Inversion of a SIR-based model: a critical analysis

about the application to COVID-19 epidemic

M. Giudici⋄ , A. Comunian⋄ , R. Gaburro⋆
arXiv:2004.07738v1 [q-bio.PE] 16 Apr 2020

⋄
Università degli Studi di Milano, Italy
⋆
University of Limerick, Ireland

April 17, 2020

Abstract
Calibration of a SIR (Susceptibles-Infected-Recovered) model with of-
ficial data at international level for the COVID-19 pandemics provides
a good example of the difficulties inherent the solution of inverse prob-
lems. Inverse modeling is set up in a framework of discrete inverse
problems, which explicitly considers the role and the relevance of data.
Together with a physical vision of the model, this is very useful to dis-
cuss the uncertainties on the data and how they influence the reliability
of calibrated model parameters and, ultimately, of model predictions.

Keywords: Inverse problems; Mathematical modelling; Model calibration;

Epidemic modeling

1 Introduction
Epidemic modeling is usually performed with compartmental models, often
called SIR models, which are claimed to go back to the works by Ronald
Ross and Hilda P. Hudson more than one century ago [13, 14] and by Ander-
son Gray McKendrick and William Ogilvy Kermack ten years later [9, 10].
This class of models shares several characteristics with models of population
dynamics and with conceptual lumped models in hydrology. These models
simulate the temporal evolution of some compartments of the population,
which is normally subdivided among Susceptibles (i.e., those persons who
have not yet been affected by the virus and which could be subject to in-
fection), Infected (i.e., those persons who have been infected by the virus)
and Recovered (i.e., those persons who have recovered, after having been in-
fected). For this reason these models are usually referred to as SIR models.
They are based on simple laws to describe the transfer of individuals from
one class to the others.
These models have found wide application both in life sciences, mostly in
epidemiology, and in the field of economic, political and social sciences, e.g.,

1
2 Giudici et al., Inversion of a SIR-based model: application to COVID-19 epidemic

to assess the costs of different policies to block epidemics and the diffusion
of viruses. Most papers consider academic issues and are rarely calibrated
against real data.
Model calibration is a common problem in geophysical and environmen-
tal modeling. A general framework to handle discrete inverse problems for
model calibration is proposed in [7] and can be useful to discuss some char-
acteristics of SIR models and the role of data, also following the discussion
in [6].
The wide number of data which are collected during the COVID-19
pandemic due to the diffusion of the SARS-CoV-2 virus (also called “coro-
navirus”) provides an exceptional basis to perform some modeling exercises
and to test different calibration methods.
The objectives of this paper are to fix some concepts about SIR models
and their calibration and to discuss the relevance of data for reliability of
model outcomes.
The paper is designed to advance the knowledge on the functioning,
potentialities and limitations of epidemic models. Also it is expected to
provide further insights in epidemic model calibration. Instead, it is not
designed to provide forecasts of the pandemic evolution. In authors’ opinion,
the quality of available data does not permit to perform reliable forecast and
model outcomes should be used with high prudence.
The paper is organized as follows. Next section is devoted to the descrip-
tion of the model, in the continuous and discrete case, of the methods used
for the calibration of the numerical model, and of the data for the applica-
tion to the COVID-19 pandemic; in particular, inverse modeling, i.e., model
calibration, will be set up and discussed within the framework proposed by
[7]. Some results of the model with reference to COVID-19 pandemic will
be shown in the third section, whereas the fourth section will be devoted
to a discussion of several topics: the assumptions at the basis of the SIR
model; some remarks about model calibration; some remarks about data
uncertainty. The concluding section will also include some hints for future
developments of this work.

2 Methods and materials

2.1 The continuous model
This section is devoted to the description of the SIR model considered in
this paper.
First of all, S(t), I(t), R(t) and D(t) represent the number of, respec-
tively, susceptible, infected, recovered and dead individuals of the population
under study as a function of time. Notice that D includes only those persons
who died while being infected. P = S + I + R denotes the total number of
population individuals.
 Giudici et al., Inversion of a SIR-based model: application to COVID-19 epidemic 3

The coefficients β and δ denote the birth and death rate, respectively,
under normal conditions, i.e., without considering deaths caused by the
epidemic. These coefficients are rarely considered in epidemic modeling, as
the variation due to the normal evolution of the population is negligible or
at least smoother than the variations caused by epidemics.
The following equations, based on historical papers [13, 14, 9, 10], are
used to describe the time evolution of S, I, D and R:

dS IS
= βS − γ − δS,
dt P
dI IS
= γ − ρI − φI + βI,
dt P
(1)
dD
= φI,
dt
dR
= βR + ρI − δR.
dt

The second term on the right hand side of the first equation of (1) repre-
sents the number of individuals who are infected per unit time. It is based
on the assumption that each infected person has contacts with a given num-
ber of persons in a certain time interval and that the fraction of them who
are susceptible is given by S/P , whereas (I + R)/P is the fraction of those
persons who cannot be infected, if it is assumed that recovered people are
immunized. The coefficient γ is the infection coefficient, i.e., the rate of
potential infection.
The coefficients ρ and φ represent the recovery and fatality rate, re-
spectively. The coefficients β, γ, δ, ρ and φ are assumed to be constant
and their dimension is [time−1 ]. The assumptions behind this model are
discussed thoroughly in section 4.
The initial conditions of the model are given by S(0) = P (0), I(0) = 1,
R(0) = 0 and D(0) = 0. This means that t = 0 corresponds to the time at
which the first individual is infected.
Notice that from equations (1), it follows that

dP
= βP − δP − φI. (2)
dt

2.2 The discrete model

The discrete model is a simple forward-time finite-differences discretization

of equations (1). Therefore, if t′ ∈ N is the index used to denote the discrete
time steps and a uniform spacing ∆t is considered, then the following explicit
4 Giudici et al., Inversion of a SIR-based model: application to COVID-19 epidemic

iterative equations are obtained:

I(t′ )
   
S(t′ + 1) = 1+ β−γ − δ ∆t S(t′ ),
P (t′ )
S(t′ )
   
′
I(t + 1) = 1+ γ − ρ − φ + β ∆t I(t′ ),
P (t′ ) (3)

D(t′ + 1) = D(t′ ) + φI(t′ )∆t,

R(t′ + 1) = [1 + (β − δ)∆t] R(t′ ) + ρI(t′ )∆t.

