.
Educ. Res.: Am. J. Phys. Suppl. 68(7), S37 (2000)
44. Tall, D.: A child thinking about infinity. J. Math. Behav. 20, 7–19 (2001)
45. Tall, D.: Mathematical intuition, with special reference to limiting processes. In:
Proceedings of the Fourth International Conference for the Psychology of Mathe-
matics Education, pp. 170–176 (1980)
46. Tall, D.: The transition to advanced mathematical thinking: functions, limits, infin-
ity and proof. In: Grouws, D. (ed.) Handbook for Research on Mathematics Teach-
ing and Learning, pp. 495–511. Macmillan, New York (1992)
47. Tall, D., Vinner, S.: Concept image and concept definition in mathematics with
particular reference to limits and continuity. Educ. Stud. Math. 12, 151–169 (1981)
48. Williams, S.R.: Models of limit held by college calculus students. J. Res. Math.
Educ. 22(3), 219–236 (1991)
 A Multi-factor RSA-like Scheme
with Fast Decryption Based on Rédei
Rational Functions over the Pell
Hyperbola

Emanuele Bellini1 and Nadir Murru2(B)

1
Technology Innovation Institute, Abu Dhabi, UAE
[email protected]
2
Department of Mathematics, University of Turin, Turin, Italy
[email protected]

Abstract. We propose a generalization of an RSA-like scheme based on

Rédei rational functions over the Pell hyperbola. Instead of a modulus
which is a product of two primes, we define the scheme on a multi-factor
modulus, i.e. on a product of more than two primes. This results in a
scheme with a decryption which is quadratically faster, in the number of
primes factoring the modulus, than the original RSA, while preserving
a better security. The scheme reaches its best efficiency advantage over
RSA for high security levels, since in these cases the modulus can contain
more primes. Compared to the analog schemes based on elliptic curves,
as the KMOV cryptosystem, the proposed scheme is more efficient. Fur-
thermore a variation of the scheme with larger ciphertext size does not
suffer of impossible group operation attacks, as it happens for schemes
based on elliptic curves.

Keywords: Cryptography · Pell conic · Rédei rational functions · RSA

1 Introduction
RSA is the most widespread asymmetric encryption scheme. Its security is based
on the fact that the trapdoor function τN,e (x) = xe mod N , where N = pq is the
product of two large prime integers, and e an invertible element in Zφ(N ) (φ(N )
being the Euler totient function), cannot be inverted by a polynomial-time in N
algorithm without knowing either the integers p, q, φ(N ) or the inverse d of e
modulo φ(N ). Thus the pair (N, e), called the public key, is known to everyone,
while the triple (p, q, d), called the secret key, is only known to the receiver of an
encrypted message. Both encryption and decryption are performed through an
exponentiation modulo N . Precisely, the ciphertext C is obtained as C = M e
(mod N ), and the original message M is obtained with the exponentiation M =
C d (mod N ). While usually the encryption exponent is chosen to be small, the
decryption exponent is about the size of N , implying much slower performances
c Springer Nature Switzerland AG 2020
Y. D. Sergeyev and D. E. Kvasov (Eds.): NUMTA 2019, LNCS 11973, pp. 343–357, 2020.
https://doi.org/10.1007/978-3-030-39081-5_30
344 E. Bellini and N. Murru

during decryption with respect to encryption. Through the years many proposal
have been presented trying to speed up the decryption process.
In this work we present the fastest, to the authors knowledge, of such decryp-
tion algorithms whose security is based on the factorization problem. The pre-
sented scheme exploits different properties of Rédei rational functions, which
are classical functions in number theory. The proposed decryption algorithm is
quadratically, on the number of primes composing the modulus N , faster than
RSA.
The work is divided as follows. In Sect. 2 an overview of the main schemes
based on the factorization problem which successfully improved RSA decryp-
tion step is presented. In Sect. 3 the main theoretical results underlying our
scheme are described. Section 4 is devoted to the presentation of the crypto-
graphic scheme, and in Sects. 5 and 6 its security and efficiency are discussed,
respectively. Section 7 concludes the work.

2 Related Work
In this section we briefly overview the main cryptographic schemes based on
the factorization problem that have been introduced in order to improve RSA
decryption step.
Usually, the general technique to speed up the RSA decryption step C = M e
(mod N ) is to compute the exponentiation modulo each factor of N and then
obtain N using the Chinese Remainder Theorem.

2.1 Multifactor RSA

There exists variants of RSA scheme which exploit a modulus with more than
2 factors to achieve a faster decryption algorithm. This variants are sometimes
called Multifactor RSA [6], or Multiprime RSA [8,10]. The first proposal exploit-
ing a modulus of the form N = p1 p2 p3 has been patented by Compaq [9,10] in
1997. About at the same time Takagi [30] proposed an even faster solution using
the modulus N = pr q, for which the exponentiation modulo pr is computed
using the Hensel lifting method [11, p. 137]. Later, this solution has been gener-
alized to the modulus N = pr q s [28]. According to [10], the appropriate number
of primes to be chosen in order to resist state-of-the-art factorization algorithms
depends from the modulus size, and, precisely, it can be: up to 3 primes for 1024,
1536, 2048, 2560, 3072, and 3584 bit modulus, up to 4 for 4096, and up to 5 for
8192.

2.2 RSA-like Schemes

Another solution which allows to obtain even faster decryption is to use RSA-like
schemes based on isomorphism as [3,16,17,26]. As an additional property, these
schemes owns better security properties with respect to RSA, avoiding small
exponent attacks to either d [31] or e [12,13], and vulnerabilities which appear
 Multi-factor RSA-like Scheme over Pell Hyperbola with Fast Decsryption 345

when switching from one-to-one communication scenario to broadcast scenario

(e.g., see [14]). The aforementioned schemes are based on isomorphism between
two groups, one of which is the set of points over a curve, usually a cubic or a
conic. A complete overview on RSA-like schemes based on conics can be found
in [3]. In general, schemes based on cubic curves have a computationally more
expensive addition operation compared to schemes based on conic equations.

2.3 Generalizing RSA-like Scheme with Multifactor Modulus

As done when generalizing from RSA to Multiprime RSA, in [7] a generalization
of [16,17] has been proposed, thus generalizing a RSA-like scheme based on
elliptic curves and a modulus N = pq to a similar scheme based on the generic
modulus N = pr q s .
In this paper we present a similar generalization of the scheme [3], which
is based on the Pell’s equation, to the modulus N = pe11 · . . . · perr for r > 2,
obtaining the fastest decryption of all schemes discussed in this section.

3 Product of Points over the Pell Hyperbola

In [3], we introduced a novel RSA–like scheme based on an isomorphism between
certain conics (whose the Pell hyperbola is a special case) and a set of parameters
equipped with a non–standard product. In Sect. 4, we generalize this scheme
considering a prime power modulus N = pe11 · · · perr . In this section, we recall
some definitions and properties given in [3] in order to improve the readability
of the paper. Then, we study properties of the involved products and sets in Zpr
and ZN .

3.1 A Group Structure over the Pell Hyperbola over a Field

Let K be a field and x2 − D an irreducible polynomial over K[x]. Considering
the quotient field A[x] = K[x]/(x2 − D), the induced product over A[x] is

(p + qx)(r + sx) = (pr + qsD) + (qr + ps)x.

The group of unitary elements of A∗ [x] = A[x] − {0A[x] }1 is {p + qx ∈ A∗ [x] :
p −Dq 2 = 1}. Thus, we can introduce the commutative group (HD,K , ⊗), where
2

HD,K = {(x, y) ∈ K × K : x2 − Dy 2 = 1}
and
(x, y) ⊗ (w, z) = (xw + yzD, yw + xz), ∀(x, y), (w, z) ∈ HD,K . (1)
It is worth noting that (1, 0) is the identity and the inverse of an element (x, y)
is (x, −y).
Remark 1. When K = R, the conic HD,K , for D a non–square integer, is called
the Pell hyperbola since it contains all the solutions of the Pell equation and ⊗
is the classical Brahamagupta product, see, e.g., [15].
1
The element 0A[x] is the zero polynomial.
346 E. Bellini and N. Murru

3.2 A Parametrization of the Pell Hyperbola

From now on let A = A[x].
Starting from A∗ , we can derive a parametrization for HD,K . In particular,
let us consider the group A∗ /K∗ , whose elements are the equivalence classes of
A∗ and can be written as

{[a + x] : a ∈ K} ∪ {[1K∗ ]}.

The induced product over A∗ /K∗ is given by

[a + x][b + x] = [ab + ax + bx + x2 ] = [D + ab + (a + b)x]

and, if a + b = 0, we have

D + ab
[a + x][b + x] = [ + x]
a+b
else
[a + x][b + x] = [D + ab] = [1K∗ ].
This construction allows us to define the set of parameters PK = K ∪ {α},
with α not in K, equipped with the following product:
⎧
⎨ D + ab
ab= , a + b = 0
a+b . (2)
⎩
a  b = α, a + b = 0

We have that (PK , ) is a commutative group with identity α and the inverse
of an element a is the element b such that a + b = 0. Now, consider the following
parametrization for the conic HD,K :

1
y= (x + 1) .
m
It can be proved that the following isomorphism between (HD,K , ⊗) and (PK , )
holds: ⎧
⎪
⎪ HD,K → PK
⎪
⎪
⎪
⎨ 1+x
(x, y) → ∀(x, y) ∈ HD,K , y = 0
ΦD : y (3)
⎪
⎪
⎪
⎪ (1, 0) →
α
⎪
⎩
(−1, 0) → 0 ,
and ⎧
⎪
⎪PK →H  D,K 
⎪
⎨ m2 + D 2m
Φ−1 : m → , ∀m ∈ K , (4)
D
⎪
⎪ m2 − D m2 − D
⎪
⎩
α → (1, 0) ,
see [1] and [3].
 Multi-factor RSA-like Scheme over Pell Hyperbola with Fast Decsryption 347

Proposition 1. When K = Zp , p prime, (PK , ) and (HD,K , ⊗) are cyclic

groups of order p + 1 and
m(p+2) = m (mod p), ∀m ∈ PZp
or, equivalently
(x, y)⊗(p+2) = (x, y) (mod p), ∀(x, y) ∈ HD,Zp ,
where powers are performed using products  and ⊗, respectively. See [3].
The powers in PK can be efficiently computed by means of the Rédei rational
functions [27], which are classical functions in number theory. They are defined
by considering the development of
√ √
(z + D)n = An (D, z) + Bn (D, z) D,
for z integer and D non–square positive integer. The polynomials An (D, z) and
Bn (D, z) defined by the previous expansion are called Rédei polynomials and
can be evaluated by

n An (D, z) DBn (D, z)
M =
Bn (D, z) An (D, z)
where 
zD
M= .
1 z
From this property, it follows that the Rédei polynomials are linear recurrent
sequences with characteristic polynomial t2 − 2zt + (z 2 − D). The Rédei rational
functions are defined by

An (D, z)
Qn (D, z) = , ∀n ≥ 1.
Bn (D, z)
Proposition 2. Let mn be the n–th power of m ∈ PK with respect to , then
mn = Qn (D, m).
See [2].
Remark 2. The Rédei rational functions can be evaluated by means of an algo-
rithm of complexity O(log2 (n)) with respect to addition, subtraction and mul-
tiplication over rings [24].

3.3 Properties of the Pell Hyperbola over a Ring

In this section, we study the case K = Zpr that we will exploit in the next
section for the construction of a cryptographic scheme. In what follows, we will
omit from HD,K the dependence on D when it will be clear from the context.
First, we need to determine the order of HZpr in order to have a result similar
to Proposition 1 also in this situation.
348 E. Bellini and N. Murru

Theorem 1. The order of the cyclic group HZpr is pr−1 (p + 1), i.e., the Pell
equation x2 − Dy 2 = 1 has pr−1 (p + 1) solutions in Zpr for D ∈ Z∗pr quadratic
non–residue in Zp .
Proof. Since, by Proposition 1, the Pell equation in Zp has p + 1 solutions, then
we need to prove the following
1. any solution of the Pell equation in Zp , generates pr−1 solutions of the same
equation in Zpr ;
2. all the solutions of the Pell equation in Zpr are generated as in the previous
step.
(1) Let (x0 , y0 ) be a solution of x2 − Dy 2 ≡ 1 (mod p). We want to prove that
for any integer 0 ≤ k < pr−1 , there exists one and only one integer h such
that (x0 + kp, y0 + hp) is solution of x2 − Dy 2 ≡ 1 (mod pr ).
Indeed, we have
(x0 + kp)2 − D(y0 + hp)2 = 1 + vp + 2x0 kp + k 2 p2 − 2Dy0 hp − Dh2 p2 ,
since x20 − Dy02 = 1 + vp for a certain integer v. Thus, we have that (x0 +
kp, y0 + hp) is solution of x2 − Dy 2 ≡ 1 (mod pr ) if and only if
Dph2 + 2Dy0 h − v − 2x0 k − k 2 p ≡ 0 (mod pr−1 ).
Hence, we have to prove that there is one and only one integer h that satisfies
the above identity. The above equation can be solved in h by completing the
square and reduced to
(2Dph + 2Dy0 )2 ≡ s (mod pr−1 ), (5)
where s = (2Dy0 )2 + 4(v + 2x0 k + k 2 p)Dp. Let us prove that s is a quadratic
residue in Zpr−1 . Indeed,
s = 4D((x0 + kp)2 − 1)
   r−1
s s
and surely the Jacobi symbol r−1
= = 1 if r is odd. If r is even
p p
we have that
     
s 4 D (x0 + kp)2 − 1
= =1
pr−1 pr−1 pr−1 pr−1
     r−1
4 D D
since = 1, = = −1 by hypothesis on D,
pr−1 pr−1 p
 
(x0 + kp)2 − 1
= −1, since (x0 + kp)2 − 1 ≡ Dy02 (mod p). Now, let ±t
pr−1
be the square roots of s. It is easy to note that
t ≡ 2Dy0 (mod p), −t ≡ −2Dy0 (mod p)
 Multi-factor RSA-like Scheme over Pell Hyperbola with Fast Decsryption 349

or
−t ≡ 2Dy0 (mod p), t ≡ −2Dy0 (mod p).
Let us call t̄ the only one between t and −t that is equal to 2Dy0 in Zp .
Hence, Eq. (5) is equivalent to the linear equation

ph ≡ (t̄ − 2Dy0 )(2D)−1 (mod pr−1 ),

which has one and only one solution, since t̄ − 2Dy0 ≡ 0 (mod p). Note that,
if t̄ is not equal to 2Dy0 in Zp the above equation has no solutions. Thus,
we have proved that any solution of the Pell equation in Zp generates pr−1
solutions of the Pell equation in Zpr .
(2) Now, we prove that all the solutions of the Pell equation in Zpr are generated
as in step 1.
Let (x̄, ȳ) be a solution of x2 − Dy 2 ≡ 1 (mod pr ), i.e., x̄2 − Dȳ 2 = 1 + wpr ,
for a certain integer w. Then x0 = x̄ − kp and y0 = ȳ − hp, for h, k integers,
are solutions of x2 − Dy 2 ≡ 1 (mod p). Indeed,

(x̄ − kp)2 − D(ȳ − hp)2 = 1 + wpr − 2x̄kp + k 2 p2 + 2Dȳhp − Dh2 p2 .

As a consequence of the previous theorem, an analogous of the Euler theorem

holds for the product ⊗.

Theorem 2. Let p, q be prime numbers and N = pr q s , then for all (x, y) ∈ HZN
we have r−1 s−1
(x, y)⊗p (p+1)q (s+1) ≡ (1, 0) (mod N )
for D ∈ Z∗N quadratic non–residue in Zp and Zq .

Proof. By Theorem 1, we know that

r−1
(x, y)⊗p (p+1)
≡ (1, 0) (mod pr )

and s−1
(x, y)⊗q (s+1)
≡ (1, 0) (mod q s ).
r−1
(p+1)q s−1 (s+1)
Thus, said (a, b) = (x, y)⊗p , we have

(a, b) ≡ (1, 0) (mod pr ),

i.e., a = 1 + kpr and b = hpr for some integers h, k. On the other hand, we have

(a, b) ≡ (1, 0) (mod q s ) ⇔ (1 + kpr , hpr ) ≡ (1, 0) (mod q s ).

We can observe that 1 + kpr ≡ 1 (mod q s ) if and only if k = k  q s for a certain

integer k  . Similarly, it must be h = h q s , for an integer h . Hence, we have that
(a, b) = (1 + k  pr q s , h pr q s ) ≡ (1, 0) (mod N ).
350 E. Bellini and N. Murru

Corollary 1. Let p1 , ..., pr be primes and N = pe11 · . . . · perr , then for all (x, y) ∈
HZN we have
(x, y)⊗Ψ (N ) = (1, 0) (mod N ),
where
Ψ (N ) = pe11 −1 (p1 + 1) · . . . · perr −1 (pr + 1),
for D ∈ Z∗N quadratic non–residue in Zpi , for i = 1, ..., r.

Now, we can observe that when we work on ZN , the map ΦD is not an

isomorphism. Indeed, the orders of HD,ZN and PZN do not coincide. However,
it is still a morphism and we also have |Z∗N | = |HZ∗ N |, because of the following
proposition.
Proposition 3. With the above notation, we have that

1. ∀(x1 , y1 ), (x2 , y2 ) ∈ HZ∗ N , ΦD (x1 , y1 ) = ΦD (x2 , y2 ) ⇔ (x1 , y1 ) = (x2 , y2 );

2. ∀m1 , m2 ∈ Z∗N , Φ−1 −1
D (m1 ) = ΦD (m2 ) ⇔ m1 = m2 ;
3. ∀m ∈ ZN , we have Φ (m) ∈ HZ∗ N and ∀(x, y) ∈ HZ∗ N , we have ΦD (x, y) ∈
∗ −1

Z∗N .

See [3].
As a consequence, we have an analogous of the Euler theorem also for the
product , i.e., for all m ∈ Z∗N the following holds

mΨ (N ) = α (mod N ) ,

where  is the special product in PZN defined in Eq. 2.

4 The Cryptographic Scheme

In this section, we describe our public–key cryptosystem based on the properties

studied in the previous section.

4.1 Key Generation

The key generation is performed by the following steps:

rr prime numbers p1 , . . . , pr , r odd integers e1 , . . . , er and compute
– choose
N = i=1 pei i ;
r
– choose an integer e such that
gcd(e, lcm i=1 pei i −1 (pi + 1)) = 1;
– evaluate d = e−1 (mod lcm i=1 pei i −1 (pi + 1)).
r

The public or encryption key is given by (N, e) and the secret or decryption key
is given by (p1 , . . . , pr , d).
 Multi-factor RSA-like Scheme over Pell Hyperbola with Fast Decsryption 351

4.2 Encryption


M 2
− 1
We can encrypt pair of messages (Mx , My ) ∈ Z∗N × Z∗N , such that
x
=
N
−1. This condition will ensure that we can perform all the operations. The
encryption of the messages is performed by the following steps:
Mx2 − 1 ∗
– compute D = (mod N ), so that (Mx , My ) ∈ HD,Z ;
My2 N

Mx + 1
– compute M = Φ(Mx , My ) = (mod N );
My
– compute the ciphertext C = M e (mod N ) = Qe (D, M ) (mod N )

Notice that not only C, but the pair (C, D) must be sent through the insecure
channel.

4.3 Decryption

The decryption is performed by the following steps:

– compute C d (mod N) = Qd (D, C) (mod

 N) = M;
2
M +D 2M
– compute Φ−1 (M ) = , (mod N ) for retrieving the mes-
M2 − D M2 − D
sages (Mx , My ).

5 Security of the Encryption Scheme

The proposed scheme can be attacked by solving one of the following problems:

1. factorizing the modulus N = pe11 · . . . · perr ;

2. computing Ψ (N ) = pe11 −1 (p1 + 1) · . . . · perr −1 (pr + 1), or finding the number of
solutions of the equation x2 − Dy 2 ≡ 1 mod N , i.e. the curve order, which
divides Ψ (N );
3. computing Discrete Logarithm problem either in (HZ∗ N , ⊗) or in (PZ∗N , );
4. finding the unknown d in the equation ed ≡ 1 mod Ψ (N );
5. finding an impossible group operation in PZN ;
6. computing Mx , My from D.

5.1 Factorizing N or Computing the Curve Order

It is well known that the problem of factorizing N = pe11 · . . . · perr is equivalent

to that of computing the Euler totient function φ(N ) = pe11 −1 (p1 − 1) · . . . ·
perr −1 (pr − 1), e.g. see [23] or [29, Section 10.4].
In our case we need to show the following
352 E. Bellini and N. Murru

Proposition 4. The problem of factorizing N is equivalent to the problem of

computing Ψ (N ) = pe11 −1 (p1 + 1) · . . . · perr −1 (pr + 1) or the order of the group
PZ∗N (or equivalently of HZ∗ N ), which is a divisor of Ψ (N ).

Proof. Clearly, knowing the factorization of N yields Ψ (N ). Conversely, sup-

pose N and Ψ (N ) are known. A factorization of N can be found by applying
Algorithm 1 recursively.

Remark 3. Algorithm 1 is an adaptation of the general algorithm in [29,

Section 10.4], used to factorize N by only knowing φ(N ) (Euler totient func-
tion) and N itself. The main idea of the Algorithm 1 comes from the fact that
xΨ (N ) = 1 (mod N ) for all x ∈ Z∗N , which is the analog of the Euler theorem
in PZN . Notice that, because of Step 7, Algorithm 1 is a probabilistic algorithm.
Thus, to find a non-trivial factor, it might be necessary to run the algorithm
more than once. We expect that a deeper analysis of the algorithm will lead
to a similar probabilistic behaviour than the algorithm in [29], which returns a
non-trivial factor with probability 1/2.

Algorithm 1. Find a factor of N by knowing N and Ψ (N )

1: function Find factor(N ,Ψ (N ))
2: h=0
3: t = Ψ (N )
4: while IsEven(t) do
5: h=h+1
6: t=t/2
7: a = Random(N − 1)
8: d = gcd(a, N )
9: if d = 1 then
10: return d
11: b = at mod N
12: for j = 0, . . . , h − 1 do
13: d = gcd(b + 1, N )
14: if d = 1 or d = N then
15: return d
16: b = b2 mod N
17: return 0

Since we proved that the problems 1 and 2 are equivalent, we can only focus
on the factorization problem.
According to [10], state-of-the-art factorization methods as the Elliptic Curve
Method [18] or the Number Field Sieve [4,19] are not effective if in the following
practical cases

with e1 +e2 +e3 ≤

– |N | = 1024, 1536, 2048, 2560, 3072, 3584 and N = pe11 pe22 pe33 √
3 and pi , i = 1, 2, 3 greater than approximately the size of 3 N .
 Multi-factor RSA-like Scheme over Pell Hyperbola with Fast Decsryption 353

– |N | = 4096 and N = pe11 pe22 pe33 pe44 with e√

1 +e2 +e3 +e4 ≤ 4 and pi , i = 1, . . . , 4
greater than approximately the size of 4 N .
– |N | = 8192 and N = pe11 pe22 pe33 pe44 pe55 with e1 + e2 + √e3 + e4 + e5 ≤ 5 and
pi , i = 1, . . . , 5 greater than approximately the size of 5 N .
Notice that currently, the largest prime factor found by the Elliptic Curve
Method is a 274 bit digit integer [32]. Note also that the Lattice Factoring
Method (LFM) of Boneh, Durfee, and Howgrave-Graham [5] is designed to fac-
tor integers of the form N = pu q only for large u.

5.2 Computing the Discrete Logarithm

Solving the discrete logarithm problem in a conic curve can be reduced to the
discrete logarithm problem in the underlying finite field [22]. In our case the
curve is defined over the ring ZN . Solving the DLP over ZN without knowing
the factorization of N is as hard as solving the DLP over a prime finite field of
approximately the same size. As for the factorization problem, the best known
algorithm to solve DLP on a prime finite field is the Number Field Sieve. When
the size of N is greater than 1024 then the NFS can not be effective.

5.3 Solving the Private Key Equation

In the case of RSA, small exponent attacks [12,13,31] can be performed to
find the unknown d in the equation ed ≡ 1 mod Ψ (N ). Generalization of these
attacks can be performed on RSA variants where the modulus is of the form
N = pe11 pe22 [20]. It has already been argued in [3,16] and [16] that this kind of
attacks fails when the trapdoor function is not a simple monomial power as in
RSA, as it is in the proposed scheme.

5.4 Finding an Impossible Group Operation

In the case of elliptic curves over ZN , as in the generalized KMOV cryptosystem
[7], it could happen that an impossible addition between two curve points occurs,
yielding the factorization of N . This is due to the fact that the addition formula
requires to perform an inversion in the underlying ring ZN . However, as shown
by the same authors of [7], the occurrence of an impossible addition is very
unlikely for N with few and large prime factors.
In our case an impossible group operation may occur if a + b is not invertible
in ZN , i.e. if gcd(a + b, N ) = 1, yielding in fact a factor of N . However, also in
our case, if N contains a few large prime factors, impossible group operations
occur with negligible probability, as shown by the following proposition.
Proposition 5. The probability to find an invertible element in PZN is approx-
imately  
1 1
1− 1− · ... · 1 −
p1 pr
354 E. Bellini and N. Murru

Proof. The probability to find an invertible element in PZN is given by dividing

the number of non-invertible elements in PZN by the total number of elements
of this set, as follows:
|PZN | − #{invertible elements in PZN }
(6)
|PZN |
|ZN | + 1 − (#{invertible elements in ZN } + 1)
= (7)
|ZN | + 1
N − φ(N )
= (8)
N +1
 
1 1
∼1 − 1 − · ... · 1 − (9)
p1 pr
   
where we used N ∼ N + 1 and φ(N ) = N 1 − p11 · . . . · 1 − p1r .

This probability tends to zero for large prime factors.

Let us notice that, in the Pell curve case, it is possible to avoid such sit-
uation, by performing encryption and decryption in HZ∗ N , without exploiting
the isomorphism operation. Here the group operation ⊗ is defined between two
points on the Pell curve, as in Eq. 1, and does not contain the inverse operation.
In the resulting scheme the ciphertext is obtained as (Cx , Cy ) = (Mx , My )⊗e ,
where the operation ⊗ depends on D. Thus the triple (Cx , Cy , D) must be trans-
mitted, resulting in a non-compressed ciphertext.

5.5 Recovering the Message from D

Mx2 −1
To recover the message pair (Mx , My ) from D = My2 (mod N ), the attacker
must solve the quadratic congruence Mx2 − DMy2 − 1 = 0 (mod N ) with respect
to the two unknowns Mx and My . Even if one of the two coordinates is known
(partially known plaintext attack), it is well known that computing square roots
modulo a composite integer N , when the square root exists, is equivalent to
factoring N itself.

5.6 Further Comments

As a conclusion to this section, we only mention that as shown in [3], RSA-
like schemes based on isomorphism own the following properties: they are more
secure than RSA in the broadcast scenario, they can be transformed to semanti-
cally secure schemes using standard techniques which introduce randomness in
the process of generating the ciphertext.

6 Efficiency of the Encryption Scheme

Recall that our scheme encrypts and decrypts messages of size 2 log N . To
decrypt a ciphertext of size 2 log N using CRT, standard RSA requires four full
 Multi-factor RSA-like Scheme over Pell Hyperbola with Fast Decsryption 355

exponentiation modulo N/2-bit primes. Basic algorithms to compute xd mod p

requires O(log d log2 p), which is equal to O(log3 p) if d ∼ p.
Using CRT, if N = pe11 · . . . · perr , our scheme requires at most r exponentiation
modulo N/r-bit primes.
This means that the final speed up of our scheme with respect to RSA is

4 · (N/2)3
= r2 /2 (10)
r · (N/r)3
When r = 2 our scheme is two times faster than RSA, as it has already been
shown in [3]. If r = 3 our scheme is 4.5 time faster, with r = 4 is 8 times faster,
and with r = 5 is 12.5 times faster.

7 Conclusions
We generalized an RSA-like scheme based on the Pell hyperbola from a modulus
that was a product of two primes to a generic modulus. We showed that this gen-
eralization leads to a very fast decryption step, up to 12 times faster than original
RSA for the security level of a modulus of 8192 bits. The scheme preserves all
security properties of RSA-like schemes, which are in general more secure than
RSA, especially in a broadcast scenario. Compared to similar schemes based on
elliptic curves it is more efficient. We also pointed that a variation of the scheme
with non-compressed ciphertext does not suffer of impossible group operation
attacks.

References
1. Barbero, S., Cerruti, U., Murru, N.: Generalized Rédei rational functions and ratio-
nal approximations over conics. Int. J. Pure Appl. Math 64(2), 305–317 (2010)
2. Barbero, S., Cerruti, U., Murru, N.: Solving the Pell equation via Rédei rational
functions. Fibonacci Q. 48(4), 348–357 (2010)
3. Bellini, E., Murru, N.: An efficient and secure RSA-like cryptosystem exploiting
Rédei rational functions over conics. Finite Fields Appl. 39, 179–194 (2016)
4. Bernstein, D.J., Lenstra, A.K.: A general number field sieve implementation. In:
Lenstra, A.K., Lenstra, H.W. (eds.) The development of the number field sieve.
LNM, vol. 1554, pp. 103–126. Springer, Heidelberg (1993). https://doi.org/10.
1007/BFb0091541
5. Boneh, D., Durfee, G., Howgrave-Graham, N.: Factoring N = pr q for large r.
Crypto 1666, 326–337 (1999)
6. Boneh, D., Shacham, H.: Fast variants of RSA. CryptoBytes 5(1), 1–9 (2002)
7. Boudabra, M., Nitaj, A.: A new generalization of the KMOV cryptosystem. J.
Appl. Math. Comput. 57, 1–17 (2017)
8. Ciet, M., Koeune, F., Laguillaumie, F., Quisquater, J.: Short private exponent
attacks on fast variants of RSA. UCL Crypto Group Technical Report Series CG-
2003/4, Université Catholique de Louvain (2002)
9. Collins, T., Hopkins, D., Langford, S., Sabin, M.: Public key cryptographic appa-
ratus and method. Google Patents, US Patent 5,848,159 (1998)
356 E. Bellini and N. Murru

10. Compaq: Cryptography using Compaq multiprime technology in a parallel process-

ing environment. ftp://15.217.49.193/pub/solutions/CompaqMultiPrimeWP.pdf.
Accessed 2019
11. Cohen, H.: A Course in Computational Algebraic Number Theory. Springer, Hei-
delberg (2013)
12. Coppersmith, D., Franklin, M., Patarin, J., Reiter, M.: Low-exponent RSA with
related messages. In: Maurer, U. (ed.) EUROCRYPT 1996. LNCS, vol. 1070, pp.
1–9. Springer, Heidelberg (1996). https://doi.org/10.1007/3-540-68339-9 1
13. Coppersmith, D.: Small solutions to polynomial equations, and low exponent RSA
vulnerabilities. J. Cryptol. 10(4), 233–260 (1997)
14. Hastad, J.: N using RSA with low exponent in a public key network. In: Williams,
H.C. (ed.) CRYPTO 1985. LNCS, vol. 218, pp. 403–408. Springer, Heidelberg
(1986). https://doi.org/10.1007/3-540-39799-X 29
15. Jacobson, M.J., Williams, H.C., Taylor, K., Dilcher, K.: Solving the Pell Equation.
Springer, New York (2009). https://doi.org/10.1007/978-0-387-84923-2
16. Koyama, K.: Fast RSA-type schemes based on singular cubic curves y 2 + axy ≡ x3
(mod n). In: Guillou, L.C., Quisquater, J.-J. (eds.) EUROCRYPT 1995. LNCS,
vol. 921, pp. 329–340. Springer, Heidelberg (1995). https://doi.org/10.1007/3-540-
49264-X 27
17. Koyama, K., Maurer, U.M., Okamoto, T., Vanstone, S.A.: New public-key schemes
based on elliptic curves over the Ring Zn . In: Feigenbaum, J. (ed.) CRYPTO 1991.
LNCS, vol. 576, pp. 252–266. Springer, Heidelberg (1992). https://doi.org/10.1007/
3-540-46766-1 20
18. Lenstra Jr., H.W.: Factoring integers with elliptic curves. Ann. Math. 126, 649–673
(1987)
19. Lenstra, A.K., Lenstra, H.W., Manasse, M.S., Pollard, J.M.: The number field
sieve. In: Lenstra, A.K., Lenstra, H.W. (eds.) The development of the number field
sieve. LNM, vol. 1554, pp. 11–42. Springer, Heidelberg (1993). https://doi.org/10.
1007/BFb0091537
20. Lu, Y., Peng, L., Sarkar, S.: Cryptanalysis of an RSA variant with moduli N = pr q l .
J. Math. Cryptol. 11(2), 117–130 (2017)
21. McEliece, R.J.: A public-key cryptosystem based on algebraic coding theory. Deep
Space Netw. Prog. Rep. 44, 114–116 (1978)
22. Menezes, A.J., Vanstone, S.A.: A note on cyclic groups, finite fields, and the discrete
logarithm problem. Appl. Algebr. Eng. Commun. Comput. 3(1), 67–74 (1992)
23. Miller, G.L.: Riemann’s hypothesis and tests for primality. In: Proceedings of Sev-
enth Annual ACM Symposium on Theory of Computing, pp. 234–239 (1975)
24. More, W.: Fast evaluation of Rédei functions. Appl. Algebr. Eng. Commun. Com-
put. 6(3), 171–173 (1995)
25. NIST: Round 1 Submissions. https://csrc.nist.gov/Projects/Post-Quantum-
Cryptography/Round-1-Submissions. Accessed 2019
26. Padhye, S.: A public key cryptosystem based on pell equation. IACR Cryptology
ePrint Archive, p. 191 (2006)
27. Rédei, L.: Über eindeutig umkehrbare polynome in endlichen körpern redei. Acta
Sci. Math. 11, 85–92 (1946)
28. Lim, S., Kim, S., Yie, I., Lee, H.: A generalized takagi-cryptosystem with a modulus
of the form pr q s . In: Roy, B., Okamoto, E. (eds.) INDOCRYPT 2000. LNCS,
vol. 1977, pp. 283–294. Springer, Heidelberg (2000). https://doi.org/10.1007/3-
540-44495-5 25
29. Shoup, V.: A Computational Introduction to Number Theory and Algebra. Cam-
bridge University Press, Cambridge (2009)
 Multi-factor RSA-like Scheme over Pell Hyperbola with Fast Decsryption 357

30. Takagi, T.: Fast RSA-type cryptosystem modulo p k q. In: Krawczyk, H. (ed.)
CRYPTO 1998. LNCS, vol. 1462, pp. 318–326. Springer, Heidelberg (1998).
https://doi.org/10.1007/BFb0055738
31. Wiener, M.J.: Cryptanalysis of short RSA secret exponents. IEEE Trans. Inf. The-
ory 36(3), 553–558 (1990)
32. Zimmermann, S.: 50 largest factors found by ECM. https://members.loria.fr/
PZimmermann/records/top50.html. Accessed 2017
 Paradoxes of the Infinite and Ontological
Dilemmas Between Ancient Philosophy
and Modern Mathematical Solutions

Fabio Caldarola1(B) , Domenico Cortese2 , Gianfranco d’Atri1 ,

and Mario Maiolo3
1
Department of Mathematics and Computer Science, Università della Calabria,
Cubo 31/A, 87036 Arcavacata di Rende, CS, Italy
{caldarola,datri}@mat.unical.it
2
Via degli Orti, 12, 89861 Tropea, VV, Italy
[email protected]
3
Department of Environmental and Chemical Engineering (DIATIC),
Università della Calabria, Cubo 42/B, 87036 Arcavacata di Rende, CS, Italy
[email protected]

Abstract. The concept of infinity had, in ancient times, an indistin-

guishable development between mathematics and philosophy. We could
also say that his real birth and development was in Magna Graecia, the
ancient South of Italy, and it is surprising that we find, in that time,
a notable convergence not only of the mathematical and philosophical
point of view, but also of what resembles the first “computational app-
roach” to “infinitely” or very large numbers by Archimedes. On the other
hand, since the birth of philosophy in ancient Greece, the concept of
infinite has been closely linked with that of contradiction and, more pre-
cisely, with the intellectual effort to overcome contradictions present in
an account of Totality as fully grounded. The present work illustrates the
ontological and epistemological nature of the paradoxes of the infinite,
focusing on the theoretical framework of Aristotle, Kant and Hegel, and
connecting the epistemological issues about the infinite to concepts such
as the continuum in mathematics.

Keywords: Infinite · Pythagorean school · Greek philosophy ·

Grossone · Unimaginable numbers

1 Introduction
“Mathematics is the science of infinity” says the very famous sentence of Her-
mann Weyl, 1930 (see [53, p. 17]). And a quite surprising fact is that the South of
Italy, in particular Calabria and Sicily, played a historic role of the highest impor-
tance in the development of the idea of infinity in mathematics and philosophy,
disciplines that at the time were often not distinguishable in many respects. One
could almost say that the infinite from the mathematical-philosophical point of
c Springer Nature Switzerland AG 2020
Y. D. Sergeyev and D. E. Kvasov (Eds.): NUMTA 2019, LNCS 11973, pp. 358–372, 2020.
https://doi.org/10.1007/978-3-030-39081-5_31
 Paradoxes of the Infinite Between Philosophy and Mathematics 359

view was born in Magna Graecia as well as from the computational point of view,
in fact, “unimaginable numbers” were born in the Greek colonies of Sicily by the
greatest mathematician of antiquity, Archimedes of Syracuse. It is therefore very
interesting to investigate how the concept of infinity was born, developed and
evolved on the border between mathematics and philosophy and this is what the
present paper wants to do, at least in part. A further coincidence is that recent
speculations on the idea and use of infinity, in mathematics but not only, once
again see Calabria as the protagonist, as we will see later.
From Sect. 2, we will analyze and explain why the idea of the infinite has
always been perceived as a contradiction in relation with the philosophical and
scientific necessity of systematizing the entire reality within a complete and
univocal set of measures and forces. Such a contradiction took two forms:
1. The idea of infinite as a being which pushes itself beyond any given limit in
size or time. Once one assumes an infinite being, there is no longer a way
to hypothesize a “center” in order to univocally identify the position of the
parts in the existent, nor there is a way to understand a foundation which
explains being in its entirety, since entirety can never be comprehended and,
therefore, led back to one picture or principle.
2. The idea of infinite as a process which pushes itself beyond any given limit in
size or time. This is the case of the infinite divisibility of matter or of time.
Similarly to the preceding case, there is no way to identify a unity of measure
(such as integer numbers) through which to construct the entire reality: there
will always be a possible incommensurability.
In the last sections we will see that modern philosophy and mathematics -
because of the necessity of supposing both discrete unities of measure and a
continuum matter and irrational numbers - cannot overcome this conceptual
short circuit which leads to the paradox of the infinite (in which, because of the
recalled incommensurability, the infinite processes imply that the “part” is as
large as the “whole”). Idealistic philosophy, instead, disarms such a concept of
infinity by considering it as an intellectual trap and as an unimportant process
from a dialectical or “pragmatic” point of view. The issue becomes to create
a convention which puts in harmony all human experiences indifferently to a
possible infinite process.

2 The Infinite as an Ontological Problem

and the Limitless Size
The need to resolve contradictions in ancient philosophy is the one which is most
linked to the typically philosophical necessity of outlining the Totality, that is to
say of finding out the unique, original reason or ground (arche) which can explain
the existence and the behavior of “everything”. The occurrence of a contradiction
stays in the fact that if something remains unexplained and unconnected to the
rest of totality the very notion of knowledge of totality and episteme as pursued
by philosophy collapses (see, for instance, [14,17,50,54]). This can happen if not
everything can be reduced to one principle and we need at least two disconnected
360 F. Caldarola et al.

ones to make sense of phenomena, or if not all becoming can be explained with
one principle. Such an unexplained differences would be arbitrary gaps which
would deprive philosophy (or epistemology) of its ultimate goal: to trace the
rational origin of reality in order to inscribe all its aspects within “the sense of
the Whole”. This goal is not at all extraneous to the intrinsic nature of modern
science and physics - for instance, in its effort to locate elements which are more
and more fundamental and to ultimately unify all physical laws into one great
theory (theory of everything, see [21,23]).
Pythagoras was maybe the first philosopher mathematician who really had to
deal with the concept of infinite and with the disruption of the idea of a “struc-
tured whole” which its existence entails. Pythagorean mathematics is based on
the idea of “discontinuity”, as it is exclusively anchored in integer numbers and,
therefore, the increase of a magnitude proceeds by “discontinuous leaps”. In such
a worldview all objects were constituted by a finite number of monads, parti-
cles similar to atoms. Two magnitudes could be expressed by an integer num-
ber and were mutually commensurable, they admitted a common denominator.
Pythagorean thought will be put in crisis by the discovery of incommensurable
magnitudes (that is to say, which do not admit a common denominator), devel-
oped within the school itself as the relationship between diagonal and side of a
square resulted to be irrational, and safeguarded as an unspeakable secret. This
entailed that diagonal and side are composed not by a finite amount of points,
but by infinite points: for the first time actual infinite and not only potential
infinite was discussed.
Within this conceptual framework, the idea of infinite becomes an issue in
several senses, which are strongly interrelated. The simplest of these senses is
infinite as what pushes itself beyond any given limit in size or time. Can an
“infinite” size (or time) be considered consistent with the - metaphysical or
even physical - necessity of postulating a Totality which must be, at least in
principle, wholly grounded? Is it more consistent to suppose a finite universe,
with the danger of a possible call for “something which would always be beyond
its Whole”, or to envisage an infinite one, with the risk of the impossibility to
make sense of a Being which is never complete and, thus, never totally established
by one rational foundation? This last is the objection put forward by the Eleatic
Parmenides, for whom Being is eternal in the sense of being without past and
future, because it cannot come from not-being, but it is hypothesized as a finite
sphere - as opposed to the idea of his disciple Melissus. For Parmenides Being
is also immutable: becoming can never be rationally justified from the existence
of the necessary one original founding ground, because the alteration of this
“one” should be explained by another principle otherwise it would be arbitrary
and contradictory, and such a double principle should be explained by another
original principle and so on. Becoming must only be explained as an illusion. The
ultimate paradox of the infinite stays here. If you suppose its existence - in the
extension of time and space or in the divisibility of time and space - you cannot
hypothesize an original element which explains why reality organizes itself in a
certain way or in another. In fact, if you suppose an infinity divisibility of time
 Paradoxes of the Infinite Between Philosophy and Mathematics 361

and space you cannot locate or imagine such an original element, and if you
suppose an infinite extension of time or space you can never ensure the entire
consistency of time and space with the structure dictated by such an element.
On the other hand, if you suppose the non-existence of the infinite, it would be
impossible to explain why a Being stops being, or why a certain force stops acting
(in its capacity to divide matter or time). The latter case can only be explained
by the existence of two counterposed original elements, but their contraposition
would remain unexplained, betraying the very aim of episteme and science.

3 Infinite as Infinite Divisibility of a Certain Size

or Time and the Problem of Continuum

While Parmenides, maybe the first philosopher to clearly recognize the question
above, disarms the paradox of the infinite by excluding - as an “illusion” - the
possibility of time, becoming, alteration, infinity but even of something outside
the sphere of being (not preventing some logical inconsistency), Aristotle philos-
ophy tries to circumvents the paradox by accepting the duplicity of the original
element. In his case, they are “form” and “matter”. The “ground” which for
the Stagirite is at the basis of ontology, in fact, does not concern so much the
mere physical and material constituents of being as, rather, the formal principles
which imprint the underline indefinite matter and which build individuals and
characters of reality. First of all, in the Aristotelian ontology this formal prin-
ciple can stand “on its own” in the sense that a form, an identity of a primary
substance is complete and it does not draw its characters or existence from other
structures. Also, primary substances, which correspond to the individuals, are
not mere aggregates of their parts. To stand on its own, in this case, does not
mean that substance cannot change. Substance can be generated, altered and
corrupted by means of a “substantial transformation” (that is to say through
the process in which an individual assumes or loses its form on or from its mat-
ter - which is, in turn, an union of another form and another matter until the
indefinite primary matter in reached).
The transformations of substances are explained by the four interpretative
causes, material, efficient, formal and final one, with the final cause being the
crucially prevalent one in explaining the existence and change of things, as clar-
ified in Physics II and in Parts of Animals I. The causes, in fact, are ultimately
explained as an irresistible tension which the worldly matter has towards the
supreme immobile mover, a de facto divinity which has, for Aristotle, the same
qualities of the Eleatic Being and act as final cause of everything. Aristotle,
like Plato before him, solves the paradox of movement by assuming a necessary
“axiomatic” duplicity of first principle, in his case the supreme first mover and
indeterminate matter which continuously needs to be “informed”; a duplicity
which explains the relentless finalistic tension which corresponds to becoming.
Indeed, Aristotle’s metaphysics presents the interesting characteristic of inter-
preting all kinds of movements, included division and enlargement, as expres-
sion of a finalistic push which is itself the tension between the two “fundamental
362 F. Caldarola et al.

archai” of Being: in this way the contradiction of the existence of movement,

division or alteration which the necessity of a unique ground brings is steril-
ized. Further, one of these two “principles” seems to be, for the Stagirite, only a
principle in a negative sense, interpretable as an absence of a definite form (see
Physics, 204a8–204a16, and Metaphysics, IX.6, 1048a–b, in [4, Vol. I]).
Such a theoretical scenario entails one important thing: Aristotle cannot be
interested in infinite division of a magnitude (and in the infinite backward reduc-
tion it implies) as an action corresponding per se to an attempt to locate the
Origin. The prospect of an infinite reduction is theoretically unimportant because
the transformations in which its steps consist (as every kind of transformation)
would already be a reflection and a function of what is for him the arche, the
tension towards immobile mover. The ontological gaps (in the sense explained in
the previous section) which a similar series of actions manifests are already any-
way “endless”; in the sense that the inevitable existence of matter - the “second
principle” - makes the tension towards perfection endless. What really counts
from a scientific and an ontological point of view is, therefore, the comprehension
of the nature - that is to say of the causes - of the daily phenomena and their
ultimate direction. The existence of infinite division is therefore ontologically
relevant in the possible daily progression of its stages (infinite in potentiality),
not in its concept as “already actually present” (infinite in actuality), which
stops being idealized.
Such a rational solution is no longer possible in the philosophical era which
Severino characterizes for the dissolution of the cohesion of “certainty” and
“truth” (see [50, vol. II]). Modern philosophy, from Descartes to Kant included,
introduced the opposition between “certainty” and “truth”: the majority of
philosophers had to deal with the fact that what appears rational to human
thought may be at odds with what is outside it. Kant fulfills this attitude by
claim in that the chain of conditions, and also the world, are a subjective repre-
sentations and their epistemological value stays in how we manage - by means
of sensibility and understanding, the subjective categories - to build a consistent
picture of experience. Reason is still an arbiter of empirical truths, useful to
assess consistency and harmony of the set of concepts which our understand-
ing produces. But outside experience it does not make sense to rationally argue
for one solution or the other of the aporia earlier recalled, for finitude or for
infinity of original conditions and constituents: it is outside human reaching, for
instance, to verify whether a series of causes is actually infinite, such a totality
is never to be met with in experience, it’s only the “natural attitude” of human
reason to relentlessly chase this task (see [28, p. 460–461]). As a consequence, the
first two antinomies of pure reason involve the very issue of infinity in relation
to entirety: “Thesis 1: The world has a beginning in time, and is also limited as
regards space. Antithesis 1: The world has no beginning, and no limits in space;
it is infinite as regards both time and space. Thesis 2: Every composite substance
in the world is made up of simple parts, and nothing anywhere exists save the
simple or what is composed of the simple. Antithesis 2: No composite thing in the
 Paradoxes of the Infinite Between Philosophy and Mathematics 363

world is made up of simple parts, and there nowhere exists in the world anything
simple.” (See [34, p. 48] and also [5,6]).
This coincides to the fact that Kant’s philosophy cannot outflank the para-
doxes of the infinite or, better to say, he cannot propose a theory of knowledge
which metaphysically or even existentially systematizes the world. Once the fac-
ulties of reason are limited in their scrutiny of radical backward causes and onto-
logical elements by the inevitable gap between certainty and truth, one cannot
claim any definitive solution to potential contradictions. Theoretical reason and
categories cannot embrace the knowledge of the first cause and of the totality -
and, consequently, of the solutions to the paradoxes of the infinite.
Modern mathematics and modern philosophy seem to share the impossibil-
ity to rationally unarm the tension caused by the existence of something which
is incommensurable to the notion of entireness and complete systematicity. As
hinted before, if one supposes an infinity divisibility of time and space one cannot
locate or imagine an original element which structures the fabric of reality with
its regularities, and if one supposes an infinite extension of time or space it can
never been ensured the entire consistency of time and space with the structure
dictated by such an element. In mathematics such a problem corresponds with
that of the continuum. If empirical reality is rationally perceived as a “contin-
uum” both in its temporal or spatial extension and in its subdivision, in other
words, you can always pick and isolate a further fragment of magnitude which
disrupts what has been systematized so far in term of relations among different
magnitudes to picture the structure of reality, without rationality being legiti-
mate in restricting such a process with a justifications which has not empirical
bases. All this makes reciprocal commensurability among unities of measures and
imagines of completeness impossible. Hence the paradox of the infinite whereby
the part is as “large” as the entire. Mathematics and modern philosophy issues
may be synthesized in this way. The character of the incommensurable to entire-
ness and complete systematicity is brought, in mathematics, by the very necessity
of systematicity to compare and measure reciprocally incommensurable magni-
tudes and, in modern philosophy, by the necessity not to resort to ideal rational
hypotheses in the origin of Being to contain the logical infinite ontological pro-
cesses explained in this article.
Because of the existence of the continuum in mathematics and physics, you
cannot get rid of the “actual” infinite because you cannot get rid of the incom-
mensurable. Therefore you cannot do without supposing an “already happened
infinite approach” which converges to a measure, in order to “perfectly” measure
things - as opposed to Aristotle. Limits, or better series, are a way to explicit that
entire, structurally finite magnitude to which the sum of some infinite fragments
is convergent.

4 Infinite as a “Convention” and the Conventional

Resolutions to the Contradiction
Infinite has also been thought as resolution to the intellectual and existential dis-
comfort which derives from the finitude of things within becoming. This position
364 F. Caldarola et al.

has been adopted, in the history of philosophy, first by pre-Socractic thinkers

who did not have yet a clear account of the paradox of the infinite as described
above and, secondly, by idealistic philosophers such as Hegel who decided that
the only way to circumvent such a paradox is to resort to the acceptation of
“conventions” as ontologically valid. In other words, infinite is the character
of the totality of being once we have given a reason to the chaotic turning up
of differences, to the apparent contradictions and all kinds of obstacles which
make our experience of the world appear full of uncomfortable discrepancies.
A similar rationalization aims at finding the intrinsic harmony of things, so to
show the ground, the origin (arche) which accounts for their coming into beings
and, therefore, describes the Totality as the very Logic of their opposition. A
Logic which, taken “in itself”, is not limited to one contingent thing, it does
not find arbitrary discrepancies: it is unlimited, like a circle. The philosopher
who manages to be aware of that is able to experience the non-contradiction of
existence. The rational unity of all things is infinite, therefore, in the sense that
it never presents a particular which is impossible to make sense of, and therefore
it cannot contemplate a feeling of limit or restriction due to a contradiction: it
is always already “full of Being”. Such a conception of infinite and arche first
appears in Greek philosophy in Anaximander and his idea of apeiron (see [51]),
the original identity of all things which is without peculiar determinations and,
thus, “without boundaries”. It was famously elaborated by Heraclitus and his
emphasis on the strive of oppositions as the logic principle of reality (see [27,29])
and it will be, as we will see, a crucial feature of Hegel’s idealism.
German idealism - and its major representative, Hegel - introduces an inno-
vative insight to modern philosophy: “construction” is real, what seems to be
the fruit of human mind or human culture is the expression, the reflection or,
even, an active part of the rational structure of reality and of its logic. From this
point of view it does not make any sense to distinguish between “empirical” and
“metaphysical” construction as everything draws its “epistemo-logical dignity”
from its being a logical or “practical” passage which contributes to harmonize
and rationalize experience. With Hegel we meet the conceptions of the infinite
as “rationalization” which aims at finding the intrinsic harmony of things, so
to show the origin which accounts for their coming into beings and, therefore,
describes the Totality as the very Logic of their opposition. Being infinite, in this
sense, is the character of the idealistic Absolute and its ontological sense is the
very opposite of the infinite as a limitless addition of finite aspects of reality. To
continue adding an element to a concept or to a magnitude without elaborating
a rational synthesis of the sense of its coming into being or of its telos - and the
other parts of reality means to maintain a certain “arbitrary” discrepancy. The
infinite series of this concept is what Hegel calls “bad infinity” (see [24]).
The contradiction between the necessity to find a first condition to make
our conception of the world “complete” on the one hand and the necessity not
to “arbitrarily stop” without reason to a certain element (proper of a world
which is continuum) on the other hand should be synthesized by the idea of
the universe as eternal (without time) and “circular” as rationally explained
 Paradoxes of the Infinite Between Philosophy and Mathematics 365

in itself, as the existence of a feature is always explained as the synthesis of a

previous apparent discrepancy, without the need to resort to a condition which is
out of the already “existent” dialectic of beings. If a thought manages to invent
an instrument which makes sense of a process, harmonize its meaning with the
surroundings without resorting to virtually infinite actions, such a thought is real
and rational. Even if the cosmological problem is not central to Hegel’s reflection,
his dialectics brings to the non-necessity of an infinite research of spatial and
temporal limits or conditions. As he says in the Science of Logic, “the image of
true infinity, bent back onto itself, becomes the circle” (see [24, p. 149, §302]).
The most original and pregnant insight of Hegel’s rationalization of reality
is that it does not matter, for dialectical speculation, how many “finite” and
irrational characters are present in a certain context, or how deep a contradiction
is structured in a concept. Since the aim of philosophy - and, therefore, of human
existence - is to make thought and Being coincide (see [15]) and since thought and
Being are not two separate entities, any practical convention which fulfills the
task of making such irrational characters unimportant and of making our acting
and experience “satisfactory” has the dignity of a real character of reality. One
way to “synthesize” the necessity of modern mathematics calculus of resorting
to the concept of incommensurability between magnitudes - and, thus, of limits
and “actual” infinite - and the necessity of logic and common sense to avoid
the relative paradoxes is the concept of grossone, proposed by Sergeyev. The
grossone appears to be what has just be named as a conventional systematization
which interprets the infinite as an Entire in order to outflank its contradictions,
and which consciously overlooks the fact of its being constituted by “infinite
finitudes” because practically unimportant - and for this reason “irrational” in
a dialectical sense - in the context of its applications.
Sergeyev starts from a literary interpretation of the principle “the part is
less than the whole” and he applies it both to finite and, especially, infinite
quantities, in strong contrast with the conception of Cantor’s infinite, for which
the infinite processes imply that the “part” is as large as the “whole”. As we
have seen, Idealistic philosophy too, agrees with such a principle that leads to
a “multiple concept of infinity”, that results also in accord with the dialecti-
cal or “pragmatic” point of view. At the University of Calabria, in the early
2000, Sergeyev began to develop his idea of an infinite with good computational
properties and, at the same time, easy to use at every level, from schools to
academic research. In a few words, Sergeyev developed a new computational
system based on two fundamental units or “atoms”: the ordinary unit 1 to gen-
erate finite quantities as the familiar natural or integer numbers, and a new
infinite unit, named “grossone”, to generate infinite and infinitesimal numbers.
For example, [38,40,43,46,48] are introductory surveys to this new system and
the book [39] is also written in a popular way and well understandable for high
school students. This new computational methodology has even been applied
in a number of areas both in mathematics, applied sciences, philosophy, and
recently also in mathematical education. For instance, optimization and numer-
ical differentiation [1,16,20,33,42,44,49], calculus of probabilities and geometry
366 F. Caldarola et al.

[13,31,32], fractals and complex systems [3,8,9,12,18,19,41,45,47], paradoxes

and supertasks [35,37], didactic experimentations [2,26,36], and others.
The peculiar main characteristic of the grossone, that leads it to a growing
diffusion, is probably his double nature of infinite number, but with the behavior
of an ordinary natural one. This intrinsic characteristic is in fact the basis of
several of its properties among which we recall, since this article deals with
paradoxes from different points of view, the ability to solve some of them focused
on the infinite. For example, consider what is probably the most famous of them,
the Hilbert’s paradox of the Grand Hotel, which was designed to highlight the
strange properties of the infinite countable cardinality ℵ0 in the Cantorian sense.
Even when the hotel is full, it is always possible to host a finite or infinite
countable number of new guests by moving the present guests in the hotel in
a suitable manner.1 This leads to many semantic problems at various levels, as
well as questions of a logical nature. To take a simple example, it is sufficient just
to consider the semantics of the word “full”: commonly, if a tank or container is
full it is not possible to add more. Adopting the grossone system instead, such
kinds of paradoxes are avoided because a hotel with ① rooms cannot accept
other guests when it is full (see for instance [39, Section 3.4]).
Another example, of different nature, that yields a paradox is the following,
known as the Thompson lamp paradox.

Example 1 (The Thompson lamp paradox). Given a lamp, assume that we turn
it on at time zero and turn it off after 1/2 minute, then we turn it on after 1/4
minute and turn it back off again after 1/8 minute, and so on, by acting on the
switch to each successive power of 1/2. At the end of a minute, will the lamp be
on or off? This puzzle was originally proposed by the philosopher J.F. Thompson
in 1954 to analyze the possibility of completing a supertask, i.e. a larger task
made up of an infinite number of simple tasks (see [52]).
Note that the classical geometric series of common ratio 1/2 and starting
from 1/2 converges to 1, in symbols
+∞  t
 1
= 1, (1)
t=1
2

hence in a minute the switch will be moved infinitely many times (admitting that
this is possible) and the question above has no answer. Using the grossone-based
system, it is offered in [39, Section 3.4] the solution that the lamp is “off” after
one minute, and it is essentially due to the parity of grossone.
We propose here a more detailed interpretation: the switch will have the first
 2
motion at the time zero, the second at the time 12 , the third at the time 12 + 12 ,
 2  n−1
and the n-th at the time 12 + 12 + . . . + 12 . We recall that a sequential

1
If there are n new guests the simplest choice is to use the function IN → IN, m → m+
n, instead, in case of an infinite countable number, the function IN → IN, m → 2m
(see [22, 39]).
 Paradoxes of the Infinite Between Philosophy and Mathematics 367

process, in Sergeyev’s theory, cannot have more than ① steps, hence the last
action on the switch, i.e. the ①-th action, will be made at the time

− 1  t
① − 1  t
①
1 1
= −1 +
t=1
2 t=0
2
 1 ①  ① − 1
1− 1
= −1 + 1
2
=1−
2
2

(cf. Eq. (1)). In other words this means that the last action on the switch will
be made at an infinitesimal time before the end of the minute, more precisely
 1 ① − 1
2 minute before, and from this moment on, the switch will remain off
(hence it will be off as well at the time 1 min).

Sergeyev, by means of an “infinite” which behaves as an ordinary natural

finite number, disarms the anarchy of the continuum. Such an “anarchy” springs
from the absence of a limit in a process of division – of space or time - or in the
growth of a size, absence which is the reason why you cannot locate an original
element which can justify a univocal system of references capable of ultimately
describing the features of reality or – which is the same thing – of “creating a
stable totality”. This infinite process of addition of “fragments of finitude” is
what create the inconvenience whereby the “part is as big as the entire” and the
consequent impossibility to systematize the entire. To sterilize such a process
means to conventionally set aside the traditional and problematic sense of the
infinite to embrace what idealistic philosophy and Hegel refers to as “true infi-
nite” (see [24]), that is to say a practical convention which makes thought and
experience more fluid and comforting (without ontological “finitudes”). “Infi-
nite” is here used, of course, in a completely unusual acceptation if compared
to its common mathematical meaning. The function of grossone is therefore the
same as a dialectical (in a philosophical sense) device which ensures that our
approach to a certain concept is deprived of discrepancies: we perceive it with-
out ontological “limits” in the sense that we do not find compromising alteration
or obstacle in its logic.
Unimaginable numbers2 seem to respond to the same pragmatic and, there-
fore, ontological necessity: even if they does not contemplate actually “infinite”
quantities, they give “form” to numbers which are as large as to question their
commensurability to a rational systematization of our picture of reality, gener-
ating the same discomfort of “relentless process” as an infinite process does. But
from a closer point of view, these notions can assume a more specific ontological

2
The unimaginable numbers are numbers extremely large so that they cannot be
written through the common scientific notation (also using towers of exponents) and
are behind every power of imagination. To write them some special notations have
been developed, the most known of them is Knuth’s up-arrow notation (see [30]). A
brief introduction to these numbers can be found in [10], while more information is
contained in [7, 11, 25].
368 F. Caldarola et al.

connotation, which communicate with Aristotle’s approach as we have already

hinted.
We have seen how Aristotle is not interested in infinite division of a magnitude
(and in the infinite backward reduction it implies) as an action corresponding
per se to an attempt to locate the Origin. The prospect of an infinite reduction
is theoretically unimportant for him because the transformations in which its
steps consist (as every kind of transformation for Aristotle’s philosophy) would
already be a function of what is the arche, the tension towards immobile mover,
the de facto divinity external to the empirical perimeter, at the origin of all
causes.
Aristotle’s philosophy can afford to outflank the issues about the infinite
because to conceive it would be impossible and redundant since this very idea is
incommensurable to the idea of first cause and first principle, whose action will
always be “in progress”. The infinite is always only potential, never actual.
The rejection of an ontological hypostatization of the concept of the infinite is
also confirmed in its relation to the notion of substance. The infinite is not a for-
mal principle which “is” in the primary sense of the term, it is not a substance
whose structure cannot be divided because if so it would lose its unique and
original identity: it is an accident which can be “divided” and may potentially
happen and be altered by applying to completely different circumstances: “it is
impossible that the infinite should be a thing which is in itself infinite, separable
from sensible objects. If the infinite is neither a magnitude nor an aggregate, but
is itself a substance and not an accident, it will be indivisible; for the divisible
must be either a magnitude or an aggregate. But if indivisible, then not infinite,
except in the way in which the voice is invisible. But this is not the way in which
it is used by those who say that the infinite exists, nor that in which we are
investigating it, namely as that which cannot be gone through.”3 The infinite
is reduced to what we would call today “an idea in a Kantian sense” - with
the difference that, as said, its paradoxes does not even have a crucial role in
the determination of ultimate metaphysical truths: “the infinite does not exist
potentially in the sense that it will ever actually have separate existence; it exists
potentially only for knowledge. For the fact that the process of dividing never
comes to an end ensures that this activity exists potentially, but not that the
infinite exists separately.”4 Similarly, the unimaginable numbers can be inter-
preted as an idea in “a Kantian sense”, having only a regulatory function in the
sense that they own a linguistic and rational denotation and can give a sensible,
“reasonable” context to the regularities which we ordinarily experiences, but
that will never actually exist because no thoughts or no computing machine will
never be able to “see” (or even think) them. Obviously, in the logical context of
modern mathematics, the idea of original immobile mover is note relevant as the
concept of “natural number”, which share with the first one a similar regulatory
theoretical (and, therefore, pragmatic) “action”.

3
See Physics, 204a8–204a16, in [4, Vol. I].
4
See Metaphysics, IX.6, 1048a-b, in [4, Vol. I].
 Paradoxes of the Infinite Between Philosophy and Mathematics 369

5 Conclusions
The sense of the introduction of grossone into a mathematical formula cannot be
reduce to a “technical stratagem”, and to understand it is necessary to under-
stand the ontological relevance of the problem of the infinite in the context of a
consistent picture of the world which any discipline try to achieve. In this paper
we have seen how the original philosophical effort to comprehend the entireness
of reality, both in the sense of its fundamental elements which justify its struc-
ture and in the sense of the “total amount” of its features, inevitably clashes
with the problem that the process of locating these elements or this totality can
be thought as without an end. This problem produces the paradox whereby a
“part” of reality can contain as many amounts of hypothetical “fundamental
elements” or features as the entire. This is an issue which belongs to ontology in
a broad sense as to mathematics and its elements in a more specific acceptation,
since mathematics - in order to work and to be applied to “experience” - has to
create its specific reality consistent in its entireness and in its axioms. One of the
solutions to this philosophical paradox elaborated by modern philosophy is to
take this problem as a very “problem of thought”, making reality coincide with
our very conception of reality and with our practical and cultural instruments.
In this sense, for Hegel’s philosophy, reality is dialectical not “for our thought”:
reality coincides with our thought and vice versa, and any ontological problem is
reduced to a problem of internal consistency and contradiction. But if thought
and reality are two sides of the same coin, then to create a logic which allows our
thought to work with the “reality” of a discipline without any logical arrest or
short circuit which would generate practical issues is to create a new dialectical
advance in reality - or, at least, in the reality internal to our discipline. The
innovation of grossone is that it seems to engage with such a task in the con-
text of the paradoxes of the infinite, by proposing an explicitly “conventional”
notion which challenges the very necessity of the ontological mathematical tur-
moil experienced so far.

Aknowledgments. This work is partially supported by the research projects

“IoT&B, Internet of Things and Blockchain”, CUP J48C17000230006, POR Calabria
FESR-FSE 2014–2020.

References
1. Amodio, P., Iavernaro, F., Mazzia, F., Mukhametzhanov, M.S., Sergeyev, Y.D.: A
generalized Taylor method of order three for the solution of initial value problems in
standard and infinity floating-point arithmetic. Math. Comput. Simul. 141, 24–39
(2017)
2. Antoniotti, A., Caldarola, F., d’Atri, G., Pellegrini, M.: New approaches to basic
calculus: an experimentation via numerical computation. In: Sergeyev, Y.D.,
Kvasov, D.E. (eds.) NUMTA 2019. LNCS 11973, pp. 329–342. Springer, Cham
(2020)
3. Antoniotti, L., Caldarola, F., Maiolo, M.: Infinite numerical computing applied to
Hilbert’s, Peano’s, and Moore’s curves. Mediterr. J. Math. (to appear)
370 F. Caldarola et al.

4. Barnes, J. (ed.): The Complete Works of Aristotle. The Revised Oxford Transla-
tion, vol. I and II. Princeton University Press, Princeton (1991). 4th printing
5. Bennet, J.: Kant’s Dialectic. Cambridge University Press, Cambridge (1974)
6. Bird, G.: Kant’s Theory of Knowledge: An Outline of One Central Argument in
the Critique of Pure Reason. Routledge & Kegan Paul, London (1962)
7. Blakley, G.R., Borosh, I.: Knuth’s iterated powers. Adv. Math. 34(2), 109–136
(1979). https://doi.org/10.1016/0001-8708(79)90052-5
8. Caldarola, F.: The Sierpiński curve viewed by numerical computations with infini-
ties and infinitesimals. Appl. Math. Comput. 318, 321–328 (2018). https://doi.
org/10.1016/j.amc.2017.06.024
9. Caldarola, F.: The exact measures of the Sierpiński d-dimensional tetrahedron in
connection with a Diophantine nonlinear system. Commun. Nonlinear Sci. Numer.
Simul. 63, 228–238 (2018). https://doi.org/10.1016/j.cnsns.2018.02.026
10. Caldarola, F., d’Atri, G., Maiolo, M.: What are the unimaginable numbers? Sub-
mitted for publication
11. Caldarola, F., d’Atri, G., Mercuri, P., Talamanca, V.: On the arithmetic of Knuth’s
powers and some computational results about their density. In: Sergeyev, Y.D.,
Kvasov, D.E. (eds.) NUMTA 2019. LNCS 11973, pp. 381–388. Springer, Cham
(2020)
12. Caldarola, F., Maiolo, M., Solferino, V.: A new approach to the Z-transform
through infinite computation. Commun. Nonlinear Sci. Numer. Simul. 82, 105019
(2020). https://doi.org/10.1016/j.cnsns.2019.105019
13. Calude, C.S., Dumitrescu, M.: Infinitesimal probabilities based on grossone.
Spinger Nat. Comput. Sci. 1, 36 (2019). https://doi.org/10.1007/s42979-019-0042-
8
14. Cavini, W.: Ancient epistemology naturalized. In: Gerson, L.P. (ed.) Ancient Epis-
temology. Cambridge University Press, Cambridge (2009)
15. Cesa, C.: Guida a Hegel. Laterza, Bari (1997)
16. Cococcioni, M., Pappalardo, M., Sergeyev, Y.D.: Lexicographic multi-objective lin-
ear programming using grossone methodology: theory and algorithm. Appl. Math.
Comput. 318, 298–311 (2018)
17. Crivelli, P.: Aristotle on Truth. Cambridge University Press, Cambridge (2004)
18. D’Alotto, L.: A classification of one-dimensional cellular automata using infinite
computations. Appl. Math. Comput. 255, 15–24 (2015)
19. D’Alotto, L.: Cellular automata using infinite computations. Appl. Math. Comput.
218(16), 8077–8082 (2012)
20. De Leone, R.: The use of grossone in mathematical programming and operations
research. Appl. Math. Comput. 218(16), 8029–8038 (2012)
21. David Peat, F.: Superstrings and the Search for the Theory of Everything. Con-
temporary Books, Chicago (1988)
22. Faticoni, T.G.: The Mathematics of Infinity: A Guide to Great Ideas. Wiley, Hobo-
ken (2006)
23. Gribbin, J.: The Search for Superstrings, Symmetry, and the Theory of Everything.
Back Bay Books, New York (2000)
24. Hegel, G.W.F.: The Science of Logic. Cambridge University Press, Cambridge
(2010 [1817]). Ed. by G. Di Giovanni
25. Hooshmand, M.H.: Ultra power and ultra exponential functions. Integral
Transforms Spec. Funct. 17(8), 549–558 (2006). https://doi.org/10.1080/
10652460500422247
 Paradoxes of the Infinite Between Philosophy and Mathematics 371

26. Ingarozza, F., Adamo, M.T., Martino, M., Piscitelli, A.: A grossone-based numer-
ical model for computations with infinity: a case study in an Italian high school.
In: Sergeyev, Y.D., Kvasov, D.E. (eds.) NUMTA 2019. LNCS 11973, pp. 451–462.
Springer, Cham (2020)
27. Kahn, C.H.: The Art and Thought of Heraclitus. Cambridge University Press,
Cambridge (1979)
28. Kant, E.: Critique of Pure Reason. Cambridge University Press, Cambridge (1998
[1791]). Transl. by P. Guyer, A.W. Wood (eds.)
29. Kirk, G.S.: Heraclitus: The Cosmic Fragments. Cambridge University Press, Cam-
bridge (1978)
30. Knuth, D.E.: Mathematics and computer science: coping with finiteness. Science
194(4271), 1235–1242 (1976). https://doi.org/10.1126/science.194.4271.1235
31. Margenstern, M.: An application of grossone to the study of a family of tilings of
the hyperbolic plane. Appl. Math. Comput. 218(16), 8005–8018 (2012)
32. Margenstern, M.: Fibonacci words, hyperbolic tilings and grossone. Commun. Non-
linear Sci. Numer. Simul. 21(1–3), 3–11 (2015)
33. Mazzia, F., Sergeyev, Y.D., Iavernaro, F., Amodio, P., Mukhametzhanov, M.S.:
Numerical methods for solving ODEs on the Infinity Compute. In: Sergeyev, Y.D.,
Kvasov, D.E., Dell’Accio, F., Mukhametzhanov, M.S. (eds.) 2nd International Con-
ference “NUMTA 2016 - Numerical Computations: Theory and Algorithms”, AIP
Conference Proceedings, vol. 1776, p. 090033. AIP Publishing, New York (2016).
https://doi.org/10.1063/1.4965397
34. Nicolau, M.F.A., Filho, J.E.L.: The Hegelian critique of Kantian antinomies: an
analysis based on the Wissenchaft der Logik. Int. J. Philos. 1(3), 47–50 (2013)
35. Rizza, D.: A study of mathematical determination through Bertrand’s Paradox.
Philos. Math. 26(3), 375–395 (2018)
36. Rizza, D.: Primi passi nell’aritmetica dell’infinito (2019, preprint)
37. Rizza, D.: Supertasks and numeral system. In: Sergeyev, Y.D., Kvasov, D.E.,
Dell’Accio, F., Mukhametzhanov, M.S. (eds.) 2nd International Conference
“NUMTA 2016 - Numerical Computations: Theory and Algorithms”, AIP Confer-
ence Proceedings, vol. 1776, p. 090005. AIP Publishing, New York (2016). https://
doi.org/10.1063/1.4965369
38. Sergeyev, Y.D.: A new applied approach for executing computations with infinite
and infinitesimal quantities. Informatica 19(4), 567–596 (2008)
39. Sergeyev, Y.D.: Arithmetic of infinity. Edizioni Orizzonti Meridionali, Cosenza
(2003)
40. Sergeyev, Y.D.: Computations with grossone-based infinities. In: Calude, C.S., Din-
neen, M.J. (eds.) UCNC 2015. LNCS, vol. 9252, pp. 89–106. Springer, Cham (2015).
https://doi.org/10.1007/978-3-319-21819-9 6
41. Sergeyev, Y.D.: Evaluating the exact infinitesimal values of area of Sierpinskinski’s
carpet and volume of Menger’s sponge. Chaos Solitons Fractals 42(5), 3042–3046
(2009)
42. Sergeyev, Y.D.: Higher order numerical differentiation on the Infinity Computer.
Optim. Lett. 5(4), 575–585 (2011)
43. Sergeyev, Y.D.: Lagrange lecture: methodology of numerical computations with
infinities and infinitesimals. Rend Semin Matematico Univ Polit Torino 68(2), 95–
113 (2010)
44. Sergeyev, Y.D.: Solving ordinary differential equations by working with infinitesi-
mals numerically on the infinity computer. Appl. Math. Comput. 219(22), 10668–
10681 (2013)
372 F. Caldarola et al.

