This page intentionally left blank

C AMB RI D G E TR ACTS I N M AT H E M AT I C S

General Editors

B . B O LLO B Á S , W . FU LTO N , A . K AT O K , F. K I RWA N ,

P. S A RN A K, B . S I M O N , B . T O TA RO

181 Totally Positive Matrices

Totally Positive Matrices

ALLAN PINKUS

Technion, Israel
CAMBRIDGE UNIVERSITY PRESS
Cambridge, New York, Melbourne, Madrid, Cape Town, Singapore,
São Paulo, Delhi, Dubai, Tokyo

Cambridge University Press

The Edinburgh Building, Cambridge CB2 8RU, UK

Published in the United States of America by Cambridge University Press, New York

www.cambridge.org
Information on this title: www.cambridge.org/9780521194082
© A. Pinkus 2010

This publication is in copyright. Subject to statutory exception and to the

provision of relevant collective licensing agreements, no reproduction of any part
may take place without the written permission of Cambridge University Press.
First published in print format 2009

ISBN-13 978-0-511-68889-8 eBook (EBL)

ISBN-13 978-0-521-19408-2 Hardback

Cambridge University Press has no responsibility for the persistence or accuracy

of urls for external or third-party internet websites referred to in this publication,
and does not guarantee that any content on such websites is, or will remain,
accurate or appropriate.
This monograph is dedicated to the memory of
I. J. Schoenberg, M. G. Krein, F. R. Gantmacher and S. Karlin,
the four pioneers of the theory of total positivity.

We work in the dark – we do what we can – we give what we have.

Our doubt is our passion, and our passion is our task.
 Contents

Foreword page ix
1 Basic properties of totally positive and strictly
totally positive matrices 1
1.1 Preliminaries 1
1.2 Building (strictly) totally positive matrices 5
1.3 Nonsingularity and rank 12
1.4 Determinantal inequalities 24
1.5 Remarks 33
2 Criteria for total positivity and strict total
positivity 36
2.1 Criteria for strict total positivity 37
2.2 Density and some further applications 41
2.3 Triangular total positivity 47
2.4 LDU factorizations 50
2.5 Criteria for total positivity 55
2.6 “Simple” criteria for strict total positivity 60
2.7 Remarks 74
3 Variation diminishing 76
3.1 Main equivalence theorems 76
3.2 Intervals of strict total positivity 83
3.3 Remarks 85
4 Examples 87
4.1 Totally positive kernels and Descartes systems 87
4.2 Exponentials and powers 88
4.3 Cauchy matrix 92
4.4 Green’s matrices 94
4.5 Jacobi matrices 97

vii
viii Contents

4.6 Hankel matrices 101

4.7 Toeplitz matrices 104
4.8 Generalized Hurwitz matrices 111
4.9 More on Toeplitz matrices 117
4.10 Hadamard products of totally positive matrices 119
4.11 Remarks 125
5 Eigenvalues and eigenvectors 127
5.1 Oscillation matrices 127
5.2 The Gantmacher–Krein theorem 130
5.3 Eigenvalues of principal submatrices 140
5.4 Eigenvectors 144
5.5 Eigenvalues as functions of matrix elements 149
5.6 Remarks 152
6 Factorizations of totally positive matrices 154
6.1 Preliminaries 154
6.2 Factorizations of strictly totally positive matrices 156
6.3 Factorizations of totally positive matrices 164
6.4 Remarks 167
Afterword 169
References 174
Author index 180
Subject index 182
 Foreword

In this monograph was present the central properties of finite totally

positive matrices. As such, the monograph has only six main chapters.
We consider the basic properties of such matrices, determinantal criteria
for when a matrix is totally positive, their variation diminishing properties,
various examples of totally positive matrices, their eigenvalue/eigenvector
properties, and factorizations of such matrices. Numerous topics are
excluded from this exposition. Total positivity is a theory of considerable
consequence, and the most glaring omissions of this monograph are
undoubtedly its various applications to diverse areas. Aside from the
many applications mentioned in Gantmacher, Krein [1950] and Karlin
[1968], applications can be found in approximation theory (see Schumaker
[1981], Pinkus [1985c]), combinatorics (see Brenti [1989], [1995], [1996]),
graph theory (see Fomin, Zelevinsky [2000], Berenstein, Fomin, Zelevinsky
[1996]), Lie group theory (see Lusztig [1994]), majorization (see Marshall,
Olkin [1979]), noncommutative harmonic analysis (see Gross, Richards
[1995]), shape preservation (see Goodman [1995]), computing using totally
positive matrices (see de Boor, Pinkus [1977], Koev [2005], Demmel, Koev
[2005], Koev [2007]), refinement equations and subdivision (see Cavaretta,
Dahmen, Micchelli [1991], Micchelli, Pinkus [1991]), and infinite totally
positive banded matrices (see Cavaretta, Dahmen, Micchelli, Smith [1981],
de Boor [1982], Smith [1983], Dahmen, Micchelli, [1986]). See also the many
references in these papers and also the many references to these papers.
There has been no attempt to make this monograph all-encompassing, and
we apologize to all who feel that their contributions to the theory have
been slighted as a consequence.
The theory of totally positive matrices is an odd bird in the matrix
theory aviary. Much of the motivation for its study has come from problems

ix
x Foreword

in analysis, and the main initiators and contributors to the theory were
analysts. I. J. Schoenberg was interested in the problem of estimating
the number of real zeros of a polynomial, and this led him to his work
on variation diminishing transformations (in the early 1930s) and Pólya
frequency sequences, functions, and kernels (late 1940s and early 1950s).
These, together with his work on splines (1960s and 1970s), are central
topics in the theory of total positivity. M. G. Krein was led to the
theory of total positivity via ordinary differential equations whose Green’s
functions are totally positive (mid 1930s). S. Karlin came to the theory
of total positivity (in the 1950s and 1960s) by way of statistics, reliability
theory, and mathematical economics. The two major texts on the subject
Oscillation Matrices and Kernels and Small Vibrations of Mechanical
Systems, by F. R. Gantmacher and M. G. Krein (see Gantmacher, Krein
[1950]), and Total Positivity. Volume 1, by S. Karlin (see Karlin [1968]),
are a blend of analysis and matrix theory (and in the latter case the
emphasis is most certainly on analysis). (Their companion volumes The
Markov Moment Problem and Extremal Problems, by M. G. Krein and
A. A. Nudel’man (see Krein, Nudel’man [1977]) and Tchebycheff Systems:
with Applications in Analysis and Statistics, by S. Karlin and W. J. Studden
(see Karlin, Studden [1966]), are totally devoted to topics of analysis.)
Thankfully we have the short monograph of T. Ando that eventually
appeared as Ando [1987] (it was written a few years earlier) and was
devoted to totally positive matrices. The present monograph is an attempt
to update and expand upon Ando’s monograph. A considerable amount of
research has been devoted to this area in the past twenty years, and such
an update is certainly warranted.
It was Schoenberg, in Schoenberg [1930], who coined the term total positiv
(in German). Krein and Gantmacher (see Gantmacher, Krein [1935]),
unaware of Schoenberg’s earlier paper, used the term complètement non
négative and complètement positive (French) for totally positive and strictly
totally positive, respectively. As such, many authors use the term totally
nonnegative and totally positive for totally positive and strictly totally
positive, respectively, which, aside from the lack of consistency and order,
all too often leads to confusion. We follow the Schoenberg/Karlin/Ando
terminology.
It is a pleasure to acknowledge the help of Carl de Boor and David
Tranah. All errors, omissions and other transgressions are the author’s
responsibility.
 Foreword xi

I would like to close this short foreword with a personal note. My first
mathematical paper (jointly written with my doctoral supervisor Sam
Karlin) was in the area of total positivity. It is said that as one gets old(er)
one often returns to one’s first love. I plead guilty on both counts.

Haifa, 2008.
 1
Basic properties of totally positive and
strictly totally positive matrices

In this chapter, we introduce some of the notation, basic definitions, and

various classic facts and formulæ. Many of the results of this chapter
will be used in subsequent chapters. Matrix notation, more especially the
notation used for submatrices and minors, is both clumsy and problematic.
It definitely takes getting used to. But rest assured that one does eventually
get used to it. The medium here is not the message.
In Section 1.2 we consider some simple and less simple operations that
preserve total positivity and strict total positivity. Section 1.3 is about
nonsingularity and rank. Here we immediately see results that are less
than obvious. Finally, in Section 1.4, we present a few basic determinantal
inequalities that are valid for totally positive and strictly totally positive
matrices.

1.1 Preliminaries
For a positive integer n, and for each p ∈ {1, . . . , n}, we define the simplex

Ipn := {i = (i1 , . . . , ip ) : 1 ≤ i1 < · · · < ip ≤ n}

in Zp+ . That is, Ipn denotes the set of strictly increasing sequences of p
integers in {1, . . . , n}.
We use the following notation to define submatrices and minors of a
matrix. If A = (aij )ni=1 m j=1 is an n × m matrix, then for each i ∈ Ip and
n

j ∈ Iq we let
m



i i1 , . . . , ip p
A[i, j] = A =A := (aik j
)k=1 q
=1
j j1 , . . . , jq
denote the p × q submatrix of A determined by the rows indexed i1 , . . . , ip

1
2 Basic properties of totally positive matrices

and columns indexed j1 , . . . , jq . When p = q then

   
i i1 , . . . , ip p
A(i, j) = A =A := det (aik j
)k,
=1
j j1 , . . . , jp
denotes the associated minor, i.e., the determinant of the submatrix.
This monograph is about totally positive and strictly totally positive
matrices. They are defined as follows:

Definition 1.1 An n × m matrix A is said to be totally positive (TP) if

all its minors are nonnegative, i.e.,
 
i1 , . . . , ip
A(i, j) = A ≥0 (1.1)
j1 , . . . , jp
for all i ∈ Ipn , j ∈ Ipm , and all p = 1, . . . , min{n, m}. It is said to be strictly
totally positive (STP) if strict inequality always holds in (1.1).

We use and reuse various classic facts and formulæ. We list some of them
here for easy reference.

Cauchy–Binet and p th compound matrices The p th compound

 mof the n × m matrix A is denoted by A[p] and is defined as the
matrix
p × p matrix with entries
n

(A(i, j))i∈Ipn ,j∈Ipm

where the i ∈ Ipn and j ∈ Ipm are arranged in lexicographic order, i.e., for
distinct i, k ∈ Ipn we set i > k if the first nonzero term in the sequence
i1 − k1 , . . . , ip − kp is positive.
Assume B = CD, where B is an n × m matrix, C is an n × r matrix,
and D is an r × m matrix. The Cauchy–Binet formula may be written as
follows. For each i ∈ Ipn , j ∈ Ipm ,

B(i, j) = C(i, k)D(k, j) ,
k∈Ipr

i.e.,
      
i1 , . . . , ip i1 , . . . , ip k1 , . . . , kp
B = C D ,
j1 , . . . , jp k1 , . . . , kp j1 , . . . , jp
1≤k1 <···<kp ≤r

or, alternatively,
B[p] = C[p] D[p] .
 1.1 Preliminaries 3

This is, of course, a generalization of the formula for matrix multiplication.

For p = 1 the above reduces to

r
bij = cik dkj .
k=1

This Cauchy–Binet formula is valid when p ≤ min{n, m, r}. For p > r (p ≤

min{n, m}) the set Ipr is empty, and in the above sum we set B(i, j) = 0.
This is the “correct convention” as rank B ≤ r.
For p vectors u1 = (u1,1 , . . . , un,1 ), . . . , up = (u1,p , . . . , un,p ) ∈ Cn , and
each i ∈ Ipn we set
 1  p
u ∧ · · · ∧ up (i) = det (ui
,j )
,j=1 .
n
We consider u1 ∧ · · · ∧ up as a vector in C( p ) . It is termed the Grassman
product or wedge product or exterior product of u1 , . . . , up . Obviously
u1 ∧ · · · ∧ up = 0 if and only if the u1 , . . . , up are linearly dependent. From
the Cauchy–Binet formula it easily follows that

A[p] (u1 ∧ · · · ∧ up ) = Au1 ∧ · · · ∧ Aup .

Sylvester’s Determinant Identity Let A be an n × m matrix, and let

(α1 , . . . , αp ) ∈ Ipn and (β1 , . . . , βp ) ∈ Ipm .

For each i ∈ {1, . . . , n}\{α1 , . . . , αp } and j ∈ {1, . . . , m}\{β1 , . . . , βp } we

set
 
k1 , . . . , kp+1
bij = A
1 , . . . , p+1
where {k1 , . . . , kp+1 } is the set of integers {α1 , . . . , αp , i} arranged in
natural (increasing) order, and {1 , . . . , p+1 } is the set of integers
{β1 , . . . , βp , j} arranged in natural order. We generally abuse notation by
writing
 
α1 , . . . , αp , i
bij = A .
β1 , . . . , βp , j
But it is always to be understood that we have arranged these row and
column indices in natural order.
Sylvester’s Determinant Identity states that the minors of the (n − p) ×
(m − p) matrix B = (bij ) satisfy
 
 r−1  
i1 , . . . , ir α1 , . . . , αp α1 , . . . , αp , i1 , . . . , ir
B = A A .
j1 , . . . , jr β1 , . . . , βp β1 , . . . , βp , j1 , . . . , jr
4 Basic properties of totally positive matrices

The submatrix


α1 , . . . , αp
A
β1 , . . . , βp
is called the pivot block.

Inverses If A = (aij ) is an n × n nonsingular matrix, then the elements

of its inverse A−1 = (cij ) satisfy

i+j 1, . . . , j, . . . , n
(−1) A
1, . . . , i, . . . , n
cij =  
1, . . . , n
A
1, . . . , n

where we use j to indicate that we have deleted the jth index. Thus, in
the numerator above, we have taken the determinant of the submatrix of A
obtained by deleting the jth row and ith column. More generally we have


p
i +j j1 , . . . , jn−p

  (−1) k=1 k k A i , . . . , i
i1 , . . . , ip 1 n−p
A−1 =  
j1 , . . . , jp 1, . . . , n
A
1, . . . , n
where i1 < · · · < ip and i1 < · · · < in−p are complementary indices in
{1, . . . , n}, as are the j1 < · · · < jp and j1 < · · · < jn−p

.

Laplace expansion by minors We will make use of the classical Laplace

expansion of a determinant given by
      
1, . . . , n
p i1 , . . . , ip i , . . . , in−p
A = (−1) r=1 ir +jr
A A 1 
.
1, . . . , n j1 , . . . , jp j1 , . . . , jn−p
1≤j <···<j ≤n
1 p

In the above, i1 < · · · < ip and i1 < · · · < in−p are complementary indices
in {1, . . . , n}, as are the j1 < · · · < jp and j1 < · · · < jn−p

; p is fixed; and
the summation is over all ordered p-tuples j1 < · · · < jp .

Principal minors If A is an n × n matrix, then its principal submatrices

are those submatrices of the form


i1 , . . . , ip
A .
i1 , . . . , ip
 1.2 Building (strictly) totally positive matrices 5

That is to say, principal submatrices are the square submatrices of A, all of

whose diagonal elements are diagonal elements of A. The principal minors
of A are their determinants
 
i1 , . . . , ip
A .
i1 , . . . , ip

Determinantal identity The following determinantal identity will prove

very useful. We state it here for easy reference. Let A be an n × m matrix.
Let 1 ≤ i1 < · · · < ir ≤ n and 1 ≤ j1 < · · · < jr+1 ≤ m. Then for any
k ∈ {1, . . . , r} and  ∈ {2, . . . , r}
  
i1 , . . . , ir i1 , . . . , ik , . . . , ir
A A
j1 , . . . , j
, . . . , jr+1 j2 , . . . , jr
  
i1 , . . . , ir i1 , . . . , ik , . . . , ir
=A A (1.2)
j2 , . . . , jr+1 j1 , . . . , j
, . . . , jr
  
i1 , . . . , ir i1 , . . . , ik , . . . , ir
+A A .
j1 , . . . , jr j2 , . . . , j
, . . . , jr+1

Proof Let B be the (r + 1) × (r + 1) matrix given by

 
ai1 j1 ··· ai1 jr ai1 jr+1
 .. .. .. .. 
 . 
B= . . . .
 ai j ··· air jr ai j 
r 1 r r+1

0 ··· 0 1

Apply Sylvester’s Determinant Identity, where the pivot block of size (r −

1) × (r − 1) is given by all but the kth and last row, and all but the first
and th column of B.

1.2 Building (strictly) totally positive matrices

If A is a totally positive or strictly totally positive matrix, then there are
other (strictly) totally positive matrices associated with A and derived
from A. There are also various operations that preserve the class of totally
positive and strictly totally positive matrices. We review some of these here.
The first series of propositions as presented here are easily verified. Their
proofs are left to the reader.
6 Basic properties of totally positive matrices

Proposition 1.2 Assume A is a (strictly) totally positive matrix. Then AT

(the transpose of A), as well as every submatrix of A and AT is (strictly)
totally positive.