The initial conditions of the discrete model are given by:

S(0) = P0 − 1, I(0) = 1, D(0) = 0, R(0) = 0, (4)

where P0 is the population at t′ = 0.

The results presented in this paper consider the application of the model
to a nation. In other words, the population of the whole nation is considered,
without any further subdivision in provinces, regions or states. Moreover,
the time spacing is 1 day, in agreement with the sampling of the available
data set on COVID-19 pandemic (see section 2.4).

2.3 The inverse problem

As stated in the introduction, the inverse problem is defined by making use
of the conceptual framework and the notation of [7].
The time-varying state of the system is included in an array s:

s = S(t′ ), I(t′ ), R(t′ ), D(t′ ), t′ = 0, . . . , Nmod − 1 ,


(5)

where Nmod is the number of modeled time steps.

The available data are collected in an array d that, in the specific case
considered here, includes the number of infected, recovered and dead per-
sons, released by sanitary official organizations:

{Iobs (t), Robs (t), Dobs (t), t = 0, . . . , Nobs − 1} ⊂ d. (6)

Nobs denotes the number of time steps for which data are available. Notice
that in the discrete case, t ∈ Z is used as the time-index which denotes days,
starting from the reference date, taken as t = 0, which corresponds to the
first day for which epidemic data are available.
The model parameters are included in an array p:

p = {β, δ, ∆t, ρ, φ, γ, t0 , P0 } , (7)

where t0 represents the day, at which the first individual of the population
is infected and the total population is P0 .
 Giudici et al., Inversion of a SIR-based model: application to COVID-19 epidemic 5

The discrete model given by equations (3) and (4), can be written as the
following system of equations

f (p, s) = 0. (8)

If the numbers of model parameters and state parameters are N(p) and
N(s) , respectively, then p ∈ P ⊆ RN(p) and s ∈ S ⊆ RN(s) , where the
subset P takes into account some conditions on the parameters, e.g., that
the parameters ρ, φ, γ and P0 must satisfy the following constraints:

0 ≤ ρ ≤ 1 − φ ≤ 1, 0 ≤ φ ≤ 1 − ρ ≤ 1, 0 ≤ γ, 0 ≤ P0 . (9)

The unique solution of the forward problem, i.e., of (8), can be expressed
as
s = g(p). (10)
The array p can be subdivided in the two sub-arrays

p(fix) = {β, δ, ∆t} , (11)

which includes the model parameters, whose values are fixed before the
simulation, and
p(cal) = {ρ, φ, γ, t0 , P0 } , (12)
which includes the model parameters, whose values are obtained from the
solution of an inverse problem. Therefore
t t t
 
p = p(fix) , p(cal) . (13)

The array of fixed parameter could be a function of d: p(fix) (d).

The model outcome, i.e., the state of the system, is used to forecast
the number of infected, recovered and dead individuals at the times for
which some estimates are available from epidemiological data. Therefore,
the model forecast is expressed as an array y, which is function of s, p and
d: y (d, s, p).
Then, model calibration requires that the model forecast be close to
a calibration target, an array t that collects the values which should be
attained by the model forecast, if the model were physically “correct” and
the model parameters were “optimal”. Recall that t may
depend on d and
p(fix) , but should be independent of p(cal) : t = t d, p(fix) .
The correspondence between model forecast and data must take into
account that t′ = t − t0 , so that t′ = 0 when t = t0 , i.e., the time at which
the first individual is infected. Moreover, the objective of model calibration
will be to find the parameter values which best fit the observations in a given
time interval, t1 ≤ t < t2 . Since data and model outcomes refer to different
time intervals, the following inequalities must be verified:

0 ≤ t − t0 < Nmod , 0 ≤ t1 ≤ t < t2 ≤ Nobs . (14)

6 Giudici et al., Inversion of a SIR-based model: application to COVID-19 epidemic

Equation (14) are satisfied if one picks

t1 ≥ t0 and t2 ≤ min (Nmod + t0 , Nobs ) . (15)

Then, the arrays y and t are given by:

y = {I(t′ ), R(t′ ), D(t′ ), t1 − t0 ≤ t′ < t2 − t0 } ,

(16)
t = {Iobs (t), Robs (t), Dobs (t), t1 ≤ t < t2 } .

The misfit between model predictions and the target values is computed
by means of the following objective function:

O p(cal) = y (d, s, p) − t d, p(fix)

(17)
= OI p(cal) + OR p(cal) + OD p(cal) ,

where each of OI , OR and OD is defined by

tX
)1/2
2 −1 
(
X(t − t0 ) − Xobs (t) 2

  1
OX p(cal) = , (18)
t2 − t1 t=t max(ξ, Xobs (t))
1

with X ∈ {I, R, D}, Xobs being the corresponding element of {Iobs , Robs , Dobs },
and ξ ≥ 1 is a threshold. In other words, O is the sum of three functions,
each of which considers one of the three observed quantities, separately.
The model calibration is performed by solution of the following inverse
problem: given p(fix) and d, given the solution s = g (p) to (8), given the
⋆
functions y (d, g (p) , p) and t, find p(cal) ∈ P (cal) , such that