45. Sergeyev, Y.D.: The exact (up to infinitesimals) infinite perimeter of the Koch
snowflake and its finite area. Commun. Nonlinear Sci. Numer. Simul. 31, 21–29
(2016)
46. Sergeyev, Y.D.: Un semplice modo per trattare le grandezze infinite ed infinitesime.
Mat. Soc. Cult. Riv. Unione Mat. Ital. 8(1), 111–147 (2015)
47. Sergeyev, Y.D.: Using blinking fractals for mathematical modelling of processes of
growth in biological systems. Informatica 22(4), 559–576 (2011)
48. Sergeyev, Y.D.: Numerical infinities and infinitesimals: methodology, applications,
and repercussions on two Hilbert problems. EMS Surv. Math. Sci. 4(2), 219–320
(2017)
49. Sergeyev, Y.D., Mukhametzhanov, M.S., Mazzia, F., Iavernaro, F., Amodio, P.:
Numerical methods for solving initial value problems on the Infinity Computer.
Int. J. Unconv. Comput. 12(1), 55–66 (2016)
50. Severino, E.: La Filosofia dai Greci al nostro Tempo, vol. I, II, III. RCS Libri,
Milano (2004)
51. Theodossiou, E., Mantarakis, P., Dimitrijevic, M.S., Manimanis, V.N., Danezis,
E.: From the infinity (apeiron) of Anaximander in ancient Greece to the theory of
infinite universes in modern cosmology. Astron. Astrophys. Trans. 27(1), 162–176
(2011)
52. Thompson, J.F.: Tasks and super-tasks. Analysis 15(1), 1–13 (1954)
53. Weyl, H.: Levels of Infinity/Selected Writings on Mathematics and Philosophy.
Dover (2012). Ed. by P. Pesic
54. Zanatta, M.: Profilo Storico della Filosofia Antica. Rubettino, Catanzaro (1997)
 The Sequence of Carboncettus Octagons

Fabio Caldarola1(B) , Gianfranco d’Atri1 , Mario Maiolo2 ,

and Giuseppe Pirillo3
1
Department of Mathematics and Computer Science, Università della Calabria,
Cubo 31/A, 87036 Arcavacata di Rende, CS, Italy
{caldarola,datri}@mat.unical.it
2
Department of Environmental and Chemical Engineering (DIATIC),
Università della Calabria, Cubo 42/B, 87036 Arcavacata di Rende, CS, Italy
[email protected]
3
Department of Mathematics and Computer Science U. Dini, University of Florence,
viale Morgagni 67/A, 50134 Firenze, Italy
[email protected]

Abstract. Considering the classic Fibonacci sequence, we present in

this paper a geometric sequence attached to it, where the word “geo-
metric” must be understood in a literal sense: for every Fibonacci num-
ber Fn we will in fact construct an octagon Cn that we will call the
 Carboncettus octagon, and in this way we obtain a new sequence
n-th
Cn n consisting not of numbers but of geometric objects. The idea of
this sequence draws inspiration from far away, and in particular from a
portal visible today in the Cathedral of Prato, supposed work of Carbon-
cettus marmorarius, and even dating back to the century before that of
the writing of the Liber Abaci by Leonardo Pisano called Fibonacci (AD
1202). It is also very important to note that, if other future evidences will
be found in support to the historical effectiveness of a Carboncettus-like
construction, this would mean that Fibonacci numbers were known   and
used well before 1202. After the presentation of the sequence Cn n , we
will give some numerical examples about the metric characteristics of the
first few Carboncettus octagons, and we will also begin to discuss some
general and peculiar properties of the new sequence.

Keywords: Fibonacci numbers · Golden ratio · Irrational numbers ·

Isogonal polygons · Plane geometric constructions

1 Introduction
The names here proposed of “n-th Carboncettus octagon” and “Carboncettus
sequence/family of octagons”, or better, the inspiration for these names, comes
from far away, sinking its roots in the early centuries of the late Middle
Ages. They are in fact connected to the cathedral of Prato, a jewel of Ital-
ian Romanesque architecture, which underwent a profound restructuring in the
11th century, followed by many others afterwards. The side portal shown in
c Springer Nature Switzerland AG 2020
Y. D. Sergeyev and D. E. Kvasov (Eds.): NUMTA 2019, LNCS 11973, pp. 373–380, 2020.
https://doi.org/10.1007/978-3-030-39081-5_32
374 F. Caldarola et al.

Fig. 1 (which we will later call simply the portal) at the time of its construction
seems to have been the main portal of the cathedral. The marble inlays on its
sides and the figures represented have aroused many discussions among scholars
for many years and in particular have always aroused the attention and interest
of G. Pirillo, an interest that he recently transmitted also to the other authors.
Pirillo studied the figures of the portal for a long time and traced a fascinating
symbolism, typical of medieval culture (see for example [11]). According to these
studies, the right part of the portal, for instance, through a series of very regular
and symmetrical figures, would recall the divine perfection, while the left part,
through figures that approximate the regular ones but are not themselves regu-
lar, the imperfection and the limits of human nature. The very interesting fact
is that the artist/architect who created the work (which is thought to be a cer-
tain Carboncettus Marmoriarius, very active at that time and in those places,
[11]) seems to have been in part used the mathematical language to express
these concepts and ideas, and this thing, if confirmed, would assume enormous
importance, because before the 12th century we (and many experts of the field)
have no knowledge of similar examples. The construction of the Carboncettus
octagon (or better, of the Carboncettus octagons, since they are infinitely many)
originates from Fibonacci numbers and yields a sequence not of numbers but of
geometrical figures: we will explain the details starting from Sect. 2.
From the historical point of view we cannot avoid to note an interesting,
particular coincidence: probably, the most known and most important octagonal
monument existing in Calabria dates back to the same period as the construction
of the portal of the Duomo of Prato, and it is the octagonal tower of the Norman-
Swabian Castle in Cosenza. But it is important to specify, for the benefit of the
reader, that, in Cosenza, on the site of the actual Norman-Swabian Castle, a
fortification had existed from immemorial time, which underwent considerable
changes over the years: first a Bruttuan fortress, then Roman, Norman and
Swabian, when it had the most important restructuring due to Frederick II
of Swabia. In particular, it is Frederick who wanted the octagonal tower visible
today, his preferred geometric shape: remember, for example, the octagonal plan
of the famous Castel del Monte near Bari, in Apulia.
With regard to Fibonacci numbers, we would like to point out to the reader
for completeness of information, a recent thesis by G. Pirillo often and many
times discussed within this group of authors. In [10,12–14] Pirillo presented the
audacious thesis that the first mathematicians who discovered Fibonacci num-
bers were some members of the Pythagorean School, well documented and active
in Crotone in the 6th, 5th and 4th centuries B.C., hence about 1,700 years before
that Leonardo Pisano, known as “Fibonacci”, wrote his famous Liber Abaci in
1202. Such a thesis is mainly supported by computational evidences arising
from pentagon and pentagram about the well-known Pythagorean discovery of
the existence of incommensurable numbers. The interested reader can find fur-
ther information and references on the Pythagorean School, incommensurable
lengths, Fibonacci numbers and some recent developments in [6,8,10,14–17].
 The Sequence of Carboncettus Octagons 375

Fig. 1. The side portal of the cathedral of Prato. The two topmost figures have octag-
onal shape: the one on the right is based on a regular octagon, while the one on the
left seems to allude to a very particular construction that inspires thus paper and the
now called Carboncettus octagons.
376 F. Caldarola et al.

Similarly to the above thesis note that, since the portal in Prato is dating
back to the 12th century, if other future evidences will support the employ of
Fibonacci numbers in its geometries, this would mean that they were known
before 1202 as well, even if only a few decades.
A final remark on notations: we denote by N the set of positive integers and
by N0 the set N ∪{0}. A sequence
 of numbers or other mathematical objects
is denoted by an n∈N , an n , or simply {an }. If, moreover, A, B, C are three
points of the plane, AB denotes the line segment with endpoints A and B, |AB|
its length, and ABC the measure of the angle with vertex in B.

2 The Carboncettus Family of Octagons

If r is any positive real number, we denote by Γr the circumference of radius
r centered in the origin. As usual, let Fn be the n-th Fibonacci number for all
n ∈ N0 , i.e.,

F0 = 0, F1 = 1, F2 = 1, F3 = 2, F4 = 3, F5 = 5, etc.

If n ∈ N we consider a couple of concentric circumferences having radii of

length Fn and Fn+2 , respectively. If n = 1 they are represented in green in
Fig. 2, were the radius of the inner circumference is 1 and that of the outer
one is 2, i.e. F3 . Then we draw two couples of parallel tangents, orthogonal
between them, to the inner circumference and we consider the eight intersec-
tion points A, B, C, D, E, F, G, L with the outer circumference ΓFn+2 , as in
Fig. 2. The octagon obtained by drawing the polygonal through the points
A, B, C, D, E, F, G, L, A, in red in Fig. 2, is called the n-th Carboncettus octagon
and is denoted by Cn . Therefore, the red octagon in Fig. 2, is the first Carbon-
cettus octagon C1 .
From a geometrical point of view, the Carboncettus octagon Cn is more
than a cyclic polygon; it is in fact an isogonal octagon for all n ∈ N, that is, an
equiangular octagon with two alternating edge lengths.1 More recently it is also
used to say a vertex-transitive octagon: all the vertices are equivalent under the
symmetry group of the figure and, in the case of Cn , for every couple of vertices,
the symmetry which send the first in the second is unique. The symmetry group
  group D4 .
of Cn is in fact isomorphic to the one of the square, the dihedral 2

An interesting property of the Carboncettus sequence Cn n∈N is the fact

that, with the exception of the first three elements C1 , C2 , C3 (or, at most,
also C4 ), all the subsequent ones are completely indistinguishable from a regular
octagon (see, for example, Fig. 3 representing C2 : it is yet relatively close to a
regular octagon). Due to the lack of space, we will deepen these and other impor-
tant aspects mentioned in the following, in a subsequent paper in preparation.
1
In this view, a recent result established that a cyclic polygon is equiangular if and
only if is isogonal (see [7]). Of course, an equiangular octagon is not cyclic in general,
while it is true for 3- and 4-gons (see [2]).
2
Note, for didactic purposes, how the multiplication table of D4 emerges much more
clearly to the mind of a student thinking to C1 than thinking to a square.
 The Sequence of Carboncettus Octagons 377

Fig. 2. The construction of the Carboncettus octagon. In the picture, in particular, it

is shown in red the octagon C1 . (Color figure online)

3 The First Four Octagons of the Carboncettus Sequence:

Geometric Properties and Metric Data
In this section we will give
 some
 numerical examples looking closely at the first
elements of the sequence Cn n∈N .

Example 1 (The octagon C1 ). The first Carboncettus octagon C1 is built starting

from the circumferences Γ1 and Γ2 , as said in Sect. 2. In this case we obtain a
very particular isogonal octagon: drawing the eight radii

OA, OB, OC, OD, OE, OF , OG, OL (1)

of the circumference Γ2 as in Fig. 2, the resulting shape has commensurable angle

measures, in fact all them are integer multiples of π/12 = 15◦ . Not only; in this
way C1 results formed by 4 equilateral triangles (congruent to ABO, see Fig. 2)
and 4 isosceles triangles (congruent to BCO). The lengths of their sides and
heights are
√
|AB| = |OA| = 2, |OH| = |KC| = 3,
√ √ √ √ (2)
|BC| = 6 − 2, |ON | = 6+ 2
2
,

which, for example, are all incommensurable in pairs. Instead, for the widths of
the angles we trivially have

AOB = OBA = π/3 = 60◦ , BOC = HOB = π/6 = 30◦ ,

(3)
OBC = 5π/12 = 75◦ .
378 F. Caldarola et al.
 
Discussing the commensurability of the angles for all the sequence Cn n is
interesting, but we are forced to postpone this elsewhere. The same, as well,
considering the commensurability, along all the sequence, of some of the side
lengths made explicit in (2). Note lastly that perimeter and area are
√ √ √
Per(C1 ) = 8 + 4 6 − 4 2, Area(C1 ) = 4 + 4 3.

The second Carboncettus octagon C2 originates from the circumferences Γ1

and Γ3 , with radii F2 = 1 and F4 = 3, respectively, and the result is the black
octagon in Fig. 3, compared with a red regular one inscribed in the circumference
Γ3 itself. Using the letters disposition of Fig. 2, the lengths of the correspondent
sides and heights considered in (2), the angle widths, perimeter and area, are
those listed in the second column of Table 1.

Table 1. Some metric data relative to the first three elements of the Carboncettus
sequence, after C1 . The letters are displayed in the construction as in Fig. 2.

C2 C3 C4
|OK| 1 2 3
|OA| 3 5 8
|AB| 2 4 6
√ √ √ √ √
|BC| 4− 2 42 − 2 2 110 − 3 2
√ √ √
|OH| 2 2 21 55
√ √ √ √ √
|ON | 2 + 2/2 2 + 42/2 (3 2 + 110)/2
AOB ≈38.942◦ ≈47.156◦ ≈44.049◦
BOC ≈51.058◦ ≈42.844◦ ≈45.951◦
OAB ≈70.529◦ ≈66.421◦ ≈67.976◦
√ √ √ √ √
Perim. 24 − 4 2 16 + 4 42 − 8 2 24 + 4 110 − 12 2
√ √ √
Area 14 + 8 2 34 + 8 21 92 + 12 55

4 The “limit octagon” and Future Researches

 
Many aspects of the new sequence Cn n are interesting to investigate. For
example, scaling the octagon Cn by a factor equal to the n-th Fibonacci number
N
Fn , the sequence will converge to a limit octagon C∞ (where the top N stands
for “normalized”) that can be drawn through the “Carboncettus construction”
described at the beginning of Sect. 2, by starting from the circumferences with
radii given by the following limit ratios
Fn Fn+2
lim =1 and lim , (4)
n→∞ Fn n→∞ Fn
 The Sequence of Carboncettus Octagons 379

Fig. 3. The second element of the Carboncettus sequence, the octagon C2 , is drawn in
black. A regular octagon inscribed in the same circumference Γ3 , is also represented in
red. (Color figure online)

respectively. It is well known that the limit of the ratio of two consecutive
Fibonacci numbers Fn+1 /Fn converges to the golden ratio
√
φ := (1 + 5)/2 ≈ 1.618033987, (5)

hence, the second limit in (4) is simple to compute as follows3

Fn+2 Fn+2 Fn+1

lim = lim · = φ2 ≈ 2.618033987, (6)
n→∞ Fn n→∞ Fn+1 Fn
N
and we conclude that C∞ can be constructed using the circumferences Γ1 and
Γφ2 .
Another approach to directly study the “limit octagon” C∞ instead of the
N
“limit normalized octagon” C∞ , could come by using the computational system
introduced for example in [18–20] and applied as well to limit curves, limit poly-
topes, fractals and similar geometric shapes in [1,3–5,20] (or even to Fibonacci
numbers in [9]).

Aknowledgments. This work is partially supported by the research projects

“IoT&B, Internet of Things and Blockchain”, CUP J48C17000230006, POR Calabria
FESR-FSE 2014–2020.

3
The reader certainly remembers the well know property φ2 = 1 + φ of the golden
ratio that causes the coincidence of the fractional parts of (5) and (6).
380 F. Caldarola et al.

References
1. Antoniotti, L, Caldarola, F., Maiolo, M.: Infinite numerical computing applied to
Hilbert’s, Peano’s, and Moore’s curves. Mediterr. J. Math. (in press)
2. Ball, D.: Equiangular polygons. Math. Gaz. 86(507), 396–407 (2002)
3. Caldarola, F.: The Sierpiński curve viewed by numerical computations with infini-
ties and infinitesimals. Appl. Math. Comput. 318, 321–328 (2018). https://doi.
org/10.1016/j.amc.2017.06.024
4. Caldarola, F.: The exact measures of the Sierpiński d-dimensional tetrahedron in
connection with a Diophantine nonlinear system. Commun. Nonlin. Sci. Numer.
Simul. 63, 228–238 (2018). https://doi.org/10.1016/j.cnsns.2018.02.026
5. Caldarola, F., Maiolo, M., Solferino, V.: A new approach to the Z-transform
through infinite computation. Commun. Nonlin. Sci. Numer. Simul. 82, 105019
(2020). https://doi.org/10.1016/j.cnsns.2019.105019
6. Caldarola, F., Cortese, D., d’Atri, G., Maiolo, M.: Paradoxes of the infinite and
ontological dilemmas between ancient philosophy and modern mathematical solu-
tions. In: Sergeyev, Y., Kvasov, D. (eds.) NUMTA 2019. LNCS, vol. 11973, pp.
358–372. Springer, New York (2020)
7. De Villiers, M.: Equiangular cyclic and equilateral circumscribed polygons. Math.
Gaz. 95, 102–107 (2011)
8. Koshy, T.: Fibonacci and Lucas Numbers with Applications. Wiley, New York
(2001)
9. Margenstern, M.: Fibonacci words, hyperbolic tilings and grossone. Commun. Non-
lin. Sci. Numer. Simul. 21(1–3), 3–11 (2015)
10. Pirillo, G.: A characterization of Fibonacci numbers. Chebyshevskii Sbornik 19(2),
259–271 (2018)
11. Pirillo, G.: Figure geometriche su un portale del Duomo di Prato. Prato Storia e
Arte 121, 7–16 (2017). (in Italian)
12. Pirillo, G.: La scuola pitagorica ed i numeri di Fibonacci. Archimede 2, 66–71
(2017). (in Italian)
13. Pirillo, G.: L’origine pitagorica dei numeri di Fibonacci. Periodico di Matematiche
9(2), 99–103 (2017). (in Italian)
14. Pirillo, G.: Some recent results of Fibonacci numbers, Fibonacci words and Stur-
mian words. Southeast Asian Bull. Math. 43(2), 273–286 (2019)
15. Pirillo, G.: Fibonacci numbers and words. Discret. Math. 173(1–3), 197–207 (1997)
16. Pirillo, G.: Inequalities characterizing standard Sturmian and episturmian words.
Theoret. Comput. Sci. 341(1–3), 276–292 (2005)
17. Pirillo, G.: Numeri irrazionali e segmenti incommensurabili. Nuova Secondaria 7,
87–91 (2005). (in Italian)
18. Sergeyev, Y.D.: Arithmetic of Infinity. Edizioni Orizzonti Meridionali, Cosenza
(2003)
19. Sergeyev, Y.D.: Lagrange lecture: methodology of numerical computations with
infinities and infinitesimals. Rend. Semin. Matematico Univ. Polit. Torino 68(2),
95–113 (2010)
20. Sergeyev, Y.D.: Numerical infinities and infinitesimals: methodology, applications,
and repercussions on two Hilbert problems. EMS Surv. Math. Sci. 4(2), 219–320
(2017)
 On the Arithmetic of Knuth’s Powers
and Some Computational Results About
Their Density

Fabio Caldarola1 , Gianfranco d’Atri1 , Pietro Mercuri2 ,

and Valerio Talamanca3(B)
1
Department of Mathematics and Computer Science, Università della Calabria,
Cubo 31/A, 87036 Arcavacata di Rende, Italy
{caldarola,datri}@mat.unical.it
2
Università “Tor Vergata”, 00133 Rome, Italy
[email protected]
3
Università Roma Tre, 00146 Rome, Italy
[email protected]

Abstract. The object of the paper are the so-called “unimaginable

numbers”. In particular, we deal with some arithmetic and computa-
tional aspects of the Knuth’s powers notation and move some first steps
into the investigation of their density. Many authors adopt the conven-
tion that unimaginable numbers start immediately after 1 googol which
is equal to 10100 , and G.R. Blakley and I. Borosh have calculated that
there are exactly 58 integers between 1 and 1 googol having a nontrivial
“kratic representation”, i.e., are expressible nontrivially as Knuth’s pow-
ers. In this paper we extend their computations obtaining, for example,
that there are exactly 2 893 numbers smaller than 1010 000 with a non-
trivial kratic representation, and we, moreover, investigate the behavior
of some functions, called krata, obtained by fixing at most two arguments
in the Knuth’s power a ↑b c.

Keywords: Unimaginable numbers · Knuth up-arrow notation ·

Algebraic recurrences · Computational number theory

1 Introduction: The Unimaginable Numbers

An unimaginable number, intuitively and suggestively, is a number that go
beyond the human imagination. There is not a completely accepted standard
formal definition of unimaginable numbers, but one of the most used is the fol-
lowing: a number is called unimaginable if it is greater than 1 googol, where
1 googol is equal to 10100 . To better understand the size of numbers like these,
consider that it is estimated that in the observable universe there are at most
1082 atoms; this justifies the term unimaginable. The first appearance of the
unimaginable numbers was, to our knowledge, in Magna Graecia in the work of
Archimedes of Syracuse. Archimedes in his work called “Arenarius” in Latin, or
c Springer Nature Switzerland AG 2020
Y. D. Sergeyev and D. E. Kvasov (Eds.): NUMTA 2019, LNCS 11973, pp. 381–388, 2020.
https://doi.org/10.1007/978-3-030-39081-5_33
382 F. Caldarola et al.

the “Sand Reckoner” in English, describes, using words of the natural language,
an extremely large number that, in exponential notation, is equal to
16
108·10 = 1080 000 000 000 000 000 . (1)

Obviously, writing this number without any kind of modern mathematical nota-
tion, as Archimedes did, is very very difficult. Let us jump to modern times and
introduce the most used notation that allows to write numbers so large that are
definitely beyond the common experience of a human being.

Definition 1 (Knuth’s up-arrow notation). For all non-negative integers

a, b, n, we set
⎧
⎪
⎨a · b if n = 0;
n
a ↑ b := 1 if n ≥ 1 and b = 0; (2)
⎪
⎩ n−1 n
a↑ (a ↑ (b − 1)) if n ≥ 1 and b ≥ 1.

For n = 1 we obtain the ordinary exponentiation, e.g., 3 ↑ 4 = 34 ; for n = 2

we obtain tetration, for n = 3 pentation, and so on. Hence (2) represents the so
called n-hyperoperation. Now, using Knuth’s notation, the Archimedes’ number
(1) can be easily written as follows
 
(10 ↑ 8) ↑2 2 ↑ (10 ↑ 8).

In this paper we use an alternative notation to that introduced in Definition 1.

Denoting by IN the set of natural numbers (i.e., non-negative integers) we define
the Knuth’s function k as follows

k : IN × IN × IN −→ IN
(3)
(B, d, T ) → k(B, d, T ) := B ↑d T

and we call the first argument of k (i.e., B) the base, the second (i.e., d) the
depth and the third (i.e., T ) the tag (see [3]).
The paper is organized as follows: in Sect. 2 we introduce some general com-
putational problems, while in Sect. 3, which is the core of this work, we deal with
density and representational problems related to Knuth’s powers. In particular,
Proposition 2 and Corollary 1 give some simple results which characterize the
difference of digits in base 10 between two “consecutive” Knuth’s powers of the
simplest “non-trivial” type, i.e., a ↑2 2. Proposition 3 extends, instead, a compu-
tation by Blakley and Borosh (see [3, Proposition 1.1]): they found that there
are exactly 58 numbers smaller than 1 googol (= 10100 ) nontrivially expressible
through the Knuth’s function k. We obtained that such number increases to 2893
if we consider integers lesser than 1010 000 . Among these 2893 numbers, 2888 are
expressible through the aforementioned form a ↑2 2.
We conclude the introductory section by giving the reader some brief infor-
mation on some useful references to deepen the issues related to unimaginable
numbers. In addition to article [3] which, for our purposes, represents the main
 On the Arithmetic of Knuth’s Powers 383

reference, the same authors investigate the modular arithmetic of Knuth’s pow-
ers in [4]. Knuth himself had instead introduced the notation (2) a few years
earlier in [16] (1976), but these ideas actually date from the beginning of the
century (see [1,2,13,19]). More recent works that start from “extremely large” or
“infinite” numbers are [7,8,12,14,15,17,18,20,21,23]. There are also the online
resources [5,6,22]. Finally [10] provides the reader with a brief general introduc-
tion with some further reference.

2 Representational and Computational Problems

It is obvious that almost all the numbers are unimaginable, hence a first natural
question is: can we write every unimaginable number using Knuth’s up-arrow
notation? The answer is trivial: we cannot. Actually there are just very few
numbers that are expressible using this notation. More precisely let K0 denote
the image of the map k, i.e., the set of those natural numbers that are expressible
via Knuth’s notation. As customary one can consider the ratio
 
# K0 ∩ {m ∈ IN : m < x}
ρ0 (x) = . (4)
x
The computed values of ρ0 (x) are very close to zero and ρ0 (x) appears to be
quickly converging to zero as x → +∞. In the next section we compute some
values of a ratio related to this.
In recent years dozens of systems and notations have been developed to write
unimaginable numbers (for example see [1,6,7,14,15]), most of them can reach
bigger numbers in a more compact notation than Knuth’s notation can, but the
difference between two consecutive numbers with a compact representation in a
specific notation often increases quicker than in Knuth’s notation. Hence, almost
all unimaginable numbers remain inaccessible to write (and to think about?) and
the problem of writing an unimaginable number in a convenient way is open.
A strictly related open problem is to find a good way to represent an unimag-
inable number on a computer. It is not possible to represent with usual decimal
notation numbers like 3 ↑3 3 on a computer at the present time. Hence, to do
explicit computations involving these numbers is quite hard. Therefore finding
a way, compatible with classical operations, to represent these kind of numbers
on a computer not only would make many computations faster but it would also
help to deeper develop the mathematical properties and the applications related
to unimaginable numbers.
We recall, for the convenience of the reader, some basic properties of Knuth’s
up-arrow notation that will be used in the next section.
Proposition 1. For all positive integers a, b, n, with b > 1, we have:
(i) a ↑n b < (a + 1) ↑n b;
(ii) a ↑n b < a ↑n (b + 1);
(iii) a ↑n b ≤ a ↑n+1 b, where the equality holds if and only if a = 1 or a = b = 2.
Proof. See [3, Theorem 1.1].
384 F. Caldarola et al.

3 About the Density of Numbers with Kratic

Representation
We follow the nomenclature used in [3] and we say that a positive integer x has
a non-trivial kratic representation if there are integers a, b, n all greater than 1
such that x = a ↑n b. Note that a kratic representation should not be confused
with a kratos: a kratos (pl. krata)1 is a function h that comes from the Knuth’s
function k by fixing at most two arguments (see [3]). It is then a natural question
to ask “how many” numbers have a non-trivial kratic representation.

Example 1. The least positive integer with non-trivial kratic representation is 4,

in fact 2 ↑n 2 = 4 for all positive integers n.

It is easy to see that numbers with kratic representation of the form a ↑2 2 =

a
a are more frequent than those with other types of kratic representation. The
following proposition states how often they appear with respect to the number
of digits, i.e., it calculates the increment of the number of digits between two
“consecutive” numbers with kratic representation of that form. We need a further
piece of notation: as usual Log denotes the logarithm with base 10,
α the floor
of a real number α and ν(a) the number of digits of a positive integer a (in base
10). Using these notation we have

ν(a) =
Log a + 1 (5)

for all positive integers a.

Proposition 2. For every integer a ≥ 1 we have



 a+1
 a 1
ν (a + 1) − ν (a ) = Log(a + 1) + a Log 1 + (6)
a
or 

  1
ν (a + 1)a+1 − ν (aa ) = Log(a + 1) + a Log 1 + + 1. (7)
a
Proof. The proposition states that the difference between the number of digits
of (a + 1)a+1 and aa is given by Formula (6) or (7). For any integer a ≥ 1 we
have
  
ν (a + 1)a+1 − ν (aa ) = Log(a + 1)a+1 −
Log aa 
=
(1 + a)Log(a + 1) −
a Log a
=
Log(a + 1) + a Log(a + 1) − a Log a
+ a Log a −
a Log a


1
= Log(a + 1) + a Log 1 + + a Log a (8)
a
−
a Log a .
1
Kratos, written in ancient Greek κράτ oς, indicated the “personification of power”.
 On the Arithmetic of Knuth’s Powers 385

Since for all real numbers α and β the following inequalities hold

α +
β ≤
α + β ≤
α +
β + 1, (9)

then, combining (9) with (8) we obtain



1  
Log(a + 1) + a Log 1 + ≤ ν (a + 1)a+1 − ν (aa )
 a

1
≤ Log(a + 1) + a Log 1 + + 1,
a
proving the proposition.
Corollary 1. For every integer a ≥ 1 the following inequalities hold
 

Log(a + 1) ≤ ν (a + 1)a+1 − ν (aa ) ≤
Log(a + 1) + 2. (10)

Proof. The first inequality in (10) is an immediate consequence of (6). For the
second one note that, using the previous proposition, the second inequality in
(9) and the well-known bound

a
1
1+ < e, for all a ≥ 1, (11)
a
we obtain

a 
  1
ν (a + 1)a+1 − ν (aa ) ≤ Log(a + 1) + Log 1 + +1
 a
a 
1
≤
Log(a + 1) + Log 1 + +2
a
=
Log(a + 1) + 2.

The two possibilities given

 by (6) and
 (7) in Proposition 2 and the three given
by Corollary 1 (that is, ν (a + 1)a+1 − ν (aa ) −
Log(a + 1) = 0, 1, 2) are all
effectively realized: it is sufficient to look at the values a = 1, 2, 7 in Table 1.

Table 1. The first 10 values of aa .

a 1 2 3 4 5 6 7 8 9 10
a
a 1 4 27 256 3 125 46 656 823 543 16 777 216 387 420 489 10 000 000 000

Remark 1. Note that using (11) and the lower bound 2 ≤ (1 + 1/a)a , for a ≥ 1,
we obtain
(a + 1) ↑2 2
2(a + 1) ≤ < e(a + 1) (12)
a ↑2 2
for all integers a ≥ 1. It is also interesting that the ratio of two consecutive
numbers of that form can be approximated by a linear function in the base a.
386 F. Caldarola et al.

The previous remark implies that given a number with kratic representation
of the form a ↑2 2, the subsequent one, (a + 1) ↑2 2, is rather close to it. Instead,
numbers with kratic representation of other forms are much more sporadic: the
following proposition gives a more precise idea of this phenomenon.
Proposition 3. There are exactly 2893 numbers smaller than 1010 000 that
admit a non-trivial kratic representation. Among them, 2888 have a representa-
tion of the form a ↑2 2, and only 5 do not have such a representation.
Proof. By [3, Proposition 1.1] there are exactly 58 numbers with less than 102
digits in decimal notation that have a non-trivial kratic representation; we collect
them in the following set
E2 = {a ↑2 2 : 2 ≤ a ≤ 56}  {2 ↑2 3, 3 ↑2 3, 2 ↑2 4}.
Note also that some of them have more than one representation:
2 ↑2 2 = 4 = 2 ↑d 2 ∀d ≥ 2, 3 ↑2 3 = 327 = 3 ↑3 2, 2 ↑2 4 = 216 = 2 ↑3 3.
We look for the numbers we need to add to E2 to obtain the desired set
 
E := n ∈ IN : n < 1010 000 and n has a non-trivial kratic representation .
We consider different cases depending on the depth d.
(i) “d = 2”. Since
Log(2889 ↑2 2) ≈ 9998.1 and Log(2890 ↑2 2) ≈ 10001.99,
we have to add to E2 the numbers from 57 ↑2 2 to 2889 ↑2 2. Then, since
Log(5 ↑2 3) ≈ 2184.28 and Log(6 ↑2 3) ≈ 36305.4,
the numbers 4 ↑2 3 and 5 ↑2 3 belong to E as well. Instead,
Log(3 ↑2 4) ≈ 3638334640024.1 and Log(2 ↑2 5) ≈ 19728.3 (13)
guarantee, by using Proposition 1, that there are no other elements with
d = 2 in E.
(ii) “d = 3”. Note that 4 ↑3 2 = 4 ↑2 4 > 3 ↑2 4, and 3 ↑3 3 = 3 ↑2 3 ↑2 3 > 3 ↑2 4,
and 2 ↑3 4 = 2 ↑2 216 , hence, by using (13), we have that E does not contain
any new element with d = 3.
(iii) “d = 4”. Since 3 ↑4 2 = 3 ↑3 3 and 2 ↑4 3 = 2 ↑3 4, part (ii) yields that they
do not belong to E. Therefore, it has no new elements with d = 4.
Now, only the (trivial) case “d ≥ 5” remains, but since 3 ↑d 2 > 3 ↑4 2, (iii)
yields that there are no new elements with d ≥ 5 in E. In conclusion, we have
proved that
E = E2  {a ↑2 2 : 57 ≤ a ≤ 2889}  {4 ↑2 3, 5 ↑2 3}
and its cardinality is 2893. From the proof, it is also clear that the only elements
of E having no representation of the type a ↑2 2 are 2 ↑2 3, 3 ↑2 3, 2 ↑2 4, 4 ↑2 3
and 5 ↑2 3.
 On the Arithmetic of Knuth’s Powers 387

We conclude the paper with some observations about the frequency of num-
bers with non-trivial kratic representation. Let K denote the set of integers that
admit a non-trivial kratic representation. Define the kratic representation ratio
ρ(x) as: 
#(K ∩ {m ∈ IN : m < x}
ρ(x) := .
x
(Note the differences with respect to (4).) We find the following values:
1
ρ(10) = = 0.1 · 10−1 ,
10
  3
ρ 102 = = 0.3 · 10−1 ,
102
  5
ρ 104 = = 0.5 · 10−3 ,
104
  9
ρ 1010 = = 0.9 · 10−9 ,
1010
 2
 58
ρ 1010 = = 0.58 · 10−98 ,
10102
 4
 2893
ρ 1010 = = 0.2893 · 10−9996 .
10104
These data seems to indicate that ρ(x) tends rapidly to zero for x → +∞.
However, it is not known, to our knowledge, an explicit formula for ρ(x) or,
equivalently, for the cardinality of the set K(x) = K ∩ {m ∈ IN : m < x} itself.

Aknowledgments. This work is partially supported by the research projects

“IoT&B, Internet of Things and Blockchain”, CUP J48C17000230006, POR Calabria
FESR-FSE 2014–2020. The third author was also supported by the research grant “Ing.
Giorgio Schirillo” of the Istituto Nazionale di Alta Matematica “F. Severi”, Rome.

References
1. Ackermann, W.: Zum hilbertschen aufbau der reellen zahlen. Math. Ann. 99, 118–
133 (1928). https://doi.org/10.1007/BF01459088
2. Bennett, A.A.: Note on an operation of the third grade. Ann. Math. Second Series
17(2), 74–75 (1915). https://doi.org/10.2307/2007124
3. Blakley, G.R., Borosh, I.: Knuth’s iterated powers. Adv. Math. 34(2), 109–136
(1979). https://doi.org/10.1016/0001-8708(79)90052-5
4. Blakley, G.R., Borosh, I.: Modular arithmetic of iterated powers. Comp. Math.
Appl. 9(4), 567–581 (1983)
5. Bowers, J.: Exploding Array Function. Accessed 25 Apr 2019. http://www.
polytope.net/hedrondude/array.htm
6. Bowers, J.: Extended operator notation. Accessed 21 Apr 2019. https://sites.
google.com/site/largenumbers/home/4-1/extended operators
7. Bromer, N.: Superexponentiation. Mathematics Magazine 60(3), 169–174 (1987).
JSTOR 2689566
388 F. Caldarola et al.

8. Caldarola, F.: The exact measures of the Sierpiński d-dimensional tetrahedron in

connection with a Diophantine nonlinear system. Commun. Nonlinear Sci. Numer.
Simul. 63, 228–238 (2018). https://doi.org/10.10.16/j.cnsns.2018.02.026
9. Caldarola, F., Cortese, D., d’Atri, G., Maiolo, M.: Paradoxes of the infinite and
ontological dilemmas between ancient philosophy and modern mathematical solu-
tions. In: Sergeyev, Y.D., Kvasov, D.E. (eds.) NUMTA 2019. LNCS 11973, pp.
358–372. Springer, Cham (2020). https://doi.org/10.1007/978-3-030-39081-5 31
10. Caldarola, F., d’Atri, G., Maiolo, M.: What are the ”unimaginable numbers”?
(submitted)
11. Donner, J., Tarski, A.: An extended arithmetic of ordinal numbers. Fundamenta
Mathematicae 65, 95–127 (1969)
12. Friedman, H.M.: Long finite sequences. J. Comb. Theor. Series A 95(1), 102–144
(2001). https://doi.org/10.1006/jcta.2000.3154
13. Goodstein, R.L.: Transfinite ordinals in recursive number theory. J. Symbolic Logic
12(4), 123–129 (1947). https://doi.org/10.2307/2266486
14. Hooshmand, M.H.: Ultra power and ultra exponential functions. Integr.
Transforms Spec. Functions 17(8), 549–558 (2006). https://doi.org/10.1080/
1065246050042224
15. Knobel, R.A.: Exponentials reiterated. American Mathematical Monthly 88(4),
235–252 (1981). https://doi.org/10.2307/2320546
16. Knuth, D.E.: Mathematics and computer science: coping with finiteness. Science
194(4271), 1235–1242 (1976). https://doi.org/10.1126/science.194.4271.1235
17. Lameiras Campagnola, M., Moore, C., Costa, J.F.: Transfinite ordinals in recur-
sive number theory. J. Complex. 18(4), 977–1000 (2002). https://doi.org/10.1006/
jcom.2002.0655
18. Leonardis, A., d’Atri, G., Caldarola, F.: Beyond Knuth’s notation for ”Unimagin-
able Numbers” within computational number theory. arXiv:1901.05372v2 [cs.LO]
(2019)
19. Littlewood, J.E.: Large numbers. Math. Gaz. 32(300), 163–171 (1948). https://
doi.org/10.2307/3609933
... ...
20. MacDonnell, J.F.: Some critical points of the hyperpower function xx xx . Int.
J. Math. Educ. 20(2), 297–305 (1989). https://doi.org/10.1080/0020739890200210
21. Marshall, A.J., Tan, Y.: A rational number of the form a a with a irrational. Math.
Gaz. 96, 106–109 (2012)
22. Munafo, R.: Large Numbers at MROB. Accessed 19 May 2019
23. Nambiar, K.K.: Ackermann functions and transfinite ordinals. Appl. Math. Lett.
8(6), 51–53 (1995). https://doi.org/10.1016/0893-9659(95)00084-4
 Combinatorics on n-sets: Arithmetic
Properties and Numerical Results

Fabio Caldarola , Gianfranco d’Atri, and Marco Pellegrini(B)

Department of Mathematics and Computer Science, Università della Calabria,

Cubo 31/A, 87036 Arcavacata di Rende, CS, Italy
{caldarola,datri}@mat.unical.it, [email protected]

Abstract. The following claim was one of the favorite “initiation ques-
tion” to mathematics of Paul Erdős: for every non-zero natural number
n, each subset of I(2n) = {1, 2, . . . , 2n}, having size n + 1, contains at
least two distinct elements of which the smallest divides the largest. This
can be proved using the pigeonhole principle. On the other side, it is easy
to see that there are subsets of I(2n) of size n without divisor-multiple
pairs; we call them n-sets, and we study some of their combinatorial
properties giving also some numerical results. In particular, we give a
precise description of the elements that, for a fixed n, do not belong to
every n-set, as well as the elements that do belong to all the n-sets. Fur-
thermore, we give an algorithm to count the n-sets for a given n and,
in this way, we can see the behavior of the sequence a(n) of the num-
ber of n-sets. We will present some different versions of the algorithm,
along with their performances, and we finally show our numerical results,
that is, the first 200 values of the sequence a(n) and of the sequence
q(n) := a(n + 1)/a(n).

Keywords: n-sets · Natural numbers · Divisibility relations ·

Combinatorics on finite sets · n-tuples

1 Introduction
During a dinner Paul Erdős posed a question to the young Lajos Pósa: is it
true that, for every integer n ≥ 1, each subset of I(2n) = {1, 2, . . . , 2n} having
size n + 1 contains at least two distinct elements of which the smallest divides
the largest? Before the dinner ended, he proved this fact using the pigeonhole
principle and equivalence classes of different cardinalities (see [1]).
On the other side, it is easy to see that there are subsets of I(2n) of size
n without divisor-multiple pairs; we call them n-sets (see Definition 1) and we
asked ourselves some questions about them: How many are they? What can we
say about their elements? If we fix n, are there elements included in every n-set?
Are there elements not included in any n-set? There is already some research
about n-sets, see for instance [2–4].
This paper is structured as follows. In Sect. 2 we introduce the problem and
we describe an equivalence relation that is the key to answer to many of these
c Springer Nature Switzerland AG 2020
Y. D. Sergeyev and D. E. Kvasov (Eds.): NUMTA 2019, LNCS 11973, pp. 389–401, 2020.
https://doi.org/10.1007/978-3-030-39081-5_34
390 F. Caldarola et al.

questions, as presented in [5]. In Sect. 3 we find some characterizations of the

elements that are not included in any n-set, as well as the elements that are
included in any n-set; some results are already present in [6]. Finally, in Sect. 4,
we describe an algorithm to count the n-sets for a given integer n > 0. In this
way we can see the behavior of the sequence a(n) of the number of n-sets and
we will also show some different versions of the algorithm.

2 Notations and Preliminaries

2.1 Erdős Problem

Throughout the paper, lowercase letters indicate natural numbers, that is non-
negative integers. Given two integers y > x > 0, we say that they form a
divisor-multiple pair if x divides y, in symbols x|y. Moreover, the cardinality of
a set A is denoted by |A| and, for short, we indicate the set {1, 2, . . . , n} of the
first n positive natural numbers by I(n).
For the following proposition see [1, Chap. 25.1] or [5].

Proposition 1. For any integer n ≥ 1, every subset X of I(2n) such that |X| ≥
n + 1 contains at least a divisor-multiple pair.

For instance, if n = 3 the set {3, 4, 5, 6} ⊂ I(6) contains only one divisor-
multiple pair, while the set {1, 2, 4, 6} ⊂ I(6) contains exactly 5 divisor-multiple
pairs. Instead, what can happen if X ⊂ I(2n) and |X| = n? For instance,
{2, 3, 5} ⊂ I(6) and {6, 7, 8, 9, 10} ⊂ I(10) do not contain any divisor-multiple
pair, and these examples lead us to the following definition.

Definition 1. If X is a subset of I(2n) such that |X| = n and X does not

contain any divisor-multiple pair, we say that X is a n-set.

Remark 1. We observe that for every positive integer n, n-sets do exist, that is,
the value n + 1 in Proposition 1 is optimal: indeed, for every n ≥ 1, the sets
Yn = {n + 1, n + 2, . . . , 2n − 1, 2n} and Zn = {n, n + 1, . . . , 2n − 2, 2n − 1} are
n-sets.

2.2 An Interesting Equivalence Relation

The proof of Proposition 1 uses some of the following facts, that will be also useful
to study the structure of n-sets. First, we observe that any positive integer x can
be written in a unique way as x = 2k d, with k ≥ 0 and d odd. Then, we define
a function ε such that ε(x) = d.

Definition 2. If x, y ∈ I(2n), we say that x ∼ y if ε(x) = ε(y); we denote [x]n

the set {y ∈ I(2n) : y ∼ x}, that is, the equivalence class of x.
 Combinatorics on n-sets: Arithmetic Properties and Numerical Results 391

It is easy to see that ∼ is an equivalence relation and each equivalence class

[x]n is a set of the form {d, 2d, . . . , 2t d} with d odd, 2t d ≤ 2n and 2t+1 d > 2n;
thus, every equivalence class [x]n contains exactly one odd number. This means
that here are n equivalence classes in I(2n) because there are exactly n odd
numbers in I(2n). We moreover observe that if x ∼ y and x < y then x|y, since

x = 2k d and y = 2k d, with k < k  .

Example 1. For n = 10 we have [1]10 = {1, 2, 4, 8, 16}, [3]10 = {3, 6, 12},

[5]10 = {5, 10, 20}, [7]10 = {7, 14}, [9]10 = {9, 18}, [11]10 = {11}, [13]10 = {13},
[15]10 = {15}, [17]10 = {17}, [19]10 = {19}.

Definition 3. Given n ≥ 1 we denote the family of all the n-sets by N (n) and
its cardinality, |N (n)|, by a(n), i.e., a(n) represents the number of (unordered)
n-tuples contained in I(2n) without divisor-multiple pairs.

For example, if n = 1 we have N (1) = {{1}, {2}} and a(1) = 2. Instead, if

n = 3 we have N (3) = {{2, 3, 5}, {3, 4, 5}, {4, 5, 6}}, hence a(3) = 3.
For future references we state the following

Remark 2. Fixed n ≥ 1, every n-set contains exactly one element of any equiv-
alence class [x]n .

This gives a (rather unpractical) method to construct an n-set: we choose

exactly one element from every equivalence class and we check if there are divisor-
multiple pairs. If not, we have an n-set, otherwise we change our choices.
If we study what happens when increasing n, we see that, for any n and for
any x, we have [x]n ⊆ [x]n+1 , that is, equivalence classes are “nested”. Moreover,
we can easily observe that if d is odd with n < d < 2n, we have [d]n = {d}: so
|[d]n | = 1 and d belongs to every n-set. This motivates the following

Definition 4. Let n ≥ 1 be fixed. The set S(n) := X∈N (n) X is called the n-

kernel and the set E(n) := I(2n)− X∈N (n) X is called the set of the n-excluded
elements.

For instance, S(1) = ∅, S(3) = {5}, while E(1) = ∅ and E(3) = {1} (see
the example after Definition 3). Some simple results about S(n) and E(n) are
hence the following (see [5])

Proposition 2. (a) If d is odd with n < d < 2n, we have d ∈ S(n).

(b) If 1 ≤ n < m, then E(n) ⊆ E(m).

3 Excluded Integers and Kernel

Proposition 2 (b) implies that if x does not belong to any n-set and 1 ≤ n <
m, then x does not belong to any m-set. Then, what is the minimum n that
“excludes” a given x? We start to consider odd numbers and, from now on, d
will be always an odd natural number.
392 F. Caldarola et al.

Definition 5. Given d, we call d-threshold value the number nd := (3d + 1)/2.

The reason of this definition comes from the following (see [5,6]).

Proposition 3. For every n ≥ 1, for every d, we have that d ∈ E(n) if and only
if n ≥ nd , i.e. n is equal or greater than the d-threshold value nd = (3d + 1)/2.

The previous proposition represents an effective criterion to evaluate if an odd

number d is n-excluded for a certain n or, vice versa, to calculate the minimum
n that makes a certain odd number d an n-excluded. In order to see whether
any integer is n-excluded or not, we need further considerations. From now on it
could be useful to label odd numbers, so we will introduce the following notation:
di := 2i − 1, ∀i ≥ 1. Obviously, di < dj if and only if i < j.

Definition 6. Fixed n ≥ 1, we set Mi = Mi,n := max[di ]n and mi = mi,n :=

min([di ]n \ E(n)) for all integers i with 1 ≤ i ≤ n.

Example 2. For n = 5 we have [1]5 = {1, 2, 4, 8}, [3]5 = {3, 6}, [5]5 = {5, 10},
[7]5 = {7} and [9]5 = {9}. Thus M1 = 8, M2 = 6, M3 = 10 M4 = 7, M5 = 9 and
Y5 = {6, 7, 8, 9, 10} = {Mi : 1 ≤ i ≤ 5}.

Observe that the Mi are never n-excluded because they are exactly all the
elements of the n-set Yn described in Remark 1. On the other hand, enumerating
the set {mi,n : 1 ≤ i ≤ n} completely is not as simple as enumerating the set
{Mi,n : 1 ≤ i ≤ n} because we previously need to determine E(n). Nevertheless
these numbers will play an important role in obtaining the following (see [6]).

Proposition 4. Let n ≥ 1 be fixed, x ∈ I(2n) and ε(x) = di . Then, the following

are equivalent:

(a) x ∈ E(n);
(b) x is a divisor of a certain mj , the minimum non-excluded element of the
class [dj ]n , with j > i.

Corollary 1. If x = 2k d ∈ E(n) and h < k, then 2h d ∈ E(n).

This means
 that, in every class [di ]n , the set
 of non-excluded elements is of
the form mi = 2k di , 2k+1 di , . . . , 2k+t di = Mi , for some k, t ≥ 0. In particular,
if k = 0 all the elements of [di ]n are not excluded, instead if t = 0 we have
mi = Mi ∈ S(n). In any case, the maximum element is never excluded.
We now need a generalized definition of the threshold value, connected to any
even or odd integer, as follows:

Definition 7. Given x = 2k d, k ≥ 0, we call x-threshold value the number

nx := (3k+1 d + 1)/2.

Note that if k = 0 then x = d and the x-threshold value in Definition 7 coin-

cides with the d-threshold value in Definition 5. Hence nx is a true generalization
of nd , extended to even integers x.
 Combinatorics on n-sets: Arithmetic Properties and Numerical Results 393

Remark 3. It is interesting to note that different integers can generate the same
threshold value. In fact, for any integer h with 0 ≤ h ≤ k, we have that all the
numbers of the form yh = 2k−h 3h d generate the same threshold value nyh =
(3(k−h)+1 3h d + 1)/2 = (3k+1 d + 1)/2.

Finally, we have a criterion to decide whether an integer is n-excluded or not

(see [6]).

Proposition 5. For every x = 2k d, we have that x ∈ E(n) if and only if n ≥ nx ,

where nx is the x-threshold value defined as nx = (3k+1 d + 1)/2.

Example 3. Now we take n = 35 and we want to determine the set E(35) ⊂

I(70). Note that the condition n ≥ nx is equivalent to d ≤ (2n − 1)/3k+1 .
– Excluded odd numbers: 1, 3, 5, . . . , 21, 23 (because 23 ≤ 69/3 and 25 >
69/3).
– Excluded numbers of the form 2d (k = 1): (2 · 1), (2 · 3), (2 · 5), (2 · 7) (because
7 ≤ 69/32 and 9 > 69/32 ).
– Excluded numbers of the form 4d (k = 2): (4 · 1) (because 1 ≤ 69/33 and
3 > 69/33 ).
– Excluded numbers of the form 2k d with k ≥ 3: none (because 1 > 69/3k+1 ).
 
In conclusion we have E(35) = 1, 2, 3, 4, 5, 6, 7, 9, 10, 11, 13, 14, 15, 17, 19, 21, 23 .

The previous proposition, about excluded integers, can also be helpful to give
a precise description of the n-kernel S(n).

Proposition 6. For every n ≥ 1 we have

 
n+1 2n − 1
S(n) = {d : n + 1 ≤ d ≤ 2n − 1} ∪ 2d : ≤d≤ .
2 3

Proof. Observe that x ∈ S(n) if and only if |[x]n \ E(n)| = 1, then we distinguish
between odd and even numbers.
For odd numbers, by Proposition 2 (a), we already know that {d : n+1 ≤ d ≤
2n − 1} ⊂ S(n); instead, if d ≤ n we know that |[d]n | ≥ 2, then the maximum of
[d]n is even and d ∈
/ S(n).
For even numbers: if x = 2k d with k ≥ 1, we have that 2k d ∈ S(n) if and
only if 2k−1 d ∈ E(n) and 2k+1 d ≥ 2n + 2. By Proposition 5, we know that
n ≥ (3k d + 1)/2 and, in addition, 2k d ≥ n + 1. Putting all together, we obtain
 k
n+1 2n − 1 3 2n − 1
≤d≤ =⇒ ≤ < 2.
2k 3k 2 n+1

This forces k = 1, because if k ≥ 2 then (3/2)k > 2; thus, x = 2d and

n+1 2n − 1
≤d≤ .
2 3
394 F. Caldarola et al.

Example 4. As in Example 3 we take n = 35 and we want to determine the set

S(35) ⊂ I(70).
– Odd numbers: 37, 39, 41, . . . , 67, 69.
– Even numbers: with 36/2 ≤ d ≤ 69/3, we get 38, 42, 46.
In conclusion we have:
S(35) = {37, 38, 39, 41, 42, 43, 45, 46, 47, 49, 51, 53, 55, 57, 59, 61, 63, 65, 67, 69}.

4 Counting n-sets
In this section we come to the problem of counting n-sets. By Remark 2 we must
choose exactly one element from every equivalence class, in order to get an n-set.
Now, if an element belongs to the n-kernel its choice is forced. We can ignore
all the n-excluded numbers, then we can define the restricted equivalence classes
Ci := [di ]n \ E(n). These Ci are actually equivalence classes in I(2n) \ E(n).
Working “by hand” with small integers n ≥ 1, we observe a new phenomenon:
some classes contain exactly two non-excluded numbers, say y and 2y, and none
of them is a divisor or a multiple of every other non-excluded numbers.
Definition 8. We say that two classes Ci and Cj are related if there are two
non-excluded integers x, y such that x ∈ Ci , y ∈ Cj , and x, y form a divisor-
multiple pair. If a class Ci is not related to any other class, we say that Ci is
unrelated.
Proposition 7. If C is an unrelated class, then it contains exactly one or two
non-excluded elements.
Proof. Suppose, by reductio ad absurdum, that C is an unrelated class with at
least 3 non-excluded elements. Such 3 elements are then of the form 2k d, 2k+1 d,
2k+2 d and, by Remark 3, we know that 2k+1 d has the same threshold value of
2k 3d. Now this gives 2k 3d < 2k 4d = 2k+2 d ≤ 2n and, consequently, 2k 3d ∈
I(2n), but since 2k 3d is a multiple of 2k d, then we obtain 2k 3d ∈ [3d]n = C  ,
while 2k d ∈ [d]n = C. In conclusion this means that C and C  are related, hence
a contradiction.
For completeness, we notice that unrelated classes with exactly one non-
excluded element actually exist (i.e., the classes that contain the elements of the
kernel) as well as unrelated classes with exactly two non-excluded elements (see,
for instance, Example 5).
Definition 9. An unrelated class with exactly two non-excluded elements is
called an independent pair.
The reason for this name is the following: for an independent pair {y, 2y},
whether we choose y or we choose 2y, the choices of elements of other classes
are not affected. Then, if we count the total number of choices of elements from
other classes and we multiply it by 2 we get a(n). Obviously, this can be done
for every independent pair.
 Combinatorics on n-sets: Arithmetic Properties and Numerical Results 395

Proposition 8. Fixed n ≥ 1, if we have r independent pairs and x choices of

the elements that belong to the other classes, then

a(n) = 2r x.

The proof is trivial from the previous discussion. We now describe a first
version (version 1.0) of our algorithm to count n-sets:

(1) Divide I(2n) in n equivalence classes, according to ∼.

(2) If the minimum element of a class is a divisor of the minimum element of a
following class, remove it (by Proposition 4 it is in fact excluded) and repeat.
Alternatively, remove excluded elements using Proposition 5.
(3) Remove the classes with only one element (these elements form the kernel,
then their choice is forced).
(4) Find and count the independent pairs. If r is their number, remove then the
corresponding classes.
(5) Using all remaining elements, count the choices for them, one element for
each class so that no element is a divisor or a multiple of anyone else (for
this purpose we can use a simple backtracking algorithm). Let x the number
of such choices.
(6) Finally, write out a(n) = 2r x.

In order to execute step (5), we consider a graph G(n) whose vertices are
the numbers of I(2n) that are both non-excluded and in no unrelated class. In
G(n) there is an edge between the vertices y and z if and only if y, z belong to
different classes and form a divisor-multiple pair. Sometimes we can have two
classes Ci and Cj with i < j, an edge between mi and Mj , and no edges going
“outside” these classes. In this case Ci is related only to Cj and vice versa. This
means that the choices involving the elements of these two classes do not affect
the choices of the elements in other classes. Therefore it is easy to count these
choices: since there is only one forbidden combination, namely (mi , Mj ), they
are |Ci | · |Cj | − 1 (we recall that we already removed excluded elements).
Definition 10. If Ci , Cj are two classes as above, we say that they are strictly
related.
By our software computations, in this case we see that |Cj | is always 2, while
|Ci | can be 2 or 3; then, |Ci | · |Cj | − 1 can only take the value 3 or 5. Hence we
have a number (possibly zero) of factors 3 and 5. After removing these classes,
experimentally we see that all the other classes seem to be “connected”, and this
means that there is not a simple way to calculate the number of choices... In
any case, we have that a(n) = 2r 3e3 5e5 x, where r is the number of independent
pairs, ep is the number of factors p that come from the strictly related classes
and x is the number of choices in the remaining part of the graph G(n). We
remark that the number x can still have some (or all) of the prime factors 2, 3, 5.
Thus, the previous exposed algorithm can be improved by adding the follow-
ing step (4.2) after the old step (4) called here (4.1), and modifying the last step
(6) as shown below:
396 F. Caldarola et al.

(4.1) Find and count the independent pairs.If r is their number, remove then
the correspondent classes.
(4.2) Find if there are strictly related classes Ci , Cj . If |Ci | · |Cj | − 1 = 3, add 1
to e3 , instead, if |Ci | · |Cj | − 1 = 5, add 1 to e5 . Remove then Ci and Cj .
(6) Write out a(n) = 2r 3e3 5e5 x.

We refer to such a modified algorithm as the version 1.1 of the algorithm to

compute a(n). We now give a complete and elucidating example of computing
for n = 44.

Example 5. Take n = 44.

– Excluded elements (see Proposition 5):

• odd numbers: 1, 3, 5, . . . , 29;

• numbers of the form 2d: 2, 6, 10, 14, 18;
• numbers of the form 4d: 4, 12;
• numbers of the form 8d: 8.
– Kernel: {45, 47, 49, . . . , 87} ∪ {46, 50, 54, 58} (see Proposition 6).
– Independent pairs: {34, 68}, {38, 76}, {31, 62}, {35, 70}, {37, 74}, {41, 82},
{43, 86}. Hence r = 7.
– Strictly related classes: {28, 56} and {42, 84}, {26, 52} and {39, 78}, {22, 44,
88} and {33, 66}. Thus e3 = 2 and e5 = 1.
– Remaining numbers: 16, 20, 24, 30, 32, 36, 40, 48, 60, 64, 72, 80. See also the
graph G(44) in Fig. 1.
– By using a backtracking algorithm, as in step (5), we find that for the
graph G(44) in Fig. 1 there are 33 choices, from (16, 24, 20, 36, 30) up to
(64, 48, 80, 72, 60), then we have x = 33.

In conclusion, counting the 44-sets we obtain a(44) = 27 · 32 · 5 · 33 = 190080.

Fig. 1. Remaining numbers and their divisibility relations: the graph G(44).
 Combinatorics on n-sets: Arithmetic Properties and Numerical Results 397

The version 1.0 of our algorithm was able to calculate a(156) in 12 c.a. by
using a common desk pc, while version 1.1 can calculate a(156) in about 0.4 ,
a(200) in 1 72 c.a, a(229) in 1h 13 20 c.a. To go further, we developed a version
2.0 of the algorithm, in order to get a better performance of step (5), and to do
it we use a recursion process. More precisely, we replace step (5) in version 1.1
by the following three new steps:

(5.1) Consider the first column of the graph.

(5.2) For each element of the first column, remove all of its multiples and count
the choices for the subgraph starting from the second column (the subgraph
can have unrelated classes, of any size, and strictly related classes; we
consider these facts in order to improve the performance).
(5.3) Sum the choices for every subgraph.

Further improvements were obtained considering that, for n ≥ 122, the main
graph is not anymore connected, then we can count separately the choices for
any connected component and multiply them. This is also true for subgraphs
considered in Step (5.2), even for lower values of n. Moreover, we can store data
about any connected component, because a graph can contain more components
with the same structure, especially small components, so they are studied only
once. In fact, once we considered the graph, one can ignore the numbers written
inside any vertex and focus only on its structure. The version 2.0 can calculate
a(2000) in less than 1 , a(3000) in about 6 and a(4000) in about 59 .
Now we show a table with the values of a(n) for n between 141 and 200. We
notice that a table for n ≤ 140 can be found in [5].

Table 1. The values of r, e3 , e5 , x and a(n), for 141 ≤ n ≤ 200.

n r e3 e5 x a(n)
141 23 6 1 2874768 87900275217530880
142 24 6 1 2874768 175800550435061760
143 24 6 1 2874768 175800550435061760
144 24 6 1 3535116 216182780193669120
145 25 6 1 3535116 432365560387338240
146 25 6 1 3535116 432365560387338240
147 24 7 1 3535116 648548340581007360
148 24 6 2 3535116 1080913900968345600
149 26 7 2 207948 762998047742361600
150 26 7 2 289731 1063074361717555200
151 27 7 2 289731 2126148723435110400
152 26 6 3 289731 1771790602862592000
153 26 6 2 2028117 2480506844007628800
154 27 6 2 2028117 4961013688015257600
155 27 6 2 2028117 4961013688015257600
156 27 6 2 2652153 6487479438173798400
(continued)
398 F. Caldarola et al.

Table 1. (continued)

n r e3 e5 x a(n)
157 28 6 2 2652153 12974958876347596800
158 29 6 2 1041250 10188081856512000000
159 28 7 2 1041250 15282122784768000000
160 28 7 2 1570324 23047187687748403200
161 28 7 2 1570324 23047187687748403200
162 28 7 2 2097494 30784308137636659200
163 29 7 2 2097494 61568616275273318400
164 28 6 3 2097494 51307180229394432000
165 27 7 3 2097494 76960770344091648000
166 28 7 3 2097494 153921540688183296000
167 30 7 2 2097494 123137232550546636800
168 30 7 2 2696778 158319298993559961600
169 31 7 2 2696778 316638597987119923200
170 30 7 2 5008302 294021555273754214400
171 30 7 1 35058114 411630177383255900160
172 30 6 2 35058114 686050295638759833600
173 30 6 2 35058114 686050295638759833600
174 29 7 2 35058114 1029075443458139750400
175 30 7 2 35058114 2058150886916279500800
176 31 7 2 10296468 1208946080003339059200
177 30 8 2 10296468 1813419120005008588800
178 31 8 2 10296468 3626838240010017177600
179 31 8 2 10296468 3626838240010017177600
180 31 8 2 12106458 4264390937307355545600
181 32 8 2 12106458 8528781874614711091200
182 32 8 2 12106458 8528781874614711091200
183 31 9 2 12106458 12793172811922066636800
184 31 8 3 12106458 21321954686536777728000
185 33 8 2 12106458 17057563749229422182400
186 32 9 2 12106458 25586345623844133273600
187 33 9 2 12106458 51172691247688266547200
188 32 8 3 12106458 42643909373073555456000
189 32 8 3 16351517 57596747872934559744000
190 32 8 3 30367103 106965388906878468096000
191 32 8 3 30367103 106965388906878468096000
192 32 8 3 38775698 136583908471798235136000
193 33 8 3 38775698 273167816943596470272000
194 35 8 2 38775698 218534253554877176217600
195 34 9 2 38775698 327801380332315764326400
196 34 8 2 190430552 536619592706793091891200
197 34 8 2 190430552 536619592706793091891200
198 34 8 2 266586008 751220187789456624844800
199 35 8 2 266586008 1502440375578913249689600
200 34 8 2 405038413 1141369102446680132812800
 Combinatorics on n-sets: Arithmetic Properties and Numerical Results 399

We see clearly from Table 1 that the sequence a(n) is not monotonic, in fact,
sometimes we have a(n) < a(n + 1), sometimes a(n) = a(n + 1) and other times
a(n) > a(n + 1). Moreover, the ratio a(n + 1)/a(n) seems to be bounded. After
a first look at the table, one can state the following conjecture.

Conjecture 1. Let q(n) := a(n + 1)/a(n), then:

(a) n ≡ 1 (mod 3) yields 1/2 < q(n) ≤ 1;

(b) n ≡ 2 (mod 3) yields 1 < q(n) ≤ 3/2;
(c) n ≡ 0 (mod 3) yields 7/5 < q(n) ≤ 2.

Now we show a table with some values of q(n) (Table 2).

Table 2. Some rounded values of q(n), with 1 ≤ n ≤ 200, divided according to the
congruence class of n mod. 3

n q(n) n q(n) n q(n)

1 1 2 1.5 3 1.6667
4 0.8 5 1.5 6 2
7 0.8333 8 1.4 9 1.8571
10 1 11 1.3077 12 2
13 0.7059 14 1.5 15 1.6667
16 1 17 1.4 18 2
19 0.7857 20 1.5 21 2
22 0.7879 23 1.3077 24 2
25 1 26 1.3529 27 1.6377
28 1 29 1.3982 30 2
31 0.6519 32 1.5 33 2
34 1 35 1.233 36 2
37 1 38 1.5 39 1.5354
40 0.6769 41 1.5 42 2
43 0.8333 44 1.3939 45 2
46 1 47 1.3043 48 2
49 0.74 50 1.5 51 1.6667
52 1 53 1.3514 54 2
55 0.8178 56 1.5 57 2
58 0.8 59 1.2446 60 2
61 1 62 1.3985 63 1.6284
64 1 65 1.5 66 2
67 0.5769 68 1.5 69 1.856
70 1 71 1.2328 72 2
73 1 74 1.5 75 1.6667
76 0.8 77 1.5 78 2
79 0.7676 80 1.3383 81 2
82 1 83 1.2876 84 2
85 0.8 86 1.5 87 1.6645
(continued)
400 F. Caldarola et al.

Table 2. (continued)

n q(n) n q(n) n q(n)

88 1 89 1.3038 90 2
91 0.8333 92 1.5 93 2
94 0.7024 95 1.2822 96 2
97 1 98 1.3999 99 1.5568
100 1 101 1.5 102 2
103 0.6665 104 1.5 105 2
106 1 107 1.1986 108 2
109 0.9286 110 1.5 111 1.6237
112 0.7866 113 1.5 114 2
115 0.8333 116 1.4 117 2
118 1 119 1.2228 120 2
121 0.6544 122 1.5 123 1.6667
124 1 125 1.3985 126 2
127 0.8142 128 1.5 129 1.8571
130 0.8 131 1.3077 132 2
133 1 134 1.3461 135 1.6667
136 1 137 1.5 138 2
139 0.6177 140 1.5 141 2
142 1 143 1.2297 144 2
145 1 146 1.5 147 1.6667
148 0.7059 149 1.3933 150 2
151 0.8333 152 1.4 153 2
154 1 155 1.3077 156 2
157 0.7852 158 1.5 159 1.5081
160 1 161 1.3357 162 2
163 0.8333 164 1.5 165 2
166 0.8 167 1.2857 168 2
169 0.9286 170 1.4 171 1.6667
172 1 173 1.5 174 2
175 0.5874 176 1.5 177 2
178 1 179 1.1758 180 2
181 1 182 1.5 183 1.6667
184 0.8 185 1.5 186 2
187 0.8333 188 1.3506 189 1.8571
190 1 191 1.2769 192 2
193 0.8 194 1.5 195 1.637
196 1 197 1.3999 198 2
199 0.7597 200 1.5
 Combinatorics on n-sets: Arithmetic Properties and Numerical Results 401

We can see that the upper inequalities in Conjecture 1 (a)–(c) become equal-
ities very often, while we can not say anything about lower bounds in each
congruence class mod. 3. In fact, by way of example, a previous version of this
conjecture stated that “if n ≡ 0 (mod 3) then 3/2 < q(n) ≤ 2”, but this is false
because, for instance, q(639) ≈ 1.4945 and q(1119) ≈ 1.4669.

Acknowledgments. This work is partially supported by the research projects

“IoT&B, Internet of Things and Blockchain”, CUP J48C17000230006, POR Calabria
FESR-FSE 2014-2020.

References
1. Aigner, M., Ziegler, G.M.: Proofs from the BOOK, 4th edn. Springer, Heidelberg
(2010). https://doi.org/10.1007/978-3-642-00856-6
2. The On-Line Encyclopedia of Integer Sequences. http://oeis.org/A174094. Accessed
3 June 2019
3. Liu, H., Pach, P.P., Palincza, R.: The number of maximum primitive sets of integers
(2018). arXiv:1805.06341
4. Vijay, S.: On large primitive subsets of 1,2,...,2n (2018). arXiv:1804.01740
5. Bindi, C., et al.: Su un risultato di uno studente di Erdős. Periodico di Matematiche,
Organo della Mathesis (Società italiana di scienze matematiche e fisiche fondata nel
1895), Serie 12, Anno CXXVI, N. 1 Gen-Apr 2016, vol. 8, pp. 79–88 (2016)
6. Bindi, C., Pellegrini, M., Pirillo, G.: On a result of a student of Erdős (submitted
for publication)
 Numerical Problems in XBRL Reports
and the Use of Blockchain as Trust
Enabler

Gianfranco d’Atri1,4 , Van Thanh Le2 , Dino Garrı̀3 , and Stella d’Atri4(B)
1
Department of Mathematics and Computer Science, University of Calabria,
Rende, Italy
[email protected]
2
Faculty of Computer Science, Free University of Bolzano, Bolzano, Italy
[email protected]
3
BlockchainLab, Milan, Italy
[email protected]
4
Blockchain Governance, Cosenza, Italy
[email protected]

Abstract. Financial statements are formal records of the financial

activities that companies use to provide an accurate picture of their
financial history. Their main purpose is to offer all the necessary data
for an accurate assessment of the economic situation of a company and its
ability to attract stakeholders. Our goal is to investigate how Benford’s
law can be used to detect fraud in a financial report with the support of
trustworthiness by blockchain.

Keywords: Financial models · Blockchain · Laws of Bradford and Zipf

1 Introduction
The eXtensible Business Reporting Language (XBRL) [9] is the world-leading
standard for business reporting, which opens a new era for report digitalization
that enables machines to read reports as humans. By scanning reports, we can
investigate financial aspects as well as doing forecasts for the organization, in
the paper, we will examine a well-known and proven Benford law, apply into an
XBRL report to review its consistency, also give recommendations for analyzers.
Traditionally, a company or an organization release their financial reports
based on their own standard and only for the internal usages, that results in
statistic issues for authorities, especially when governments would like to inves-
tigate their gross national metrics as GDP or GDI. Widely, exchanging informa-
tion in different reporting languages ruins business professionals over the world.
Therefore a new standard of XBRL is needed and could be spread for any kind
of report in any language. We will explore XBRL concept in details in next
subsections.
The work is the extension of the two conference papers [3,4] with the apply
of DLV. The DLV system is already used in a number of real-world applications
c Springer Nature Switzerland AG 2020
Y. D. Sergeyev and D. E. Kvasov (Eds.): NUMTA 2019, LNCS 11973, pp. 402–409, 2020.
https://doi.org/10.1007/978-3-030-39081-5_35
 Enabling Trust with Blockchain in XBRL Reports 403

including agent-based systems, information extraction, and text classification.

DLV is a rather solid candidate to constitute the basis of a system supporting
automatic validation of XBRL data in concrete application domains. An intro-
duction to DLV can be found in [5,8] and an in-depth description of OntoDLP
is contained in [14,15].

1.1 XBRL

Financial reports contain sensitive data that might have a huge impact on the
future of an organization in terms of investments and collaborations. This man-
dates careful management and control mechanisms able to capture any inconsis-
tencies or manipulation of the published reports. The first step towards this goal
started with the introduction of the eXtensible Business Reporting Language [9],
which is the world’s leading standard for financial reporting. It facilitates inter-
organization communication and enables automatic reports processing and anal-
ysis. XBRL was started from an XML standard from Charles Hoffman in 1998,
the version was developed and then the first international meeting for XBRL
was held in London, 2001 until now the recent release version is XBRL 2.0.
XBRL relies on XML and XML based schema to define all its constructs. Its
structure consists of two main parts:

1. XBRL instance, containing primarily the business facts being reported (see
Fig. 1).

Fig. 1. Facts

2. XBRL taxonomy, a collection of arcs which define metadata about these facts
and their relationship with other facts (see Fig. 2).

Fig. 2. XBRL Linkbase example

404 G. d’Atri et al.

1.2 Blockchain

Blockchain is a distributed, decentralized ledger that stores transactions and

it addresses the double-spending problem in a trust-less peer-to-peer network,
without the need for a trusted third party or an administrator. Blockchain is
maintained by a network of computers called nodes. Whenever a new transaction
arrives, the nodes verify its validity and broadcast it to the rest of the network.
The main building elements of a Blockchain are [11]:

– Transactions, which are signed pieces of information created by the partici-

pating nodes in the network then broadcast to the rest of the network.
– Blocks, that are collections of transactions that are appended to the
blockchain after being validated.
– A blockchain is a ledger of all the created blocks that make up the network.
– The blockchain relies on Public keys to connect the different blocks together
(similar to a linked list).
– A consensus mechanism is used to decide which blocks are added to the
blockchain.