Proposition 1.3 Assume A is an n × m (strictly) totally positive matrix.

Let B denote the matrix obtained from A by reversing the order of both its
rows and columns, i.e., if A = (aij )ni=1 m n m
j=1 , then B = (bij )i=1 j=1 , where
bij = an+1−i,m+1−j , i = 1, . . . , n, j = 1, . . . , m. The matrix B is (strictly)
totally positive.

This next proposition immediately follows from an application of the

Cauchy–Binet formula.

Proposition 1.4 If A is an n × m totally positive matrix and B an m × r

totally positive matrix, then AB is an n × r totally positive matrix. If m ≥
min{n, r}, A is an n × m strictly totally positive matrix and B an m × r
totally positive matrix of rank r, then AB is an n×r strictly totally positive
matrix. Similarly, if m ≥ min{n, r}, A is an n × m totally positive matrix
of rank n and B an m × r strictly totally positive matrix, then AB is an
n × r strictly totally positive matrix.

Note that if m < min{n, r}, then rank AB ≤ m and so AB cannot

possibly be strictly totally positive.

Proposition 1.5 The following operations preserve the class of (strictly)

totally positive matrices.
(i) Multiplying a row (column) by a positive scalar.
(ii) Adding a positive multiple of a row (column) to the preceding or the
succeeding row (column).
(iii) Adding a positive value to the (1, 1) entry of the matrix (and to the
(n, m) entry for an n × m matrix).

From the formulæ for minors of the inverse we also have the following.

Proposition 1.6 Assume A is a square strictly totally positive matrix.

Then DA−1 D is a strictly totally positive matrix, where D is the diagonal
matrix with diagonal entries alternately 1 and −1. If A is a nonsingular
totally positive matrix, then DA−1 D is a nonsingular totally positive
matrix.

In addition, from Sylvester’s Determinant Identity we have the following.

 1.2 Building (strictly) totally positive matrices 7

Proposition 1.7 Assume A is an n × m (strictly) totally positive matrix.

Fix 1 ≤ i1 < · · · < ik ≤ n and 1 ≤ j1 < · · · < jk ≤ m, and let
 
i1 , . . . , ik , i
bij = A
j1 , . . . , jk , j
for i ∈ {1, . . . , n}\{i1 , . . . , ik } and j ∈ {1, . . . , m}\{j1 , . . . , jk }. (Recall that
it is to be understood that we have arranged these row and column indices
in natural order.) Then B = (bij ) is an (n − k) × (m − k) (strictly) totally
positive matrix.
A result that looks similar is the following:

Theorem 1.8 Assume A is an n × m totally positive matrix. Given k, set

   
1, . . . , k 1, . . . , k, i
bij = aij A −A
1, . . . , k 1, . . . , k, j
for i = k +1, . . . , n, j = k +1, . . . , m. Then B = (bij ) is an (n−k)×(m−k)
totally positive matrix. Furthermore, rank B ≤ k, and if A is a strictly
totally positive matrix then all r × r minors of B are strictly positive for
r = 1, . . . , min{k, n − k, m − k}.
We defer the proof of this theorem to the end of this chapter. There are
additional operations that preserve strict total positivity and total positivity,
but they are not as obvious or as immediate as some of the previous
operations. We list two of them here, but defer their proof to Section 2.2.

Proposition 1.9 Assume A is an n × m strictly totally positive matrix.

For given 1 ≤ r < n set

aij,  i = 1, . . . , r, j = 2, . . . , m
cij = i−1,i
A 1,j , i = r + 1, . . . , n, j = 2, . . . , m.

Then C = (cij ) is an n × (m − 1) strictly totally positive matrix. If A is a

totally positive matrix, then C is a totally positive matrix.

Proposition 1.10 Assume A is an n × m strictly totally positive matrix.

Fix p < min{m, n} and set
 
i − p, . . . , i − 1, i
cij = A
1, . . . , p, j
for i = p + 1, . . . , n, j = p + 1, . . . , m. Then C = (cij ) is an (n − p) × (m − p)
strictly totally positive matrix. If A is a totally positive matrix, then C is
a totally positive matrix.
8 Basic properties of totally positive matrices

When p = 1, the above matrix C is a submatrix of the matrix C (with

r = 1) in Proposition 1.9.
A totally different operation that preserves total positivity and strict
total positivity is the following form of iteration. Let A = (aij )ni,j=1 be an
n × n matrix. We define the matrix B = (bij )ni,j=1 as follows:

b1j = a1j , j = 1, . . . , n,

and for i ≥ 2

n
bij = bi−1,k akj , j = 1, . . . , n.
k=1

Theorem 1.11 Let A and B be as defined above. If A is a (strictly) totally

positive matrix, then B is a (strictly) totally positive matrix.

Proof We prove the theorem assuming A is strictly totally positive. The

same proof holds, verbatim, for A totally positive.
Our proof will use induction arguments. From the definition we see that
bij > 0 for all i, j. Assume that we have proven that all p × p minors of B
are strictly positive for p = 1, . . . , r − 1. We prove that the same holds for
all r × r minors.
Given 1 ≤ i1 < · · · < ir ≤ n, 1 ≤ j1 < · · · < jr ≤ n, consider
 
i1 , . . . , ir
B .
j1 , . . . , jr

We first assume that i1 = 1. Expanding the above minor by its first row
we obtain
  
r  
i1 , . . . , ir s−1 i2 , . . . , ir
B = (−1) a1js B .
j1 , . . . , jr s=1
j1 , . . . , js , . . . , jr

As 1 < i2 < · · · < ir ≤ n, it follows from the Cauchy–Binet formula that

 
i2 , . . . , ir
B =
j1 , . . . , js , . . . , jr
    
i2 − 1, . . . , ir − 1 k1 , . . . , kr−1
B A .
k1 , . . . , kr−1 j1 , . . . , js , . . . , jr
1≤k1 <···<kr−1 ≤n
 1.2 Building (strictly) totally positive matrices 9

Thus
 
i1 , . . . , ir
B =
j1 , . . . , jr

r   
i2 − 1, . . . , ir − 1
(−1)s−1 a1js B
k1 , . . . , kr−1
s=1 1≤k1 <···<kr−1 ≤n
 
k1 , . . . , kr−1
A
j1 , . . . , js , . . . , jr
   r
i2 − 1, . . . , ir − 1 
= B (−1)s−1 a1js
k1 , . . . , kr−1
1≤k1 <···<kr−1 ≤n s=1
 
k1 , . . . , kr−1
A
j1 , . . . , js , . . . , jr
    
i2 − 1, . . . , ir − 1 1, k1 , . . . , kr−1
= B A .
k1 , . . . , kr−1 j1 , . . . , jr
2≤k1 <···<kr−1 ≤n

As each of the factors in the last sum is strictly positive (we use here the
induction hypothesis) we have that
 
i1 , . . . , ir
B > 0.
j1 , . . . , jr
We complete the proof, for this fixed r, by applying an induction
argument based on the value i1 . We have proved the result for i1 = 1.
Now assume that i1 > 1. From the Cauchy–Binet formula,
      
i1 , . . . , ir i1 − 1, . . . , ir − 1 k1 , . . . , kr
B = B A .
j1 , . . . , jr k1 , . . . , kr j1 , . . . , jr
1≤k1 <···<kr ≤n

By our assumption on A and induction hypothesis on B each factor in the

sum is strictly positive. This proves the theorem.

This next result is interesting as we only vary columns and we only

consider n × n minors of A (and thus do not really need the full total
positivity of A to obtain our result).
Let A be an n × m matrix where n < m, and assume that
 
1, . . . , n
A = 0.
1, . . . , n
We define the n × (m − n) matrix B = (bij ) as follows:
 
1, . . . , n
bij = A , i = 1, . . . , n, j = 1, . . . , m − n.
1, . . . , i, . . . , n, n + j
10 Basic properties of totally positive matrices

Then,

Theorem 1.12 For A and B as defined above,

 
 r−1
i1 , . . . , ir r(r−1) 1, . . . , n
B = (−1) 2 A
j1 , . . . , jr 1, . . . , n
 
1, . . . , n
A 
i1 , . . . , in−r , n + j1 , . . . , n + jr
where i1 , . . . , in−r is complementary to i1 , . . . , ir in {1, . . . , n}. Set C =
(cij )ni=1 m−n
j=1 where

cij = bn−i+1,j , i = 1, . . . , n, j = 1, . . . , m − n.

If A is a (strictly) totally positive matrix, then C is a (strictly) totally

positive matrix.

Proof Let D be the 2n×m matrix whose first n rows are the unit vectors ei ,
i = 1, . . . , n, and whose last n rows are A. We apply Sylvester’s Determinant
Identity with pivot block


n + 1, . . . , 2n
D .
1, . . . , n
That is, set
 
i, n + 1, . . . , 2n
eij = D , i = 1, . . . , n, j = 1, . . . , m − n.
1, . . . , n, n + j
Note that
 
i+1 1, . . . , n
eij = (−1) A = (−1)i+1 bij .
1, . . . , i, . . . , n, n + j
Therefore, from Sylvester’s Determinant Identity,
 

 
i1 , . . . , ir r+ rk=1 ik i1 , . . . , ir
B = (−1) E
j1 , . . . , jr j1 , . . . , jr

r

 r−1
n + 1, . . . , 2n
= (−1)r+ k=1 ik D
1, . . . , n
 
i1 , . . . , ir , n + 1, . . . , 2n
D .
1, . . . , n, n + j1 , . . . , n + jr
Now
   
n + 1, . . . , 2n 1, . . . , n
D =A
1, . . . , n 1, . . . , n
 1.2 Building (strictly) totally positive matrices 11

while
 
i1 , . . . , ir , n + 1, . . . , 2n
D =
1, . . . , n, n + j1 , . . . , n + jr

r
 
1, . . . , n
(−1) k=1 ik +k A 
i1 , . . . , in−r , n + j1 , . . . , n + jr
where i1 , . . . , in−r is complementary to i1 , . . . , ir in {1, . . . , n}.
Thus
 
i1 , . . . , ir
B =
j1 , . . . , jr

 r−1  
r(r−1) 1, . . . , n 1, . . . , n
(−1) 2 A A  .
1, . . . , n i1 , . . . , in−r , n + j1 , . . . , n + jr
The matrix C is obtained from B by simply interchanging the order of the
rows and therefore
 
 r−1  
i1 , . . . , ir 1, . . . , n 1, . . . , n
C = A A  .
j1 , . . . , jr 1, . . . , n i1 , . . . , in−r , n + j1 , . . . , n + jr
The result now follows.

There is another construction of strictly totally positive matrices, related

to the previous construction. Assume A is an n × m strictly totally positive
matrix. Form the (n + m) × m matrix D by adjoining to A the m × m
matrix C given by
 
0 0 ··· 0 1
 0 0 · · · −1 0 
 
C= .. .. . . .. ..  ,
 . . . . . 
(−1)m−1 0 ··· 0 0
i.e.,


A
D= .
C
It is readily verified that
   
α1 , . . . , αm i1 , . . . , ik
D =A
1, . . . , m j1 , . . . , jk

where {α1 , . . . , αm } = {i1 , . . . , ik , n + m + 1 − jm−k , . . . , n + m + 1 − j1 },
 
and {j1 , . . . , jm−k } is complementary to {j1 , . . . , jk } in {1, . . . , m}. When
k = 0, i.e., (α1 , . . . , αm ) = (n + 1, . . . , n + m), we set A(∅) = 1 in agreement
12 Basic properties of totally positive matrices

with the above equality. Thus the strict total positivity property of A is
equivalent to the fact
 
α1 , . . . , αm
D >0
1, . . . , m
for all 1 ≤ α1 < · · · < αm ≤ n + m.
Two elementary operations that preserve this latter property of D are
the following. First, we can cyclically shift the rows of D, where we multiply
the row going from the first to the last (or last to first) by (−1)m−1 . Second,
we can multiply D by any m × m matrix M with det M > 0.
Let E be the (n + m) × m matrix obtained from D by a simple forward
cyclic rotation of the rows, i.e., shift row r to row r + 1 and row n + m to
row 1, and multiply the new first row of E by (−1)m−1 . In addition, since
   
n + 1, . . . , n + m n, n + 1, . . . , n + m − 1
E =D >0
1, . . . , m 1, . . . , m
there exists an m × m matrix M with det M > 0 such that


B
F = EM =
C
where C is as was previously defined. Thus B is an n × m strictly totally
positive matrix. A calculation shows that
 
an,2  an,3  ··· an,m  1
 
 A 1,n 1,2 A 1,n
1,3 ··· A 1,m 1,n
a1,1 
 
B= .. .. .. .. .. .
 . . . . . 
       
n−1,n
A 1,2 n−1,n
A 1,3 · · · A 1,mn−1,n
an−1,1

It is also easily verified by direct calculation that this matrix is strictly

totally positive. In the above it is not sufficient that A be totally positive.

1.3 Nonsingularity and rank

Totally positive nonsingular matrices enjoy surprising and useful properties,
one of which is the following:

Theorem 1.13 Let A be an n×n totally positive nonsingular matrix. Then

all principal minors of A are strictly positive.

Prior to proving this result we first explain and prove an ancillary result
that will be of independent interest.
 1.3 Nonsingularity and rank 13

Zero entries of totally positive matrices and zero values of their minors
are evidence of boundary behavior within the class of totally positive
matrices and, as such, are not arbitrary in nature. A zero entry of a totally
positive matrix A or a zero minor of this totally positive matrix portends
linear dependence or “throws a shadow.” That is, under suitable linear
independence assumptions all minors of the same order to the right and
above it, or to the left and below it, are also zero. Let us define these notions
more precisely.

Definition 1.14 Let A = (aij )ni=1 m

j=1 .

(a) The right shadow of the entry aij is the i × (m − j + 1) submatrix

(ars )ir=1 ms=j .
(b) The left shadow of the entry aij is the (n − i + 1) × j submatrix
(ars )nr=i js=1 .
(c) The right shadow of the submatrix


i + 1, . . . , i + r
A
j + 1, . . . , j + r
is the (i + r) × (m − j) submatrix


1, . . . , i + r
A .
j + 1, . . . , m
(d) The left shadow of the submatrix


i + 1, . . . , i + r
A
j + 1, . . . , j + r
is the (n − i) × (j + r) submatrix


i + 1, . . . , n
A .
1, . . . , j + r
Note that in this definition we only consider submatrices with consecutive
rows and columns. This is not without reason, as will become evident from
Theorem 1.19. We can now state and prove:

Proposition 1.15 If A is an n × m totally positive matrix and


i + 1, . . . , i + r
rank A =r−1
j + 1, . . . , j + r
then at least one of the following holds. Either the rows i + 1, . . . , i + r or
the columns j + 1, . . . , j + r of A are linearly dependent, or the right or left
14 Basic properties of totally positive matrices

shadow of


i + 1, . . . , i + r
A
j + 1, . . . , j + r
has rank r − 1.

Proof We first prove the case r = 1. That is, we prove that if A is totally
positive and aij = 0, then at least one of the following holds. Either the ith
row or jth column of A is zero, or the right or left shadow of aij is zero.
Assume neither the ith row nor the jth column of A is zero. Let ai
> 0
for some . If  < j we prove that the right shadow of A is zero as follows.
For any r < i
 
r, i
0≤A = ar
aij − arj ai
= −arj ai
≤ 0.
, j
Since ai
> 0 we have arj = 0 for all r < i. As the jth column of A is not
zero we have a k > i for which akj > 0. Now, for any r ≤ i, s ≥ j,
 
r, k
0≤A = arj aks − ars akj = −ars akj ≤ 0,
j, s
and since akj > 0 we have ars = 0. Similarly, if  > j then it follows that
the left shadow of aij is zero.
Let us now assume that r > 1. As


i + 1, . . . , i + r
A
j + 1, . . . , j + r
is of rank r − 1 there are p, q ∈ {1, . . . , r} such that

i + 1, . . . , i

+ p, . . . , i + r
A > 0. (1.3)


j + 1, . . . , j + q, . . . , j + r

Set

i + 1, . . . , i

+ p, . . . , i + r, k
bk
= A
j + 1, . . . , j

+ q, . . . , j + r, 

for k ∈ {1, . . . , n}\{i + 1, . . . , i
+ p, . . . , i + r} and  ∈ {1, . . . , m}\{j +

1, . . . , j + p, . . . , j + r}, where we understand that the row and column
indices are rearranged in increasing order. Let B = (bk
). By Sylvester’s
Determinant Identity, we see that B is an (n − r + 1) × (m − r + 1) totally
positive matrix and bi+p,j+q = 0. As such we can apply the result proved
in the case r = 1.
 1.3 Nonsingularity and rank 15

If bi+p,
= 0 for all  ∈ {1, . . . , m}\{j + 1, . . . , j
+ q, . . . , j + r}, i.e.,

 
i + 1, . . . , i + r
A =0
j + 1, . . . , j

+ q, . . . , j + r, 
for all , then from (1.3) it follows that the rows i + 1, . . . , i + r are
linearly dependent. Similarly, if bk,j+q = 0 for all k ∈ {1, . . . , n}\{i +
1, . . . , i

+ p, . . . , i + r}, then the columns j + 1, . . . , j + r are linearly
dependent.
If the right shadow of bi+p,j+q vanishes, i.e.,

i + 1, . . . , i

+ p, . . . , i + r, k
A =0
j + 1, . . . , j

+ q, . . . , j + r, 

for all k ≤ i + p,  ≥ j + q, then from (1.3) we see that


1, . . . , i + r
A
j + 1, . . . , m
is of rank r − 1. Similarly, if the left shadow of bi+p,j+q vanishes, then


i + 1, . . . , n
A
1, . . . , j + r
is of rank r − 1.