⋆
 
p(cal) = arg min O p(cal) ,
p(cal) ∈P (cal)
(19)
⋆
 
i.e., O p(cal) ≤ O p(cal) , ∀p(cal) ∈ P (cal) ,

 
t t t
 
where P (cal) = p(cal) : p(fix) ,p (cal) ∈P .
The threshold ξ ∈ R plays a double role. First of all, it keeps positive the
denominator of the fraction appearing in (18). Furthermore, it controls some
characteristics of the objective function. For ξ = 1, OX is nothing but the
root-mean-squared relative difference between observed and modeled values
of X. If a large value of ξ is used, then relative errors corresponding to
large values of Xobs will be dominant; from the practical point of view, this
means that early time behavior has a minor relevance for the model fitting.
In particular, if ξ > max{Xobs (t), t1 ≤ t < t2 }, then OX reduces to the
standard root-mean-squared error.
 Giudici et al., Inversion of a SIR-based model: application to COVID-19 epidemic 7

2.4 Data and computer implementation for COVID-19

The application of the model introduced in section 2.2 and of the model
calibration introduced in section 2.3 can be attempted thanks to publicly
available data on COVID-19 pandemic. The application will be performed at
national level, i.e., the considered population will be the whole population
of some countries. For each country, the array d is populated with data
coming from two basic sources.
Data on COVID-19 pandemic are available from the GitHub reposi-
tory managed by the John Hopkins University [3]. This is a collection of
publicly available data from multiple sources, which are processed and de-
livered by the Johns Hopkins University Center for Systems Science and
Engineering (JHU CSSE). Notice that the data are provided to the pub-
lic strictly for educational and academic research purposes. The data are
updated daily and the files used in this paper have been downloaded from
https://github.com/CSSEGISandData/COVID-19 on April 16, 2020. From
those files, the array t is easily built.
A tailored code has been developed under Python 3.7.6 to download
data from the Github repository, perform forward modeling and calibrate
the model by solution of the inverse problem. In particular, inversion is
based on the functions of the optimize module from SciPy v1.4.1.
Figure 1 shows the trend of confirmed cases, recovered and deceased peo-
ple for some countries, among those that have been considered as the most
relevant for the analysis of COVID-19 pandemic not only by the scientific
community, but also by mass media. These plots show different trends for
different countries and for the different quantities.
Aside from China, for which the starting phase is not reported, because
the virus diffusion started earlier than the first date for which data are
available in the data set, the number of confirmed cases (top plot in Figure 1)
shows a first slow increase, followed by an exponential increase and possibly
a slowdown after few weeks. It is highly questionable whether this behavior
is related to the number of tests performed to confirm virus infection.
The most regular trends are clearly the ones describing the number of
deceased people (bottom plot in Figure 1), after about one week since the
first reported case in each country considered in this study. Doubts about
comprehensiveness of official data on deaths caused by coronavirus have
been raised by several sources of information and by some commentators.
Nevertheless, it seems safe to state that the number of deaths represents the
smoothest time series and possibly the less affected by uncertainties.
Notice that the daily sampling rate of these data induces to choose ∆t =
1 day, and the coefficients β, γ, δ, φ and ρ share the same measurement
units, namely day−1 .
The second data source is the most updated version of the UN Demo-
graphic Yearbook [12]. Demographic data have been extracted from this
firmed infections. Middle: recovered patients. Bottom: deaths.
Figure 1: Data about COVID-19 pandemic in selected countries. Top: con-

8
Giudici et al., Inversion of a SIR-based model: application to COVID-19 epidemic
Number of cases Number of cases Number of cases
100

101

102

103

104

100

101

102

103

104

105

100

101

102

103

104

105

106
1/22/20 1/22/20 1/22/20
1/23/20 1/23/20 1/23/20

Korea, South
Japan
Iran
Ireland
Germany
France
Spain
Italy
United Kingdom
US
China

Korea, South
Japan
Iran
Ireland
Germany
France
Spain
Italy
United Kingdom
US
China
1/24/20 1/24/20 1/24/20
1/25/20 1/25/20 1/25/20
1/26/20 1/26/20 1/26/20
1/27/20 1/27/20 1/27/20
1/28/20 1/28/20 1/28/20
1/29/20 1/29/20 1/29/20
1/30/20 1/30/20 1/30/20
1/31/20 1/31/20 1/31/20
2/1/20 2/1/20 2/1/20
2/2/20 2/2/20 2/2/20
2/3/20 2/3/20 2/3/20
2/4/20 2/4/20 2/4/20
2/5/20 2/5/20 2/5/20
2/6/20 2/6/20 2/6/20
2/7/20 2/7/20 2/7/20
2/8/20 2/8/20 2/8/20
2/9/20 2/9/20 2/9/20
2/10/20 2/10/20 2/10/20
2/11/20 2/11/20 2/11/20
2/12/20 2/12/20 2/12/20
2/13/20 2/13/20 2/13/20
2/14/20 2/14/20 2/14/20
2/15/20 2/15/20 2/15/20
2/16/20 2/16/20 2/16/20
2/17/20 2/17/20 2/17/20
2/18/20 2/18/20 2/18/20
2/19/20 2/19/20 2/19/20
2/20/20 2/20/20 2/20/20
2/21/20 2/21/20 2/21/20
2/22/20 2/22/20 2/22/20
2/23/20 2/23/20 2/23/20
2/24/20 2/24/20 2/24/20
2/25/20 2/25/20 2/25/20
2/26/20 2/26/20 2/26/20
2/27/20 2/27/20 2/27/20
2/28/20 2/28/20 2/28/20
2/29/20 2/29/20 2/29/20
3/1/20 3/1/20 3/1/20