Blockchain contains mathematical aspects that make it become trust and

stable over the network:

– Cryptographic Hash Functions: is a function H if and only if it is infeasible

to find two value x, y that x = y and H(x) = H(y)
– Hiding: the function H is hiding if y = H(r||x) with r is chosen from a
probability distribution, it is infeasible to find x with known value of y.
– Puzzle friendliness: with known value of y in y = H(r||x), the H is called
puzzle friendly if we have n bit of y, we can not find x in 2n times.
In order to build a blockchain-based solution for financial report, we need a
blockchain model supporting smart contract, there are several possible models,
such as Hyperledger Fabric [1], Bitcoin [10] or NEO [12], but we choose Ethereum
[6] as the best candidate, since it is an open-source blockchain platform that
enables developers to build and deploy decentralized applications, interaction
with web-based application is also much more easier than other models. The plat-
form runs smart contracts, a computer protocol running on top of a blockchain,
performing as a contract. Smart contracts can include data structures and func-
tion calls that are executed in a centralized fashion. This guarantees the fact
that the contract execution will persist on the chain.

1.3 Benford’s Law

Benford’s law was discovered by Simon Newcomb in 1881 and was published in
the American Journal of Mathematics [2], he observed that in logarithm tables
the earlier pages of library copies were much more worn than the other pages.
It is similar with the situation for scientists preferred using the table to look up
numbers which started with number one more than others.
 Enabling Trust with Blockchain in XBRL Reports 405

It has been shown that this result applies to a wide variety of data sets and
applications [13], including electricity bills, street addresses, stock prices... It
tends to be most accurate when values are distributed across multiple orders of
magnitude.
Using Benford’s law to detect fraud was investigated in [2]. Following the
law, a set of numbers is said to satisfy Benford’s law if the leading first digit d
(d ∈ 1, ..., 9) occurs with probability:
d+1 1
P (d) = log10 (d + 1) − log10 (d) = log10 ( = log10 (1 + )
d d
That presents for the distribution (Fig. 3):

Fig. 3. Benford’s law for the first digit

Besides this, the distribution for the second and following digits also are
demonstrated but the most well-known one is the first. The law has a brief
explanation in [7] where R. Fewster showed that: any number X could be written
by this way: X = r ∗ 10n . Where r is a real number with 1 ≤ r < 10 and n is an
integer. The leading digit of X is the same as r, the value will be 1 if and only
if 1 ≤ r < 2. We can isolate r by taking logs to base 10:
log10 X = log10 (r ∗ 10n ) = log10 (r) + n
The leading digit of X is 1 when 1 ≤ r < 2:
0 ≤ log10 (r) < log10 2 = 0.301 Thus n ≤ log10 (X) < n + 0.301
Benford laws are widely discussed in the academic area with its application.
Our contribution is threefold: (i) providing a methodology to automati-
cally evaluate and validate the consistency of the generated reports, (ii) to use
Blockchain to store information as an immutable and uninterruptible worldwide
database, (iii) to apply numerical tools to detect possible errors or even frauds
in the reports.
406 G. d’Atri et al.

For (i) we use ASP (answer set programming) to automatize reasoning on

the XBRL data. For (ii) we analyze the implementation of smart contracts on
the Ethereum Blockchain. For (iii) we explore the so-called Bradford’s and Zipf’s
laws as tools for checking numerical values in the specific domain of company
balance sheets.

2 Architecture Based Blockchain

As discussed in our previous work [4], the architecture composes three main
parts: XBRL Reader, XBRL Evaluator and XBRL Storage as Fig. 4.

Fig. 4. Architecture

– XBRLReader reads XBRL components and then convert them to DLVFact

and DLVArc
– XBRL Evaluator qualifies DLV objects above following metrics
– XBRL Storage will store essential result data to blockchain.

3 Numerical Methodology
In our scope, we use Benford’s law as a metric to make a recommendation for
the system management in case their dataset does not fit the law. The law will
be applied in different size of dataset to have a comprehensive view of analyzing.
Our queries are presented in Fig. 5.

Fig. 5. iDLV queries for Benford’s law

 Enabling Trust with Blockchain in XBRL Reports 407

3.1 Experiment Environment

These queries are executed in our dataset with 122 Financial reports of Ital-
ian companies from 2000 to 2015 and provided by a top provider of financial
information in Italy.
With evaluation aspects related to the law:
– All first digits in all reports
– All first digits from conto economico group
– All first digits from stato patrimoniale group
– All first digits for individual reports
For the group of all report and individual reports, it facilitate us to have a
view from general to detail, the other two groups presented for two major and
indispensable parts in an Italian financial report.

3.2 Numerical Results

We set scarto value as the mean of the differences between Benford’s law and
our actual results (Table 1 for the group of all financial reports)
 9
2
i=1 (Pi −Bi )
Scarto = 9

The scarto we got is 0.33 for all documents in our dataset, this value is accept-
able for an accountancy company. In reality, the auditors could set a baseline
for them to follow the value of scarto.

Table 1. Costs of smart contract functions execution

Number Count Percentage ( Pi ) Benford’s law( Bj ) Differences

9 750 4.72 4.6 0.12
8 806 5.08 5.1 −0.02
7 873 5.5 5.8 −0.3
6 985 6.21 6.7 −0.49
5 1303 8.21 7.9 0.31
4 1469 9.25 9.7 −0.45
3 2001 12.61 12.5 0.11
2 2820 17.76 17.6 0.16
1 4867 30.66 30.1 0.56
Negative 1731
Zero 4945
Scarto 0.33

Considering other evaluation aspects: conto economico group and stato

patrimoniale group (Fig. 6). The result is affected by the dataset, the smaller
dataset the higher scarto value we get.
408 G. d’Atri et al.

Fig. 6. conto economico and stato patrimoniale

– conto economico: Number of negative is 505, Number of Zero is 483, Scarto

is 0.96
– stato patrimoniale: Number of negative is 135, Number of Zero is 195, Scarto
is 1.33

The scarto value becomes even worse when we scale down the dataset for
each report. With each report, the value of scarto is commonly in the range of
5.0 to 9.0 with around 55 reports having the nearest values.

4 Conclusion

In the paper, we demonstrated an application of Benford’s law for financial

reports in XBRL format. The tests showed that the i-DLV tool is adaptable
for big data with a combination of complex calculations. For Benford’s law, it
is more sufficient with a big database and could become a useful technique for
accountants to do investigation for the whole dataset of a specific company. The
result from the law enables them to have a big picture of the data status and
could become a baseline for further analysis. However, with a small dataset,
we need other tools or other standards to qualify. Blockchain performs a trust
enabler role in the system that prevents insecurity behaviors and maintains
system consistency.

Acknowledgements. This research is partially supported by POR Calabria FESR-

FSE 2014-2020, research project IoT&B, CUP J48C17000230006.
 Enabling Trust with Blockchain in XBRL Reports 409

References
1. Cachin, C.: Architecture of the hyperledger blockchain fabric. In: Workshop on
Distributed Cryptocurrencies and Consensus Ledgers (2016)
2. Durtschi, C., Hillison, W., Pacini, C.: The effective use of benford’s law to assist
in detecting fraud in accounting data. J. Forensic Account. 5(1), 17–34 (2004)
3. D’Atri, G.: Logic-based consistency checking of XRBL instances. IJACT 3–6, 126–
131 (2014)
4. D’Atri, G., Le, V.T., Ioini, N.E., Pahl, C.: Towards trustworthy financial reports
using blockchain. Cloud Comput. 2019, 47 (2019)
5. Eiter, T., Faber, W., Leone, N., Pfeifer, G.: Declarative problem-solving using the
DLV system. In: Minker, J. (ed.) Logic-Based Artificial Intelligence. The Springer
International Series in Engineering and Computer Science, vol. 597, pp. 79–103.
Springer, Boston (2000). https://doi.org/10.1007/978-1-4615-1567-8 4
6. Ethereum Foundation: Ethereum (2013). www.ethereum.org. Accessed 10 Apr 2019
7. Fewster, R.: A simple explanation of Benford’s law. Am. Stat. 63, 26–32 (2009)
8. Gelfond, M., Lifschitz, V.: Classical negation in logic programs and disjunctive
databases. New Gener. Comput. 9(3), 365–385 (1991). https://doi.org/10.1007/
BF03037169
9. Malhotra, R., Garritt, F.: Extensible business reporting language: the future of
e-commerce-driven accounting. Int. J. Bus. 9(1) (2004)
10. Nakamoto, S., et al.: Bitcoin: a peer-to-peer electronic cash system (2008)
11. Narayanan, A., Bonneau, J., Felten, E., Miller, A., Goldfeder, S.: Bitcoin and
Cryptocurrency Technologies: A Comprehensive Introduction. Princeton Univer-
sity Press, Princeton (2016)
12. NEO: NEO Smart Economy (2013). https://docs.neo.org/docs/en-us/index.html.
Accessed 06 Oct 2019
13. Nigrini, M.J.: Benford’s Law: Applications for Forensic Accounting, Auditing, and
Fraud Detection, vol. 586. Wiley, Hoboken (2012)
14. Ricca, F., Gallucci, L., Schindlauer, R., Dell’Armi, T., Grasso, G., Leone, N.:
Ontodlv: an ASP-based system for enterprise ontologies. J. Logic Comput. 19,
643–670 (2009)
15. Ricca, F., Leone, N.: Disjunctive logic programming with types and objects: the
DLV+ system. J. Appl. Logic 5, 545–573 (2007)
 Modelling on Human Intelligence
a Machine Learning System

Michela De Pietro(B) , Francesca Bertacchini , Pietro Pantano ,

and Eleonora Bilotta

University of Calabria, via Pietro Bucci, 17b, Arcavacata di Rende, Italy

{michela.depietro,francesca.bertacchini,pietro.pantano,
eleonora.bilotta}@unical.it
http://www.unical.it

Abstract. Recently, a huge set of systems, devoted to emotions recog-

nition has been built, especially due to its application in many work
domains, with the aims to understand human behaviour and to embody
this knowledge into human-computer interaction or human-robot interac-
tion. The recognition of human expressions is a very complex problem for
artificial systems, caused by the extreme elusiveness of the phenomenon
that, starting from six basic emotions, creates a series of intermediate
variations, difficult to recognize by an artificial system. To overcome these
difficulties, and expand artificial knowledge, a Machine Learning (ML)
system has been designed with the specific aim to develop a recognition
system modelled on human cognitive functions. Cohn-Kanade database
images was used as data set. After training the ML, it was tested on a
representative sample of unstructured data. The aim is to make com-
putational algorithms more and more efficient in recognizing emotional
expressions in the faces of human subjects.

Keywords: Machine learning · Artificial intelligence · Emotion

recognition

1 Introduction
The face is the first instrument of non-verbal communication and is the first
interaction between human beings. The face changes based on what a person
feels at a given moment. Starting from the slightest change in the facial mus-
cles and continuing to change until emotion is expressed. This change provides
information about the emotional state of a person [1].
When we talk about the analysis of facial expressions, we refer to the recogni-
tion of face, the different facial movements and the changes in the face. Emotion
is often expressed through subtle changes in facial features, such as in stiffening
the lips when a person is angry or in lowering the corners of the lips when a
person is sad [2] or again in the different change of eyebrows or of eyelids [3].
Facial expressions are an important instrument in non-verbal communica-
tion. The role of the facial expressions’ classification could be a helpful used in
c Springer Nature Switzerland AG 2020
Y. D. Sergeyev and D. E. Kvasov (Eds.): NUMTA 2019, LNCS 11973, pp. 410–424, 2020.
https://doi.org/10.1007/978-3-030-39081-5_36
 Modelling on Human Intelligence a Machine Learning System 411

behavioural analysis [1]. For effective human-computer interaction, automated

analysis of facial expression is important.
Ekman & Friesen devised the coding of actions of the human face system,
called FACS [4]. FACS allows different facial movements to be coded in Action
Units (AU) based on the muscular activity of the face that generates temporary
changes in facial expression. They identified 44 FACS AUs, of which 30 AUs
are related to the contractions of specific facial muscles: 12 for the upper face
and 18 for the lower face. AUs can happen singly or in combination [5]. The
face movement takes place in a space and takes into consideration landmarks
points of a face. The dynamics of change in this space can reveal emotions, pain
and cognitive states and regulate social interaction. The movement of landmark
points of a face, such as the corners of the mouth and eyes, constitutes a ‘reference
point space’ [6].
A lot of researchers [7,8] have used different neural networks for facial expres-
sion analysis. The performance of a neural network depends on several factors
including the initial random weights, the training data, the activation function
used, and the structure of the network including the number of hidden layer
neurons, etc.
The main purpose of this study is to create a system of recognition of facial
emotions in some images. The work is carried out using a computational algo-
rithm capable of performing an automatic recognition job pre-training the sys-
tem using a set of data and a set of tests. In particular, the system has been
trained in trying to recognise emotions already from the first change in the facial
muscles in the subjects, that is at the onset of expression. In this way the com-
putational algorithms become more and more efficient in recognizing emotional
expressions in the faces of human subjects and they are being trained to the
complexity of the dynamics of emotional expressions.

2 Background

Since Darwin’s revolutionary book [9] on the expression of the emotions, neu-
rophysiological and cognitive theories advanced the processes of recognition of
emotions in human subjects. The peripheral theory of James [10] and the central
theory of Cannon [11], with different visions, belong to the first strand. Accord-
ing to James, an emotional response occurs when there is an emotional stimulus
that can be internal or external to our body. Both the emotional stimulus and
the information that it brings with it are perceived by the sense organs, thus
causing a neurogenerative modification and, only later, in the central nervous
system, the processing at the cortical level is activated, becoming an emotionally
felt event.
According to Cannon [11] emotions originate from the central nervous sys-
tem and, later, they reach the periphery. Then they are displayed through the
expression of emotions, revealed by the face and the whole body. The coding of
information takes place in the central nervous system, precisely in the thalamus,
the centre for gathering, processing and sorting information from the centre to
412 M. De Pietro et al.

the periphery and from the periphery to the centre and the amygdala, a struc-
ture located in the brain, considered as an “emotional computer”, which allows
the sorting and recognition of emotions. Patients with amygdala damage present
difficulties in recognising emotions (whether pleasant or not) and abnormalities
in emotional behaviour [12]. The amygdala is therefore recognised as the centre
for the exchange of emotional messages [13]. Starting from the seventies, cogni-
tive theories introduce the psychological element, considered the most relevant
aspect in emotional research. Both the physiological and psychological compo-
nents, the latter defined arousal, interact with each other, causing the emotional
event.
According to the cognitive-activation theory of Schachter and Singer [8], by
assigning to the emotional state a specific emotional event, our brain succeeds
in contextualising its own emotional experience [14]. A further contribution is
provided by Lazarus [15], with the introduction of the concept of appraisal, the
process that allows the cognitive processing of the event. Indeed, as Frijda [16]
will extend later, emotions originate from the causal evaluation of the events,
related to the meanings that people attach to the event itself. They do not appear
by chance, but from an in-depth analysis of the situations that arise, determining
the influence that emotions can ultimately have on a person’s well-being [17].
Events that meet expectations and desires enable positive emotions; on the con-
trary, events that can harm, activate negative emotions. For this reason, each
individual, based on his/her experience, may experience the same situations
or events expressing different and sometimes divergent emotions. Emotions are
therefore placed in a subjective dimension and are very flexible and variable.
Some studies talk about the 3D reconstruction [18], they tell about the
expression of a face can be reconstructed with 3D techniques that allow to sim-
ulate and help to understand the emotions present on a face [19,20]. Often, the
psychological conflict is represented by the division and multiplication of the
inner characters of a subject and the images of “strange attractors” [21,22],
which represent chaos [23–25]. Each cerebral hemisphere maintains a network
without scales that generates and maintains a global state of chaos [26]. Some-
times chaos can be approached in cases of people suffering from neurological
and psychological disorders, often in these people it is difficult to recognise the
emotions they express [27]. Numerous studies show that 3D virtual learning envi-
ronments and information communication technologies can be excellent tools to
support people, for example, suffering from autism spectrum disorders [28,29]. In
particular, they can be very effective in helping to understand social behaviours
and emotions [30].
Very significant advances in the study of emotions and subsequent models
of interpretation in computer vision were obtained thanks to the FACS (Facial
Action Coding System) coding system, devised by Ekman & Friesen [4]. Accord-
ing to this system, emotional expressions are expressed from the face muscular
contraction according to specific organisations. The FACS system makes it pos-
sible to identify the individual muscles (called Action Units) (and their combina-
tion in emotional units), involved in all the emotional expressions that humans
 Modelling on Human Intelligence a Machine Learning System 413

produce. Thus, the movements of one or more facial muscles determine the six
expressions recognised as universal, that are neutral, happy, sad, angry, disgust-
ing and surprise facial expressions. These movements play a very important role
in conveying the individual’s emotional states to the observer, especially in face-
to-face social interaction. The research has developed different approaches and
methods for the analysis of fully automatic facial expressions, useful in human-
computer interaction or computer-robotic systems [31–33]. In automatic recog-
nition, the facial expression of the image is processed to extract this information
from it, which can help to recognise the six basic expressions. The steps of the
process are as follows: image acquisition, extraction of basic features, and finally
classification of expressions.

3 Materials and Methods

3.1 Facial Image Database

In this research we used is the Cohn-Kanade database which contains 486

sequences of 97 subjects extended in the 2010 with the name CK+ database
that added 107 sequences of other 26 subjects, with a total of 593 sequences
from 123 subjects [34]. Each sequence contains images ranging from onset (neu-
tral frame) to peak expression (last frame). The peak frame belongs to the
FACS coded for the facial action unit (AUs). In particular, the FACS system,
that was designed by Ekman and Friesen [4], devised the coding of the actions
of the human face system. FACS allows different facial movements to be coded
in Action Units (AU) based on the muscular activity of the face that generates
temporary changes in facial expression. They identified 44 FACS AUs, of which
30 AUs are related to the contractions of specific facial muscles: 12 for the upper
face and 18 for the lower face. AUs can happen singly or in combination [35].
At all subjects in CK database were asked to perform a series of expressions,
with a neutral background behind them and only one person was present in
the diagram [36]. Image sequences have been digitised into 640 arrays of 480
or 490 pixels with 8-bit precision for gray-scale values. In order for the internal
validity of the data to be guaranteed, they have been manually coded and the
reliability of the coding has subsequently been verified [34]. Only 118 subjects
are annotated with the principal emotions (anger, disgust, fear, happy, sad and
surprise) [35].
The database was composed by subjects were between 18 and 50 years, 69%
women, 81%, Euro-American, 13% African Americans and 6% other groups.
Their facial behaviour was recorded using two Panasonic AG-7500 cameras syn-
chronised to the hardware.
Images that were not labelled by the original CK database system and CK+
were deleted and were not used in this study.
The Dataset of this study is composed by 1544 photos (Fig. 1).
414 M. De Pietro et al.

Fig. 1. Part of dataset.

3.2 Mathematica Wolfram Machine Learning

To analyse data, we applied Wolfram Mathematica software that use its own
computer language, called Wolfram Language.
Wolfram Mathematica is able to learn what to do only by looking at exam-
ples, using the idea of machine learning. In particular, it is possible to construct
algorithms to analyse data or images automatically [36]. There are two types of
machine learning: supervised and unsupervised learning. Supervised learning is
defined as the process of an algorithm that learns from the training data set. The
algorithm formulates forecasts on the control set based on what was described
by the supervisor. The unsupervised learning algorithms, instead, model the
structure of the input data or their distribution to obtain other information on
the data. Wolfram Mathematica includes several types of integrated automatic
learning. In particular, functions like Predict and Classify are automated and
work on many types of data, including numerical, categorical, time series, text,
images and audio [36]. In this study, Classify supervised automatic learning was
used.
In our work, we processed image using automatic machine learning feature
provided by the software. The neural network to which Wolfram Mathematica
refers is the LeNet Network. The software allows you to immediately use the
available pre-trained networks or manipulate and reassemble them to train new
data. Indeed, through the use of Wolfram language, it is possible to implement
the automatic learning system of the neural network contained in it.
 Modelling on Human Intelligence a Machine Learning System 415

3.3 Image Processing

A research proposal is to ensure that the software analyse the images and auto-
matically recognises the emotions expressed in the image.
To understand how the software takes out the emotion in the images, we must
refer to the slight changes of face or, in particular, to the changes of landmarks.
Indeed, subjects modify their facial expression by moving the fundamental char-
acteristics of their faces, that is, by modifying the position of the reference points.
The fundamental facial points are used to locate and to represent the salient
regions of face, such as:

– Eyes (right and left);

– Eyebrows (right and left);
– Nose;
– Mouth (internal and external);
– Outline of the face.

These points are starting points for the software’s automatic recognition of
emotions. Wolfram Mathematica extracts all parameters of the reference points
(Fig. 2).

Fig. 2. Different colours of different landmarks. (Color figure online)

The first step was to manipulate and reassemble the pre-trained networks
available in the Wolfram Mathematica software. This in order to create a more
solid and functional network by training the machine with database described
above.
To implement the training of the machine, the images corresponded a par-
ticular emotion have been assigned to each emotion, taking the images of each
subject from the onset to the peak of emotion (excluding neutral emotion). In
this way the software learned to understand the corresponding emotion already
from the first variation of the landmarks in the subjects’ faces (Fig. 3).
To start recognition process we applied FindFaces module in Wolfram Math-
ematica that allows to find people’s faces in the images proposed and returns
a list of bounding boxes. In addition, we use Classify command was assigned
416 M. De Pietro et al.

Fig. 3. Change of landmarks with variation of emotion.

to each emotion the corresponding package of photos (Fig. 1), manipulating and
reassembling the pre-trained Mathematica’s networks. Classify generates Clas-
sifierFunction based on an association of classes with their examples. From here
it is possible to extract information about the method used by the calculator to
classify emotions. The same procedure was carried out for all six emotions.

3.4 Image Analysis

The pre-trained Wolfram Mathematica network has been increased using a total
of 1544 samples adopted for machine training. The machine took about 1 min
to do the learning, having 6 classes to learn from (Fig. 4). In this case Wolfram
Mathematica Machine Learning used the “Forecasting method” for its learn-
ing. This method, as defined by Wolfram [36], predicts the value or class of an
example using a set of decision trees. It is a type of machine learning algorithm
named Bootstrap Aggregation or Bagging (Bootstrap Aggregation Algorithm).
The Bootstrap method is used for estimating statistical quantities from sam-
ples and creates models from a single set of training data. In particular, it is
a learning method for classification and regression that operates by building a
multitude of decision trees. The forest forecast is obtained by taking the most
common class or the tree’s predictions of the average value. Each decision tree
is then trained on a random subset of the training set [36].
The total accuracy of the method is 91.3% (Fig. 5); the accuracy of individual
emotions is high for some emotions (89% happiness, 81% disgust), while low due
to negative emotions (0.02% anger, 27% fear) (Fig. 6).
The last step was to test the trained machine to verify its validity. Having
chosen a group of unused photos in the initial database, named Test Dataset.
The Test Dataset is composed by 37 images for each emotion. The FacialEx-
pression module was used, which allows to recognise the expression of a face
displayed in an image. FacialExpression has an integrated classifier that auto-
matically chooses the best algorithm to adopt based on the available models
built and trained through a large database called Wolfram’s Net Neural Repos-
itory, an archive consisting of 70 neural network models [36]. In particular, the
output is a prediction of what the neural network interprets or returns the most
probable alternative to the results, using the trained network in step 2. In this
way using the network trained in step 2 the Mathematica Machine Learning
identifies the emotions included in each image and the relative probability of
correct identification of the expressed emotion.
 Modelling on Human Intelligence a Machine Learning System 417

Fig. 4. Information about training.

Fig. 5. Method accuracy. Fig. 6. Classes accuracy.

418 M. De Pietro et al.

4 Results
Emotion detection is provided in Fig. 7 and it can be seen that Disgust, Happi-
ness and Surprise perform better than the other emotions.

Fig. 7. Confusion matrix for every emotion.

This result is common and intuitive as these 3 are particular emotions that
cause great deformation in the facial features. Indeed, the movement of the
facial area for these emotions was easily detected by the trained system. On
the contrary, other emotions (i.e. Anger, Sadness and Fear) do not behave as
well, as the movements of the face are less than those of positive emotions. An
explanation of this arises from the fact that emotions like Rage and Fear have
 Modelling on Human Intelligence a Machine Learning System 419

subtle movements and are easily confused with other stronger emotions. This is
also confirmed by other studies such as Lucey, Cohn, Kanade [5].
The machine has now been tested on a single subject to further verify its
validity. The results obtained are discrete; for each subject the right emotion
is recognised. From here it can be stated that the method works well on small
image tests. In particular, in Fig. 8 the recognition of the fear emotion is fulfilled,
in Fig. 9 the recognition of the disgust emotion, in Fig. 10 the recognition of the
fear emotion.

Fig. 8. Recognition of anger emotion in one subject.

Fig. 9. Recognition of disgust emotion in one subject.

Fig. 10. Recognition of fear emotion in one subject.

Figure 11 shows the recognition of the happiness emotion, Fig. 12 the recog-
nition of the sad emotion, Fig. 13 the recognition of the surprised emotion. For
each subject the emotion is already recognised at the onset of emotion.
Now, Mathematica Machine Learning has been asked to provide information
about a single emotion, taken happiness as an example, it displays the progress
420 M. De Pietro et al.

Fig. 11. Recognition of happiness emotion in one subject.

Fig. 12. Recognition of sadness emotion in one subject.

Fig. 13. Recognition of surprise emotion in one subject.

 Modelling on Human Intelligence a Machine Learning System 421

of the percentage of recognition of the happiness. As shown in Fig. 14, the recog-
nition rate is low for the first images, i.e. the onset of emotion, this is because
at first a few facial muscles move slowly and the emotion can be confused with
another emotion (such as example surprise), the percentage of recognition how-
ever goes back up to about 40% when the emotion is actually recognised.

Fig. 14. Recognition of surprise emotion in one subject.

5 Conclusion
In this work, the recognition of emotions using Wolfram Mathematica software
has been realised.
The reasons that led to this work concern:

a. making computational algorithms more and more efficient in recognising emo-

tional expressions in the faces of human subjects;
b. train algorithms to the complexity of the dynamics of emotional expressions.

In fact, to overcome the traditional stillness of the data, usually used in

this sector, we have used the different articulations of the onset of emotional
behaviour, until its complete appearance in the face of the subject, up to the
disappearance to allow for other types of expressions. This allowed us to train a
computational machine first on static faces and then to train it precisely on the
expressive dynamics, using a specially organised experimental set.
The results show that:
– The system is able to recognise the six main facial expressions used in an
extremely efficient way;
422 M. De Pietro et al.

– Recognises with reliable percentages almost all the emotions expressed in

dynamic form, even if the recognition percentages are fluctuating and do not
progress linearly from the onset of the expression until the return of the facial
muscles to the neutral expression;
– It does not properly recognise the facial expressions related to the anger
emulation: in this case the performance of the computational system is very
low.

Evolutionary explanations of why this emotion that was not adequately

recognised by the system may reside in the fact that, according to some authors,
anger is not functional to social relations, so it may not be adaptive from the
Darwinian point of view or because it is very insignificant from the point of view
of the facial dynamics that this emotion realises, being able to manifest even only
with the look and leaving the rest of the facial muscles completely immobile.
Further system training and data set choices focused on this emotion could
give us different results, which we will pursue in other experiments.

References
1. Kulkarni, S.S., Reddy, N.P., Hariharan, S.I.: Facial expression (mood) recognition
from facial images using committee neural networks. Biomed. Eng. Online 8(1),
16 (2009)
2. Carroll, J.M., Russell, J.A.: Facial expressions in Hollywood’s protrayal of emotion.
J. Pers. Soc. Psychol. 72(1), 164 (1997)
3. Eibl-Eibesfeldt, I.: Human Ethology. Aldine de Gruvter, New York (1989)
4. Ekman, P., Friesen, W.V.: Facial Action Coding System: Investigator’s Guide.
Consulting Psychologists Press, Palo Alto (1978)
5. Lucey, P., Cohn, J.F., Kanade, T., Saragih, J., Ambadar, Z., Matthews, I.:
The extended Cohn-Kanade dataset (CK+): a complete dataset for action unit
and emotion-specified expression. In: Computer Society Conference on Computer
Vision and Pattern Recognition-Workshops, pp. 94–101. IEEE (2010)
6. Sebe, N., Lew, M.S., Sun, Y., Cohen, I., Gevers, T., Huang, T.S.: Authentic facial
expression analysis. Image Vis. Comput. 25(12), 1856–1863 (2007)
7. Kobayashi, H., Hara, F.: Facial interaction between animated 3D face robot and
human beings. In: International Conference on Systems, Man, and Cybernetics.
Computational Cybernetics and Simulation, vol. 4, pp. 3732–3737. IEEE, Orlando
(1997)
8. Schachter, S., Singer, J.: Cognitive, social, and physiological determinants of emo-
tional state. Psychol. Rev. 69(5), 379 (1962)
9. Darwin, C.: On the Origin of Species. D. Appleton & Company, New York (1860)
10. James, W.: Essays in Psychology, vol. 13. Harvard University Press, Cambridge
(1983)
11. Cannon, W.B.: The James-Lange theory of emotions: a critical examination and
an alternative theory. Am. J. Psychol. 39(1/4), 106–124 (1927)
12. Walker, D.L., Davis, M.: The role of amygdala glutamate receptors in fear learning,
fear-potentiated startle, and extinction. Pharmacol. Biochem. Behav. 71(3), 379–
392 (2002)
 Modelling on Human Intelligence a Machine Learning System 423

13. Doré, B.P., Tompson, S.H., O’Donnell, M.B., An, L.C., Strecher, V., Falk, E.B.:
Neural mechanisms of emotion regulation moderate the predictive value of affective
and value-related brain responses to persuasive messages. J. Neurosci. 39(7), 1293–
1300 (2019)
14. LeDoux, J.E., Hofmann, S.G.: The subjective experience of emotion: a fearful view.
Curr. Opin. Behav. Sci. 19, 67–72 (2018)
15. Lazarus, R.S.: Thoughts on the relations between emotion and cognition. Am.
Psychol. 37(9), 1019 (1982)
16. Frijda, N.H.: Appraisal and beyond. Cogn. Emot. 7(3–4), 225–231 (1993)
17. Bourgais, M., Taillandier, P., Vercouter, L., Adam, C.: Emotion modeling in social
simulation: a survey. J. Artif. Soc. Soc. Simul. 21(2), 1–5 (2018)
18. Wibowo, H., Firdausi, F., Suharso, W., Kusuma, W.A., Harmanto, D.: Facial
expression recognition of 3D image using facial action coding system (FACS).
Telkomnika 17(2), 628–636 (2019)
19. Gabriele, L., Bertacchini, F., Tavernise, A., Vaca-Cárdenas, L., Pantano, P.,
Bilotta, E.: Lesson planning by computational thinking skills in Italian pre-service
teachers. Inform. Educ. 18(1), 69–104 (2019)
20. Gabriele, L., Tavernise, A., Bertacchini, F.: Active learning in a robotics labora-
tory with university students. In: Increasing Student Engagement and Retention
Using Immersive Interfaces: Virtual Worlds, Gaming, and Simulation, pp. 315–339.
Emerald Group Publishing Limited (2012)
21. Abdechiri, M., Faez, K., Amindavar, H., Bilotta, E.: The chaotic dynamics of high-
dimensional systems. Nonlinear Dyn. 87(4), 2597–2610 (2017)
22. Freeman, W.J.: Emotion is from preparatory brain chaos; irrational action is from
premature closure. Behav. Brain Sci. 28(2), 204–205 (2005)
23. Bilotta, E., Bossio, E., Pantano, P.: Chaos at school: Chua’s circuit for students in
junior and senior high school. Int. J. Bifurcat. Chaos 20(01), 1–28 (2010)
24. Bilotta, E., Pantano, P., Vena, S.: Artificial micro-worlds part I: a new approach
for studying life-like phenomena. Int. J. Bifurcat. Chaos 21(2), 373–398 (2011)
25. Bilotta, E., Pantano, P.: Artificial micro-worlds part II: cellular automata growth
dynamics. Int. J. Bifurcat. Chaos 21(03), 619–645 (2011)
26. Bilotta, E., Stranges, F., Pantano, P.: A gallery of Chua attractors: part III. Int.
J. Bifurcat. Chaos 17(3), 657–734 (2007)
27. Lorincz, A., Jeni, L., Szabo, Z., Cohn, J., Kanade, T.: Emotional expression clas-
sification using time-series kernels. In: Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition Workshops, pp. 889–895. IEEE (2013)
28. Bertacchini, F., et al.: An emotional learning environment for subjects with
Autism Spectrum Disorder. In: International Conference on Interactive Collab-
orative Learning (ICL), pp. 653–659. IEEE (2013)
29. Cárdenas, L.A.V., et al.: An educational coding laboratory for elementary pre-
service teachers: a qualitative approach. Int. J. Eng. Pedagogy (iJEP) 6(1), 11–17
(2016)
30. Bertacchini, F., Bilotta, E., Gabriele, L., Pantano, P., Servidio, R.: Using Lego
MindStorms in higher education: cognitive strategies in programming a quadruped
robot. In: Workshop Proceedings of the 18th International Conference on Comput-
ers in Education, pp. 366–371 (2010)
31. Bertacchini, F., Tavernise, A.: Knowledge sharing for cultural heritage 2.0: pro-
sumers in a digital agora. Int. J. Virtual Communities Soc. Netw. 6(2), 24–36
(2014)
32. Bertacchini, F., Bilotta, E., Pantano, P.: Shopping with a robotic companion. Com-
put. Hum. Behav. 77, 382–395 (2017)
424 M. De Pietro et al.

33. Mane, S., Shah, G.: Facial recognition, expression recognition, and gender identi-
fication. In: Balas, V.E., Sharma, N., Chakrabarti, A. (eds.) Data Management,
Analytics and Innovation. AISC, vol. 808, pp. 275–290. Springer, Singapore (2019).
https://doi.org/10.1007/978-981-13-1402-5 21
34. Kanade, T., Cohn, J.F., Tian, Y.: Comprehensive database for facial expression
analysis. In: Proceedings Fourth IEEE International Conference on Automatic Face
and Gesture Recognition, pp. 46–53. IEEE (2000)
35. Tian, Y.I., Kanade, T., Cohn, J.F.: Recognizing action units for facial expression
analysis. IEEE Trans. Pattern Anal. Mach. Intell. 23(2), 97–115 (2001)
36. Wolfram, S.: An Elementary Introduction to the Wolfram Language. Wolfram
Media, Incorporated, Champaign (2017)
 Algorithms for Jewelry Industry 4.0

Francesco Demarco1(B) , Francesca Bertacchini2 , Carmelo Scuro1 ,

Eleonora Bilotta1 , and Pietro Pantano1
1
Department of Physics, University of Calabria, Rende, Italy
{francesco.demarco,carmelo.scuro,pietro.pantano}@unical.it
2
Department of Mechanical, Energy and Management Engineering,
University of Calabria, Rende, Italy
[email protected]

Abstract. The industrial and technological revolution and the use of

innovative software allowed to build a virtual world from which we
can control the physical one. In particular, this development provided
relevant benefits in the field of jewelry manufacturing industry using
parametric modeling systems. This paper proposes a parametric design
method to improve smart manufacturing in 4.0 jewelry industry. By using
constrained collection of schemata, the so called Direct Acyclic Graphs
(DAGs) and additive manufacturing technologies, we created a process
by which customers are able to modify 3D virtual models and to visu-
alize them, according to their preferences. In fact, by using the software
packages Mathematica and Grasshopper, we exploited both the huge
quantity of mathematical patterns (such as curves and knots), and the
parametric space of these structures. A generic DAG, grouped into a unit
called User Object, is a design tools shifting the focus from final shape
to digital process. For this reason, it is capable to returns a huge number
of unique combinations of the starting configurations, according to the
customers preferences. The configurations chosen by the designer or by
the customers, are 3D printed in wax-based resins and, later, ready to be
merged, according to artisan jewelry handcraft. Two cases studio are pro-
posed to show empirical evidences of the designed process to transform
abstract mathematical equations into real physical forms.

Keywords: Parametric jewelry · Algorithm for jewelry · Smart

manufacturing

1 Introduction
Since their introduction CAD (Computer Aided Design) software has expanded
to all fields of design including jewelry one. CAD system allows designers to gen-
erate and display extremely complex objects and they are generally employed in
detailed design stage rather than in conceptual design stage. It’s well estab-
lished that most of these systems cannot allow design exploration because
shape transformation and reinterpretations aren’t supported. So, the concep-
tual design stage, in which designers generates ideas and explore possibilities, is
c Springer Nature Switzerland AG 2020
Y. D. Sergeyev and D. E. Kvasov (Eds.): NUMTA 2019, LNCS 11973, pp. 425–436, 2020.
https://doi.org/10.1007/978-3-030-39081-5_37
426 F. Demarco et al.

often assigned to hand sketches. However, these systems Parametric Design (PD)
are emerging as a distinctive method of design. PD is a tool involving an explicit
visual dataflow in the form of a graph and it supports the designers by offering a
wide design exploration. This dataflow model a design as a constrained collection
of schemata [1] in the form of a graph so called Direct Acyclic Graph (DAG).
Parametric design aid human creativity by providing the opportunity to explore
a larger range of design possibilities than in manual production [4], design varia-
tions can be made simply by adjusting the input parameters. In the recent years,
parametric modeling has become a widespread tool in computational designer
community, basically, two main approaches have been developed for this tool.
First approach, adopted in the present case, is based on a visual explicit dataflow
program called Directed Acyclic Graph (DAG). While the second approach is
an implicit bottom-up method based on cognitive natural system in the form
of heuristic algorithm. In the first phase of the development, parametric design
found its application in the design of large-scale object and therefore in architec-
ture design. Now, the technological development with the consequent introduc-
tion of the numerical control machine and the 3D printer strictly linked to the
development of tools and software applications allowing the physicalizations of
small-scale objects. These innovations lend designers to research more complex
ideas, parametric design has had an impact not only on formal characteristic but
also influences the logic of approach to design, indeed, we can talk about a new
paradigm of design thinking [5]. Nowadays, the parametric modelling is present
in every design area and its use increases day after day. Therefore, we can assist
at a continuous metamorphosis and redefinition of parametric design theories
and, simultaneously, at the development of parametric tools and machineries.
This paper introduces an approach to parametric design thinking oriented
to the creation of jewels inspired by mathematical shapes, Mathematics can
offer a huge library of aesthetic patterns, that integrated with algorithmic logic,
adapt well to applications in art and design. The advantages of this approach
are demonstrated using two example of real parametric modeling that maintains
the benefit of design exploration until the detailed design stage, this allows final
customer to explore different combination and personalizes the model. The paper
is organized in five main section. Section 2 provide a brief review on the related
works. Section 3 described the methodology adopted for this research. Section 4
provide the experimental results and their discussion. The conclusions of this
work are presented in Sect. 5.

2 Literature Review

Parametric design has found its maximum applications in the industry 4.0 para-
digm, it represents the systemic transformation of production thanks to the
combination of physical-digital system. This industrial revolution is linked to
the technological phenomenon of digitalization that allows us to build a virtual
world from which we can control the physical one. In this paradigm virtual sim-
ulations, IOT strategies [34,35] and additive manufacturing technologies [44,45]
 Algorithms for Jewelry Industry 4.0 427

guide product development processes allowing the study and realization of com-
plex objects, a complexity that can only be achieved through parametric model-
ing processes. “Parametric design is about to change”[6] this is the simplest and
most effective description given to parametric design, it allows to overcome the
static design system. In literature, we can find several researches and different
new approaches and theories, however, an approach considered more appropri-
ate and efficacy than other does not exist. The scientific community is still
engaged in research for the definition of a general theory of parametric design
thinking [5,6,9], theoreticians are defining and characterizing the formalizations
of parametric design processes and their related key concepts [7,41]. Research
developed towards both parametric design approaches generative design (GD)
and graph based [1,31]. GD is a method that uses computational capabilities to
support design. GD, mainly diffused among the scientific community, uses Five
techniques that are spread in all fields of design. Those techniques have been
described by Singh et al. [8,9] and those are: cellular automata, shape gram-
mar, L-systems, swarm intelligence and genetic algorithms. Several researchers
have studied various issues of generative design. Bilotta et al. propose a design
based on cellular automata, to represent chaos and hyperchaos object [10–29].
There are few publications on parametric design system related to jewelry appli-
cations. An example is the work of Kielarova et al. who proposed an approach
on generative design based on a shape grammar applied to jewelry ring design
[2]. It also provides an effective description of the designer’s main aim who is the
development of a generative design system that applies affine transformations to
original elements and create new 3D shapes based on a finite set of shape rules.
Another example is the work of Sansri et al. [30] they studied the applications
of a genetic algorithm to an Art Deco double clip broach.

2.1 Lack of Literature on Design and Industry Interaction

If it is true that there is a broad scientific discussion on parametric modeling soft-
ware, such as Generative Components (Bentley Systems), Dynamo (Autodesk)
and Grasshopper (McNeel and Associates) [30], the analysis of the sources
revealed a complete lack of paper dealing with the design theme in the paradigm
of industry 4.0 [44]. The presence of this connection is demonstrated by the
development of tools and software applications designed to directly connect the
virtual world to the productive field. These processes, which have been kept as
industrial secrets, are beginning to impact not only the productive system but
also the society on a large scale. An important example of this phenomenon is
Shapewa is a platform designed for direct physicalization of virtual model, any-
one using this platform can become a designer and physicalize a unique object.
Another example is Shapediver a tool that automatically turns parametric CAD
file into interactive 3D model accessible through any web browser. This tool
allows companies to provide a product with a high rate of customization by
keeping the DAG hidden and eliminating high application development costs.
The cause of the lack of scientific production on this theme is probably to be
found on its multidisciplinary nature, indeed, the two topics are part of two
428 F. Demarco et al.

distinct fields of research. Parametric design research is mainly carried out by

researchers from architecture and civil engineering departments [36–40], while
industry 4.0 is mainly treated by completely different classes of researchers [42–
46]. Therefore, this article introduces itself in a multidisciplinary sector with
the main purpose of creating a scientific discussion on new production processes
starting from design up to the finished object.

3 Methodology
Mathematics leads man’s thinking to the limit and becomes a source of inspira-
tion for design research, there is an infinite catalog of mathematical shapes and
models that can be adopted in jewelry design. Jewelry design, as any other indus-
trial design process, need to consider the balance of aesthetic and functionality,
therefore, for the correct design of an object it is not enough to create an aesthet-
ically attractive shape, but it is necessary to take into account many ergonomic
parameters. This last process implies not only the generation of inferences and
judgements, but also the planning of actions in order to act on shapes and trans-
forming them. The aim of a designer is the individuating of users’ needs, think-
ing about the functions that fit well with these needs, for creating objects that
are capable of embodying these functions by using some formal properties [10].
Explicit Direct Acyclic Graph generated in parametric design acts as a cognitive
artifact shifting the focus from the final form to the digital process, [3,30–33]
or rather it acts as a generative algorithm. Building the structure of the DAG
designers define the geometric constrain and the relationship between the parts,
this involves the creation of a shape grammar. Most of the resulting rules in this
grammar are based on affine transformations, such as, translation, scaling, rota-
tion, and repetition [1]. The resulting constrained collection of schemata allows
to explore the design space when input parameters are changed in real-time [38].
In this case study, we started from analyzing some mathematical shapes and

Fig. 1. A schematic representation of the adopted methodology: DAG for jewelry

design. (Color figure online)
 Algorithms for Jewelry Industry 4.0 429

we have chosen a couple with completely different characteristics. Then, shape

transformations in jewelry design process were studied to identify how shapes are
transformed from one state to another with the aim of guaranteeing aesthetics
and functionality, this involves creativity, analysis and development. In the two
examples, presented in the fourth section, has been developed the methodology
described in Fig. 1. This methodology corresponds to the creation of DAG devel-
oped in Grasshopper with the aid of a series of extensions such as peacock and
lunchbox. Each of the graph-based models can be divided into six parts, since
the parametric model is well-structured it has the potential to be understood
by others due to their explicit representation. We can define the methodology
adopted as a digital workflow that can always be replicated by the following six
steps:
– Step 1. Creation of the conditions of existence necessary to define any math-
ematical function;
– Step 2. Entry of parametric equations;
– Step 3. Introduction of the shape modifiers;
– Step 4. Introduction of sliders and panels for transformations management;
– Step 5. Definition of the generative system;
– Step 6. Graphic output interface.
The main advantage in this graph-based application is to keep the possibility of
exploring the design space even when the detailed model has been completed.
This allows final users to move sliders, type new numbers or press buttons and
their eyes see the design morph and evolve, so user can find his unique solution
in the parameter space. In this process a .STL file is generated and transmitted
through a data network and the resulting object can be physicalized in polymeric
resin using additive manufacturing technologies. The prototype, so generated, is
used in classic jewelry manufacturing process, raising the technological threshold
of manufacturing industry without making any change in the productive process.

4 Results
In our analysis we considered the development of two generative DAGs, each
of them consists of a generative algorithm structured the six main parts. These
structuring parts of the algorithm are those introduced in Fig. 1. The most impor-
tant part of this generative structure is the Step 3: the geometric constrains are
introduced. The constraints constitute of a set of rules that allow to transform
a pure shape into an object capable to satisfy a specific function.
The role of the designer is not just to choose the shape of the material object,
but he has to be able to identify the suitable transformations that can be applied
to the original shape. For this aim, the shapes and the operations used in jewelry
manufacturing have been deeply analyzed in order to define certain rules and
constraints that could guide the designer in the parametric tanking process. From
the analysis conducted it emerged that the most commonly used transformations
are:
430 F. Demarco et al.

– Translation: moving of a shape;

– Rotation; turning a shape around a center or an axis;
– Scaling: changing size of a shape;
– Mirror: reflecting a shape;
– Repetition: create a pattern of shape;
– Combination: combining different shape.

These shape alterations are based on the affine transformations and very often
they are combined between them for the creation of the so-called ornamental
groups. These transformations are controlled by a group of input parameters.
These parameters constitute a semi-automatic user communication interface.
Although the basic forms and transformation rules were previously defined by the
designer the user is free to make explorations in the design space by manipulating
the input data. The system is capable to calculate all the solutions obtained as
combinations of the different input slider and provide, almost, in real time the
shape transformation, in fact the maximum computational time is less than 5
second. In the logic diagram shown in Fig. 1 the group that allows the interaction
with the shape is the one in step 4 represented in blue. The two case studies
proposed in the paper are analyzed below. Of both cases studied the generative
systems, using graph based method is presented. The generative systems were
implemented following the logic set out in Sect. 3 of the paper, it is possible to see
the correspondence between the logical scheme and the corresponding functions
in the DAG by the use of color groups. In order to create an easy-to-use user
interface the input parameter, which are introduced in step 4, are moved to the
beginning of explicit digital dataflow.

4.1 Ring with Two Stones DAG

The first shape analyzed was derived from a heuristic algorithm created by Wol-
fram Mathematica. This algorithm is based on chaotic development systems,
unlike what was expected by these complex systems the patterns that emerged
were regular and they presented alternation of symmetries and breaks of the
symmetries themselves.
As you can see this DAG has eight input parameters, each modification of
the input parameters corresponds to a variation, in real time, of the geometry,
therefore, there are countless outputs rings that this single algorithm is capable
to generate. In Fig. 2 is showed the Grasshopper3D workflow designed for mod-
elling the Ring with stones. Figure 3 shows only a few examples of the results
that can be obtained, the user can evolve the model by varying the shape and
size of the stone, changing the stone’s material as well as the material of the
ring, its size and the number of faces that delimit the surface.
 Algorithms for Jewelry Industry 4.0 431

Fig. 2. Ring with stones generative DAG.

Fig. 3. Five different possible results of ring with stone generative DAG.
432 F. Demarco et al.

4.2 Lissajous Ring DAG

Further experimentation was conducted on the Lissajous pattern. This latter is

representative of the graph of a curve given by a system of parametric equations.
The curves represent one of the main areas of investigation of mathematics for
their ability to represent physical phenomena. This system of equations is often
used as a means of studying oscillatory phenomena. The variation of the fre-
quencies of this system of equations generates harmonic patterns of particular
beauty that are well suited to aesthetic purposes as the use in jewelry man-
ufacturing. Figure 4 shows the DAG generator of the ring in its explicit form
in which the different functions can be identified. This generator is capable to
enlarge the exploration in the design space by providing the user multiple alter-
natives. Countless sequences of this figure are generated while maintaining the
characteristics and functions that the ring have to satisfy. The main transforma-
tions related to the DAG are based on the frequency variations in the harmonic
functions generating the curve. The variation of the frequency as in every har-
monic equation the variation of the frequency leads to a drastic variation of the
curve. In particular, as we can notice in Fig. 5, these types of variations cause
a substantial change in the grid that contracts or expands. Many other param-
eters act on the transformation of the pure shape, these bring variation in the
number of modules making up the object, the overall width of the ring and the
thickness of the section. This gives the possibility to personalize the object and
also the possibility of acting on the ergonomic characteristics and on the weight
of the object. The user can then choose to generate completely different rings
from extremely rarefied objects to complex plots as shown in Fig. 5.

Fig. 4. Lissajous ring generative DAG.

 Algorithms for Jewelry Industry 4.0 433

Fig. 5. Possible Lissajous pattern and their transformation through DAG.

5 Conclusion and Future Work

This paper is based on an experimental investigation to connect parametric and
generative design with o field of industrial design in the new production frame-
work of Industry 4.0 (I4.0). In our opinion, the jewelry sector is the manufactur-
ing sector with the greatest potential for development compared to these new
technologies. However, it still remains tied to traditional low-tech methodologies.
This investigation helps to understand how abstract mathematical equation sys-
tems, algorithmic and genetic theories can be used to become effective tools
of industrial design. This is due to the fact that such fusion allows creating
unique and beautiful forms with a high scientific and mathematical background.
The generative system conceived by the designer helps the user to automati-
cally create his own configuration of a specific project allowing a high level of
customization. Product customization has become an essential attribute in the
industrial framework created by the recent I4.0 revolution.
The study of the transformations of shapes and the ergonomic functions used
for the correct design of jewels has led to the definition of a correct grammar. A
so created grammar is used precisely for the transformation of pure shapes into
objects capable of satisfying specific features. The conducted research showed
that it is possible to: (1) adopt any abstract mathematical models as a source
to produces real objects; (2) show how new industrial designers can use the field
of parametric and computational modelling to foster their creativity.
434 F. Demarco et al.

For the experimental part of this work we used the explicit parametric mod-
eling method, since the necessity of a common language between humans and
machines leaded to the creation of a clear interface that could be easily used.
However, we referred also to implicit systems in order to extend and deepen the
research on the parametric design by applying genetic algorithms. Indeed, further
developments of this work regard the use of heuristic algorithms to optimize the
solutions provided by generative systems, applying further mathematical con-
cepts to the jewelery manufacturing field.

References
1. Aish, R., Woodbury, R.: Multi-level interaction in parametric design. In: Butz, A.,
Fisher, B., Krüger, A., Olivier, P. (eds.) SG 2005. LNCS, vol. 3638, pp. 151–162.
Springer, Heidelberg (2005). https://doi.org/10.1007/11536482 13
2. Kielarova, S.W., et al.: An approach of generative design system: jewelry design
application. In: 2013 IEEE International Conference on Industrial Engineering and
Engineering Management, Piscataway, New Jersey, USA, pp. 1329–1333 (2013)
3. Harding, J.E., Shepherd, P.: Meta-parametric design. Des. Stud. 52, 73–95 (2017)
4. Oxman, R.: Theory and design in the first digital age. Des. Stud. 27, 229–265
(2006)
5. Oxman, R.: Thinking difference: theories and models of parametric design thinking.
Des. Stud. 52, 4–39 (2017)
6. Woodbury, R.G., Peters, B., Sheikholesami, M.: Element of Parametric Design, 1st
edn. Taylor & Francis, Abingdon (2010)
7. Oxman, R.: Parametric design thinking. Des. Stud. 52, 1–3 (2017)
8. Singh, V., Gu, N.: Towards an integrated generative design framework. Des. Stud.
33(2), 185–207 (2012)
9. Bentley, P., Kumar, S.: Three ways to grow designs: a comparison of embryogenies
for an evolutionary design problem. In: Proceedings of the 1st Annual Confer-
ence on Genetic and Evolutionary Computation, pp. 35–43. Morgan Kaufmann
Publishers (1999)
10. Bilotta, E., Pantano, P., Stranges, F.: Computer graphics meets chaos and hyper-
chaos: some key problems. Comput. Graph. 30(3), 359–367 (2006)
11. Gabriele, L., Tavernise, A., Bertacchini, F.: Active learning in a robotics labora-
tory with university students. In: Increasing Student Engagement and Retention
Using Immersive Interfaces: Virtual Worlds, Gaming, and Simulation, pp. 315–339.
Emerald Group Publishing Limited (2012)
12. Adamo, A., Bertacchini, P.A., Bilotta, E., Pantano, P., Tavernise, A.: Connecting
art and science for education: learning through an advanced virtual theater with
“talking heads”. Leonardo 43(5), 442–448 (2010)
13. Bertacchini, F., Bilotta, E., Pantano, P.: Shopping with a robotic companion. Com-
put. Hum. Behav. 77, 382–395 (2017)
14. Bertacchini, F., Bilotta, E., Caldarola, F., Pantano, P., Bustamante, L.R.: Emer-
gence of linguistic-like structures in one-dimensional cellular automata. In: AIP
Conference Proceedings, vol. 1776, p. 090044. AIP Publishing (2016)
15. Bertacchini, F., Bilotta, E., Carini, M., Gabriele, L., Pantano, P., Tavernise, A.:
Learning in the smart city: a virtual and augmented museum devoted to Chaos
theory. In: Chiu, D.K.W., Wang, M., Popescu, E., Li, Q., Lau, R. (eds.) ICWL
2012. LNCS, vol. 7697, pp. 261–270. Springer, Heidelberg (2014). https://doi.org/
10.1007/978-3-662-43454-3 27
 Algorithms for Jewelry Industry 4.0 435

16. Bertacchini, F., Tavernise, A.: Knowledge sharing for cultural heritage 2.0: pro-
sumers in a digital agora. Int. J. Virtual Communities Soc. Network. (IJVCSN)
6(2), 24–36 (2014)
17. Bertacchini, F., Bilotta, E., Gabriele, L., Pantano, P., Tavernise, A.: Toward the
use of Chua’s circuit in education, art and interdisciplinary research: some imple-
mentation and opportunities. Leonardo 46(5), 456–463 (2013)
18. Bilotta, E., Bossio, E., Pantano, P.: Chaos at School: Chua’s circuit for students
in junior and senior High School. Int. J. Bifurcat. Chaos 20(1), 1–28 (2010)
19. Bilotta, E., et al.: ImaginationTOOLS (TM)-A 3D environment for learning and
playing music. In: Eurographics Italian Chapter Conference Proceedings, pp. 139–
144, (2007)
20. Bilotta, E., Di Blasi, G., Stranges, F., Pantano, P.: A gallery of Chua attractors
part VI. Int. J. Bifurcat. Chaos 17(6), 1801–1910 (2007)
21. Bilotta, E., Stranges, F., Pantano, P.: A gallery of Chua attractors: part III. Int.
J. Bifurcat. Chaos 17(3), 657–734 (2007)
22. Bilotta, E., Di Blasi, G., Stranges, F., Pantano, P.: A gallery of Chua attractors:
part IV. Int. J. Bifurcat. Chaos 17(4), 1017–1077 (2007)
23. Bilotta, E., Pantano, P., Vena, S.: Artificial micro-worlds part I: a new approach
for studying life-like phenomena. Int. J. Bifurcat. Chaos 21(2), 373–398 (2011)
24. Bilotta, E., Pantano, P.: Artificial micro-worlds part II: cellular automata growth
dynamics. Int. J. Bifurcat. Chaos 21(3), 619–645 (2011)
25. Lombardo, M.C., Barresi, R., Bilotta, E., Gargano, F., Pantano, P., Sammartino,
M.: Demyelination patterns in a mathematical model of multiple sclerosis. J. Math.
Biol. 75(2), 373–417 (2017)
26. Abdechiri, M., Faez, K., Amindavar, H., Bilotta, E.: The chaotic dynamics of high-
dimensional systems. Nonlinear Dyn. 87(4), 2597–2610 (2017)
27. Bertacchini, F., et al.: An emotional learning environment for subjects with Autism
Spectrum Disorder. In: 2013 International Conference on Interactive Collaborative
Learning (ICL), pp. 653–659. IEEE, September 2013
28. Vaca-Cárdenas, L.A., et al.: Coding with Scratch: the design of an educational
setting for Elementary pre-service teachers. In: 2015 International Conference on
Interactive Collaborative Learning (ICL), pp. 1171–1177. IEEE, September 2015
29. Bertacchini, F., Bilotta, E., Gabriele, L., Pantano, P., Servidio, R.: Using Lego
MindStorms in higher education: cognitive strategies in programming a quadruped
robot. In: Workshop Proceedings of the 18th International Conference on Comput-
ers in Education, ICCE, pp. 366–371 (2010)
30. Sansri, S., Kielarova, S.W.: Multi-objective shape optimization in generative
design: art deco double clip brooch jewelry design. In: Kim, K.J., Kim, H., Baek,
N. (eds.) ICITS 2017. LNEE, vol. 449, pp. 248–255. Springer, Singapore (2018).
https://doi.org/10.1007/978-981-10-6451-7 30
31. Aish, R., Hanna, S.: Comparative evaluation of parametric design systems for
teaching design computation. Des. Stud. 52, 144–172 (2017)
32. Hudson, R., Shepherd, P., Hines, D.: Aviva Stadium: a case study in integrated
parametric design. Int. J. Architect. Comput. 9(2), 187–203 (2011)
33. Oxman, R., Gu, N.: Theories and models of parametric design thinking. In: Pro-
ceedings of the 33rd eCAADe Conference, pp. 477–482 (2015)
34. Ashton, K., et al.: That ‘internet of things’ thing. RFID J. 22(7), 97–114 (2009)
35. Sundmaeker, H., et al.: Vision and challenges for realising the Internet of Things.
In: Cluster of European Research Projects on the Internet of Things, vol. 3, no. 3,
pp. 34–36. European Commission (2010)
436 F. Demarco et al.

36. Prats, M., et al.: Transforming shape in design: observations from studies of sketch-
ing. Des. Stud. 20(5), 503–520 (2009)
37. Derix, C.: Mediating spatial phenomena through computational heuristics. In: Pro-
ceedings of the 30th Annual Conference of the Association for Computer Aided
Design in Architecture, pp. 61–66 (2010)
38. Gero, J.S., Kumar, B.: Expanding design spaces through new design variables. Des.
Stud. 14(2), 210–221 (1993)
39. Harding, J., Joyce, S., Shepherd, P., Williams, C.: Thinking Topologically at Early
Stage Parametric Design. na. (2012)
40. French, M.J., Gravdahl, J.T., French, M.J.: Conceptual Design for Engineers.
Springer, Heidelberg (1985). https://doi.org/10.1007/978-3-662-11364-6
41. Cho, S.B.: Towards creative evolutionary systems with interactive genetic algo-
rithm. Appl. Intell. 16(2), 129–138 (2002)
42. Gibson, I., Rosen, D.W., Stucker, B.: Additive Manufacturing Technologies, 2nd
edn. Springer, New York (2014). https://doi.org/10.1007/978-1-4939-2113-3
43. Kruth, J.P., Leu, M.C., Nakagawa, T.: Progress in additive manufacturing and
rapid prototyping. CIRP Ann. 47(2), 525–540 (1998)
44. Huang, S.H., Liu, P., Mokasdar, A., Hou, L.: Additive manufacturing and its soci-
etal impact: a literature review. Int. J. Adv. Manuf. Technol. 47(2), 1191–1203
(2013)
45. Lee, J., Bagheri, B., Kao, H.A.: A cyber-physical systems architecture for industry
4.0-based manufacturing systems. Manuf. Lett. 3, 18–23 (2015)
46. Rüßmann, M., et al.: Industry 4.0: the future of productivity and growth in man-
ufacturing industries. Boston Consult. Group 9(1), 54–89 (2015)
 Clustering Analysis to Profile Customers’
Behaviour in POWER CLOUD Energy
Community

Lorella Gabriele(B) , Francesca Bertacchini , Simona Giglio ,

Daniele Menniti , Pietro Pantano , Anna Pinnarelli , Nicola Sorrentino ,
and Eleonora Bilotta

University of Calabria, via Pietro Bucci, 17b, Arcavacata di Rende, Italy

{lorella.gabriele,francesca.bertacchini,simona.giglio,daniele.menniti,
pietro.pantano,anna.pinnarelli,nicola.sorrentino,
eleonora.bilotta}@unical.it

Abstract. This paper presents a cluster analysis study on energy con-

sumption dataset to profile “groups of customers” to whom address
POWERCLOUD services. POWER CLOUD project (PON I& C2014–
2020) aims to create an energy community where each consumer can
become also energy producer (PROSUMER) and so exchange a surplus
of energy produced by renewable sources with other users, or collec-
tively purchase or sell wholesale energy. In this framework, an online
questionnaire has been developed in order to collect data on consumers
behaviour and their preferences. A clustering analysis was carried on the
filled questionnaires using Wolfram Mathematica software, in particular
FindClusters function, to automatically group related segments of data.
In our work, clustering analysis allowed to better understand the energy
consumption propensity according the identified demographic variables.
Thus, the outcomes highlight how the availability to adopt technologies
to be used in PowerCloud energy community, increases with the growth
of the family unit and, a greater propensity is major present in the age
groups of 18–24 and 25–34.

Keywords: Machine learning · Cluster analysis · Consumer behaviour

1 Introduction
A sustainable energy transition with the adoption of highly advanced technolo-
gies such as Smart Grid involves changes in a wide range of customers’ energy
behaviors, including the adoption of sustainable energy sources, energy efficient
technologies, investments in energy efficiency processes in buildings, and above
all, changes in direct and indirect behavior by costumers in energy consump-
tion. Some research has shown that such measures still struggle to take off for
a number of reasons, including the acceptance of technologies by users, of cru-
cial importance for the development of a new culture of energy saving. These
c Springer Nature Switzerland AG 2020
Y. D. Sergeyev and D. E. Kvasov (Eds.): NUMTA 2019, LNCS 11973, pp. 437–450, 2020.
https://doi.org/10.1007/978-3-030-39081-5_38
438 L. Gabriele et al.

technologies are more advanced than the previous ones. Smart Grids contain
clusters of electrical engineering technologies merged with network technologies,
connected with powerful instrumentation of smart sensors and meters, placed
between the system that supplies electricity and the end users [1]. The advan-
tages are noteworthy for users. In fact, the main features of Smart Grids are their
ability to recover following damage, the possibility to all types of costumers of
actively participating for energy saving, adaptability to different types of both
wholesale and retail market, for industrial or domestic uses, resistance to attacks
and natural disasters, a higher quality supply [2]. Since the interfaces of these
systems are highly specialized and specific, they require a high level of knowledge,
and certainly a non-trivial interaction by costumers. Therefore, communication
plays a crucial role in the adoption of smart grid technologies. Only 5–6 years
ago, some demographic surveys have shown that many adult users have not only
never heard of such technologies, but that they did not even have an understand-
ing of the amount of energy consumed in their home. However, they claimed to
be willing to cooperate if they had the information they needed. So usually, the
acceptance of new technologies for energy saving is directly linked to the amount
of information a costumer has, accompanied by the perceived usefulness, in rela-
tion to how much this technology positively influences the quality of life [3,4].
Thus, human behavior is complex and acceptance rarely is in line with central-
ized energy saving policies. In fact, unlike the decision-making choices of daily
products, where consumers make rational choices in line with their values and
intentions, it is not the case for the choice of energy products-services, on which
the lack of knowledge or misinformation consider the costs and benefits of all the
existing and optimal technological alternatives. Moreover, even if the choices, at
the social level, have been implemented and the end users have accepted the tech-
nologies, the construction of significant relationships between energetic behavior
and daily behavior is still scarce. However, an increasingly refined knowledge
about all the customers’ behavior allows a very important process because, with
the liberalization of electricity markets, sales companies increasingly need to
define depletion patterns for their electricity customers. Many studies have been
carried out precisely to define consumption models, using Machine Learning and
clustering systems in order to classify types of customers connected with their
consumption profiles. These results have been gathered by using some interest-
ing technological tools such as AMR (Automatic Meter Reading) [5] and the
demand response management (DRM) schemes to manage energy for residential
buildings in smart grids [6]. This allowed an optimization of the energy mar-
kets. Currently, however, new needs are emerging in the energy communities.
Therefore, it is not only necessary to have an idea of customers’ electricity con-
sumption behavior, but also how users exchange electricity in the community,
how sell and buy again, redefining energy production and marketing methods.
In short, in order to optimize the energy reserves produced more and more, from
the research point of view it is interesting to know which are the social dynamics
of energy exchange, whether it is for sale or purchase. Therefore, in this article,
a clustering analysis is proposed to profile the customers’ behavior useful for the
 Clustering Analysis to Profile Customers’ Behaviour 439

PowerCloud project aims. The remainder of the paper is organised as follows.

After this introduction, Sect. 2 informs about the related works on the topic.
Section 3 introduces the PowerCloud project at a glance, highlighting the Real-
ization Objective 3 (OR3). Section 4 argues about Machine Learning and cluster
analysis; Sect. 5 presents the methodology adopted to implement questionnaire
and to analyze dataset. Section 6 discusses Data and Statistical analysis and
Sect. 7 summarizes the Clustering results. Finally, concluding remarks close the
work.

2 Related Work

Smart grids (SGs) for energy saving are now a reality in almost all countries of
the world. Although there is no a commonly accepted meaning for intelligent
network, the US Department of Energy (DOE) classifies it as a two-way flow of
electricity and information that can be used to monitor everything from power
plants to customers’ preferences. According to DOE, this definition embodies one
of the main prerogatives of an intelligent network or SGs. The SGs are usually
connected with the Advanced Metering Infrastructure (AMI), distributed world-
wide, which is why we talk about Big Data and in general of all the connected
4.0 Technologies for the energy sector [7–9]. A large number of heterogeneous
data have been produced by these infrastructures, distributed all over the world.
In a simplified model of the SG, each consumer is considered as an actor who
has the main aim to reduce his/her energy consumption. To evaluate the con-
sumers’ harvesting behavior, they are endowed with a Smart Meter (SM), which
is a device between the local source and the grid, capable of tracking the energy
consumed by load or injected into the grid. These processes are promptly trans-
ferred to a billing system capable of reporting on each energy withdrawal or
injection, establishing from time to time the intervals of interest for data record-
ing or recording all the consumption behavior of customers throughout the entire
period of a day. In order to give to the end-customers a more detailed feedback
on their electricity consumption, many methods have been tested, from price
reduction to monetary incentives, creating elastic tariffs and dynamic pricing or
by direct-load control [10], by allowing to others the control of all consuming
devices and machines. The information that can be extracted from the SM are
of enormous value for all the players of the SGs domain, both for consumers
and for stakeholders, returning into economic benefits and improved services for
both ones. However, analyzing such information flow presents any problems. As
Big Data, they are huge and complex data sets, which are difficult for traditional
tools to store, process and analyze [11], while the computational capabilities to
analyze them should be extraordinarily powerful and time-consuming. This is
why several automatic methods of analysis have been developed [12]. Measured
in the field, collected for a long enough period on the costumers’ load behav-
ior, data are useful to formulate algorithms capable of clustering customers into
macro-categories, based on their behavioral models. In the related literature,
the used methods foresee the following steps: (1) Data collection and processing
440 L. Gabriele et al.

which allow to progress with the association of each customer with its represen-
tative load model, to be considered for categorization purposes. Data must be
indicative of customer’s consumptions behavior collected on a daily load model.
The duration must be long enough to ensure that behavioral patterns emerge.
Therefore, no less than two or three weeks of observation are required in the
same loading condition. At the end of data collection, they are processed and
those that present anomalies are eliminated; (2) In the second phase of the clus-
tering process, the definition of the main functions used as input for allowing the
analysis of the customers’ grouping algorithms is carried out, making a selection
of the most representative features of each user; (3) In the third phase of the
clustering method, the evaluation of the clustering procedures, illustrating the
types of algorithms and the validity of the chosen indicators is fulfilled in order to
assess clustering effectiveness; (4) In the last phase, called post-clustering phase,
there is the emergence of the customer classes, on the attributes used to estab-
lish the client classes based on the results of algorithms applied. In our research
we applied Wolfram Mathematica clustering techniques, since it allows to run a
number of automatically operations in several areas of applications and it can
be applied in the estimating, assessing and monitoring of data processing [13].
In Zotos [13] opinion, Wolfram Mathematica is a tool developed primarily to
handle several aspects of technical computing and engineering in a coherent and
unified way and it is a “faster and more capable functional language approach”.

3 POWER CLOUD Project and User Profile

POWER CLOUD project aims to create an energy community, where each con-
sumer can become also energy producer (PROSUMER) and so exchange a sur-
plus of energy produced by renewable sources with other users, or collectively
purchase or sell wholesale energy. In particular, the project will implement the
POWERCLOUD platform that will be capable to manage energy consumption
and energy production of users in two ways: working in real-time data or on
planned data. In the planning way, it will be made an estimation of the con-
sumption profiles of each user, allowing the optimal scheduling of the energy
loads according both to the availability of energy produced and the price of
the electricity market. In real-time way, any surplus or energy deficit will be
managed, allowing the sale or supply, according to the market modality at a
local level among the users of the energy community. Both modalities, allow to
obtain an economic optimization of energy supply and consumption. The plat-
form collect a large amount of data and process them according specific needs
and returning useful suggestions regarding the to the different energy players
of the energy community. The idea is to interact with each PROSUMER (con-
sumers/producers) so that he can modulate his consumption and possibly, as
energy producers, exchange the surplus of energy with other users without fur-
ther intermediaries; or by collectively buying the wholesale energy being able to
dynamically manage their loads optimally to obtain considerable savings in terms
of tariffs. As regards the Realization Objective 3 (OR3), it aims to collect and
 Clustering Analysis to Profile Customers’ Behaviour 441

analyse data on users’ habits who are part of the energy community, grouping
them through advanced mathematical and statistical techniques, to elaborate
an energy profile for individual and/or class of users. The Realization Objec-
tive 3 is managed by the ESG group. ESG is a research group that operates
at the University of Calabria and has a long tradition in carrying out inter-
disciplinary research activities, conjugating different knowledges belonging to
different fields, such as Mathematics [14–20], Complex systems [21–25], Machine
learning [26], Artificial Intelligence [27] and Cognitive psychology [28–35], as
required by the activities foreseen by the OR3. In this regard, an online ques-
tionnaire has been developed and data has been analysed in order to understand
consumer’s behaviour and their preferences. The ultimate purpose is to profile
“groups of customers” to whom address POWERCLOUD services taking into
account their daily behaviour regarding electricity consumption and other exter-
nal factors (daily and seasonal rhythms on the one hand, meteorological factors
etc.).

4 Machine Learning and Cluster Analysis

The advent of big data era lead some researchers to generate innovate techniques
and apply new processing models named Machine Learning. Machine Learning is
an area of computer science connected to Artificial Intelligence (AI) and focused
on the development of computer systems able to learn and act as humans do [36].
Through Machine Learning algorithms, we can to carry out research on database
of structured and unstructured data, acquiring information about different sys-
tems including multimedia systems such as images, texts, videos, building pre-
dictive models and extracting values from Big Data. Machine Learning tasks
are classified into two main categories: supervised learning (supervised machine)
and unsupervised learning (unsupervised machine). In supervised learning it
is given examples in the form of possible inputs and the respective requested
outputs, the aim is to extract a general rule that associates the input to the
correct output creating a model; in unsupervised learning, the model aims to
find a structure in the inputs provided, without the inputs being labelled in
any way [37]. Currently, Machine Learning analysis has become one of the most
important topics in several fields, such as engineering, economics, physics, and
many others [26,38] due the development of advanced analytics techniques that
provide to leverage Big Data as a competitive advantage. Advanced statistical
methods are used to conduct Big Data analysis including Naive Bayes for classi-
fication, Logistic Regression for regression and Support Vector Machines (SVM)
for regression and classification [39]. In recent years, the scientific research has
increasingly considered Machine Learning algorithms a useful method to analyse
electrical load consumption data [40,41], in particular using cluster analysis, a
powerful method to explore patterns structures within data. Indeed, cluster anal-
ysis, based on an unsupervised learning process, is one of the most commonly
applied data mining techniques and useful to form partitioning a dataset into
sub-groups that are similar to each other and are quite dissimilar to the other
442 L. Gabriele et al.

groups [42,43]. The most popular clustering approaches are Density-Based Clus-
tering (DBSCAN - density-based spatial clustering of applications with noise),
Hierarchical Clustering (algorithm that groups similar objects into groups named
clusters) and K-Means Clustering (algorithm that allows to clustering N data
points into K disjoint subsets based on their attributes) [44]. In our research,
given a set of data composed by the most relevant information about energy
production and consumption at home, we have applied cluster analysis, using
Wolfram Mathematica [45].