Proposition 1.15, in the case r = 1, is sometimes referred to as the

“shadow lemma.” We now prove Theorem 1.13.

Proof of Theorem 1.13 We first prove directly that arr > 0 for all
r ∈ {1, . . . , n}. Assume arr = 0. From Proposition 1.15 we have four
options. But all four options contradict the nonsingularity of A. Obviously
we cannot have that the rth row or column of A is zero. Thus either the
left or right shadow of arr is zero. Assume it is the right shadow. Then
aij = 0 for all i ≤ r and all j ≥ r, implying that the first r rows of A are
linearly dependent. This is a contradiction and therefore arr > 0.
We derive the general result by applying an induction argument on the
size of the minor and using Sylvester’s Determinant Identity. We assume
that for any totally positive nonsingular n × n matrix (any n) all principal
minors of order at most p − 1 are strictly positive (p ≤ n). We prove that
this same result holds for all principal minors of order p . We have proved
the case p = 1. For any 1 ≤ i1 < · · · < ip ≤ n set
 
i1 , . . . , ip−1 , k
bk
= A ,
i1 , . . . , ip−1 , 
16 Basic properties of totally positive matrices

for k,  ∈ {1, . . . , n}\{i1 , . . . , ip−1 }, and let B = (bk
). As an immediate

consequence of Sylvester’s Determinant Identity and our induction
hypothesis, it follows that B is totally positive and nonsingular. Thus the
diagonal entries of B are strictly positive. As

 p−2  
i1 , . . . , ip−1 i1 , . . . , ip
0 < bip ip = A A
i1 , . . . , ip−1 i1 , . . . , ip
and, by our induction hypothesis, we have
 
i1 , . . . , ip−1
A >0
i1 , . . . , ip−1
it therefore follows that
 
i1 , . . . , ip
A > 0.
i1 , . . . , ip

For totally positive matrices there is also an interaction between its rank,
which of its rows and columns can be linearly (in)dependent, and the strict
positivity of specific minors. This we detail in the next series of results.

Proposition 1.16 Let A be an n × m totally positive matrix, and let ak

denote the kth row of A, k = 1, . . . , n. Given 1 = i1 < · · · < ir+1 = n,
assume that the r + 1 vectors ai1 , . . . , air+1 are linearly dependent, while
the r vectors ai1 , . . . , air and ai2 , . . . , air+1 are each linearly independent.
Then A is necessarily of rank r.

Proof Since the ai1 , . . . , air+1 are linearly dependent, while the ai2 , . . . , air+1
are linearly independent we can write

r+1
a i1 = cs ais .
s=2

As the ai1 , . . . , air are linearly independent there exist 1 ≤ j1 < · · · < jr ≤
m for which  
i1 , . . . , ir
A > 0.
j1 , . . . , jr
Substituting for ai1 we obtain
  
r+1    
i1 , . . . , ir is , i2 , . . . , ir ir+1 , i2 , . . . , ir
A = cs A = cr+1 A
j1 , . . . , jr s=2
j1 , j2 , . . . , jr j1 , j2 , . . . , jr
 
i2 , . . . , ir , ir+1
= (−1)r+1 cr+1 A .
j1 , . . . , jr−1 , jr
 1.3 Nonsingularity and rank 17

The matrix A is totally positive and the left-hand side is strictly positive.
Thus (−1)r+1 cr+1 > 0.
By assumption A is of rank at least r. If r = m there is nothing to
prove, while if r + 1 = n there is also nothing to prove. As such we assume
that r + 1 ≤ m and r + 1 < n. Let  ∈ {1, . . . , n}\{i1 , . . . , ir+1 }. Thus
1 = i1 <  < ir+1 = n. For every choice of 1 ≤ k1 < · · · < kr+1 ≤ m
  
r+1  
i1 , . . . , , . . . , ir is , i2 , . . . , , . . . , ir
A = cs A
k1 , k2 , . . . , kr+1 s=2
k1 , k2 , . . . , kr+1
 
ir+1 , i2 , . . . , , . . . , ir
= cr+1 A
k1 , k2 , . . . , kr+1
 
i2 , . . . , , . . . , ir , ir+1
= (−1)r cr+1 A .
k1 , k2 , . . . , kr+1
As A is totally positive and (−1)r cr+1 < 0, it follows that
 
i1 , . . . , , . . . , ir
A = 0.
k1 , k2 , . . . , kr+1
Since this is true for every choice of 1 ≤ k1 < · · · < kr+1 ≤ m, we have
that a
∈ span{ai1 , . . . , air }. From our assumption we also have air+1 ∈
span{ai1 , . . . , air }. Thus A is of rank r.

Proposition 1.17 Let A be an n × m totally positive matrix. Assume

1 = i1 < · · · < ir+1 = n and 1 = j1 < · · · < jr+1 = m. If
 
i1 , . . . , ir+1
A = 0,
j1 , . . . , jr+1
while
   
i1 , . . . , ir i2 , . . . , ir+1
A ,A > 0,
j1 , . . . , jr j2 , . . . , jr+1
then A is of rank r.

Proof Let ak denote the kth row of A. By assumption the r vectors

ai1 , . . . , air and the r vectors ai2 , . . . , air+1 are linearly independent. Thus
it suffices, by Proposition 1.16, to prove that the r + 1 vectors ai1 , . . . , air+1
are linearly dependent.
This latter result follows from an application of Proposition 1.16 to the
(r + 1) × m matrix


i1 , . . . , ir+1
B=A
1, . . . , m
18 Basic properties of totally positive matrices

where we exchange the roles of the rows and columns. That is, if bk denotes
the kth column of B, then, by assumption, the r + 1 vectors bj1 , . . . , bjr+1
are linearly dependent, while the r vectors bj1 , . . . , bjr and bj2 , . . . , bjr+1
are each linearly independent. Thus B is of rank r, implying that the r + 1
vectors ai1 , . . . , air+1 are linearly dependent.

Another consequence of Proposition 1.16 is the following, which tells

us something about the possible singularity structure of totally positive
matrices.

Proposition 1.18 Let A be an n × m totally positive matrix, and let ak

denote the kth row of A, k = 1, . . . , n. Given 1 = i1 < · · · < ir+1 < n,
assume that the r + 1 vectors ai1 , . . . , air+1 are linearly independent. Let
k > ir+1 (k ≤ n) and assume that the r + 1 vectors ai1 , . . . , air , ak are
linearly dependent. Then ak = 0.

Proof This proposition can be proved using the method of proof in

Proposition 1.16. But it also follows as a consequence of Proposition 1.16.
We explain this latter proof here.
We have that the rows ai1 , . . . , air , ak are linearly dependent, while the
a , . . . , air are linearly independent. If the rows ai2 , . . . , air , ak are linearly
i1

independent, then, from Proposition 1.16, the rank of the matrix composed
from the rows indexed i1 , . . . , ir , ir+1 , k is exactly r. But, by assumption,
the ai1 , . . . , air+1 are linearly independent. Thus the ai2 , . . . , air , ak are
necessarily linearly dependent. We repeat this argument, each time lopping
off the first vector, until we arrive at the desired result that ak is linearly
dependent, i.e., ak is the zero vector.

The above results allow us to totally characterize the possible vanishing

minors of a nonsingular totally positive matrix. We have the following.

Theorem 1.19 Let A be an n × n nonsingular totally positive matrix. Let

(αk , βk , rk )
k=1 be the set of all triples such that for each k ∈ {1, . . . , }
 
αk + 1, . . . , αk + rk
A = 0,
βk + 1, . . . , βk + rk

and no principal minors of the rk × rk submatrix


αk + 1, . . . , αk + rk
A
βk + 1, . . . , βk + rk
 1.3 Nonsingularity and rank 19

vanish. For such submatrices we have αk = βk . Moreover

 
i1 , . . . , ip
A =0
j1 , . . . , jp
if and only if for some k ∈ {1, . . . , } there is an rk ×rk principal submatrix
of


i1 , . . . , ip
A
j1 , . . . , jp
that lies in the right shadow of an above


αk + 1, . . . , αk + rk
A
βk + 1, . . . , βk + rk
if αk < βk , or in its left shadow if αk > βk .
In other words, what we are proving here is that for a nonsingular totally
positive matrix there are certain basic zero minors formed from consecutive
rows and columns and all other zero minors are derived from them. This
derivation is also of a very specific nature. It is a consequence of a vanishing
principal minor of the zero minor lying in the shadow of one of these basic
zero minors.

Proof Assume
 
αk + 1, . . . , αk + rk
A = 0,
βk + 1, . . . , βk + rk
and no principal minor of


αk + 1, . . . , αk + rk
A
βk + 1, . . . , βk + rk
vanishes.
As A is nonsingular it follows from Theorem 1.13 that αk = βk . In
addition, from Proposition 1.15 we have that each such vanishing minor
either throws a right or a left shadow. If αk < βk then it must throw a
right shadow since the left shadow of


αk + 1, . . . , αk + rk
A
βk + 1, . . . , βk + rk
is

αk + 1, . . . , n
A
1, . . . , βk + rk
which contains the nonsingular rk × rk principal submatrix


αk + 1, . . . , αk + rk
A .
αk + 1, . . . , αk + rk
20 Basic properties of totally positive matrices

But this contradicts Theorem 1.13. Similarly if αk > βk then


αk + 1, . . . , αk + rk
A
βk + 1, . . . , βk + rk
must throw a left shadow.
Now if

i1 , . . . , ip
A
j1 , . . . , jp
is such that it contains an rk × rk principal submatrix that lies in the right
shadow of one of the


αk + 1, . . . , αk + rk
A
βk + 1, . . . , βk + rk
if αk < βk , or in its left shadow if αk > βk , then from Proposition 1.15 this
principal minor of


i1 , . . . , ip
A
j1 , . . . , jp
must vanish. It now follows from Theorem 1.13 that we must have
 
i1 , . . . , ip
A = 0.
j1 , . . . , jp
This proves the easier direction of the theorem.
Let us now assume that
 
i1 , . . . , ip
A =0
j1 , . . . , jp
for some choice of 1 ≤ i1 < · · · < ip ≤ n and 1 ≤ j1 < · · · < jp ≤ m.
We assume, in what follows, that no principal minors of this minor vanish.
(Otherwise replace the above by a principal minor with this same property.)
As  
i1 , . . . , ip
A =0
j1 , . . . , jp
while    
i1 , . . . , ip−1 i2 , . . . , ip
A ,A >0
j1 , . . . , jp−1 j2 , . . . , jp
it follows from Proposition 1.17 that the (ip − i1 + 1) × (jp − j1 + 1) totally
positive matrix


i1 , i1 + 1, . . . , ip
A
j1 , j1 + 1, . . . , jp
(composed of consecutive rows and columns) has rank p − 1. In this proof
 1.3 Nonsingularity and rank 21

we can assume p > 1 as the case p = 1 is a direct consequence of Proposition

1.15.
We claim that

jp − i1 ≤ p − 2 or ip − j1 ≤ p − 2.

Assume not. Then

jp − i1 ≥ p − 1 and ip − j1 ≥ p − 1. (1.4)

Set
α = max{i1 , j1 } − 1

We claim that
i1 ≤ α + 1 < · · · < α + p ≤ ip

and
j1 ≤ α + 1 < · · · < α + p ≤ jp .

To see this note that if α = i1 − 1, then α + p = i1 + p − 1 ≤ ip (since

the i1 , . . . , ip are increasing integers), while α + 1 = i1 ≥ j1 and from (1.4)
α + p = i1 + p − 1 ≤ jp . If α = j1 − 1 the same analysis applies. As


i1 , i1 + 1, . . . , ip
A
j1 , j1 + 1, . . . , jp
has rank p − 1, it therefore follows that
 
α + 1, . . . , α + p
A = 0,
α + 1, . . . , α + p
which contradicts Theorem 1.13. Thus

jp − i1 ≤ p − 2 or ip − j1 ≤ p − 2.

Let us assume that jp − i1 ≤ p − 2. The matrix


i1 , i1 + 1, . . . , i1 − p + 1
A
jp − p + 1, jp − p + 2, . . . , jp
composed of p consecutive rows and columns is of rank, at most, p−1. Thus
by Proposition 1.15 this submatrix throws a right or left shadow. From the
analysis of the first part of the proof of this theorem we see that it throws
a left shadow since i1 > jp − p + 1. That is,


i1 , . . . , n
A
1, . . . , jp
22 Basic properties of totally positive matrices

is of rank p − 1. As


i1 , . . . , ip
A
j1 , . . . , jp
lies in this submatrix we have that


i1 , . . . , ip
A
j1 , . . . , jp
lies in the left shadow of


α + 1, . . . , α + p
A
β + 1, . . . , β + p
where α = i1 − 1 and β = jp − p. In fact from our assumption that no
principal minors of


i1 , . . . , ip
A
j1 , . . . , jp
vanish, it necessarily follows that


α + 1, . . . , α + p
A
β + 1, . . . , β + p
is of rank exactly p − 1.
The case where ip − j1 ≤ p − 2 is handled similarly. That is, it follows
that the


i1 , . . . , ip
A
j1 , . . . , jp
is in the right shadow of the matrix


ip − p + 1, ip − p + 2, . . . , ip
A
j1 , j1 + 1, . . . , j1 + p − 1
of rank p − 1. This proves the theorem.

One simple application of Theorem 1.19 is the following. As we have seen

from Theorem 1.13, if A is totally positive and
 
i1 , . . . , ip
A >0
j1 , . . . , jp
then aik ,jk > 0, k = 1, . . . , p. In general the converse need not hold. We
claim that if the converse holds for consecutive row and column indices,
then it holds in general.
 1.3 Nonsingularity and rank 23

Proposition 1.20 Let A be an n × n nonsingular totally positive matrix.

Assume that  
i + 1, . . . , i + p
A >0
j + 1, . . . , j + p
if ai+k,j+k > 0, k = 1, . . . , p, for all possible i, j and p. Then for all
1 ≤ i1 < · · · < ip ≤ n, 1 ≤ j1 < · · · < jp ≤ n, and all p we have
 
i1 , . . . , ip
A >0
j1 , . . . , jp
if aik ,jk > 0, k = 1, . . . , p.

Proof Assume
 
i1 , . . . , ip
A = 0.
j1 , . . . , jp
From Theorem 1.19 there exist (α, β, r) such that
 
α + 1, . . . , α + r
A = 0,
β + 1, . . . , β + r
no principal submatrix of


α + 1, . . . , α + r
A
β + 1, . . . , β + r
vanishes, and some r × r principal submatrix of


i1 , . . . , ip
A
j1 , . . . , jp
lies in the right shadow of


α + 1, . . . , α + r
A
β + 1, . . . , β + r
if α < β, or in its left shadow if α > β.
The assumption of the proposition implies that r = 1 and thus
aα+1,β+1 = 0. We therefore have an α = β such that for some k ∈ {1, . . . , p}
the aik ,jk lies in the right shadow of aα+1,β+1 = 0 if α < β, or in its left
shadow if α > β. This implies that
aik ,jk = 0.

Totally positive matrices that satisfy the above property, i.e.,

 
i1 , . . . , ip
A >0
j1 , . . . , jp
24 Basic properties of totally positive matrices

if and only if
aik ,jk > 0, k = 1, . . . , p,

are called almost strictly totally positive matrices.

1.4 Determinantal inequalities

There are various determinantal inequalities that hold for totally positive
and strictly totally positive matrices. In this section we detail a few of
them.

Generalized Hadamard Inequalities We start with the generalized

Hadamard inequalities that hold for a wide class of matrices. In this
monograph we prove these inequalities only for totally positive matrices.
Let i, j, and k be disjoint sets in {1, . . . , n}. We prove the following, where
we slightly abuse notation.