recovered

confirmed
3/2/20 3/2/20 3/2/20

deaths
3/3/20 3/3/20 3/3/20
3/4/20 3/4/20 3/4/20
time

time

time
3/5/20 3/5/20 3/5/20
3/6/20 3/6/20 3/6/20
3/7/20 3/7/20 3/7/20
3/8/20 3/8/20 3/8/20
3/9/20 3/9/20 3/9/20
3/10/20 3/10/20 3/10/20
3/11/20 3/11/20 3/11/20
3/12/20 3/12/20 3/12/20
3/13/20 3/13/20 3/13/20
3/14/20 3/14/20 3/14/20
3/15/20 3/15/20 3/15/20
3/16/20 3/16/20 3/16/20
3/17/20 3/17/20 3/17/20
3/18/20 3/18/20 3/18/20
3/19/20 3/19/20 3/19/20
3/20/20 3/20/20 3/20/20
3/21/20 3/21/20 3/21/20
3/22/20 3/22/20 3/22/20
3/23/20 3/23/20 3/23/20
3/24/20 3/24/20 3/24/20
3/25/20 3/25/20 3/25/20
3/26/20 3/26/20 3/26/20
3/27/20 3/27/20 3/27/20
3/28/20 3/28/20 3/28/20
3/29/20 3/29/20 3/29/20
3/30/20 3/30/20 3/30/20
3/31/20 3/31/20 3/31/20
4/1/20 4/1/20 4/1/20
4/2/20 4/2/20 4/2/20
4/3/20 4/3/20 4/3/20
4/4/20 4/4/20 4/4/20
4/5/20 4/5/20 4/5/20
4/6/20 4/6/20 4/6/20
4/7/20 4/7/20 4/7/20
4/8/20 4/8/20 4/8/20
4/9/20 4/9/20 4/9/20
4/10/20 4/10/20 4/10/20

Korea, South
Japan
Iran
Ireland
Germany
France
Spain
Italy
United Kingdom
US
China
4/11/20 4/11/20 4/11/20
4/12/20 4/12/20 4/12/20
4/13/20 4/13/20 4/13/20
4/14/20 4/14/20 4/14/20
4/15/20 4/15/20 4/15/20
 Giudici et al., Inversion of a SIR-based model: application to COVID-19 epidemic 9

Table 1: Parameter values for test case 1.

run β δ γ ρ φ P0
(a) 0.009 y−1 0.0011 y −1 0.2 day −1
0.01 y −1 0.001 day −1 109
(b) idem idem idem 0.05 y −1 idem idem
(c) idem idem idem 0.1 y −1 idem idem

volume. The values of population, birth and death rate of each country, for
which the model has been tested, are included in d. They are used to fix
the values of β and δ, which are expressed on a daily basis, and to provide
a first estimate of P0 .

3 Results
3.1 Model results
First of all the behavior of the model is shown with test case 1, which in-
cludes three model runs for which all the model parameters are kept fixed,
but ρ: the list of parameter values is given in Table 1; the results of the
model for a one-year-long simulation period are shown in Figure 2. The
general behavior shows an exponential increase in the number of infected
persons (notice that the vertical axis is in logarithmic scale) followed by an
exponential decrease but with a longer characteristic time. The number of
deaths obviously decreases if ρ increases and in particular, we have three
different situations for the three runs: (a) for the smallest value of ρ, the
curve of susceptible persons dramatically decreases from some days before
the peak of infections and reaches very small values after few weeks; (b) for
the intermediate value of ρ, the chosen values of model parameters yield a
stationary conditions after about 8 months from the start of the epidemic
for the number of susceptible and dead people, which reach almost the same
value; (c) for the highest value of ρ, the number of susceptible people de-
creases with time, but remains consistent. Notice that, for this test case,
the reduction of the total population is limited, less than 10%, and after
one year almost all the living population is recovered. It is important to
stress that this test case has the goal of showing how the model can predict
different behavior and these results should not be considered as a forecast
of the actual behavior of any real pandemic.
SIR models are sometimes applied using the ratio of the number of indi-
viduals in each category with respect to the total population as a variable.
Even if test case 1 showed that for three sets of model parameters, which
differ only for the value of ρ, the total population shows only a limited vari-
ation, nevertheless, the term used to compute the infection rate introduces
a non-linearity in the model. Therefore test case 2 is designed to assess
the effect of P0 on model results. P0 values span four orders of magnitude,
10 Giudici et al., Inversion of a SIR-based model: application to COVID-19 epidemic

(a) ρ = 10−2 day−1

1010

108

Number of cases 106

Susceptibles (7)
Infected (67676400)
Dead (84655160)
Recovered (845760640)
Population (913364351)
104

102

100
0 50 100 150 200 250 300 350
Time (days)

(b) ρ = 5 · 10−2 day−1

1010

108

106
Number of cases

104

102
Susceptibles (18400416)
Infected (33395)
Dead (19227934)
Recovered (960371072)
Population (978799487)
100
0 50 100 150 200 250 300 350
Time (days)

(c) ρ = 10−1 day−1

1010

108

106
Number of cases

104

102
Susceptibles (200472208)
Infected (48206)
Dead (7902054)
Recovered (789595520)
Population (990110463)
100
0 50 100 150 200 250 300 350
Time (days)

Figure 2: Model results for test case 1. Numbers in the legends refer to the
values at the end of the simulation period.
 Giudici et al., Inversion of a SIR-based model: application to COVID-19 epidemic 11

from 106 to 109 , whereas the other parameters are fixed at the values of run
(a) of test case 1. The results are shown in Figure 3 as functions of the
normalized quantities versus time. The values of each function at the end
of the simulation period are very similar. The main differences are in the
evolving phase, for which the response of a small population appears to be
more rapid than that of a large population. Roughly speaking, the curves
corresponding to high populations show a delay with respect to the curve
for the smallest population of about 15 days per an increase in P0 by an
order of magnitude. This remark, if confirmed by runs with more reliable
parameter sets, could have fundamental consequences in the design of early
warning systems.