5 Methodology
5.1 Scope
The OR3 POWERCLOUD aimed at investigate which are users available to get
closer to the renewable energy market, to domotic and to be part of an energy
community and to recognize users with different approaches towards energy con-
sumption and use.

5.2 Procedure
To achieve the scope above descripted, an online survey was implemented and
submitted to representative sample of households energy consumption. A first
draft of the questionnaire has been assembled and administered to a small group
of five households, through an interview. The aim of this paper is to identify
energy consumption dataset to profile “groups of customers” to whom address
POWERCLOUD services. In this view, Jakob Nielsen, one of the world’s leading
experts on usability, states that to conduct an effective usability test of any prod-
uct, no more than five users are needed [46,47]. A large sample is not necessary,
but it is sufficient that the level of experience of the chosen subjects corresponds
with a good approximation to that of the population for which the product is
addressed. The usability test with users is an empirical evaluation methodology
that foresees observation during the survey, it allows to obtain valuable infor-
mation on the degree of understanding of the “product” without excessively
wasting resources. Thus, users were asked to complete the questionnaire high-
lighting whether the questions had elements that were difficult to understand
and/or were contradictory, in order to individuate strengths and weaknesses of
the pilot questionnaire. This phase was followed by a brainstorming with the
project group to obtain clarifications on the problematic or neglected points and
lastly to elaborate the final questionnaire. The online questionnaire developed
with Google module consists of 4 section and 22 items: the first section was
devoted to collect demographic information, the second section to collect data
relating to user preferences on propensity for energy saving, the third section
concerns willingness to adopt renewable energies, the fourth section on domotics
and automation systems knowledge. The questions are in mixed mode (open and
closed). For closed-ended questions, subjects were asked to express their opin-
ion using a 5-point Likert scale, where: 1 = Strongly disagree 2 = Disagree 3 =
 Clustering Analysis to Profile Customers’ Behaviour 443

Uncertain 4 = Agree 5 = Strongly agree. The average duration of completing the

questionnaire was approximately 8 min. 261 questionnaires have been compiled
on households’ energy consumption useful to study users’ habits. Questionnaire
answers have been encoded numerically both for multiple choices and several
categories. Answers with missing input has been eliminated (4 questionnaires).
This is a very important and critical process for the construction of the data set.
Consequently, 257 vectors of answers have been analysed with Mathematica.
We followed two-step of data analysis: a descriptive analysis to examine a raft of
consumers attitudes, motivations and knowledge variables, along with traditional
covariates such as online service used and other demographic characteristics, and
a cluster analysis to identify distinct groups with similar features in the sam-
ple. As previously done by Frank et al. [48], we carried out Descriptive analysis
on the results of the clusters to understand the relationship of these groups of
consumers (allocated in the different clusters) with different levels of Propen-
sity for energy saving and opinion on current energy supplier, Willingness to
adopt renewable energies and Domotics and automation systems knowledge. In
Fig. 1 the macro steps that allowed to obtain the “Energy customers’ behaviour
profile” are schemed.

Fig. 1. A schematic representation of the adopted methodology: from the collection of

the dataset to the customers’ profile.

6 Findings
6.1 Descriptive Analysis
Descriptive analysis carried out on a datataset of 257 answers, shows that the
sample is equally composed by women and man. Predominantly they are aged
between 18 to 34. 61% and live in an apartment, while 39% in a detached house.
42% of them have energy class A house. They use the web to carry out numerous
activities, from finding travel and vacation packages to using online services,
even if, in the majority, they use social networks (59%). The house is mostly
heated by an autonomous boiler, in fact there is a very high percentage 84% and
use Gas for heat water (79%). The demographic characteristics of the sample
444 L. Gabriele et al.

are summarized in Table 1. The survey instrument employed a five-point Likert

scale to solicit level of agreement with three factors towards Propensity for energy
saving and opinion on current energy supplier, Willingness to adopt renewable
energies, Domotic and automation systems knowledge. The average Likert scale
scores and the standard deviation are displayed in Table 2. Most respondents
“agree” or “strongly agree” (mean score is 4.32) if “in my city more energy will
be produced by renewable sources”.

Table 1. Characteristics of the study sample (percentages).

6.2 Clustering Analysis

A cluster analysis procedure is applied to three identified demographic variables:
age, family unit and online services used, in relation with the different levels of
Propensity for energy saving and opinion, Willingness to adopt renewable ener-
gies and Domotic and automation systems knowledge. Some results are displayed
in Tables 3 and 4. We used a specific function “FindClusters” to select and group
homogeneous elements in a set of data to reduce the search space and to find the
optimal solutions [45]. By default, “FindClusters” function identifies clusters on
 Clustering Analysis to Profile Customers’ Behaviour 445

Table 2. Propensity for energy saving and opinion on current energy supplier, Will-
ingness to adopt renewable energies, Domotic and automation systems knowledge.

Table 3. Cluster Analysis on Age and Propensity for energy saving and opinion on
current energy supplier [a], Willingness to adopt renewable energies [b] and Domotic
and automation systems knowledge [c].

the basis of the shortest distance comparing data without acquiring any infor-
mation about it. To generate the clusterization on the input of dataset, Wolfram
Mathematica software used Gaussian Mixture modelling the probability density
446 L. Gabriele et al.

Table 4. Cluster Analysis on Online services and Propensity for energy saving and
opinion on current energy supplier [a], Willingness to adopt renewable energies [b],
Domotic and automation systems knowledge [c].

Fig. 2. Dendrogram shows the number of clusters related Age and Domotic analysis
 Clustering Analysis to Profile Customers’ Behaviour 447

of a numeric space using a mixture of multivariate normal distribution. Further-

more, as regarding the distance, the software clusterized data according to a
similarity function. The similarity function automatically apply the Euclidean
distance, which groups elements into clusters automatically according to their
proximity. The different analyses carried out on dataset show similar results
both on mean and standard deviation and number of clusters created. In fact,
emerged 2 or 3 main clusters and only in one case “Map [Length, cluster]” is
equal to 11. However, even in this latter case the greatest number of elements is
grouped in the first 3 segments. A segment (133 respondents) belong to 18–24
year age range are aware that “Domotic” is not a valid aid only for the elderly
or disabled people but can contribute to energy saving and therefore reduce
consumption and waste and capable to make home safer. A strong awareness
is noted in the 25 and 34 year age range (74) as regards Domotic and automa-
tion systems knowledge. While the population belonging to an age group over
45 does not show a clear knowledge of these sistems. On the base of the sam-
ple’s demographic characteristics, we identified the following consumer profiles:
A first profile with 18–24 year age range, family unit composed by 3 persons,
uses both e-commerce services and social networks. This subjects are inclined to
adopt technologies useful to energy saving and to employ renewable source. The
second profile with 25–34 year age range, family unit composed by 3 persons,
strongly agree to energy saving and the adoption of technologies for renewable
energy. They believe that is important to have clarity and transparency on the
invoice and are willing to use A++ household appliances. However, they con-
sider today’s energy market cheap. The third profile with 35 to 44 age, family
unit composed by 3 persons, are not agree with production and use of renewable
energy sources and consider their energy supplier unreliable. Furthermore, they
are not willing to pay extra in the invoice for renewable energy production and
are strongly discouraged by the energy market. Figure 2 shows the dendrogram
of the performed hierarchical cluster analysis using age and Domotic variables.
The dendrogram represents the similarities between respondents, grouped into
11 main clusters.

7 Conclusions

Cluster analysis determines a reduction of the total quantity of information

to reach homogeneous groups. Each consumer analysed has x variables (demo-
graphic, behavioural or affective) that can be quantitative and qualitative.
Hence, dividing into categories means considering, in a group of individuals com-
mon characteristics with the group to which he belongs. This analysis is applied
in different pattern-analysis and fields and it is relevant to make assessments
on data structure [49]. In this work the purpose is to detect similarities among
dataset collected by the questionnaire in order to identify distinct patterns capa-
ble to describe customers behaviour profile. In particular, this is useful to define
the profile for each customer class, and the calculation of the global power and
energy information for the customer classes in order to adapt the billing tariffs.
448 L. Gabriele et al.

As regards POWERCLOUD project, it is important to collect information about

consumer’s profile, on their preferences and their habits in order to promote
specific marketing activities, to handle customer needs and formulate effective
business strategies and identify firms strengths and weaknesses. For example,
advertising could be targeted to persuade consumers that are unwilling to pay
extra to adopt renewable energy. It will be useful to highlight that the initial
outlay, will allow to obtain economic advantages later. Moreover, specific cookies
could be showed on websites usually used by consumers for “booking holidays”.
Broader audience could be directly engaged through social networks in acquiring
POWERLCOUD community benefits also in real time.

Acknowledgements. This research was supported by the following grants POWER-

CLOUD (PON I&C2014-2020-MISE F/050159/01-03/X32).

References
1. Yu, X., Xue, Y.: Smart grids: a cyber-physical systems perspective. Proc. IEEE
104(5), 1058–1070 (2016)
2. Dada, J.O.: Towards understanding the benefits and challenges of Smart/Micro-
Grid for electricity supply system in Nigeria. Renew. Sustain. Energy Rev. 38,
1003–1014 (2014)
3. Davis, F.D.: Perceived usefulness, perceived ease of use, and user acceptance of
information technology. MIS Q. 13(3), 319–340 (1989)
4. Park, C.K., Kim, H.J., Kim, Y.S.: A study of factors enhancing smart grid con-
sumer engagement. Energy Policy 72, 211–218 (2014)
5. Mutanen, A., Ruska, M., Repo, S., Jarventausta, P.: Customer classification and
load profiling method for distribution systems. IEEE Trans. Power Delivery 26(3),
1755–1763 (2011)
6. Li, W.T., et al.: Demand response management for residential smart grid: from
theory to practice. IEEE Access 3, 2431–2440 (2015)
7. Tu, C., He, X., Shuai, Z., Jiang, F.: Big data issues in smart grid-a review. Renew.
Sustain. Energy Rev. 79, 1099–1107 (2017)
8. Zhou, K., Fu, C., Yang, S.: Big data driven smart energy management: from big
data to big insights. Renew. Sustain. Energy Rev. 56, 215–225 (2016)
9. Wang, K., et al.: Wireless big data computing in smart grid. IEEE Wirel. Commun.
24(2), 58–64 (2017)
10. Strüker, J., Dinther, C.: Demand response in smart grids: research opportunities
for the IS discipline. In: Proceedings of the Eighteenth Americas Conference on
Information Systems, Seattle, Washington, 9–12 August, pp. 1–10 (2012)
11. Chen, H.H., Chen, S., Lan, Y.: Attaining a sustainable competitive advantage in
the smart grid industry of China using suitable open innovation intermediaries.
Renew. Sustain. Energy Rev. 62, 1083–1091 (2016)
12. Chicco, G.: Overview and performance assessment of the clustering methods for
electrical load pattern grouping. Energy 42(1), 68–80 (2012)
13. Zotos, K.: Performance comparison of Maple and Mathematica. Appl. Math. Com-
put. 188(2), 1426–1429 (2007)
14. Alfano, I., Carini, M., Gabriele, L.: Building SCIENAR, a virtual community of
artists and scientists: usability testing for the system improvement. In: Lazakidou,
 Clustering Analysis to Profile Customers’ Behaviour 449

A. (ed.) Virtual Communities, Social Networks and Collaboration, pp. 147–161.

Springer, New York (2012). https://doi.org/10.1007/978-1-4614-3634-8 8
15. Bilotta, E., Di Blasi, G., Stranges, F., Pantano, P.: A gallery of Chua attractors:
Part VI. Int. J. Bifurcat. Chaos 17(06), 1801–1910 (2007)
16. Bilotta, E., Stranges, F., Pantano, P.: A gallery of Chua attractors: Part III. Int.
J. Bifurcat. Chaos 17(3), 657–734 (2007)
17. Bilotta, E., Di Blasi, G., Stranges, F., Pantano, P.: A gallery of Chua attractors:
Part IV. Int. J. Bifurcat. Chaos 17(04), 1017–1077 (2007)
18. Bilotta, E., Bossio, E., Pantano, P.: Chaos at school: Chua’s circuit for students in
junior and senior High School. Int. J. Bifurcat. Chaos 20(01), 1–28 (2010)
19. Bilotta, E., Pantano, P., Vena, S.: Artificial micro-worlds part I: a new approach
for studying life-like phenomena. Int. J. Bifurcat. Chaos 21(2), 373–398 (2011)
20. Bilotta, E., Pantano, P.: Artificial micro-worlds part II: cellular automata growth
dynamics. Int. J. Bifurcat. Chaos 21(03), 619–645 (2011)
21. Bertacchini, F., Bilotta, E., Carini, M., Gabriele, L., Pantano, P., Tavernise, A.:
Learning in the smart city: a virtual and augmented museum devoted to chaos
theory. In: Chiu, D.K.W., Wang, M., Popescu, E., Li, Q., Lau, R. (eds.) ICWL
2012. LNCS, vol. 7697, pp. 261–270. Springer, Heidelberg (2014). https://doi.org/
10.1007/978-3-662-43454-3 27
22. Bertacchini, F., Bilotta, E., Gabriele, L., Pantano, P., Tavernise, A.: Toward the
use of Chua’s circuit in education, art and interdisciplinary research: some imple-
mentation and opportunities. Leonardo 46(5), 456–463 (2013)
23. Abdechiri, M., Faez, K., Amindavar, H., Bilotta, E.: The chaotic dynamics of high-
dimensional systems. Nonlinear Dyn. 87(4), 2597–2610 (2017)
24. Abdechiri, M., Faez, K., Amindavar, H., Bilotta, E.: Chaotic target representation
for robust object tracking. Sig. Process. Image Commun. 54, 23–35 (2017)
25. Bertacchini, F., Bilotta, E., Caldarola, F., Pantano, P.: The role of computer sim-
ulations in learning analytic mechanics towards chaos theory: a course experimen-
tation. Int. J. Math. Educ. Sci. Technol. 50(1), 100–120 (2019)
26. Giglio, S., Bertacchini, F., Bilotta, E., Pantano, P.: Using social media to identify
tourism attractiveness in six Italian cities. Tour. Manag. 72, 306–312 (2019)
27. Firouznia, M., Faez, K., Amindavar, H., Koupaei, J.A., Pantano, P., Bilotta, E.:
Multi-step prediction method for robust object tracking. Digit. Signal Proc. 70,
94–104 (2017)
28. Adamo, A., Bertacchini, P.A., Bilotta, E., Pantano, P., Tavernise, A.: Connecting
art and science for education: learning through an advanced virtual theater with
“talking heads”. Leonardo 43(5), 442–448 (2010)
29. Bertacchini, F., Bilotta, E., Gabriele, L., Pantano, P., Servidio, R.: Using Lego
MindStorms in higher education: cognitive strategies in programming a quadruped
robot. In: Workshop Proceedings of the 18th International Conference on Comput-
ers in Education, pp. 366–371 (2010)
30. Bertacchini, F., et al.: An emotional learning environment for subjects with
Autism Spectrum Disorder. In: International Conference on Interactive Collab-
orative Learning (ICL), pp. 653–659. IEEE (2013)
31. Bertacchini, F., Tavernise, A.: Knowledge sharing for cultural heritage 2.0: pro-
sumers in a digital agora. Int. J. Virtual Commun. Soc. Netw. 6(2), 24–36 (2014)
32. Vaca-Cárdenas, L.A., et al.: Coding with scratch: the design of an educational set-
ting for elementary pre-service teachers. In: International Conference on Interactive
Collaborative Learning (ICL), pp. 1171–1177. IEEE (2015)
450 L. Gabriele et al.

33. Gabriele, L., Marocco, D., Bertacchini, F., Pantano, P., Bilotta, E.: An educational
robotics lab to investigate cognitive strategies and to foster learning in an arts and
humanities course degree. Int. J. Online Eng. (iJOE) 13(04), 7–19 (2017)
34. Bertacchini, F., Bilotta, E., Lombardo, M.C., Sammartino, M., Pantano, P.: Brain-
like large-scale cognitive networks and dynamics. Eur. Phys. J. Spec. Top. 227(7–
9), 787–797 (2018)
35. Gabriele, L., Bertacchini, F., Tavernise, A., Vaca-Cárdenas, L., Pantano, P.,
Bilotta, E.: Lesson planning by computational thinking skills in Italian pre-service
teachers. Inf. Educ. 18(1), 69–104 (2019). https://doi.org/10.15388/infedu.2019.
04
36. Zhou, L., Pan, S., Wang, J., Vasilakos, A.V.: Machine learning on big data: oppor-
tunities and challenges. Neurocomputing 237, 350–361 (2017)
37. Love, B.C.: Comparing supervised and unsupervised category learning. Psychon.
Bull. Rev. 9(4), 829–835 (2002)
38. Bertacchini, F., Bilotta, E., Pantano, P.: Shopping with a robotic companion. Com-
put. Hum. Behav. 77, 382–395 (2017)
39. Duque, J., Patino, J., Betancourt, A.: Exploring the potential of machine learning
for automatic slum identification from VHR imagery. Remote Sens. 9(9), 895 (2017)
40. Tanwar, A.K., Crisostomi, E., Ferraro, P., Raugi, M., Tucci, M., Giunta, G.: Clus-
tering analysis of the electrical load in European countries. In: 2015 International
Joint Conference on Neural Networks (IJCNN), pp. 1–8. IEEE (2015)
41. Benı́tez, I., Quijano, A., Dı́ez, J.L., Delgado, I.: Dynamic clustering segmentation
applied to load profiles of energy consumption from Spanish customers. Int. J.
Electr. Power Energy Syst. 55, 437–448 (2014)
42. Antoniadis, A., Brossat, X., Cugliari, J., Poggi, J.M.: Clustering functional data
using wavelets. Int. J. Wavelets Multiresolut. Inf. Process. 11(01), 1350003 (2013)
43. Raju, S., Chandrasekaran, M.: Performance analysis of efficient data distribution in
P2P environment using hybrid clustering techniques. Soft Comput. 23(19), 9253–
9263 (2019)
44. Nerurkar, P., Shirke, A., Chandane, M., Bhirud, S.: Empirical analysis of data
clustering algorithms. Procedia Comput. Sci. 125, 770–779 (2018)
45. Wolfram, S.: An elementary introduction to the Wolfram Language. Wolfram
Media, Incorporated (2017)
46. Nielsen, J., Tahir, M.: Homepage Usability: 50 Websites Deconstructed. New Riders
Publishing, Thousand Oaks (2001)
47. Nielsen, J.: Designing Web Usability. Verlag: Markt+Technik Verlag; Adresse:
München (2001)
48. Frank, A.G., Dalenogare, L.S., Ayala, N.F.: Industry 4.0 technologies: implemen-
tation patterns in manufacturing companies. Int. J. Prod. Econ. 210, 15–26 (2019)
49. Jain, A.K., Murty, M.N., Flynn, P.J.: Data clustering: a review. ACM Comput.
Surv. (CSUR) 31(3), 264–323 (1999)
 A Grossone-Based Numerical Model
for Computations with Infinity: A Case
Study in an Italian High School

Francesco Ingarozza(B) , Maria Teresa Adamo, Maria Martino,

and Aldo Piscitelli

Liceo Scientifico Statale “Filolao”, via Acquabona, 88821 Crotone, KR, Italy
[email protected], [email protected],
[email protected], [email protected]

Abstract. The knowledge and understanding of abstract concepts sys-

tematically occur in the studies of mathematics. The epistemological
approach of these concepts gradually becomes of higher importance as
the level of abstraction and the risk of developing a “primitive concept”
which is different from the knowledge of the topic itself increase. A typical
case relates to the concepts of infinity and infinitesimal. The basic idea is
to overturn the normal “concept-model” approach: no longer a concept
which has to be studied and modeled in a further moment but rather a
model that can be manipulated (from the calculation point of view) and
that has to be associated to a concept that is compatible with the cal-
culus properties of the selected model. In this paper the authors want to
prove the usefulness of this new approach in the study of infinite quan-
tities and of the infinitesimal calculus. To do this, they expose results of
an experiment being a test proposed to students of a high school. The
aim of the test is to demonstrate that this new solution could be useful
in order to enforce ideas and acknowledgment about infinitesimal cal-
culus. In order to do that, the authors propose a test to their students
a first time without giving any theoretical information but only using
an arithmetic/algebraic model. In a second moment, after some lectures,
the students repeat the test showing that new better results come out.
The reason is that after lessons, students could join new basic ideas or
primitive concepts to their calculus abilities. By such doing they do not
use a traditional “concept–model” but a new “model–concept” solution.

Keywords: Mathematics education · Teaching/learning methods and

strategies · Grossone · Computer tools

1 Introduction

The understanding and knowledge of abstract concepts or of infrequent concepts,

are that of an element which systematically occurs in the study of mathematics
c Springer Nature Switzerland AG 2020
Y. D. Sergeyev and D. E. Kvasov (Eds.): NUMTA 2019, LNCS 11973, pp. 451–462, 2020.
https://doi.org/10.1007/978-3-030-39081-5_39
452 F. Ingarozza et al.

(see [16,23]). Therefore the epistemological approach in regards to these con-

cepts, becomes fundamental to give guarantees of success in terms of didactic
results. The sense of this observation is very important. Besides it increases as the
same level of abstraction of mathematics concepts and related models increases.
The reason for that sentence is that as the level of abstraction increases and
becomes considerable, it increases the risk to develop those fundamental ideas
which could be different form the real being of knowledge. Those fundamental
ideas are defined primitive concepts; they could be very far from the real subject
of knowledge. Not only, that ideas could be misleading with the respect to the
whole ideas of knowledge (see [3,5,13–15,17,18]). One of the abstraction ele-
ments that high school students encounter is related to the concepts of infinity
and infinitesimal (see, for example, [12,16,22,24,27,28]). Therefore it becomes
mandatory a path of knowledge, of discussion about traditional teaching. Such
a strategy can be one, which makes use of a symbolic association of a “symbol-
concept” model. This way it becomes possible to normalize a process based on an
abstract concept, whose in-depth discussion requires a convergence of different
signifiers.
This normalization process can be reached through the assumption of a single
shared model, such a syllabus based on an epistemological model (see [19,21]).
The basic idea is to overturn the normal “concept-model” approach: no longer
a concept to be studied and modelled (only in that of a second time), but a
model that can be manipulated (from the point of view of calculation) and
to which it is possible to associate a concept. The associated (to the model)
concept is compatible with the calculation properties of the identified model and
it’s also compatible with the theoretical results which can be presented (see, for
instance, [3,5,21]). The described approach is the one which has been used in this
didactic experiment. It represents a new idea for the presentation of the concepts
of infinity and infinitesimal. The objective is therefore the ability to express a
concept regardless of the intrinsic difficulties of the same representation. For the
infinite quantity, the symbol ① (“grossone”) is used. By using this model-based
approach, the “symbol-concept” association overcomes the operational limits
of a no adequate language. In the following discussion we present the results
of a didactic experimentation. This experiment is aimed at demonstrating the
validity and effectiveness, in terms of didactic repercussions. The demonstration
is based on a real didactic case which is the study of infinitesimal calculus.
Through this case we want to prove the validity and effectiveness of the “symbol-
concept” approach and of the model based approach in computational arithmetic
(see [4,12,19,21,23,25]). It is important to note that another similar experiment
has been carried out independently during the same periods (see [1]).
It is important to remark that the content of this paper concerns about
the Arithmetic of infinite (based on the use of the symbol ①). This idea was
developed starting from 2003 by Sergeyev (see, e.g., [22,23,25,26]). We should
notice that there are different examples in the scientific literature that can con-
firm the usefulness of such an approach in different fields of mathematics (see
[2,6–11,19,20,22–27]). Infinite Arithmetic represents a new and different point
 A Grossone-Based Numerical Model for Computations with Infinity 453

of view on the concept of number and its related concept of counting. Starting
from their own experiences, the authors can confirm the difficulties of the tradi-
tional calculus that students face, especially when they deal with counting. For
example, difficulties related to the study of probabilities are originated by the
difficulties of counting events. The new approach can be applied both to already
known problems and to new ones. For the first class of problems this method can
confirm well known results while for new ones it can represent an alternative way
for finding solutions. Regarding the study of infinitesimal calculus, the grossone
based model can represent a good solution which can be easily understood by
high school students.

2 Description of the Experimentation

In the following two subsections we deal with the sample and the pursued objec-
tives and purposes.

2.1 The Sample

The didactic experiment was directed to the students in their 4th year at the
secondary school Liceo Scientifico Filolao, in Crotone, Italy. In particular, the
sample was evaluated by three classes (from now we call them class A, class B and
class C in order to save the privacy of the students) and involved a total of about
70 students; it lasted about 8 h and was carried out in the period of April/May
2019. The choice of the sample and of the period is not random but it has a
precise reason: during this period those students started to approach infinitesimal
calculus and exactly limits operations. Because of these considerations, it is
correct to suppose that there are at least four reasons to consider the chosen
sample and the period to be good choices:
– Confidence with the concepts exposed in the experiment;
– Students know the difficulties in regards to a symbolic-arithmetic manipula-
tion of these subjects;
– The possibility to verify, after the experiment, that it is possible to have an
algebraic and computational manipulation of the topics;
– The possibility to verify the advantages deriving from this model-based app-
roach for infinitesimal calculus.
It should be reminded that all the classes involved in the experiment have, as
planned, already operated with elements of infinitesimal calculus and with the
study of indeterminate forms.

2.2 Objectives and Purposes

The experiment carried out on the selected sample was not only a didactic exper-
iment but also an epistemological experiment: the target of the experiment was
twofolded. On the one hand, the well-known importance of presenting concepts
454 F. Ingarozza et al.

that require a good deal of abstraction as clearly as possible and far from ambi-
guities, doubts and perplexities, and jointly, a good mastery of arithmetical,
algebraic and computational skills with regard to presented concepts. Near to
this aspect, purely didactic, from an epistemological point of view, it was/is
important to verify the didactic efficacy of the approach used, starting from the
verification and clear comprehension, and with the mastery of the calculation
properties of the concepts presented to be able to draw appropriate considera-
tions about the usefulness of the presented approach (see [13,15,18]).

3 Results of the Experiment

The sample was submitted to a didactic experiment which was made up by three
distinct phases:
– Test administration;
– Didactic lectures oriented on commenting the test;
– Test administration.
The basic idea of the experiment is to highlight some shadow areas which are
related to some aspects inherited by infinitesimal calculus and especially to the
formal treatment of the concept of infinity. In order to demonstrate the didactic
efficiency of the model based approach, a preliminary test was submitted to the
students; the aim of this test was to verify good computational capacities linked
to some symbols in the face of a limited understanding of some concepts or of
some obvious gaps in the complete and full understanding of those same concepts.
In particular, the grossone symbol was presented within the preliminary test but
a complete description of grossone and its properties was not given. Only one
information was provided to the students, in the introduction of the preliminary
test. The information is the clarification of the property ①> n, ∀ finite n ∈ IN
(for obvious reasons of formal completeness, see also [19]). After doing that
clarification, the students were given a test of 46 questions divided into four
macro areas:
– Section 1 - Order relationships;
– Section 2 - Arithmetic;
– Section 3 - Equality;
– Section 4 - Counting of the elements of known numerical sets.
In Section 1 there are questions underlying the order relationships con-
structed starting from the symbol ①. This section presents 5 open questions. It
is therefore possible to answer each question by selecting one of the 5 proposed
solutions. With regard to the questions contained in Section 1 it is reasonable
to expect sufficiently high results. Moreover, the almost non-existent knowledge
about ① is not a limitation for a successful result in Section 1. Indeed, the
starting hypothesis (①> n, ∀ finite n ∈ IN) is more than sufficient to answer
with a certain degree of accuracy and with a reduced limit of error, to the
questions contained in Section 1 concerning trivial order relationships between
 A Grossone-Based Numerical Model for Computations with Infinity 455

arbitrary quantities containing ①. In Section 2, six algebraic questions/exercises

were submitted. In particular, to pupils were given some expressions to solve.
The expressions contain within them the symbol ① (which must be treated as
a trivial and simple algebraic quantity, starting from the previous consideration
①> n, ∀n ∈ IN). It is evident that since these are exercises of a purely alge-
braic nature (expressions!) the submission method of Section 2 is that of closed
questions: the students have to perform the required calculations and enter the
results. We can confirm the same observations made for Section 1 even for Section
2; in this section of a pure arithmetic/algebraic exercises are submitted. More
precisely they (the exercises) are expressions containing the quantity ①. They
can be treated as expressions of literal calculation that contain precisely the
quantity ① instead of a monomial with any literal part (no clarifications or fur-
ther properties are required on the numerical set which contains the quantity ①).
It is reasonable to expect quite high results for the entire Section 2 and it
is quite obvious to suppose that different results (far from expected ones) may
be due to trivial calculation errors. In Section 3, questions were proposed in the
form of open ended questions, and they are related to the “parity” of quanti-
ties obtained starting from ①. Finally, in Section 4, students have to answer to
questions regarding the counting of elements of known numerical sets or parts
of them. In Section 3 and Section in 4 results change considerably. In fact, the
only given information about ① is not more sufficient to provide exhaustive clar-
ifications that allow unequivocal and correct answers to the questions which are
proposed in Section 3 and Section 4 (apart from rarely cases such as the first
7/8 questions of Section 4 in which it is possible to answer by using the normal
notions of infinitesimal calculation and limit operation). Besides it is important
to remember that the test was made in such a way to avoid the incidence on
the final result of randomly right answers (especially in sections in which open
ended questions are submitted). In order to avoid false positive cases (randomly
correct answers) which could distort the evaluation on the didactic impact of
the experimentally presented model, it was decided to assign a highly penalizing
negative score for each wrong answer. In particular, wrong answers generate a
score of -1, which therefore cancels the score generated by a right answer. In this
way it is reasonable to expect that the students won’t answer the questions which
are related to unknown subjects and topics. By using this strategy is possible
to suppose a strong reduction of false positive cases; in this way it is possible to
immunize the experiment against this potential risk that could affect its actual
validity.
Starting from the previous considerations regarding Section 3 and Section 4,
including the strategy of answer evaluation, it is therefore reasonable to expect
a significant lowering of the results obtained by the students in Section 3 and
Section 4. Notice also that in a total of 46 total applications, Section 4 evaluates
more than 54% and Section 3 evaluates more than 20% of the total; for this
reason the two sections jointly account for over 75% and therefore it is reasonable
to expect very low results from the preliminary test. The fundamental idea of
the experiment is just this: to verify a radical change of the results of the same
456 F. Ingarozza et al.

test proposed during the initial and final phases, on the same sample. As was
already said, there is only one significant difference between these two phases
of the test: the test was submitted for the second time to the same sample of
students only after they had followed a cycle of lectures regarding the Grossone
theory and model. In such way the students could develop a new basic idea and
a new way to approach infinitesimal calculus.
The aim of the experiment in its whole is to demonstrate that this new
approach could improve the development of new primitive concepts for students.
The first aim of the lesson is to speak to the students about the model used
for the representation of infinite quantities, that is, the symbol ① which starts
as a numerical approach and extends to all its properties (see [19,23,25]). The
target of the test is precisely to demonstrate that starting form an easy algebraic
manipulation of a symbol, an association with more abstract concepts can be
useful in order to understand aspects related to the concept. In particular, and
in detail (Section 3 and Section 4) a few expedients are sufficient to start from
the innovation associated with a mere and new definition. This definition is that
of a new numerical set, made up of an infinite number of elements some of
which are infinite elements related to infinite quantities. Thanks to this new and
different approach, (grossone model based) students could be able to reach a very
high level of knowledge, in a new way which could be a logical unexceptionable
manner. This is probably the only way, or one of the few available solutions in
order to reach this knowledge, in regard to these subjects. The same concepts
could be precluded or not easily deducible by using a different strategy so far
from that described on this paper, for example, the traditional way (as used and
intended in Italian schools).
The last consideration can be confirmed by results of the test which had been
submitted for the first time to students. In fact, these students were in the same
conditions as the most of students of Italian high schools. They had never used
the new approach which is proposed in this article. Such knowledge allows, in the
final phase of the experimentation, to be able to carry out the same preliminary
test (even if it’s called final test) from which better results come out. It can
be said that this experiment is an evaluation and measurement model for the
behavior of didactic efficiency of the model based approach.
In the following, some examples of questions of the test are reported.
For Section 1:
Let us consider ① symbol as a positive quantity such that ①> n (∀ finite n ∈ IN).
Choose the right order relationship among the different proposed solutions

(1) ① and -①
(A) ①> -①
(B) We can’t establish an order relationship
(C) -①≤ ①
(D) -①= ①
(E) ①< -①
(2) 2① and 7①
(A) 2①> 7①
 A Grossone-Based Numerical Model for Computations with Infinity 457

(B) 7①= 2①
(C) We can’t establish an order relationship
(D) 2①≥ 7①
(E) 2①< 7①

(3) (1/2) ① and (2/3)①

(A) (1/2)①> (2/3)①
(B) (1/2)①= (2/3)①
(C) (2/3)①> (1/2)①
(D) We can’t establish an order relationship
(E) -(2/3)①> (1/2)①

For Section 2:
Solve the following exercises:

(1) 6(① − 3) − 8(9 + 3①)

(2) ①(3 + ①) − 4①2 − 3①(1 − ①)
(3) 2[①(3① − 5) + 3(① + 3) − ①(2① + 4① − 11)]
 2
1
(4) 41 (① − 2)(① + 2) − ①−1
2
(5) (①2 + ① + 1)2 − (①2 + ①)2
(6) (①2 + ① + 1)(① − 1) − (① + 1)(①2 − ① + 1) + 2

For Section 3:
Determine whether the following quantities are even or odd numbers, by indicating the
letter E (even) or O (odd).

(1) 2①
(2) 5①
(3) 7① + 1
(4) 2① − 3
1
(5) 2
①
1
(6) 5
①+3
3
(7) 7
①
5
(8) 4
①−2
1
(9) 2
①−3
(10) ①

For Section 4 (remind that in this Section 25 questions have been asked):

(1) Indicate the correct value of the sum ∞ + 2, justifying your answer
(2) Indicate the correct value of the sum ∞ + ∞, justifying your answer
(3) Indicate the correct value of the product 2 · ∞, justifying your answer
(4) Indicate the correct value of the product ∞ · ∞, justifying your answer
(5) Indicate the correct value of the sum ∞ − ∞, justifying your answer
...
(9) Determine the number of elements of the set IN
(10) Determine the number of elements of the set ZZ
(11) Let us consider E as the set of even numbers. Determine if the number of elements
of the set IN \ P is even or odd.
458 F. Ingarozza et al.

(12) Let us consider O as the set of odd numbers. Determine if the number of elements
of the set ZZ \ O is even or odd.
(13) Determine the nature (even or odd) of elements of the set ZZ \ {0}
(14) Determine the nature (even or odd) of elements of the set made up by the first
100 natural numbers minus the first 30 odd numbers

It should be noticed that the whole test is made up of 46 questions: 5 ques-

tions for Section 1 (order relationships), 6 questions, more exactly, 6 exercises
(or expressions) for Section 2, 10 questions for Section 3 (parity problems) and
25 questions for Section 4. It is important to note that Section 3 and Section 4
are strictly related each to other. In Section 3 students have to find the nature
(even or odd) of some quantities made up by the symbol ①. In Section 4 they
are called to determine the nature of the number of elements of sets which have
infinite elements. How can they do this? How can they count an infinite num-
ber of elements? The only way is to use the ① based model as described in the
following section.
Another important thing to be noted in Section 4 is the following: students
should be able to answer very easily to some questions by studying infinitesimal
calculus according to the traditional approach which is used in Italian school. On
the other hand, they are not prepared to overcame the second part of Section 4.
According to the consideration which has been just done, it is possible to use the
results of Section 4 to have an idea of the behavior of the new didactic approach
for the study of infinitesimal calculus.

4 The Lessons
Lessons took place in the form of frontal teaching. They were characterized by
the use of methodologies based on the model of social collaboration and by using
a metacognitive approach. Due to these reasons, the model of the Socratic lesson
was used. By doing this choice students were led by the teacher to build their
knowledge of concepts step by step. The starting point was the commentary of
the answers given (and above all not given) during the preliminary test. This
comment has revealed two facts: few difficulties in performing simple complex
arithmetic operations and at the same time a very big difficulty in counting (this
problem is often ignored in the literature, see [3,13,18]). This finding was derived
from questions such as:

– Is the number of elements of the set IN even or odd?

– How many are the elements of the set ZZ?

To these questions we tried to give an answer by using a logically unexceptionable

path: it is possible not to know if the number of elements of the set IN is even or
odd but surely it will be the sum of two numbers which constitute a partition of
IN, i.e. even numbers and odd numbers. These two numerical sets have the same
 A Grossone-Based Numerical Model for Computations with Infinity 459

number of elements and so, it is possible to deduce that whatever the number of
these elements (let’s call α this number), the number of the elements of IN will
always be double (and so 2α). Thus, it is easy to understand that the number
of elements of IN is even.
From this consideration we then moved to the definition of ①: the scale value
at which all the elements of IN are counted. From this point on the road has been
downhill: the set IN of natural numbers having elements from 1 to ① is defined.
This set contains infinitely many elements (precisely, ① elements), some of them
 is
are finite, the others are infinite (see, e.g., [19,23,25,26]). Then, the set IN

defined as the set which contains positive integers larger than ①, i.e., IN ⊂ IN.
From there it was possible with a few tricks that referred to the properties of
the rest classes, to go on to discuss all the aspects which allow the students to
answer to the most of the question of the test, especially to those of Section 3
and Section 4.

5 Analysis of the Results

In this section we show that the experimentation led to results which are broadly
in line with the forecasts. The preliminary test has scores which never reach
the level of sufficiency (related as the half of the maximum value, equal to 46
points). It is important to note that the most of scores levels are widely achieved
in sections 1 and 2 concerning traditional arithmetic. As it was expected, it can
be seen from Tables 1, 2 and 3 below, there is a considerable difference between
the results which are carried out before the lectures and the results related to
the second administration of the test (results related to the first administration
of the test are always on the left for each comparison). On the one hand, it
is possible to note that there are no significant differences (between the first
administration of the test and the second one) related to Section 1 or Section
2. Moreover, in these sections, students have reached a score really closed to
the maximum allowed value. On the other hand, it is not possible to say the
same thing for Sections 3 and 4. In these cases, differences between the first
administration of the test and the second one are very strong and they are in
the direction to confirm the positive behavior of the proposed approach in terms
of didactic efficiency.
The negative trend of the initial test results is mainly produced by the results
of section 3 and section 4 which are based on questions that require a mini-
mum level of knowledge at a conceptual level. The approach described above
demonstrates, at the outcome of the test, its effectiveness since the same test,
administered after only two lessons, produced completely different results. As
it is possible to see, the difference is mainly due to the results of the questions
posed in sections 3 and 4 which, in the preliminary test, provided the major
difficulties to the students and had affected the final result.
460 F. Ingarozza et al.

Table 1. The mean score obtained by the students. The rows refer to the classes and
the columns refer to the 4 different sections in which the tests have been divided.
Columns for each section are double, “before” and “after” the cycle of lectures.

Class Section 1 Section 1 Section 2 Section 2 Section 3 Section 3 Section 4 Section 4

before/5 after/5 before/6 after/6 before/10 after/10 before/25 after/25
A 3.14 3.52 3.86 4.56 1.1 9.71 2.71 24.33
B 1.64 4.45 3.86 5.77 1.82 9.45 0.27 22.5
C 3.08 4.83 4.8 4.63 0.32 9.75 0.68 23.96

Table 2. The standard deviations relative to the means listed in Table 1.

Class Section 1 Section 1 Section 2 Section 2 Section 3 Section 3 Section 4 Section 4

before/5 after/5 before/6 after/6 before/10 after/10 before/25 after/25
A 2.41 1.01 2.89 1.28 1.23 0.39 3.06 0.7
B 6.23 0.79 8.66 0.36 2.6 0.43 0.74 1.25
C 1.19 0.22 2.24 1.98 6.37 0.35 1.26 1.96

Table 3. Table presents the mean vote of all the sections 1–4 with weights. The last
two columns yield the relative standard deviations as in Table 2.

Class Mean, sections 1–4 Mean, sections 1–4 Standard dev. before Standard dev after
before lectures/46 after lectures/46 lectures/46 lectures/46
A 10.81 42.52 9.58 4.82
B 7.59 42.1 29.7 1.48
C 8.88 43.17 11.63 4.81

6 Conclusions

The experimentation performed in Liceo Scientifico “Filolao” using the grossone-

based methodology has shown that the symbol-concept approach can be an
improved solution with respect to the traditional concept-symbol approach at
least in the discussion of those topics that are intrinsically difficult to treat and
understand. It has been shown that the ①-based methodology allows students
to understand better concepts related to infinity and infinitesimals. The authors
hope that this work can be a starting point for similar educational experiments
carried out on a large scale and even in different initial conditions regarding
periods of the test, age of the students of the sample, etc. In this way it is
possible to have a better overview of the real effectiveness of this strategy in
terms of educational impact. Starting from these considerations, the authors
intend to identify new didactic paths for the introduction in teaching of new
computational algorithms that make use of the acquired knowledge.

Acknowledgements. The authors thank Fabio Caldarola, University of Calabria,

for the supervision of the project and the Headmistress of Liceo Scientifico “Filolao”,
Antonella Romeo, for the economic support. The authors thank the anonymous review-
ers for their useful comments that have improved the presentation. Special thanks go
to Irene Dattolo for her valuable support provided for the translation of the text.
 A Grossone-Based Numerical Model for Computations with Infinity 461

References
1. Antoniotti, L., Caldarola, F., d’Atri, G., Pellegrini, M.: New approaches to basic
calculus: an experimentation via numerical computation. In: Sergeyev, Ya.D.,
Kvasov, D.E. (eds.) NUMTA 2019. LNCS, vol. 11973, pp. 329–342. Springer, Hei-
delberg (2019)
2. Antoniotti, L., Caldarola, F., Maiolo, M.: Infinite numerical computing applied to
Hilbert’s, Peano’s, and Moore’s curves. Mediterranean J. Math. (to appear)
3. Asubel, D.: Educazione e processi cognitivi. Franco Angeli (1978)
4. Bertacchini, F., Bilotta, E., Caldarola, F., Pantano, P.: The role of computer sim-
ulations in learning analytic mechanics towards chaos theory: a course experimen-
tation. Int. J. Math. Educ. Sci. Technol. 50(1), 100–120 (2019)
5. Bonaiuti, G., Calvani, A., Ranieri, M.: Fondamenti di didattica. Teoria e prassi dei
dispositivi formativi. Carrocci, Roma (2007)
6. Caldarola, F.: The exact measures of the Sierpiński d-dimensional tetrahedron in
connection with a Diophantine nonlinear system. Commun. Nonlinear Sci. Numer.
Simul. 63, 228–238 (2018). https://doi.org/10.1016/j.cnsns.2018.02.026
7. Caldarola, F.: The Sierpiński curve viewed by numerical computations with infini-
ties and infinitesimals. Appl. Math. Comput. 318, 321–328 (2018). https://doi.
org/10.1016/j.amc.2017.06.024
8. Caldarola, F., Cortese, D., d’Atri, G., Maiolo, M.: Paradoxes of the infinite and
ontological dilemmas between ancient philosophy and modern mathematical solu-
tions. In: Sergeyev, Y.D., Kvasov, D.E. (eds.) NUMTA 2019. LNCS, vol. 11973,
pp. 358–372. Springer, Heidelberg (2019)
9. Caldarola, F., Maiolo, M., Solferino, V.: A new approach to the Z-transform
through infinite computation. Commun. Nonlinear Sci. Numer. Simul. 82, 105019
(2020). https://doi.org/10.1016/j.cnsns.2019.105019
10. Cococcioni, M., Pappalardo, M., Sergeyev, Y.D.: Lexicographic multi-objective lin-
ear programming using grossone methodology: theory and algorithm. Appl. Math.
Comput. 318, 298–311 (2018)
11. De Cosmis, S., De Leone, R.: The use of grossone in mathematical programming
and operations research. Appl. Math. Comput. 218(16), 8029–8038 (2012)
12. Ely, R.: Nonstandard student conceptions about infinitesimals. J. Res. Math. Educ.
41(2), 117–146 (2010)
13. Faggiano, E.: “Integrare” le tecnologie nella didattica della Matematica: un com-
pito complesso. Bricks 2(4), 98–102 (2012)
14. Gastaldi, M.: Didattica generale. Mondadori, Milano (2010)
15. Gennari, M.: Didattica generale. Bompiani, Milano (2006)
16. Iannone P., Rizza D., Thoma A.: Investigating secondary school students’ episte-
mologies through a class activity concerning infinity. In: Bergqvist E., et al. (eds.)
Proceedings of the 42nd Conference of the International Group for the Psychology
of Mathematics Education, Umeå, Sweden, vol. 3, pp. 131–138. PME (2018)
17. La Neve, C.: Manuale di didattica. Il sapere sull’insegnamento. La Scuola, Brescia
(2011)
18. Palumbo, C., Zich, R.: Matematica ed informatica: costruire le basi di una nuova
didattica 2(4), 10–19 (2012)
19. Rizza, D.: Primi Passi nell’Aritmetica dell’Infinito. Un nuovo modo di contare e
misurare (2019, Preprint)
20. Rizza, D.: A study of mathematical determination through Bertrand’s Paradox.
Philosophia Mathematica 26(3), 375–395 (2018)
462 F. Ingarozza et al.

21. Scimone, A., Spagnolo, F.: Il caso emblematico dell’inverso del teorema di Pitagora
nella storia della trasposizione didattica attraverso i manuali. La matematica e la
sua didattica 2, 217–227 (2005)
22. Sergeyev, Y.D.: A new applied approach for executing computations with infinite
and infinitesimal quantities. Informatica 19, 567–596 (2008)
23. Sergeyev, Y.D.: Arithmetic of infinity. 2nd electronic ed. 2013. Edizioni Orizzonti
Meridionali, Cosenza (2003)
24. Sergeyev, Y.D.: Numerical point of view on Calculus for functions assuming finite,
infinite, and infinitesimal values over finite, infinite, and infinitesimal domains.
Nonlinear Anal. Ser. A: Theory Methods Appl. 1(12), 1688–1707 (2009)
25. Sergeyev, Y.D.: Un semplice modo per trattare le grandezze infinite ed infinitesime.
Matematica, Società Cultura: Rivista dell’Unione Matematica Italiana 8(1), 111–
147 (2015)
26. Sergeyev, Y.D.: Numerical infinities and infinitesimals: Methodology, applications,
and repercussions on two Hilbert problems. EMS Surv. Math. Sci. 4(2), 219–320
(2017)
27. Sergeyev, Y.D., Mukhametzhanov, M.S., Mazzia, F., Iavernaro, F., Amodio, P.:
Numerical methods for solving initial value problems on the infinity computer.
Int. J. Unconv. Comput. 12, 3–23 (2016)
28. Tall, D.: A child thinking about infinity. J. Math. Behav. 20, 7–19 (2001)
 A Computational Approach with MATLAB
Software for Nonlinear Equation Roots
Finding in High School Maths

Annarosa Serpe(&)

Department of Mathematics and Computer Science, University of Calabria,

Rende, CS, Italy
[email protected]

Abstract. The paper focuses on solving the nonlinear equation f ð xÞ ¼ 0; one

of the classic topics of Numerical Analysis present in the syllabus of experi-
mental sections of Italian high schools in secondary education. The main
objective of this paper is to propose an example of constructivist teaching
practice emphasizing the computational approach with the use of MATLAB
software.
MATLAB is a high-performance language for technical computing, but it is
also suitable for high school maths class teaching because of its powerful
numeric engine, combined with interactive visualization tools. All this helps to
keep teaching and learning of this part of mathematics alive and attractive.

Keywords: Iterative methods  Math education  MATLAB software

1 Introduction

Solving nonlinear equations has been a part of the Maths syllabus of the experimental
sections of the Italian secondary school for three decades and is often present in the
questions of the written test of the State Exams (National High School Examination).
Despite its considerable relevance, in classroom teaching practice, this topic is often
overlooked or treated more from a theoretical than from a computational point of view.
It is believed that this constitutes a serious gap in educational standards because the
topic is understood in its essence only if various examples are resolved using a com-
puter, and not by reducing the topic to a series of calculations, mostly done by hand,
often insignificant and detached from real contexts. Nowadays, the computer can act as
mediator between the concrete and the abstract thanks to the relationship between
Mathematics and computer technology [14, 18]. For example, computer programming
favours a highly educational mental training [12, 20, 27] at the same time sensible
reality problems can effectively be solved or simulated. Programming involves the
ability to generate a solution to a problem. Generating solutions means that one of the
learning outcomes is the ability to solve problems and also, if the problem is a big
problem, the ability to split the problem into sub-problems and create a generalizable
central solution. In addition, the student achieves the ability to create usable, readable
and attractive solutions. Programming with MATLAB software not only supports an

© Springer Nature Switzerland AG 2020

Y. D. Sergeyev and D. E. Kvasov (Eds.): NUMTA 2019, LNCS 11973, pp. 463–477, 2020.
https://doi.org/10.1007/978-3-030-39081-5_40
464 A. Serpe

education of Applied Mathematics but at the same time provides teachers with tools
and ideas conducive to engaging the students in activities-based learning of mathe-
matics encouraging several aspects of maths - empirical, speculative, formal, and
reflexive [6].
This paper is part of the debate surrounding programming in high school Maths,
with special focus on implementing algorithms related to appropriate methods for
solving the nonlinear equation f ð xÞ ¼ 0; with the use of MATLAB software. The aim
is to favour the algorithm as a synthesis between the syntactic and semantic aspects of
mathematical objects as well as to encourage computer simulation interpreted as a
‘physical experiment’ and a source of conjecture. The use of algorithms is a compu-
tational attempt to model and thus make problems in the real world ‘effectively
computable’.
The computational approach is very relevant to high school Maths education, and
in this specific instance it helps students to relate the different semiotic representations
of the iterative methods studied, to compare them and to analyse the pros and cons of
applying them. In this way the students discover maths concepts and make general-
izations, thus developing and promoting a different kind of mathematical thinking.
The paper is structured as follows. Section 2 presents the methodological frame-
work used. In Sect. 3 the teaching design of the computational approach is imple-
mented, followed by reflections on the comparative analysis of the algorithms, and the
advantages and disadvantages of iterative methods to estimate roots of equations.
Finally, in Sect. 4 conclusions and recommendations are suggested.

2 Methodological Framework

In general, the appropriate methods for solving the nonlinear equation f ð xÞ ¼ 0 are
iterative methods. They are divided into two groups: Open methods (OM) and
Bracketing methods (BM).
The OMs require only a single starting value or two starting values that do not
necessarily bracket a root. They may diverge as the computation progresses, but when
they do converge, they usually do so much faster than BMs. Some of the best known
OMs are Secant method, Newton-Raphson method, and Muller’s method; whereas the
BMs are Bisection method and False Position Method. In Italian high school, the
teaching of iterative methods is often characterized by over-emphasizing algorithmic
developments and procedural handling of the symbolic aspects of mathematical
objects. In this style of teaching, the students construct a partial mathematical
knowledge consisting mainly of algorithms, a situation that makes them manipulate
symbols routinely, without giving significance to the basic concepts of iterative
methods. The lack of articulation between the different semiotic registers [9] that
should be acquired does not allow students to form an adequate comprehension of the
mathematical concepts involved.
The difficulty lies in making students understand the main difference between a
numerical method and a numerical algorithm. The immediate consequence of this
distinction is that the same numerical method can be implemented using different
algorithms; in particular, the different algorithms will not be equivalent, but some will
 A Computational Approach with MATLAB Software 465

be better than others. The support of mathematical software such as MATLAB can
facilitate this learning process because it provides a means to articulate the different
semiotic registers of a concept. In fact, a computational approach in the classroom
helps to bring out two fundamental aspects:
1. a numerical method allows us to approximate a value otherwise impossible to
determine with algebraic methods;
2. a numerical method is often characterized by the need to perform many, relatively
simple, calculations based on a repetitive scheme, for this reason numerical algo-
rithms are almost exclusively used for the realization of calculating programs
performed by a computer.
The algorithm is a strong conceptual tool in that it involves all sorts of technical and
intellectual inferences, interventions and filters [11]; therefore, it has always been a
driving force in the history of mathematics [4]. Due to this, the choice of a particular
algorithm for solving a given problem is the result of an analysis of the problem and a
comparative analysis of the various algorithms, based on their cost and their accuracy,
as well as intuition refined by experience. Therefore, computer programming is an
unavoidable part in the design of algorithm.
The contribution of programming in the learning of school mathematics has been
demonstrated in numerous projects and research settings [2, 5, 10, 17, 18, 25, 30].
Computer programming has been described by Nickerson (1982) [21] as a creative
endeavour that requires planning, accuracy in language use, generation and testing of
hypotheses, and ability to identify action sequences. It ought to represent a fundamental
part of the literacy for twenty-first century citizens [26] as it is now a skill required for
most jobs and spare time activities. While this is certainly true, one should not forget
the psychological perspectives of computing in Mathematics [16].
The value of programming in educational and pedagogical terms has been recog-
nized internationally [1, 5, 8, 19, 22–24, 28, 29]. To this end, it is important to stress
the importance that algorithms are independent of the programming language used, and
each algorithm can be expressed in different programming languages. The design of an
algorithm is a demanding intellectual activity, significantly more difficult than
expressing the algorithm as a program.
Programming with MATLAB not only helps to reinforce traditional mathematics
learning, but also the teaching of numerical methods for solving some classical
mathematical problems using computers. The students can create the algorithm to get
the solution or they can plot the graph of the function and view it as a graphic
visualization. In this way, students assimilate the notions and procedures of modelling
real-world problems as mathematical problems as well as translating mathematical
solutions into real-world solutions. All this motivates the students to appreciate the
relevance of computational components in classical mathematics materials. In bringing
students closer to the culture of analysis and numerical methods, the teacher plays a
fundamental role because she/he must harmonize theory, methods and algorithms in a
constructive, lively and understandable way. This requires a change of vision in the
teaching which - in most cases - prefers a theoretical style which is scarcely usable for
students as it lacks models and examples. A different approach, such as the compu-
tational one, makes it easier for students to make sense of what they study by maturing
466 A. Serpe

the awareness that the numerical methods constitute an aspect of applied mathematics.
The teaching of iterative methods for solving the nonlinear equation f ð xÞ ¼ 0, effec-
tively, is dependent upon a wide range of factors, but among the most important are
those which are associated with activities and practice within the educational process.
Coming from an applied mathematics perspective, simulations with the computer can
be seen as a part of the modelling process. As a result of this, the teaching practice
design proposed is based on the ‘Numerical modelling cycle’ (Fig. 1) by Sonar [31]
with the use of MATLAB.

Fig. 1. Numerical modelling cycle [31].

From Fig. 1, it is clear that the conclusion of each modelling cycle is the imple-
mentation of a computer simulation program. Of course, numerical algorithms play a
central role in this implementation of the model because they encode a particular type
of abstraction. The algorithm bridges the gap between code and implementation,
between software and experience. An implemented algorithm is, on the one hand, an
intellectual gesture, and, on the other, a functioning system that incorporates in its
structure the material assumptions about perception, decision and communication. In
the end, MATLAB is used to translate the language of mathematics into the language
of the computer [15].

3 Teaching the Iterative Methods with MATLAB

One of the basic principles of numerical analysis is iteration. In general, the idea of
iteration indicates the repetition of a simple process to improve the estimation of the
solution of a more complicated problem. Iterative methods are important in solving
many of today’s real-world problems [7], so it is important that your first approach to
these methods be as positive and simple, as well as informative and educational, as
possible. Based on the outlined methodological framework, the first step is to introduce
a real-world problem.
 A Computational Approach with MATLAB Software 467

3.1 Real-World Problem

The houses of Charles and David are located on two adjacent plots separated by a 2-
meter-high fence. As summer approaches, Charles and David decide to paint the
exterior walls of their houses overlooking the fence. To do this, Charles and Davis use
two ladders respectively 8 m and 6 m long (see Fig. 2).
Calculate the distance between the walls of the two houses.

Fig. 2. Real task representation.

3.2 Mathematical Model

The position of the two ladders recalls two right-angled triangles with a common
cathetus (see Fig. 3). From a closer observation it is possible to identify four right-
angled triangles two by two similar to each other.

Fig. 3. Geometrical task representation.

Let AB ¼ x, applying the Pythagorean theorem to right-angled triangles ABD and

ACB we have:
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2 2 pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 2 2 pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
BD ¼ AD  AB ¼ 64  x2 ; and AC ¼ CB  AB ¼ 36  x2 :

Applying Thales’ theorem to triangles ACB, HEB, and ABD, AHE respectively,
we have the following proportions:
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
x : HB ¼ 36  x2 : 2; x : AH ¼ 64  x2 : 2
468 A. Serpe

By hypothesis AH ¼ AB  HB, and after the easy substitutions we have that:

 
1 1 1
2x pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi þ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi  ¼0
64  x2 36  x2 2

Now by the cancellation law for multiplication:

 
1 1 1
2x ¼ 0 or pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi þ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi  ¼0
64  x2 36  x2 2

In the first case, 2x ¼ 0 is not possible since x 6¼ 0 by hypothesis (x ¼ AB is the

distance between the walls of the two houses); then this implies that:
 
1 1 1
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi þ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi  ¼0 ð1Þ
64  x 2 36  x 2 2

Quantitative Analysis. Equation (1) is not easy to solve; In fact, in this case we do
not have any solving formula. Numerically, (1) can be equivalently expressed as
f ð xÞ ¼ 0, where f : ½a; b ! R is a real function in a finite interval ½a; b  R.
The domain of f ð xÞ is: 8x 2 6; 6½  R. Since x represents the distance between
the walls of the two houses, then the interval will be considered 0; 6½  R. With the
use of the MATLAB, it is possible to draw the graph of f ð xÞ. To do this let us consider
a grid of the domain by point xi 2 ½0; 6½; i ¼ 1; . . .:; n, and let us compute the corre-
sponding values yi ¼ f ðxi Þ; i ¼ 1; . . .; n: The naturalness of the method is accompanied
by a theoretical support that guarantees its convergence property: the plot of the
approximant is approaching the exact one by increasing n, i.e. the number of consid-
ered points1. Implementing MATLAB we need few instructions on the ‘prompt of
command’:
>> x = [0:0.001:5.9];
>> y = 1./sqrt(64-x.^2) + 1./sqrt(36-x.^2)-1./2;
>> plot(x,y);
>> plot(x,y,’LineWidth’,1.4);
>> plot(x,y,’b’,x,0*x,’k’,’LineWidth’,1.2);
The boundary point of the domain and the number of equispaced points are fixed,
the function and vector of points are created (abscissa and ordinate of the points,
respectively) and then the function is displayed through the code plot(x,y).
This creates a 2-D line plot of the data in Y versus the corresponding values in X.
The last two commands are needed to set the plot’s properties of the displayed curve
(colour, kind of line, etc.).

1
This is polynomial interpolation field.
 A Computational Approach with MATLAB Software 469

Fig. 4. Plot of function (1) in the interval [0, 5.9), for n ¼ 100.

This plot (Fig. 4) help the students to visualize the proprieties of function (1), and
to explicit the fact that a function typically changes sign in the vicinity of a root.
The presence of one root is suggested at about x ¼ 5:4 where f ð xÞ appears to be
tangent to the x axis. The graphic interpretation, as well as offering a rough estimate of
the root, is an important tool for understanding the properties of the function and
anticipating the pitfalls of numerical methods. Beyond its usefulness, the graphic
technique used has a limited practical value because it is not precise.

3.3 The Bisection Method

Quantitative analysis returns a rough estimate of the root. This estimate can be used as
an initial hypothesis to introduce iterative methods. In fact, function (1) is continuous
on the interval ½0; 6½  R and changes sign on opposite sides of the root. Starting from
this observation it possible to more precisely identify the root by dividing the interval
into a number of subintervals. Each of these subintervals is searched to locate the sign
change. This process can obviously be continued until we hit the point where the
interval has become so small that the root is determined with sufficient accuracy [3].
For example, see Fig. 5.

Fig. 5. Plot of function (1) in the interval ½4:5; 5:5for n ¼ 100:

470 A. Serpe

By using this information, most numerical methods for f ð xÞ ¼ 0 compute a

sequence of increasingly accurate estimates of the root.
These methods are called iterative methods. We assume that f : R ! R i.e.,
f ð xÞ ¼ 0 is a function that is real valued and that x is a real variable. Suppose that
f ð xÞ ¼ 0 is continuous on an interval ½a; b, and f ðaÞ  f ðbÞ\0:
Then f ð xÞ changes sign on ½a; b, and f ð xÞ has at least one root on the interval. The
simplest numerical procedure for finding a root is to repeatedly halve the interval ½a; b,
keeping the half for which f ð xÞ changes sign. This procedure is called the Bisection
method, and is guaranteed to converge to a root denoted here by a.
Specifically, the procedure is as follows. Define a1 ¼ a; b1 ¼ b. Then
for n ¼ 1; 2; 3; . . . do

 xn ¼ 1 ð an þ bn Þ
 2
 if f ðxn Þ ¼ 0 then a ¼ xn :

 if f ðxn Þ \ 0 then an þ 1 ¼ bn þ 1 else

 an þ 1 ¼ an ; bn þ 1 ¼ x n

Since bn  an ¼ 2ðn1Þ ðb  aÞ; n ¼ 1; 2; 3; . . .:; and xn is the midpoint of ½an ; bn 

if a is the root eventually captured, we have

1 ba
jxn  aj  ðan þ bn Þ ¼ n : ð2Þ
2 2

Although convergence of an iterative process is certainly desirable, it takes more

than just convergence to make it practical. What one wants is fast convergence. A basic
concept to measure the speed of convergence is the order of convergence [13].
Therefore, it is necessary to recall the definition of linear convergence.
Definition. Linear convergence [13]. We can say that xn converges to a (at least)
linearly if

jxn  aj  en ; ð3Þ

where fen g is a positive sequence satisfying

en þ 1
lim ¼ c; 0\c\1: ð4Þ
n!1 en

If (3) and (4) hold with the inequality in (3) replaced by an equality, then c is called
the asymptotic error constant.
Thus, (3) holds with en ¼ 2n ðb  aÞ and

en þ 1 1
¼ ; all n: ð5Þ
en 2
 A Computational Approach with MATLAB Software 471

This shows that the Bisection method converges (at least) linearly with asymptotic
error constant (for the bound en ) equal to 12.
Given an (absolute) error tolerance Tol > 0, the error in (2) will be less than or
equal to Tol if

ba
 Tol:
2n

Solved explicitly for n, this will be satisfied if

log ba
n¼ Tol
;
log 2

where d xe denotes the “ceiling” of x (i.e., the smallest integer  x).

Thus, we know a priori how many steps are necessary to achieve a prescribed
accuracy [13]. The Bisection method is also known as dichotomy or binary chopping or
the half-interval method. The procedure is called the bracketing method because two
initial guesses for the root are required. As the name implies, the guesses must
‘bracket’, or be on either side of, the root.
Bisection Algorithm. The construction of the algorithm is an important phase to
define the ‘finite sequence of steps’ that allows the computer to get to the solution.
A simple algorithm for bisection calculation is as follows. The step are as follows:

½Initialization n ¼ 1; an ¼ a; bn ¼ b
1
½Bisection Iteration xn ¼ ðan þ bn Þ
2
½Convergence Test If jf ðxn Þj  e then the zero is xn : Stop:
If f ðan Þ  f ðxn Þ\0 then bn ¼ xn ; else an ¼ xn :
n ¼ n þ 1 and repeat step 2 until convergence is achieved:

3.4 Computer Simulation of Bisection Method in MATLAB Code

When implementing the procedure on a computer, it is clearly unnecessary to
provide arrays to store an ; bn ; xn ; one simply keeps overwriting. Assuming a and b have
been initialized and TOL assigned, one could use the following MATLAB program
code:
472 A. Serpe

% Bisection method
function [alfa,n]=bisection(f,aa,bb,TOL)
n=1;
a(n)=aa; b(n)=bb;
error(n)=(bb-aa)/2;
x(n)=(a(n)+b(n))/2;
while error(n)>=TOL
if f(x(n))==0
alfa=x(n);
return
else if f(a(n))*f(x(n))<0
a(n+1)=a(n);
b(n+1)=x(n);
else
a(n+1)=x(n);
b(n+1)=b(n);
end
n=n+1;
error(n)=(b(n)-a(n))/2;
x(n)=(a(n)+b(n))/2;
end
alfa=x(n-1);
% The analysis of error
[a',b',x',f(a)',f(b)',f(x)',abs(b-a)', error']
subplot(1,2,1)
plot(1:n,error,'-r.')
title('error')
xlabel('iterations')
subplot(1,2,2)
semilogy(1:n,error,'-k')
grid on
title('log(error)')
xlabel('iterations');
% Check on the number of iterations
ceil(log((b(1)-a(1))/TOL)/log(2));

For the above mentioned MATLAB code an M-file used as a function was
created2.

2
To write our program in MATLAB it is important to distinguish between a program (or MATLAB
script) and a function. MATLAB scripts are “main programs” and functions are that which is written
and can be used in them. The functions must be stored in the same directory where the script is which
calls them. A MATLAB-script is stored as an M-file (a file with the suffix .m) and is executed in the
command window by typing its name without the suffix.
 A Computational Approach with MATLAB Software 473

A function has input parameters and delivers results as output parameters. So for
our programme code (Bisection method), the function utilized in the ‘prompt of
command’ is: [alfa,n] = bisection(f,aa,bb,TOL).
This function has four input parameters and two output parameters. The input
parameters are the boundary points a and b, the required tolerance (TOL) and the
function (1). The output parameters are the root of the equation and the number of
iterations.
In this contest it is important to highlight that in MATLAB the index number 0
(zero) is not allowed. Therefore, the iterative numbering (and that of the vector com-
ponents) always starts from 1 (one).

3.5 Measurement and Output with MATLAB for Bisection Method

From the code proposed in Sect. (3.4), executing the following commands
>> f = @(x) 1./sqrt(64-x.^2) + 1./sqrt(36-x.^2)-1./2;
>> [alfa,n] = bisection(f,5,5.6,10^(-4))
you obtain the information requested. The output (Fig. 6) shows in the columns:
• 1 through to 3 the value of an ; bn ; xn :
• 4 through to 6 the value of f ðan Þ; f ðbn Þ; f ðxn Þ:
• 7 through to 8 the absolute value jbn  an j and ba
2n .

The numerical value of the root and the iteration are shown at the end of the output.
After thirteen iterations the estimation of the root searched for is:

a ¼ 5:206689453125000;

and the Error ¼ 0:000073242187500.

The fact that the error has a linear trend on a logarithmic scale tells us that the
Bisection method has a convergence order of 1 or simply converges in a linear way: it
tends to 0 with the same speed as 1/n. This translates into the fact that at each step you
always gain the same number of exact significant digits. In particular, we can observe
from the output (Fig. 6) that in about every 3 steps a correct decimal figure is gained.
The explanation lies in the fact that the Bisection method divides the current interval
into two equal parts. Translated into machine terms3, a binary figure (BIT = BInary
digiT) is exactly arrived at in each step. Since it takes more than 3 bits to do 104 this
explains why you get a figure in base 10 roughly every 3 steps of the method. In this
way the students “touch” computer arithmetic.

3
Which are stored in base 2 in the systems currently in use (Standard IEEE754).
4
Indeed, 23 = 8 e 24 = 16 or, more simply log210 = 3.32…..
474 A. Serpe

Fig. 6. Output for the Bisection method applied to solve (1)

Figure 7 show two plots indicating the error on the basis of number of iterations.
Bisection Method Analysis. What are the pros and cons of the Bisection method?
Since the Bisection method discards 50% of the current intervals at each step, it
brackets the root much more quickly than the incremental search method does.
The pros are as follows:
– it is always convergent. Since the method brackets the root, the method is guar-
anteed to converge;
– as iterations are conducted, the interval gets halved. So one can guarantee the error
in the solution of the equation.
 A Computational Approach with MATLAB Software 475

Fig. 7. Plot of the error as a function of the number of iterations for the Bisection method.

Since the Bisection method is like an incremental search, the following cons occur:
– it only finds roots where the function crosses the x axis. It cannot find roots where
the function is tangent to the x axis;
– it can be fooled by singularities in the function;
– it cannot find complex roots of polynomials.
Analyzing the pros and cons, we can assert that the advantage of the Bisection
method is that it is guaranteed to be convergent.
A disadvantage of the Bisection method is that it cannot detect multiple roots. In
general, the Bisection method is used to get an initial rough approximation of the
solution. Then faster converging methods are used to find the solution.

4 Conclusions

The computational approach utilized can be part of a more general didactical strategy
and in no case can substitute for the teacher who has to choose and develop the
didactical practice to support -step by step- the activity of the students, helping them to
organize data and encourage creativity. Particularly, when software such as MATLAB
is used, the teacher must plan the process of designing which starts from the defining
step and she/he must structure the learning material in order to provide the students
with tools to critically probe the concepts and to validate them with formal proofs in
order and draw conclusions. These tools are important when searching for ways to
work with students by simulation with mathematical models.
The computational approach brings students closer to the culture of analysis and
numerical methods in an easy and understandable way because emphasis is given to the
knowledge and to accompanying justifications. The basic pedagogy is that of making
the students ‘experience’ rather than trying to provide them with all the concepts and
results on the subject. The experience leads the students to a careful reading of the
476 A. Serpe

cultural aspects of the computation thus maturing the awareness that all these aspects
are held together in the humble container of the algorithm. In fact, through the con-
struction and implementation by computer the student has the real possibility of
thinking in computational terms (understanding what estimating, analysing the accu-
racy and execution of that estimation means).
The algorithm then becomes a “cultural machine” because it produces objects,
processes and cultural experiences [11]. Mechanically memorizing procedures is use-
less, students must learn the basic concepts.
Numerical mathematics is also experimental. Much can be learned simply by doing
tests and observing how the calculation proceeds thanks to the algorithm as a stimu-
lating and reflexive procedural criticism.