Theorem 1.21 Let A be any n × n totally positive matrix and i, j, and k

be disjoint subsets of {1, . . . , n}. Then,

A(i, j, k)A(i) ≤ A(i, j)A(i, k). (1.5)

If A is strictly totally positive, then strict inequality holds in (1.5).

We recall what we mean by the above inequality. If i = (i1 , . . . , ip ) ∈ Ipn ,

j = (j1 , . . . , jq ) ∈ Iqn , k = (k1 , . . . , kr ) ∈ Irn are disjoint sets of ordered
indices in {1, . . . , n}, then the following inequalities hold between the
principal minors of A:
   
i1 , . . . , ip , j1 , . . . , jq , k1 , . . . , kr i1 , . . . , ip
A A
i1 , . . . , ip , j1 , . . . , jq , k1 , . . . , kr i1 , . . . , ip
   
i1 , . . . , ip , j1 , . . . , jq i1 , . . . , ip , k1 , . . . , kr
≤A A ,
i1 , . . . , ip , j1 , . . . , jq i1 , . . . , ip , k1 , . . . , kr
where we (always) assume that the row and column indices have been
rearranged in natural (increasing) order. This inequality is called a
Generalized Hadamard Inequality. It is sometimes written

A(s ∪ t)A(s ∩ t) ≤ A(s)A(t)

where s and t are arbitrary subsets of {1, . . . , n}. In our notation s = i ∪ j

and t = i ∪ k. If s ∩ t = i = ∅, then (1.5) remains valid and is generally
called a Hadamard Inequality. (For convenience we always set A(∅) = 1.)
 1.4 Determinantal inequalities 25

Proof We prove the result for a totally positive matrix. If A(i, j, k) = 0 then
(1.5) is certainly valid. As such we may assume that A(i, j, k) > 0. Thus
from Theorem 1.13 all the other minors in (1.5) are strictly positive.
Let
 
i, k
bk
= A .
i, 
Set |j| = q and |k| = r, where | · | denotes the cardinality of the set. From
Sylvester’s Determinant Identity
B(j) = A(i, j)A(i)q−1
B(k) = A(i, k)A(i)r−1
and
B(j, k) = A(i, j, k)A(i)q+r−1 .

Thus (1.5) is equivalent to the simpler Hadamard inequality

B(j, k) ≤ B(j)B(k), (1.6)

where j and k are disjoint sets of ordered indices in {1, . . . , n}, B is a totally
positive matrix, and B(j, k) > 0. It is this inequality that we now prove.
Our proof will be by induction on q + r = m. We always assume that
q, r ≥ 1. For m = 2 we must show that
     
j, k j k
B ≤B B = bjj bkk
j, k j k
where j = (j) and k = (k). As
 
j, k
B = bjj bkk − bjk bkj
j, k
and bjk , bkj ≥ 0, this inequality is immediate.
Assume m > 2 and q > 1. Let j1 ∈ j and j = j\{j1 } and set
 
j1 , k
ck
= B
j1 , 
for k,  ∈ j ∪ k. From Sylvester’s Determinant Identity, C = (ck
) is totally
positive and
C(j , k) > 0.

Now
C(j ) = B(j)bq−2
j1 j1
C(k) = B(k, j1 )br−1j1 j1
26 Basic properties of totally positive matrices

and
C(j , k) = B(j, k)bjq+r−2
1 j1
.

From the induction hypothesis and since |j | + |k| = m − 1

C(j , k) ≤ C(j )C(k),

implying, by the above set of equalities,

B(j, k)bj1 j1 ≤ B(j)B(k, j1 ).

Since |k| + 1 = r + 1 < m, we can again apply the induction hypothesis to

obtain
B(k, j1 ) ≤ B(k)bj1 j1 .

Substituting in the above gives us the desired inequality (1.6).

This same proof, with minor modifications in order to verify strict
inequality, is valid for strictly totally positive matrices.

Remark In all the determinantal inequalities in this section we consider

only principal minors. Recall (Proposition 1.2) that as all submatrices of
(strictly) totally positive matrices are themselves (strictly) totally positive
matrices these inequalities are also valid for nonprincipal minors. But be
careful with the bookkeeping.

There are various consequences of this generalized Hadamard

inequalities. We present two of them here.

Theorem 1.22 Let A be an n × n totally positive matrix. Let i1 , . . . , iq be

arbitrary subsets of {1, . . . , n}. Set

j
= {r : r belongs to at least  of the i1 , . . . , iq }.

Then
A(j1 ) · · · A(jq ) ≤ A(i1 ) · · · A(iq ).

If A is strictly totally positive, then strict inequality holds in the above.

Proof We prove this result for a totally positive matrix. The same proof,
with minor modifications, is valid for strictly totally positive matrices. Our
proof will be by induction on q. Note that the case q = 2 is exactly Theorem
1.21. Set

k
= {r : r belongs to at least  of the i1 , . . . , iq−1 }.

 1.4 Determinantal inequalities 27

From the induction hypothesis

A(k1 ) · · · A(kq−1 ) ≤ A(i1 ) · · · A(iq−1 ). (1.7)
Let
m
= {r : r belongs to iq and at least  of the i1 , . . . , iq−1 }.
Thus
j
= k
∪ m
−1 and m
= k
∩ m
−1 ,  = 1, . . . , q,
where m0 = iq , mq = kq = ∅ and A(∅) = 1. Applying Theorem 1.21 we
obtain
A(j
)A(m
) ≤ A(k
)A(m
−1 ),  = 1, . . . , q.
Thus

q 
q
A(j
)A(m
) ≤ A(k
)A(m
−1 ),

=1
=1

which reduces to
q−1

q 
A(j ) ≤


A(k
) A(m0 )

=1
=1

and from (1.7) this implies that

q−1
q  
q
A(j ) ≤


A(i
) A(iq ) = A(i
).

=1
=1
=1

Let A be an n × n matrix and for r = 1, . . . , n, set

  1
   (n−1
i1 , . . . , ir  r−1 )
Qr =  A .
i1 , . . . , ir
1≤i1 <···<ir ≤n

Thus

n
Q1 = aii
i=1
  1
   n−1
i1 , i2 
Q2 =  A
i1 , i2
1≤i1 <i2 ≤n

..
.  
1, . . . , n
Qn = A .
1, . . . , n
28 Basic properties of totally positive matrices

We prove that the Qr are a decreasing sequence of values.

Theorem 1.23 If A is an n × n totally positive matrix, then

Qn ≤ Qn−1 ≤ · · · ≤ Q1 .

If A is strictly totally positive, then strict inequalities hold in the above.

Proof We prove this result for totally positive matrices. The same proof,
with minor modifications, is valid for strictly totally positive matrices. Let
  1
   (nr)
i1 , . . . , ir 
Pr =  A , r = 1, . . . , n,
i1 , . . . , ir
1≤i1 <···<ir ≤n

and set P0 = 1. It is more convenient to work with the Pr .

We prove that Pr+1 Pr−1 ≤ Pr2 for r = 1, . . . , n − 1. Let us first consider
the case r = 1, i.e., P2 ≤ P12 . From the Hadamard inequality
 
r, s
A ≤ arr ass .
r, s

Thus, as is easily verified,

   
n
r, s
A ≤ an−1
ii
r, s
1≤r<s≤n i=1

from which
  2  n  n2
   n(n−1) 
r, s 
P2 =  A ≤ aii = P12 .
r, s
1≤r<s≤n i=1

For r = 2, . . . , n − 1 the proof is similar, but the bookkeeping is more

complicated. From the generalized Hadamard inequality we have
   
i1 , . . . , ir−1 , ir , ir+1 i1 , . . . , ir−1
A A
i1 , . . . , ir−1 , ir , ir+1 i1 , . . . , ir−1
   
i1 , . . . , ir−1 , ir i1 , . . . , ir−1 , ir+1
≤A A .
i1 , . . . , ir−1 , ir i1 , . . . , ir−1 , ir+1

Fixing i1 , . . . , ir−1 and taking products over ir < ir+1 (distinct from
 1.4 Determinantal inequalities 29

i1 , . . . , ir−1 ) we obtain
 
  
  n−r+1
 i1 , . . . , i , i , i
r−1 r r+1  i1 , . . . , ir−1 ( 2 )
A A
ir <ir+1
i1 , . . . , ir−1 , ir , ir+1 i1 , . . . , ir−1
 
  i1 , . . . , ir−1 , ir 
n−r

≤ A .
i
i1 , . . . , ir−1 , ir
r

We now vary over i1 , . . . , ir−1 to get

  r+1
   (r−1)
 i1 , . . . , ir−1 , ir , ir+1 
A ×
i1 , . . . , ir−1 , ir , ir+1
1≤i1 <···<ir+1 ≤n
  n−r+1
   ( 2 )
 i1 , . . . , ir−1 
A
i1 , . . . , ir−1
1≤i1 <···<ir−1 ≤n
 (n−r)(r−1
r
)
  
i1 , . . . , ir 
≤ A ,
i1 , . . . , ir
1≤i1 <···<ir ≤n

i.e.,
n r+1 n
(r+1)(r−1 ) (r−1 )(n−r+1) (n)(n−r)(r−1
r
)
Pr+1 Pr−1 2
≤ Pr r .

A simple calculation reduces this to

Pr+1 Pr−1 ≤ Pr2 .

Now
Qr/n
r = Pr , r = 1, . . . , n.

Thus we have
2r
r+1 Qr−1 ≤ Qr ,
Qr+1 r = 1, . . . , n − 1,
r−1

where Q0 = 1. For r = 1 this is

Q22 ≤ Q21

implying
Q2 ≤ Q1 .

Assume 2 ≤ r ≤ n − 1 and

Qr ≤ · · · ≤ Q1 .
30 Basic properties of totally positive matrices

From the above

2r
r+1 Qr−1 ≤ Qr ≤ Qr
Qr+1 r−1 r+1 r−1
Qr−1

and therefore
Qr+1 ≤ Qr .

Thus
Qn ≤ · · · ≤ Q1 .

There are other determinantal inequalities valid for totally positive

matrices. We here note one such additional inequality.

Proposition 1.24 Assume A = (aij ) is an n × n strictly totally positive

matrix, and we are given 0 ≤ ji ≤ n − i, i = 1, . . . , n − 1, with

0 ≤ jn−1 ≤ · · · ≤ j1 .

Define the n × n matrix C = (cij ) where, for each i,


0, j = 1, . . . , ji
cij =
aij , j = ji + 1, . . . , n.

Then
(−1)j1 +···+jn−1 det C > 0.

Proof We prove this result by induction and by the use of Sylvester’s

Determinant Identity. It is easily seen to hold for n = 2.
If jk = n − k for some k, then
   
k(n−k) 1, . . . , k k + 1, . . . , n
det C = (−1) C C
n − k + 1, . . . , n 1, . . . , n − k
and one can apply induction directly to obtain our result.
Let us therefore assume that n ≥ 3, the ji are nonincreasing in i, and
0 ≤ ji < n − i, i = 1, . . . , n − 2 (note that jn−1 = 0). Set
 
2, . . . , n − 1, i
bij = C , i, j ∈ {1, n}.
2, . . . , n − 1, j
Thus, from Sylvester’s Determinant Identity,
   
2, . . . , n − 1 1, . . . , n
b11 bnn − b1n bn1 = C C .
2, . . . , n − 1 1, . . . , n
 1.4 Determinantal inequalities 31

Let r denote the number of zeros in the first column of C. Then by the
induction hypothesis
 
j2 +···+jn−2 −(r−1) 2, . . . , n − 1
(−1) C > 0,
2, . . . , n − 1
and applying the induction hypothesis to each of the bij we get,
 
j1 +···+jn−2 j1 +···+jn−2 1, . . . , n − 1
(−1) b11 = (−1) C > 0,
1, . . . , n − 1
 
2, . . . , n
(−1)j2 +···+jn−2 −(r−1) bnn = (−1)j2 +···+jn−2 −(r−1) C > 0,
2, . . . , n
 
1, . . . , n − 1
(−1)j1 +···+jn−2 −r b1n = (−1)j1 +···+jn−2 −r C > 0,
2, . . . , n
 
2, . . . , n
(−1)j2 +···+jn−2 bn1 = (−1)j2 +···+jn−2 C > 0.
1, . . . , n − 1
Thus
(−1)j1 −(r−1) [b11 bnn − b1n bn1 ] > 0

and therefore
(−1)j1 +···+jn−1 det C > 0.

If A is only totally positive then the strict inequality in the statement of

Proposition 1.24 should be replaced by an inequality. This is most easily
verified by an appeal to Theorem 2.6 of the next chapter.
We now apply Proposition 1.24 (after interchanging row and column
indices) in order to prove Theorem 1.8.

Proof of Theorem 1.8 Recall that

   
1, . . . , k 1, . . . , k, i
bij = aij A −A
1, . . . , k 1, . . . , k, j
for i = k + 1, . . . , n, j = k + 1, . . . , m. Expand the rightmost minor by its
last row to obtain
 k  
s+k 1, . . . , k
bij = ais (−1) A .
s=1
1, . . . , s, . . . , k, j

From this expression we see that rank B ≤ k.

We shall prove that if A is strictly totally positive then all r × r minors
of B are strictly positive for r = 1, . . . , min{k, n − k, m − k}. (By appealing
32 Basic properties of totally positive matrices

to Theorem 2.6 we then can show that if A is only totally positive then B
is totally positive.)
Set
 
1, . . . , k
csj = (−1)s+k A , s = 1, . . . , k, j = k + 1, . . . , m.
1, . . . , s, . . . , k, j

Thus for r ≤ min{k, n − k, m − k} we have

      
i1 , . . . , ir i1 , . . . , ir s1 , . . . , sr
B = A C
j1 , . . . , jr s1 , . . . , sr j1 , . . . , jr
1≤s1 <···<sr ≤k

by the Cauchy–Binet formula.

From Theorem 1.12
 
 r−1
s1 , . . . , sr r(r−1)
1, . . . , k
kr+ 2 + r
=1 s

C = (−1) A
j1 , . . . , jr 1, . . . , k
 
1, . . . , k
A  ,
s1 , . . . , sk−r j1 , . . . , jr

where s1 , . . . , sk−r is complementary to s1 , . . . , sr in {1, . . . , k}.