3.2 Model calibration

Model calibration for the COVID-19 pandemic by solution of the inverse
problem is a very challenging problem. This is not surprising at all, because
the comparison of the trends of the model time series (Figure 2) with those
observed from the data and drawn in Figure 1 shows that the SIR model
can hardly reproduce the observed trend.
In particular, this paper is focused on the results obtained with data
from Italy, but the same qualitative remarks apply also to other cases.
The best “trial-and-error” estimate is shown in Figure 4 and is obtained
for the following values of the parameters to be calibrated: γ = 0.33 day −1 ,
ρ = 0.015 day −1 , φ = 0.0025 day −1 , t0 = 10.
Starting from this initial set of parameters, minimization of the objective
function O was performed with a SciPy function which implements several
methods to find a minimum, also by taking into account possible bounds
on the independent variable of the objective function. Several tests have
been conducted and have shown that the best results were obtained with
the L-BFGS-B method, which is a variation of the BroydenFletcherGoldfarb-
Shanno (BFGS) algorithm [5] to reduce memory requirements and to handle
simple constraints. The bounds assigned for the parameters to be calibrated
are listed in Table 2. If the whole set of data is used, i.e., t1 = t0 = 10,
t2 = Nobs , the application of this algorithm leads to a parameter set for
which γ is very small, in fact close to the lower bound.
The application of differential evolution, an algorithm of global
minimization [16], yields the following set of parameters when ξ = 106 ,
which is equivalent to considering root-mean-squared error for OX : γ =
(0.1958 ± 2 · 10−5 ) day −1 , ρ = (1.289 · 10−2 ± 2 · 10−5 ) day−1 , φ = (8.14 ·
10−3 ± 2 · 10−5 ) day−1 , t0 = −9, P0 = 218, 200 ± 200. The mean value and
its standard deviation of each parameter has been estimated from 10 runs of
this stochastic algorithm, which introduces small variations in the returned
results. The comparison between observed and fitted time series is shown
in Figure 5. Two facts should be mentioned: t0 < 0, i.e., it seems that the
12 Giudici et al., Inversion of a SIR-based model: application to COVID-19 epidemic

(a) P0 = 106
100
Susceptibles
Infected
Dead
Reco ered
10−1

10−2

Fraction of cases o er population

10−3

10−4

10−5

10−6
0 50 100 150 200 250 300 350
Time (days)

(b) P0 = 107
100
Susceptibles
Infected
Dead
Reco ered
10−1

10−2
Fraction of cases o er population

10−3

10−4

10−5

10−6
0 50 100 150 200 250 300 350
Time (days)

(c) P0 = 108
100
Susceptibles
Infected
Dead
Reco ered
10−1

10−2
Fraction of cases o er population

10−3

10−4

10−5

10−6
0 50 100 150 200 250 300 350
Time (days)

(d) P0 = 109
100
Susceptibles
Infected
Dead
Reco ered
10−1

10−2
Fraction of cases o er population

10−3

10−4

10−5

10−6
0 50 100 150 200 250 300 350
Time (days)

Figure 3: Model results for test case 2.

 Giudici et al., Inversion of a SIR-based model: application to COVID-19 epidemic 13

108

106

Number of cases

104

Observed infected (156363)

102 Observed dead (19899)
Observed recovered (34211)
Susceptibles (7775381)
Infected (46702604)
Dead (852824)
Recovered (5116738)
Population (58277329)
100
1/22/20
1/23/20
1/24/20
1/25/20
1/26/20
1/27/20
1/28/20
1/29/20
1/30/20
1/31/20
2/1/20
2/2/20
2/3/20
2/4/20
2/5/20
2/6/20
2/7/20
2/8/20
2/10/20
2/9/20
2/11/20
2/12/20
2/13/20
2/14/20
2/15/20
2/16/20
2/17/20
2/18/20
2/19/20
2/20/20
2/21/20
2/22/20
2/23/20
2/24/20
2/25/20
2/26/20
2/27/20
2/28/20
2/29/20
3/1/20
3/2/20
3/3/20
3/4/20
3/5/20
3/6/20
3/7/20
3/8/20
3/9/20
3/10/20
3/11/20
3/12/20
3/13/20
3/14/20
3/15/20
3/16/20
3/17/20
3/18/20
3/19/20
3/20/20
3/21/20
3/22/20
3/23/20
3/24/20
3/25/20
3/26/20
3/27/20
3/28/20
3/29/20
3/30/20
3/31/20
4/1/20
4/2/20
4/3/20
4/4/20
4/5/20
4/6/20
4/7/20
4/8/20
4/10/20
4/9/20
4/11/20
4/12/20
Time (days)

Figure 4: Comparison of observations for Italy and modelled data with the
parameters obtained from subjective “trial-and-error” calibration.

Table 2: Intervals of variation fixed for the parameters to be calibrated for

inversion of data referred to Italy.
γ ρ φ t0 P0
minimum 10 day −4 −1 −5
10 day −1 −6
10 day −1
−10 2 · 105
maximum 1 day −1
0.1 day −1
0.1 day −1
30 108

infection started before the official appearance of the first confirmed case; P0
is close to the lower bound, so that the model predicts that the population
which has been involved in the infection could be relatively small.

These results suggested a further test case; in particular, the minimum

admissible value for t0 has been lowered to −30. The optimal set of values for
this test are: γ = (0.1543±1.5·10−5 ) day−1 , ρ = (1.237·10−2 ±1·10−5 ) day−1 ,
φ = (7.924 · 10−3 ± 2 · 8.5 · 10−6 ) day−1 , t0 = −29, P0 = 247, 490 ± 96. The
comparison between observed and modeled data is shown in Figure 6. The
value of the function O decreases from 8.3·10−3 for the results of Figure 5 to
5.2 · 10−3 for those of Figure 6, but the visual inspection shows a moderate
difference.