References
1. Aydin, E.: The use of computers in mathematics education: a paradigm shift from “computer
assisted instruction” towards “student programming”. Turk. Online J. Educ. Technol.
TOJET 4(2), 27–34 (2005)
2. Breed, E.A., Monteith, J.L. de K., Mentz, E.: Effective learning in computer programming:
the role of learners’ reflective thinking. In: Samways, B. (ed.) Proceedings 8th World
Conference on Computers in Education – WCCE 2005, Western Cape, South Africa (2005)
3. Chapra, S.C., Canale, R.P.: Numerical Methods for Engineers. McGraw-Hill Higher
Education, Boston (2010)
4. Costabile, F.A., Serpe, A.: Archimedes in secondary schools: a teaching proposal for the
math curriculum. In: Paipetis, S., Ceccarelli, M. (eds.) The Genius of Archimedes–23
Centuries of Influence on Mathematics, Science and Engineering, vol. 11, pp. 479–491.
Springer, Dordrecht (2010). https://doi.org/10.1007/978-90-481-9091-1_36
5. Costabile, F.A., Serpe, A.: Computer-based mathematics instructions with MATCOS: a
pedagogical experiment. In: Handbook of Research on Didactic Strategies and Technologies
for Education: Incorporating Advancements, pp. 724–738. IGI Global (2013). https://doi.
org/10.4018/978-1-4666-2122-0.ch063
6. Cretchley, P., Harman, C., Ellerton, N., Fogarty, G.: MATLAB in early undergraduate
mathematics: an investigation into the effects of scientific software on learning. Math. Educ.
Res. J. 12(3), 219–233 (2000). https://doi.org/10.1007/BF03217086
7. Daponte, P., Grimaldi, D., Molinaro, A., Sergeyev, Y.D.: Fast detection of the first zero-
crossing in a measurement signal set. Measurement 19(1), 29–39 (1996). https://doi.org/10.
1016/S0263-2241(96)00059-0
8. Dubinsky, E., Tall, D.: Advanced mathematical thinking and the computer. In: Tall, D. (ed.)
Advanced Mathematical Thinking. Mathematics Education Library, vol. 11, pp. 231–248.
Springer, Dordrecht (2002). https://doi.org/10.1007/0-306-47203-1_14
9. Duval, R.: Representation, Vision and Visualization: Cognitive Functions in Mathematical
Thinking. Basic Issues for Learning (1999)
10. Feurzeig, W., Papert, S., Bloom, M., Grant, R., Solomon, C.: Programming-language as a
conceptual framework for teaching mathematics. Newslett. SIGCUE Outlook 4(2), 13–17
(1970)
11. Finn, E.: What Algorithms Want: Imagination in the Age of Computing. MIT Press, Boca
Raton (2017)
 A Computational Approach with MATLAB Software 477

12. Frassia, M.G., Serpe, A.: Learning geometry through mathematical modelling: an example
with GeoGebra. Turk. Online J. Educ. Technol. 2017, 411–418 (2017). (November Special
Issue INTE)
13. Gautschi, W.: Numerical Analysis. Springer, Heidelberg (2012). https://doi.org/10.1007/
978-0-8176-8259-0
14. Goos, M., Galbraith, P., Renshaw, P., Geiger, V.: Perspectives on technology mediated
learning in secondary school mathematics classrooms. J. Math. Behav. 22(1), 73–89 (2003).
https://doi.org/10.1016/S0732-3123(03)00005-1
15. Guide, Getting Started. “MATLAB® 7” (2009)
16. Hatfield, L.: Toward comprehensive instructional computing in mathematics. Comput. Math.
Educ. 1–9 (1984)
17. Johnson, D.C. Educ. Inf. Technol. 5, 201 (2000). https://doi.org/10.1023/A:1009658802970
18. Kurland, D.M., Pea, R.D., Clement, C., Mawby, R.: A study of the development of
programming ability and thinking skills in high school students. J. Educ. Comput. Res. 2(4),
429–458 (1986). https://doi.org/10.2190/BKML-B1QV-KDN4-8ULH
19. Liao, Y.K.C., Bright, G.W.: Effects of computer programming on cognitive outcomes: a
meta-analysis. J. Educ. Comput. Res. 7(3), 251–268 (1991). https://doi.org/10.1016/j.chb.
2014.09.012
20. Lye, S.Y., Koh, J.H.L.: Review on teaching and learning of computational thinking through
programming: what is next for K-12? Comput. Hum. Behav. 41, 51–61 (2014). https://doi.
org/10.1016/j.chb.2014.09.012
21. Nickerson, R.S.: Computer programming as a vehicle for teaching thinking skills. Thinking:
J. Philos. Children 4(3/4), 42–48 (1983). https://doi.org/10.5840/thinking19834310
22. Oprea, J.M.: Computer programming and mathematical thinking. J. Math. Behav. 7, 175–
190 (1988)
23. Papert, S.: Mindstorms: Children, Computers, and Powerful Ideas. Basic Books, Inc., New
York (1980)
24. Pea, R.D., Kurland, D.M.: On the cognitive effects of learning computer programming. New
Ideas Psychol. 2(2), 137–168 (1984). https://doi.org/10.1016/0732-118X(84)90018-7
25. Robins, A., Rountree, J., Rountree, N.: Learning and teaching programming: a review and
discussion. Comput. Sci. Educ. 13(2), 137–172 (2003). https://doi.org/10.1076/csed.13.2.
137.14200
26. Rushkoff, D.: Program or be Programmed: Ten Commands for a Digital Age. Or Books
(2010)
27. Saeli, M., Perrenet, J., Jochems, W.M., Zwaneveld, B.: Teaching programming in secondary
school: a pedagogical content knowledge perspective. Inform. Educ. 10(1), 73–88 (2011).
https://academic.microsoft.com/paper/1509940907
28. Serpe, A., Frassia, M.G.: Computer-based activity’s development for probability education
in high school. Turk. Online J. Educ. Technol. 2017, 613–621 (2017). (October Special
Issue INTE)
29. Serpe, A., Frassia, M.G.: Technology will solve student’s probability misconceptions:
integrating simulation, algorithms and programming. In: Dooley, T., Gueudet, G. (eds.)
Proceedings of 10th Congress of the European-Society-for-Research-in-Mathematics-
Education (CERME 10), pp. 828–835 (2017). http://www.mathematik.uni-dound.de/ieem/
erme_temp/CERME10_Proceedings_final.pdf
30. Sfard, A., Leron, U.: Just give me a computer and i will move the earth: programming as a
catalyst of a cultural revolution in the mathematics classroom. Int. J. Comput. Math. Learn. 1
(2), 189–195 (1996). https://doi.org/10.1007/BF00571078
31. Sonar, T.: Angewandte Mathematik Modellbildung und Informatik, Vieweg-Verlag,
Braunschweig. Wiesbaden (2001). https://doi.org/10.1007/978-3-322-80225-5
 Task Mathematical Modelling Design
in a Dynamic Geometry Environment:
Archimedean Spiral’s Algorithm

Annarosa Serpe1(&) and Maria Giovanna Frassia2

1
Department of Mathematics and Computer Science, University of Calabria,
Rende, CS, Italy
[email protected]
2
IIS IPSIA-ITI “E. Aletti”, Trebisacce, CS, Italy
[email protected]

Abstract. Over the last twenty years, several research studies have recognized
that integrating, not simply adding, technology (Computer Algebra System -
CAS, Dynamic Geometry Software - DGS, spreadsheets, programming envi-
ronments, etc.) in the teaching of mathematics helps students develop essential
understandings about the nature, use, and limits of the tool and promotes deeper
understanding of the mathematical concepts involved. Moreover, the use of
technology in the Mathematics curricula can be important in providing the
essential support to make mathematical modelling a more accessible mathe-
matical activity for students. This paper presents an example of how technology
can play a pivotal role in providing support to explore, represent and resolve
tasks of mathematical modelling in the classroom. Specifically, a mathematical
modelling task design on the tracing of Archimedean spiral with use of a
Dynamic Geometry Environment is shown. The aim is to emphasize the
meaning and the semantic value of this rich field of study that combines tangible
objects and practical mechanisms with abstract mathematics.

Keywords: Archimedes’ spiral
Dynamic Geometry Environment

Mathematical modelling task

1 Introduction

Research on the use of technology in secondary Mathematics education has proliferated

over the last 30 years and has demonstrated that the strategic use of technological tools
such as graphing calculators, Dynamic Geometry Environment (DGE), programming
environment and spreadsheets support students’ mathematical thinking and discourse
[3, 4, 6, 8, 16]. At the same time, awareness and interest in students’ mathematical
thinking, reasoning, and sense-making has increased [7], so much so that in the last
twenty years many researchers have focused their research studies on how the use of
technology can support students’ mathematical thinking and reasoning [31].
Three themes emerge as researchers have examined mathematical thinking in the
context of mathematical activity that occurs in technological environments [17]:

© Springer Nature Switzerland AG 2020

Y. D. Sergeyev and D. E. Kvasov (Eds.): NUMTA 2019, LNCS 11973, pp. 478–491, 2020.
https://doi.org/10.1007/978-3-030-39081-5_41
 Task Mathematical Modelling Design in a Dynamic Geometry Environment 479

– the extent to which students develop tools for mathematical thinking and learning in
these environments;
– the ways in which students engage in metacognitive activity as a result of mathe-
matical activity in these environments;
– the level of generality of the mathematical thinking of students in these
environments.
As regards the first theme, some technological instruments facilitate symbolic
reasoning and/or conceptual understanding, but they can also inhibit mathematical
thought. Some studies indeed highlight how symbolic manipulation skills do not
automatically develop in a technology-based environment [34], but are the result of
adequate integration of technology into classroom practice.
As far as the second theme is concerned, technological environments have two
features that may enhance metacognitive activity: when used as tools they have the
capacity to offload some of the routine work associated with mathematical activity
leaving more time for reflection, and with their strict communication requirements they
may help bring to consciousness mathematical ideas and procedures [17].
Regarding the third theme, some studies reported a higher level of generality in the
thinking of students in the technology-based activity.
In the context of geometry, Hollebrand and colleagues in [17] observed that
learners’ generalizations were likely to be situated abstractions, not generalizing
beyond the situation in which they were developed. Doerr and Pratt in [17, p. 428]
noted ‘the lack of evidence for any claim that the students reasoned better about real-
world phenomena, even when they reasoned appropriately within the microworld’ and
that ‘there is however little evidence that students can abstract beyond the modelling
context’. The DGE were developed in the mid-1980s, in the thread of the powerful idea
of ‘direct manipulation’ [11] to simulate ruler and compass constructions and
assist/help in the precise design of geometric figures. A DGE is a computer microworld
with Euclidean geometry as the embedded infrastructure. In this computational envi-
ronment, a person can evoke geometrical figures and interact with them [18]. It is a
virtual mathematical reality where abstract concepts and constructions can be visually
reified. In particular, the traditional deductive logic and linguistic-based representation
of geometrical knowledge can be re-interpreted, or even refined, in DGE as dynamic
and process-based interactive ‘motion picture’ in real time [23]. The rapid evolution of
Geometer’s Sketchpad and Cabri Géomètre, the first two DGEs, highlighted the need
for increased dynamism. Sketchpad, for example, was initially conceived as a program
for drawing accurate static figures of Euclidean geometry. In the course of their
development, the initial idea for the different DGEs was adapted to the emerging needs.
The result was software able to build dynamic geometric shapes, in which points and
segments could be dragged maintaining the properties that characterize the constructed
figures. The relatively fast drag and drop operation (dragging) defined the future of
DGE: the functional versatility and corresponding complexity of the operation were not
anticipated, and have only gradually been dealt with [1]. Interest in DGE for learning
geometry was evident right from the start: DGE through dragging offered new possi-
bilities for the visualization and verification of the properties of geometric objects.
480 A. Serpe and M. G. Frassia

This has been recognized by the various educational areas that have integrated
DGE in their curricula [20]. Direct manipulation, a special feature of DGE, allows the
student to have a simultaneous response; according to [22], this simultaneity is a key
element which can shorten the distance between experimental and theoretical Mathe-
matics, or for switching between conjecture and formalization.
Laborde in [16] illustrates this potential for the teaching of geometry with an
example on the “black box” situations:
In the black box situations, the students are given a diagram on the screen of the computer and
they are asked questions about it. This kind of situation was used in our scenarios for intro-
ducing new trasformations. A point P and its image P’ through an unknown transformation
were given to the students. They could move P and observe the subsequent effect on P0 . Students
were asked to find the properties of the unknown transformation by means of this black box. In
such a task, students must ask themselves questions about the transformation.

This process emphasizes the students’ responsibility in the formulation of questions

concerning the transformation, but at the same time provides food for thought on the
fact that this may well distract learners from the main goal, which is to identify
geometric properties. On the other hand, before they can use the software properly, the
students require a preliminary phase of instruction on the functions offered by DGE
(including the dragging1); only at a later time should they experience the constraints
and relations that restrict the movement of the points and of the figures in general [19].
This procedure adds useful meaning to the validation of the constructed figures.
In addition to the function of dragging, the DGE also have the following
features/affordances:
– measuring (of lengths of segments, the amplitudes of angles, shapes of areas, …);
– tracing, place, animation (that let you see the evolution of models);
– The representation of functions and the investigation of their graph, at local or
general level;
– The integration of different representation registers (such as the geometrical and
analytical), that allows you to model problematic situations.
The potential of DGE is not just restricted to the field of geometry, but it can be
exploited to introduce other notions of mathematics (e.g. analysis, probability, etc.).
Furthermore, the geometrical situations can also be associated with specific prob-
lems, in which the DGE help in the visualization of the dynamic aspects of geometrical
objects [11]. In this prospective, the paper presents a mathematical modelling task
design in a DGE. The paper is structured in four sections.
Section 2 covers the theoretical framework adopted. Section 3 covers all actions of
task based on the tracing method – in Euclidean plane – of Archimedean spiral with the
use of GeoGebra software. Section 4 covers the conclusions.

1
Dragging in DGE is a powerful dynamic tool to acquire mathematical knowledge. It serves as a kind
of interactive amplification tool for leaner to see global behavior via variation.
 Task Mathematical Modelling Design in a Dynamic Geometry Environment 481

2 Theoretical Framework

Mathematics education is characterized by a high conceptual standard in regard to the

development of central terms and through the construction of numerous algorithmic
processes. In both areas, great importance can be attached to the use of digital math-
ematics tools. As mentioned above, the digital tools are not only a pedagogical medium
for organizing processes in education, in particular they strengthen the activity of doing
mathematics, such as experimenting, visualizing, applying, etc. [2, 33].
According to a well-considered use of technology, the tools used are and will
remain cognitive tools because they help to represent and work on the individual
problem (after input by the person that is working on them). Naturally, the work on this
should be designed so that it supports the mental processes of the learners, who control
the learning process, however, they should by no means be restricted. From the point of
view of the discovery of mathematical interrelations, digital mathematics tools are of
particular importance for example in simulations, understood as experimenting with
models [15].
The International Commission on Mathematical Instruction (ICMI2) has promoted
and continues to promote cultural debate in order to really support the students’
development of mathematical modelling competences because relevant for their further
education and for their subsequent professions. Several research studies recognize that
the development of technology creates more opportunities for practicing mathematical
modelling in the classroom [14, 32]. The use of technology can lead to the simplifi-
cation of difficult and complex modeling operations, especially when solving, as [13]
has shown (Fig. 1).

Fig. 1. Mathematical modelling including the use of technology.

2
https://www.mathunion.org/activities/international-commission-mathematical-instruction-icmi.
482 A. Serpe and M. G. Frassia

In the teaching process, the use of technology allows the teacher not only to cope
with traditional content using different methods, but also to explore and encounter new
content. Especially, in the teaching and learning of Euclidean geometry, and solving
problems related to geometry concepts, the DGE are appropriate tools to help enhances
conceptual understanding [12, 28–30].
Duval [9] argued that DGE are superior to paper-and-pencil based (PPB) methods
as they dissociate the “figure” from the “process of drawing”. This allows students to
understand the properties of the figure before it is sketched on the screen.
The use of DGE continually opens up new teaching perspectives in the teaching -
learning of geometry because it enhances the constructive aspect without detracting
from deductive accuracy, from clarity of hypotheses and related consequences per-
taining to the discipline [1, 21]. Thanks to DGE the graphic-constructive phase, both
prior to the acquisition of some concepts and geometrical properties, and subsequently
as verification and/or further study, is not only enjoyable, but also greatly helps
teaching, as it offers both visualization and exemplification and/or exploration.
On the other hand, according to Federigo Enriques [10]:
[…] It does not help to develop with impeccable deduction series of theorems of Euclidean
geometry, if you do not return to contemplate the constructed edifice, inviting the disciples to
distinguish the truly significant geometric properties (for example: the sum of the angles of a
triangle and Pythagoras’ theorem) from those that have value only as links in the chain.

It is not superfluous to recall here that though geometry was created as modelling of
the physical world around us, the Italian teaching tradition has followed the
hypothetical-deductive Euclidean interpretation and has progressively emphasized its
formal character, ‘decontaminated’ from the figural and constructive aspect, even in the
most harmless of terms. In short, the surveyor’s traditional tools (ruler, square ruler,
compasses), retrieved and simulated by DGE, on the one hand facilitate geometrical
intuitions, while on the other raise and stimulate interest and learners’ imagination,
enabling speculation, which is sometimes immediately verifiable, thanks to the
simultaneous computer feedback [28, 29]. The connection between a drawing and a
geometric object in everyday teaching practice is nearly always established through a
process of approximation. This is based on the idea that with subsequent, better
attempts the drawing can eventually achieve something close to the ideal figure.
The essential didactic value of the Euclidean frame has always been the perception
of its nature as a comprehensive frame, which begins with the ‘simple and evident’ and
progresses to the complex and ‘non-evident’. The integrated tools offered by a DGE
represent a valid aid along the way as they progress in the same way from what is
predefined to what is made by the user. According to this perspective, the authors have
chosen GeoGebra software because it is a free open-source DGE for mathematics
teaching and learning. GeoGebra3 combines features of DGE and CAS in a single,
integrated, and easy to-use system. It is may be the bridge for the cognitive gap that
hinders a student from carrying out a modelling task.

3
GeoGebra was created by Markus Hohenwater and now has been translated into 40 languages. Users
all over the world can freely download this software from the official GeoGebra website at http://
www.geogebra.org.
 Task Mathematical Modelling Design in a Dynamic Geometry Environment 483

However, it should also be noted that this software (and technology in general)
should never replace the mathematics, much less the teacher; it should be viewed as a
timely, and sometimes temporary, means of overcoming a difficulty.

3 Mathematical Modelling Task Design

The mathematical modelling task design aims to facilitate the understanding of the
concrete-abstract relationship, in order to the acquisition of meaning of geometrical
objects. The task implies the study of the geometrical method for tracing Archimedean
spiral in Euclidean register using drawing tools such as a ruler, set square and com-
passes and later also with a computer as an educational tool. We chose to use Geo-
Gebra because it is a constructive and creative activity, which reinforces the acquisition
of concepts as well as abstraction skills.
The reason for the choice of the Archimedean spiral lies in the fact that the spiral is
an ambiguous and double curve, on one part it gives the image of expansion and
totality; a curve that, by rotating, always remain similar to itself, but at the same time it
widens and extends to infinity, as if in rotating new parts are always being born from
the centre to move around towards the periphery. However next to this clear and serene
face, the spiral shows a second dark and disturbing one: the movement is transformed
from an expansion to a continuous contraction that hypnotically pulls you into the
centre.
Despite these contrasting images of it, the generation of spiral lines is very simple, a
rare example of depth and at the same time of geometric evidence; two reasons that
have led mathematicians to study their properties since ancient times, regardless of
their scarce applications.
The first and simplest of the spiral lines is the one studied by Archimedes, which
bears his name. Furthermore, the spirals are also at the base of the fractals [5, 24–27],
which may constitute a further topic of further study in the direction of the theoretical
framework outlined. Archimedean spiral is a curve of great charm and beauty. To date
there is no simple and precise tool capable of tracing it and for this reason it is not
easily used for practical purposes4. To draw it is necessary to identify a series of points
on the plane that must be connected by a curve or simply, as more often happens,
tracing the curve freehand. Although governed by a geometric principle5, the Archi-
medean spiral cannot be drawn simply. Some points can be identified, but the task
remains, of having to trace the form continuously. A substantial aid for tracing spiral-
shaped curves comes from the polycentric family, that is from those curves defined by
the appropriate combination of different arcs of circumference, arranged in such a way
that there are no discontinuities in the connections between the various parts. To obtain
this condition it is necessary that at the points of contact between two arcs they have a

4
A concrete image of the Archimedean spiral is the grooves of a vinyl disc, equidistant from each
other and separated by a very small constant distant.
5
The mechanical generation of the spiral is described in a plane by a point that moves with a uniform
motion along a straight line, while the line rotates in a uniform circular motion around a point.
484 A. Serpe and M. G. Frassia

common tangent, so that their rays belong to the same line. In this way the passages
between the curves will be fluid and continuous.
The tracing method for the construction of the Archimedean spiral in the Euclidean
plane represents a strategic phase from the teaching point of view because it helps the
learner to perceive the complexity of the whole, which starts from ‘simple and evident’
and arrives at the ‘complex and not evident’. The interpretation of data with GeoGebra
enables us to turn the meaning of a representation into its construction through a model
which shows and communicates geometrical synthesis. Mathematical modelling has
expressive and evocative potential from the metric and formal point of view; the virtual
tracing of a curve through simple steps adds value to the learning experience.
Table 1 summarizes the competencies that the mathematical modeling task design
aims to achieve, with an indication of the levels that can be pursued.

Table 1. Levels of competencies of mathematical modelling task.

Levels Interpretation Task Mathematical
of task solution reasoning and
computation
1st-level Failure to identify the Lack of Lack of control in the
Initial significant information autonomy in resolution process.
of the task solution Representation with software
strategies and analysis of mathematical
objects with appropriate
guidance
2nd-level Partial identification of Partial Lack of control in the
Basic significant information autonomy in resolution process.
of the task solution Representation with software
strategies and analysis of mathematical
objects partially correct
3rd-level Identification of the Autonomy in Control in the resolution
Intermediate significant information the solution process. Correct
of the task strategies representation of
mathematical objects with
software and analysis
4th-level Identification and Autonomy and Rigor, coherence and logic in
Advanced contextualization of the creativity in the the resolution process.
significant information solution Correct and original
of the task strategies representation of
mathematical objects with
software and analysis

Such a methodology avoids the traditional pitfalls of Mathematics classes, and

gives students a true chance to improve their understanding.
Based on the theoretical framework described above, the mathematical modelling
task design is shown in Fig. 2.
 Task Mathematical Modelling Design in a Dynamic Geometry Environment 485

Fig. 2. Mathematical modelling task design.

The class is divided into groups, and each group gets the required material and tools
(graph paper, ruler, compasses, protractor, pencils and rubbers); then the teacher sets
the task for the classic geometric construction of Archimedean spiral (Interpretation
and solution of the task without the technology). Then, the students do the construction
process with paper and pencil and this is followed by use of the GeoGebra spreadsheet.
The last part of the mathematical modelling task design, when the student use
GeoGebra – mathematical modelling computer-based environment – is divided in to
three actions.

3.1 Action 1 – Classic Construction of Archimedes’ Spiral

Classic construction of Archimedes’ spiral through the “predefined objects” available
in the GeoGebra toolbar: point, regular polygon, median point, half line, circumference
and arc. This procedure will allow an analysis of the construction from a mathematical
point of view.
Algorithm 1 – Archimedes’ spiral. The steps are as follows:
1. Draw two points A and B;
2. Construct the square ABCD with extreme side A and B;
3. Determine the median point (E, F, G, H) of the sides of the square;
4. Construct the rays (e, h, g, f) with origin the median point of the square and passing
through one of the vertexes of the square and pertaining to the side on which the
median point lies (anticlockwise direction);
5. Construct the circumference with centre A and passing through point B;
6. Determine the point of intersection I between the circumference and the ray with
origin the median point of side AB and passing through A;
486 A. Serpe and M. G. Frassia

7. Draw the arc p with centre A and points B and I as extremes;

8. Repeat from 5 to 7:
a. Construct the circumference CC with centre A, B, C, D (in this order) and radius
the distance between the vertex of the square and the point of intersection
between the old circumference and the rays e, h, g, f (in this order);
b. Construct the arc with centre in the vertex of the square (Fig. 3).

Fig. 3. Output of Archimedes’ spiral (initial construction).

At this point the teacher tells the students that the spiral curve is called spiral of
Archimedes6 - the great mathematician from Syracuse - exhibited for the first time in
his work “On the Spirals”, a treatise addressed to mathematician Dositeo of
Alessandria. This treatise was much admired but little read, since it was considered as
the most difficult of all of Archimedes’ works. The spiral is defined as the flat place of a
point which, starting from the end of a ray or half-line, moves uniformly along this
radius while the radius in turn rotates uniformly around its end.
The study that Archimedes made of the spiral was set in the wake of research,
typical of Greek mathematics, aimed at finding solutions to the three famous classic
problems. Ancient testimonies of the spiral are found in some Minoan paintings dating
back to 1650 BC., Although the Minoan were not aware of the properties of the curve,
their representation was realized in an extraordinary way.
The use of history in mathematics education is important for experimenting method
in order to obtain a full learning. Indeed, the history of mathematics offers the possi-
bility to explore the development of mathematical knowledge.

6
Archimedes, (born c. 287 BCE, Syracuse, Sicily [Italy]-died 212/211 BCE, Syracuse), the most-
famous mathematician and inventor in ancient Greece.
 Task Mathematical Modelling Design in a Dynamic Geometry Environment 487

3.2 Action 2 – The Improved Archimedes’ Spiral

In the Archimedean spiral the execution of a turn requires the construction of four arcs
traditionally constructed with ruler and compasses; so the need to simplify the repe-
tition process about the construction of the connected arcs becomes apparent. As a
result, the students must find a “geometrical strategy” for the solution of the problem.
In this case, GeoGebra represents a valid help as it allows for the creation of new
computational tools in the spreadsheet interface, which can then be used as predefined
objects. Students are then asked to design a new tool called ArcSpiral used for the
construction of all the connected arcs so as to reduce the execution time of the previous
algorithm. This action is very sensitive because the students have to face yet another
abstraction leap. They have to plan the design of a new tool called ArcSpiral with the
aim of the constructing the spiral of Archimedes starting from the construction of the
first arc, which requires the identification of the initial (points A, B, median point
E between A and B and the half line EA) and final (point I and arc p) objects.
Algorithm 2 – The improved Archimedes’ spiral. After the first seven steps of
algorithm 1, we have:
1. Create a new tool “ArcSpiral”, having the points A, B and median point E between
A and B, and the half line passing through E and A as initial objects, and points I and
arc p as final objects. The tool thus created can be used by clicking on its icon;
2. Construct the connected spiral arcs, through the tool “ArcSpiral” created, with
appropriate choice of points (Fig. 4).

Fig. 4. Output of Archimedes’ spiral (8 spires).

Refining the previous process of construction is relevant because the students are
required to make a considerable abstractive leap: what is difficult is the identification of
the construction, within the algorithm, which enables the repetition of an operation
(construction of joined arcs) as long as a certain condition is true.
488 A. Serpe and M. G. Frassia

3.3 Action 3 – Optimized Archimedes’ Spiral

Design a new tool “ArcSpiral” which contains a lower number of initial objects.
Planning the “ArcSpiral” tool with only two initial objects (points A and B) and one
final object (arc with centre B and points A and A0 as extremes). This step has the aim of
further perfecting the construction of the curve.
Such a highly educational process trains the students in the appreciation of the
potential of the language of mathematics and at the same time offers them a reading
awareness of theories.
Algorithm 3 – Optimized Archimedes’ spiral. The steps are follows:
1. Draw two points A and B;
2. Draw point A0 , obtained from the rotation of A with respect to centre B by an
anticlockwise angle;
3. Construct the arc with centre B and points A and A0 as extremes;
4. Create a new tool called “Arcofspiral”, with points A and B as initial objects and as
final objects the arc with centre B and A and A0 as extremes; the tool created can be
used by clicking on its icon (Fig. 5).

Fig. 5. Output of the optimized Archimedes’ spiral (12 spires).

4 Conclusions

Today it is important to plan and experiment new modalities for the teaching of
Euclidean geometry, bearing in mind the potential of new technological devices.
The mathematical modelling tasks design suggests important messages for math-
ematics education.
This practical contribution to mathematics education can prove that computer-based
classroom activities can be effectively used in the teaching and learning environment.
Firstly, the students can become familiar with DGE; secondly, geometric facts, figures,
 Task Mathematical Modelling Design in a Dynamic Geometry Environment 489

shapes and their proprieties with the constructions can be observed by using the
software’s features. Thus, the students can have the chance to verify the conditions by
exploring and observing the geometric properties of the shapes with all sufficient
conditions. This also can give opportunities to check and prove all features dynamically
with the program itself. Consequently, the students have the chance to prove the terms
and to observe construction conditions of geometric features for each case. Then,
starting by simply drawing the geometric shapes and figures and providing all con-
ditions for the construction requires students to consider all of the related facts and
features together with associated geometric realities.
The implementation of the first construction algorithm for the curve, repeated
several times, facilitates the understanding and use of the geometrical objects; however,
at the same time the need to shorten the repetition sequence for the construction of the
joined arcs emerges for the first time.
The use of GeoGebra, suitably exploited in teaching practice, favours the struc-
turing of knowledge in meaningful networks thus maturing the students’ skills. It offers
an effective impact on mathematics education and has the potential to promote student-
centred learning and active learning. However, the mathematical modelling task design
is in no way limiting. The creative teacher can use the design as a springboard for new
teaching initiatives which are instructive and engaging.

References
1. Arzarello, F., Olivero, F., Paola, D., Robutti, O.: A cognitive analysis of dragging practices
in Cabri environments. ZDM 34(3), 66–72 (2002). https://doi.org/10.1007/BF02655708
2. Barzel, B., Hußmann, T., Leuders, T.: Computer, internet & co im Mathematik-Unterricht.
Scriptor-Cornelsen, Berlin (2005)
3. Bonanno, A., Camarca, M., Sapia, P., Serpe, A.: Archimedes and caustics: a twofold
multimedia and experimental approach. In: Paipetis, S., Ceccarelli, M. (eds.) The Genius of
Archimedes – 23 Centuries of Influence on Mathematics, Science and Engineering. History
of Mechanism and Machine Science, vol. 11, pp. 45–56. Springer, Dordrecht (2010). https://
doi.org/10.1007/978-90-481-9091-1_4
4. Burrill, G., Allison, J., Breaux, G., Kastberg, S., Leatham, K., Sanchez, W.: Handheld
graphing technology in secondary mathematics: Research findings and implications for
classroom practice. Retrieved from the Texas Instruments website (2002). http://education.ti.
com/sites/UK/downloads/pdf/References/Done/Burrill,G.%20(2000).pdf
5. Caldarola, F.: The Sierpinski curve viewed by numerical computations with infinities and
infinitesimals. Appl. Math. Comput. 318, 321–328 (2018). https://doi.org/10.1016/j.amc.
2017.06.024
6. Costabile, F.A., Sapia, P., Serpe, A.: Archimedes in secondary schools: a teaching proposal
for the math curriculum. In: Paipetis, S., Ceccarelli, M. (eds.) The Genius of Archimedes –
23 Centuries of Influence on Mathematics, Science and Engineering. History of Mechanism
and Machine Science, vol. 11, pp. 479–491. Springer, Dordrecht (2010). https://doi.org/10.
1007/978-90-481-9091-1_36
7. Dick, T.P., Hollebrands, K.F.: Focus in high school mathematics: Technology to support
reasoning and sense making, pp. xi-xvii. National Council of Teachers of Mathematics,
Reston, VA (2011)
490 A. Serpe and M. G. Frassia

8. Drijvers, P., Ball, L., Barzel, B., Heid, M.K., Cao, Y., Maschietto, M.: Uses of Technology
in Lower Secondary Mathematics Education. ITS. Springer, Cham (2016). https://doi.org/10.
1007/978-3-319-33666-4
9. Duval, R.: Geometry from a cognitive point of view. In: Mammana, C., Villani, V. (eds.)
Perspectives on the teaching of geometry for the 21st century, pp. 37–52 (1998)
10. Enriques, F.: Insegnamento dinamico. Periodico di Matematica Serie IV, vol. I, pp. 6–16
(1921)
11. Gueudet, G., Trouche, L.: Mathematics teacher education advanced methods: an example in
dynamic geometry. ZDM 43(3), 399–411 (2011). https://doi.org/10.1007/s11858-011-0313-x
12. Frassia, M.G., Serpe, A.: Learning geometry through mathematical modelling: an example
with GeoGebra. Turk. Online J. Educ. Technol. 2017(Nov Spec. Issue INTE), 411–418
(2017)
13. Galbraith, P., Goos, M., Renshaw, P., Geiger, V.: Technology enriched classrooms: some
implications for teaching applications and modelling. In: Mathematical modelling in
education and culture, pp. 111–125. Woodhead Publishing, Cambridge (2003). https://doi.
org/10.1533/9780857099556.3.111
14. Rodríguez Gallegos, R., Quiroz Rivera, S.: Developing modelling competencies through the
use of technology. In: Stillman, G.A., Blum, W., Biembengut, M.S. (eds.) Mathematical
Modelling in Education Research and Practice. IPTLMM, pp. 443–452. Springer, Cham
(2015). https://doi.org/10.1007/978-3-319-18272-8_37
15. Greefrath, G., Weigand, H.-G.: Simulieren: Mit Modellen experimentieren. Mathematic
lehren 174, 2–6 (2012)
16. Laborde, C.: The use of new technologies as a vehicle for restructuring teachers’
mathematics. In: Lin, F.L., Cooney, T.J. (eds.) Making sense of mathematics teacher
education, pp. 87–109. Springer, Dordrecht (2001). https://doi.org/10.1007/978-94-010-
0828-0_5
17. Heid, M.K., Blume, G.W.: Research on technology and the teaching and learning of
mathematics: Vol. 1. Research syntheses. Information Age, Charlotte (2008)
18. Hohenwarter, J., Hohenwarter, M., Lavicza, Z.: Introducing dynamic mathematics software
to secondary school teachers: the case of GeoGebra. J. Comput. Math. Sci. Teach. 28(2),
135–146 (2009). Retrieved 18 September 2019, from https://www.learntechlib.org/primary/
p/30304/
19. Hollebrands, K.F.: The role of a dynamic software program for geometry in the strategies
high school mathematics students employ. J. Res. Math. Educ. 38(2), 164–192 (2007).
https://doi.org/10.2307/30034955
20. Kotenkamp, U., Blessing, A.M., Dohrmann, C., Kreis, Y., Libbrecht, P., Mercat, C.:
Interoperable interactive geometry for Europe-First technological and educational results and
future challeges of the Intergeo project. In: Durand-Guerrier, V., Soury-Lavergne, S.,
Arzarello, F. (eds.) Proceeding of the Sixty European Conference on Research on
Mathematics Education, pp. 1150–1160 (2010). http://www.inrp.fr/publications/edition-
electronique/cerme6/wg7-11-kortenkamp.pdf. Accessed 18 September 2019
21. Leikin, R., Grossman, D.: Teachers modify geometry problems: from proof to investigation.
Educ. Stud. Math. 82, 515–531 (2013). https://doi.org/10.1007/s10649-012-9460-4
22. Leung, A.: Dynamic geometry and the theory of variation. In: Pateman, N.A., Dougherty, B.
J., Zillox, J. (eds.) Proceedings of the 27th Conference of the International Group for the
Psychology of Mathematics Education, Honolulu, USA, vol. 3, pp. 197–204 (2003). https://
eric.ed.gov/?id=ED501009. Accessed 18 September 2019
23. Lopez-Real, F., Leung, A.: Dragging as a conceptual tool in dynamic geometry environments.
Int. J. Math. Educ. Sci. Technol. 37(6), 665–679 (2006). https://doi.org/10.1080/0020739060
0712539
 Task Mathematical Modelling Design in a Dynamic Geometry Environment 491

24. Mandelbrot B.B.: Fractal Geometry of Nature. Freeman, San Francisco (1983). https://doi.
org/10.1002/esp.3290080415
25. Pickover, C.A.: Mathematics and beauty: a sampling of spirals and ‘Strange’ spirals in
science. Nat. Art. Leonardo 21(2), 173–181 (1988). https://doi.org/10.5539/jmr.v3n3p3
26. Sergeyev, Y.D.: Blinking fractals and their quantitative analysis using infinite and
infinitesimal numbers. Chaos, Solitons Fractals 33(1), 50–75 (2007). https://doi.org/10.
1016/j.chaos.2006.11.001
27. Sergeyev, Y.D.: Un semplice modo per trattare le grandezze infinite ed infinitesime.
Matematica nella Società e nella Cultura: Rivista della Unione Matematica Italiana 8(1),
111–147 (2015)
28. Sergeyev, Y.D.: The exact (up to infinitesimals) infinite perimeter of the Koch snowflake and
its finite area. Commun. Nonlinear Sci. Numer. Simul. 31(1-3), 21–29 (2016). https://doi.
org/10.1016/j.cnsns.2015.07.004
29. Serpe, A., Frassia, M.G.: legacy and influence in Mathematics and Physics with educational
technology: a laboratory example. In: Magazù, S. (ed.) New Trends in Physics Education
Research, pp. 77–96. Nova Science Publisher, New York (2018)
30. Serpe, A.: Geometry of design in high school - an example of teaching with Geogebra. In:
GomezChova, L., LopezMartinez, A., CandelTorres, I. (eds.), INTED2018, Proceedings
12th International Technology, Education and Development Conference, Valencia, Spain,
pp. 3477–3485 (2018). https://doi.org/10.21125/inted.2018.0668
31. Sherman, M.: The role of technology in supporting students’ mathematical thinking:
Extending the metaphors of amplifier and reorganizer. Contemporary Issues in Technology
and Teacher Education, 14 (3), 220–246 (2014). https://www.learntechlib.org/primary/p/
130321/. Accessed 18 September 2019
32. Siller, H.S., Greefrath, G.: Mathematical modelling in class regarding to technology. In:
Durand-Guerrier, V., Soury-Lavergne, S., Arzarello, F. (eds.) Proceeding of the Sixty
European Conference on Research on Mathematics Education, pp. 1150–1160 (2010). http://
www.inrp.fr/editions/cerme6. Accessed 18 September 2019
33. Weigand, H.G., Weth, T.: Computer im Mathematikunterricht: Neue Wege zu alten Zielen.
Spektrum, Heidelberg (2002)
34. Yerushalmy, M.: Student perceptions of aspects of algebraic function using multiple
representation software. J. Comput. Assist. Learn. 7, 42–57 (1991). https://doi.org/10.1111/j.
1365-2729.1991.tb00223.x
Optimization and Management of Water
Supply
 Numerical Experimentations for a New
Set of Local Indices of a Water Network

Marco Amos Bonora1 , Fabio Caldarola2 , Joao Muranho3 , Joaquim Sousa4 ,

and Mario Maiolo1(B)
1
Department of Environmental and Chemical Engineering (DIATIC),
Università della Calabria, Cubo 42/B, 87036 Arcavacata di Rende, CS, Italy
{marcoamos.bonora,mario.maiolo}@unical.it
2
Department of Mathematics and Computer Science, Università della Calabria,
Cubo 31/B, 87036 Arcavacata di Rende, CS, Italy
[email protected]
3
Department of Computer Science, Universidade da Beira Interior,
Rua Marquês d’Ávila e Bolama, 6201-001 Covilhã, Portugal
[email protected]
4
Department of Civil Engineering, Instituto Politécnico de Coimbra,
Rua Pedro Nunes, 3030-199 Coimbra, Portugal
[email protected]

Abstract. Very recently, a new set of local performance indices has

been proposed for an urban water supply system together with a useful
mathematical model or, better, framework that organizes and provides
the tools to treat the complex of these local parameters varying from
node to node. In this paper, such indices are considered and examined
in relation to hydraulic software using Demand Driven Analysis (DDA)
or Pressure Driven Analysis (PDA). We investigate the needed hypothe-
ses to obtain effective numerical simulations employing, in particular,
EPANET or WaterNetGen, and the concrete applicability to a real water
supply system known in literature as the KL network.

Keywords: Urban water supply systems · Performance indices ·

Mathematical modeling · EPANET 2.0.12 · WaterNetGen 1.0.0.942

1 Introduction

In the resolution of the hydraulic problem associated with drinking water dis-
tribution networks (WDNs), a hydraulic software model solves the continuity
equations in the junction nodes and the energy equations in the links. There
are two main resolutive approaches in literature: if the flow rate supplied in the
demand nodes is considered constant and defined upstream of the simulation,
the solver software will look for a solution that will guarantee that the flow and
load regime will meet the required supply in the nodes. In this case we speak of
Demand Driven Analysis (DDA). If instead the model foresees that the supply

c Springer Nature Switzerland AG 2020

Y. D. Sergeyev and D. E. Kvasov (Eds.): NUMTA 2019, LNCS 11973, pp. 495–505, 2020.
https://doi.org/10.1007/978-3-030-39081-5_42
496 M. A. Bonora et al.

may differ from the request in the nodes, depending on the pressure regime, it
is a Pressure Driven Analysis (PDA).
The use of PDA models is more expensive in computational terms but
presents results more representative of reality in the case of WDNs characterized
by a poor regime of pressures; in the event that the pressure regime is sufficient
to guarantee the supply in all the demand nodes there are no real advantages to
using a PDA approach, which will provide very similar results if not identical to
DDA models.
A Demand Driven approach is typical of software such as EPANET (see [23]).
This software allows to model the hydraulic behavior of the water distribution
network (WDN) and also to perform water quality simulations. EPANET is one
of the most widespread software in WDNs simulation. On the other hand, a
well-known software for PDA analysis is WaterNetGen, an EPANET extension
developed by Muranho et al. (see [21,22]).
The aim of this paper is to study in real contexts a new set of local perfor-
mance indices recently developed by Caldarola and Maiolo in [1] (cf. also [2]) and,
in particular, analysing and applying them to the WDN described by Kang and
Lansey in [11]. In the following sections will be examined the hypotheses needed
for a computational assessment and a practical application of these indices to
the considered WDN, the results obtained by using DDA and PDA approaches,
and the relation between them and some well-known indices as the resilience
measures proposed and discussed in [5–8,24,25].
For similar indices concerning the vulnerability of infrastructures, the sus-
tainability of water resources and various types of hydropotable risk, the reader
can see [3,4,12–19].

2 Performance Indices

For a given water network we denote by n the number of its junction nodes, by r
the number of tanks (or reservoirs), by qi , hi and hi = hi + zi the discharge, the
pressure head and the piezometric head at the i-th node, respectively, where zi
stands for the elevation head. Hence pi = γ qi hi represents the delivered power at
the node i, where γ is the specific weight of water. We also use the notations qi∗ ,
hi∗ , h∗i and p∗i to indicate the (minimal) project requests relative to the above
defined quantities, as is usual in much current literature (see, e.g., [5–10,25] and
the references therein).
Inside a structured mathematical framework described in [1], the following
local indices are proposed as “elementary building bricks” to construct (new)
local and global indices for the needs of a WDN, and also useful to recover the
well-known global indices ordinarily used in WDN analysis and implemented in
many hydraulic simulation software:
 Numerical Experimentations for a New Set of Local Indices 497

qi − qi∗
qis := Local discharge surplus index,
qi∗
∗
h i − hi
his := ∗ Local pressure head surplus index,
hi
hi − h∗i (1)
s
hi := Local piezometric head surplus index,
h∗i
p i − p∗i qi hi − qi∗ h∗i
psi := = Local power surplus index,
p∗i qi∗ h∗i
where i = 1, 2, . . . , n (see [1,2] for more details).
Example 1. An example of how it is possible to recover many well-known global
indices of a WDN using (1) and the mathematical framework exposed in [1],
is provided by the new formulation, given in (4) and (5), of the following two
resilience indices
n  
qi∗ hi − h∗i
i=1
Ir =  r n (2)
Qk Hk − qi∗ h∗i
k=1 i=1

and
n 
 
qi hi − qi∗ h∗i
i=1
IR = 
r 
n , (3)
Qk Hk − qi∗ h∗i
k=1 i=1

where Qk and Hk are the discharge and the head, respectively, from the tank k.
The former has been introduced by Todini in [25], while the second is a modified
version of Ir used by Di Nardo et al. in [5–10].
The resilience indices Ir and IR are written in [1,2] as follows
hs • p∗
Ir = (4)
γ (Q • H − q ∗ • h∗ )
and
ps • p∗
IR = , (5)
γ (Q • H − q ∗ • h∗ )
where
– Q := (Q1 , Q2 , . . . , Qr ),
– H := (H1 , H2 , . . . , Hr ),
– q ∗ := (q1∗ , q2∗ , . . . , qn∗ ),
– h∗ := (h∗1 , h∗2 , . . . , h∗n ),
– p∗ := (p∗1 , p∗2 , . . . , p∗n ),
– hs := (hs1 , hs2 , . . . , hsn ),
– ps := (ps1 , ps2 , . . . , psn ),
and “•” denotes the standard scalar product between real vectors (of dimension
n or r in our case).
For more details and examples the reader can see [1] and [2].
498 M. A. Bonora et al.

3 Hydraulic Solvers
In this work, two different approaches are studied using known software for
the hydraulic modeling of WDNs. To obtain a solution, EPANET solves the
continuity Eq. (6)(a) in each junction node and the energy law one, expressed in
(6)(b), for each pipe connecting two nodes. The energy law links the headloss
to the flow, depending on the pipe characteristics. Such equations are generally
expressed as in the following form
⎧
⎪
⎪
n(i)
⎨ Qij − qi = 0 for all i = 1, 2, . . . , n , (a)
(6)
⎪
⎪ j=1
⎩h − h = h = R · Q e + m · Q 2 , (b)
i j ij ij l ij

where qi is the flow demand and hi the nodal head at note i, n the total number
of junction nodes, n(i) the number of those linked to the node i, hij and Qij
the headloss and the flow rate in the pipe between the linked nodes i and j
respectively, R the resistance coefficient, e the flow exponent in the headloss
formula (resistance law) and ml a minor loss coefficient (see for example [23]).
EPANET uses a DDA approach, so the water demand at junction nodes is
a known term for solving the continuity Eq. (6)(a). If the solver fails to find a
combination of heads in the junction nodes and flows in the links that satisfy
the total demand, it will stop the simulation without obtaining a full solution.
In order to simulate the network with a PDA approach, WaterNetGen is
used. In addition to continuity equations on nodes (6)(a) and energy for links
(6)(b), WaterNetGen makes changes to the EPANET solver in order to simulate
a difference between supply and demand in case of insufficient pressure. This
software adds a third equation (see [21,22]), which expresses the water supply
according to:
⎧
⎪
⎪1 if hi ≥ hiref ,
⎪
⎨
α
hi − hi min
qiavl (hi ) = qireq · if himin < hi < hiref , (7)
⎪
⎪ h ref − hmin
⎪
⎩
i i
0 if hi ≤ himin ,

where the measures explained below are referred to the node i of the network:
hi = node pressure
qiavl = available water supply
qireq = water demand
hiref = service pressure (necessary to completely satisfy the demand)
himin = minimum pressure below which there is no supply
α = exponent of the pressure-demand relationship.
The local surplus and the resilience indices recalled in Sect. 2 use some design
conditions (qi∗ , hi∗ and h∗i ) as reference requests (see, e.g., [5–8,24,25]) that serve
as comparison terms for the actual functioning parameters of the network (qi ,
hi and hi , respectively).
 Numerical Experimentations for a New Set of Local Indices 499

A problem related to the use of these indices is the difficulty in defining

and identifying those conditions. Scientific literature shows, in fact, that it is
rare to have precise indications on design pressures and water demand. Some
authors use a pressure value equal for the whole network [5–10,25]) but it was
not possible to identify publications that gave explicit indications on the design
values of the water demand.

3.1 Demand Driven Analysis

In absence of indications on the design water demand, it seems that some authors
have placed the design water demand equal to that assigned to the software as
known terms. This means that, the design water demand is equal on all nodes
to the actual water supply, that is

qi = qi∗ for all i = 1, 2, . . . , n. (8)

Then, it is immediate that the local discharge surplus index defined in (1) will
be null everywhere,
qi
qis = −1=1−1=0 for all i = 1, 2, . . . , n,
qi∗
and the local power surplus index will be equal to the local head surplus one:
pi qi hi hi
psi = ∗ − 1 = ∗ ∗ − 1 = ∗ − 1 = hsi . (9)
pi qi hi hi

In particular, Eq. (8) implies that there cannot be surplus or deficit on water
supply since the values will always be identical. No predictions or assumptions
can be made on the sign of the local head and pressure head indices in (1) since
it will solely depend on the regime of the water heads in the WDN.
The design water head (or pressure) assigned to the network will not affect the
EPANET results, because these values will only affect the performance indices.
Moreover, the resilience index Ir proposed by Todini and the resilience index IR
by Di Nardo et al., will in this case coincide:
n 
 
qi hi − qi∗ h∗i
i=1
IR = 
r 
n
Qk Hk − qi∗ h∗i
k=1 i=1

n  
qi∗ hi − h∗i
i=1
= 
r 
n = Ir .
Qk Hk − qi∗ h∗i
k=1 i=1

It is also immediate to notice equality between the two formulations in (4) and
(5), if we recall that the local head surplus index and the local power surplus
index coincide (see (8)).
500 M. A. Bonora et al.

3.2 Pressure Driven Analysis

In PDA approach there is an effective difference between water demand and
supply in the nodes. Taking into account the relation (7), the water demand and
service pressure coincide with the design conditions used in the performance
indices. The relation (7) becomes
⎧
⎪
⎪1
if hi ≥ hi∗ ,
⎪
⎨ 
min α
hi − hi
qi = qi∗ · if himin < hi < hi∗ , (10)
⎪ hi∗ − himin
⎪
⎪
⎩0 if hi ≤ himin .

With this relation no longer worth (8), in fact, the supply may differ from the
design demand and the local discharge surplus index in (1) can get non-zero
values.
In particular, in WaterNetGen analysis, two situations can occur:
– the network is characterized by a good hydraulic head regime, so the behavior
of PDAs is similar to that of DDAs;
– there are head deficits and the water supply is lower compared to the demand.
Note moreover that in the first scenario above, Eq. (8) remains still valid.
For a head-deficient WDN we have instead

qi = qi∗ if hi ≥ hi∗
,
qi < qi∗ if hi < hi∗

whence
qis ≤ 0 for all i = 1, 2, . . . , n.
This means that in the second scenario, the local discharge surplus index can
only get non-positive values.

4 Application to the KL Network

The local surplus indices were tested in a practical application on a real network,
well-known in literature. The used WDN is the network proposed by Kang and
Lansey in [11] (KL network, for short) which consists of 935 nodes and 1274
pipes. In average condition the network has a consumption of 177 l/s (2808
gal/min) and the authors propose in [11] a peak factor of 1.75 to take into
account hourly peaks. The authors also provide an indication of the minimum
pressure in the network (design pressure) which is equal to 28 m (40 psi).
In this work a higher peak coefficient is used in order to establish a
low-pressure regime and analyze the differences between the DDA and PDA
approaches in the surplus index and in the resilience indices. The peak factor
used is 2.5 (see [20, pag. 22]). The minimum pressure is 15 m and the design one
is 28 m.
 Numerical Experimentations for a New Set of Local Indices 501

The local surplus indices are plotted for each node of the network for both
analysis types and the results are shown in Figs. 1 and 2, respectively. A colori-
metric scale that takes into account 3 intervals was used:
− Red: Deficit conditions;
− Yellow: Conditions close to the project requests;
− Green: Condition of surplus.
For the local discharge surplus index and the local pressure head surplus
index the bands are:
[−1; −0.1] Deficit conditions, for values that are less than 90% of the
design values;
[−0.1; 0.1] Design conditions, for values that do not differ more than
10% from the design values;
[0.1; 1] Surplus conditions, for values that are greater than 110%
of the design values.
For the local power surplus index and the local head surplus index, the bands
are different because the presence of the geodetic elevation reduces the variability
of the index, therefore they are:
[−1; −0.01] Deficit conditions, for values that are less than 99% of
the design values;
[−0.01; 0.01] Design conditions, for values that do not differ more
than 1% of the design values;
[0.01; 1] Surplus conditions, for values that are greater than
101% of the design values.
In correspondence to Fig. 1, hence with the DDA approach, both resilience
indices IR and Ir coincide to the following value

IR = Ir = 0.0532.

Instead, by using the PDA approach and hence referring to Fig. 2, we obtained

IR = −1.46 and Ir = 0.448.

For the results interpretation of Figs. 1 and 2, it is clear that the multiplicative
peak coefficient of the flows establishes in the network a condition characterized
by a poor pressures regime. The increase of the elevation parameter, which grows
further away from the reservoir (SE direction), contributes to the load losses
along the pipelines, and this creates a deficit condition especially in the areas
farthest from the reservoir. Recall that Fig. 1 shows the results obtained from the
application of a solver that uses the DDA model and, as explained in Subsect. 3.1,
the information provided by the local discharge surplus index qis and the local
power surplus index psi (i.e., the first and the last index of the four defined
in (1)) are not significant. The former, in fact, is everywhere equal to zero by
the hypotheses made in (8) about the coincidence between nodal discharges qi
and minimal design discharge requests qi∗ . The reader can note that this agrees
502 M. A. Bonora et al.

Fig. 1. The graphical display of local indices for the Kang and Lansey WDN, where the
results are obtained from a DDA hydraulic simulation. Subfigures (a) and (b) show the
local head and the local pressure head surplus index, respectively. Subfigures (c) and
(d), instead, show the local power and the local discharge surplus index, respectively.

with the bottom right picture in Fig. 1 where all the junction nodes are colored
in yellow. Similarly, the local power surplus index psi coincides with the local
head surplus index hsi (recall (9)) and this agrees, as well, with the same nodes
coloration between the two left pictures Fig. 1.
The surplus indices relating to pressure and load give, instead, an immedi-
ate graphical information of the network status. Due to the distance from the
reservoir, the nodes in the southern area of the network will be characterized by
deficit conditions compared to the design ones.
 Numerical Experimentations for a New Set of Local Indices 503

Fig. 2. The graphical display of local indices for the Kang and Lansey WDN, where
the results are obtained from a PDA hydraulic simulation. The Subfigures (a), (b), (c)
and (d) are in full correspondence with those of Fig. 1, for the same local indices.

On the other hand, the information provided by local indices, calculated for
the result of a PDA model (see Fig. 2), with the same design conditions and peak
coefficients, give different results. In this case, the local discharge surplus index
qis provides information on the nodes that, due to the pressure deficit, cannot
guarantee that the supply meets the request. As highlighted in the Subsect. 3.2,
there cannot be nodes in surplus. The shown condition of pressure and load
deficit is less critical than the one obtained with the DDA model. It is clear
that, being minor the supply than the previous case, there are lower flow rates,
lower speeds and consequently the network contains larger load losses. Finally,
504 M. A. Bonora et al.

the local power surplus index psi provides a new set of information which, in this
case, will not coincide with the ones provided by the local head surplus index.

5 Conclusions

The new set of local indices proposed by Caldarola and Maiolo in [1] are assessed
for a WDN in addition to two well-known resilience ones. The mathemati-
cal framework described there allowed to simplify the automatic calculation
of resilience indices. It was also possible to visualize the local surplus indices
graphically and this approach allowed to have an immediate feedback on the
state of the network. The assessment of these indices on a WDN leads moreover
to a series of observations about their application limits. Both softwares that use
DDA and PDA models were employed. The models hypotheses limit the pos-
sible results, sometimes preventing the achievement of indices representative of
the WDN situation. The type of solution model used influences the value of the
indices and the resilience assessment. The lack of a precise definition of design
conditions is the main factor of uncertainty in the results.

Aknowledgments. This work is partially supported by the research projects

“origAMI, Original Advanced Metering Infrastructure”, CUP J48C17000170006, POR
Calabria FESR-FSE 2014–2020, and “I-BEST”, CUP B28I17000290008, PON “Inno-
vazione e competitività” 2014/2020 MiSE Horizon 2020.

References
1. Caldarola, F., Maiolo, M.: Local indices within a mathematical framework for
urban water distribution systems. Cogent Eng. 6, 1643057 (2019). https://doi.
org/10.1080/23311916.2019.1643057
2. Caldarola, F., Maiolo, M.: Algebraic tools and new local indices for water networks:
some numerical examples. In: Sergeyev, Y., Kvasov, D. (eds.) NUMTA 2019. LNCS,
vol. 11973, pp. 517–524. Springer, New York (2020)
3. Carini, M., Maiolo, M., Pantusa, D., Chiaravalloti, F., Capano, G.: Modelling and
optimization of least-cost water distribution networks with multiple supply sources
and users. Ric. Matem. 67(2), 465–479 (2018). https://doi.org/10.1007/s11587-
017-0343-y
4. Cervarolo, G., Mendicino, G., Senatore, A.: Re-modulating water allocation in a
complex multi-reservoir system under current and climate change scenarios. Eur.
Water 37, 47–57 (2012)
5. Di Nardo, A., Di Natale, M.: A heuristic design support methodology based on
graph theory for district metering of water supply networks. Eng. Optim. 12, 193–
211 (2011)
6. Di Nardo, A., Di Natale, M.: A design support metodology metodology for dis-
trict metering of water supply networks. Water Distrib. Syst. Anal. 2010, 870–887
(2012)
7. Di Nardo, A., et al.: Redundancy features of water distribution systems. Procedia
Eng. 186, 412–419 (2017)
 Numerical Experimentations for a New Set of Local Indices 505

8. Di Nardo, A., Di Natale, M., Santonastaso, G.: A comparison between different

techniques for water network sectorization. Water Sci. Technol. Water Supply 14,
961–970 (2014)
9. Di Nardo, A., Di Natale, M., Santonastaso, G., Tzatchkov, V., Alcocer-Yamanaka,
V.: Water network sectorization based on graph theory and energy performance
indices. J. Water Res. Planning Manag. 140, 620–629 (2014)
10. Di Nardo, A., Di Natale, M., Santonastaso, G., Tzatchkov, V., Alcocer-Yamanaka,
V.: Performance indices for water network partitioning and sectorization. Water
Sci. Technol. Water Supply 15, 499–509 (2015)
11. Kang, D., Lansey, K.: Revisiting optimal water-distribution system design: issues
and a heuristic hierarchical approach. J. Water Res. Planning Manag. 138, 208–217
(2012)
12. Maiolo, M., Carini, M., Capano, G., Piro, P.: Synthetic sustainability index (SSI)
based on life cycle assessment approach of low impact development in the Mediter-
ranean area. Cogent Eng. 4, 1410272 (2017). https://doi.org/10.1080/23311916.
2017.1410272
13. Maiolo, M., Martirano, G., Morrone, P., Pantusa, D.: Assessment criteria for a sus-
tainable management of water resources. Water Pract. Technol. 1 (2006). https://
doi.org/10.2166/wpt.2006.012
14. Maiolo, M., Mendicino, G., Pantusa, D., Senatore, A.: Optimization of drinking
water distribution systems in relation to the effects of climate change. Water 9,
803 (2017). https://doi.org/10.3390/w9100803
15. Maiolo, M., Pantusa, D.: A methodological proposal for the evaluation of potable
water use risk. Water Pract. Technol. 10, 152–163 (2015). https://doi.org/10.2166/
wpt.2015.017
16. Maiolo, M., Pantusa, D.: An optimization procedure for the sustainable manage-
ment of water resources. Water Sci. Technol. Water Supply 16, 61–69 (2016).
https://doi.org/10.2166/ws.2015.114
17. Maiolo, M., Pantusa, D.: Combined reuse of wastewater and desalination for the
management of water systems in conditions of scarcity. Water Ecol. 72, 116–126
(2017)
18. Maiolo, M., Pantusa, D.: Infrastructure vulnerability index of drinking water sup-
ply systems to possible terrorist attacks. Cogent Eng. 5, 1456710 (2018). https://
doi.org/10.1080/23311916.2018.1456710
19. Maiolo, M., Pantusa, D., Chiaravalloti, F., Carini, M., Capano, G., Procopio, A.: A
new vulnerability measure for water distribution network. Water 10, 1005 (2019).
https://doi.org/10.3390/w10081005
20. Milano, V.: Acquedotti. Hoepli Editore, Milano (1996)
21. Muranho, J., Ferreira, A., Sousa, J., Gomes, A., Sa Marques, A.: Waternetgen -
an epanet extension for automatic water distribution network models generation
and pipe sizing. Water Sci. Technol. Water Supply 12, 117–123 (2012)
22. Muranho, J., Ferreira, A., Sousa, J., Gomes, A., Sa Marques, A.: Pressure depen-
dent demand and leakage modelling with an epanet extension - waternetgen. Pro-
cedia Eng. 89, 632–639 (2014). 16th Conference on Water Distribution System
Analysis, WDSA
23. Rossman, L.A.: Epanet 2: users manual (2000)
24. Shin, S., et al.: A systematic review of quantitative resilience measures for water
infrastructure systems. Water 10, 164 (2018)
25. Todini, E.: Looped water distribution networks design using a resilience index
based on heuristic approach. Urban Water 12, 115–122 (2000)
 Performance Management of Demand
and Pressure Driven Analysis in a
Monitored Water Distribution Network

Marco Amos Bonora1 , Manuela Carini1 , Gilda Capano1 , Rocco Cotrona1 ,

Daniela Pantusa1 , Joaquim Sousa2 , and Mario Maiolo1(B)
1
Department of Environmental and Chemical Engineering (DIATIC),
Università della Calabria, Cubo 42/B, 87036 Arcavacata di Rende, CS, Italy
{marcoamos.bonora,manuela.carini,gilda.capano,daniela.pantusa,
mario.maiolo}@unical.it, [email protected]
2
Department of Civil Engineering, Polytechnic Institute of Coimbra,
Coimbra, Portugal
[email protected]

Abstract. A smart management of water distribution networks requires

the infrastructure to be operated with high efficiency. For many years the
hydraulic modelling of water distribution networks has been conditioned
by the scarce availability of quality data but the technological advances
contributed to overcome this drawback. The present work describes the
research activity carried out about water distribution network modelling,
focusing on model construction and calibration. For this purpose, Water-
NetGen, an extension of the Epanet hydraulic simulation software, has
been used. EPANET simulation model assumes that the required water
demand is always fully satisfied regardless the pressure (Demand Driven
Analysis - DDA), while WaterNetGen has a new solver assuming that
the required water demand is only fully satisfied if the pressure con-
ditions are adequate (Pressure Driven Analysis - PDA). A comparison
between the software outputs is the starting point for a new method
of allocating and distributing water demand and water losses along the
network, leading to model results closer to the measurements obtained
in the real network. The case study is the water distribution network of
the municipality of Nicotera, in Southern Italy.

Keywords: Water distribution network (WDN) · Water

management · Pressure Driven Analysis (PDA) · Demand Driven
Analysis (DDA) · Calibration

1 Introduction
The water demand increase and the water scarcity require the use of manage-
ment practices sensitive to the importance of water in human life. Consequently,
all the aspects related to the proper allocation of available resources, to the eval-
uation of the climate change effects on water resources and schemes, to the use of
c Springer Nature Switzerland AG 2020
Y. D. Sergeyev and D. E. Kvasov (Eds.): NUMTA 2019, LNCS 11973, pp. 506–516, 2020.
https://doi.org/10.1007/978-3-030-39081-5_43
 Performance Management of Demand and PDA in a Monitored WDN 507

unconventional water resources, to the correct system design and operational effi-
ciency, to the assess of vulnerability to natural and man-made disasters, are sci-
entific topics of current interest [2,6–8,10,18]. In this context, research activities
concerning the design and management of water distribution networks (WDN)
are of the utmost importance. The efficient management of water resources is
related to the hydraulic balances of the parameters which describe the network
behaviour. In the scientific literature there are many hydraulic-mathematical
models which allow evaluating the water distribution network efficiency. Among
these there are models based on the performance indices that play an impor-
tant role, e.g. resilience, entropy and vulnerability [9,13,21]. These models allow
monitoring of the WDN correct functioning also for risk management. However,
the number and variability of the parameters which determine the risk in WDN
require the use of specific analytical methods, capable of quantifying it. In the
literature there are other models which allow increasing the knowledge of the
network, facilitating the planning, design and management phases. Simulation,
skeletonization, calibration and optimization models are very important in this
context, for the ability to show the basic behavioural structure of a WDN, allow-
ing the implementation of improvement measures (optimization of characteris-
tic parameters) and the WDN analysis in various scenarios. The improvement
of WDN management is frequently achieved by using simulation models, built
with the help of new software tools. In most cases, the model validity depends
on the ability to guarantee a distribution suited to the requests; therefore the
water demand is an essential parameter. Recently, due to the increasing need
of more realistic models, many of these incorporate special pressure-demand
relationships, that enable real prediction of the WDN behaviour [5]. In general,
simulation models may include a DDA or a PDA approach. In the first case
water demand is fixed while in the second it depends on the nodal pressures.
The traditional DDA approach can be used for planning, design and operation
of WDN working under normal conditions, while the PDA approach is used
in particular scenarios in which pressure conditions restrict water availability.
In the literature there are many models based on these approaches. The best-
known DDA application is presented in [17]. This is the open source software
Epanet, one of the most used tools for hydraulic simulation. The original node-
pipe equations, derived by [22], are solved by the Generalized Reduced Gradient
(GRG) method, which the same as the Newton-Raphson method. [23] presented
a unified framework for deriving simultaneous equations algorithms for WDN,
comprising all sets of equations solved by the Newton-Raphson method. The
Epanet software uses a DDA approach and so assumes that water demand can
always be met regardless the network pressures. However, if the network has
to operate in unusual operating conditions (excessive demands, pipe breaks or
fire scenarios, causing low pressures) it is shown that the DDA approach is not
able to correctly simulate operation [1]. In these situations, PDA simulation is
preferred, although often the analysis of pressure variability in a WDN can be
temporary and unpredictable [4]. Among the PDA approaches, [16] proposed a
methodology which takes into account the residual head versus outflow relation-
508 M. A. Bonora et al.

ship having no clamping of the outflow. [15] presented a pressure-driven method

relying on a co-energy minimization problem formulation that do not requires
to define any pressure-consumption relationship and does not rely on the detec-
tion of topological changes. This method permits better hydraulic predictions
for network sections supplied by nodes with negative pressures. [20] developed a
new head-discharge relationship which is based on volumetric demand (pressure
independent) and head dependent demand, separately. [14] proposed a method
based on the insertion of a sequence of devices at each water demand node (a
General Purpose Valve, a fictitious junction, a reach with a check valve and a
reservoir) to transform the DDA into a PDA. In this context it may be useful to
have tools available to facilitate the implementation of a PDA approach also in
Epanet software. There are many studies that deal with increasing its potential,
also introducing applications which enable the PDA analysis. [3] proposed an
Epanet extension to insert the pressure/demand functions into OOTEN toolkit,
which aimed to simulate WDN under failures conditions. This study showed that
when WDN work under abnormal pressure conditions the DDA approach fails
and the PDA approach produces realistic results. Among these tools, WaterNet-
Gen [11] plays an important role, it is an EPANET extension for the automatic
generation and the sizing of synthetic models for WDN, particularly useful for
the introduction of the PDA solver [19]. The extension, which can also be used
to assess the technical performance of WDN [12] and analyse critical scenarios
(abnormal pressure conditions), maintains the original user interface of EPANET
to preserve the user experience, with the introduction of new functionalities
particularly useful in contexts where it is necessary to modify elements from
imported models, such as those obtained from CAD drawings [11]. An impor-
tant characteristic is the ability to use this interface to split pipes (by automat-
ically introducing intermediate nodes), change nodes to tanks/reservoirs, and
easily insert pressure reduction valves to avoid or prevent the pressure exceeding
the maximum value allowed [11]. Furthermore, WaterNetGen considers pressure
independent demands (DDA) and three different types of pressure dependent
demands (PDA): node consumption, pipe consumption and pipe losses. This
software enables assigning the pipe consumption and losses through the use of
pipe consumption and losses coefficients and respective patterns, very useful for
large size WDN. In this paper, the Nicotera WDN (Southern Italy) is used to
compare the calibration results from Epanet with those from WaterNetGen. The
objective of this procedure is to define reliable data for a correct flow distribu-
tion, thus defining a new method of allocating and distributing demand and
water losses along the network.

2 Case Study: Nicotera WDN

 
The Nicotera municipality (38o 33 20” N, 15o 56 15” E) is located in the Calabria
region, in southern Italy (Fig. 1), with a territorial extension equal to 28.25 km2 .
The municipal water supply system, which must serve 6,353 inhabitants (the
per capita water consumption is about 475 l/inhab/day), extends for about
 Performance Management of Demand and PDA in a Monitored WDN 509

Fig. 1. Nicotera municipality position.

51.6 km and includes 6 tanks, 4 wells and 3 reservoirs (Figs. 2 and 3). The
Nicotera centre WDN, which is the subject of this work, is directly fed by 2
tanks (Piraino and Madonna della Scala). The average flow rate entering the
Nicotera centre WDN from both tanks are as follows:

– out of the Piraino tank 27.8 l/s,

– out of the Madonna della Scala tank 0.4 l/s.

During the data collection, the flow rates coming from both tanks (2), the
tank water levels (2) and some network pressures (3) and flows (3) were moni-
tored. These measurements were conducted simultaneously and continuously for
a period of two consecutive days, with an acquisition frequency equal to five
minutes, using the following technologies:
– Flow - ultrasonic flow meters with transit time;
– pressure - LoLog Vista data logger with internal pressure sensors;
– level - LoLog Vista data logger with Radcom depth sensors.
The pressure meters were installed far from the tanks so that the measure-
ments represent the general behaviour of the WDN. The water supply system
model has 7 water sources (4 wells and 3 reservoirs) and comprises 6 District
Metered Areas (DMAs), although this study is focused only on two of them
(Nicotera centre): Nicotera Est and Nicotera Ovest. The model of these two
510 M. A. Bonora et al.

DMAs has 2 tanks (Piraino and Madonna della Scala), 186 nodes and 202 pipes
with diameters in the range between 12 and 150 mm and 4 different materials
(steel, cast iron, galvanized iron, HDPE).

3 Numerical Results and Discussions

The first simulation step, carried out using the Epanet software, required the pipe
roughness coefficient, the allocation of water demand (with a spatial approach
for each junction in the network, in this case calculated with Thiessen polygons
obtained with QGIS), calculation of the water loss and its allocation for all
junctions in the network, and 24 hours’ demand patterns. This input data was
used to simulate the Nicotera centre WDN behaviour for a 24 h period, between
the 0:00 h of May 21st 2009 and 0.00 h of May 22nd 2009.
The following analysis showed great differences between the simulation
results and the measurements (gaps bigger than 30% in the tank levels, 40%
in pressures and 80% in flows). The chart in Fig. 4 presents some of the pressure
results for node P2 from the first simulation. The average simulated pressure
was equal to 26.07 m, while that observed was 30.15 m, and this was the best
fit obtained in this simulation). To reduce the gap between the simulated and
observed data, it was necessary to proceed with the calibration, which was carried
out with WaterNetGen, an Epanet extension that has some interesting features
to simplify the calibration work. The pressure and flow results obtained in the
first simulation shown that it was necessary to correct the flow distribution in
the network, namely increase it in the network served by the Piraino municipal
tank (Nicotera East) and decrease them by the same amount for the network
served by Madonna della Scala municipal tank (Nicotera West).
Firstly, demand coefficients were assigned to the pipes based on the building
density observed in Google Earth. Secondly, water loss coefficients were assigned
to the pipes in order to balance water loss in both networks and obtain good fit-
tings with the flow measurements. Finally, roughness coefficients were adjusted
to achieve good agreements with the pressure measurements. At the end of this
procedure the simulation results were quite similar to the measurements (Figs. 5,
6, 7 and 8). Taking again node P2 as an example, the average simulated pressure
is now equal to 30.41 m which compares with the observed value 30.15m, and
the mean error is 0.424 m (1.4%). The biggest mean errors were: tank level -
0.035 m (0.7% Fig. 9), pressure - 1.253 m (4.0%) and flow - 0.436 l/s (8.4%). A
considerable reduction of the gaps between the observed and simulated values
is evident, confirming a good calibration. The results obtained show the advan-
tage of using WaterNetGen in the calibration process, mainly due to a proper
allocation of water demand to the nodes (water consumption and water losses
assigned to the pipes). This issue, as previously mentioned, is essential to obtain
good calibration results. In the case of WaterNetGen, the assignment of the
nodal demands becomes quite simple by assigning to the pipes water demand
and water loss coefficients and respective patterns. The pipe water demand and
water loss are automatically assigned to the pipe end nodes using specific nodal
demand categories for their representation [11].
Performance Management of Demand and PDA in a Monitored WDN 511

Fig. 2. (a) Nicotera municipality water distribution network scheme.

512 M. A. Bonora et al.

Fig. 3. Zoom-in of Nicotera municipality water distribution network scheme north part

Fig. 4. Comparison of measured and simulated pressures for Node P2 obtained with
Epanet.
 Performance Management of Demand and PDA in a Monitored WDN 513

Fig. 5. Comparison of measured and simulated pressures for Node P1 obtained with
Epanet.

Fig. 6. Comparison of measured and simulated pressures for Node P2 obtained with
WaterNetGen.

Fig. 7. Comparison of measured and simulated flows for Link 49 obtained with Water-
NetGen.
514 M. A. Bonora et al.

Fig. 8. Comparison of measured and simulated flows for Link 21 obtained with Water-
NetGen.

Fig. 9. Comparison of measurements and simulation results from WaterNetGen:

Piraino Tank Level.

4 Conclusions
The DDA simulation, despite representing the traditional methodology, does
not provide realistic results for networks under abnormal operating conditions.
In this context, the PDA approach produces better results, especially in the
implementation through software tools such as WaterNetGen, a useful Epanet
extension. This software easily enables the assessment of the WDN performance,
and analysis of critical scenarios, by applying a simple method to allocate and
distribute water demand and losses along the network. The reliability of Water-
NetGen is justified by obtaining simulation values similar to the measurements
(which in some cases are perfect matches), resulting in a good calibration. This
methodology applied to the real case of the Nicotera WDN showed how the
EPANET simulator with the WaterNetGen extension is a useful tool to achieve
accurate simulations. The calibration results provide useful data to define some
new criteria for allocating and distributing water demand and losses along the
network.
 Performance Management of Demand and PDA in a Monitored WDN 515

Aknowledgments. This work is partially supported by the Italian Regional Project

(POR CALABRIA FESR 2014–2020): “origAMI, Original Advanced Metering Infras-
tructure” CUP J48C17000170006.