Therefore
 
i1 , . . . , ir
B
j1 , . . . , jr
    
i1 , . . . , ir s1 , . . . , sr
= A C
s1 , . . . , sr j1 , . . . , jr
1≤s1 <···<sr ≤k

 r−1
r(r−1) 1, . . . , k
= (−1)kr+ 2 A ×
1, . . . , k

r
   
i1 , . . . , ir 1, . . . , k
(−1)
=1 s
A A  .
s1 , . . . , sr s1 , . . . , sk−r , j1 , . . . , jr
1≤s1 <···<sr ≤k

We now apply Proposition 1.24 to the (k + r) × (k + r) strictly totally

positive matrix


1, . . . , k, i1 , . . . , ir
A
1, . . . , k, j1 , . . . , jr

where we set as zero the r × r bottom right corner of this matrix, i.e., set as
zero the (i, j) entries for i, j = k + 1, . . . , k + r. It follows from Proposition
1.3 that Proposition 1.24 also holds in this case as interchanging all rows
and columns preserves strict total positivity. The resulting matrix has sign
 1.5 Remarks 33

(−1)r . Now apply Laplace’s expansion by minors to this matrix to obtain

  
i1 , . . . , ir
(−1)r (−1)(k+1+s1 )+···+(k+r+sr ) A
s1 , . . . , sr
1≤s1 <···<sr ≤k
 
1, . . . , k
×A  > 0.
s1 , . . . , sk−r , j1 , . . . , jr
Thus  
i1 , . . . , ir
B > 0.
j1 , . . . , jr

1.5 Remarks
The study of total positivity and strict total positivity for continuous
kernels predates the study of total positivity and strict total positivity for
matrices (see e.g., Kellogg [1918]). Such phenomena are not uncommon.
The three pioneers of the theory of totally positive matrices are
F. R. Gantmacher, M. G. Krein, and I. J. Schoenberg. It was Schoenberg
[1930] who first considered such matrices. He did so in his study of variation
diminishing properties (see Chapter 3). In fact Schoenberg also coined
the term total positiv (in German) in his 1930 paper. Krein came to
consider such kernels and (later) matrices as a consequence of his research
on ordinary differential equations whose Green’s kernel is totally positive.
The joint paper of Gantmacher, Krein [1937] (an announcement appeared
in Gantmacher, Krein [1935]) presented most of the main results relating to
spectral properties of totally positive matrices, and many other important
results concerning totally positive matrices (except that they were then
unaware of Schoenberg’s earlier paper and the variation diminishing
properties associated with such matrices). This 1937 paper, in a slightly
expanded form, is most of Chapter II of the book Gantmacher, Krein
[1950]. In Gantmacher, Krein [1937] they used the term complètement
non négative and complètement positive (French) for totally positive and
strictly totally positive, respectively. As such many authors use the terms
totally nonnegative and totally positive for totally positive and strictly
totally positive, respectively.
For a proof and history of the Cauchy–Binet formula, Sylvester’s
Determinant Identity and the Laplace expansion by minors, see Brualdi,
Schneider [1983] and references therein. The initial propositions of
Section 1.2 can, for the most part, be found in Gantmacher, Krein
[1937], Ando [1987], and Karlin [1968]. Theorem 1.8 is from Ando [1987],
Theorem 3.9. The proof as presented here is very different. Proposition 1.9
34 Basic properties of totally positive matrices

is due to Whitney [1952]. Theorem 1.11 is in Karlin [1968], Theorem 6.1,

p. 132. The construction at the end of Section 1.2 is from Boocher, Froehle
[2008]. Theorem 1.13 is contained in Karlin [1968], p. 89. The proof therein
is different. Proposition 1.15 is from de Boor, Pinkus [1982] (see also
Koteljanskii [1950]). Propositions 1.16 and 1.17 are from Gantmacher,
Krein [1937], and can also be found in Karlin [1968]. Theorem 1.19 is in
Pinkus [2008].
Almost strictly totally positive matrices were introduced in Gasca,
Micchelli, Peña [1992]. Proposition 1.20 can be found there, although the
proof is much different.
The simplest form of the Hadamard Inequality for totally positive
matrices can already be found in Gantmacher, Krein [1937], p. 450. The
form of the generalized Hadamard inequalities as given here is from
Koteljanskii [1950], where general conditions are given for a generalized
Hadamard inequality to hold. Generalized Hadamard inequalities hold for
positive definite matrices and M -matrices. A square matrix is an M -matrix
if all its principal minors are positive and all its off-diagonal entries are
nonpositive.
Theorem 1.22 is contained in Carlson [1968]. A generalization to

q
 (k−1 
r−1 )
A(jk ) ≤ A(i
1 ∩ · · · ∩ i
r )
k=r 1≤
1 <···<
r ≤q

for any r ≤ q may be found in Fan [1968]. (The case r = 1 is Theorem

1.21.) Theorem 1.23 for positive definite matrices is known as Szasz’s
inequality (see Beckenbach, Bellman [1961], p. 64). It only depends upon
the nonnegativity of the principal minors and the generalized Hadamard
inequalities and thus is also valid for totally positive matrices, as was noted
by Koteljanskii [1950]. It was generalized, for this same class of matrices,
in Fan [1967] to the following. Let A be an n × n matrix and let s1 , . . . , sq
be pairwise disjoint subsets of {1, . . . , n}, q ≥ 2. Set
  q−1
1

  i1  (r−1)
Qr =  A s ∪ · · · ∪ sir  .
1≤i1 <···<ir ≤q

Then
Qq ≤ · · · ≤ Q1 .
Other determinantal inequalities exist for (strictly) totally positive
matrices (see e.g., Mühlbach, Gasca [1985], Fallat, Gekhtman, Johnson
[2003], Skandera [2004] and Boocher, Froehle [2008]). The construction at
 1.5 Remarks 35

the end of Section 1.2 permits us to generalize many of the determinantal

inequalities of Section 1.4 (there must exist the same number of factors on
both sides of the inequality) to where we have nonprincipal minors (see
Boocher, Froehle [2008]).
Boocher, Froehle [2008], building on work of Skandera [2004] and Fallat,
Gekhtman, Johnson [2003], were able to characterize all determinantal
inequalities of the form
       
i1 , . . . , ik m1 , . . . , mp r1 , . . . , rt u1 , . . . , uq
A A ≤A A (1.8)
j1 , . . . , jk 1 , . . . , p s1 , . . . , st v1 , . . . , vq
valid for all n×n totally positive matrices. The analogous problem for more
than two factors has yet to be solved, although numerous results can be
found in these papers.
One characterization of the inequalities (1.8) is the following that appears
in Boocher, Froehle [2008]. For ease of exposition we introduce the following
notation. For each α = (α1 , . . . , αn ), a vector of n distinct integers in
{1, . . . , 2n}, we let
 
i1 , . . . , ik
A(α) = A ,
j1 , . . . , jk
where {α1 , . . . , αn } = {i1 , . . . , ik , 2n + 1 − j1 , . . . , 2n + 1 − jn−k

}, and
 
{j1 , . . . , jn−k } is complementary to {j1 , . . . , jk } in {1, . . . , n}. For k = 0,
i.e., α = (n + 1, . . . , 2n), we set A(α) = 1. Based on results of this chapter
we have shown or can easily show that
A(i, j + 1, ∆)A(i + 1, j, ∆) ≤ A(i, j, ∆)A(i + 1, j + 1, ∆) (1.9)
where i, j ∈ {1, . . . , 2n}, ∆ ⊂ {1, . . . , 2n} with |∆| = n − 2, i + 1, j + 1 are
understood mod 2n, and i, i+1, j, j +1, ∆ are all distinct. Boocher, Froehle
[2008] proved that a determinantal inequality of the form (1.8) holds for all
n × n totally positive matrices if and only if the inequality can be factored
as a product of inequalities of the form (1.9), where we vary over i, j, and ∆
as above. Other characterizations may be found in Boocher, Froehle [2008].
 2
Criteria for total positivity and strict
total positivity

In this chapter we discuss the problem of establishing determinantal

criteria for when a matrix is totally positive or strictly totally positive.
Totally different criteria will be discussed in the next chapter on variation
diminishing.
In Section 2.1 we prove Fekete’s Lemma and some of its consequences.
The most notable thereof is Theorem 2.3, which states that for a matrix to
be strictly totally positive it suffices to prove that all its minors composed
of the first k rows and k consecutive columns and all its minors composed
of the first k columns and k consecutive rows are strictly positive for all
possible k. We apply the results of Section 2.1 in Section 2.2, where we
prove that strictly totally positive matrices are dense in the class of totally
positive matrices, and provide proofs of Propositions 1.9 and 1.10 from
Chapter 1.
In Section 2.3 we discuss triangular matrices, detail determinantal
criteria for when such matrices are totally positive, and in Section 2.4
we consider the LDU -factorization of strictly totally positive and totally
positive matrices. We will return to the study of factorizations of totally
positive matrices in Chapter 6. In Section 2.5 we consider determinantal
criteria for when a matrix is totally positive. The results are nowhere near
as elegant as those valid for strictly totally positive matrices.
In Section 2.6 we prove a recent surprising and beautiful result of
O. M. Katkova and A. M. Vishnyakova (completing work initiated by
T. Craven and G. Csordas). They prove that if A = (aij ) is an n × n
matrix, all of whose entries are strictly positive, and if

 
π
aij ai+1,j+1 > 4 cos2 ai,j+1 ai+1,j
n+1

36
 2.1 Criteria for strict total positivity 37

for all i, j = 1, . . . , n − 1, then A is strictly totally positive. Furthermore,

the constant 4 cos2 n+1 π
in the above inequality is best possible.

2.1 Criteria for strict total positivity

A matrix A = (aij )ni=1 m
j=1 is strictly totally positive if and only if
 
i1 , . . . , ip
A >0
j1 , . . . , jp
for all 1 ≤ i1 < · · · < ip ≤ n, 1 ≤ j1 < · · · < jp ≤ m, and all p =
1, . . . , min{n, m}. Thus, formally, we must verify
min{n,m} 
    
n m n+m
= −1
p=1
p p n

conditions. For n = m this number is on the order of 4n n−1/2 . Thankfully,

it is not necessary to check all these conditions. For example, if ai1 > 0 for
i = 1, . . . , n, a1j > 0 for j = 1, . . . , m, and
 
1, i
A = a11 aij − ai1 a1j > 0,
1, j
for i = 2, . . . , n and j = 2, . . . , m, then we immediately obtain aij > 0 for
all i, j > 1. In other words, many of the inequalities in the definition of a
strictly totally positive matrix are superfluous. In this section we search for
reasonable determinantal criteria establishing strict total positivity.
We start with a classic result due to Fekete.

Lemma 2.1 (Fekete’s Lemma). Assume C is an n × m matrix (n ≥ m)

such that all (m − 1)st order minors with columns 1, . . . , m − 1 are strictly
positive and the mth order minors composed from consecutive rows are also
strictly positive. Then all mth order minors of C are strictly positive.

Proof For a given strictly increasing sequence i = (i1 , . . . , ik ) of k integers

in {1, . . . , n}, i.e., i ∈ Ikn , we define d(i), the dispersion of i, by

k
d(i) := (ij − ij−1 − 1) = (ik − i1 ) − (k − 1) ≥ 0.
j=2

d(i) counts the number of integers between i1 and ik that are not in the
sequence i1 , . . . , ik . Thus d(i) = 0 if and only if the sequence i is composed
of consecutive integers. We will prove this lemma by induction on d(·).
38 Criteria for total positivity and strict total positivity

Let i = (i1 , . . . , im ) be as above. We wish to prove that

 
i1 , . . . , im
C > 0.
1, . . . , m

By our assumption, this is valid when d(i) = 0.

For n = m there is nothing to prove, and as such we assume that n > m.
Let i = (i1 , . . . , im ) with d(i) = r > 0, and assume the result holds for all
i with d(i) < r. As the sequence i is not composed of consecutive integers,
we can add to this sequence an integer between i1 and im . Let 1 ≤ j1 <
· · · < jm+1 ≤ n be the resulting sequence assuming j
to be the new index.
Consider the determinantal identity (1.2) from Chapter 1:
  
j2 , . . . , jm j1 , . . . , j
, . . . , jm+1
C C
1, . . . , m − 1 1, . . . , m
  
j1 , . . . , j
, . . . , jm j2 , . . . , jm+1
= C C
1, . . . , m − 1 1, . . . , m
  
j2 , . . . , j
, . . . , jm+1 j1 , . . . , jm
+C C .
1, . . . , m − 1 1, . . . , m

In the above equation the three m − 1 order minors are all positive by
assumption, as they are all based on the columns 1, . . . , m − 1. Now d(i) =
(im − i1 ) − (m − 1) = (jm+1 − j1 ) − (m − 1) = r, while (jm+1 − j2 ) − (m − 1)
and (jm − j1 ) − (m − 1) are strictly less than r. Thus, by our induction
hypothesis, we have
   
j2 , . . . , jm+1 j1 , . . . , jm
C , C > 0.
1, . . . , m 1, . . . , m

This implies that

  
j1 , . . . , j
, . . . , jm+1 i1 , . . . , im
C =C > 0.
1, . . . , m 1, . . . , m

We will use the method of proof of Fekete’s Lemma at least twice more in
this chapter. An immediate consequence of the above result is the following.

Theorem 2.2 Let A be an n×m matrix. Assume that all kth order minors
of A composed from consecutive rows and consecutive columns are strictly
positive for k = 1, . . . , min{n, m}. Then A is strictly totally positive.
 2.1 Criteria for strict total positivity 39

However, we need not verify all these inequalities in order to determine

if A is strictly totally positive. It suffices to check the strict positivity of an
even smaller number of determinants.

Theorem 2.3 Let A be an n×m matrix. Assume that all kth order minors
of A composed from the first k rows and k consecutive columns, and also all
kth order minors of A composed from the first k columns and k consecutive
rows, are strictly positive for k = 1, . . . , min{n, m}. Then A is strictly
totally positive.

There are thus exactly nm determinantal inequalities to be checked. For

n = m this n2 is a significantly smaller value than 4n n−1/2 .

Proof We prove that the conditions in Theorem 2.3 imply that all kth
order minors of A composed from consecutive rows and columns are strictly
positive for k = 1, . . . , min{n, m}. We then apply Theorem 2.2 to obtain
our desired result. In other words we prove that
 
i + 1, . . . , i + k
A >0 (2.1)
j + 1, . . . , j + k
for i = 0, . . . , n − k, j = 0, . . . , m − k, and k = 1, . . . , min{n, m}. Our
assumption is that this result holds when i = 0 and when j = 0.
Let us first prove that (2.1) also holds when j = 1, i.e.,
 
i + 1, . . . , i + k
A > 0.
2, . . . , k + 1
We prove this by induction on i and on k. Assume i = 1. For k = 1 we
must prove that
 
2
A = a22 > 0.
2
Now
 
1, 2
A = a11 a22 − a21 a12 > 0.
1, 2
By assumption a11 , a21 , a12 > 0. Thus a22 > 0. Assume the result holds for
k − 1. By an application of Sylvester’s Determinant Identity,
       
1, . . . , k 2, . . . , k + 1 1, . . . , k 2, . . . , k + 1
A A −A A
1, . . . , k 2, . . . , k + 1 2, . . . , k + 1 1, . . . , k
   
2, . . . , k 1, . . . , k + 1
=A A .
2, . . . , k 1, . . . , k + 1
40 Criteria for total positivity and strict total positivity

All determinants containing initial consecutive rows or columns are strictly

positive by assumption, and from the induction hypothesis
 
2, . . . , k
A > 0.
2, . . . , k
Thus
 
2, . . . , k + 1
A > 0.
2, . . . , k + 1
Let us now consider the case i = 2. For k = 1 we must prove that a32 > 0.
Now
 
2, 3
A = a21 a32 − a31 a22 > 0.
1, 2
By assumption a21 , a31 > 0. From the previous analysis a22 > 0. Thus
a32 > 0. Assume the result holds for k − 1. From Sylvester’s Determinant
Identity,
       
2, . . . , k + 1 3, . . . , k + 2 2, . . . , k + 1 3, . . . , k + 2
A A −A A
1, . . . , k 2, . . . , k + 1 2, . . . , k + 1 1, . . . , k
   
3, . . . , k + 1 2, . . . , k + 2
= A A .
2, . . . , k 1, . . . , k + 1
All determinants containing initial consecutive rows or columns are strictly
positive, and by the above case i = 1
 
2, . . . , k + 1
A > 0.
2, . . . , k + 1
From the induction hypothesis
 
3, . . . , k + 1
A > 0.
2, . . . , k
Thus
 
3, . . . , k + 2
A > 0.
2, . . . , k + 1
We continue this process for all possible i. We do the same for i = 1 and
all j and k. We then go through the argument again for j = 2, then i = 2,
etc. . . , or alternatively, since the above result now holds for i = 1 and
j = 1, we can apply an induction argument to the (n − 1) × (m − 1) matrix
obtained from A by deleting its first row and column.

In Section 2.4 of this chapter we give a different and more transparent

proof of this result.
 2.2 Density and some further applications 41

When dealing with strictly totally positive matrices we can reverse the
order of the rows and columns (see Proposition 1.3). As a consequence
thereof, we have the following.

Corollary 2.4 Let A be an n×m matrix. Assume that all kth order minors
of A composed from the last k rows and k consecutive columns, and also all
kth order minors of A composed from the last k columns and k consecutive
rows, are strictly positive for k = 1, . . . , min{n, m}. Then A is strictly
totally positive.

Similar corollaries hold for many of the results of this chapter. They
should be understood. We will not bother to state them again.
Is this the minimal number of determinants that must be checked when
determining whether A is a strictly totally positive matrix? It seems that
it must be minimal as the number of conditions is nm, which is the number
of entries in the matrix A.
The following is presented here as it fits in with our present discussion
and can be seen to be a simple consequence of Theorem 2.3. It is also an
immediate corollary of Theorem 1.19. We state it without proof.

Proposition 2.5 Let A be an n × n totally positive matrix. Then A is

strictly totally positive if
 
1, . . . , k
A > 0
n − k + 1, . . . , n
 
n − k + 1, . . . , n
A > 0
1, . . . , k
for k = 1, . . . , n.

2.2 Density and some further applications

Strictly totally positive matrices are easier to handle than totally positive
matrices and fortuitously the class of strictly totally positive matrices is
dense in the class of all totally positive matrices. That is, if A is an n × m
totally positive matrix, then there are sequences (Ak )∞k=1 of n × m strictly
totally positive matrices such that elementwise

lim Ak = A.
k→∞

This result will prove very useful.