If ξ = 1, then γ = (0.31 ± 0.03) day −1 , ρ = (1.9 ± 0.3) · 10−2 day−1 ,

φ = (1.7 ± 0.4) · 10−2 day−1 , t0 = 19 ± 2, P0 = 16, 000, 000 ± 10, 000, 000.
Indeed, for this test case, the returned “optimal” value of O has a coefficient
of variation equal to 13%, much larger than those obtained in the test cases
for ξ = 106 , which were less than 0.06%.
set used for model calibration, i.e., they correspond to t1 and t2 .
minimization algorithm, for t0 = −29. Vertical dotted line delimit the data
the parameters obtained by solution of the inverse problem with a global
Figure 6: Comparison of observations for Italy and modelled data with

set used for model calibration, i.e., they correspond to t1 and t2 .

minimization algorithm, for t0 = −9. Vertical dotted line delimit the data
the parameters obtained by solution of the inverse problem with a global
Figure 5: Comparison of observations for Italy and modelled data with

14
Giudici et al., Inversion of a SIR-based model: application to COVID-19 epidemic
Number of cases Number of cases

100

101

102

103

104

105

106

100

101

102

103

104

105

106
1/22/20
1/23/20 1/22/20
1/23/20
1/24/20
1/25/20 1/24/20
1/25/20
1/26/20
1/27/20 1/26/20
1/27/20
1/28/20
1/29/20 1/28/20
1/29/20
1/30/20
1/31/20 1/30/20
1/31/20
2/1/20
2/2/20 2/1/20
2/2/20
2/3/20
2/4/20 2/3/20
2/4/20
2/5/20
2/6/20 2/5/20
2/6/20
2/7/20
2/8/20 2/7/20
2/8/20
2/9/20
2/10/20 2/9/20
2/10/20
2/11/20
2/12/20 2/11/20
2/12/20
2/13/20
2/14/20 2/13/20
2/14/20
2/15/20
2/16/20 2/15/20
2/16/20
2/17/20
2/18/20 2/17/20
2/18/20
2/19/20
2/20/20 2/19/20
2/20/20
2/21/20
2/22/20 2/21/20
2/22/20
2/23/20
2/24/20 2/23/20
2/24/20
2/25/20
2/26/20 2/25/20
2/26/20
2/27/20
2/28/20 2/27/20
2/28/20
2/29/20
3/1/20 2/29/20
3/1/20
Time (days)

Time (days)
3/2/20
3/3/20 3/2/20
3/3/20
3/4/20
3/5/20 3/4/20
3/5/20
3/6/20
3/7/20 3/6/20
3/7/20
3/8/20
3/9/20 3/8/20
3/9/20
3/10/20
3/11/20 3/10/20
3/11/20
3/12/20
3/13/20 3/12/20
3/13/20
3/14/20
3/15/20 3/14/20
3/15/20
3/16/20
3/17/20 3/16/20
3/17/20
3/18/20
3/19/20 3/18/20
3/19/20
3/20/20
3/21/20 3/20/20
3/21/20
3/22/20
3/23/20 3/22/20
3/23/20
3/24/20
3/25/20 3/24/20
3/25/20
3/26/20
3/27/20 3/26/20
3/27/20
3/28/20
3/29/20 3/28/20
3/29/20
3/30/20 3/30/20
Population (244966)
Recovered (39928)
Dead (25580)
Infected (154512)
Susceptibles (27238)
Observed recovered (38092)
Observed dead (21645)
Observed infected (165155)

Population (202162)
Recovered (40271)
Dead (25434)
Infected (143076)
Susceptibles (9275)
Observed recovered (38092)
Observed dead (21645)
Observed infected (165155)
3/31/20
4/1/20 3/31/20
4/1/20
4/2/20
4/3/20 4/2/20
4/3/20
4/4/20
4/5/20 4/4/20
4/5/20
4/6/20
4/7/20 4/6/20
4/7/20
4/8/20
4/9/20 4/8/20
4/9/20
4/10/20
4/11/20 4/10/20
4/11/20
4/12/20
4/13/20 4/12/20
4/13/20
4/14/20
4/15/20 4/14/20
4/15/20
 Giudici et al., Inversion of a SIR-based model: application to COVID-19 epidemic 15

4 Discussion
4.1 Remarks about the model
Some basic assumptions, on which the model developed in this work is
funded, deserve to be recalled and discussed.
The developed model basically assumes “homogeneity” of the popula-
tion. In other words, no distinction is done in terms of sex, age, economic
wealth, health and wellness, working conditions, life style, home state, and
any other, including genetic background. Also, the model assumes that
the population under study is a closed system, thus disregarding variations
induced by short-time, touristic or business travels, by intermediate-time
mobility of students and workers and by long-time effects of migrant fluxes.
The model is also independent of the climatic and environmental condi-
tions, i.e., the processes considered by the model are assumed to be indepen-
dent of the variability of weather conditions and environmental quality at
any temporal and space scale. In particular, this means that neither sharp
and rapid variations nor annual or seasonal cycling should affect these pro-
cesses.
Epidemic models rarely consider birth and death rates, because the cor-
responding terms in the equations are usually negligible. In this work, these
terms have been kept, in order to facilitate this discussion. In particular,
following the assumption of population homogeneity, it is assumed that in-
fected pregnant women give birth to infected babies and that this occur at
the same rate as for susceptible women.
With regard to infection rate, which is described by the term γIS/P in
(1), some remarks are in order. This term is computed by assuming that
each infected individual has a given, constant number of contacts with other
persons per unit time. The fraction of contacted persons which cannot be
infected is given by (I + R)/P , which assumes that recovered people become
immune to the virus, an aspect which is not confirmed by the scientific
community (see, e.g., [15]). Moreover, recovered people are assumed to be
not infectious, which is the case if the response of their immune system
is so fast that, once they come in contact with the virus again, the virus
is destroyed by the immune system before it can be spread to susceptible
persons. On the other hand, the fraction of contacted individuals which can
be infected is given by S/P . The γ coefficient, due to the “homogeneity”
assumption, is considered to be independent of the factors which have been
recalled at the beginning of this subsection; in particular, working and living
conditions could control the distance and the duration of contacts of infected
- and therefore infectious - individuals with other persons.
The so-called recovery and fatality coefficients ρ and φ are assumed to
be constant. This is not based on the “homogeneity” assumption only. In
fact, this implies that recovery and fatality are modeled as instantaneous
16 Giudici et al., Inversion of a SIR-based model: application to COVID-19 epidemic