References
1. Ackley, J.R.L., Tanyimboh, T.T., Tahar, B., Templeman, A.B.: Head-driven anal-
ysis of water distribution systems. Water Softw. Syst. Theor. Appl. 1(3), 183–192
(2001)
2. Barthel, R., Janisch, S., Nickel, D., Trifkovic, A., Hörhan, T.: Using the multiactor-
approach in Glowa-Danube to simulate decisions for the water supply sector under
conditions of global climate change. Water Resour. Manag. 23(2), 239–275 (2010).
https://doi.org/10.1007/s11269-009-9445-y
3. Cheung, P.B., Van Zyl, J.E., Reis, L.F.R.: Extension of EPANET for pressure
driven demand modeling in water distribution system. Comput. Control Water
Indus. 1, 311–316 (2005)
4. Ciaponi, C., Creaco, E.: Comparison of pressure-driven formulations for WDN
simulation. Water 10(4), 523–537 (2018). https://doi.org/10.3390/w10040523
5. Giustolisi, O., Walski, T.M.: Demand components in water distribution network
analysis. J. Water Resour. Plan. Manag. 138(4), 356–367 (2011). https://doi.org/
10.1061/(ASCE)WR.1943-5452.0000187
6. Maiolo, M., Pantusa, D.: Combined reuse of wastewater and desalination for the
management of water systems in conditions of scarcity. Water Ecol. 4(72), 116–126
(2017). https://doi.org/10.23968/23053488.2017.22.4.116-126
7. Maiolo, M., Pantusa, D.: Infrastructure Vulnerability Index of drinking water sys-
tems to terrorist attacks. Cogent Eng. 5(1), 1456710 (2018). https://doi.org/10.
1080/23311916.2018.1456710
8. Maiolo, M., Mendicino, G., Pantusa, D., Senatore, A.: Optimization of drinking
water distribution systems in relation to the effects of climate change. Water 9(10),
803–817 (2017). https://doi.org/10.3390/w9100803
9. Maiolo, M., Pantusa, D., Carini, M., Capano, G., Chiaravalloti, F., Procopio, A.:
A new vulnerability measure for water distribution network. Water 10(8), 1005
(2018). https://doi.org/10.3390/w10081005
10. Minville, M., Brissette, F., Krau, S., Leconte, R.: Behaviour and performance of a
water resource system in Québec (Canada) under adapted operating policies in a
climate change context. Water Resour. Manag. 24, 1333–1352 (2010). https://doi.
org/10.1007/s11269-009-9500-8
11. Muranho, J., Ferreira, A., Sousa, J., Gomes, A., Marques, A.S.: WaterNetGen: an
EPANET extension for automatic water distribution network models generation
and pipe sizing. Water Sci. Technol. Water Supply 12(1), 117–123 (2012). https://
doi.org/10.2166/ws.2011.121
12. Muranho, J., Ferreira, A., Sousa, J., Gomes, A., Marques, A.S.: Pressure-dependent
demand and leakage modelling with an EPANET extension-WaterNetGen. Proce-
dia Eng. 89, 632–639 (2014). https://doi.org/10.1016/j.proeng.2014.11.488
13. Nachtnebel, H.P.: Irreversibility and Sustainability in Water Resources Sys-
tems. Risk, Reliability, Uncertainty, and Robustness of Water Resources Systems.
International Hydrology Series. Cambridge University Press, Cambridge (2002).
https://doi.org/10.1017/CBO9780511546006
516 M. A. Bonora et al.

14. Pacchin, E., Alvisi, S., Franchini, M.: Analysis of non-iterative methods and pro-
posal of a new one for pressure-driven snapshot simulations with EPANET. Water
Resour. Manag. 31(1), 75–91 (2017). https://doi.org/10.1007/s11269-016-1511-7
15. Piller, O., Van Zyl, J.E.: Pressure-driven analysis of network sections supplied via
high-lying nodes. In: Boxal, J., Maksimovic, C. (eds.) Proceedings of the Comput-
ing and Control in the Water Industry. Taylor and Francis Group, London (2009)
16. Reddy, L.S., Elango, K.: Analysis of water distribution networks with head-
dependent outlets. Civil Eng. Syst. 6(3), 102–110 (1989)
17. Rossman, L.A.: EPANET 2: users manual (2000)
18. Samani, H.M.V., Mottaghi, A.: Optimization of water distribution networks using
integer linear programming. J. Hydraul. Eng. 132(5), 501–509 (2006). https://doi.
org/10.1061/(ASCE)0733-9429(2006)132:5(501)
19. Sousa, J., Muranho, J., Sá Marques, A., Gomes, R.: Optimal management of
water distribution networks with simulated annealing: the C-Town problem. J.
Water Resour. Plan. Manag. 142(5), C4015010 (2015). https://doi.org/10.1061/
(ASCE)WR.1943-5452.0000604
20. Tabesh, M., Shirzad, A., Arefkhani, V., Mani, A.: A comparative study between
the modified and available demand driven based models for head driven analysis of
water distribution networks. Urban Water J. 11(3), 221–230 (2014). https://doi.
org/10.1080/1573062X.2013.783084
21. Todini, E.: Looped water distribution networks design using a resilience index
based heuristic approach. Urban Water 2(2), 115–122 (2000). https://doi.org/10.
1016/S1462-0758(00)00049-2
22. Todini, E., Pilati, S.: A gradient method for the solution of looped pipe networks.
In: Coulbeck, B., Orr, C.H. (eds.) Computer Applications in Water Supply, vol. 1,
pp. 1–20. Wiley, London (1988)
23. Todini, E., Rossman, L.A.: Unified framework for deriving simultaneous equa-
tions algorithms for water distribution networks. J. Hydraul. Eng. 139(5), 511–526
(2013). https://doi.org/10.1061/(ASCE)HY.1943-7900.0000703
 Algebraic Tools and New Local Indices
for Water Networks:
Some Numerical Examples

Fabio Caldarola1 and Mario Maiolo2(B)

1
Department of Mathematics and Computer Science, Università della Calabria,
Cubo 31/B, 87036 Arcavacata di Rende, CS, Italy
[email protected]
2
Department of Environmental and Chemical Engineering (DIATIC),
Università della Calabria, Cubo 42/B, 87036 Arcavacata di Rende, CS, Italy
[email protected]

Abstract. Very recently, a new set of local indices for urban water net-
works has been proposed by the authors, within a mathematical frame-
work which is unprecedented for this field, as far as we know. Such indices
can be viewed as the “elementary bricks” that can be used to construct
as many global (and local) indices as one needs or wants, where the glue,
or mortar, is given by the mathematical tools of the aforementioned
framework coming mostly from linear algebra and vector analysis. In
this paper, after a brief description of the setting as explained above, we
recover, through new formulations, some well-known global indicators
like the resilience index Ir introduced by Todini. Then we also give some
explicit numerical computations and examples, sometimes with the help
of the hydraulic software EPANET 2.0.12.

Keywords: Mathematical modeling · Linear algebra · Urban water

networks · Performance indices · EPANET 2.0.12

1 Introduction
In recent years many authors have introduced, sometimes with considerable suc-
cess, a multitude of indices, especially of energetic-hydraulic nature (for example,
indices of resilience, robustness, pressure, failure, flow deficit, mechanical redun-
dancy, balance, reliability, entropy, etc.) to characterize and summarize in a sin-
gle parameter some of the most important peculiar characteristics of a complex
water network (for instance, a recent review of 21 different resilience measures
was given last year in paper [20]). Therefore these indices, which are expressly
designed to be of global nature, do not adapt very well to local analysis even
applying them to a small portion of the network: a small portion of a network,
in fact, is not the same as a small independent network.
In [3] the authors propose a new set of local indicators within a mathemat-
ical framework which is also unprecedented, as far as we know, for hydraulic-
engineering purposes. Such indices, besides providing the basis for a local analysis
c Springer Nature Switzerland AG 2020
Y. D. Sergeyev and D. E. Kvasov (Eds.): NUMTA 2019, LNCS 11973, pp. 517–524, 2020.
https://doi.org/10.1007/978-3-030-39081-5_44
518 F. Caldarola and M. Maiolo

of the water distribution network (WDN), can be seen as the “elementary bricks”
with which, by means of the mathematical tools offered by the aforementioned
framework, which act as glue or mortar for the bricks, one can construct as many
global (and local) indices as one needs or wants, for the study of the considered
WDN. Moreover, he can also recover many already known global indicators,
often even giving a deeper structural interpretation of the same. This possibility
is explicitly illustrated in [3] in several cases, giving per se a strong automatic
validation to the new proposed machinery.
In this paper we first give a brief description of the local indices introduced
in [3], then we examine their building relations with some resilience indices like
that proposed by Todini in [21], or the one used by Di Nardo et al. in [5–10],
or others (see Sect. 2). Section 3 instead is devoted to numerical computations
and explicit examples. In particular, with the help of the hydraulic software
EPANET 2.0.12, we will examine some complementary cases of those considered
in [3] for the prototypical looped system known as the two-loop network (TLN)
and we will computing explicitly the family of local indices, deriving also from
them the global resilience measures mentioned above. Finally, in Sect. 4, we
will remove the ubiquitous assumption of a uniform minimal design pressure on
the network, and we will consider analogous pipes calibrations of those of the
previous section. Once the calculations are done again, we will briefly compare
the obtained results.

2 The Local Performance Indices for a WDN

As noticed in the Introduction, the contemporary trend of designing and devel-

oping indices of a global nature for water distribution networks (WDNs) is
strongly growing and involves a number of aspects of WDNs. But not only;
multiple lines of research have introduced a wide range of similar indices and
measures concerning as well the vulnerability of infrastructures, the sustainabil-
ity of water resources, various types of hydropotable risk, etc. (the reader can
see, for instance, [4,12–19] and the references therein). On the contrary, instead,
indices, measures and parameters of local nature are really very little present in
the literature of this field.
To introduce local indices for a WDN we need a certain amount of notations.
Let N be a given WDN and we denote by n the number of its junction nodes, by
m the number of pipes in N connecting two nodes, by r the number of reservoirs,
by Hk and Qk the head and the discharge, respectively, outgoing from the k-th
reservoir, k = 1, 2, . . . , r, by zi , qi and hi , the elevation height, the discharge
and the pressure head, respectively, at the node i, where i = 1, 2, . . . , n. Then we
also pose hi = zi + hi for the piezometric or hydraulic head and pi = γ qi hi for
the power delivered at the node i, where γ = ρg denotes, as usual, the specific
weight of water. We moreover use the widespread star notation, i.e., qi∗ , hi∗ ,
h∗i = zi + hi∗ and p∗i = γ qi∗ h∗i , to indicate the minimal requests or the design
conditions relative to the above defined quantities (see [3,5–10,21] and many
references mentioned there).
 Algebraic Tools and New Local Indices for Water Networks 519

In [3] the following local performance indices are defined and proposed as
elementary factors to perform a local-global analysis on the WDN:
qi −qi∗
qis := qi∗ Local discharge surplus index ,
hi −hi∗
his := hi∗ Local pressure head surplus index ,
hi −h∗ (1)
hsi := h∗
i
Local piezometric head surplus index ,
i

pi −p∗ qi hi −qi∗ h∗
psi := p∗
i
= qi∗ h∗
i
Local power surplus index ,
i i

where i = 1, 2, . . . , n. We moreover collect such local indices in vectors obtaining

the following local surplus vectors

q s := (q1s , q2s , . . . , qns ) Local discharge surplus vector ,

s
h := (h1s , h2s , . . . , hns ) Local pressure head surplus vector ,
s (2)
h := (hs1 , hs2 , . . . , hsn ) Local piezometric head surplus vector ,
ps := (ps1 , ps2 , . . . , psn ) Local power surplus vector .

Finally let Dj be the diameter of the j-th pipe, j = 1, 2, . . . , m, and we also

pose D := (D1 , D2 , . . . , Dm ), Q := (Q1 , Q2 , . . . , Qr ), H := (H1 , H2 , . . . , Hr ),
q ∗ := (q1∗ , q2∗ , . . . , qn∗ ), h∗ := (h∗1 , h∗2 , . . . , h∗n ) and p∗ := (p∗1 , p∗2 , . . . , p∗n ), to have
all the data in vectorial form.
Example 1. The well-known resilience index introduced by E. Todini in [21] is
defined as
n  
qi∗ hi − h∗i
i=1
Ir = Ir (N) :=  r n (3)
Qk Hk − qi∗ h∗i
k=1 i=1

and is used and implemented in various hydraulic software. Using local vectors
and some elementary mathematical tools, (3) can be written as
hs • p∗
Ir = (4)
γ (Q • H − q ∗ • h∗ )
where “•” denotes the standard scalar product between real vectors (of dimension
n or r in this case).
Example 2. The resilience index used by Di Nardo et al. in [5–10] is defined as
follows
n 
 
qi hi − qi∗ h∗i
i=1
IR = IR (N) :=  r 
n , (5)
Qk Hk − qi∗ h∗i
k=1 i=1

and using (2) and our mathematical framework we have

520 F. Caldarola and M. Maiolo

ps • p∗
IR = . (6)
γ (Q • H − q ∗ • h∗ )
Meaning, applications and, for instance, some advantages of the new formulas
(4) and (6), are discussed in [3] and also in [2]. Here we just notice as the linear
algebra language allows an easy and effective implementation in many engineer-
ing software and numerical computing systems as, for example, MATLAB1 .
Remark 1. Among the local indices defined in (1) it is easy to find relations as
the following
psi = qis hsi + qis + hsi ,
which yields
ps = q s ◦ hs + q s + hs , (7)
where “◦” denotes the Hadamard product defined entrywise between vectors and
matrices (see [3] or [11]).
For future use we recall here another simple algebraic tool (for several others
see [3]); for an n-tuple x = (x1 , x2 , . . . , xn ) belonging to Rn , the 1-norm, or
taxicab norm  · 1 is defined as

n
x1 = (x1 , x2 , . . . , xn )1 := |xi |. (8)
i=1

3 Local Indices Applied to the TLN

The very simple TLN, after its appearance in [1], has become very widespread
in literature for various types of examples, both theoretical and computational.
It is represented in Fig. 1 where each of the 8 pipes is 1000 m long, the altimetric
nodal data are given by

z = (150 m, 160 m, 155 m, 150 m, 165 m, 160 m),

and the reservoir R has Q = 1120 m3 /h and H = 210 m (coincident with its
geodetic height).
Example 3. If we take
 
q ∗ = q = 100 m3 /h, 100 m3 /h, 120 m3 /h, 270 m3 /h, 330 m3 /h, 200 m3 /h
= (1, 1, 1.2, 2.7, 3.3, 2) · 102 m3 /h, (9)
∗
h = (30 m, 30 m, 30 m, 30 m, 30 m, 30 m)

as in [21], we immediately obtain

h∗ = (180 m, 190 m, 185 m, 180 m, 195 m, 190 m),

(10)
p∗ = (18, 19, 22.2, 48.6, 64.35, 38) · 103 γ m4 /h,
1
The linear algebra foundation, in this case, is clear even from the name: MATrix
LABoratory.
 Algebraic Tools and New Local Indices for Water Networks 521

Fig. 1. A schematic representation of the TLN.

where γ = ρg = 9810 N/m3 is the specific weight of water. By using the Hazen-
Williams formula with a coefficient of 130 and the software EPANET 2.0.12 with
the following input diameters (in inches)

D = (D1 , . . . , D8 ) = (20, 14, 14, 6, 14, 1, 14, 10) (11)

(see [21, Table 3, Sol. B]), the pressure output is

h = (55.96 m, 40.39 m, 45.82 m, 46.63 m, 33.43 m, 31.79 m), (12)

hence, consequently,

h = (205.96 m, 200.39 m, 200.82 m, 196.63 m, 198.43 m, 191.79 m). (13)

We therefore obtain the following explicit values for the local surplus vectors (2)

q s = (since q = q ∗ ) = (0, 0, 0, 0, 0, 0),

hs ≈ (0.8653, 0.3463, 0.5273, 0.5543, 0.1143, 0.0597),
(14)
hs ≈ (0.1442, 0.0547, 0.0855, 0.0924, 0.0176, 0.0094),
ps = (using (7) and q s = 0 ) = hs .

Thus we can compute the resilience index Ir using (14) and (4), (9), (10) as
follows2
hs • p∗
Ir = γ(Q •H −q ∗ •h ∗ )
(0.1442,0.0547,0.0855,0.0924,0.0176,0.0094) • (18,19,22.2,48.6,64.35,38)·103 γ m4 /h
≈ γ [1120 (m3 /h) · 210 m − ((1,1,1.2,2.7,3.3,2)102 (m3 /h)) • (18,19,18.5,18,19.5,19)10 m]

= 11 513.4
25 050 ≈ 0.45962.

The index IR coincides with Ir as well: see (4), (6), and the last equation in (14).
2
Our value for Ir is very close to 0.47, the one computed in [21, Tab. 3].
522 F. Caldarola and M. Maiolo

Example 4. A different choice of the diameters in (11) affects all the local vectors
except q s which continues to be null. If, for example, we alter just the diameter
of the eighth pipe adding 2 in., i.e. replacing D8 = 10 with D8 = 12 in (11) (see
[21, Table 3, Sol. D]), then, by using the software EPANET 2.0.12, we obtain

h = (55.96 m, 40.38 m, 45.83 m, 46.63 m, 33.43 m, 34.63 m),

h = (205.96 m, 200.38 m, 200.83 m, 196.63 m, 198.43 m, 194.63 m),

and the following slight differences in the local surplus vectors defined in (2)

hs ≈ (0.8653, 0.346, 0.5277, 0.5543, 0.1143, 0.1543),

(15)
h = ps ≈ (0.1442, 0.0546, 0.0856, 0.0924, 0.0176, 0.0244).
s

Using (4), (9), (10) and (15), we then compute3

hs • p∗
Ir = γ(Q •H −q ∗ •h ∗ ) = 12 081.6
25 050 ≈ 0.48223.

4 The TLN with Non-uniform Minimal Design Pressure

Almost all WDNs in literature have uniform design conditions, when they are
present, for all the junction node. Here we have not the space to develop several
or new examples, hence we consider the cases studied in the previous section
but adding some alterations on the minimal design pressure request, i.e. on the
vector h∗ in such a way as to make its components no longer constant.

Example 5. Referring to the conditions of Example 3, we replace the vector h∗

with the following

h∗ = (30 m, 31 m, 30 m, 24 m, 33 m, 32 m),

hence
h∗ = (180 m, 191 m, 185 m, 174 m, 198 m, 192 m). (16)
This change affects the local head surplus vectors (and the local power one) as
follows
hs ≈ (0.8653, 0.3029, 0.5273, 0.9429, 0.013, −0.0066),
(17)
h = ps ≈ (0.1442, 0.0492, 0.0856, 0.1301, 0.0022, −0.0011),
s

and using (4), (16) and (17), we compute the resilience indices obtaining
hs • p∗
IR = Ir = γ(Q •H −q ∗ •h ∗ ) = 11 643.4
25 180 ≈ 0.46241. (18)

Comparing Example 3 and 5 is very interesting and it is enlightening about

the limits of global indices and the need for a local-global analysis of WDNs.
For brevity, let N3 be the network in Example 3 and N5 the one in Example 5
(N3 and N5 differ just for the design pressure minimal condition in the junction
3
The result, this time, agrees perfectly with 0.48 calculated in [21, Tab. 3].
 Algebraic Tools and New Local Indices for Water Networks 523

nodes 2,4,5,6). Note, first of all, that N5 is slightly more resilient than N3 ; more
precisely
Ir (N5 ) ≈ 1.006 % more than Ir (N3 ).
But N5 certainly has many more criticalities4 than the other: we just highlight
the two major ones. First of all, it has a pressure deficit in node 6 (note the
minus sign in the last component of hs in (17)). Then, as it can be shown using
EPANET 2.0.12, any failure, even very small at any point of the network, causes
a pressure deficit at node 5 as well, long before a pressure deficit occurs in the
network N3 . But it is also very important to note that the greater fragility of
N5 is not due to a higher request for minimum pressure in total, in fact it is
identical for both the networks:

h∗ (N3 )1 = h∗ (N5 )1 = 180 m (19)

(see (8) for the definition of the 1-norm  · 1 ). We therefore conclude that the
greater vulnerability of N5 , despite its resilience index, is mainly due to the
worst distribution between the nodes of the same total pressure design request
of 180 m as shown by (19). Such a worst distribution could be immediately
noted by comparing the local pressure surplus vectors hs (N3 ) and hs (N5 ): for
instance, just note that in the node 4 of N5 we have a pressure surplus of 94,29 %
(i.e. about twice the design request) vs the 55,43 % for the same node of N3 .

Aknowledgments. This work is partially supported by the research projects

“origAMI, Original Advanced Metering Infrastructure”, CUP J48C17000170006, POR
Calabria FESR-FSE 2014–2020, and “I-BEST”, CUP B28I17000290008, PON “Inno-
vazione e competitività” 2014/2020 MiSE Horizon 2020.

References
1. Alperovits, E., Shamir, U.: Design of optimal water distribution systems. Water
Resour. Res. 13, 885–900 (1977)
2. Bonora, M., Caldarola, F., Muranho, J., Sousa, J., Maiolo, M.: Numerical experi-
mentations for a new set of local indices of a water network. In: Sergeyev, Y. D. and
Kvasov, D. E. (eds.) Proceedings of the 3rd International Conference “Numerical
Computations: Theory and Algorithms”. Lecture Notes in Computer Science. vol.
11973. Springer, New York (2020). https://doi.org/10.1007/978-3-030-39081-5 42
3. Caldarola, F., Maiolo, M.: Local indices within a mathematical framework for
urban water distribution systems. Cogent Eng. 6, 1643057 (2019). https://doi.
org/10.1080/23311916.2019.1643057
4. Carini, M., Maiolo, M., Pantusa, D., Chiaravalloti, F., Capano, G.: Modelling and
optimization of least-cost water distribution networks with multiple supply sources
and users. Ricerche di Matematica 67(2), 465–479 (2018)

4
Being the two WDNs so similar and being dimension and complexity so small, a
“criticality” of one WDN with respect to the other must obviously be understood in
an appropriate sense...
524 F. Caldarola and M. Maiolo

5. Di Nardo, A., Di Natale, M.: A heuristic design support methodology based on

graph theory for district metering of water supply networks. Eng. Optim. 12, 193–
211 (2011)
6. Di Nardo, A., Di Natale, M.: A design support metodology metodology for dis-
trict metering of water supply networks. Water Distrib. Syst. Anal. 2010, 870–887
(2012)
7. Di Nardo, A., et al.: Redundancy features of water distribution systems. Proc. Eng.
186, 412–419 (2017)
8. Di Nardo, A., Di Natale, M., Santonastaso, G.: A comparison between different
techniques for water network sectorization. Water Sci. Technol.: Water Supply 14,
961–970 (2014)
9. Di Nardo, A., Di Natale, M., Santonastaso, G., Tzatchkov, V., Alcocer-Yamanaka,
V.: Water network sectorization based on graph theory and energy performance
indices. J. Water Resour. Plann. Manage. 140, 620–629 (2014)
10. Di Nardo, A., Di Natale, M., Santonastaso, G., Tzatchkov, V., Alcocer-Yamanaka,
V.: Performance indices for water network partitioning and sectorization. Water
Sci. Technol.: Water Supply 15, 499–509 (2015)
11. Horn, R., Johnson, C.: Matrix Analysis. Cambridge University Press, Cambridge
(2012)
12. Maiolo, M., Carini, M., Capano, G., Piro, P.: Synthetic sustainability index (SSI)
based on life cycle assessment approach of low impact development in the Mediter-
ranean area. Cogent Eng. 4, 1410272 (2017)
13. Maiolo, M., Martirano, G., Morrone, P., Pantusa, D.: Assessment criteria for a sus-
tainable management of water resources. Water Pract. Technol. 1(1): wpt2006012
(2006). https://doi.org/10.2166/wpt.2006.012
14. Maiolo, M., Mendicino, G., Pantusa, D., Senatore, A.: Optimization of drinking
water distribution systems in relation to the effects of climate change. Water 9,
803 (2017). https://doi.org/10.3390/w9100803
15. Maiolo, M., Pantusa, D.: A methodological proposal for the evaluation of potable
water use risk. Water Pract. Technol. 10, 152–163 (2015)
16. Maiolo, M., Pantusa, D.: An optimization procedure for the sustainable manage-
ment of water resources. Water Sci. Technol.: Water Supply 16, 61–69 (2016).
https://doi.org/10.2166/ws.2015.114
17. Maiolo, M., Pantusa, D.: Combined reuse of wastewater and desalination for the
management of water systems in conditions of scarcity. Water Ecol. 72, 116–126
(2017)
18. Maiolo, M., Pantusa, D.: Infrastructure vulnerability index of drinking water sup-
ply systems to possible terrorist attacks. Cogent Eng. 5, 1456710 (2018)
19. Maiolo, M., Pantusa, D., Chiaravalloti, F., Carini, M., Capano, G., Procopio, A.: A
new vulnerability measure for water distribution network. Water 10, 1005 (2019)
20. Shin, S., et al.: A systematic review of quantitative resilience measures for water
infrastructure systems. Water 10, 164 (2018)
21. Todini, E.: Looped water distribution networks design using a resilience index
based on heuristic approach. Urban Water 12, 115–122 (2000)
 Identification of Contamination Potential
Source (ICPS): A Topological Approach
for the Optimal Recognition of Sensitive
Nodes in a Water Distribution Network

Gilda Capano, Marco Amos Bonora, Manuela Carini , and Mario Maiolo(B)

Department of Environmental and Chemical Engineering,

University of Calabria, Rende, Italy
{gilda.capano,marcoamos.bonora,manuela.carini,mario.maiolo}@unical.it

Abstract. The correct management of urban water networks have to be

supported by monitoring and estimating water quality. The infrastruc-
ture maintenance status and the possibility of a prevention plan avail-
ability influence the potential risk of contamination. In this context, the
Contamination Source Identification (CSI) models aim to identify the
contamination source starting from the concentration values referring to
the nodes. This paper proposes a methodology based on Dynamics of
Network Pollution (DNP). The DNP approach, linked to the pollution
matrix and the incidence matrix, allows a topological analysis on the net-
work structure in order to identify the nodes and paths most sensitive
to contamination, namely those that favor a more critical diffusion of
the introduced contaminant. The procedure is proposed with the aim of
optimally identifying the potential contamination points. By simulating
the contamination of a synthetic network, using a bottom-up approach,
an optimized procedure is defined to trace back to the chosen node as
the most probable contamination source.

Keywords: Water quality · Contamination sources · Graph theory

1 Introduction
The sustainable management of water resources requires an efficient control of
the distribution systems performance to guarantee an adequate supply to the
users. A measure of non-achievement of qualitative and quantitative standards
is associated with the assessment of the potable water risk, which is particularly
useful for the careful planning of infrastructural and management interventions.
The potable water risk evaluation is complex because it depends on many fac-
tors that are sometimes difficult to estimate, for example, the source pollution,
obstructions or dysfunctions, the water quality alterations and water losses [18].
The evaluation of the water quality alterations due to accidental or intentional
events (natural or artificial) is of equal complexity. In fact, especially in recent
c Springer Nature Switzerland AG 2020
Y. D. Sergeyev and D. E. Kvasov (Eds.): NUMTA 2019, LNCS 11973, pp. 525–536, 2020.
https://doi.org/10.1007/978-3-030-39081-5_45
526 G. Capano et al.

years, the water systems vulnerability is considered a priority, for the impor-
tance dedicated to the critical infrastructures [2,21]. The water quality in the
drinking water networks, their distribution, utilization, discharge and purifica-
tion, aimed at reuse in the agricultural and industrial field, conditions typical
processes of the circular economy [19,20]. For these aspects, it is necessary to
equip the water systems with the monitoring of qualitative and quantitative
parameters using increasingly reliable instruments, based on real-time control,
in order to facilitate forecasting and risk prevention operations [17]. Being able
to continuously acquire reliable data on electro-filters placed in the network is a
very ambitious result also depending on the correct location of the water quality
sensors in the strategic network points. In this regard, in the literature, there
are different modeling approaches based on the correct positioning of the sen-
sors in the network [15,25]. This observation confirms the attention to this topic.
A correct survey of the location of the measuring instruments has to take into
account the objective (easy identification of the contamination source) allow-
ing the contaminant tracing. The contamination source identifying is a priority
problem. In scientific literature this problem has been widely discussed and inter-
preted using various methodologies, which are generally called Contamination
Source Identification (CSI) [2]. These models provide an adequate calculation
of three parameters: the location of the contamination source, the pollute con-
centration and the intrusion time, through different modeling approaches. [9]
propose a simulation-optimization method for complexes water distribution sys-
tem, which does not focus on a topological view. This method is based on an
optimal predictor-corrector algorithm to locate the sources and their release his-
tories. The optimization approach is used to define the similarity between the
simulated and measured output response data at monitoring points. [24] charac-
terize of the contaminant source with an optimization approach using a genetic
algorithm linked to the EPANET simulation software. The sources characteriza-
tion is based on the three different sensors types, assuming that contamination
intrusions are associated with a single location. [5] propose a methodology for
identifying the contamination source through an optimization problem using
the water fraction matrix concept. [4] propose a Bayesian belief network (BBN)
method which comparison the sensors data with other simulation of contami-
nation scenarios. The approach presented clarified how the uncertainties on the
mass and the position of the source influence the probability of sensors detect-
ing. [28] use a methodological approach based on the probability of interaction
between pollutant concentration data and injection duration with the feedback
provided by consumers. [16] propose an integrated simulation-optimization pro-
cedure with a logistic regression and a local improvement method to accelerate
convergence. This method is based on the pre and post screening technique with
the aim of accelerating convergence by reducing the investigation field. [13], using
Artificial Neural Networks (ANN), have developed a methodology to identify the
position of release of contaminants in the network. The water systems sensitivity,
in relation to the risk of the resource alteration, can detect erroneous dosages of
the reagents which are spilled into drinking water: among these, chlorine has a
 Identification of Contamination Potential Source (ICPS) 527

role of particular scientific interest and in the management practice. Chlorine,

which is often used as a water disinfectant, can be used as an indicator of the
gradual deterioration of the water quality. In fact, the chlorine concentration
released into the network by the treatment plant is progressively reduced due
to the reactions with the bacterial component. For this reason, it is important
to estimate the correct content of residual chlorine in the water network nodes.
The analysis models of chlorine decay in the network are many and focus their
attention on the characteristics of the decay reaction kinetics. In principle, most
of these models refer to [26], on which based the criteria set for quality simu-
lations in Epanet software. For these models, the chlorine decay occurs due to
the reactions that are generated within the mass flow and to the reactions that
are activated along the pipe wall. For this reason, the scientific interest focuses
on the estimation of the Bulk decay coefficient (kb) and the Wall decay coeffi-
cient (kw). The methods of calculating these parameters can be classification in
direct methods [1,22,27] which are based on the application of specific formula-
tions, and indirect methods [6,14,23] which mainly use calibration techniques.
In this work, an expeditious methodology is proposed for an easy identification
of sensitive nodes within an urban distribution network based on the Dynamics
of Network Pollution (DNP). The DNP summarizes a series of essential aspects
in such assessments as the concentration of the pollutants in the single node,
topology and the number of times a node appears in the potential pollutant
paths in the network (node occurrence). This last aspect represents an essential
element for the evaluation, because the most sensitive node (the node to which,
in the event of contamination, a worse pollution dynamic is associated) will be
identified in the list of nodes with the greatest number of occurrences. One of the
salient aspects of the proposed methodology is based on the predictive capacity
of the contamination source. The DNP represents a tool to obtain, information
on the potential location of meters (for this reason it can influence the definition
of monitoring costs). For this reason, the proposed methodological scheme is
not inventible but is bound to the logic of the detailed setting in the following
sections. The methodology is applied in the Kang and Lansey network (KL)
[11] which, according to a black box logic is contaminated providing a chlorine
spill outside the node to which, in the event of contamination, a worse pollution
dynamic is associated limits (0.2 mg/l).

2 Materials and Methods

The DNP is a useful tool to carry out a preliminary screening on the contaminant
diffusion in the network, contributing to the identification of sensitive paths and
nodes, starting from topological information. On an analytical level the DNP is
defined by the following matrix product:

DN P = IM x P M (1)

where: IM indicates the incidence matrix, with dimension (n, t), with n = nodes
number and t = pipelines number connecting nodes (arcs), and P M indicates
528 G. Capano et al.

the pollution matrix, with size (t, 1). DN P , having dimensions (n, 1), is a vector
that, starting from the concentration values in the pipelines, provides an estimate
of the pollutants concentration in the nodes. The DN P vector refers to a single
time instant and for this reason, for the time discretization k, it is necessary to
define a DN P for each k-th time sampling chosen. Regarding the relation (1),
P M is a vector that contains the average concentration values in the pipelines,
in a generic instant To + dt after the contamination occurred at in To time. This
quantity is weighed with respect to the water volume passed through the generic
pipeline and for this reason, it can be interpreted as a mass flow rate. This vector
is descriptive of a brief contamination scenario, defined by a criterion based on
the node sensitivity. The contamination scenario is defined in the hypothesis of
a chlorine overdose and is determined by the setting of the parameters of the
chlorine reaction kinetics. In the specific case, using the [26] approach, a first
order reaction is identified for the bulk coefficient, using a kb value of −0.55d−1
and neglecting, instead, the mass transport linked to the interaction with the
pipe walls. PM, therefore, contains values of the pollulant concentration in terms
of mg/l. The matrix differs from similar literature cases [5,12] because it does
not provide binary information, but identifies the pollutant quantity in the node,
defining the contribution of each node to the dynamic of contaminant diffusion.

2.1 Identification of Contamination Potential Sources

The definition of contamination potential sources in the KL network is aimed at
studying the sensitivity of the nodes involved in the pollution dynamics. In order
to carry out this investigation, it is necessary that the network be considered as
a weighted oriented graph (see [8])

N = (J, P )

where

J: the set of n vertices;

P : the set of m edges or arcs;
ρij : the non-negative weight associated to the arc ij if ij ∈ P
and ρij = 0 if ij ∈
/ P (i.e., if i, j are disconnected vertices in N ).

More in detail, the graph of the water network association considers the
demand nodes and the reservoirs/tanks as Junction and the pipeline, together
with the longitudinal elements (Valves and Pumps), as arches. The arcs direction
in the directed graph is defined by orientation of the pipes through the flow in
conditions of average flow (HYP). The arches weight is defined as the inverse of
the pipeline volume. This choice is linked to an application requirement, because
it is wanted to select the path with the highest pollutant volume by an algorithm
that identifies the paths with minimum weight:
1
ρij = (2)
Wij
 Identification of Contamination Potential Source (ICPS) 529

π(Dij )2
Wij = Lij (3)
4
Where:
ρij weight of the ij pipe [m3 ]
Wij volume of the ij pipe [m3 ]
Lij length of the ij pipe [m3 ]
Dij diameter of the ij pipe [m3 ]

However, as a preface to the definition of the procedure the following defini-

tions are useful. Given an oriented graph:

G = (V, E) (4)

with v ∈ V is a G vertex, the In-Degree of v in G is equal to the number of

incident arcs in v, while the Out-Degree of v in G is equal to the number of
incident arcs from v. Notations:

indeg(v) o indegG (v) outdeg(v) o outdegG (v) (5)

From a purely computational point of view, the implementation of the DN P

estimation algorithm is based on the identification of the Source and Sink nodes
[3,10] in the weighted oriented graph describing the network. Specifically, in a
directed graph G = (V, E) a node v ∈ V , which has only arcs starting from it,
is called Source:

indegG (v) = 0 outdegG (v) > 0 (6)

While, a node v ∈ V that has only incident arcs in it, is called Sink:

indegG (v) > 0 outdegG (v) = 0 (7)

This definition highlights a hierarchical criterion between nodes, based on the

definition of dominator. In a oriented graph G = (V, E) taken the nodes d, ∈ V ,
it can say that d dominates n if every path that starts from the entry node to
reach n node, has to pass through d. Notation:

d dom n (8)

d dom n con d = n (9)

To identify the most sensitive nodes within the network an automatic procedure
has been implemented that develops in two phases. In the first phase the nodes
to be discarded are identified, according to the following criterion: all the nodes
that are “directly dependent” from the reservoir will be discarded. A node is said
to be directly dependent on the reservoir nodes if it is Strictly Dominated by
one of them or if it is Strictly Dominated by a node dominated by the reservoir.
More in detail it can hypothesize the following synthetic schematization
530 G. Capano et al.

– source nodes are placed in a list;

– source nodes will be classified: (i) strictly Dominated by the nodes in the list;
(ii) with indeg = 1; (iii) that are not terminal nodes.

These nodes are defined as “directly dependent” on the reservoir and are placed
in the list of excluded nodes. The “directly dependent” nodes identified in the
previous phase are added to the list. The procedure is repeated identifying the
nodes Strictly Dependent from other nodes which are Strictly Dependent on the
reservoir. In the second phase, a minimum path search algorithm is applied, using
the inverse of the volume to identify the paths with maximum water volume in
the pipes. Considering that the weights are always positive, the Dijkstra algo-
rithm is used [7]. This is applied to the chosen network, interpreted as an oriented
graph, and the minimum weight path that connects, if possible, each Source node
to each Sink node is identified. From the list of paths with maximum volume
the number of occurrences of each node is obtained. The occurrences number is
the number of times a node appears in the paths found. The importance linked
to the quantification of node occurrences depends on the objective of the work,
which, in a network composed of n nodes with different topological hierarchical
levels of participation in the pollution dynamics, wants to identify the single
most significant nodes. These nodes are the nodes that appear many times, that
is those present in the more sensitive paths with maximum volume. In fact, the
most sensitive node or path will be identified in the nodes list with the greatest
occurrences number. The procedure is applied to KL network (Fig. 1), which has
935 nodes and 1274 pipes. The average total demand is 177 l/s. The network is
simulated under peak conditions with the total consumption is 336 l/s.

3 Numerical Results
In order to identify the most sensitive node in the KL network, classified as a
potential source of contamination, it is necessary to discuss the results obtained
for individual application phases. The first part of the procedure, aimed at car-
rying out a preliminary screening of the nodes hierarchy, allows to eliminate the
obligatory passage nodes (the source nodes and the nodes Strictly Dependent on
the reservoir) and the terminal nodes (Fig. 2a). Then proceed with the calcula-
tion of the occurrences from the paths list with maximum volume (Fig. 2b). The
Fig. 2b has an important role, because it makes clear the criterion for identifying
the interesting node for contamination. This node will be chosen among those
that have the highest number of filtered occurrences. The total occurrences (or
simply occurrences) are obtained by summing the number of times that the node
appears in all paths. The filtered occurrences instead discriminate the reservoir
node and do not count the occurrences of the “obligatory passage nodes” in the
single Dijkstra application, that is those excluded in the first selection phase.
 Identification of Contamination Potential Source (ICPS) 531

Fig. 1. Kang and Lansey network.

The histogram shows that the most critical situation is associated with nodes
249 and 250, which have a 24 value for a filtered occurrence. According to the
criterion identified, the choice between the two might seem equivalent, however,
in the network topology it is clear that 249 is upstream of 250, so it will be
chosen as a “node to be polluted”. The hydraulic simulation is carried out on
a permanent motion setting, forecasting an impulsive input of chlorine in the
chosen node, lasting 30 min. The pollutant diffusion is shown in Fig. 3, where it
has been chosen to show the trend of the chlorine concentration.
Now it is possible to proceed with the DNP calculation. The DNP values
are calculated for each node and for each time interval. For reasons of synthesis
and clarity, only the DNP values associated with the nodes identified in Fig. 2a
will be shown for a period of 8 hours. It is important to specify that the DNP
vector takes positive and negative values. The DNP values are indicative of
the pollutant balance in the single node. Negative values indicate the pollutant
mass leaving the node and vice versa the positive values. In the case of the
path identified in Fig. 2a, the trend of the DNP can be summarized by the
following graph. The trends of the curves oscillate according to the pipes number
and their characteristics, which converge in the single node, determining the
contaminant dilution more or less fastly. The information that can be obtained
from the resulting DNP values are of two types: the peaks values, both positive
and negative, provide quantitative information that allows an indication of the
decay degree and mixing of the pollutant. Their position and shape, on the
532 G. Capano et al.

Fig. 2. (a) The occurrences location in the KL network. The red crosses indicate the
obligatory passage nodes, the black crosses the terminal nodes, the yellow circles indi-
cate the nodes with the maximum filtered occurrences and the green circle indicates
the chosen node (ID 249). (b) Histogram of total and filtered occurrences of the nodes
in the most sensitive path. (Color figure online)

other hand, provides an indication of the network pollution dynamics. As an

example, Fig. 5 shows the DNP graph for some particularly significant nodes. As
mentioned above, with the DNP there is information on the pollution dynamics
in the network, which for some nodes is slow and lasts over time, while for
others it is impulsive. Through the graphic representation of the DNP it can
therefore have immediate feedback on the behavior of some nodes. Some nodes,
among those closest to the immission area, show an impulsive trend, indicative
of a temporally rapid and highly concentrated pollution dynamics, while, other
downstream nodes are distinguished by a trend in which the peak less high,
lasts over time. The first situation indicates the pollution periods with high
concentrations with minor duration, while, in the second case, the “polluted”
period continues for longer times. From the evaluation of the results it is clear
that the interesting node cannot be one, since two different pollution dynamics
have been identified. These nodes are the 253, with the maximum DNP value
(impulsive behavior) and the node 256, which corresponds to a contamination
dynamic more durable in overtime (Fig. 4).
 Identification of Contamination Potential Source (ICPS) 533

Fig. 3. Network contamination status. (a) situation after 30 min, (b) situation after
1 h and (c) situation after 4 h.

Fig. 4. DNP values for the nodes along the path shown (highlighted in yellow) in Fig. 2.
(Color figure online)
534 G. Capano et al.

Fig. 5. DNP trends for the most significant nodes, useful for peak identifying.

4 Conclusions

The water quality of a drinking water network can be easily compromised during
transport to users due to the intrusion of a pollutant through tanks, nodes and
broken pipes. In this work, an expeditious procedure is proposed for the identi-
fication of one or more sensitive nodes in a network, which are the nodes that
can determine a worse contaminant distribution if contaminated. The method is
based on the DNP calculation, a vector that for each node and for each chosen
time sampling, provides an indicator of the pollutant quantity that switches.
DNP is based on a quick procedure that summarizes hydraulic and topological
evaluations of the network. The importance linked to the topology determines
the application conditions of the method, in fact, it is not possible to define
a priori and absolutely the necessary measurement points because this infor-
mation is linked to the topological characteristics of the network chosen as a
case study. The importance linked to hydraulic evaluations, on the other hand,
does not constrain the study to the analysis of chlorine concentration alone,
but the interest can be directed towards other characteristic parameters (PH,
temperature), considering that their variability defines the conditions of decay
of the chlorine. The case study is well suited to experiment with the validity
of the proposed methodology, as it is characterized by a good number of loops
and therefore by a high degree of contaminant dilution. The particularity and
the interest of the procedure depend on the ability to manage a problem with
unknown input parameters and evaluations on partially deductible outputs.
 Identification of Contamination Potential Source (ICPS) 535

Aknowledgments. This work is partially supported by the Italian Regional Project

(POR CALABRIA FESR 2014–2020): “origAMI, Original Advanced Metering Infras-
tructure” CUP J48C17000170006.

References
1. Al-Jasser, A.O.: Chlorine decay in drinking-water transmission and distribution
systems: pipe service age effect. Water Res. 41(2), 387–396 (2007). https://doi.
org/10.1016/j.watres.2006.08.032
2. Adedoja, O.S., Hamam, Y., Khalaf, B., Sadiku, R.: Towards development of an
optimization model to identify contamination source in a water distribution net-
work. Water 10(5), 579–606 (2018). https://doi.org/10.3390/w10050579
3. Borowski, E.J., Borwein, J.M.: The HarperCollins Dictionary of Mathematics.
HarperCollins, New York (USA) (1991)
4. Dawsey, W.J., Minsker, B.S., VanBlaricum, V.L.: Bayesian belief networks to
integrate monitoring evidence of water distribution system contamination. J.
Water Resour. Plann. Manage. 132(4), 234–241 (2006). https://doi.org/10.1061/
(ASCE)0733-9496(2006)132:4(234)
5. Di Cristo, C.D., Leopardi, A.: Pollution source identification of accidental contam-
ination in water distribution networks. J. Water Resour. Plann. Manage. 134(2),
197–202 (2008). https://doi.org/10.1061/(ASCE)0733-9496(2008)134:2(197)
6. Digiano, F.A., Zhang, W.: Pipe section reactor to evaluate chlorine-wall reaction.
J.-Am. Water Works Assoc. 7(1), 74–85 (2005)
7. Dijkstra, E.W.: A note on two problems in connexion with graphs. Numerische
Mathematik 1(1), 269–271 (1959)
8. Di Nardo, A., Di Natale, M.: A heuristic design support methodology based on
graph theory for district metering of water supply networks. Eng. Optim. 43(2),
193–211 (2011). https://doi.org/10.1080/03052151003789858
9. Guan, J., Aral, M.M., Maslia, M.L., Grayman, W.M.: Identification of contaminant
sources in water distribution systems using simulation-optimization method: case
study. J. Water Resour. Plann. Manage. 132(4), 252–262 (2006). https://doi.org/
10.1061/(ASCE)0733-9496(2006)132:4(252)
10. Harary, F.: Graph Theory. Reading. Addison-Wesley, Boston (USA) (1994)
11. Kang, D., Lansey, K.: Revisiting optimal water-distribution system design: issues
and a heuristic hierarchical approach. J. Water Resour. Plann. Manage. 138(3),
208–217 (2012). https://doi.org/10.1061/(ASCE)WR.1943-5452.0000165
12. Kessler, A., Ostfeld, A., Sinai, G.: Detecting accidental contaminations in munic-
ipal water networks. J. Water Resour. Plann. Manage. 124(4), 192–198 (1998).
https://doi.org/10.1061/(ASCE)0733-9496(1998)124:4(192)
13. Kim, M., Choi, C.Y., Gerba, C.P.: Source tracking of microbial intrusion in water
systems using artificial neural networks. Water Res. 42(4–5), 1308–1314 (2008).
https://doi.org/10.1016/j.watres.2007.09.032
14. Kim, H., Kim, S., Koo, J.: Modelling chlorine decay in a pilot scale water distri-
bution system subjected to transient. In: Civil and Environmental Engineering,
Pusan National University (Korea) and Environmental Engineering, University of
Seoul (2015) https://doi.org/10.1016/j.proeng.2015.08.89
15. Kim, S.H., Aral, M.M., Eun, Y., Park, J.J., Park, C.: Impact of sensor measure-
ment error on sensor positioning in water quality monitoring networks. Stoch. Envi-
ron. Res. Risk Assess. 31(3), 743–756 (2017). https://doi.org/10.1007/s00477-016-
1210-1
536 G. Capano et al.

16. Liu, L., Zechman, E.M., Mahinthakumar, G., Ranji Ranjithan, S.: Identifying con-
taminant sources for water distribution systems using a hybrid method. Civil
Eng. Environ. Syst. 29(2), 123–136 (2012). https://doi.org/10.1080/10286608.
2012.663360
17. Maiolo, M., Carini, M., Capano, G., Pantusa, D., Iusi, M.: Trends in metering
potable water. Water Pract. Technol. 14(1.1), 1–9 (2019). https://doi.org/10.2166/
wpt.2018.120
18. Maiolo, M., Pantusa, D.: A methodological proposal for the evaluation of potable
water use risk. Water Pract. Technol. 10(1), 152–163 (2015). https://doi.org/10.
2166/wpt.2015.017
19. Maiolo, M., Pantusa, D.: Combined reuse of wastewater and desalination for the
management of water systems in conditions of scarcity. Water Ecol. 4(72), 116–126
(2017). https://doi.org/10.23968/2305-3488.2017.22.4.116-126
20. Maiolo, M., Pantusa, D.: A proposal for multiple reuse of urban wastewater. J.
Water Reuse Desalin. 8(4), 468–478 (2018a). https://doi.org/10.2166/wrd.2017.
144
21. Maiolo, M., Pantusa, D.: Infrastructure Vulnerability Index of drinking water sys-
tems to terrorist attacks. Cogent Eng. 5(1), 1456710 (2018b). https://doi.org/10.
1080/23311916.2018.1456710
22. Mostafa, N.G., Minerva, E., Halim, H.A.: Simulation of Chlorine Decay in Water
Distribution Networks. Public Works Department, Faculty of Engineering, Cairo
(Egypt), Using EPANET. Case Study - Sanitary and Environmental Engineering
Division (2013)
23. Nagatani, T., et al.: Residual chlorine decay simulation in water distribution sys-
tem. In: The International Symposium on Water Supply Technology, Yokohama
(Japan) (2008)
24. Preis, A., Ostfeld, A.: Genetic algorithm for contaminant source characterization
using imperfect sensors. Civil Eng. Environ. Syst. 25(1), 29–39 (2008). https://
doi.org/10.1080/10286600701695471
25. Rashid, B., Rehmani, M.H.: Applications of wireless sensor networks for urban
areas: a survey. J. Netw. Comput. Appl. 60, 192–219 (2016)
26. Rossman, L.A., Clark, R.M., Grayman, W.M.: Modeling chlorine residuals in drink-
ing water distribution systems. J. Environ. Eng. 120(4), 803–820 (1994)
27. Rossman, L.A.: EPANET 2: Users Manual (2000)
28. Tao, T., Huang, H.D., Xin, K.L., Liu, S.M.: Identification of contamination source
in water distribution network based on consumer complaints. J. Central S. Univ.
Technol. 19, 1600–1609 (2012)
 Seeking for a Trade-Off Between
Accuracy and Timeliness
in Meteo-Hydrological Modeling Chains

Luca Furnari(B) , Alfonso Senatore , and Giuseppe Mendicino

DIATIC, University of Calabria, Rende, CS, Italy

{luca.furnari,alfonso.senatore,giuseppe.mendicino}@unical.it

Abstract. The level of detail achieved by operational General Circula-

tion Models (e.g., the HRES 9 km resolution forecast recently launched
by the ECMWF) raises questions about the most appropriate use of
Limited Area Models, which provide for further dynamical downscaling
of the weather variables. The two main objectives targeted in hydro-
meteorological forecasts, i.e. accuracy and timeliness, are to some extent
conflicting. Accuracy and precision of a forecast can be evaluated by
proper statistical indices based on observations, while timeliness mainly
depends on the spatial resolution of the grid and the computational
resources used. In this research, several experiments are set up apply-
ing the Advanced Research Weather Research and Forecasting (WRF-
ARW) Model to a weather event occurred in Southern Italy in 2018.
Forecast accuracy is evaluated both for the HRES ECMWF output and
that provided by WRF dynamical downscaling at different resolutions.
Furthermore, timeliness of the forecast is assessed adding to the time
needed for GCM output availability the time needed for Limited Area
simulations at different resolutions and using varying core numbers. The
research provides useful insights for the operational forecast in the study
area, highlighting the level of detail required and the current weaknesses
hindering correct forecast of the hydrological impact of extreme weather
events.

Keywords: Hydro-meteorological modeling · Warning lead time ·

Weather model resolution

1 Introduction
Numerical weather prediction (NWP) is usually based on modeling chains where
forecasts are dynamically downscaled from a General Circulation Model (GCM)
coarse grid (of the order of ∼ 101 km) to the desired resolution (∼ 100 km).
GCM forecasts provide the initial and boundary conditions to the Limited Area
Models (LAMs), which adopt finer grids to improve the accuracy. Downscaled
fields of meteorological variables can be used then for many purposes, e.g. simply
like triggers for activating warning procedures or as input data for hydrological
c Springer Nature Switzerland AG 2020
Y. D. Sergeyev and D. E. Kvasov (Eds.): NUMTA 2019, LNCS 11973, pp. 537–544, 2020.
https://doi.org/10.1007/978-3-030-39081-5_46
538 L. Furnari et al.

models, in order to predict the ground impact of the meteorological forcings

(e.g., [4,10]).
Though very widely used, the dynamical downscaling approach is continu-
ously subject to very detailed analysis in order to understand its real benefits in
terms of forecasting accuracy (e.g., [5,7,9]), despite the greater computational
burden required and the longer calculation times, which can be an important
issue in operational contexts. This research question is particularly timely if the
recent amazing improvements of the operational GCMs are considered, whose
current spatial resolutions are comparable to that of LAMs run operationally
only a few years ago. Discussion within the scientific community about the
real improvement given by GCMs’ increased resolution [13] is providing posi-
tive feedback [8]. A very important novelty from an operational point of view is
that on March 2016 the European Centre for Medium-Range Weather Forecasts
(ECMWF) released a new version of its operative GCM forecast model, the Inte-
grated Forecasting System - High Resolution (IFS-HRES), where the horizontal
resolution of the deterministic forecasts has been increased from 16 km to 9 km.
This work presents a preliminary analysis over a single case study about
which output resolution of an NWP modeling chain could be the optimal trade-
off between the accuracy of the forecasting and the computational cost. The
selected event affected the northeastern side of the Calabrian peninsula (south-
ern Italy) during summer of 2018. From a forecasting point of view, it is very
challenging, due to the deep convective activity leading to intense and highly
localized rainfall. Nevertheless Calabria, due to its complex orography and its
position in the middle of the Mediterranean sea, is often interested by this type
of high energy events, which are being subjected to several studies (e.g., [2,11]).
The paper is structured as follows: next Sect. 2 briefly describes the event
analyzed and the modeling strategy adopted, Sect. 3 analyzes the results of sim-
ulations both in terms of accuracy and computational performances and finally
in the conclusions (Sect. 4) the main findings of the research are summarized
and briefly discussed.

2 Data and Methods

2.1 Description of the Event
On the early afternoon of 20 August 2018 a highly localized and intense event hit
the northeastern part of the Calabria region (southern Italy). Specifically, the
event affected the Raganello Creek catchment, a small Mediterranean catchment
characterized by a complex and irregular orography in its upper/mountain area,
which amplified the convective characteristics of the event. Unfortunately, 10
casualties occurred nearby the village of Civita (downstream a catchment area
of about 100 km2 ), despite the local flood warning agency (the Centro Funzionale
Multirischi) emitted a 2 out of 4 level warning message.
One important feature of this event is that the regional rain gauge network
was not able to detect significant rainfall measures, as Fig. 1b shows. The main
source publicly released about the real extent and intensity of the event is the
 Seeking for a Trade-Off Between Accuracy and Timeliness in NWP 539

technical report of the event by the Centro Funzionale Multirischi [3], where a
static image of the Surface Rainfall Total (SRT) derived from the national radar
network is shown Fig. 2a. This SRT image, showing the accumulated three-hour
rainfall from 10 to 13 UTC of 20 August 2018, highlights that only a small area
of the catchment, not covered by rain gauges, was interested by rain. This static
image will be also used as comparison with the output of the meteorological
model.

2.2 Modeling Strategy

The Limited Area Model (LAM) adopted in this study is the Weather Research
and Forecasting model, in the Advanced Research Weather core (WRF-ARW,
[12]), version 4.1. WRF and its libraries, such as NetCDF, HDF5, zlib and others,
were compiled using Intel Compiler version 12.1.4.
Two different domain configurations were used in this work. The first (called
C1 hereafter) is the same used in [1], with two nested (one-way) grids, the out-
ermost (D01) covering the central Mediterranean basin at 6 km horizontal reso-
lution with 312 × 342 cells and the innermost (D02) focusing on the Calabrian
Peninsula with 2 km resolution (200 × 200 cells) (Fig. 1a). Both computational
domains extend over 44 vertical layers. In order to guarantee numerical stability
of the model, the time step size was set equal to 36 s and to 12 s, respectively
for D01 and D02.

Fig. 1. (a) Overview of the two grids nested used in configuration C1, (b) zoom on the
Raganello Creek catchment closed at Civita (red point), showing rain gauges measure-
ments, orography and river network. (adapted from [1]) (Color figure online)

The second domain configuration (called C2 hereafter) adopts only one

domain, whose geographic coverage is approximately the same as the D02 of C1
configuration, with 3 km grid spacing (140 × 140 grid points). Also, this domain
configuration extends over 44 vertical layers. Time step size was fixed to 18 s.
540 L. Furnari et al.

All simulations forecasted 24 h, starting at 00 UTC of 20 August 2018, in

order to guarantee enough spin up time before the occurrence of the event. The
ECMWF Integrated Forecasting System (IFS) in its high-resolution determinis-
tic forecast version (HRES - 9 km horizontal resolution) was used to provide ini-
tial and boundary conditions every 3 h. The ECMWF products, even if referred
to 00 UTC, are available starting from 6 UTC, due to the time needed by GCM
to complete 24 h of simulation. Physical schemes adopted for both configurations
are summarised in Table 1. Cumulus parameterization adopted was the Tiedtke
cumulus scheme, turned on only on D01 domain in the C1 configuration, whose
horizontal resolution is greater than 5 km.
All simulations were performed on the 4-socket/16-core AMD Opteron 62xx
class CPU 2.2 GHz provided by the ReCaS infrastructure (for further details
please refer to [6]).

Table 1. WRF physical parameterization.

Component Scheme adopted

Microphysics New Thompson
Planet boundary layer Mellor-Yamada-Janjic scheme
Shortware radiation Goddard
Longwawe radiation Rapid Radiative Transfer Model (RRTM)
Land surface model Unified NOAH

3 Results
Simulations results are shown in Figs. 2b–d and are directly comparable with
the SRT radar image (Fig. 2a). The radar image shows a large area in the north-
ern part of Calabria interested by medium/high precipitation intensity, with a
peak localized in the proximity of the northern boundary of the Raganello Creek
catchment. Both radar and simulations show a kind of “C” shape for the precipi-
tation pattern and highlight the effect of surrounding mountain ranges, acting as
obstacles and contributing to enhancing the development of the convective cell.
The rainfall amount clearly increases with the resolution. Lower rainfall amounts
in the ECMWF forecast and in configuration C2 are connected to the coarser
resolutions that both flatten rainfall values over larger cell areas and smooth the
complex orographic features of the region.
In agreement with [1], WRF simulations delayed about one-hour the rain-
fall peak observed by the radar, with the highest simulated rainfall intensity
occurring between 13 and 14 UTC (this is the main reason for the discrepancies
between observations and simulations in Fig. 2). Configuration C1 forecasted
about 20 mm averaged rainfall over the catchment between 10 and 14 UTC; for
the same time, configuration C2 and ECMWF forecasts simulated only about
 Seeking for a Trade-Off Between Accuracy and Timeliness in NWP 541

Fig. 2. (a) Radar SRT estimate between 10 and 13 UTC, the black line represents
the border of the catchment and the red point the outlet at Civita; (b) ECMWF 9 km
accumulated rainfall forecast for the same time; (c) WRF 3-h accumulated precipitation
provided by configuration C1, the black line represents the catchment borders, the black
dot Civita, the grey lines are the regional administrative borders, (d) same as c, but
for configuration C2. (Color figure online)

2 mm and 3 mm average precipitation over the catchment, respectively. Specifi-

cally, the drier configuration C2 concentrated rainfall over smaller areas than C1,
thus missing the event over the Raganello Creek catchment and being not eligi-
ble to perform further hydrological simulations. However, it was able to simulate
medium/high rainfall amounts not far from the western catchment borders, with
a peak value of about 72 mm in three hours. With configuration C1 the peak
amount of 79 mm was located north of the catchment.
Both configurations C1 and C2 proved to be able to forecast enough high
precipitation amounts to trigger warning procedures, even though the level of
accuracy concerning the correct localization of the event was different.
The computational burden was evaluated through an analysis of the com-
putational scalability of the system, using 1, 2, 4, 8, 16, 32 and 64 threads
respectively. Execution times were calculated using the default WRF output con-
figuration, which produces relatively small output files (WRF allows choosing
several sets of variables, depending on the aims of the users, which could easily
generate very large output files and not negligible writing times). Of course, the
542 L. Furnari et al.

processing time is very sensitive to the different time steps (increasing with the
resolution), the number of cells in the computational domain(s) and also the
number of domains, due to the nesting procedure. Therefore, it is expected that
configuration C1, having two domains and higher resolution in the innermost
domain, is penalized with respect to configuration C2.
Figure 3a, showing the total execution times of the two different configura-
tions depending on the number of threads used, clearly highlights that, given the
same number of threads, configuration C2 is always faster than C1, with execu-
tion times reduced from 60% to 70%. The optimal number of threads resulted
in 32 for both configurations, with 83 and 23 min, respectively for C1 and C2.
Speed-up (Fig. 3b), calculated as the ratio between the time taken by the best
available serial algorithm and the time taken by the parallel algorithm running
in different threads, also provides better results for configuration C2 (speed-
up equal to 16 with 32 threads) and highlights, for both configurations, the
reduced performance using 64 threads. The latter result is most probably due to
the increasing communication time between distinct threads, overhanging and
hiding the gain given by the greater computing power. The higher speed-up of
configuration C2 also means that it is more able to exploit the available comput-
ing power. This outcome is connected to the lack of nesting between domains.
Efficiency (calculated as the ratio between speed up and number of threads)
with 32 threads is equal to 0.36 for C1 and 0.49 for C2.

Fig. 3. (a) Total execution time achieved by WRF with configurations C1 (black line)
and C2 (red line); (b) Speed-up obtained by WRF with configurations C1 (black line)
and C2 (red line) (Color figure online)

4 Conclusions
The paper presented a case study investigating the optimal configuration of the
NWP modeling chain, balancing accuracy and timeliness. The results showed
that: (1) notwithstanding the relatively high resolution of the GCM, dynamical
downscaling is required to reproduce reliably the high convective event; (2) both
configurations C1 and C2 of the LAM are capable to reproduce the main features
 Seeking for a Trade-Off Between Accuracy and Timeliness in NWP 543

of the event, forecasting enough rainfall amount to trigger warning procedures

for a larger area than the Raganello Creek catchment, but (3) lower resolution
configuration C2 almost completely misses the localization of the event over
the catchment, preventing any further meaningful hydrological simulation, and
(4) higher resolution configuration C1 takes about 1 more hour, optimizing the
available computational resources, to provide its results, therefore reducing the
forecast lead time.
With the conditions analyzed in this specific case study, it would have been
possible to exploit all the 64 available threads running simultaneously both con-
figurations C1 and C2, having right away a general framework and then more
detailed results. Of course, it is a specific solution for this case, envisaging how-
ever a more complex (and most probably needed for this kind of events) ensemble
approach. With different domain extents or different architecture of the com-
puting infrastructure (e.g., higher CPU clock speed), alternative solutions could
have been found. Also, the number of case studies analyzed needs to be highly
increased to get more general findings. Nevertheless, as a preliminary indication,
the analysis performed suggests that, fixed specific lead time, in topographi-
cally complex areas it is desirable to provide forecasts with the highest space
resolution as possible.

Acknowledgements. L. Furnari acknowledges for the program “POR Calabria

FSE/FESR 2014/2020 - Mobilitá internazionale di Dottorandi e Assegni di
ricerca/Ricercatori di Tipo A” Actions 10.5.6 and 10.5.12. We thank the “Centro
Funzionale Multirischi” of the Calabrian Regional Agency for the Protection of the
Environment, for providing the observed precipitation data. We, also, would like to
acknowledge the INFN group at University of Calabria for giving the opportunity to
use their cluster.

References
1. Avolio, E., Cavalcanti, O., Furnari, L., Senatore, A., Mendicino, G.: Brief commu-
nication: preliminary hydro-meteorological analysis of the flash flood of 20 august
2018 in Raganello Gorge, Southern Italy. Nat. Hazards Earth Syst. Sci. 19(8),
1619–1627 (2019)
2. Avolio, E., Federico, S.: WRF simulations for a heavy rainfall event in southern
Italy: verification and sensitivity tests. Atmos. Res. 209, 14–35 (2018)
3. CFM: Evento meteopluviometrico del 20 agosto 2018 - torrente raganello. Technical
report, Centro Funzionale Multirischi della Calabria (2018)
4. Davolio, S., Miglietta, M.M., Diomede, T., Marsigli, C., Morgillo, A., Moscatello,
A.: A meteo-hydrological prediction system based on a multi-model approach for
precipitation forecasting. Nat. Hazards Earth Syst. Sci. 8(1), 143–159 (2008)
5. Di Luca, A., Argüeso, D., Evans, J.P., de Elı́a, R., Laprise, R.: Quantifying the
overall added value of dynamical downscaling and the contribution from different
spatial scales. J. Geophys. Res. Atmos. 121(4), 1575–1590 (2016)
6. Guarracino, N., Lavorini, V., Tarasio, A., Tassi, E.: The ReCaS project cosenza
infrastructure, pp. 43–55 (2017). Chapter 5
544 L. Furnari et al.

7. Hong, S.Y., Kanamitsu, M.: Dynamical downscaling: fundamental issues from an

NWP point of view and recommendations. Asia Pac. J. Atmos. Sci. 50(1), 83–104
(2014)
8. Neumann, P., et al.: Assessing the scales in numerical weather and climate pre-
dictions: will exascale be the rescue? Philos. Trans. Roy. Soc. A Math. Phys. Eng.
Sci. 377(2142) (2019). 20180148
9. Rummukainen, M.: Added value in regional climate modeling. Wiley Interdisc.
Rev. Clim. Change 7(1), 145–159 (2016)
10. Senatore, A., Mendicino, G., Gochis, D.J., Yu, W., Yates, D.N., Kunstmann, H.:
Fully coupled atmosphere-hydrology simulations for the central mediterranean:
impact of enhanced hydrological parameterization for short and long time scales.
J. Adv. Model. Earth Syst. 7(4), 1693–1715 (2015)
11. Senatore, A., Mendicino, G., Knoche, H.R., Kunstmann, H.: Sensitivity of modeled
precipitation to sea surface temperature in regions with complex topography and
coastlines: a case study for the mediterranean. J. Hydrometeorol. 15(6), 2370–2396
(2014)
12. Skamarock, W.C., et al.: A description of the advanced research WRF version 3.
NCAR technical note-475+str (2008)
13. Wedi, N.P.: Increasing horizontal resolution in numerical weather prediction and
climate simulations: illusion or panacea? Philos. Trans. Roy. Soc. A Math. Phys.
Eng. Sci. 372(2018), 20130289 (2014)
 Optimization Model for Water
Distribution Network Planning
in a Realistic Orographic Framework

Mario Maiolo1(B) , Joaquim Sousa4 , Manuela Carini1 ,

Francesco Chiaravalloti2 , Marco Amos Bonora1 , Gilda Capano1 ,
and Daniela Pantusa3
1
Department of Environmental and Chemical Engineering (DIATIC),
Università della Calabria, Cubo 42/B, 87036 Arcavacata di Rende, CS, Italy
{mario.maiolo,manuela.carini,marcoamos.bonora,gilda.capano}@unical.it
2
Research Institute for Geo-Hydrological Protection (IRPI), CNR,
87036 Rende, CS, Italy
[email protected]
3
Innovation Engineering Department, University of Salento, Lecce, Italy
[email protected]
4
Department of Civil Engineering, Polytechnic Institute of Coimbra,
Coimbra, Portugal
[email protected]

Abstract. Defining criteria for correct distribution of water resource is

a common engineering problem. Stringent regulations on environmental
impacts underline the need for sustainable management and planning
of this resource usage, which is sensitive to many parameters. Opti-
mization models are often used to deal with these problems, identify-
ing the optimal configuration of a Water Distribution Network (WDN)
in terms of minimizing an appropriate function propotional to he con-
struction cost of the WDN. Generally, this cost function increases as
the distance between the source-user connection increases, therefore in
minimum cost optimization models is important to identify minimum
source-user paths compatible with the orography. In this direction, the
methodology presented in the present work proposes a useful approach to
find minimum-length paths on surfaces, which moreover respect suitable
hydraulic constraints and are therefore representative of reliable gravity
water pipelines. The application of the approach is presented in a real
case in Calabria.

Keywords: Optimization model · Water planning model · Graph

theory

1 Introduction
Due to the possible presence of multiple water sources and different types of
conflicting water users, one of the most challenging problems in water resources
c Springer Nature Switzerland AG 2020
Y. D. Sergeyev and D. E. Kvasov (Eds.): NUMTA 2019, LNCS 11973, pp. 545–556, 2020.
https://doi.org/10.1007/978-3-030-39081-5_47
546 M. Maiolo et al.

management is the optimal allocation of water resources with respect the water
demand and availability [4,11]. An useful approach to deal with this problem is to
identify the optimal configuration of a Water Distribution Network (WDN)[3,6]
in terms of minimizing an appropriate function which is proportional to the
construction cost of the WDN. In particular, in [2,8] the authors propose a
minimum-cost optimization model that determines an idealised water distribu-
tion system providing the optimal allocations of water resources among different
sources and users. The cost of the WDN, which defines the optimization problem,
is in the form: 
C∝ f (Qi,j , Li,j , hi,j ) (1)
i,j

where Qi,j is the flow rate between the source i and the user j, Li,j is the
source-user distance and hi,j the altitude difference (piezometric head difference
between i and j). Beyond the details of the expression (1), it is obvious that the
cost of a pipeline increases as its length increases. Therefore, it is reasonable to
consider minimum-length path joining i and j. A critical approximation in [2,8]
is the calculation of Li,j as Euclidean distance between the points i nd j.
With this approach the proposed procedure is expeditious, because only posi-
tions of source and destinations are needed in order to completely define the
geometry of the problem, but the resulting optimal minimum-cost WDN is less
representative from a physically feasible hydraulic infrastructure.
An improvement in this sense is obtained by taking into account the orog-
raphy of the territory and by calculating Li,j as the length of the shortest path
between i and j, lying on the orography.
The present work describes a methodology that allows to identify the shortest
path between two point on a topographic surface which moreover is compatible
with appropriate hydraulic criteria, making it possible the use of such path as a
realistic trace for a gravity water pipeline.
Finding shortest paths (geodesic) on surfaces is a challenging problem in
computational geometry, with important application in robotics, path planning,
texture mapping, computer graphics [9].
A possible approach to this problem consists in converting the geometric
problem to a graph problem and find approximate solutions [1,7,10]. In effect,
the surface under examination can be transformed into a graph and procedures,
e.g. Dijkstra’s algorithm [5], capable to identify minimum paths joining two
nodes of the graph can be used.
A 2-dimensional surface S embedded in R3 can be represented with a polyhe-
dral surface, i.e. a set of polygonal faces that constitutes a piecewise-flat approx-
imation of the surface. Two polygons do not intersect, except at a common point
v (vertex) or an a common edge e. Sampling a sufficiently large number of points
on the orographic surface,its spatial structure is preserved.
The sets of vertices (or nodes) V = {v1 , v2 , ..., vn } and edges (or arcs) E =
{e1 , e2 , ..., em } actually constitute a graph. If the weight of each arc is set equal
to the euclidean distance of the nodes it joins, then the shortest path between
 Optimization Model for Water Distribution Network Planning 547

two points on the surface will be approximated by the shortest path between
the corresponding nodes in the graph.
The degree of novelty of the present work lies in identifying minimum-length
paths that respect hydraulic limitations allowing them to constitute possible
layouts for gravity water pipelines. For example,paths joying source-user nodes
should preferably avoid the uphill sections in order to guarantee an adequate
hydraulic load along the entire route.

2 Methodology
2.1 Polyhedral Surface Construction
The presented methodology use a digital terrain elevation model in raster format
as a database. The raster format is a matrix structure that represents a rectan-
gular grid of pixels. This structure can be used to store topographic information
and is common-use format in GIS (Geographic Information System) software.
In a DTM (Digital Terrain Model) raster, the type of information stored is the
altimetry. Each pixel of the raster corresponds to the mean elevation of the area
covered by that pixel, to each of them will be associated a pair of coordinates of
a geodetic datum. The raster images used in this work consist of square pixels
of equal size. The topographic surface is modelled as a weighted oriented graph:

G = (V, E) (2)

with n nodes and m links with non-negative weight ij . The nodes are placed
in the center of the DTM cells. Each node will get a triplet of coordinates (x, y, z)
from the cells in which is placed. To identify the nodes, in addition to matrix
position indices of the corresponding cell, it can be useful to have single index
k identifying the k th node. For example, the index map k → (i(k), j(k)) can
be easily written out for the lexicographical ordering, according to which the
nodes are numbered by proceeding from left to right, from top to the bottom. In
the usual approach to approximate the geodesics on a surface with the minimal
paths on a graph, the weights ij of the arcs are set equal to the Euclidean
distances between the corresponding nodes i and j, and each arc can be travelled
indifferently in both directions
ij = ji (3)
for each corresponding arcs ∈ E. In this work, in order to take into account
the hydraulic constraints and penalize the uphill sections of the path, different
weights are assigned to uphill and downhill arches; more precisely, the weight of
the upward arc is increased:
ij > ji (4)
if zi > zj , where z is the altitude of the nodes. Both a directed and an undirected
graph can be used to model a topographic surface. Since it will be necessary
to take into account whether if an arc is rising or descending, a direct graph
is chosen. As regards the linking criterion, there are eight possible connections
548 M. Maiolo et al.

along the directions: North, South, East, West, North-West, South-West, North-
East, South-East (Fig. 2). These connections can be automatically generated
whether the position of the cell in the matrix (i, j) or the node index k is
used. The weight of a link is placed equal to the Euclidean distance between
cell centers. For a DTM with square cells it will depend only on the type of
connection (straight or diagonal) and on the elevation difference between start-
ing and ending node. The distance between two adjacent nodes is:

dxyz = dx2 + dy 2 + dz 2 (5)

For a DTM with square cells :

dx = dy = Δl (6)

Hence, if dxy indicates the horizontal distance between nodes, Eq. (5) can be
rewritten as: 
dxyz = dxy 2 + dz 2 (7)
√
where dxy = Δl for N , S, E, W links, and dxy = 2Δl for N W , SE, N E,
SW links. A dummy length is added in order to differently weight uphill and
downhill links: 
dxyz = dxy 2 + dz 2 + Pen · dz (8)

Fig. 1. Digital terrain model in raster format.

where P en ≥ 0 represents the penalty for the ascending traits of the path: for
uphill links, the penalty is positive, for others is null. If P en is equal to zero,
 Optimization Model for Water Distribution Network Planning 549

Fig. 2. Link generation method

the procedure gives the simple geodesic path. In order to reduce the extension of
the graph and reduce the calculus time, all nodes with a higher elevation than
the source node are excluded from the calculation. It is possible to exclude these
nodes because the shortest paths that will represent the route of the supply
pipes have to work only by gravity. Once the constrained geodesic path has been
identified, it will be necessary to verify that this is usable as a path of a gravity
supply pipe. Hence, a hydraulic check is implemented that takes into account
the slope and the burial of the pipeline.