42 Criteria for total positivity and strict total positivity

Theorem 2.6 Strictly totally positive matrices are dense in the class of
totally positive matrices.
2
Proof For each q ∈ (0, 1) the matrix Qn (q) = (q (i−j) )ni=1 nj=1 is a strictly
2
totally positive matrix. This may be proven as follows. As q (i−j) =
2 2
q i q j q −2ij it suffices to prove that pij is strictly totally positive for
p = q −2 > 1. Let P = (pij )ni=1 m
j=1 . Consider
 
i + 1, . . . , i + k
P .
j + 1, . . . , j + k
Set xr = pi+r , r = 1, . . . , k. It then easily follows, factoring out from the
rth row of this minor p(i+r)(j+1) , that
   k
i + 1, . . . , i + k
P = p(i+r)(j+1) V (x1 , . . . , xk ) ,
j + 1, . . . , j + k r=1

where V (x1 , . . . , xk ) is the Vandermonde of the points x1 , . . . , xk and thus

equals
 
(xt − xs ) = (pi+t − pi+s ).
1≤s<t≤k 1≤s<t≤k

The terms in the product on the right are strictly positive since p > 1. This
implies that
 
i + 1, . . . , i + k
P >0
j + 1, . . . , j + k
and therefore, from Theorem 2.2, the matrix Qn (q) is strictly totally
positive. The matrix Qn (q) has the further property that
lim Qn (q) = In
q→0+

(elementwise) where In is the n × n identity matrix.

Now, let A be an arbitrary n × m totally positive matrix of rank r. Form
Bq = Qn (q)AQm (q).
Then each Bq is of rank r, and as is easily verified using the Cauchy–Binet
formula, all the k × k minors of Bq are strictly positive for every k ≤ r.
Furthermore
lim Bq = A.
q→0+

If r = min{n, m} we are finished. Assume not and let Bq be identical

with the matrix Bq except that we add an arbitrarily small ε > 0 to the
 2.2 Density and some further applications 43

(1, 1)-element of Bq . The matrix Bq is totally positive and of rank r + 1.
As such, by the same method as employed above, we can approximate Bq
arbitrarily well by a matrix of rank r + 1, all of whose k × k minors are
strictly positive for every k ≤ r + 1. We continue this process to obtain the
result.

Determinantal criteria for when a matrix is totally positive will be

discussed in Section 2.4. But we present here a particular case of Theorem
2.13 since it can be easily proven by an application of Theorem 2.2 and
the method of proof of Theorem 2.6. This is not the best possible result,
as will be shown in Proposition 2.15.

Proposition 2.7 Let A be an n × n nonsingular matrix. Assume that all

kth order minors of A composed from arbitrary k rows and k consecutive
columns, or all kth order minors of A composed from arbitrary k columns
and k consecutive rows are nonnegative for k = 1, . . . , n. Then A is totally
positive.

Proof Assume all kth order minors of A composed from arbitrary k rows
and k consecutive columns are nonnegative. Let Q(q) be the n × n matrix
as used in the proof of Theorem 2.6, q ∈ (0, 1). Set

Aq = Q(q)A.

From the Cauchy–Binet formula

 
i + 1, . . . , i + k
Aq
j + 1, . . . , j + k
    
i + 1, . . . , i + k 1 , . . . , k
= Q(q) A .
1 , . . . , k j + 1, . . . , j + k
1≤
1 <···<
k ≤n

By assumption we always have

 
1 , . . . , k
A ≥ 0,
j + 1, . . . , j + k
and as A is nonsingular there exist 1 ≤ 1 < · · · < k ≤ n for which
 
1 , . . . , k
A > 0.
j + 1, . . . , j + k
Furthermore Q(q) is strictly totally positive. Thus
 
i + 1, . . . , i + k
Aq > 0,
j + 1, . . . , j + k
44 Criteria for total positivity and strict total positivity

for all i, j ∈ {0, . . . , n−k}. From Theorem 2.2 Aq is strictly totally positive.
In addition,

lim Aq = A,
q→0+

and therefore A is totally positive.

Theorems 2.2 and 2.3 are very useful tools for proving strict total
positivity. We now settle a previous debt by presenting proofs of
Propositions 1.9 and 1.10 from Section 1.2. We recall the following.

Proposition 1.9 Assume A is an n × m strictly totally positive matrix.

For given 1 ≤ r < n set

aij,  i = 1, . . . , r, j = 2, . . . , m
cij = i−1,i
A 1,j , i = r + 1, . . . , n, j = 2, . . . , m.

Then C = (cij ) is an n × (m − 1) strictly totally positive matrix. If A is

totally positive, then C is totally positive.

Proof Let C = (cij ) be as above. We prove our desired result by obtaining

an explicit formula for the arbitrary minors of C composed of consecutive
rows, i.e.,
 
i + 1, . . . , i + k
C .
j1 , . . . , jk

The first statement in Proposition 1.9 will then follow from either Theorem
2.2 or 2.3.
Obviously if i + k ≤ r, then
   
i + 1, . . . , i + k i + 1, . . . , i + k
C =A .
j1 , . . . , jk j1 , . . . , jk

We prove that for i + 1 ≤ r < i + k

   
i + 1, . . . , i + k i + 1, . . . , i + k
C = ar,1 · · · ai+k−1,1 A
j1 , . . . , jk j1 , . . . , jk

and for r < i + 1

   
i + 1, . . . , i + k i, i + 1, . . . , i + k
C = ai+1,1 ai+2,1 · · · ai+k−1,1 A .
j1 , . . . , jk 1, j1 , . . . , jk
 2.2 Density and some further applications 45

The first equality can be explained as follows.

ai+1,j1 ··· ai+1,jk
.. ..
. .
  ···
i + 1, . . . , i + k ar,j1  ar,jk 
C = A r,r+1
··· A r,r+1 .
j1 , . . . , jk 1,j1 1,jk
.. ..
 .   . 
A i+k−1,i+k
1,j1 ··· A i+k−1,i+k
1,jk

By adding ar+1,1 times the (r − i)th row (which is the row with entries
(ar,j1 , . . . , ar,jk )) to the succeeding row, we get a new (r − i + 1)st row of

(ar,1 ar+1,j1 , . . . , ar,1 ar+1,jk ).

Factor out ar,1 and add ar+2,1 times this row to the next row to obtain a
new (r − i + 2)nd row of

(ar+1,1 ar+2,j1 , . . . , ar+1,1 ar+2,jk ),

etc. The formula now easily follows.

The second equality is similarly obtained by considering the determinant
ai1 aij1  ··· aijk 
0 A i,i+1
1,j1 ··· A i,i+1
1,jk
.. .. .. .
.  .   . 
0 A i+k−1,i+k
1,j1 ··· A i+k−1,i+k
1,jk

On the one hand this determinant equals

 
i + 1, . . . , i + k
ai1 C .
j1 , . . . , jk
On the other hand, adding ai+1,1 times the first row to the second row we
get a new second row of

(ai1 ai+1,1 , ai1 ai+1,j1 , . . . , ai1 ai+1,jk ).

Factor out ai1 and add ai+2,1 times the second row to the third row etc.
The formula now easily follows.
From these three formulæ it follows from Theorems 2.2 or 2.3 that if A
is strictly totally positive then C is strictly totally positive. Now assume
that A is only totally positive. From Theorem 2.6 there exists a sequence
(Ak ) of n × m strictly totally positive matrices whose elements converge
46 Criteria for total positivity and strict total positivity

to the corresponding elements of A. Construct the associated Ck (in the

same way in which C is constructed from A). Each Ck is strictly totally
positive and the elements of Ck converge to the corresponding elements of
C. Thus C is totally positive.

Proposition 1.10 Assume A is an n × m strictly totally positive matrix.

Fix p < min{m, n} and set
 
i − p, . . . , i − 1, i
cij = A
1, . . . , p, j

for i = p + 1, . . . , n, j = p + 1, . . . , m. Then C = (cij ) is an (n − p) × (m − p)

strictly totally positive matrix.

Proof We prove Proposition 1.10 as a consequence of Proposition 1.9. As A

is a strictly totally positive matrix, Proposition 1.10 holds for p = 1 since
the associated C is a submatrix of the matrix C in Proposition 1.9 with
r = 1. We apply an induction argument on p. Let us assume the validity
of Proposition 1.10 for p − 1. That is, setting
 
i − p + 1, . . . , i − 1, i
eij = A
1, . . . , p − 1, j

for i = p, . . . , n, j = p, . . . , m, we assume that E = (eij ) is an (n − p + 1) ×

(m − p + 1) strictly totally positive matrix. Now define
 
i − 1, i
dij = E
p, j

for i = p + 1, . . . , n, j = p + 1, . . . , m. Applying Proposition 1.9 with r = 1

it follows that D = (dij ) is an (n − p) × (m − p) strictly totally positive
matrix. In addition, from Sylvester’s Determinant Identity,
   
  i−p,...,i−2,i−1 i−p,...,i−2,i−1
i − 1, i A
 1,...,p−1,p 
A
 1,...,p−1,j 
dij = E = i−p+1,...,i−1,i
p, j A 1,...,p−1,p A i−p+1,...,i−1,i
1,...,p−1,j
   
i − p + 1, . . . , i − 1 i − p, i − p + 1, . . . , i − 1, i
= A A .
1, . . . , p − 1 1, . . . , p − 1, p, j

Note that the values

 
i − p + 1, . . . , i − 1
A
1, . . . , p − 1
 2.3 Triangular total positivity 47

are strictly positive and independent of j. Thus the matrix D being strictly
totally positive is equivalent to the matrix C with elements
 
i − p, i − p + 1, . . . , i − 1, i
cij = A
1, . . . , p − 1, p, j
being strictly totally positive. This advances the induction step and proves
the proposition.

2.3 Triangular total positivity

A square matrix A = (aij ) is said to be upper triangular if aij = 0 for all
i > j. It is said to be lower triangular if aij = 0 for all i < j. We say that
the n × n matrix A = (aij ) is upper strictly totally positive if it is upper
triangular and
 
i1 , . . . , ik
A >0
j1 , . . . , jk
whenever im ≤ jm , m = 1, . . . , k. (As usual we always assume that 1 ≤
i1 < · · · < ik ≤ n and 1 ≤ j1 < · · · < jk ≤ n.) Note that for an upper
triangular matrix we always have
 
i1 , . . . , ik
A =0
j1 , . . . , jk
if im > jm for some m ∈ {1, . . . , k}. We say that the n ×n matrix A = (aij )
is upper totally positive if it is upper triangular and
 
i1 , . . . , ik
A ≥0
j1 , . . . , jk
whenever im ≤ jm , m = 1, . . . , k. (This latter restriction has no meaning
here.)
Similarly we say that the n × n matrix A = (aij ) is lower strictly totally
positive if it is lower triangular and
 
i1 , . . . , ik
A >0
j1 , . . . , jk
whenever im ≥ jm , m = 1, . . . , k. We say that the n × n matrix A = (aij )
is lower totally positive if it is lower triangular and
 
i1 , . . . , ik
A ≥0
j1 , . . . , jk
whenever im ≥ jm , m = 1, . . . , k.
There are relatively simple criteria for determining when a triangular
48 Criteria for total positivity and strict total positivity

matrix is strictly totally positive in the above sense. These criteria are very
reminiscent of the criteria in Theorem 2.3, and for good reason.

Theorem 2.8 Let A = (aij ) be an n × n upper triangular matrix satisfying

 
1, . . . , k
A >0 (2.2)
j + 1, . . . , j + k

for j = 0, 1, . . . , n−k, k = 1, . . . , n. Then A is upper strictly totally positive.

Similarly, if A = (aij ) is an n × n lower triangular matrix satisfying
 
i + 1, . . . , i + k
A >0
1, . . . , k

for i = 0, 1, . . . , n − k, k = 1, . . . , n, then A is lower strictly totally positive.

Proof Obviously the two claims are equivalent. We prove the first claim.
That is, we show that if A is an n × n upper triangular matrix satisfying
(2.2), then
 
i1 , . . . , ik
A >0
j1 , . . . , jk

whenever im ≤ jm , m = 1, . . . , k.
From (2.2) and Fekete’s Lemma 2.1, it follows that
 
1, . . . , k
A >0 (2.3)
j1 , . . . , jk
for all 1 ≤ j1 < · · · < jk ≤ n, and k = 1, . . . , n. Since A is upper triangular
and satisfies (2.3) we also have that for any 1 ≤ j1 < · · · < jk ≤ n, where
i + m ≤ jm , m = 1, . . . , k,
     
1, . . . , i i + 1, . . . , i + k 1, . . . , i, i + 1, . . . , i + k
A A =A > 0.
1, . . . , i j1 , . . . , jk 1, . . . , i, j1 , . . . , jk

This implies that

 
i + 1, . . . , i + k
A >0 (2.4)
j1 , . . . , jk

for any 1 ≤ j1 < · · · < jk ≤ n, where i + m ≤ jm , m = 1, . . . , k. It thus

remains to prove that we can replace consecutive rows by arbitrary rows
(within the confines of the upper triangularity).
To this end we rework the method of proof of Fekete’s Lemma 2.1. For
 2.3 Triangular total positivity 49

a given i = (i1 , . . . , ik ) ∈ Ikn we recall that d(i), the dispersion of i, is given

by

k
d(i) := (ij − ij−1 − 1) = (ik − i1 ) − (k − 1) ≥ 0.
j=2

Then d(i) = 0 if and only if the sequence i is composed of consecutive

integers. We prove the result by induction on both k and on the value of
d(·).
We assume that  
i1 , . . . , is
A >0
j1 , . . . , js
whenever im ≤ jm , m = 1, . . . , s, and s ≤ k and d(i) < p. From (2.4) the
result is valid for every k when d(i) = 0.
Let i = (i1 , . . . , ik ) with d(i) = p > 0, and im ≤ jm , m = 1, . . . , k. We
claim that  
i1 , . . . , ik
A > 0.
j1 , . . . , jk
As the sequence i is not composed of consecutive integers, we can add to
this sequence an integer between i1 and ik . Let 1 ≤ r1 < · · · < rk+1 ≤ n be
the resulting sequence, assuming r
to be the new index (1 <  < k + 1).
Thus
rm ≤ jm , m = 1, . . . ,  − 1
rm < rm+1 ≤ jm , m = , . . . , k. (2.5)
By the determinantal identity (1.2) of Section 1.1, we have
   
r1 , . . . , r
, . . . , rk+1 r2 , . . . , rk
A A
j1 , . . . , jk j2 , . . . , jk
   
r2 , . . . , rk+1 r1 , . . . , r
, . . . , rk
= A A
j1 , . . . , jk j2 , . . . , jk
   
r1 , . . . , rk r2 , . . . , r
, . . . , rk+1
+A A .
j1 , . . . , jk j2 , . . . , jk
Now  
r2 , . . . , rk
A >0
j2 , . . . , jk
by the induction assumption on k and since rm ≤ jm , m = 2, . . . , k (see
(2.5)), while
 
r1 , . . . , rk
A >0
j1 , . . . , jk
50 Criteria for total positivity and strict total positivity

since rm ≤ jm , m = 1, . . . , k (see (2.5)) and by the induction assumption on

d(i). That is, since r1 = i1 and rk < ik it follows that d((r1 , . . . , rk )) < d(i).
Furthermore
 
r2 , . . . , r
, . . . , rk+1
A >0
j2 , . . . , jk
by the induction hypothesis on k and since rm ≤ jm , m = 2, . . . ,  − 1, and
rm+1 ≤ jm , m = , . . . , k (see (2.5)). Similarly we have that
 
r1 , . . . , r
, . . . , rk
A > 0.
j2 , . . . , jk
Now
 
r2 , . . . , rk+1
A ≥ 0,
j1 , . . . , jk
because this determinant vanishes if we have rm+1 > jm for some
m ∈ {1, . . . , −1} (which is possible), while if not then it is strictly positive
since d((r2 , . . . , rk+1 )) < d(i).
It therefore follows that
   
i1 , . . . , ik r1 , . . . , r
, . . . , rk+1
A =A >0
j1 , . . . , jk j1 , . . . , jk
which advances the induction hypothesis and proves the theorem.

In parallel with Proposition 2.5 we have the following result, which we

state for upper triangular matrices. It is also an immediate consequence of
Theorem 1.19.

Proposition 2.9 Let A be an n × n upper totally positive matrix. Then A

is upper strictly totally positive if
 
1, . . . , k
A >0
n − k + 1, . . . , n
for k = 1, . . . , n.

2.4 LDU factorizations

Assume A = (aij ) is an n × n strictly totally positive matrix. Since
 
1, . . . , k
A >0
1, . . . , k
 2.4 LDU factorizations 51

for k = 1, . . . , n, then, as is well known, A has a unique LU factorization.

In fact A can be decomposed into the form

A = LDU,

where L = (ij ) is a unit diagonal lower triangular matrix, D = (dij ) is a

diagonal matrix, and U = (uij ) is a unit diagonal upper triangular matrix.
The nonzero elements of L, D and U are explicitly given by
 
1, . . . , j − 1, i
A 1, . . . , j − 1, j
ij =  , i ≥ j,
1, . . . , j − 1, j
A 1, . . . , j − 1, j
 
1, . . . , i − 1, i
A 1, . . . , i − 1, j
uij =  , i ≤ j,
1, . . . , i − 1, i
A 1, . . . , i − 1, i
 
1, . . . , i − 1, i
A 1, . . . , i − 1, i
dii =   .
1, . . . , i − 1
A 1, . . . , i − 1

Using Theorem 2.8 we prove that L is a lower strictly totally positive matrix
and U is an upper strictly totally positive matrix.