processes, i.e., independent of the time passed since infection; moreover, no

distinction is done among death or healing of infected people according to
the strength of their symptoms and to the location where they are treated
(home and hospitals non-intensive, or Intensive Care Units – ICUs). The
latter condition could be modeled by subdividing the class of infected people
among sub-classes, e.g., asymptomatic, with light symptoms, admitted to
hospital non-intensive care units, admitted to ICUs.
Accounting for the time from infection is slightly more complex, but
could be handled, for instance, by introducing functions (φ̃, ρ̃) of elapsed
time since infection. Such functions should enter in a deconvolution product
involving the number of persons who have been infected at a given time and
are still infected, i.e., are not yet recovered or passed away. With this
approach, φI and ρI in (1) could be replaced by
Z τmax Z τmax
˜ − τ ) dτ and
φ̃(τ )I(t ˜ − τ ) dτ,
ρ̃(τ )I(t
0 0
 Z τh  (20)
dI ′ ′
i
′
where I(t − τ ) = (t − τ ) exp − ρ̃(τ ) + φ̃(τ ) dτ − δτ .
dt 0

Notice that the fatality coefficient, φ, accounts for the deaths related to
the pandemic, i.e., it represents the increase in the death rate due to the
pandemic. The normal death rate is considered through δ.

4.2 Remarks about model calibration by solution of the in-

verse problem
The results presented in section 3.2 show some of the classical, well known
difficulties of non-linear least-squares inversion, in particular the dependence
of the solution on the starting values, related to the existence of multiple
local minima, and the flatness of the objective function around the local
minima.
Better results have been obtained with the “differential evolution” al-
gorithm. Obviously, different algorithms for global optimization could be
tested, like, e.g., genetic algorithms [2], particle swarm optimization [8],
simulated annealing [11].
With reference to the specific example under study, it is necessary to
stress some aspects, mostly related to the role of data on model calibration
[6].
First of all, the solutions obtained by means of a global optimization
algorithm for high values of the threshold ξ and different intervals of ad-
missible values for t0 (Figures 5 and 6), show that the optimal value of P0
is smaller than the total Italian population. This parameter P0 has been
included in p(cal) with the objective of assessing the extension of the refer-
ence population. In other words, including P0 among the parameters to be
 Giudici et al., Inversion of a SIR-based model: application to COVID-19 epidemic 17

calibrated might provide a, possibly very rough, estimate of the width of

the initial population whose evolution is represented by the model. In this
particular instance, the results suggest that the reference initial population
does not cover the whole country, but only a limited portion.
The latter remark seems to go in tandem with the well-known fact that
in the countries most affected by COVID-19, the epidemic spread of the
virus had mostly concentrated in specific areas: the province of Hubei, and
above all the city of Wuhan, in China; the Lombardy region, and above all
the provinces of Bergamo, Brescia, Lodi and Milan, in Italy; the city of New
York in the USA; Île-de-France in France; Madrid and Catalunya in Spain.
Finally, it is quite difficult to assess the quality of data on the COVID-19
pandemia, but their uncertainty is expected to be very high. For instance,
the true number of infected people “remains unknown because asymptomatic
cases or patients with very mild symptoms might not be tested and will not
be identified”, as recognized, e.g., by [1]. In an interview published on March
23rd 2020 by the Italian newspaper “La Repubblica”, Angelo Borrelli, head
of Dipartimento della Protezione civile (national civil protection depart-
ment) stated that a ratio of one certified case out of every 10 total cases
is credible. Furthermore, different criteria have been adopted by different
countries and institutions to define the various categories of infected, recov-
ered and deceased people by or with COVID-19. This fact has been widely
recognised as a cause of uncertainty in the collected data. Finally, censor-
ship on COVID-19 pandemics is reported by journalists and organization in
some countries.
As a consequence, the use of official data to perform reliable estimates is
questionable. In principle, stochastic approaches, e.g., the Bayesian frame-
work, could be very helpful to handle discrepancies between model pre-
dictions and observations. Unfortunately, in this case the systematic and
random errors could be so high as to make it very difficult to handle them
even in a stochastic framework.

5 Conclusions
The modeling exercis conducted within this work supports a series of re-
marks, which are summarized in this conclusive section, together with some
future perspectives.
Starting from some remarks about modeling aspects, the limitations of
classical SIR models have been recalled. These should be always recalled
and carefully considered especially for applications and when these models
are used as engines of decision support systems.
The main limitation is related to the “homogeneity” assumption, accom-
panied by the steadiness of the recovery and fatality coefficients. The latter
aspect could be taken care of as discussed in subsection 4.1 and might yield
18 Giudici et al., Inversion of a SIR-based model: application to COVID-19 epidemic

terms of the form given in (20).