2.2 Hydraulic Check

To find out if the elevation trace of supplying pipe can work, it is necessary to
solve the hydraulic problem related to it. The novelty of this work is the insertion
of a procedure that influences the automatic choice of the curves that represent
the supply paths taking into account their purpose. The motion formulas related
to the problem, use the geometrical and hydraulic characteristics of the pipe. If
the piezo metric line is always higher than the topographic surface, the pipeline
can work only with gravity. The hydraulic check of a geodesic takes place imme-
diately after finding the curve. In this phase, there is no information about the
flows that will pass through the pipes and neither on diameters and roughness.
A complete hydraulic check is impossible. The check that will be carried out
imposes a limit slope, which must be ensured. In the preliminary design phases
of supply pipes the engineering practice uses a constant slope for the piezometric
line. This is done to consider a constant headloss along the pipe. The constant
slope line will follow the geodesic curvilinear abscissa. If the line intersects the
topographic surface then the verification is not satisfied. Given a tridimensional
curve defined by a set of point triplets: (xi , yi , zi ), then the line will follow the
planar trace of the line defined by (xi , yi )
550 M. Maiolo et al.

Fig. 3. Hydraulic check decision algorithm with an increasing-penalty geodesic curve

set.

i 
 2 2
si = (xj−1 − xj ) + (yj−1 − yj ) (9)
j=2

Given
E : possible excavation depth
slopetg : target slope
 Optimization Model for Water Distribution Network Planning 551

the line will start from (x1 , y1 , z1 ) with the following equation:
⎧
⎨ xi
yi (10)
⎩ tg
zi = z1 + si · slopetg + E

It is worth to note that the hydraulic check is a process that can be carried out
only after the constrained geodesic is identified. It would be necessary an iterative
calculation methodology that increases the penalty until the hydraulic verifica-
tion is satisfied. This algorithm is too expensive in computational resources, as
it requires the calculus of the Eq. (1), for each links of the graph. In this work 5
values of increasing penalities were considered:

Pen = {0, 5, 10, 50, 100}; (11)

and the hydraulic control is carried out starting from a simple geodesic (P en = 0)
and then increasing the penalty until the corresponding path does not respect
the imposed hydraulic limitation (Fig. 3).

3 Application
The constrained geodesics have been searched for the DTM showed in Fig. 1. An
altimetry model of an area in the city of Rende (Italy, Calabria) was used, in
which possible source points and demand points were identified. 5 source nodes
and 4 demand nodes have been placed on this area. The source node (S5) is on a
high ground, the demand node on a flat area. On the straight path that connects

Fig. 4. DTM of the city of Rende (Italy), position of the source and demande nodes
552 M. Maiolo et al.

Fig. 5. set of 5 geodesic curves with increasing penalties found from a source point to
destination point. (a) Plan view. (b) Tridimensional view
 Optimization Model for Water Distribution Network Planning 553

Fig. 6. Section view of the 5 geodesic curves with the hydraulic check line

points, there is another high ground lower than the first one. For each of the
considered 5 penalties corresponding minimum-path are obtained (Fig. 7). For
null penalties, the output is a simple geodesic, which in plan view is an almost
straight line. As the penalty increases, the resulting constrained geodesic tries to
avoid the uphill areas by moving away from the geodesic path, making it longer.
As long as the starting point is at a height higher than the end point, there is
a constrained geodesic connecting them. The hydraulic control allows choosing
a suitable curve for the path of a supply pipe. Setting the possible excavation
depth and a target slope:
E = 2m
(12)
slopetg = 0.025
The simple geodesic and the lowest penalty does not pass the check (see
Fig. 6). The results for the entire set of user nodes and source nodes are shown
below. The information obtained is the length of the path that connects the
couples of points. The length values of the constrained geodesics are compared
with the Euclidean distances. It is immediate to notice some things: For the
Euclidean distance, the only condition for the existence of a connection is that
the source node is at a higher altitude than the destination one. In constrained
geodesics, this condition is necessary but not sufficient. This reduces the number
of possible connections. The use of (constrained or simple) geodesics implies an
554 M. Maiolo et al.

Fig. 7. Tridimensional view of a hydraulic check examples. (a) Geodesic curve that does
not pass the check, highlighted the point where the check line intersects the surface.
(b) Constrained geodesic that pass the check.

increase in length. The greater is the penalty used, the greater is the length of the
optimal curve. The application of the hydraulic check, in addition to leading to
the choice of routes at higher penalty, therefore longer, introduces the possibility
that a path between a source node with a higher altitude and a destination with
a lower altitude may not exist. By removing a certain number of links from the
set of usable links in the optimization models, the final solution, in some cases,
could differ much from that obtainable with methods that estimate the distance
without take into account the hydraulics aspects. In the application carried out,
 Optimization Model for Water Distribution Network Planning 555

as shown in Table 1, the use of a very high target slope led to the elimination of
7 links, and to the increase in the distance of one of them (Figs. 4 and 5).

Table 1. Table showing the connection matrix containing the distance between source
and users nodes. The Table matrix assess the increase in distance due to the use of
geodesic curves compared to the Euclidean ones.

Connection matrix: Constrained Geodesic length [m] Connection matrix: Euclidean line [m]
U1 U2 U3 U4 U1 U2 U3 U4
S1 4351.6 1386.6 3620.7 5562.1 S1 4292.9 1279.1 3446.0 5241.1
S2 S2 4211.3 5338.0 5401.3
S3 S3 1386.7 2497.2
S4 1035.5 S4 1351.9 977.7
S5 2510.4 4316.1 3682.2 S5 2425.3 2527.7 3904.4 3612.2
Geodesic penalty information Length increase using constrained geodesics
U1 U2 U3 U4 U1 U2 U3 U4
S1 Pen 0 Pen 0 Pen 0 Pen 0 S1 1.37% 8.40% 5.07% 6.13%
S2 HC TP HC HC S2
S3 HC TP TP HC S3
S4 HC TP TP Pen 0 S4 5.91%
S5 Pen 0 HC Pen 0 Pen 0 S5 3.51% 10.54% 1.94%
HC: Hydraulic check not passed.
TP: Impossible path because the source node is lower than the demand node.

4 Conclusion
The present work describes a methodology aimed at identifying the minimum
paths that follows the topological surface to connect a source node with a des-
tination node. The novelty of the work consists in the use of an approach that
provides distances and paths in a more representative way than reality with
respect to Euclidean distances and geodetic curves, and also allows to take into
account hydraulic constraints in the identification of the paths. The use of this
methodology in the context of Water Distribution Systems optimization models
allows to obtain more detailed and realistic solutions for the subsequent design
phase. The application carried out to a real surface model has allowed to make a
comparison between the Euclidean distances and the lengths of the constrained
paths obtained with the proposed approach, showing how the distances that take
into account the topographic surface are longer of the Euclidean ones. The use
of Euclidean distances therefore leads to underestimating quantities of materials
and excavations necessary for the realization of the optimization configuration
identified by the model.

Acknowledgments. Research supported by the Italian Regional Project (POR

CALABRIA FESR 2014-2020): origAMI original Advanced Metering Infrastructure
[J48C17000170006].
556 M. Maiolo et al.

References
1. Balasubramanian, M., Polimeni, J.R., Schwartz, E.L.: Exact geodesics and shortest
paths on polyhedral surfaces. IEEE Trans. Pattern Anal. Mach. Intell. 31(6), 1006–
1016 (2008). https://doi.org/10.1109/TPAMI.2008.213
2. Carini, M., Maiolo, M., Pantusa, D., Chiaravalloti, F., Capano, G.: Modelling and
optimization of least-cost water distribution networks with multiple supply sources
and users. Ricerche Mat. 67(2), 465–479 (2018)
3. Coelho, A.C., Labadie, J.W., Fontane, D.G.: Multicriteria decision support sys-
tem for regionalization of integrated water resources management. Water Resour.
Manag. 26(5), 1325–1346 (2012). https://doi.org/10.1007/s11269-011-9961-4
4. Davijani, M.H., Banihabib, M.E., Anvar, A.N., Hashemi, S.R.: Multi-objective
optimization model for the allocation of water resources in arid regions based on
the maximization of socioeconomic efficiency. Water Resour. Manag. 30(3), 927–
946 (2016). https://doi.org/10.1007/s11269-015-1200-y
5. Dijkstra, E.W.: A note on two problems in connexion with graphs. Numer. Math.
1(1), 269–271 (1959)
6. Hsu, N.S., Cheng, K.W.: Network flow optimization model for basin-scale water
supply planning. J. Water Resour. Plan. Manag. 128(2), 102–112 (2002). https://
doi.org/10.1061/(ASCE)0733-9496(2002)128:2(102)
7. Kanai, T., Suzuki, H.: Approximate shortest path on a polyhedral surface based
on selective refinement of the discrete graph and its applications. In: Proceedings
Geometric Modeling and Processing 2000. Theory and Applications, pp. 241–250.
IEEE (2000). https://doi.org/10.1109/GMAP.2000.838256
8. Maiolo, M., Pantusa, D.: An optimization procedure for the sustainable manage-
ment of water resources. Water Supply 16(1), 61–69 (2015). https://doi.org/10.
2166/ws.2015.114
9. Porazilova, A.: The Geodesic Shortest Path (2007)
10. Surazhsky, V., Surazhsky, T., Kirsanov, D., Gortler, S.J., Hoppe, H.: Fast exact and
approximate geodesics on meshes. ACM Trans. Graphics (TOG) 24(3), 553–560
(2005). https://doi.org/10.1145/1073204.1073228
11. Zhanping W., Juncang T.: Optimal allocation of regional water resources based
on genetic algorithms. J. Converg. Inf. Technol. (JCIT) 7(13) (2012). https://doi.
org/10.4156/jcit.vol7.issue13.51
 Scenario Optimization of Complex Water
Supply Systems for Energy Saving
and Drought-Risk Management

Jacopo Napolitano(&) and Giovanni M. Sechi

University of Cagliari, Via Marengo 2, Cagliari, Italy

{jacopo.napolitano,sechi}@unica.it

Abstract. The management of complex water supply systems needs a close

attention to economic aspects concerning high costs related to energy require-
ments in water transfers. Specifically, the optimization of activation schedules of
water pumping plants is an important issue, especially managing emergency and
costly water transfers under drought-risk. In such optimization context under
uncertainty conditions, it is crucial to assure simultaneously energy savings and
water shortage risk alleviating measures. The model formulation needs to
highlight these requirements duality to guarantee an adequate water demand
fulfillment respecting an energy saving policy. The proposed modeling approach
has been developed using a two stages scenario optimization in order to consider
a cost-risk balance, and to achieve simultaneously energy and operative costs
minimization assuring an adequate water demand fulfillment for users. The
optimization algorithm has been implemented using GAMS interfaced with
CPLEX solvers. An application of the proposed optimization approach has been
tested considering a water supply system located in a drought-prone area in
North-West Sardinia (Italy). By applying this optimization procedure, a robust
strategy in pumping activation was obtained for this real case water supply
system.

Keywords: Scenario analysis
Energy optimization
Water management

1 Introduction

The management optimization of complex water supply systems, aimed to the energy
saving, is an interesting and actual research topic [2, 10, 16]. Problems pertaining to
water system management policies and specifically concerning the effectiveness of
emergency and costly water transfers activation to alleviate droughts, are faced with
different methodological approaches [8, 9, 12]. Solving these optimization problems
frequently leads to complex computational models: their solution needs efficient
approaches to deal with many uncertainties which arise modeling real systems and
trying to achieve optimal decision rules, in order to provide robust solutions to the
water resource system’s Authorities [17].
In the water resource modeling field, problems affected by uncertainty have been
treated implementing several computational solutions, especially with application of

© Springer Nature Switzerland AG 2020

Y. D. Sergeyev and D. E. Kvasov (Eds.): NUMTA 2019, LNCS 11973, pp. 557–569, 2020.
https://doi.org/10.1007/978-3-030-39081-5_48
558 J. Napolitano and G. M. Sechi

stochastic dynamic programming to multi-reservoir systems. Under scarcity conditions,

the system reliability evaluation is intimately related to a quantitative evaluation of
future water availability and the opportunity to provide water through the activation of
emergency and costly water transfers. Hence, the water system optimization problem
needs to deal with uncertainties particularly in treating the effectiveness of emergency
measures activation to face droughts.
On the occasion of drought occurrences, water managers must be able to manage
this criticality, alleviating the effect of shortages on water users. A useful solution could
be to define some emergency policies, supplying additional water to demand centers.
Modern water supply systems could count on the presence of emergency sources with
enough capacity for these critical occurrences. Frequently, the activation of these
emergency transfers requires additional costs, such as for activation of pumping
schedules. Optimal strategies on this costly transfer activation is a hard decision
problem: it is conditioned by uncertainties, mainly due to future demands behavior and
hydrological inputs that are normally characterized by high variability during the time
horizon normally considered in modeling the systems.
According some authors [9, 11, 14], the uncertainty could be described through
different scenarios, which occurrence follows a probability distribution. In case of
previous pessimistic forecasts, uncertainty and variability can generate water excess or
subsequent spilling from water storage reservoirs, causing losses and therefore
resulting as regrets costs [6]. On the other hand, the definition of emergency policies in
reservoirs management could consider early warning measures, taking advantages of
lower energy prices in some time-periods and achieving economic savings also related
to avoid the occurrence of water shortages.
The hereafter described research aims to develop an optimization under uncertainty
modeling approach, in order to deal with water resources management problems
especially referring to the definition of optimal activation rules for emergency acti-
vation of pumping stations in drought conditions. Therefore, this study aims to define a
cost-risk trade-off considering the minimization of water shortage damages and the
pumping operative costs, under different hydrological and demand scenarios occur-
rence possibilities. The expected results should be able to provide the water system’s
authority with a strategic information, defining optimal rules, and specifically optimal
activation reservoir-storage triggers for water pumping stations.
The formulation of the related optimization model needs to highlight this duality: to
guarantee water demands fulfillment respecting an energy saving policy.
Optimal rules in order to minimize the emergency transfer’s costs will be defined
through a cost/risk balanced management and formulating a multistage scenario opti-
mization model. The proposed modeling approach has applied to a real case concerning
a multi-reservoir and multi-user water supply system in a drought-prone area, located in
the North-West Sardinia (Italy) region, characterized by South-Mediterranean climate.
 Scenario Optimization of Complex Water Supply Systems 559

2 Scenario Optimization for a Water Supply System

Management

Scenario analysis approach considers that all future events can be described through a
set of different and statistically independent scenarios [14]. A single scenario describes
a possible sequence in the realization of some sets of uncertain data along the analyzed
time horizon. Considering all together the structure of the scenarios temporal evolution,
it is possible to obtain a robust decision policy, minimizing the risk of wrong future
decisions.
This modelling approach can be represented as a tree-graph (Fig. 1), according to
appropriate aggregation rules, called congruity constraints. Some scenarios sharing a
common initial portion of data can be considered partially aggregated with the same
decision variables for the aggregated part, taking into account possible evolutions in the
subsequence of different behaviors.

Fig. 1. Scenario-tree aggregation

In order to perform correctly the scenario aggregation, must be defined stages and
branching-times as represented in Fig. 1. A branching-time identifies the time-period in
which some scenarios, that are identical up to that period, begin to be different.
The root of the scenario-tree corresponds to the time at which decisions (common
to all scenarios) have been taken, while the leaves represent the state of the system in
the last stage. Each path from the root to a leaf identifies a possible sequence of
occurrences along the entire time horizon.
Therefore, each possible scenario corresponds with a dynamic multi-period graph
[11], following a particular sequence of decisions. E.g., in the water resource man-
agement, it is possible to consider a hydrological series of water inflows or a sequence
of management decisions related to a reservoir [16].
560 J. Napolitano and G. M. Sechi

2.1 Trigger Values Optimization Model and Cost-Risk Trade Off

Dealing with uncertainty, in order to find a balanced water system management cri-
terion between alternatives represented through the scenario tree and to improve the
decision-making process robustness, a cost-risk balancing optimization approach has
been used.
This model formulation tries to attain to a robust decision policy minimizing the risk
to assume wrong and harmful decisions in the system management. Therefore, the water
system’s authority should be able to define a reliable decision system, minimizing the
energetic and management system costs and reducing the deficit hardships for users.
Considering the entire set g 2 G of scenarios, the objective function (1) minimizes
the energy and management costs, optimizing the configuration of the flows xg along
the network. Moreover, it minimizes the weighted distance between a single scenario
activation threshold b
S g and the optimal barycentric value Sb .
X X X   2 
Minimizeð1  kÞ p c x þk
g g g g
p w g b
S gi  Sbi ð1Þ
xg ;b
g2G g2G i2P
S g ;Sb

subject to

Ag xg ¼ bg 8g 2 G ð2Þ

lg  xg  ug 8g 2 G ð3Þ

x 2 U ð4Þ

All decision variables and data are scenario dependent, hence the index g. A weight
pg can be assigned to each scenario characterizing its relative importance. Weights
could represent the probability of occurrence of each scenario.
The vector cg describes the unit cost of different activities like delivery cost,
opportunity cost related to unsatisfied demand, opportunity cost of spilled water,
energy cost and so on. The set of standardized equality constraints Ag describes the
relationships between storage, usage, spill, and exchange of water at different nodes
and in subsequent time periods. The RHS values bg are given from scenario occur-
rences and are related to data of inflows and demands. The lower and upper bounds lg
and ug are defined by structural and policy constraints on operating the system. All
constraints (2–4) are collected from all scenarios and must be considered in the
aggregated model. The additional set of constraints (4) are called non-anticipative
constraints and x 2 U represents the congruity constraints derived by the scenario
aggregation rules [11].
In general terms, the first part of the objective function (1) can be defined as a cost
function and it tries to look for the system flows configuration that allows minimizing
the costs supported during the water system management. The second part can be
considered as a risk function, it is quadratic and it tries to minimize the quadratic
weighted distance between the barycentric value and each single scenario trigger value.
The weight wg is the cost related to the risk occurrences of each scenario g 2 G.
 Scenario Optimization of Complex Water Supply Systems 561

In this way, giving a weighted value to both terms of the objective function, we can
find a solution of the cost-risk balancing problem.
The relationship between cost function and risk function is regulated by the
parameter k called weight factor, which can vary between 0 and 1. Intermediate values
of k provide different tradeoffs between costs and risks.
In order to guarantee a correct operation of pumping stations, the model (1–4)
should be completed introducing a new set of constraints (5–9).
XRi
hgi BM  j¼1
xvgj  b
S gi  ð1  hgi ÞBM 8i 2 P; 8g 2 G ð5Þ

xgi ¼ ð1  hgi ÞPi 8i 2 P; 8g 2 G ð6Þ

xgi ¼ ð1  hgi ÞTk 8k 2 K; 8i 2 P; 8g 2 G ð7Þ

XRi
b
S gi \ Kj 8i 2 P; 8g 2 G ð8Þ
j¼1

XRi
Sbi \ j¼1
Kj 8i 2 P; 8g 2 G ð9Þ

Where:
i 2 Pf1; . . .; nP g Pumping stations in the system;
k 2 K f1; . . .; nK g Diversion dams in the system;
j 2 Rf1; . . .; nR g Reservoirs in the system;
hgi 2 f0; 1g Binary variable
The activation of pump stations is supposed to be dependent on the stored volume
levels in reservoirs that could supply the downstream demand nodes by gravity or,
anyway, without recurring to emergency and costly water transfer activation. There-
fore, to model the pump activation, a binary variable hgi to each i-th pump station
should be assigned. This variable represents the on/off condition for a single pump
station as it can assume one or zero value.
In the optimization model hgi is dependent on the sum of the stored levels xvgj in the
j-th reservoirs supplying water by gravity, according the activation dependences shown
for the real case in Table 5. Therefore, constraints (5) allows the i-th pump station
activation if the sum of the stored volume in reservoir j is under the threshold value b S gi .
In this constraint, the parameter BM is a large scalar.
Constraint (6) guarantees that, in the case of activation of the i-th pump station, the
flow along the pumping arc starting from the i-th station will be equal to its capacity P.
If the pump station i-th is located downstream to a transshipment node k, the
constraint (7) assures that, in case of activation, the flow along the arc will be equal to
the potential water volumes withdrawal from this node k 2 K [15], which should be
lower to the pump capacity P.
Constraints (8–9) impose an upper bound on the activation storage levels of the i-th
pumping station equal to the sum of the reference reservoir’s capacity Kj .
The adopted optimization procedure, defining trigger values in the pumps activa-
tion, is summarized in the flowchart shown in Fig. 2.
562 J. Napolitano and G. M. Sechi

Fig. 2. Scenario-optimization: main modelling steps

In a first step of the analysis, it is possible to work using a single scenario opti-
mization in order to calibrate the main parameters of the process (spilling costs, storage
benefit and so on). In a second phase, the model examines the set of different scenarios.
Once evaluated optimal values using scenario analysis, a re-optimization phase could
be performed in order to verify the reliability of the water system and to obtain the
network’s flows in single scenario configurations. Through this last phase is possible to
reach the sensitivity analysis and to verify output caused by assumptions about the
adopted parameters.

3 Case Study: North-West Sardinia Water Supply System

Considering a water management problem, the graph representation approach [3] has
been considered as an efficient support for the mathematical modeling [1, 5]. According
to this, J ¼ ðN; LÞ are sets satisfying L  ½N 2 , where the elements of N are the nodes of
the graph J, while the elements of L are its arcs. In the common water system notation,
nodes can represent groundwater, sources, reservoirs, demands, etc. Arcs represent the
connections between nodes, where water could flow.
This approach allows drafting a complex water system problem through a simple
flow network on a graph. In the single period, we can represent the physical system and
the static situation by a direct network called basic graph, as reported in Fig. 3.
This kind of analysis could be extended to a wide time-horizon T, assuming a time
step (month) t. By replicating the basic graph for each period of the time horizon, it is
possible to generate a dynamic multi-period network [11].

3.1 Water System’s Features

An application of the proposed optimization procedure has been tested considering the
draft of a real water supply system located in a drought-prone area in North-West
Sardinia (Italy).
 Scenario Optimization of Complex Water Supply Systems 563

Fig. 3. North-West Sardinia water supply system

The main water sources are provided by 2 artificial reservoirs and 2 diversion dams.
Their main features are reported in Tables 1 and 2.
The evaluation of potentiality in the water supply arises from historically hydro-
logical inflows evaluated from 1922 to 1992 considering the values reported in the
Sardinia Region Water Plan [13].

Table 1. Reservoir’s features

Code Reservoir Capacity[106 m3] Spilling cost [€/m3]
R1 Temo Rocca Doria 70 0.001
R2 Bidighinzu 10.9 0.001

Reservoirs are represented in the sketch by system’s storage nodes. Diversion dams
do not have a large storage volume: therefore, only a partial incoming flow can be
diverted to demand centers or to larger capacity reservoirs.

Table 2. Diversion dams

Code Transshipment Spilling cost [€/m3]
T1 Cumone - Badu Crabolu 0.001
T2 Ponte Valenti - Calambru 0.001

Water demands have been grouped in four centers, according to two different users:
civil and irrigation. As shown in Table 3, an annual volume of demand, rate and deficit
564 J. Napolitano and G. M. Sechi

costs are associated to each demand center. Deficit costs quantify the possible damages
supported by the users in the case of shortage occurrences. These costs have been
evaluated starting from the water annual rates for unit of volume applied by each
stakeholder.

Table 3. Demand centers

Code Demand center Demand Rate cost Planned Unplanned
[106 m3/year] [€/m3] deficit cost deficit cost
[€/m3] [€/m3]
D1 Irrigation Cuga 15.36 0.006 0.06 0.6
D2 Civil Temo 6.93 0.025 0.25 2.5
D3 Irrigation Valle 0.95 0.006 0.06 0.6
Giunchi
D4 Civil 15.56 0.025 0.25 2.5
Bidighinzu

Deficits can be categorized in two classes: planned and unplanned. Planned deficits
can be forecasted and communicated in advance to water users, while the unplanned
ones arise when the realization of water inflows follow hydrological scenarios of
unpredictable scarcity, affecting and harming several users. In order to take into
account these costs in the mathematical model, in case of shortage occurrences, the
initial 15% of the monthly demand not satisfied will be compute as planned deficit
while remaining surplus of shortages should be considered as unplanned deficits.

Table 4. Pump stations’ features

Code Pumping station Capacity [106 m3/month] Pumping cost [€/m3]
P1 Padria 7.88 0.071
P2 Rocca Doria 1.66 0.196
P3 Su Tulis 3.15 0.087

Pump stations reported in Table 4 allow demand centers to be supplied with an

increased economic burden, namely incurring energy costs in addition to the ordinary
management.

Table 5. Pump stations’ activation dependences

Reservoir ! R1 R2
Pump station #
P1 1 0
P2 0 1
P3 0 1
 Scenario Optimization of Complex Water Supply Systems 565

Modeling the system, threshold levels for pumps activation refer to the stored
volume in reservoirs that supply the downstream demand nodes. These functional
dependencies are reported in the Table 5, where 1 means dependence, while 0 inde-
pendence for pump activation.

3.2 Optimization of the Pumping Schedules Thresholds

For each pump station, two season’s activation thresholds should be identified, in order
to define a barycentric seasonal value. These trigger values refer specifically to the dry
(in Mediterranean countries months from April to September) and wet (months from
October to March) hydrological semesters.
By applying the scenario optimization process, a robust decision strategy for sea-
sonal trigger values of pumps activation was retrieved. Reference scenarios were settled
considering the regional hydrological database [13] modified in order to take into
account climatological trends observed during the last decades in this region. Starting
from the river-runoff historical database of 50 years horizon, scenarios of different
criticism are extracted from the observed data: each one has been of 20 years, defined at
monthly step. Inside the 240 monthly periods the branching time is located at the 120th
period. As shown in Fig. 1, the resulting scenario-tree has been composed. Therefore,
the historical database has been used in order to generate 4 hydrological scenarios
characterized by the same length, organized with a common root of 10 years and the
following data diversified in the following 10 years by climate criticism.
The scenario analysis approach needs a complex mathematical formulation in terms
of model dimensions and number of variables and constraints considered. Moreover,
the mathematical model described in the Eqs. (1–4) should be solved through a Mixed
Integer Quadratically Constraints Programming. Therefore, in order to afford effica-
ciously this problem, guaranteeing a reliable solution with a reasonable computational
time, it has been implemented by the software GAMS [4] calling CPLEX [7] Branch
and Cut algorithm as optimization solver.
The optimized activation thresholds reported in Table 6 have been evaluated
considering this scenario-tree and solving the optimization model (1–4) using a k value
equal to 0.5, then assuring an equal weight to cost and risk elements in the objective
function.

Table 6. Optimized activation thresholds

Activation threshold [106 m3] S1 S2 S3
Wet Dry Wet Dry Wet Dry
60.22 67.13 1.04 3.26 7.64 8.9

According the scenario analysis, these optimized values are barycentric among all
hydrological scenarios, and they will guarantee a compromise among different water
resource availabilities. These barycentric values have been adopted during the re-
optimization phase, where the whole process has been developed assigning these
thresholds as fixed parameters.
566 J. Napolitano and G. M. Sechi

Table 7 reports the average annual costs evaluated by an economic post-processor

taking into account deficit and pumping costs. Only the real costs of the system, related
to energy consumption and drought occurrences have been considered during this
analysis.

Table 7. Economical post-processor

Mean costs 106 €/year
Pumping 1.82
Deficit 0.19
Total 2.01

Results show that almost the total cost amount is due to the energy costs, while only
the 10% is caused by deficit occurrences. It highlights that the cost-risk balancing
approach has been able to assure simultaneously energy costs minimization and a
reduction of possible damages caused by water shortages.
Figures 4, 5 and 6 show the percentage of water deficit along the considered four
scenarios for the affected demand centers. As expected, deficits affect especially irri-
gation users in order to preserve the civil demands, which represent the water system’s
major priority users.
This representation highlights planned and unplanned deficit occurrences, indeed:
planned deficit are included under the 15%, while unplanned deficits across this
threshold.

Deficit - D1: Irrigation Cuga

g1 g2 g3 g4 15%

100%

80%
Deficit [%]

60%

40%

20%

0%

Months

Fig. 4. Water deficit at demand node D1 – Irrigation Cuga

 Scenario Optimization of Complex Water Supply Systems 567

Deficit - D3: Irrigation Valle dei Giunchi

g1 g2 g3 g4 15%

100%

80%
Deficit [%]

60%

40%

20%

0%

Months

Fig. 5. Water deficit at demand node D3 – Irrigation Valle dei Giunchi

Deficit - D4: Civil Bidighinzu

g1 g2 g3 g4 15%

100%

80%
Deficit [%]

60%

40%

20%

0%

Months

Fig. 6. Water deficit at demand node D4 – Civil Bidighinzu

568 J. Napolitano and G. M. Sechi

4 Conclusions

The scenario-optimization approach confirmed its potentiality when applied to the real
case of a water resource management problem and specifically optimizing the activa-
tion schedules of water pumping plants and managing emergency and costly water
transfers under drought-risk. It allowed identifying two barycentric seasonal optimal
activation thresholds for each pumping plant located in the water system.
These results are obtained using a two stage stochastic programming taking into
account expected water demand and hydrological series.
This optimization approach was implemented using GAMS that could be consid-
ered as an excellent support during the model development. The software allowed
writing easily the optimization models and interfacing with CPLEX solvers.
Considering the real case application, the cost-risk balancing approach minimized
the operational and management costs and contextually restricted risks and conflicts
between users in shortage conditions. Costs and penalties have been evaluated in a re-
optimization phase through an economic post-processor taking into account water
shortage penalties and pumping costs and assuring a trade-off between cost and risk
elements. For the considered water scheme, the proposed methodologies guarantees
almost the complete fulfilment of the water demand: unplanned deficits still remain but
only for few periods of the considered time horizon.
Using a more adherent to the reality simulation approach, the effectiveness of the
obtained results has been tested interacting with the regional water system’s Authority
by comparison with the occurred management behavior.

References
1. Ahuja, R., Magnanti, T., Orlin, J.: Network Flows: Theory, Algorithms, and Applications.
Prentice Hall, Englewood Cliffs (1993)
2. D’Ambrosio, C., Lodi, A., Wiese, S., Bragalli, C.: Mathematical programming techniques in
water network optimization. Eur. J. Oper. Res. 243(3), 774–788 (2015)
3. Diestel, R.: Graph Theory, 3rd edn. Springer, New York (2005)
4. GAMS: A user’s guide. GAMS Development Corporation. Washington DC, USA (2008)
5. Jensen, P., Barnes, J.: Network Flow Programming. Wiley, New York (1980)
6. Kang, D., Lansey, K.: Multiperiod planning of water supply infrastructure based on scenario
analysis. ASCE J. Water Resour. Plann. Manag. 140, 40–54 (2014)
7. IBM: Cplex Optimization Studio (2017). http://www-03.ibm.com/software
8. Lerma, N., Paredes-Arquiola, J., Andreu, J., Solera, A., Sechi, G.M.: Assessment of
evolutionary algorithms for optimal operating rules design in real water resource systems.
Environ. Model Softw. 69, 425–436 (2015)
9. Napolitano, J., Sechi, G.M., Zuddas, P.: Scenario optimization of pumping schedules in
complex water supply system considering a cost-risk balancing approach. Water Resour.
Manag. 30, 5231–5246 (2016)
10. Nault, J., Papa, F.: Lifecycle assessment of a water distribution system pump. ASCE J. Water
Resour. Plann. Manag. 141(12), A4015–004 (2015)
11. Pallottino, S., Sechi, G.M., Zuddas, P.: A DSS for water resource management under
uncertainty by scenario analysis. Water Resour. Manag. 28(12), 3975–3987 (2014)
 Scenario Optimization of Complex Water Supply Systems 569

12. Pasha, M.F.K., Lansey, K.: Strategies to develop warm solutions for real-time pump
scheduling for water distribution systems. Environ. Model Softw. 20, 1031–1042 (2004)
13. RAS: Piano stralcio di bacino regionale per l’utilizzazione delle risorse idriche. Regione
autonoma della Sardegna, Italy (2006)
14. Rockafellar, R.T., Wets, R.J.B.: Scenario and policy aggregation in optimization under
uncertainty. Math. Oper. Res. 16, 119–147 (1991)
15. Sassu, E., Zucca, R., Sechi, G.M.: Calibration of regional flow-duration curves evaluating
water resource withdrawal from diversion dam. In: Garrote, L., Tsakiris, G., Tsihrintzis, V.
A., Vangelis, H., Tigkas, D. (eds.) Managing Water Resources for a Sustainable Future,
Proceedings of the 11th World Congress of EWRA on Water Resource and Environment,
Madrid, 25–29 June 2019 (2019)
16. Sechi, G.M., Gaivoronski, A.A., Napolitano, J.: Optimizing pumping activation in multi-
reservoir water supply systems under uncertainty with stochastic quasi-gradient methods.
Water Resour. Manag. 33(2), 1881–1895 (2019)
17. Sechi, G.M., Sulis, A.: Water system management through a mixed optimization-simulation
approach. ASCE J. Water Resour. Plann. Manag. 135, 160–170 (2009)
 Optimizing Rainwater Harvesting Systems
for Non-potable Water Uses and Surface
Runoff Mitigation

Stefania Anna Palermo1(&) , Vito Cataldo Talarico1 ,

and Behrouz Pirouz2
1
Department of Civil Engineering, University of Calabria, Rende, CS, Italy
{stefania.palermo,vitocataldo.talarico}@unical.it
2
Department of Mechanical, Energy and Management Engineering,
University of Calabria, Rende, CS, Italy
[email protected]

Abstract. Rainwater harvesting systems represent sustainable solutions that

meet the challenges of water saving and surface runoff mitigation. The collected
rainwater can be re-used for several purposes such as irrigation of green roofs and
garden, flushing toilets, etc. Optimizing the water usage in each such use is a
significant goal. To achieve this goal, we have considered TOPSIS (Technique for
Order Preference by Similarity to Ideal Solution) and Rough Set method as Multi-
Objective Optimization approaches by analyzing different case studies. TOPSIS
was used to compare algorithms and evaluate the performance of alternatives,
while Rough Set method was applied as a machine learning method to optimize
rainwater-harvesting systems. Results by Rough Set method provided a baseline
for decision-making and the minimal decision algorithm were obtained as six
rules. In addition, The TOPSIS method ranked all case studies, and because we
used several correlated attributes, the findings are more accurate from other simple
ranking method. Therefore, the numerical optimization of rainwater harvesting
systems will improve the knowledge from previous studies in the field, and pro-
vide an additional tool to identify the optimal rainwater reuse in order to save water
and reduce the surface runoff discharged into the sewer system.

Keywords: Rainwater harvesting
Water supply
Flood risk mitigation

1 Introduction

There are many benefits in Rainwater harvesting (RWH) systems mainly water saving
for non-potable water uses and surface runoff mitigation. Moreover, the collected
rainwater can be re-used for several purposes including green roofs and garden,
flushing toilets, etc. In previous studies, the optimization of rainwater harvesting sys-
tems was mostly limited to optimum size of the tankers according to hydrological and
hydraulic analysis and in some cases combined with economic analysis.
However, the design of RWH systems depends on many elements and even opti-
mizing different water usages is significant. Therefore, in this paper, Multi-Objective
Optimization approaches have been applied, and ranking methods such as TOPSIS

© Springer Nature Switzerland AG 2020

Y. D. Sergeyev and D. E. Kvasov (Eds.): NUMTA 2019, LNCS 11973, pp. 570–582, 2020.
https://doi.org/10.1007/978-3-030-39081-5_49
 Optimizing RWH Systems for Non-potable Water Uses 571

(Technique for Order Preference by Similarity to Ideal Solution) have been considered
to compare algorithms and evaluate the performance of alternatives to reach the ideal
solution. Moreover, the attributes analysis such as Rough Set method has been used in
analysis of vague description of decisions.

1.1 Rainwater Harvesting (RWH) Systems

The combined effect of global urbanization, climate change and water scarcity, requires
a transition towards a sustainable, smart and resilient urban water management. In this
regard, nowadays the sustainability concept is a basilar element for scientific, technical
and socio-economic discussion. Therefore, the implementation of decentralized
stormwater controls systems, as LID (Low Impact Development) systems, represents a
promising strategy to achieve several benefits at multiple scales [1, 2].
Among these techniques, Rainwater Harvesting (RWH), considered an ancient
practice used all over the world to meet the water demand, is now supported by many
countries as a suitable solution to limit potable water demand, reduce frequency, peaks
and volumes of stormwater runoff at the source, and participate in the restoration of
natural hydrological cycle [3–6].
The principal component of a conventional RWH system is the rainwater tank
which temporally stores the water from a capturing surface, normally the building roof
or others impervious surfaces closely to the building. In a single-family building,
above-ground tank, named “rain barrels”, are generally used for irrigation and runoff
control, while, in the case of multi-story building, above or below-ground concrete
cisterns are implemented. In addition, a system consisting of gutters and downspouts
lead the runoff from the collecting surface to the tank, while a dedicated piping network
is needed for rainwater reuse. One or more pumps can be used to assure the pressure
head for different usages, while other devices as first flus diverters, debris screen, and
filters are generally implemented for water quality control. These information and more
specific detail regarding this technique can be found in several studies [4, 7, 8].
Recent advances have showed the possibility for real-time monitoring and control
of these systems in order to increase their efficiency in terms of reduction of urban
flooding or combined sewer overflows [9] and optimize the rainwater reuse.
Harvested rainwater can be considered a renewable water source that is perfect for
different non-potable water uses, as toilet flushing, laundry, car washing, terrace
cleaning, private garden irrigation and green roof irrigation [7, 10–15].
In addition, as source control technology distributed at urban catchment scale, these
systems are suitable to reduce stormwater runoff volume. In this regard, several studies
have evaluated also the hydrological efficiency of RWH in terms of reduction of the
runoff volume and peak discharged to the sewer system [5, 7, 11, 14, 16].
Several studies have been carried out to show the RWH efficiency for water saving
and runoff mitigation, as study of Campisano and Modica [11] showed that the per-
formance depends of site-specific factors, such as roof type and surface, precipitation
regime, demand usage, tank size, number of people in the household, etc.
Based on a deeper literature review, in this paper some factors, here after called
“attributes” have been selected and considered in a mathematical optimization of
RWH.
572 S. A. Palermo et al.

2 Methodology

In the current study, the Rough Set method applied as a machine learning method to
optimize rainwater-harvesting systems. The process is reviewed in details and the result
is achieved with analysis of different case studies.

2.1 Rough Set Theory

The Rough Set theory is attributes analysis based on data or knowledge about a
decision. The results of analysis can provide clear rules for similar decisions [17–19].
The Rough Set rules in a given approximation space apr = (U, A) can be calculated as
follow:

The lower approximations of X in A apr ðAÞ ¼ fx j x 2 U; U=ind ðAÞ  Xg ð1Þ

The upper approximations of X in A apr ðAÞ ¼ fx j x 2 U; U=ind ðAÞ \ X 6¼ ug ð2Þ

U=ind ðaÞ ¼ fðxi ; xj Þ 2 U  U; f ðxi ; aÞ ¼ f ðxj ; aÞ; 8a 2 Ag ð3Þ

Boundary is BN ðAÞ ¼ aprðAÞ  aprðAÞ ð4Þ

The reduct ¼ minimal set of attributes B  A such that rB ðUÞ ¼ rA ðUÞ ð5Þ
P
card ðBðXi ÞÞ
The quality of approximation of U by B rB ðU Þ ¼ ð6Þ
card ðU Þ

Decision rule is u ) h ð7Þ

Where:
U is a set,
A is attributes of the set,
u is the conjunction of elementary conditions,
h is the disjunction of elementary decisions.

2.2 TOPSIS Method

TOPSIS (Technique for Order Preference by Similarity to Ideal Solution) method
developed by Hwang and Yoon in 1981 is a method to solve the ranking and com-
paring decisions [20–22] and can be applied to wide range of multi-attribute decision
makingwith
 several attributes [22–24]. The ranking in this method can be done for
matrix nij m  n according to the similarity to ideal solution as follow:
 Optimizing RWH Systems for Non-potable Water Uses 573

aij
Normalized decision matrix: N ¼ nij ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Pm ; i ¼ 1; 2; . . .; m ; j ¼ 1; 2; . . .; n ð8Þ
i ¼ 1aij
2

Weighted normalized decision matrix: V ¼ N  Wn  n ð9Þ

Determining the solutions:

     
Ideal solution: A þ ¼ maxi aij jj 2 J ; mini aij jj 2 J þ ð10Þ
     
Negative-ideal solutions: A ¼ mini aij jj 2 J ; maxi aij jj 2 J þ ð11Þ

J þ ¼ fj ¼ q; 2; . . .; njjg Associated with positive impact criteria

J ¼ fj ¼ q; 2; . . .; njjg Associated with negative impact criteria
Determining the distances from the solutions:
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Xn

Distance from ideal solution: diþ ¼ j¼1

v ij  v j
þ
; i ¼ 1; 2; . . .; m ð12Þ

rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Xn

Distance from negative-ideal solution: di ¼ j¼1

vij  vjþ ; i ¼ 1; 2; . . .; m ð13Þ

Ranking in order to the highest closeness to the negative-ideal condition:

di
CLi ¼ ; 0  CLi  1 & i ¼ 1; 2; . . .; m ð14Þ
di þ diþ

2.3 Case Studies

To carry the analysis by rough set and TOPSIS a set of data is required such as case
studies. The selection of the case studies has been done in a way that considers the main
possible attributes/factors confronting in rainwater harvesting systems. The first and
second case studies (CS1–CS2) are taken from the study carried out by Herrmann and
Schimida [7] that considers the development and performance of rainwater utilization
systems in Germany, and specifically in Bochum where the mean annual precipitation is
787 mm. More in detail, the data considered there for CS1, are related to a one-family
house (with an effective roof area of 150 m2, 4 persons, combined demand of 160 l/d, a
storage volume of 6 m3 and an additional retention volume of 15 m3, with a covering
efficiency of 98%). The CS2 case refers to a multi-story building (CS2) (with an
effective roof area of 320 m2, 24 persons, toilet flushing demand of 480 l/d, a storage
volume of 14 m3 and an additional retention volume of 35 m3).
A study carried out by Domènech and Saurí [13] was considered to select the third
(CS3) and the fourth (CS4) case study. Both case studies are in Sant Cugant del Vallès –
Spain. More in detail, CS3 is a single-family house with a rooftop catchment area of
574 S. A. Palermo et al.

107 m2, 3 residents, a toilet and laundry usage demand of 27 LCD and 16 LCD,
respectively. While CS4 refers to a multi-family building with a rooftop catchment area
of 625 m2, 42 residents, a toilet and laundry usage demand of 30 LCD and 16 LCD,
respectively. For CS3, we considered a model scenario in which a tank of 13 m3 can
meet 80% of the combined demand of toilet flushing and laundry; while for CS4 the
model scenario is a tank of 31 m3 covering 59.5% of the combined demand of toilet
flushing and laundry.
Case studies CS5, CS6 and CS7 are considered from the study of Palla et al. [5].
These three case studies are located in Genoa (Italy) with a mean annual precipitation of
1340 mm. More in detail, CS5 is a 4-flat house with 16 inhabitants, a roof area of
420 m2, an annual toilet flushing demand of 233.6 m3/y and a tank capacity of 14 m3.
CS6 is a 6-flat house with 24 inhabitants, a roof area of 420 m2, an annual toilet flushing
demand of 350.4 m3/y and a tank capacity of 21 m3. CS7 is a condominium with 32
inhabitants, a roof area of 680 m2, an annual toilet flushing demand of 467.2 m3/y and a
tank capacity of 28 m3. For the three case studies, the modeling results show a water
saving efficiency of 0.83, 0.79 and 0.76 for CS5, CS6 and CS7, respectively.
The CS8 case study considers the values of the example of application found in
Campisano and Modica [25], where a 4 people residential house with a daily toilet
flushing demand of 0.168 m3, a roof area of 186 m2, daily precipitation of 0.0018 m
and a size tank of 2.93 m3, achieving a water saving of 67%, was considered.
The CS9 case study refers to a real case study at University of Calabria in Southern
Italy [15], where a tank of 1.5 m3 is located at the base of an university building to
collect the water for an experimental full-scale green roof implementation and the water
is reused to irrigate the same green roof in the dry period. Finally, three hypothetical
cases, CS10, CS11 and CS12, have been considered to evaluate remain factors under
different conditions. Specifically, CS10 represent the hypothetical implementation of
RWH systems in the old town of Cosenza, CS11 in the old town of Matera, and CS12
in the new area of Quattromiglia, in the town of Rende, respectively.

Table 1. Case studies

Locations of case studies Case study
Bochum – Germany [7] CS1
Bochum – Germany [7] CS2
Sant Cugant del Vallès – Spain [13] CS3
Sant Cugant del Vallès – Spain [13] CS4
Genoa – Italy [5] CS5
Genoa – Italy [5] CS6
Genoa – Italy [5] CS7
Sicily – Italy [25] CS8
University of Calabria (Rende) – Italy [15] CS9
Old town of Cosenza – Italy CS10
Old town of Matera – Italy CS11
New Area of Quattromiglia (Rende) – Italy CS12
 Optimizing RWH Systems for Non-potable Water Uses 575

3 Results
3.1 Application of Rough Set Theory in Optimizing Rainwater-
Harvesting Systems
In real projects there is an enormous quantity of data that may be considered and this
makes hard the decision making process. In Rough Set method, all data should be
categorized. In this regard, the correlated RWH attributes must be determined. All the
information about the case studies in form of determined attributes, classification of
attributes and decision level for each of them should be provided. According to the data
gathered in Table 1, the main RWH attributes have been determined and are presented
in Table 2. The attributes have been classified based on 3 classes which denote the
suitability conditions for decisions and are high (H), medium (M) and low (L).

Table 2. Conditional attributes for ranking decisions of selected case studies

Conditional attributes Classification of individual situations Decision
(a) Building type 1 - One-family building/one-office with H
garden
2 - One-family building/one-office without M
garden
3 - Multi-family building/multi offices with
garden
4 - Multi-family building/multi offices L
without garden
(b) Roof type 1 - Slope roof with tiles, corrugated plastic, H
plastic or metal sheets
2 - Flat roof covered with plastic or metal
sheet
3 - Flat roof with concrete or asphalt slabs
4 - Flat roofs with gravel M
5 - Extensive green roof
6 - Intensive green roof L
(c) Roof Size (collecting area) 1 - Big surface capture (>250) H
2 - Average surface capture (100–250) M
3 - Small surface capture (<100) L
(d) The age of the building 1 - New building H
2 - Average age building M
3 - Old building L
(e) Average Annual Precipitation 1 - >700 or <300 H
2 - 300 to 700 M
(f) Number of building residents 1 - 1 to 4 H
(based on demand) 2 - 5 to 20 M
3 - more 20 L
(continued)
576 S. A. Palermo et al.

Table 2. (continued)
Conditional attributes Classification of individual situations Decision
(g) Density of city (based on the 1 - Low Density H
location of the barrels) 2 - Medium Density M
3 - High Density L
(h) Type of urban area 1 - New urban area H
2 - Average age urban area M
3 - Old urban area L
(i) Demand usage (m3/y) 1 - combined usage H
2 - one usage (toilet flushing or garden M
irrigation)
3 - laundry L
4 - terrace cleaning
5 - car washing
(j) Tank size 1 - Big Tank (>20 m3) H
2 - Medium Tank (6–20 m3) M
3 - Low Tank (<6 m3) L
(k) Economic 1 - Very Economic H
2 - Partly Economic M
3 - Expensive L

According to 11 attributes and classes, the selected case studies have been ranked
from 1 to 3 and the results are presented in Table 3. For instance, in the first case study
(CS1), since the conditional attribute (a) that is “Building type” is “One-family build-
ing”, the highest rank, i.e. 3, has been selected. Since the table represents the correlation
between the case studies and conditional attributes it is named “decision rules”.

Table 3. Data Inspection for analysis of site selection decision ranking

Case study Conditional attributes Decision level
a b c d e f g h i j k
CS1 3 3 2 2 3 3 1 2 3 1 2 M
CS2 1 3 3 2 3 1 1 2 2 1 2 M
CS3 3 3 2 2 2 3 1 2 3 2 1 H
CS4 2 3 3 2 2 1 1 2 3 3 1 H
CS5 2 3 3 2 3 2 1 2 2 2 2 M
CS6 2 3 3 2 3 1 1 2 2 3 2 H
CS7 2 3 3 2 3 1 1 2 2 3 2 H
CS8 2 3 2 2 2 3 2 2 2 1 3 M
CS9 2 2 2 2 3 1 2 2 2 1 3 H
CS10 1 3 2 1 3 2 1 1 2 2 1 L
CS11 1 3 1 1 3 2 1 1 3 3 1 L
CS12 1 3 3 3 3 1 1 3 2 3 2 H
 Optimizing RWH Systems for Non-potable Water Uses 577

All the decisions and attributes have been checked to find out the existence of non-
deterministic rules that means that for case studies of similar attributes decisions are
different. The number of non-deterministic rules in Table 3 was zero. Therefore, the
number of conditional attributes is sufficient for determining the decisions. The found
reduction in the data is presented in Table 4.

Table 4. The founded reduction

Raw Reduction Raw Reduction Raw Reduction
1 {a, b, d, j} 13: {a, b, i, j} 25: {a, f, j}
2 {a, c, e, j} 14: {a, b, j, k} 26: {d, e, g, j}
3 {b, c, e, j} 15: {a, b, h, j} 27: {c, f, j}
4: {a, b, c, j} 16: {b, c, h, j} 28: {b, h, i, j}
5: {b, d, e, j} 17: {a, e, h, j} 29: {e, g, h, j}
6: {a, d, e, j} 18: {b, e, h, j} 30: {f, g, h, j}
7: {b, c, d, j} 19: {c, e, i, j} 31: {b, c, i, j}
8: {a, d, e, f 20: {a, e, i, j} 32: {b, d, j, k}
9: {a, e, f, h} 21: {d, f, g, j} 33: {b, h, j, k}
10: {b, d, f, j} 22: {b, d, i, j} 34: {e, j, k}
11: {a, d, f, k} 23: {c, e, g, j} 35: {f, j, k}
12: {a, f, h, k} 24: {b, f, h, j}

After deriving the reducts, the decision rules can be achieved by overlaying the
determined reducts on the data. A decision table free of contradiction and determining a
minimal decision algorithm can be achieved after elimination of all non-deterministic
rules that was zero in this study. The contradictions have been analyzed based on the
conditional attributes and the decisions in selected case studies. Moreover, if the
attributes do not cause any contradiction they can be removed. In order to check the
impact of an attribute on the result, the attributes can be removed one by one. For
example, if the conditional attributes (a), (b), (c), and (d) be removed, the decision rules
of case studies 1 and 2 might be contradictory that means the decision levels of these
two case studies are subordinate to the mentioned conditional attributes. In this regard,
and after elimination of all removable conditional attribute or classes, the minimal
decision algorithm has been obtained and is presented in Table 5.

Table 5. Minimal decision algorithm

Rules
Rule 1 (d = 1) => (Decision = L)
Rule 2 (b = 3) & (j = 1) => (Decision = M)
Rule 3 (a = 2) & (j = 2) => (Decision = M)
Rule 4 (f = 1) & (j = 3) => (Decision = H)
Rule 5 (b = 2) => (Decision = H)
Rule 6 (a = 3) & (e = 2) => (Decision = H)
578 S. A. Palermo et al.

The validation of the rules is presented in Tables 6 and 7. It must be mentioned

that, since the selected case studies are only 12, the accuracy of the rules might not be
high. To be able to extend the result of the method to other similar case studies in RWH
systems, more field data might be required.

Table 6. Confusion matrix (sum over 10 passes)

1 2 3 None
1 2 0 0 0
2 0 2 2 0
3 1 4 1 0

Table 7. Average accuracy [%]

Correct Incorrect None
Total 40.00 ±43.59 60.00 ± 43.59 0.00 ± 0.00
1 20.00 ± 40.00 0.00 ± 0.00 0.00 ± 0.00
2 15.00 ± 32.02 15.00 ± 32.02 0.00 ± 0.00
3 5.00 ± 15.00 45.00 ± 47.17 0.00 ± 0.00

3.2 Application of TOPSIS in Ranking of Rainwater-Harvesting Systems

In this section, the TOPSIS method has been used to rank the selected case studies and
the results are compared with those of a simple ranking. The results of simple ranking
are presented in Tables 8 and 9 and those obtained by TOPSIS method in Tables 10,
11 and 12.

Table 8. Values of each attribute for each case study

Case study Attributes
a b c d e f g h i j k
CS1 3 3 150 2 787 4 1 2 58.4 6 2
CS2 1 3 320 2 787 24 1 2 175.2 14 2
CS3 3 3 107 2 612 3 1 2 47.1 13 1
CS4 2 3 625 2 612 42 1 2 705.2 31 1
CS5 2 3 420 2 1086 16 1 2 233.6 14 2
CS6 2 3 420 2 1086 24 1 2 350.4 21 2
CS7 2 3 680 2 1086 32 1 2 467.2 28 2
CS8 2 3 186 2 657 4 2 2 61.3 2.93 3
 Optimizing RWH Systems for Non-potable Water Uses 579

Table 9. Simple ranking of the factors for each case study

Case study Rank of attributes Sum of ranking Final rank
a b c d e f g h i j k
CS1 1 1 6 1 2 2 4 1 7 6 2 33 6
CS2 3 1 4 1 2 4 4 1 5 4 2 23 4
CS3 1 1 7 1 4 1 1 1 8 5 3 24 5
CS4 2 1 2 1 4 6 1 1 1 1 3 18 2
CS5 2 1 3 1 1 3 3 1 4 4 2 19 3
CS6 2 1 3 1 1 4 3 1 3 3 2 18 2
CS7 2 1 1 1 1 5 3 1 2 2 2 17 1
CS8 2 1 5 1 3 2 2 1 6 7 1 23 4

Table 10. TOPSIS matrix without scale (Normalized)

Case study Attributes
a b c d e f g h i j k
CS1 0.48 0.35 0.13 0.35 0.32 0.11 0.46 0.35 0.06 0.11 0.36
CS2 0.16 0.35 0.27 0.35 0.32 0.26 0.46 0.35 0.18 0.26 0.36
CS3 0.48 0.35 0.09 0.35 0.25 0.25 0.03 0.35 0.05 0.25 0.18
CS4 0.32 0.35 0.53 0.35 0.25 0.59 0.03 0.35 0.73 0.59 0.18
CS5 0.32 0.35 0.36 0.35 0.44 0.26 0.44 0.35 0.24 0.26 0.36
CS6 0.32 0.35 0.36 0.35 0.44 0.40 0.44 0.35 0.36 0.40 0.36
CS7 0.32 0.35 0.58 0.35 0.44 0.53 0.44 0.35 0.48 0.53 0.36
CS8 0.32 0.35 0.16 0.35 0.27 0.06 0.04 0.35 0.06 0.06 0.54

Table 11. TOPSIS matrix without scale and equal weighted

Case study Attributes
a b c d e f g h i j k
CS1 0.044 0.032 0.012 0.032 0.029 0.010 0.041 0.032 0.005 0.010 0.033
CS2 0.015 0.032 0.025 0.032 0.029 0.024 0.041 0.032 0.016 0.024 0.033
CS3 0.044 0.032 0.008 0.032 0.023 0.022 0.003 0.032 0.004 0.022 0.016
CS4 0.029 0.032 0.048 0.032 0.023 0.053 0.003 0.032 0.066 0.053 0.016
CS5 0.029 0.032 0.033 0.032 0.040 0.024 0.040 0.032 0.022 0.024 0.033
CS6 0.029 0.032 0.033 0.032 0.040 0.036 0.040 0.032 0.033 0.036 0.033
CS7 0.029 0.032 0.053 0.032 0.040 0.048 0.040 0.032 0.044 0.048 0.033
CS8 0.029 0.032 0.014 0.032 0.024 0.005 0.003 0.032 0.006 0.005 0.049
V+ 0.044 0.032 0.053 0.032 0.040 0.053 0.003 0.032 0.066 0.053 0.049
V− 0.015 0.032 0.008 0.032 0.023 0.005 0.041 0.032 0.004 0.005 0.016
580 S. A. Palermo et al.

Table 12. Ranking in TOPSIS based on higher CL and comparison with simple ranking
Case study d+ d− CL Rank in TOPSIS Simple rank method Decision level
CS4 0.040 0.109 0.731 1 2 H
CS7 0.049 0.090 0.645 2 1 H
CS6 0.063 0.064 0.504 3 3 H
CS5 0.077 0.049 0.389 4 4 M
CS3 0.095 0.054 0.363 5 7 H
CS8 0.101 0.053 0.342 6 6 M
CS2 0.088 0.038 0.302 7 5 M
CS1 0.105 0.035 0.250 8 8 M

Despite the fact that in some case studies the difference in ranking methods are
minor since in TOPSIS all correlated attributes and the differences among the values
are taken into consideration the results could be more accurate.

4 Conclusions

There are many benefits in Rainwater harvesting (RWH) systems mainly water saving
for non-potable water uses and surface runoff mitigation. Moreover, the collected
rainwater can be re-used for several purposes including green roofs and garden, flushing
toilets, etc. Our analysis showed that, in previous studies, some important factors in the
analysis and the feasibility of the RWH systems have been neglected and the opti-
mization of RWH systems mostly is limited to optimize the size of the tankers according
to hydrological and hydraulic analysis and in some cases, this is combined with an
economic analysis. In this paper, multi-objective optimization approaches have been
considered for comparing algorithms and evaluating the performance of alternatives to
identify the ideal solution. For this, a limited set of data extracted from several case
studies has been used. The selection of the case studies has been made considering the
main possible attributes/factors confronting in rainwater harvesting systems. The results
show that the Rough Set method is a suitable way for analysis of RWH systems and the
outcomes can be useful in decision making by decreasing the uncertainties, reducing the
cost, and increasing the efficiency. According to the results, TOPSIS ranking method
showed good agreement with the decision levels in the case studies. This may be due to
the consideration of all correlated attributes and of the differences between the values of
this ranking method. In conclusion, the numerical optimization of RWH systems may
improve previous studies in the field. Moreover, the Rough Set and TOPSIS methods
could be applied as a useful approach in rainwater harvesting systems investigations and
provide an additional tool to identify the optimal system and the best site.

Acknowledgements. The study was co-funded by the “Innovative Building Envelope through
Smart Technology (I-Best)” Project funded by the Italian National Operational Program
“Enterprise and Competitiveness” 2014–2020 ERDF – I AXIS “Innovation” - Action 1.1.3 –
“Support for the economic enhancement of innovation through experimentation and the adoption
of innovative solutions in processes, products and organizational formulas, as well as through the
financing of the industrialization of research results”.
 Optimizing RWH Systems for Non-potable Water Uses 581

References
1. Palermo, S.A., Zischg, J., Sitzenfrei, R., Rauch, W., Piro, P.: Parameter sensitivity of a
microscale hydrodynamic model. In: Mannina, G. (ed.) UDM 2018. GET, pp. 982–987.
Springer, Cham (2019). https://doi.org/10.1007/978-3-319-99867-1_169
2. Maiolo, M., Pantusa, D.: Sustainable water management index, SWaM_Index. Cogent Eng.
6(1), 1603817 (2019). https://doi.org/10.1080/23311916.2019.1603817
3. Christian Amos, C., Rahman, A., Mwangi Gathenya, J.: Economic analysis and feasibility of
rainwater harvesting systems in urban and peri-urban environments: a review of the global
situation with a special focus on Australia and Kenya. Water 8(4), 149 (2016). https://doi.
org/10.3390/w8040149
4. Campisano, A., et al.: Urban rainwater harvesting systems: research, implementation and
future perspectives. Water Res. 115, 195–209 (2017). https://doi.org/10.1016/j.watres.2017.
02.056
5. Palla, A., Gnecco, I., La Barbera, P.: The impact of domestic rainwater harvesting systems in
storm water runoff mitigation at the urban block scale. J. Environ. Manag. 191, 297–305
(2017). https://doi.org/10.1016/j.jenvman.2017.01.025
6. Petrucci, G., Deroubaix, J.F., De Gouvello, B., Deutsch, J.C., Bompard, P., Tassin, B.:
Rainwater harvesting to control stormwater runoff in suburban areas. An experimental case-
study. Urban Water J. 9(1), 45–55 (2012). https://doi.org/10.1080/1573062X.2011.633610
7. Herrmann, T., Schmida, U.: Rainwater utilisation in Germany: efficiency, dimensioning,
hydraulic and environmental aspects. Urban Water 1(4), 307–316 (2000). https://doi.org/10.
1016/S1462-0758(00)00024-8
8. GhaffarianHoseini, A., Tookey, J., GhaffarianHoseini, A., Yusoff, S.M., Hassan, N.B.: State
of the art of rainwater harvesting systems towards promoting green built environments: a
review. Desalin. Water Treat. 57(1), 95–104 (2016). https://doi.org/10.1080/19443994.2015.
1021097
9. Oberascher, M., Zischg, J., Palermo, S.A., Kinzel, C., Rauch, W., Sitzenfrei, R.: Smart rain
barrels: advanced LID management through measurement and control. In: Mannina, G. (ed.)
UDM 2018. GET, pp. 777–782. Springer, Cham (2019). https://doi.org/10.1007/978-3-319-
99867-1_134
10. Li, Z., Boyle, F., Reynolds, A.: Rainwater harvesting and greywater treatment systems for
domestic application in Ireland. Desalination 260(1–3), 1–8 (2010). https://doi.org/10.1016/
j.desal.2010.05.035
11. Campisano, A., Modica, C.: Rainwater harvesting as source control option to reduce roof
runoff peaks to downstream drainage systems. J. Hydroinform. 18(1), 23–32 (2016). https://
doi.org/10.2166/hydro.2015.133
12. Jones, M.P., Hunt, W.F.: Performance of rainwater harvesting systems in the southeastern
United States. Resour. Conserv. Recycl. 54(10), 623–629 (2010). https://doi.org/10.1016/j.
resconrec.2009.11.002
13. Domènech, L., Saurí, D.: A comparative appraisal of the use of rainwater harvesting in
single and multi-family buildings of the Metropolitan Area of Barcelona (Spain): social
experience, drinking water savings and economic costs. J. Clean. Prod. 19(6–7), 598–608
(2011). https://doi.org/10.1016/j.jclepro.2010.11.010
14. Cipolla, S.S., Altobelli, M., Maglionico, M.: Decentralized water management: rainwater
harvesting, greywater reuse and green roofs within the GST4Water project. In: Multidis-
ciplinary Digital Publishing Institute Proceedings, vol. 2, no. 11, p. 673 (2018). https://doi.
org/10.3390/proceedings2110673
582 S. A. Palermo et al.

15. Piro, P., Turco, M., Palermo, S.A., Principato, F., Brunetti, G.: A comprehensive approach to
stormwater management problems in the next generation drainage networks. In: Cicirelli, F.,
Guerrieri, A., Mastroianni, C., Spezzano, G., Vinci, A. (eds.) The Internet of Things for
Smart Urban Ecosystems. IT, pp. 275–304. Springer, Cham (2019). https://doi.org/10.1007/
978-3-319-96550-5_12
16. Becciu, G., Raimondi, A., Dresti, C.: Semi-probabilistic design of rainwater tanks: a case
study in Northern Italy. Urban Water J. 15(3), 192–199 (2018). https://doi.org/10.1080/
1573062X.2016.1148177
17. Pawlak, Z.: Rough set theory and its applications to data analysis. Cybern. Syst. 29(7), 661–
688 (1998). https://doi.org/10.1080/019697298125470
18. Arabani, M., Sasanian, S., Farmand, Y., Pirouz, M.: Rough-set theory in solving road
pavement management problems (Case Study: Ahwaz-Shush Highway). Comput. Res. Prog.
Appl. Sci. Eng. (CRPASE) 3(2), 62–70 (2017)
19. Arabani, M., Pirouz, M., Pirouz, B.: Optimization of geotechnical studies using basic set
theory. In: 1st Conference of Civil and Development, Zibakenar, Iran (2012)
20. Hwang, C.L., Yoon, K.P.: Multiple Attributes Decision-Making Methods and Applications.
Springer, Berlin (1981). https://doi.org/10.1007/978-3-642-48318-9
21. Balioti, V., Tzimopoulos, C., Evangelides, C.: Multi-criteria decision making using TOPSIS
method under fuzzy environment. Appl. Spillway Sel. Proc. 2, 637 (2018). https://doi.org/
10.3390/proceedings2110637
22. Krohling, R.A., Pacheco, A.G.: A-TOPSIS an approach based on TOPSIS for ranking
evolutionary algorithms. Procedia Comput. Sci. 55, 308–317 (2015). https://doi.org/10.1016/
j.procs.2015.07.054
23. Haghshenas, S.S., Neshaei, M.A.L., Pourkazem, P., Haghshenas, S.S.: The risk assessment
of dam construction projects using fuzzy TOPSIS (case study: Alavian Earth Dam). Civil
Eng. J. 2(4), 158–167 (2016). https://doi.org/10.28991/cej-2016-00000022
24. Haghshenas, S.S., Mikaeil, R., Haghshenas, S.S., Naghadehi, M.Z., Moghadam, P.S.: Fuzzy
and classical MCDM techniques to rank the slope stabilization methods in a rock-fill reservoir
dam. Civil Eng. J. 3(6), 382–394 (2017). https://doi.org/10.28991/cej-2017-00000099
25. Campisano, A., Modica, C.: Optimal sizing of storage tanks for domestic rainwater
harvesting in Sicily. Resour. Conserv. Recycl. 63, 9–16 (2012). https://doi.org/10.1016/j.
resconrec.2012.03.007
 New Mathematical Optimization Approaches
for LID Systems

Behrouz Pirouz1(&) , Stefania Anna Palermo2 ,

Michele Turco3 , and Patrizia Piro2
1
Department of Mechanical, Energy and Management Engineering,
University of Calabria, Rende, CS, Italy
[email protected]
2
Department of Civil Engineering, University of Calabria, Rende, CS, Italy
{stefania.palermo,patrizia.piro}@unical.it
3
Department of Environmental and Chemical Engineering,
University of Calabria, Rende, Italy
[email protected]

Abstract. Urbanization affects ecosystem health and downstream communities

by changing the natural flow regime. In this context, Low Impact Development
(LID) systems are important tools in sustainable development. There are many
aspects in design and operation of LID systems and the choice of the selected
LID and its location in the basin can affect the results. In this regard, the
Mathematical Optimization Approaches can be an ideal method to optimize
LIDs use. Here we consider the application of TOPSIS (Technique for Order
Preference by Similarity to Ideal Solution) and Rough Set theory (multiple
attributes decision-making method). An advantage of using the Rough Set
method in LID systems is that the selected decisions are explicit, and the method
is not limited by restrictive assumptions. This new mathematical optimization
approach for LID systems improves previous studies on this subject. Moreover,
it provides an additional tool for the analysis of essential attributes to select and
optimize the best LID system for a project.

Keywords: Optimization
LID
Rough Set Theory
TOPSIS method

1 Introduction

Low Impact Development (LID) systems are important tools in sustainable develop-
ment. There are different types of LID practices such as green roofs, green wall,
bioretention cell, permeable pavements, rainwater harvesting systems, etc. In the design
and operation of LID systems, many components must be considered. When choosing
and designing the best LID practices many factors can affect their efficiency in terms of
flooding risk mitigation, water quality improvement, water saving, urban heat island
reduction, air pollution decreasing. Previous studies are generally limited to focus the
design of a type of LID based on a determined scenario and location. However, these
elements are not fixed.
In this research, the application of mathematical optimization approaches by
TOPSIS ranking method and attributes analysis by Rough Set in evaluation of alter-
native decisions is described.
© Springer Nature Switzerland AG 2020
Y. D. Sergeyev and D. E. Kvasov (Eds.): NUMTA 2019, LNCS 11973, pp. 583–595, 2020.
https://doi.org/10.1007/978-3-030-39081-5_50
584 B. Pirouz et al.

1.1 LID Systems

In the last decades, the combined effect of urbanization and climate change produced
several environmental adverse impacts as flooding risk, water and air pollution, water
scarcity, and urban heat island effect [1–5].
Specifically, from a hydrologic point of view, land use changes led to significant
alterations of the natural hydrological cycle with a reduction of infiltration and evap-
otranspiration rates, a decreasing of groundwater recharge and baseflow, and a growth
of surface runoff. This phenomenon makes cities vulnerable to local flooding and
combined sewer overflows (CSOs) [6, 7].
In response to these issues, a novel approach, which considers the implementation
of sustainable nature-based stormwater management strategies, is necessary [8].
In this regard, several strategies, known as Low Impact Development (LID) sys-
tems [9], have gained increasing popularity. These techniques are small-scale units,
which provide several benefits at multiple scales [10–13].
LID systems consist of a series of facilities - such as green roofs, green wall, rain
gardens, bioretention cells, permeable pavements, infiltration trenches, rain water
harvesting systems, and so on - whose purpose is to infiltrate, filter, store, evaporate,
and detain runoff close to its source [14–16].
Therefore, their implementation in an urban area can create a complex system of
temporal storage facilities allowing an innovative urban stormwater management
approach [17].
In recent years, the implementation of these sustainable solutions has attracted
widespread interest from researchers, urban planners, and city managers researchers
[18] and several studies have been devoted to LID design, benefits, and simulation
models of their behavior.
From these studies, as pointed out by Eckart et al. [15], it appears that: (i) the
appropriate design of LID is affected by the site location (in terms of soil
type/conditions, plants, rainfall regime, land use types and other meteorological and
hydrological properties); (ii) the optimal selection and placement of LID is one of the
most crucial factors to consider to achieve the maximum efficiency at the minimum cost.
The implementation of these systems can improve the sustainable development and
mitigate several other environmental impacts in addition to urban flooding risk, such as
water and air pollution, water scarcity, urban heat island effect. Therefore, a large
number of factors have to be considered during the design process and for the choose of
the site location.
After a deeper analysis of different studies carried out on LID solutions, several
factors affecting their efficiency at multiple scale have been identified and are here
considered here for optimizing their use.

2 Methodology

In this paper, we discuss the application of TOPSIS (Technique for Order Preference by
Similarity to Ideal Solution) and Rough Set theory (multiple attributes decision-making
method) in optimization of LIDs.
 New Mathematical Optimization Approaches for LID Systems 585

2.1 Rough Set Theory

Low Impact Development approaches are Multi-Objectives and there are many uncer-
tainties in the selection of these objectives [19]. The Rough Set method, introduced by
Pawlak [20] in 1982, can be used as an excellent mathematical tool in analysis of vague
description of decisions such as quality of data, means variation or uncertainty that
follows from information. The rough sets philosophy is based on assumption that, for
every objects a certain information (data, knowledge) exists, that can be expressed as
attributes [21].
With respect to the available data, the objects with the same description are
indiscernible and a set of indiscernible objects, named elementary set, can be provided
to build knowledge about a real system. Deal with quantitative or qualitative data
depends on the input information and before the analysis, it is necessary to remove the
irregularities. With respect to the output data, the relevance of particular attributes and
their subsets to the quality of approximation can be acquired [22].
In this regard, the induction-based approach can provide clear rules for decision-
makers (DMs) in the form of “if…, then…”. The concept of approximation space in
rough set method can be described in a given approximation space as follows:

apr ¼ ðU; AÞ ð1Þ

U is a finite and non-empty set and A is set of attributes in the given space. Based on
the approximation space, the lower and upper approximations of a set can be defined.
Let X be a subset of U and the upper and lower approximation of X in A are:

apr ðAÞ ¼ fxjx 2 U; U/ind ðAÞ  Xg ð2Þ

apr ðAÞ ¼ fxjx 2 U; U=ind ðAÞ \ X 6¼ ug ð3Þ

where:

U/ind ðaÞ ¼ fðxi; xjÞ 2 U  U; f ðxi; aÞ ¼ f ðxj; aÞ; 8a 2 Ag ð4Þ

Equation (2), that is the best upper approximation of X in A, means the minimum
composed set in A containing X, and Eq. (3), that is the best lower approximation,
means the maximum composed set in A contained in X. The graphical illustration of
approximations in the rough set method is shown in Fig. 1.
The boundary represent as:

BN ðAÞ ¼ aprðAÞ  aprðAÞ ð5Þ

The reducts and decision rules can be defined as follows:

The reduct RED (B), is a minimal set of attributes B  A such that rB (U) = rA
(U). rB (U) indicates the quality of approximation of U by B.
586 B. Pirouz et al.