Theorem 2.10 Assume A is a square strictly totally positive matrix. Then

A has a unique factorization of the form

A = LDU,

where L is a unit diagonal lower strictly totally positive matrix, U is a unit

diagonal upper strictly totally positive matrix, and D is a diagonal matrix
whose diagonal entries are strictly positive.

Proof We need only prove that L is a lower strictly totally positive matrix
and U is an upper strictly totally positive matrix. The other facts follow
trivially. As A = LDU and D is a diagonal matrix, then from the
Cauchy–Binet formula
 
1, . . . , k
A
j + 1, . . . , j + k
      
1, . . . , k 1 , . . . , k 1 , . . . , k
= L D U .
1 , . . . , k 1 , . . . , k j + 1, . . . , j + k
1≤
1 <···<
k ≤n
52 Criteria for total positivity and strict total positivity

Since L is lower triangular

 
1, . . . , k
L =0
1 , . . . , k

except when m = m, m = 1, . . . , k. Thus

       
1, . . . , k 1, . . . , k 1, . . . , k 1, . . . , k
A =L D U .
j + 1, . . . , j + k 1, . . . , k 1, . . . , k j + 1, . . . , j + k

Now
     
1, . . . , k 1, . . . , k 1, . . . , k
L = 1, D > 0, and A > 0.
1, . . . , k 1, . . . , k j + 1, . . . , j + k

Therefore
 
1, . . . , k
U > 0.
j + 1, . . . , j + k

As this is valid for all appropriate j and k we have from Theorem 2.8 that
U is an upper strictly totally positive matrix. The analogous proof shows
that L is a lower strictly totally positive matrix.

Each square strictly totally positive matrix also has a unique

factorization of the form

A = U DL,

where U is a unit diagonal upper strictly totally positive matrix, L is a unit

diagonal lower strictly totally positive matrix, and D is a diagonal matrix
whose diagonal entries are strictly positive. The nonzero elements of U , D,
and L are explicitly given by
 
i, j + 1, . . . , n
A j, j + 1, . . . , n
uij =  , i ≤ j,
j, j + 1, . . . , n
A j, j + 1, . . . , n
 
i, i + 1, . . . , n
A j, i + 1, . . . , n
ij =  , i ≥ j,
i, i + 1, . . . , n
A i, i + 1, . . . , n
 
i, i + 1, . . . , n
A i, i + 1, . . . , n
dii =   .
i + 1, . . . , n
A i + 1, . . . , n
 2.4 LDU factorizations 53

This can be shown in a variety of ways. It also follows easily from the
previous result and the following. Let
 
0 ··· 0 1
 0 ··· 1 0 
 
Q= . .. .. .
 .. . . 
1 ··· 0 0
Then
QAQ

is the matrix obtained from A by reversing the order of its rows and
columns. As is known (see Proposition 1.3 in Chapter 1), A is strictly
totally positive if and only if QAQ is strictly totally positive. From the
LDU factorization of QAQ (Theorem 2.10) we have

A = (QLQ)(QDQ)(QU Q).

QLQ is a unit diagonal upper strictly totally positive matrix while QU Q

is a unit diagonal lower strictly totally positive matrix.
Another consequence of Theorem 2.10 is an alternative, more elementary,
proof of Theorem 2.3.

Another Proof of Theorem 2.3. Since

 
1, . . . , k
A >0
1, . . . , k
for k = 1, . . . , n, it follows that A has a unique LDU factorization. Since
   
1, . . . , k i + 1, . . . , i + k
A ,A >0
j + 1, . . . , j + k 1, . . . , k
for all appropriate j, i, and k, it follows from the proof of Theorem 2.10
that L is a unit diagonal lower strictly totally positive matrix, U is a unit
diagonal upper strictly totally positive matrix, while D is a diagonal matrix
whose diagonal entries are strictly positive. From this it easily follows that
A = LDU is strictly totally positive.

What about totally positive matrices? Do they always have a similar

unique LDU factorization? If A is also non-singular, then the answer is yes.
The proof is essentially the same as the above proof, except that L is a unit
diagonal lower totally positive matrix and U is a unit diagonal upper totally
positive matrix. Recall that since A is totally positive and nonsingular then
54 Criteria for total positivity and strict total positivity

all principal minors of A are strictly positive (Theorem 1.13). We state this
result here for easy reference.

Proposition 2.11 Assume A is a square nonsingular totally positive

matrix. Then A has a unique factorization of the form

A = LDU

where L is a unit diagonal lower totally positive matrix, U is a unit diagonal

upper totally positive matrix, and D is a diagonal matrix whose diagonal
entries are strictly positive.

However, if A is singular then we must modify our demands somewhat.

In general we lose the unit triangular nature of the L and U and the
uniqueness. What we do have is the following.

Theorem 2.12 Assume A is a square totally positive matrix. Then A has

a factorization of the form
A = LU

where L is a lower totally positive matrix and U is an upper totally positive

matrix.

Proof From Theorem 2.6 there exists a sequence of strictly totally positive
matrices (Am ) that approximate A, i.e., such that

lim Am = A.
m→∞

From Theorem 2.10, each Am = (aij (m)) has a unique factorization of the
form
Am = Lm Dm Um

where Lm is a unit diagonal lower strictly totally positive matrix, Um is a

unit diagonal upper strictly totally positive matrix, and Dm is a diagonal
matrix whose diagonal entries are strictly positive.
There are many different normalizations to this factorization that
allow us to take limits. Here is one such normalization. By pre- and
postmultiplying by positive diagonal matrices we can factor Am in the
form
Am = D ! mL
!m U!m

! m and U
where L !m are stochastic (rows sums 1). As a consequence we also
 2.5 Criteria for total positivity 55

have

aij (m) = d!ii (m)
j

for all i, where d!ii (m) is the (i, i) entry of D

! m . Thus all entries in each
of the factors are uniformly bounded. Now that we have bounds on the
elements of D! m, L! m and U!m , we may take a subsequence along which each
of the entries of these matrices converges. Let their respective limits be D,
L and U , and set L = DL. Then

A = LU.

Furthermore it follows that L is a lower totally positive matrix and U is

an upper totally positive matrix.

2.5 Criteria for total positivity

Unfortunately, nothing quite as simple as the conditions in Theorems 2.2
and 2.3 seem to be valid for determining if a matrix is totally positive.
It is possible, for example, that every minor composed from consecutive
rows and consecutive columns is nonnegative and yet A is not totally
positive. A simple example is the 4 × 4 matrix
 
1 0 0 0
 0 1 0 0 
A 0 0 1 0 .


1 0 0 1

One therefore needs somewhat more restrictive criteria.

Let us again recall that for a given strictly increasing sequence i =
(i1 , . . . , ik ) ∈ Ikn we define d(i), the dispersion of i, by


k
d(i) := (ij − ij−1 − 1) = (ik − i1 ) − (k − 1) ≥ 0.
j=2

Note that d(i) counts the number of integers between i1 and ik that are
not in the sequence i1 , . . . , ik , and thus d(i) = 0 if and only if the sequence
i is composed of consecutive integers.
The main theorem providing determinantal criteria for when a matrix
is totally positive is the following. It is based on previous reasonings and
techniques, but is technically more detailed.
56 Criteria for total positivity and strict total positivity

Theorem 2.13 An n × m matrix A = (aij ) of rank r is totally positive if

 
i1 , . . . , ik
A ≥0 (2.6)
j1 , . . . , jk

for all i = (i1 , . . . , ik ) satisfying d(i) ≤ n − r, all j = (j1 , . . . , jk ) and

k = 1, . . . , r.

Note that if A is nonsingular, i.e., r = n = m, then this is Proposition

2.7.

Proof We wish to prove that from (2.6) it follows that

 
i1 , . . . , is
A ≥0 (2.7)
j1 , . . . , js

for all 1 ≤ i1 < · · · < is ≤ n and 1 ≤ j1 < · · · < js ≤ m and all s.

The proof will be based on an induction argument. The induction is on
s. For s = 1 there is nothing to prove since d({i1 }) = 0 ≤ n − r and thus
(2.7) follows from (2.6). Let us suppose that (2.7) holds for all s ≤ k − 1,
but is not valid for some (i1 , . . . , ik ) and (j1 , . . . , jk ), as above. That is, we
have
 
i1 , . . . , ik
A < 0. (2.8)
j1 , . . . , jk

By our assumption, we must have d(i) > n − r. Choose i∗ = (i∗1 , . . . , i∗k )

and j∗ = (j1∗ , . . . , jk∗ ) such that (2.8) holds and d(i∗ ) =  is minimal, i.e.,
for given k and every i ∈ Ikn , j ∈ Ikm satisfying (2.8) we have d(i) ≥ .
If ai denotes the ith row of A, then we prove that for all p satisfying
∗ ∗
i∗1 < p < i∗k we have that ap ∈ span{ai2 , . . . , aik−1 }. As there are exactly 
such p (since d(i∗ ) = ), it follows that the rank of A is at most n − . But,
by assumption,  > n − r, which implies that the rank of A is strictly less
than r. A contradiction.
It therefore remains to prove that for all p satisfying i∗1 < p < i∗k we have
∗ ∗
ap ∈ span{ai2 , . . . , aik−1 }. This we prove in two steps. We first prove that
it is true when we restrict ourselves to the columns j1∗ , . . . , jk∗ . We then
subsequently extend it to all columns.
Let p satisfy i∗1 < p < i∗k , and let r1 < · · · < rk+1 be the sequence
obtained by adding p to the sequence i∗1 , . . . , i∗k . Assume p = rt . From the
 2.5 Criteria for total positivity 57

determinantal equality (1.2) of Section 1.1, we have for each s ∈ {1, . . . , k}:
   
r1 , . . . , rt , . . . , rk+1 r2 , . . . , rk
A A
j1∗ , . . . , jk∗ j1∗ , . . . , js∗ , . . . , jk∗
   
r2 , . . . , rk+1 r1 , . . . , rt , . . . , rk
=A A (2.9)
j1∗ , . . . , jk∗ j1∗ , . . . , js∗ , . . . , jk∗
   
r1 , . . . , rk r2 , . . . , rt , . . . , rk+1
+A ∗ A .
j1 , . . . , jk∗ j1∗ , . . . , js∗ , . . . , jk∗
By the induction assumption, all three k − 1 × k − 1 minors in (2.9) are
nonnegative, while the two k × k minors on the right-hand-side of the
equation are nonnegative since the dispersion of the rows therein is strictly
less than , and the minimality property of . Thus the value of the right-
hand-side of (2.9) is nonnegative. As (2.8) holds for i∗1 , . . . , i∗k and j1∗ , . . . , jk∗
(which is the left-most determinant in (2.9)), it necessarily follows that we
must have
  
i∗2 , . . . , p, . . . , i∗k−1 r2 , . . . , rk
A =A ∗ = 0. (2.10)
j1∗ , . . . , js∗ , . . . , jk∗ j1 , . . . , js∗ , . . . , jk∗
∗ ∗
As (2.10) is valid for all s = 1, . . . , k, we have ap ∈ span{ai2 , . . . , aik−1 } on
the columns j1∗ , . . . , jk∗ , for each p satisfying i∗1 < p < i∗k .
Since  ∗ 
i1 , . . . , i∗k
A ∗ < 0,
j1 , . . . , jk∗
∗ ∗
the rows ai1 , . . . , aik are linearly independent on the columns j1∗ , . . . , jk∗ .
∗ ∗
Thus the rows ai2 , . . . , aik−1 are linearly independent on some columns
!
j1 , . . . , !
jk−2 , where {!
j1 , . . . , !
jk−2 } = {j1∗ , . . . , jk∗ }\{js∗ , jt∗ } (we assume that
∗ ∗
js < jt in what follows). That is,
 ∗ 
i2 , . . . , i∗k−1
A = 0.
j1 , . . . , !
! jk−2
By our induction assumption, this implies that the above minor is, in fact,
strictly positive. We use

∗ 
i2 , . . . , i∗k−1
A
j1 , . . . , !
! jk−2
as a pivot block in Sylvester’s Determinant Identity. That is, set
 ∗ 
i2 , . . . , i∗k−1 , i
dij = A .
j1 , . . . , !
! jk−2 , j
Then by our induction hypothesis on k we have dij ≥ 0 for all i and j,
58 Criteria for total positivity and strict total positivity

and from (2.10) we have dpjs∗ = dpjt∗ = 0 for each p between i∗1 and i∗k .
Furthermore, by Sylvester’s Determinant Identity we have
 ∗ ∗  ∗   ∗ 
i1 , ik i2 , . . . , i∗k−1 i1 , . . . , i∗k
D =A A ∗ < 0.
js∗ , jt∗ j1 , . . . , !
! jk−2 j1 , . . . , jk∗
Thus we must have di∗1 jt∗ > 0 and di∗k js∗ > 0.
From Sylvester’s Determinant Identity, for j < jt∗ ,
 ∗   ∗   ∗ 
i1 , p i2 , . . . , i∗k−1 i1 , . . . , i∗k−1 , p
D =A A .
j, jt∗ j1 , . . . , !
! jk−2 j1 , . . . , !
! jk−2 , j, jt∗
Both terms of the right-hand side of this equality are nonnegative. The
first of these two terms is strictly positive. The second term is nonnegative
because the dispersion of i∗1 , . . . , i∗k−1 , p is strictly less than . Thus
 ∗ 
i ,p
0 ≤ D 1 ∗ = di∗1 j dpjt∗ − di∗1 jt∗ dpj .
j, jt
Now dpjt∗ = 0, di∗1 jt∗ > 0, and dpj ≥ 0. Thus dpj = 0 for all j < jt∗ . When
j > js∗ we consider
 ∗
p, ik
D
js∗ , j
and apply the same reasoning to obtain dpj = 0 for all j > js∗ . Thus for all
j we have
 ∗ 
i2 , . . . , i∗k−1 , p
0 = dpj = A .
j1 , . . . , !
! jk−2 , j
Since
 
i∗2 , . . . , i∗k−1
A > 0,
j1 , . . . , !
! jk−2
∗ ∗
this implies that row ap is in span{ai2 , . . . , aik−1 } over all columns of A,
and this is true for all p satisfying i∗1 < p < i∗k .

One fairly elementary consequence of the above is the following result.

Theorem 2.14 Let A = (aij ) be an n × n nonsingular upper triangular

matrix satisfying
 
1, . . . , k
A ≥0
j1 , . . . , jk
for all 1 ≤ j1 < · · · < jk ≤ n, k = 1, . . . , n. Then A is upper totally
 2.5 Criteria for total positivity 59

positive. Similarly, if A = (aij ) is an n × n nonsingular lower triangular

matrix satisfying
 
i1 , . . . , ik
A ≥0
1, . . . , k
for 1 ≤ i1 < · · · < ik ≤ n, k = 1, . . . , n, then A is lower totally positive.

Proof We prove the first claim. As A is a nonsingular upper triangular

matrix we necessarily have amm > 0 for m = 1, . . . , n. From Theorem 2.13
it suffices to prove that
 
i + 1, . . . , i + k
A ≥0
j1 , . . . , jk

for all appropriate i, 1 ≤ j1 < · · · < jk ≤ n and k (i.e., we must check the
case d(i) = 0). It suffices to consider only those minors with i + m ≤ jm ,
m = 1, . . . , k, since all other minors vanish due to the upper triangular
structure of A. Now, by assumption,
  
i  
1, . . . , i, i + 1, . . . , i + k i + 1, . . . , i + k
0≤A = amm A .
1, . . . , i, j1 , . . . , jk m=1
j1 , . . . , jk

As amm > 0 for all m, this proves our result.

Based on the above Theorem 2.14 and the analysis of Section 2.4 we
have the following result, which should be compared to Theorem 2.13 and
Proposition 2.7.