The assumption of “homogeneity” could be relaxed by considering “dis-
tributed” models, similar to those applied for transport phenomena, e.g., for
diffusion of contaminants in the environment. Those models can account for
“diffusive” spread and for “advective” transport. However, the required pa-
rameterisation is often much finer than the one for lumped models, so that
the number of parameters to be calibrated strongly increases, and therefore
in absence of good quality data it could be difficult to perform a reliable
calibration and validation of the model for a practical application.
Promising classes of models are given by the use of stochastic models,
either under a Monte Carlo framework or by using assimilation techniques,
e.g., the Ensemble Kalman Filter (EnKF, see, e.g., [4]). In principle, Monte
Carlo models might be adapted in a relatively easy way to account for several
phenomena and also to consider the role of some aspects (e.g., sex, age,
health and wellness, etc.) on the probability of infection. On the other
hand, EnKF could provide a firm theoretical framework to improve model
predictions by means of uncertain data.
With regard to the specific application to COVID-19 epidemic, although
it could be improvident to draw quantitative conclusions, it is nevertheless
qualitatively confirmed that infection started quite earlier than the certain
appearance of the first episodes of infection. The results of model inversion
also suggest that the calibrated model could be reliable for a portion of the
whole population. Somehow, the model itself, through its calibration, seems
to suggest the width of the population for which its approximations could
be valid.
Last, but not least for its practical importance, this paper has the ambi-
tion to provide further evidence about the great care that has to be given to
the quality of pandemic data, when used to calibrate or validate epidemic
models. In fact, poor quality data might yield unrealistic parameter values
and, therefore, unreliable model predictions.

Acknowledgments
This work presents results of a purely curiosity-driven research, which has
received support only through the standard working facilities of the authors’
institutions.

References
[1] David Baud, Xiaolong Qi, Karin Nielsen-Saines, Didier Musso,
Léo Pomar, and Guillaume Favre, Real estimates of mortality fol-
lowing covid-19 infection, The Lancet Infectious Diseases (2020),
DOI:10.1016/S1473-3099(20)30195-X.
 Giudici et al., Inversion of a SIR-based model: application to COVID-19 epidemic 19

[2] Lawrence Davis, Handbook of genetic algorithms, Van Nostrand Rein-

hold, New York, 1991.

[3] Ensheng Dong, Hongru Du, and Lauren Gardner, An interactive web-
based dashboard to track covid-19 in real time, The Lancet Infectious
Diseases (2020), DOI:10.1016/S1473-3099(20)30120-1.

[4] Geir Evensen, The ensemble kalman filter: theoretical formulation

and practical implementation, Ocean Dynamics 53 (2003), 343–367,
DOI:10.1007/s10236-003-0036-9.

[5] Roger Fletcher, Practical methods of optimization, second edition ed.,

John Wiley & Sons, Ltd, 2013.

[6] Mauro Giudici, Development, calibration and validation of physical

models, Geographic Information Systems and Environmental Model-
ing (M. C. Krane K. C. Clarke, B. O. Parks, ed.), Prentice-Hall, Upper
Saddle River (NJ), 2001, pp. 100–121.

[7] Mauro Giudici, Fulvia Baratelli, Laura Cattaneo, Alessandro Comu-

nian, Giovanna De Filippis, Cinzia Durante, Francesca Giacobbo, Sil-
via Inzoli, Mauro Mele, and Chiara Vassena, A conceptual framework
for discrete inverse problems in geophysics, 2019, arXiv:1901.07937.

[8] James Kennedy and Russell C. Eberhart, Particle swarm optimization,

Proceedings of ICNN’95 - International Conference on Neural Networks,
vol. 4, 1995, pp. 1942–1948.

[9] William Ogilvy Kermack and Anderson G McKendrick, A contribution

to the mathematical theory of epidemics, Proceedings of the Royal So-
ciety of London. Series A, Containing papers of a mathematical and
physical character 115 (1927), 700–721, DOI:10.1098/rspa.1927.0118.

[10] , Contributions to the mathematical theory of epidemics. ii. the

problem of endemicity, Proceedings of the Royal Society of London.
Series A, Containing papers of a mathematical and physical character
138 (1932), 55–83, DOI:10.1098/rspa.1932.0171.

[11] Scott Kirkpatrick, C Daniel Gelatt, and Mario P Vecchi, Op-

timization by simulated annealing, Science 220 (1983), 671–680,
DOI:10.1126/science.220.4598.671.

[12] Department of Economic and Social Affairs, 2018 demographic yearbook

annuaire démographique, sixty-ninth issue/soixante-neuvième édition
ed., United Nations, 2019.

[13] Ronald Ross, An application of the theory of probabilities to the study of

a priori pathometry.–part i, Proceedings of the Royal Society of London.
20 Giudici et al., Inversion of a SIR-based model: application to COVID-19 epidemic

Series A, Containing papers of a mathematical and physical character

92 (1916), 204–230, DOI:10.1098/rspa.1916.0007.

[14] Ronald Ross and Hilda P. Hudson, An application of the theory

of probabilities to the study of a priori pathometry.–part ii, Pro-
ceedings of the Royal Society of London. Series A, Containing pa-
pers of a mathematical and physical character 93 (1917), 212–225,
DOI:10.1098/rspa.1917.0014.

[15] Yufang Shi, Ying Wang, Changshun Shao, Jianan Huang, Jianhe Gan,
Xiaoping Huang, Enrico Bucci, Mauro Piacentini, Giuseppe Ippolito,
and Gerry Melino, Covid-19 infection: the perspectives on immune re-
sponses, Cell Death & Differentiation (2020), DOI:10.1038/s41418-020-
0530-3.

[16] Rainer Storn and Kenneth Price, Differential evolution a sim-

ple and efficient heuristic for global optimization over continu-
ous spaces, Journal of Global Optimization 11 (1997), 341–359,
DOI:10.1023/A:1008202821328.

Authors’ affiliations
Mauro Giudici
Università degli Studi di Milano, Dipartimento di Scienze della
Terra “A.Desio”, Milano, Italy
[email protected]

Alessandro Comunian
Università degli Studi di Milano, Dipartimento di Scienze della
Terra “A.Desio”, Milano, Italy
[email protected]

Romina Gaburro
University of Limerick, Department of Mathematics and Statistics,
Health Research Institute (HRI), Limerick, Ireland
[email protected]