Fig. 1. Graphical illustration of the rough set approximations

The equation is:

P
card ðBðXi ÞÞ
rB ðU Þ ¼ ð6Þ
card ðU Þ

After providing the result of reducts, the decision rules can be derived by using the
overlaying of the reducts on the information systems. An expressed decision rule can
be as follow:

u)h ð7Þ

where:

u is the conjunction of elementary conditions;

) Represents indicates
h represents the disjunction of elementary decisions.

2.2 TOPSIS Method

TOPSIS (Technique for Order Preference by Similarity to Ideal Solution) is a method
developed by Hwang and Yoon in 1981 to solve the ranking and compare problems
[23]. The ranking in this method is made according to the similarity to ideal solution
[24, 25]. The TOPSIS method can be applied to a wide range of multi-attribute decision
 New Mathematical Optimization Approaches for LID Systems 587

making with several attributes [26–28]. The graphical illustration of the TOPSIS
methodology is presented in Fig. 2.

Fig. 2. Graphical illustration of the TOPSIS methodology, (A+ represent the ideal point, A−
represent the negative-ideal point)

The ranking by TOPSIS is carried out through seven steps:

Step 1: Create the matrix
 
nij m  n ð8Þ

Step 2: Construct the normalized decision matrix

aij
N ¼ nij ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Pm ; i ¼ 1; 2; . . .; m & j ¼ 1; 2; . . .; n ð9Þ
i ¼ 1aij
2

Step 3: Construct the weighted normalized decision matrix

V ¼ N  Wn  n ð10Þ

Step 4: Determine the solutions (the ideal and negative-ideal solutions)

     
The ideal solution A þ ¼ maxi aij jj 2 J ; mini aij jj 2 J þ ð11Þ
      
The negativeideal solutions A ¼ mini aij jj 2 J ; maxi aij jj 2 J þ ð12Þ
588 B. Pirouz et al.

where,

J þ ¼ fj ¼ q; 2; . . .; njjg Associated with positive impact criteria

J ¼ fj ¼ q; 2; . . .; njjg Associated with negative impact criteria

Step 5: Determine the distance of alternatives vij from the ideal solution and the
negative-ideal solutions
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Xn

The distance from ideal solution: diþ ¼ j¼1

v ij  v j
þ
; i ¼ 1; 2; . . .; m ð13Þ

rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Xn

The distance from negativeideal solution : di ¼ v  v þ

;
j¼1 ij j ð14Þ
i ¼ 1; 2; . . .; m

Step 6: Calculate the closeness to the negative-ideal condition, CL*

di
CLi ¼ ; 0  CLi  1 & i ¼ 1; 2; . . .; m ð15Þ
di þ diþ

where,

CLi ¼ 1 if The solution has the best condition and means the highest rank
CLi ¼ 0 if The solution has the worst condition and means the lowest rank

Step 7: Rank in order to the highest CL*.

2.3 Case Studies

In order to perform the analysis of LID practices by Rough Set and TOPSIS a set of
data is required. For this purpose, sites have been chosen according to different con-
ditions and considering the main factors in the LIDs selection. The selected site along
with hypothesis are presented in Table 1.
 New Mathematical Optimization Approaches for LID Systems 589

Table 1. Site selection hypothesis to rank the attributes

Site Description
S1 An high urbanized area (78% impervious surfaces), with extremely wet condition,
average age, presenting High Flooding Risk, Low Water Scarcity, High Water
Pollution, Medium Heat Island Effect, High Air Pollution, supposing to replace 50% of
impervious area with LID (permeable pavement and green roof); and that the
inhabitants are supported to implement them, thus partially economic
S2 An urbanized area (65% of impervious surfaces), with moderately wet condition, old
area, presenting High Flooding Risk, Medium Water Scarcity, High Water Pollution,
High Heat Island Effect, High Air Pollution, where it is not possible to implement LID;
not economic
S3 A peri-urban area (45% of impervious surfaces), with extremely dry condition, new
area, presenting Low Flooding Risk, High Water Scarcity, Medium Water Pollution,
Low Heat Island Effect, Medium Air Pollution, where it is possible to implement
combined usage of LID (rainwater harvesting and biofiltration trench) for a percentage
of 65%; high economic
S4 A peri-urban area (40% of impervious surfaces), with moderate climate condition,
average age, presenting Medium Flooding Risk, Medium Water Scarcity, Low Water
Pollution, Low Heat Island Effect, Medium Air Pollution, where it is possible to
implement combined usage of LID (rainwater harvesting and green wall) for a
percentage of 40%; partially economic
S5 A normal urban area (70% of impervious surfaces), with extremely dry climate
condition, average age, presenting Low Flooding Risk, High Water Scarcity, Low
Water Pollution, High Heat Island Effect, High Air Pollution, where it is possible to
implement combined usage of LID (rainwater harvesting, green roof and green wall)
for a percentage of 50%; partially economic
S6 A rural area (20% of impervious surface), with extremely dry climate condition,
average area, presenting s Low Flooding Risk, High Water Scarcity, Low Water
Pollution, Low Heat Island Effect, Low Air Pollution, where it would be better to
implement rainwater harvesting for a percentage of 70%; partially economic

As it can be recognized, in Table 1. The main factors considered to define the case
studies include climate condition, urbanization level, age of site, flood risk, water
scarcity, water and air pollution, Heat Island Effect, LID implementation percentage
and economical condition.

3 Results
3.1 Application of Rough Set Theory in Optimizing LIDs
In Rough Set method, at the first stage all factors must be categorized in form of
attributes that are classified. This, and the decision level for each of them, can be done
according to previous standards, papers or experts. We identified 12 conditional
decision attributes for LIDs that are presented in Tables 2 and 3.
590 B. Pirouz et al.

Table 2. Conditional decision attributes in selected sites, part 1

Conditional attributes Classification Decision
(a) Type of area 1- Urban H
2- Peri-urban M
3- Rural L
(b) Climate condition based on 1- Extremely wet or extremely dry H
precipitation 2- Moderately wet or moderately dry M
(c) Age of area 1- New area H
2- Average age area M
3- Old area L
(d) Impervious surfaces of the selected 1- >75% of area H
area 2- 50%–75% of area M
3- 25%–50% of area L
4- <25% of area N
(e) Flooding risk 1- High risk H
2- Medium risk M
3- Low risk L
(f) Water scarcity 1- High water scarcity H
2- Medium water scarcity M
3- Low water scarcity L
(g) Water pollution 1- High water pollution H
2- Medium water pollution M
3- Low water pollution L

Table 3. Conditional decision attributes in selected sites, part 2

Conditional attributes Classification Decision
(h) Urban heat island effect 1- High heat island effect H
2- Medium heat island effect M
3- Low heat island effect L
(i) Air pollution 1- High air pollution H
2- Medium air pollution M
3- Low air pollution L
(j) LID percentage 1- >60% of area H
implementation 2- 30%–60% of area M
3- <30% of area L
(k) LID usage 1- Combined implementation (more than 1 LID) H
2- Only one type implementation M
3- No implementation N
(l) Economic 1- Economic H
2- Partially economic M
3- Not economic L
 New Mathematical Optimization Approaches for LID Systems 591

As it is clear from Tables 2 and 3 the conditional attributes have been categorized
in at most four classes with high (H), medium (M), low (L) and no (N) suitability
conditions for decisions. As next step, the selected sites presented in Table 1 have been
ranked according to the attributes of Tables 2 and 3. The result is presented in Table 4.
In Table 4 the ranks are based on the conditional attribute in the site and for ranks from
0 to 4. For example, in site 1 (S1), the conditional attribute (a) that is “Type of Area” is
“Urban”, and therefore the highest rank (3) has been selected.

Table 4. Ranking of decisions attributes in selected sites

Sites Conditional attributes Decision levels
a b c d e f g h i j k l
S1 3 3 2 3 3 1 3 2 3 2 3 2 H
S2 3 2 1 2 3 2 3 3 3 1 0 1 L
S3 2 3 3 1 1 3 2 1 2 3 3 3 H
S4 2 2 2 1 2 2 1 1 2 2 3 2 M
S5 3 3 2 2 1 3 1 3 3 2 3 2 H
S6 1 3 2 0 1 3 1 1 1 3 2 2 M

3.2 Determining Minimal Decision Algorithm

The finding of a minimal decision algorithm can be achieved by analysis of decision
rules in all sites and finding non-deterministic rules or sites, thus assigning different
decision levels for different sites under the same class for every conditional attribute. In
this regard, at the next stage, the contradiction between data can be checked according
to the conditional attributes ranks and the correlated decisions and some of the attri-
butes can be removed if this does not cause any contradiction. For this purpose, the
conditional attribute can be eliminated one by one to check the role of that attribute in
the result. For instance, if the conditional attributes (a), (b), and (c) were removed, the
decision rules of sites S1 and S2 might be contradictory to each other, which means
that the decision levels of these two sites are subordinate to one of the conditional
attributes. Finally, and after checking all of the combinations, the minimal decision
algorithm can be extracted that will be the main rules for each level of decisions. The
determined rules according to the attributes data of Table 4 and the six case studies of
Table 1 are as follows:

Rule 1 : ðc ¼ 1Þ ) ðD ¼ LÞ
Rule 2 : ðg ¼ 1Þ & ðh ¼ 1Þ ) ðD ¼ MÞ
Rule 3 : ðb ¼ 3Þ & ðk ¼ 3Þ ) ðD ¼ HÞ

Thus, by the determined rules it is possible to make a decision with minimum

attributes. However, these rules have been generated by the decision on the selected
case studies. Therefore, by increasing the number of the case studies, the accuracy of
592 B. Pirouz et al.

the rules will increase and after checking the validation and accuracy of the rules, it will
be possible to extend the rules for other sites.

3.3 Application of TOPSIS in Selection of LID Practices

In this section, the application of TOPSIS method in ranking the selected sites for LIDs
is presented. To provide the final ranking, the data of Tables 2 and 3 that are the
decision attributes for the selected sites have been used. The result of the process as
explained in the methodology section are presented in Tables 5, 6 and 7.

Table 5. Attributes matrix without scale

Case study Attributes
a b c d e f g h i j k l
S1 0.50 0.45 0.39 0.69 0.60 0.17 0.60 0.40 0.50 0.36 0.47 0.54
S2 0.50 0.30 0.20 0.46 0.60 0.33 0.60 0.60 0.50 0.18 0.00 0.18
S3 0.33 0.45 0.59 0.23 0.20 0.50 0.40 0.20 0.33 0.54 0.47 0.54
S4 0.33 0.30 0.39 0.23 0.40 0.33 0.20 0.20 0.33 0.36 0.47 0.36
S5 0.50 0.45 0.39 0.46 0.20 0.50 0.20 0.60 0.50 0.36 0.47 0.36
S6 0.17 0.45 0.39 0.00 0.20 0.50 0.20 0.20 0.17 0.54 0.32 0.36

Table 6. Attributes matrix without scale and with equal weight

Case study Attributes
a b c d e f g h i j k l
S1 0.04 0.04 0.03 0.06 0.05 0.01 0.05 0.03 0.04 0.03 0.04 0.04
S2 0.04 0.03 0.02 0.04 0.05 0.03 0.05 0.05 0.04 0.01 0.00 0.01
S3 0.03 0.04 0.05 0.02 0.02 0.04 0.03 0.02 0.03 0.04 0.04 0.04
S4 0.03 0.03 0.03 0.02 0.03 0.03 0.02 0.02 0.03 0.03 0.04 0.03
S5 0.04 0.04 0.03 0.04 0.02 0.04 0.02 0.05 0.04 0.03 0.04 0.03
S6 0.01 0.04 0.03 0.00 0.02 0.04 0.02 0.02 0.01 0.04 0.03 0.03

Table 7. Final ranking in based on higher CL and comparison with initial decision level
Case study d+ d− CL Rank by The initial
TOPSIS decision level
S1 0.039 0.102 0.722 1 H
S5 0.057 0.086 0.598 2 M
S3 0.066 0.080 0.548 3 H
S2 0.072 0.081 0.530 4 H
S4 0.074 0.059 0.446 5 L
S6 0.094 0.055 0.369 6 M
 New Mathematical Optimization Approaches for LID Systems 593

The results of ranking by TOPSIS in Table 7 and the comparisons with the initial
decision levels represent that in some sites the results are the same such as S1, S3, S2
and S4. However, the decision levels in S6 was M but it is at the end of TOPSIs
ranking. This might be based on the consideration of all correlated factors at the same
time and more exact.

4 Conclusions

The analysis showed that in design and operation of the LID systems, many compo-
nents can be considered and that in choosing the best LID practices and implementation
percentage many factors can affect the results. In previous studies, generally, the
attentions was limited to design a type of LID based on determined scenarios and for a
selected site that both are not fixed elements and might need to be optimized.
The results of this application of mathematical optimization approaches by TOPSIS
ranking method and attributes analysis by Rough Set in evaluation of alternative
decisions confirm the advantage of using these methods. The rules provided by Rough
Set method can improve the designing decisions. The generated decisions are explicit,
and the results are not limited to restrictive assumptions. With consideration of more
case studies, more stringent decision rules can be achieved. Moreover, the final ranks of
TOPSIS shows the advantages in compared with simple ranking method.
In conclusion, the new presented mathematical optimization approaches can
improve the previous studies about LIDs. They provide an additional tool for engineers
in analysis of essential attributes to select and optimize the best LID system for a
project and accordingly define the scenarios and hydrologic or hydraulic modeling.
This means that the presented methods would provide a baseline for decision-making
and would increase the efficiency of the systems and decrease the project cost by
preventing uncertainties.

Acknowledgements. The study was co-funded by the Italian Operational Project (PON)—
Research and Competitiveness for the convergence regions 2007/2013—I Axis “Support to
structural changes” operative objective 4.1.1.1. “Scientific-technological generators of transfor-
mation processes of the productive system and creation of new sectors” Action II: “Interventions
to support industrial research”.

References
1. Zhang, D.L., Shou, Y.X., Dickerson, R.R.: Upstream urbanization exacerbates urban heat
island effects. Geophys. Res. Lett. 36(24), 1–5 (2009)
2. Haase, D.: Effects of urbanisation on the water balance–A long-term trajectory. Environ.
Impact Assess. Rev. 29(4), 211–219 (2009)
3. Jacob, D.J., Winner, D.A.: Effect of climate change on air quality. Atmos. Environ. 43(1),
51–63 (2009)
4. Piro, P., et al.: Flood risk mitigation in a Mediterranean urban area: the case study of
Rossano Scalo (CS – Calabria, Italy). In: Mannina, G. (ed.) UDM 2018. GET, pp. 339–343.
Springer, Cham (2019). https://doi.org/10.1007/978-3-319-99867-1_57
594 B. Pirouz et al.

5. Miller, J.D., Hutchins, M.: The impacts of urbanisation and climate change on urban
flooding and urban water quality: a review of the evidence concerning the United Kingdom.
J. Hydrol.: Reg. Stud. 12, 345–362 (2017)
6. Piro, P., Turco, M., Palermo, S.A., Principato, F., Brunetti, G.: A comprehensive approach to
stormwater management problems in the next generation drainage networks. In: Cicirelli, F.,
Guerrieri, A., Mastroianni, C., Spezzano, G., Vinci, A. (eds.) The Internet of Things for
Smart Urban Ecosystems. IT, pp. 275–304. Springer, Cham (2019). https://doi.org/10.1007/
978-3-319-96550-5_12
7. Raimondi, A., Becciu, G.: On pre-filling probability of flood control detention facilities.
Urban Water J. 12, 344–351 (2015)
8. Zischg, J., Rogers, B., Gunn, A., Rauch, W., Sitzenfrei, R.: Future trajectories of urban
drainage systems: a simple exploratory modeling approach for assessing socio-technical
transitions. Sci. Total Environ. 651, 1709–1719 (2019)
9. Fletcher, T.D., et al.: SUDS, LID, BMPs, WSUD and more–The evolution and application of
terminology surrounding urban drainage. Urban Water J. 12(7), 525–542 (2015)
10. Razzaghmanesh, M., Beecham, S., Salemi, T.: The role of green roofs in mitigating Urban
Heat Island effects in the metropolitan area of Adelaide, South Australia. Urban For. Urban
Green. 15, 89–102 (2016)
11. Maiolo, M., Carini, M., Capano, G., Piro, P.: Synthetic sustainability index (SSI) based on
life cycle assessment approach of low impact development in the Mediterranean area.
Cogent Eng. 4(1), 1410272 (2017)
12. Zahmatkesh, Z., Burian, S.J., Karamouz, M., Tavakol-Davani, H., Goharian, E.: Low-impact
development practices to mitigate climate change effects on urban stormwater runoff: case
study of New York City. J. Irrig. Drain. Eng. 141(1), 04014043 (2014)
13. Jia, H., Yao, H., Shaw, L.Y.: Advances in LID BMPs research and practice for urban runoff
control in China. Front. Environ. Sci. Eng. 7(5), 709–720 (2013)
14. Piro, P., Carbone, M., Morimanno, F., Palermo, S.A.: Simple flowmeter device for LID
systems: from laboratory procedure to full-scale implementation. Flow Meas. Instrum. 65,
240–249 (2019)
15. Turco, M., Brunetti, G., Carbone, M., Piro, P.: Modelling the hydraulic behaviour of
permeable pavements through a reservoir element model. In: International Multidisciplinary
Scientific Geo Conference: SGEM: Surveying Geology & mining Ecology Management,
vol. 18, pp. 507–514 (2018)
16. Eckart, K., McPhee, Z., Bolisetti, T.: Performance and implementation of low impact
development–a review. Sci. Total Environ. 607, 413–432 (2017)
17. Palermo, S.A., Zischg, J., Sitzenfrei, R., Rauch, W., Piro, P.: Parameter sensitivity of a
microscale hydrodynamic model. In: Mannina, G. (ed.) UDM 2018. GET, pp. 982–987.
Springer, Cham (2019). https://doi.org/10.1007/978-3-319-99867-1_169
18. Jia, H., et al.: Field monitoring of a LID-BMP treatment train system in China. Environ.
Monit. Assess. 187(6), 373 (2015)
19. Zhang, G., Hamlett, J.M., Reed, P., Tang, Y.: Multi-objective optimization of low impact
development designs in an urbanizing watershed. Open J. Optim. 2, 95–108 (2013)
20. Pawlak, Z.: Rough set theory and its applications. J. Telecommun. Inf. Technol. 3, 7–10
(2002)
21. Arabani, M., Sasanian, S., Farmand, Y., Pirouz, M.: Rough-set theory in solving road
pavement management problems (case study: Ahwaz-Shush Highway). Comput. Res. Prog.
Appl. Sci. Eng. (CRPASE) 3(2), 62–70 (2017)
22. Arabani, M., Pirouz, M., Pirouz, B.: Geotechnical investigation optimization using rough set
theory. In: 9th International Congress on Civil Engineering (9ICCE), Isfahan, Iran (2012)
 New Mathematical Optimization Approaches for LID Systems 595

23. Hwang, C.L., Yoon, K.P.: Multiple Attributes Decision-Making Methods and Applications.
Springer, Berlin (1981). https://doi.org/10.1007/978-3-642-48318-9
24. Haghshenas, S.S., Neshaei, M.A.L., Pourkazem, P., Haghshenas, S.S.: The risk assessment
of dam construction projects using fuzzy TOPSIS (case study: Alavian Earth Dam). Civ.
Eng. J. 2(4), 158–167 (2016)
25. Balioti, V., Tzimopoulos, C., Evangelides, C.: Multi-criteria decision making using TOPSIS
method under fuzzy environment, application in spillway selection. In: Multidisciplinary
Digital Publishing Institute Proceedings, vol. 2, p. 637 (2018)
26. İç, Y.T.: A TOPSIS based design of experiment approach to assess company ranking. Appl.
Math. Comput. 227, 630–647 (2014)
27. Krohling, R.A., Pacheco, A.G.: A-TOPSIS an approach based on TOPSIS for ranking
evolutionary algorithms. Procedia Comput. Sci. 55, 308–317 (2015)
28. Haghshenas, S.S., Mikaeil, R., Haghshenas, S.S., Naghadehi, M.Z., Moghadam, P.S.: Fuzzy
and classical MCDM techniques to rank the slope stabilization methods in a rock-fill
reservoir dam. Civ. Eng. J. 3(6), 382–394 (2017)
 Evaluation of an Integrated Seasonal
Forecast System for Agricultural Water
Management in Mediterranean Regions

Alfonso Senatore1(B) , Domenico Fuoco1 , Antonella Sanna2 ,

Andrea Borrelli2 , Giuseppe Mendicino1 , and Silvio Gualdi2
1
University of Calabria, Rende, CS, Italy
[email protected]
2
Euro-Mediterranean Center on Climate Change, Bologna, Italy
[email protected]

Abstract. The Euro-Mediterranean Center on Climate Change (CM-

CC) seasonal forecasting system, based on the global coupled model
CMCC-CM, performs seasonal forecasts every month, producing several
ensemble integrations conducted for the following 6 months. In this study,
a performance evaluation of the skills of this system is performed in two
neighbouring Mediterranean medium-small size catchments located in
Southern Italy, the Crati river and the Coscile river, whose hydrological
cycles are particularly important for agricultural purposes.
Initially, the performance of the system is evaluated comparing
observed and simulated precipitation and temperature anomalies in the
irrigation periods of the years 2011–2017. Forecasts issued on April 1st
(i.e., at the beginning of the irrigation period) are evaluated, consider-
ing two lead times (first and second trimester). Afterward, the seasonal
forecasts are integrated into a complete meteo-hydrological system. Pre-
cipitation and temperature provided by the global model are ingested in
the spatially distributed and physically based In-STRHyM (Intermedi-
ate Space-Time Resolution Hydrological Model) model, which analyzes
the hydrological impact of the seasonal forecasts.
Though the predicted precipitation and temperature anomalies are
not highly correlated with observations, the integrated seasonal forecast
for the hydrological variables provides significant correlations between
observed and predicted anomalies, especially concerning mean discharge
(>0.65). Overall, the system showed to provide useful insights for agri-
cultural water management in the study area.

Keywords: Meteo-hydrological system · Seasonal forecast ·

Agricultural water management

1 Introduction
Filling the so-called subseasonal to seasonal (S2S) forecast gap is a challenging
issue for both weather and climate science communities. Nevertheless, improv-
c Springer Nature Switzerland AG 2020
Y. D. Sergeyev and D. E. Kvasov (Eds.): NUMTA 2019, LNCS 11973, pp. 596–603, 2020.
https://doi.org/10.1007/978-3-030-39081-5_51
 An Integrated Seasonal Forecast System in Mediterranean Regions 597

ing meteorological forecast skill up to leads out to one or more months por-
tends invaluable social and economic benefits, particularly in the field of water
resources management. Therefore, several initiatives have been running in the
last years at the international level, like the joint World Weather Research Pro-
gram (WWRP) and World Climate Research Program (WCRP) S2S Project.
Furthermore, since more than a decade many operational centers produce rou-
tine dynamical seasonal forecasts [1]. For example, in the US the North Amer-
ican Multi-Model Ensemble (NMME) real-time forecasts are incorporated as
part of National Oceanic and Atmospheric Administration’s (NOAA) opera-
tional forecast suite. In Europe, several operational forecast systems contribute
to the Copernicus Climate Change Service (C3S). Among them, the Euro-
Mediterranean Center on Climate Change (CMCC) seasonal forecast system
[2] performs seasonal forecasts every month, producing a number of ensemble
integrations conducted with the coupled model for the following 6 months. The
skill of the CMCC forecast system has been thoroughly tested from decadal [3] to
seasonal lead times, evaluating different aspects like the impact of initial condi-
tions [4] or the capability to represent the extratropical low-frequency variability
[5,6].
The usefulness of seasonal forecasts for practical applications is particularly
highlighted when they are coupled to impact models, dealing with various aspects
of the terrestrial hydrological cycle. This kind of integrated systems can provide
timely, clear and useful information to support policy decisions, with obvious
socio-economic benefits.
Unlike many meteorological variables (e.g., precipitation), time series data
of hydrological variables like river discharge or aquifer levels are subjected to a
much higher autocorrelation. The hydrologic persistence is a well known phe-
nomenon [7] and has been widely exploited for predicting catchments seasonal
behavior only with statistical tools, not necessarily relying on weather forecasts
(e.g., [8]). Nevertheless, recent advances in seasonal forecasts have allowed statis-
tical models to be profitably combined with seasonal forecasts (hybrid statistical-
dynamical forecasts) and, in several cases, to directly couple seasonal models to
cropping system models for predicting crop yield [9] or to hydrological models
for streamflow forecast (e.g., [10,11]), aimed either at overall water resources
management or at specific objectives, e.g., snow accumulation forecast, reservoir
operations or inland waterway transport management.
In this study, the performance evaluation of a forecasting chain based on
the CMCC seasonal forecast system and the spatially distributed, physically
based In-STRHyM (Intermediate Space-Time Resolution Hydrological Model)
model [12] was performed in two neighbouring Mediterranean medium-small size
catchments located in Southern Italy, where water management (especially in the
irrigation summer period) is related to remarkable agricultural activities. The
evaluation concerned the irrigation periods of the years 2011–2017, for which the
seasonal forecasts are integrated in a complete meteo-hydrological system aimed
at providing useful indications for agricultural water management in the area.
598 A. Senatore et al.

Next Sect. 2 deals with the description of the study area, the seasonal forecast
system, the hydrological model and the methodology followed for the experiment
setup, while Sect. 3 briefly summarizes the preliminary results. Finally, conclu-
sions (Sect. 4) will sketch out future steps for improving the forecast skill of the
integrated system.

2 Data and Methods

2.1 Study Area

The study area includes two catchments in Calabria (southern Italy), namely the
Crati river catchment closed at the Santa Sofia gauge station and the Coscile
river catchment closed at the Cammarata gauge station (1281 km2 and 405 km2 ,
respectively; Fig. 1). These catchments are both sub-basins of the Crati river
basin, the biggest in Calabria, and cover about 70% of its extent. Their hydro-
logical cycles are particularly important for agricultural purposes, because they
feed the Sibari plain, the main agricultural area of a region which is often subject
to drought [13].
The mean altitude of the Crati river catchment is 672 m a.s.l. (ranging from
1856 m a.s.l. on the Sila plateau to 67 m a.s.l.), while for the Coscile river catch-
ment is 698 m a.s.l. (from 2248 m a.s.l. on the Pollino Massif to 82 m a.s.l.). The
whole Crati river basin has a Mediterranean sub-humid climate. The real-time
monitoring network managed by the “Centro Funzionale Multirischi” of the Cal-
abrian Regional Agency for the Protection of the Environment provides about 15
temperature and 20 precipitation recording stations within the catchments. In
the analyzed period 2011–2017, the mean annual accumulated precipitation val-
ues for the Crati river catchment and the Coscile river catchment were 1196 mm
and 1097 mm (64.2% and 64.1% occurring between October and March), and
the mean annual temperatures were 13.0 ◦ C and 13.3 ◦ C, respectively.

2.2 The CMCC Seasonal Forecasting System

The CMCC-SPS3 is the Seasonal Prediction System developed at CMCC to

perform seasonal forecasts operationally. It is based on the CMCC-CM2 coupled
model, which consists of several independent model components, simulating at
the same time the Earth’s atmosphere, ocean, land, river routing and sea-ice.
In the SPS implementation, the number of vertical levels is increased from
30 to 46 in order to better resolve (up to 0.3 hPa) the stratosphere, which recent
evidences show to be of greater importance in the seasonal forecast framework
[14,15]. Horizontal resolution is about 1×1 degree for the atmosphere and land,
0.5×0.5 for river routing and 0.25×0.25 for ocean and sea-ice.
In order to take account of the uncertainty associated with the initial condi-
tions, 50 perturbations of the initial state are built. Perturbations are generated
by combining the initial states of the three main components: the atmosphere
(10 perturbations of the initial state), the ocean (8 perturbations) and the land
 An Integrated Seasonal Forecast System in Mediterranean Regions 599

surface (3 perturbations). Out of the possible 240, 50 unique combinations are

randomly chosen to initialize the CMCC-SPS3 each month.
The complete system has been run in retrospective forecast way over the
24-year-period 1993–2016 (hindcast period), with a slightly reduced ensemble
population of 40 members. The results of the hindcasts were used to construct a
reference climatology. Forecasts are presented in terms of anomalies with respect
to the reference climatology of the considered season (to take into account also
model climate drift). Because the model systematic error (that is the bias with
respect to observed climate) is one of the main sources of uncertainties, removing
the climatology allows analysing seasonal tendencies, where SPS exhibits positive
skills in reproducing the observed variability.
The CMCC-SPS3 contributes to C3S climate service, together with other
four providers (ECMWF, DWD, MeteoFrance and MetOffice), in the multi-
model seasonal forecast system and it is also a recognized WMO Global Pro-
ducing Center for Long-Range Forecasts. More details on model initialization,
performances, skills and shortcomings can be found in [2].

Fig. 1. Study area. The right panel zooms the framed area in the left panel: the whole
Crati river basin is shown (black contours), together with the Crati river catchment
closed at the Santa Sofia gauge station (yellow) and the Coscile river catchment closed
at the Cammarata gauge station (orange) (Color figure online)

2.3 The In-STRHyM Hydrological Model

The original distributed hydrological model In-STRHyM [12] has been rewrit-
ten in Python and integrated in the QGIS environment as a suitable tool both
for analysis over long periods, such as climate change scenarios, and for opera-
tional monitoring and subseasonal to seasonal forecasts of water resource avail-
ability (drought risk). In-STRHyM estimates the hydrological balance using a
physically-based modeling approach that explicitly simulates the hydrological
processes of canopy interception, snow accumulation and melting, actual evapo-
transpiration, root zone infiltration and percolation towards groundwater, runoff
generation due to either infiltration or saturation excess, surface and subsurface
600 A. Senatore et al.

(1D) runoff routing and baseflow (2D) routing. Being typically applied with a
time resolution of 1 day and a spatial resolution of 1 km2 , In-STRHyM allows
the distributed estimate of the main components of the hydrological balance
together with the discharge in selected outlets.
By default the model input is totally distributed, even if in the absence of
data some parameters can be assumed constant in the catchment. For the val-
idation of the In-STRHyM model in Calabria, the parameters relating to the
topographical features, the soil hydraulic properties, the depth of the root zone
and the soil uses were spatially distributed. Regarding the meteorological input,
daily accumulated precipitation and mean, minimum and maximum tempera-
ture grids are required (in the In-STRHyM model version used in this study, a
simplified method has been adopted for estimating the reference evapotranspi-
ration [16]). Furthermore, for the purpose of actual evapotranspiration estimate,
remote sensing maps (MODIS) are used providing space-time distributed infor-
mation of the NDVI (Normalized Difference Vegetation Index) and LAI (Leaf
Area Index) vegetation indices. The model was calibrated on the Crati river
catchment from October 1961 to December 1966 (Nash-Sutcliffe coefficient E2
= 0.83), then validated with more recent available discharge observations [12].

2.4 Experiment Setup

The performance evaluation concerned seven consecutive irrigation seasons (6

months, from April to September) from 2011 to 2017. The 40-member ensem-
ble hindcasts (from 2011 to 2016) and forecast (2017) issued by CMCC-SPS3
on April 1st of each of these years for the following 6 months were evaluated
comparing observed and simulated precipitation and temperature anomalies. In
this preliminary evaluation, the reference period for calculating the anomalies is
2011–2016, both for observations and hindcasts/forecast.
Afterwards, the predictive skill of the integrated system was assessed against
the hydrological impact. The In-STRHyM model was first run for the two catch-
ments from October 1st 2009 (thus ensuring more than one year of spin-up time)
to September 30th 2017, using as meteorological input the daily precipitation and
temperature grids achieved by the spatial interpolation of the available obser-
vations. Then, the state of the hydrological model on April 1st of each of the
years 2011–2017 was imposed as the initial condition for the hydrological fore-
cast in the same year. The meteorological driver for the following six months was
obtained perturbing the mean daily grids of observed precipitation and temper-
atures with the anomalies (i.e., differences from the average values) calculated at
the monthly time scale for the 40 members of the ensemble forecast. Finally, the
anomalies of the main output variables of the hydrological model calculated using
the observed meteorological input were compared with those achieved through
the ensemble seasonal forecasts.
 An Integrated Seasonal Forecast System in Mediterranean Regions 601

3 Preliminary Results
The seasonal forecasts carried out by CMCC-SPS3 do not exhibit significant skill
with the investigated meteorological variables. The correlation between observed
three-month accumulated precipitation anomalies and the corresponding average
values of the predicted anomalies during the 7 analyzed years is rather low for
both lead 0 (i.e., months of April, May and June - AMJ, Fig. 2a) and lead 1
(i.e., months of July, August and September - JAS, Fig. 2b). Similar results were
achieved comparing temperatures (Figs. 2c,d). Specifically, the model did not
seem able to correctly predict the intense summer drought occurred in 2017 in
the analyzed area (and in most of southern Italy). These results are affected, at
least partially, by the too short reference period used for calculating anomalies
and, mainly, by the reduced extent of the analyzed area (for the analysis of the
meteorological variables 11 grid points were used).

Fig. 2. Precipitation (P) and mean temperature (Tmed) anomalies achieved using
observations and SPS predictions for the Crati river catchment. In the box-and-whiskers
plots, the first and third quantiles are represented by the lower and upper edges of each
box, respectively, while the horizontal thin bar in the box is the median value. The tip
of the upper (lower) whisker represents the maximum (minimum) value

Hydrological variables provided more promising results, thanks also to their

higher persistence. Figures 3 and 4 show the monthly evolution of the observed
and predicted actual evapotranspiration (ET) and mean discharge (Qmed ) for the
Crati river catchment in the summer period (from April to September) from 2011
to 2017, highlighting a general slight underestimation of predictions. However,
to evaluate the system performance, in agreement with the SPS approach, it is
worth it to focus on the observed and predicted anomalies. ET largely depends on
predicted temperature and is relatively unaffected by the catchment “memory”:
in the AMJ period, the correlation between observed anomalies and the medians
of the predicted anomalies is –0.21, but almost in all cases the deviations from
zero are very low, such as the variances of the forecasts. In JAS, the variance
of the predicted anomalies is higher, but the median is better correlated with
observations (+0.61).
On the other hand, Qmed is a much more persistent variable and the pre-
dicted precipitation, especially in the dry summer period, can only relatively
affect mean streamflow characteristics. For the Crati river catchment, anomalies
correlation is high both in AMJ (+0.66) and, quite surprisingly, in JAS (+0.95).
602 A. Senatore et al.

Fig. 3. Monthly averaged daily ET evolution for each year from 2011 to 2017 achieved
using observed and predicted meteorological drivers for the Crati river catchment

Fig. 4. Same as Fig. 3, but for Qmed

Such promising results highlight the potential of the proposed approach for water
resources seasonal planning and management, and pave the way for further more
detailed analyses.

4 Conclusions

An integrated meteo-hydrological seasonal forecast system in a Mediterranean

region is described in this paper, together with a preliminary evaluation of its
performance. First results highlight both weak points and interesting insights.
Further investigation is required, future work will be directed towards several
aspects, e.g.: improving both the reference and application periods for calculating
anomalies; enlarging the ensemble members and the models involved; testing
dynamical downscaling of the seasonal forecasts, possibly with two-way coupled
atmospheric-hydrological models [17]. These analyses will be addressed to exploit
at its best the potential of the seasonal forecast systems, to reduce forecast
uncertainty and to concur to the long-term objective of developing valuable
tools for supporting water resources management.
 An Integrated Seasonal Forecast System in Mediterranean Regions 603

References
1. National Academies of Sciences: Engineering, and Medicine: Next Generation
Earth System Prediction: Strategies for Subseasonal to Seasonal Forecasts. The
National Academies Press, Washington, DC (2016)
2. Sanna, A., Borrelli, A., Athanasiadis, P., et al.: The CMCC Seasonal Prediction
System. CMCC Research Paper 285 (2017)
3. Bellucci, A., Haarsma, R., et al.: An assessment of a multi-model ensemble of
decadal climate predictions. Clim. Dyn. 44, 2787–2806 (2015)
4. Materia, S., Borrelli, A., Bellucci, A., et al.: Impact of atmosphere and land sur-
face initial conditions on seasonal forecasts of global surface temperature. J. Clim.
27(24), 9253–9271 (2014)
5. Athanasiadis, P.J., et al.: The representation of atmospheric blocking and the asso-
ciated low-frequency variability in two seasonal prediction systems. J. Clim. 27(24),
9082–9100 (2014)
6. Athanasiadis, P.J., et al.: A multisystem view of wintertime NAO seasonal predic-
tions. J. Clim. 30(4), 1461–1475 (2017)
7. Koutsoyiannis, D.: Hydrologic persistence and the hurst phenomenon. In: Lehr,
J.H., Keeley, J. (eds.) Water Encyclopedia. Surface and Agricultural Water, vol.
4. Wiley, New York (2005)
8. Mendicino, G., Senatore, A., Versace, P.: A Groundwater Resource Index (GRI)
for drought monitoring and forecasting in a Mediterranean climate. J. Hydrol.
357(3–4), 282–302 (2008)
9. Jha, P.A., Athanasiadis, P., Gualdi, S., et al.: Using daily data from seasonal
forecasts in dynamic crop models for yield prediction: a case study for rice in
Nepal’s Terai. Agr. Forest Meteorol. 265, 349–358 (2019)
10. Candogan Yossef, N., van Beek, R., Weerts, A., Winsemius, H., Bierkens, M.F.P.:
Skill of a global forecasting system in seasonal ensemble streamflow prediction.
Hydrol. Earth Syst. Sci. 21, 4103–4114 (2017)
11. Bell, V.A., Davies, H.N., Kay, A.L., Brookshaw, A., Scaife, A.A.: A national-scale
seasonal hydrological forecast system: development and evaluation over Britain.
Hydrol. Earth Syst. Sci. 21, 4681–4691 (2017)
12. Senatore, A., Mendicino, G., Smiatek, G., Kunstmann, H.: Regional climate change
projections and hydrological impact analysis for a Mediterranean basin in Southern
Italy. J. Hydrol. 399, 70–92 (2011)
13. Maiolo, M., Mendicino, G., Pantusa, D., Senatore, A.: Optimization of drinking
water distribution systems in relation to the effects of climate change. Water 9(10),
83 (2017)
14. Marshall, A.G., Scaife, A.A.: Impact of the QBO on surface winter climate. J.
Geophys. Res.-Atmos. 114, D18110 (2009)
15. Doblas-Reyes, F.J., Garcı́a-Serrano, J., Lienert, F., Biescas, A.P., Rodrigues,
L.R.L.: Seasonal climate predictability and forecasting: status and prospects.
WIREs Clim. Change 4(4), 245–268 (2013)
16. Mendicino, G., Senatore, A.: Regionalization of the hargreaves coefficient for the
assessment of distributed reference evapotranspiration in southern Italy. J. Irrig.
Drainage Eng. 139(5), 349–362 (2013)
17. Senatore, A., Mendicino, G., Gochis, D.J., Yu, W., Yates, D.N., Kunstmann, H.:
Fully coupled atmosphere-hydrology simulations for the central Mediterranean:
Iimpact of enhanced hydrological parameterization for short- and long-timescales.
JAMES 7(4), 1693–1715 (2015)
 Optimization of Submarine Outfalls
with a Multiport Diffuser Design

Salvatore Sinopoli, Marco Amos Bonora, Gilda Capano, Manuela Carini ,

Daniela Pantusa , and Mario Maiolo(B)

Department of Environmental and Chemical Engineering (DIATIC), Università della

Calabria, Cubo 42/B, 87036 Arcavacata di Rende, CS, Italy
{salvatore.sinopoli,marcoamos.bonora,gilda.capano,manuela.carini,
daniela.pantusa,mario.maiolo}@unical.it

Abstract. Immission of civil sewage into sea is realized to complete

the onshore depurative process or to take out the already purified waste-
water from the bathing area, ensuring a good perceived seawater quality.
Anyhow the compliance of the pollutant concentrations limits is neces-
sary to ensure safe bathing. The design of submarine pipes is usually
completed contemplating a diffuser with a series of ports for the reparti-
tion of the wastewater discharge. The real process of pollutants diffusion
into the sea, simulated with complex diffusion-dispersion models in a
motion-field dependent from environmental conditions and drift speeds,
affect the submarine pipe design. A design optimization procedure has
been realized for the marine outfall pipe-diffuser system using a simpli-
fied zone model, subjected to a sensitivity analysis on the characteristic
parameter. The method is shown using an example project for the sub-
marine outfall at service for the sewage treatment plant of Belvedere
Marittimo, on the southern Tyrrhenian Sea in Italy.

Keywords: Sea water quality · Submarine outfalls · Optimization

design · Multiport diffuser

1 Introduction
The increase in water demand and the need for environmental protection require
proper management of wastewater which is a scientific and engineering topic of
current interest [12,14,15,19]. Regarding marine outfall systems, they have long
been used for the discharge of industrial and domestic effluent as a means of
increasing its dilution and improving the assimilative capacity of, the receiving
environment [5]. The submarine outfall is a hydraulic structure that has the
purpose of discharge in the receiving marine water body the wastewater after
the wastewater treatment plant. This structure consists of onshore headwork, a
feeder pipeline, and a diffuser section [9]. A pumping station can installed into
the onshore headwork if the effluent discharge cannot take place by gravity, and
the most efficient diffuser design is a set of ports whereby effluent diffuses into
c Springer Nature Switzerland AG 2020
Y. D. Sergeyev and D. E. Kvasov (Eds.): NUMTA 2019, LNCS 11973, pp. 604–618, 2020.
https://doi.org/10.1007/978-3-030-39081-5_52
 Optimization of Submarine Outfalls with a Multiport Diffuser Design 605

the submarine environment [10,18]. The characteristics of the effluent must be

established taking into account volume, flow variations, and constituents con-
centrations. These concentrations are established by current regulations. The
design of the submarine outfall system must keep into consideration hydraulic
aspects, impact on the natural environment and economic criteria. The most
efficient hydraulic design is influenced by a uniform flow distribution through all
the ports and by the initial dilution, taking into account the necessity of avoiding
saline intrusion into the outfall and preventing scour and sedimentation in the
surrounding area [5].
Although these systems are mature and in use for many years, advancements
in computational power increase, numerical algorithms, and optimization models
allow for improved outfalls design by optimizing their performance.
In this context, this paper describes an optimization procedure that has been
realized for the submarine outfall pipe-diffuser system using a simplified zone
model, subjected to a sensitivity analysis on the characteristic parameter.
The construction of a submarine outfall system is a problem to be faced in
different ways. If on the one hand there is the need to respect the environmental
parameters, there is at the same time the need to design a diffusion system
that can ensure the discharge of the wastewater below the average sea level, and
thus guarantee the correct dilution. Furthermore, sizing must be carried out also
considering the economic criteria, given that the construction is not that of a
classic underground pipe. The proposed optimization procedure, therefore, tries
to take into account all these decision-making aspects, thus seeking a solution
that can satisfy all three of the aforesaid requirements.
Is then proposed a design criterion that first of all checks the functioning of
the hydraulic sizing, and therefore provides the design parameters of the pipeline
to a dilution model of the pollutant, used to verifies the diffuser’s capacity to
operate under the environmental aspect. From the optimization point of view,
the chosen pipeline will be the one that reaches the two previous results with
the lowest cost in terms of materials.
The optimization procedure and the sensitivity analysis carried out are
described below, applying it to the case study of the submarine outfall of
Belvedere Marittimo, in the south of Italy.

2 Materials and Methods

2.1 Hydraulic Dimensioning
The hydraulic dimensioning of this type of works is very complex. If the only
criterion to be respected in the feeder pipeline is velocity (not exceeding a few
m/s, and not much lower than 1 m/s to avoid sedimentation problems), for the
section of the diffuser the correct sizing is necessary to respect the hydraulic
operation and also to allow the complete formation of the buoyant jets.
To avoid the saline intrusion into the ports the densimetric Froude number
must be greater than 1, and for the full mouth operation, it is necessary that
the ratio between total ports area and pipe area is about 0.5:
606 S. Sinopoli et al.

A
 ports = 0, 5 (1)
AP ipe

Since their introduction in the 1950s, marine outfalls with diffusers have been
prone to saline intrusion. Saline intrusion is likely to reduce the efficiency of the
outfall and to inhibit the discharge of wastewater through long sea outfalls.
Operational difficulties arising from saline intrusion in multi-port sea outfalls
are mainly related to seawater corrosion of the pipework and reduced dilution.
The problem associated with sea water intrusion into sea outfalls has been high-
lighted in a number of papers and over the years several studies on physical
and numerical models have been carried out on this topic and its implications
[20,28,29]. Some studies have instead focused on measures to ensure rapid dilu-
tion of effluent with sea water in a wide range of prevailing conditions. Possible
design for pre-diluition devices have been studied and experiments have been
carried out to test the efficiency of the devices in increasing diluition in different
water depths [1].
The additional design criterion to be respected is to try to have a constant
flow rate for all the diffuser ports; this criterion is difficult to respect as the last
diffuser mouths are reached by a lower flow rate and are subject to a greater
head. By accepting the variability of the flow rates, the hydraulic design criterion
becomes that of identifying the head necessary for the operation of each mouth
of the diffuser.

Fig. 1. Hydraulic scheme of the diffuser.

The provided criterion solves the problem taking into account the head losses
located in the individual ports of the diffuser, and that starting from the one
upstream can guarantee a practically constant outflow. The procedure then
starts from the upstream mouth and considers a pipe with a fixed diameter
Dc, along which a number of circular ports of equal and fixed diameter D are
 Optimization of Submarine Outfalls with a Multiport Diffuser Design 607

made, and placed at equal distance between them Δs, so that the turbulences
produced by the single jet does not affect on the nearest ones.
The discharge with density ρ0 coming out of the i − th mouth is regurgitated
under a head Yi of sea water with density ρ, and can be expressed as:

Qi = Ai Cdi (2gHi ) (2)

where Cdi is an experimental outflow coefficient that can be derived from the
curve obtained by Olivotti [24] as a function of the ratio between the kinetic
head on the diffuser upstream of the port and the total head on the port itself
(Fig. 2).

Fig. 2. Hydraulic scheme of the diffuser.

The calculation method consists in fixing a head value on the first mouth,
using the value Cdi obtained from the graph and identifying with the Eq. (2)
the value of flow rate spill.
At this point it is possible to calculate the head losses distributed in the
section of pipe necessary to reach the next port, placed at a distance Δs, con-
sidering that the initial flow must be subtracted from the outflow to the first
port: 
Qk−i = Qinitial − Qports (3)
608 S. Sinopoli et al.

The value of the head on the next mouth is expressed as:

 
Δρ
Hk = Hi − (Ji Δs ) − ρ Δ (4)
ρ0
The subsequent iterations will then lead to determining the outflow from
each mouth, stopping the sizing when the outflow will be approximately equal
to that coming from the treatment plant.
However, the total flow rate may not be coincident, due to the arbitrariness
of the head chosen on the first port, and therefore does not allow a real solution.
Leaving the parameters of the pipeline unchanged (diameter and material,
which also determine the head losses) and the size of the ports, the distance
from the coast can be varied, consequently varying the depth of the diffuser
(taken into account like a linear function of sea-bed slope and distance from the
shoreline) and the head on the first port, until the flow rates coincide, and thus
obtaining the correct hydraulic operation, for the diameter of the pipe Dc and
for the diameter of the ports D both initially fixed, and thus identifying the
number of diffuser ports that must be provided to dispose the flow rate.

2.2 Verification of Environmental Parameters

The fulfillment of environmental parameters is a procedure based on a zone
model [4,21,24,27], a simplified approach that takes into account both the engi-
neering parameters related to hydraulic verification and environmental variables
such as the velocity of the current or the bacterial decay time.
The model used is the one proposed by [21,24] which takes into account
three phenomena that lead to the dilution of the wastewater, and therefore to
the reduction of concentration.
The total dilution of the wastewater can be calculated as:
C0
Sc = (5)
Ct
with C0 is the concentration of wastewater coming from the treatment plant and
Ct is the concentration near the coast.
The first contribution is given by the so-called Initial Dilution, Si, a process
extensively studied over the years both theoretically and experimentally [2,3,10,
17,24].
According to these studies the concentration of pollutant assumes a Gaussian
distribution along the axis of the jet, and the phenomenon of Initial Dilution Si
Y
can be expressed as a function of two dimensionless parameters e F , where
D
F is the densimetric Froude number, D is the diameter of the ports and Y is
the distance between the jet release point and the average sea surface.
In the present model the formulation of Cederwall [8] is used, which has
already been successfully tested in Italy [22,26]:
  35
Y 1
Si = 0.54 F 0.38 + 0.66 (6)
D F
 Optimization of Submarine Outfalls with a Multiport Diffuser Design 609

Fig. 3. Schematic representation of the buoyant jet in a quiet environment.

To apply the model to diffusers with multiple ports, the possible mutual
interference between two adjacent jets must be taken into account. According to
[3] it is possible to approximate the width of the jet as a function of its height,
provided that the number of densimetric Froude is in the range 1–20. For this
reason, it is possible to impose that the distance Δs between two ports is equal
Y
to (Fig. 4), in which Y is again evaluated like the product of sea-bed slope
3
and distance from the shoreline.
The second contribution to the reduction of concentration is given by the Sub-
sequent Dilution, SS, related to the advective-diffusive phenomena and depen-
dent on the velocity of the current, u. Also, in this case, the assumption is that
the plume can be represented by a Gaussian type distribution, and with a com-
plete vertical mixing [13] (Fig. 5).
Considering a plane flow, and velocity of the advancement of the plume
practically constant to depth, it is possible to assume a constant horizontal
dispersion coefficient according to Pearson’s law:
 1.333333
b
∈= 0.01 · (7)
10

where b is the total width of the diffusor.

610 S. Sinopoli et al.

Fig. 4. Schematic representation of the buoyant jet.

Fig. 5. Schematic representation of the far-field dispersion process.

 Optimization of Submarine Outfalls with a Multiport Diffuser Design 611

Considering these assumptions the Subsequent Dilution can be calculated

with the formulation proposed by [6,7] as:
⎡ ⎤−1


⎢  1.5
⎥
Ss = ⎣erf   3 ⎦ (8)
1+ 0.67β xb −1

where β is a dimensionless parameter that recalls ∈, and it is evaluated as:

∈
β = 12 (9)
ub
where u is the average velocity of the current incident on the diffuser.
The third dilution contribution comes instead from the bacterial decay Sb ,
due to both physical and chemical phenomena, and which can therefore be
described by the well-known Chick’s law, in which the concentration variation is
a function of the drift period, and a parameter, ke , related to reaction kinetics:
dC
= −ke · C (10)
dt
By integrating this equation:
C0
Sc = = Si · Ss · Sb (11)
Ct
Consequently:
C0 ke ·t
Sb = = 10 2.3 (12)
Si · Ss · Ct
where considering:
2.3
= T90 (13)
ke
finally obtain:
t
Sb = 10 T90 (14)
where T90 is the time necessary for the disappearance of 90% of the bacteria,
and t is the ratio between the distance x that separates the diffuser from the
coastline and the average velocity of the current u.
Once the three dilution values are known, it is, therefore, possible to calculate
the total dilution value and then evaluate the concentration of coliforms near
the coastline as:
C0 C0
Ct = = (15)
Sc Si · Ss · Sb
612 S. Sinopoli et al.

2.3 Optimization Procedure

Due to the non-continuous nature of the variables relating to the diameters of

the pipeline and the ports, the solution to the problem cannot be evaluated with
a simple minimum law.
It has been necessary to identify a logical procedure that hallows to evaluate
all the possible solutions, first from the hydraulic point of view, and subsequently
from the environmental point of view.
To do this, starting on the base offered by [22] once a pipe diameter has been
set, a calculation code has been implemented that sets a diameter for the pipe
and varies the diameters of the ports, like in example scheme in Fig. 6.

Fig. 6. Schematic representation of the algorithm logical process.

At each iteration, once the pair of diameters has been set, the program pro-
ceeds by setting a distance value from the coast. Based on the diameter and
distance, the pipe head losses are evaluated, and the flow rate by the diffuser is
checked.
If the flow is less than that coming from the treatment plant, the number
of ports is increased and the check is repeated. If the flow rate is too large, the
distance to the coast is increased and the flow rate check is repeated starting
from the minimum number of ports. If, on the other hand, the flow rate is less
than a tolerance value equal to that coming from the plant, an environmental
check is performed.
 Optimization of Submarine Outfalls with a Multiport Diffuser Design 613

If the environmental verification is not passed, the distance from the coast is
increased and the number of ports is searched. If the environmental verification
is exceeded, the solution is saved, and then, in this case, too the value of the
distance from the coast is increased.
When the maximum value allowed for the distance from the coast has been
reached, it means that all the possible solutions for the diameter of the previously
set port have been evaluated and therefore it is increased.
The algorithm then proceeds with the new pair of diameters, sharing the
minimum distance from the coast to evaluate other possible solutions.

3 Application and Results

3.1 Cost Analysis

An application of the optimization procedure has been carried out for the real
case of the Belvedere Marittimo submarine outfall, in southern Italy. The pipe
is a 600 m long HDPE, with a nominal diameter of 355 mm. The treatment
plant is equipped with a pumping station which, together with the difference in
height of the loading tank respect to the sea surface, provides a head value at
the beginning of the pipeline equal to about 19 m. Regarding the environmental
verification, a reduction up to the regulatory values has been considered, with a
coliform concentration value of 5.5 · 106 [CFU/100 ml], conveyed by a flow rate
of 0.125 m3/s.
To verify the validity of the construction of this structure, a series of diam-
eters have been evaluated based on the velocity in the pipeline. To this aim,
considering that velocity lower than 0.8 m/s are not valid for sedimentation prob-
lems, the DN280, 315, 355, 400, 450 and 500 have been selected.
Obviously for each diameter different solutions have been found, but the one
with the least length, and therefore the lowest cost in terms of materials, has
been considered the most valid solution.

Fig. 7. Comparison of minimum outfall length for each diameter.

614 S. Sinopoli et al.

As shown in Fig. 7, the system reaches the minimum necessary development

in terms of length for the DN355 with a value of 661.3 m, remaining unchanged
for higher measurements, and increasing the path for smaller diameters up to
727.2 m of DN280.
All the diffusers corresponding to the shorter overall length have only two
ports, and even in this case the minimum for the DN355 is reached.

Fig. 8. Comparison of necessary port dimension for each diameter.

Cm = 785.94 DN 1.9083 (16)

The evaluation of the structure convenience has been made based on the cost
per meter formula (16) for HDPE, already used by [11] for estimating the costs
of the systems, through which it is possible to identify the cost per meter for
every single diameter. Multiplying the values obtained for the respective lengths
leads to the cost in materials of the system.
The economically most convenient pipe to be realized is that with the DN280
with e50,356.43, while the pipeline with DN355 has a cost equal to 1.43 times
the most convenient with a total cost of e72025.89 (Fig. 9).

Fig. 9. Comparison of pipe cost for each diameter.

 Optimization of Submarine Outfalls with a Multiport Diffuser Design 615

3.2 Sensitivity Analysis

The proposed optimization model contains design values, such as the diameters,
the lengths, the capacity, which can be chosen by the designer, also through a
procedure like the one followed so far. There are, however, two parameters that
are never completely certain, the T90 bacterial decay time, which during the day
can vary from 2 to more than 20 hours, and the average velocity of the current,
which in the stretch of sea in question can vary greatly in direction and module
[23,25].
For this reason, a sensitivity analysis based on the fraction perturbation
method [16] has been carried out, to understand how the variations of these two
parameters can influence the dilution capacity.
Two systems have been analyzed, the one with DN280 and the one with
DN355, since they represent the economic optimum and the system that actually
exists, and to make the graphs more readable, the two parameters have been
made dimensionless.

Fig. 10. Sensitivity analysis for u.

The analysis of the velocity graphs shows a significant variation in the dilution
capacity, and the two plants undergo the same. In the same way, the response
to changes in T90 significantly affects the dilution capacity. Compared to the
velocity of the current the percentage variations are more moderate, but even in
this case, there are no particular differences between the two plants.
616 S. Sinopoli et al.

Fig. 11. Sensitivity analysis for T 90.

4 Conclusions

Submarine outfalls are mature hydraulic structures, but scientific advances lead
to the possibility of improving and optimizing the performance of these systems.
In particular, the optimization aspect of design and maintenance is impor-
tant, as is the economic optimization. In this context, the present work proposes
an optimization procedure for the pipe-diffuser system using a simplified zone
model subjected to a sensitivity analysis of the characteristic parameters. The
application to the Belvedere Marittimo submarine outfalls case study allowed to
evaluate the feasibility of the proposed procedure while the sensitivity analysis
carried out allowed to understand the influence on the dilution capacity of the
variations of two parameters.
The results showed that the procedure is feasible and expeditious and allows
to optimize the design of this hydraulic structures.

Aknowledgments. Research supported by the Italian Regional Project (POR CAL-

ABRIA FESR 2014–2020): “TEMAR” [J68C17000150006].

References
1. Agg, A.R., White, W.R.: Devices for the pre-dilution of sewage at submerged
outfalls. In: Proceedings of the Institution of Civil Engineers, vol. 57, no. 1, part
2, pp. 1–20 (1974)
2. Abraham, G.: Jet diffusion in stagnant ambient fluid. Delft Hydraulics Laboratory
29 (1963)
3. Abraham, G.: Horizontal jets in stagnant fluid of other density. J. Hyrdraulics Div.
91(4), 139–154 (1965)
 Optimization of Submarine Outfalls with a Multiport Diffuser Design 617

4. Benfratello, G.: Scarichi sottomarini: aspetti idraulici e fenomeni di diffusione. In:

Proceedings Immissione di acque reflue in mare. A.I.I., Ischia, Italy (1989)
5. Botelho, D.A., et al.: Marine Outfall Systems: Current Trends. Research and Chal-
lenges. IWA Specialist Group (2016)
6. Brooks, N.H.: Diffusion of sewage effluent in ocean current. In: Proceedings of
the First International Conference on Waste Disposal in the Marine Environment.
Pergamon Press (1959)
7. Brooks, N.H.: Conceptual design of submarine outfalls - jet diffusion. In: Water
Resources Engineering Educational Series, Program VII, Pollution of Coastal and
Estuarine Waters (1970)
8. Cederwall, K.: Gross parameter solution of jets and plumes. J. Hydraulics Div.
101(HY5), 489–509 (1975)
9. Charlton, J.A.: Developments in Hydraulic Engineerings. Sea Outfalls 3(3), 73–128
(1985)
10. Fisher, H.B., Imberger, J., List, E.J., Koh, R.C.Y., Brooks, N.H.: Mixing in Inland
and Coastal Waters. Academic Press, New York (1979)
11. Forrest, S., Rustum, R.: Design optimisation of marine wastewater outfalls. Int. J.
Sustainable Water Environ. Syst. 8(1), 7–12 (2016)
12. Hernández-Sancho, F., Lamizana-Diallo, B., Mateo-Sagasta, J., Qadir, M.: Eco-
nomic valuation of wastewater - the cost of action and the cost of no action-United
Nations Environment Programme (2015). ISBN: 978-92-807-3474-4
13. Holly, F.M.: Sea outfalls. In: Development in Hydraulic Engineering, vol. 3, no. 1,
pp. 1–37. Elsevier Applied Science Publisher, London (1985)
14. Maiolo, M., Pantusa, D.: A proposal for multiple reuse of urban wastewater. J.
Water Reuse Desalination 8(4), 468–478 (2017)
15. Maiolo, M., Pantusa, D.: Combined reuse of wastewater and desalination for the
management of water systems in conditions of scarcity. Water Ecol. 4(72), 116–126
(2017)
16. Mc Cuen, R.H.: The role of sensitivity analisys in hydrologic modeling. J. Hidrology
18, 37–53 (1973)
17. Occhipinti, A.G.: Bacterial disappearance experiments in Brazilian coastal waters.
In: Proceedings of Immissione di acque reflue in mare. A.I.I., Ischia, Italy (1989)
18. Rahm, S.L., Cederwall, K.: Submarine disposal of sewage. In Proceedings on the
XI Congress IAHR, Leningrad (1965)
19. Rodriguez-Garcia, G., Molinos-Senante, M., Hospido, A., Hernandey-Sancho, F.,
Moreira, M.T., Feijoo, G.: Environmental and economic profile of six typologies of
wastewater treatment plants. Water Res. 45, 5997–6010 (2011)
20. Shannon, N.R., Mackinnon, P.A., Hamill, G.: Modelling of saline intrusion in
marine outfalls. In: Proceedings of the Institution of Civil Engineers. Maritime
Engineering, vol. 158, no. MA2, pp. 47–58 (2005)
21. Veltri, P., Maiolo, M.: Un modello di zona per il trattamento marino dei liquami.
In: Proceedings on the XXII Corso di Aggiornamento in Tecniche per la Difesa
dall’Inquinamento, pp. 417–451. Bios, Cosenza (1991)
22. Veltri, P., Maiolo, M.: Un criterio ambientale per il calcolo degli scarichi dei liquami
in mare. In: Proceedings on the XXIII Hydraulics and Hydraulic infrastructures,
Firenze (1992)
23. Veltri, P., Maiolo, M.: Environmental aspects in the use of sea outfalls: a sensitivity
analysis. In: Marina Technology. Computational Mechanics Publications (1992)
24. Veltri, P., Maiolo, M.: Modelli Matematici di dispersione dei liquami in Mare e
Indagini su campo. Ingegneria Sanitaria Ambientale (1993)
618 S. Sinopoli et al.

25. Veltri, P., Morosini, A.F., Maiolo, M., Carratelli, E.P.: Analysis of Sea Current
and Diffusion Data at a Submarine Sewage Outfall, pp. 169–179. Computational
Mechanics Publications, Southampton (1997)
26. Vigliani, P.G., Sclavi, B., Olivotti, R., Visconti, A., Tartaglia, G.F.: Indagine sper-
imentale sulla diluizione iniziale dei liquami versati a mare da un diffusore, vol. 3,
pp. 129–140. Ingegneria Sanitaria (1980)
27. Viviani, G.: Il progetto delle condotte sottomarine: una semplificazione nell’uso dei
modelli complete, vol. 18, pp. 327–339. Ingegneria Ambientale (1989)
28. Wei, W., Zhong-min, Y.: Evaluation of three-dimensional numerical model for
saline intrusion and purging in sewage outfalls. J. Hydrodynamics Ser. B 19(1),
48–53 (2007)
29. Wilkinson, D.L.: Avoidance of seawater intrusion into ports of ocean outfalls. J.
Hydraulic Eng. 114(2), 218–228 (1988)
 Author Index

Acampora, Luigi II-3 Cesario, Eugenio I-223

Adamo, Maria Teresa I-451 Chiaravalloti, Francesco I-545
Addawe, Rizavel C. II-17 Cianflone, Giuseppe I-291
Agapito, Giuseppe I-215 Cicirelli, Franco I-232
Alberico, Ines I-291 Cococcioni, Marco II-337, II-370
Aleksandrov, Alexander Yu. II-271 Colombo, Giorgio II-322
Alexandre dit Sandretto, Julien II-169 Conti, Costanza I-11
Alfonsi, Giancarlo II-280, II-422 Cortese, Domenico I-358
Amodio, Pierluigi II-289, II-298 Costa, M. Fernanda P. II-59
Antonelli, Laura I-3 Cotrona, Rocco I-506
Antoniotti, Luigi I-329 Crisci, Serena I-161
Archetti, Francesco II-32 Cudazzo, Alessandro II-337
Arcuri, Paola I-243 Cuomo, Salvatore I-11, I-103, I-122
Avolio, Elenio I-291
D’Alotto, Louis II-346
Bagdasar, Ovidiu II-330, II-438 D’Ambrosio, Donato I-243, I-256, I-278
Baranov, Nikolay II-306 d’Atri, Gianfranco I-329, I-358, I-373, I-381,
Barilla, David II-314 I-389, I-402
Barkalov, Konstantin II-478 d’Atri, Stella I-402
Bellini, Emanuele I-343 D’Onghia, Mario I-243
Bertacchini, Francesca I-410, I-425, I-437 Danilenko, Olga II-72
Bevzuk, Semen II-478 De Luca, Davide Luciano II-84
Bezrukovs, Valerijs II-495 De Luca, Pasquale I-109
Bezrukovs, Vladislavs II-495 De Meo, Pasquale II-330
Bilotta, Eleonora I-410, I-425, I-437 De Pietro, Michela I-410
Bonettini, Silvia I-198 De Rango, Alessio I-243, I-256, I-278
Bonora, Marco Amos I-495, I-506, I-525, De Rosa, Rosanna I-291
I-545, I-604 De Rossi, Alessandra I-19, I-27
Borrelli, Andrea I-596 Dell’Accio, Francesco I-27, I-35, I-47
Breuza, Emanuele II-322 Demarco, Francesco I-425
Brugnano, Luigi II-289 Derevtsov, Evgeny Yu. II-97
Di Gregorio, Salvatore I-291
Caldarola, Fabio I-329, I-358, I-373, I-381, di Serafino, Daniela I-3, I-177
I-389, I-495, I-517 Di Tommaso, Filomena I-27, I-35, I-47
Calidonna, Claudia R. I-291 Dominici, Rocco I-291
Campagna, Rosanna I-11, I-146 Dzemyda, Gintautas II-354
Candelieri, Antonio II-32
Cannataro, Mario I-215
Egidi, Nadaniela I-56, I-70
Capano, Gilda I-506, I-525, I-545, I-604
Carini, Manuela I-506, I-525, I-545, I-604
Caristi, Giuseppe II-314 Falcone, Alberto II-362
Caterina, Gianluca II-43 Fatone, Lorella I-56
Cavallaro, Lucia II-330 Fernandes, Edite M. G. P. II-59
Cavoretto, Roberto I-19, I-27 Ferraro, Domenico II-280
620 Author Index

Fiaschi, Lorenzo II-370 Kahramanoğulları, Ozan I-131

Fink, Tatyana II-519 Kaneva, Olga II-519
Fiumara, Giacomo II-330 Kanzi, Nader II-314
Florio, Emilia I-307 Kel’manov, Alexander II-386, II-394
Folino, Gianluigi I-269 Khamidullin, Sergey II-386
Franchini, Giorgia I-186 Khandeev, Vladimir II-394, II-400
Francomano, Elisa I-3 Korotchenko, Anatoliy G. II-406
Frassia, Maria Giovanna I-478 Kostousova, Elena K. II-124
Fuoco, Domenico I-596 Kumar, Vivek II-139
Furnari, Luca I-256, I-537 Kurushina, Svetlana II-454
Kvasov, Dmitri E. II-413
Gabriele, Lorella I-437
Galasso, Luciano II-84 La Rosa, Salvatore I-291
Galletti, Ardelio I-109 Lauria, Agostino II-280, II-422
Gangle, Rocco II-43 Le, Van Thanh I-402
Gao, Kaifeng I-122 Liotta, Antonio II-330, II-438
Garrì, Dino I-402 Lirer, Fabrizio I-291
Garro, Alfredo II-362 Litvinas, Linas II-511
Gaudio, Roberto II-280 Lochab, Ruchika II-139
Gergel, Victor II-112, II-378 Louis, Alfred K. II-487
Giampaolo, Fabio I-103 Lupiano, Valeria I-291
Giglio, Simona I-437
Gil, Marisa I-318 Magadia, Joselito C. II-17
Giordani, Ilaria II-32 Maiolo, Mario I-358, I-373, I-495, I-506,
Giordano, Andrea I-232, I-243, I-256, I-278 I-517, I-525, I-545, I-604
Giunta, Giulio I-109 Makarov, Anton II-430
Goncharova, Marina II-503 Makarova, Svetlana II-430
Gonnelli, Domenico I-47 Malhotra, Tanya II-154
Goryachikh, Alexey II-378 Maltseva, Svetlana V. II-487
Gregoretti, Francesco I-3 Maponi, Pierluigi I-70
Gregori, Daniele II-322 Marcellino, Livia I-109
Grishagin, Vladimir II-112 Marchetti, Filippo II-322
Gualdi, Silvio I-596 Marcinkevicius, Virginijus II-195
Gulbe, Linda II-495 Marra, Francesco S. II-3
Gupta, Anjana II-154 Martino, Maria I-451
Martorell, Xavier I-318
Hahn, Bernadette N. II-446 Masi, Paolo I-146
Hancyzc, Martin M. I-131 Mastroianni, Carlo I-232
Mazzia, Francesca II-289
Iavernaro, Felice II-289 Mehra, Mani II-470
Ingarozza, Francesco I-451 Mei, Gang I-122
Israfilov, Ruslan II-112 Mendicino, Giuseppe I-243, I-256, I-537,
I-596
Jankunec, Marija II-462 Menniti, Daniele I-437
Jekabsons, Normunds II-495 Mercuri, Pietro I-381
Jurevicius, Rokas II-195 Meškauskas, Tadas II-462
 Author Index 621

Misici, Luciano I-56 Sabaliauskas, Martynas II-354

Molevich, Nonna II-454 Sanna, Antonella I-596
Mukhametzhanov, Marat S. II-362, II-413 Savkin, Vladimir II-519
Mullier, Olivier II-169 Schuster, Thomas II-97
Muranho, Joao I-495 Scuro, Carmelo I-425
Murru, Nadir I-343 Sechi, Giovanni M. I-557
Senatore, Alfonso I-243, I-256, I-537, I-596
Napolitano, Jacopo I-557 Sergeyev, Yaroslav D. II-337, II-362, II-413
Nasso, Maria Chiara II-413 Serpe, Annarosa I-463, I-478
Settanni, Giuseppina II-298
O’Neill, Sam II-438 Severino, Gerardo I-103
Occorsio, Donatella I-86 Shukla, Ankita II-470
Oliverio, Marco I-243 Singh, Anjali II-154
Sinopoli, Salvatore I-604
Palermo, Stefania Anna I-570, I-583 Smoryakova, Valentina M. II-406
Paliaga, Marta I-3 Socievole, Annalisa II-182
Panasenko, Anna II-386 Sorrentino, Nicola I-437
Pantano, Pietro I-410, I-425, I-437 Sousa, Joaquim I-495, I-506, I-545
Pantusa, Daniela I-506, I-545, I-604 Spataro, Davide I-243
Pappalardo, Massimo II-337 Spataro, William I-243, I-278
Pellegrini, Marco I-329, I-389 Straface, Salvatore I-243, I-256
Pelosi, Nicola I-291 Strongin, Roman II-478
Penkauskas, Tadas II-462 Svetov, Ivan E. II-446, II-487
Perego, Riccardo II-32
Petroselli, Andrea II-84 Talamanca, Valerio I-381
Piccialli, Francesco I-122 Talarico, Vito Cataldo I-570
Pinnarelli, Anna I-437 Themistoclakis, Woula I-86
Pirillo, Giuseppe I-373 Toraldo, Gerardo I-146, I-177
Piro, Patrizia I-583 Trillini, Carla II-208
Pirouz, Behrouz I-570, I-583 Turco, Michele I-583
Pisani, Francesco Sergio I-269
Piscitelli, Aldo I-451 Uluşeker, Cansu I-131
Pizzuti, Clara II-182 Upnere, Sabine II-495
Polyakova, Anna P. II-446 Uteshev, Alexei II-503
Pontieri, Luigi I-269 Utrera, Gladys I-318
Porfiriev, Denis II-454
Porta, Federica I-161
Valaitis, Vytautas II-195
Pospelov, Igor II-255
Valinčius, Gintaras II-462
Prato, Marco I-198
Vinci, Andrea I-223
Primavera, Leonardo I-307
Viola, Marco I-177
Viscomi, Antonio I-291
Raei, Marzie I-109
Visokolskis, Sandra II-208
Raila, Tomas II-462
Volkov, Yuriy S. II-97
Rasina, Irina II-72
Vrahatis, Michael N. II-223
Rebegoldi, Simone I-198
Rocha, Ana Maria A. C. II-59
Romano, Annalisa I-146 Wongprachan, Ratchaneewan II-239
Rongo, Rocco I-243, I-278
Ruggiero, Valeria I-161, I-186 Xu, Nengxiong I-122
622 Author Index

Zanni, Luca I-161, I-186 Zhukova, Aleksandra II-255

Zavershinsky, Igor II-454 Žilinskas, Antanas II-511
Zhu, Xiaotian I-223 Zykina, Anna II-519