Proposition 2.15 Let A = (aij ) be an n × n nonsingular matrix. Then A

is totally positive if and only if A satisfies the following:
 
1, . . . , k
(i) A 1, . . . , k > 0, k = 1, . . . , n
 
i1 , . . . , ik
(ii) A 1, . . . , k ≥ 0, 1 ≤ i1 < · · · < ik ≤ n, k = 1, . . . , n
 
1, . . . , k
(iii) A j , . . . , j ≥ 0, 1 ≤ j1 < · · · < jk ≤ n, k = 1, . . . , n.
1 k

Proof If A is totally positive then (ii) and (iii) hold. If A is also nonsingular
then (i) holds from Theorem 1.13.
Assume (i) – (iii) hold. The LDU factorization of A, as in Section 2.4,
is well defined since (i) holds. From (i), (ii), and (iii) we see that D is
a diagonal matrix whose diagonal entries are strictly positive, L is a unit
60 Criteria for total positivity and strict total positivity

diagonal lower triangular matrix, and U is a unit diagonal upper triangular

matrix. Now, from the proof of Theorem 2.10 we have,
        
i1 , . . . , ik i1 , . . . , ik 1 , . . . , k 1 , . . . , k
A = L D U
1, . . . , k 1 , . . . , k 1 , . . . , k 1, . . . , k
1≤
1 <···<
k ≤n
   
i1 , . . . , ik 1, . . . , k
=L D
1, . . . , k 1, . . . , k
and similarly
     
1, . . . , k 1, . . . , k 1, . . . , k
A =D U .
j1 , . . . , jk 1, . . . , k j1 , . . . , jk
Since  
1, . . . , k
D >0
1, . . . , k
we have as a consequence of (ii) and (iii)
   
i1 , . . . , ik 1, . . . , k
L , U ≥ 0.
1, . . . , k j1 , . . . , jk
This is valid for all 1 ≤ i1 < · · · < ik ≤ n, 1 ≤ j1 < · · · < jk ≤ n, and
k = 1, . . . , n. Thus from Theorem 2.14 L is a lower totally positive matrix
and U is an upper totally positive matrix, implying in turn that
A = LDU
is totally positive.

2.6 “Simple” criteria for strict total positivity

There is a determinantal criterion, very much different from those of the
previous sections, that suffices for total positivity. It is only sufficient and
not, of course, necessary. It is based solely on the sign of the elements of
A and a form of measure of the magnitude of the 2 × 2 minors of A with
consecutive rows and columns. It is both a surprising and a beautiful result.
But its proof is far from trivial. It is the following.

Theorem 2.16 Let A = (aij ) be an n × n matrix whose entries are strictly

positive. Assume
 
π
aij ai+1,j+1 > 4 cos2 ai,j+1 ai+1,j
n+1
for all i, j = 1, . . . , n − 1. Then A is strictly totally positive. Furthermore
the above constant 4 cos2 n+1 π
is best possible.
 2.6 “Simple” criteria for strict total positivity 61

By a simple perturbation argument and the denseness of strictly totally

positive matrices in the class of all totally positive matrices it follows that
if the inequality in Theorem 2.16 holds, but is not always strict, then A is
totally positive.
An immediate consequence of this theorem, independent of n and m, is
the more easily stated.

Corollary 2.17 Let A = (aij ) be an n×m matrix whose entries are strictly
positive. Assume
aij ai+1,j+1 ≥ 4ai,j+1 ai+1,j

for all i = 1, . . . , n − 1, j = 1, . . . , m − 1. Then A is strictly totally positive.

We start by proving that the constant 4 cos2 π

n+1 in Theorem 2.16 is best
possible.

Proposition 2.18 For any c < 4 cos2 n+1 π

there exists an n × n matrix
A = (aij ) with all entries strictly positive for which

aij ai+1,j+1 > c ai,j+1 ai+1,j , i, j = 1, . . . , n − 1,

and yet A is not totally positive.

Proof Consider the n × n symmetric Toeplitz matrix

 
2 cos θ 1 0 ··· 0
 1 2 cos θ 1 ··· 0 
 
 0 1 2 cos θ · ·· 0 
An (θ) =  ,
 . . . . .. 
 .. .. .. .. . 
0 0 0 ··· 2 cos θ

where 0 ≤ θ < π/2. Expanding by the last row we see that det An (θ)
satisfies the recurrence relation

det An (θ) = 2 cos θ det An−1 (θ) − det An−2 (θ).

Now
sin 2θ
det A1 (θ) = 2 cos θ = ,
sin θ
and
sin 3θ
det A2 (θ) = 4 cos2 θ − 1 = .
sin θ
62 Criteria for total positivity and strict total positivity

From the fact that

sin(n + 1)θ + sin(n − 1)θ = 2 cos θ sin nθ

it immediately follows that

sin(n + 1)θ
det An (θ) = ,
sin θ
implying that for θ ∈ (π/(n + 1), 2π/(n + 1)) we have

det An (θ) < 0.

For An (θ) = (aij ) we have

aij ai+1,j+1 ≥ 4 cos2 θ ai,j+1 ai+1,j

for i, j = 1, . . . , n − 1. Now for any c < 4 cos2 π

n+1 , let

θ ∈ (π/(n + 1), 2π/(n + 1))

be such that c < 4 cos2 θ. Thus

aij ai+1,j+1 ≥ c ai,j+1 ai+1,j

for all i, j and yet

det An (θ) < 0,

i.e., An (θ) is far from being totally positive.

The entries of An (θ) are not all strictly positive. But it is not a problem
to perturb An (θ) so that its entries are all strictly positive, it satisfies

aij ai+1,j+1 > c ai,j+1 ai+1,j

for all i, j = 1, . . . , n − 1, and the perturbed An (θ) is not totally positive.

In fact we can also do so while maintaining the Toeplitz character of the
matrix. For example, keep the diagonal and first off-diagonal entries of
2
An (θ) fixed and perturb all elements of the kth off-diagonal to ε(k−1) for
k = 2, . . . , n − 1. For ε > 0, sufficiently small, this new matrix satisfies all
our requirements.

The proof of the main claim of Theorem 2.16 is lengthy and complicated.
We therefore divide it into a series of lemmas. We start with the following.

Lemma 2.19 Let A = (aij ) be an m × m matrix with strictly positive

entries. Assume
aij ai+1,j+1 > c ai,j+1 ai+1,j
 2.6 “Simple” criteria for strict total positivity 63

for all i, j = 1, . . . , m − 1. Then for k,  ≥ 1

aij ai+k,j+
> ck
ai,j+
ai+k,j ,
i = 1, . . . , m − k, j = 1, . . . , m − .

Proof Multiply together the k inequalities

ai+r,j+s ai+r+1,j+s+1 > c ai+r,j+s+1 ai+r+1,j+s
for r = 0, 1, . . . , k − 1 and s = 0, 1, . . . ,  − 1.

Lemma 2.20 The following are equivalent:

"m/2 m−j 
(i) Fm (c) = j=0 j (−1)j c1j , m = 0, 1, 2, . . .

(ii) F0 (c) = F1 (c) = 1

Fm (c) = Fm−1 (c) − 1c Fm−2 (c), m = 2, 3, . . .

where x denotes the integral part of x.

It is a simple matter to verify the above lemma. Its proof is left to the
reader.

Lemma 2.21 Let Fm (c) be as above. Then

(i) For c = 4 cos2 θ,
sin(m + 1)θ
Fm (c) = .
cm/2 sin θ
(ii) For ck = 4 cos2 π
k+1 ,

1 1
Fm−1 (ck ) − 2 Fm−2 (ck ) − m ≥ Fm (ck )
ck ck
for k ≥ 3, m = 2, 3, . . . , k.

Proof (i) follows from the trigonometric identity

m/2 
sin(m + 1)θ  m − j
= (−1)j (2 cos θ)m−2j .
sin θ j=0
j

To prove (ii), note that, from (ii) of Lemma 2.20,

 
1 1 1 1 1
Fm−1 (ck ) − 2 Fm−2 (ck ) − m − Fm (ck ) = − 2 Fm−2 (ck ) − m .
ck ck ck ck ck
64 Criteria for total positivity and strict total positivity

Now
sin 3θ
4 cos2 θ = + 1.
sin θ
Thus continuing the above and using (i) we have
(m−1)π
ck − 1 sin k+1 1
= − m
c2k c m−2
2 π
sin k+1 ck
k
 
3π (m−1)π
1  sin k+1 sin k+1 1 
= m+2  π π − m−2 
ck 2 sin k+1 sin k+1 π
2 cos k+1
 
1
3π
sin k+1 sin (m−1)π
k+1
≥ m+2 π π − 1 ≥ 0,
c 2 sin k+1 sin k+1
k

for k ≥ 3 and m = 2, 3, . . . , k. Looking at the last two terms we see that

we actually have strict inequality except in the case k = 3 and m = 2.

Note that from Lemma 2.21(i) we have that Fk (ck ) = 0 and Fk (cm ) > 0
for m < k.
The proof of Theorem 2.16 is based on an induction argument. To ease
exposition we assume, in the next series of lemmas, the following.

Assumption A Let A = (aij ) be an m × m matrix with strictly positive

entries such that
aij ai+1,j+1 > c ai,j+1 ai+1,j (2.11)
for all i, j = 1, . . . , m − 1,
 
r, . . . , m
A > 0, r = 2, . . . , m, (2.12)
r, . . . , m
and
     
r, . . . , m r + 1, . . . , m r + 2, . . . , m
A > arr A − ar,r+1 ar+1,r A
r, . . . , m r + 1, . . . , m r + 2, . . . , m
(2.13)
for r = 1, . . . , m − 2.

Lemma 2.22 Under Assumption A

 
   
r, . . . , m r + 1, . . . , m 1 r + 2, . . . , m
A > arr A − ar+1,r+1 A
r, . . . , m r + 1, . . . , m c r + 2, . . . , m
for r = 1, . . . , m − 2.
 2.6 “Simple” criteria for strict total positivity 65

Proof Using (2.13) and (2.11) with i = j = r

     
r, . . . , m r + 1, . . . , m r + 2, . . . , m
A > arr A − ar,r+1 ar+1,r A
r, . . . , m r + 1, . . . , m r + 2, . . . , m
   
r + 1, . . . , m 1 r + 2, . . . , m
> arr A − ar,r ar+1,r+1 A
r + 1, . . . , m c r + 2, . . . , m

   
r + 1, . . . , m 1 r + 2, . . . , m
= arr A − ar+1,r+1 A .
r + 1, . . . , m c r + 2, . . . , m

Lemma 2.23 Under Assumption A and if Fk (c) > 0, then

   
r, . . . , m 1 r + 1, . . . , m
Fk (c)A − Fk−1 (c)arr A
r, . . . , m c r + 1, . . . , m

   
r + 1, . . . , m 1 r + 2, . . . , m
> arr Fk+1 (c)A − Fk (c)ar+1,r+1 A
r + 1, . . . , m c r + 2, . . . , m
for r = 1, . . . , m − 2.

Proof As Fk (c) > 0 we multiply the result of Lemma 2.22 by Fk (c) to

obtain
 
r, . . . , m
Fk (c)A
r, . . . , m

   
r + 1, . . . , m 1 r + 2, . . . , m
> arr Fk (c)A − Fk (c)ar+1,r+1 A
r + 1, . . . , m c r + 2, . . . , m
for r = 1, . . . , m − 2. Thus
   
r, . . . , m 1 r + 1, . . . , m
Fk (c)A − Fk−1 (c)ar,r A
r, . . . , m c r + 1, . . . , m

   
1 r + 1, . . . , m
> arr Fk (c) − Fk−1 (c) A
c r + 1, . . . , m
 
1 r + 2, . . . , m
− Fk (c)ar+1,r+1 A .
c r + 2, . . . , m
Since
1
Fk+1 (c) = Fk (c) − Fk−1 (c)
c
we have, continuing the above,

   
r + 1, . . . , m 1 r + 2, . . . , m
= arr Fk+1 (c)A − Fk (c)ar+1,r+1 A .
r + 1, . . . , m c r + 2, . . . , m
66 Criteria for total positivity and strict total positivity

Note the recursive nature of the inequality of Lemma 2.23. We use

Lemma 2.23 in the proof of the next two lemmas.

Lemma 2.24 Under Assumption A and if Fk+
(c) > 0 for  = 0, 1, . . . , m−

1 − r, then
   
r, . . . , m 1 r + 1, . . . , m
Fk (c)A − Fk−1 (c)ar,r A
r, . . . , m c r + 1, . . . , m
> arr ar+1,r+1 · · · amm Fk+m−r+1 (c)

for r = 1, . . . , m − 2.

Proof We apply Lemma 2.23 m − 1 − r times, i.e.,

   
r, . . . , m 1 r + 1, . . . , m
Fk (c)A − Fk−1 (c)ar,r A
r, . . . , m c r + 1, . . . , m

   
r + 1, . . . , m 1 r + 2, . . . , m
> arr Fk+1 (c)A − Fk (c)ar+1,r+1 A
r + 1, . . . , m c r + 2, . . . , m

 
r + 2, . . . , m
> arr ar+1,r+1 Fk+2 (c)A
r + 2, . . . , m
 
1 r + 3, . . . , m
− Fk+1 (c)ar+2,r+2 A
c r + 3, . . . , m
> ···
> arr ar+1,r+1 · · · am−2,m−2

  
m − 1, m 1
× Fk+m−1−r (c)A − Fk+m−2−r (c)am−1,m−1 amm .
m − 1, m c

Now
 
m − 1, m
A = am−1,m−1 amm − am−1,m am,m−1
m − 1, m

and as

am−1,m−1 amm > c am−1,m am,m−1

we have
 
m − 1, m 1
A > am−1,m−1 amm − am−1,m−1 amm
m − 1, m c
 
1
= 1− am−1,m−1 amm .
c
 2.6 “Simple” criteria for strict total positivity 67

Substituting in the above gives

   
r, . . . , m 1 r + 1, . . . , m
Fk (c)A − Fk−1 (c)ar,r A
r, . . . , m c r + 1, . . . , m
> arr ar+1,r+1 · · · am−2,m−2

 
1
Fk+m−1−r (c) 1 − am−1,m−1 amm
c

1
− Fk+m−2−r (c)am−1,m−1 amm
c

  
1 1
= arr ar+1,r+1 · · · amm Fk+m−1−r (c) 1 − − Fk+m−2−r (c) .
c c
From the recurrence relation of Lemma 2.20,
1
Fk+m−r+1 (c) = Fk+m−r (c) − Fk+m−r−1 (c)
c
1 1
= Fk+m−r−1 (c) − Fk+m−r−2 (c) − Fk+m−r−1 (c).
c c
Thus
   
r, . . . , m 1 r + 1, . . . , m
Fk (c)A − Fk−1 (c)ar,r A
r, . . . , m c r + 1, . . . , m
> arr ar+1,r+1 · · · amm Fk+m−r+1 (c).

Lemma 2.25 Under Assumption A and if Fk (c) > 0, k = 1, . . . , s − 1,

then
 
1, . . . , m
A >
1, . . . , m

   
s + 1, . . . , m 1 s + 2, . . . , m
a11 · · · ass Fs (c)A − Fs−1 (c)as+1,s+1 A
s + 1, . . . , m c s + 2, . . . , m
for s = 1, . . . , m − 2.

Proof We start with r = 1 in Lemma 2.22, which can be written as

 
   
1, . . . , m 2, . . . , m 1 3, . . . , m
A > a11 A − a22 A
1, . . . , m 2, . . . , m c 3, . . . , m

   
2, . . . , m 1 3, . . . , m
= a11 F1 (c)A − F0 (c)a22 A .
2, . . . , m c 3, . . . , m
We now apply Lemma 2.23 s − 1 times for r = 2, . . . , s, with k = r − 1, to
obtain our result.
68 Criteria for total positivity and strict total positivity

We are now in a position to prove the sufficiency part of Theorem 2.16.

The induction hypothesis will be that if A = (aij ) is an m × m (m ≥ 3)
matrix with strictly positive entries, for which
aij ai+1,j+1 > cm ai,j+1 ai+1,j , (2.14)
for i, j = 1, . . . , m − 1, where cm = 4 cos2 m+1 π
; then A is strictly totally
positive and
     
1, . . . , m 2, . . . , m 3, . . . , m
A > a11 A − a12 a21 A .
1, . . . , m 2, . . . , m 3, . . . , m
We first prove that this result holds in the case m = 3. It is easily verified
that the inequality
     
1, 2, 3 2, 3 3
A > a11 A − a12 a21 A
1, 2, 3 2, 3 3
holds for every strictly totally positive matrix. Thus one only need verify
(since c3 = 2) that if A is a 3 × 3 matrix with strictly positive entries, for
which
aij ai+1,j+1 > 2 ai,j+1 ai+1,j ,
for i, j = 1, 2, then A is strictly totally positive. That is, we must verify
that det A > 0. Now
     
2, 3 2, 3 2, 3
det A = a11 A − a12 A + a13 A
2, 3 1, 3 1, 2
   
2, 3 2, 3
> a11 A − a12 A .
2, 3 1, 3
From (2.14) we have
a11 a22 > 2 a12 a21 , a22 a33 > 2 a23 a32 .
Thus
a11 a22 a33 > a12 a21 a33 + a11 a23 a32 ,
from which
 
2, 3
a11 A = a11 a22 a33 − a11 a23 a32 > a12 a21 a33
2, 3
 
2, 3
> a12 a21 a33 − a12 a23 a31 = a12 A .
1, 3
This proves the result in the case m = 3.
We assume that our induction hypothesis is valid for m < n and prove
it for m = n. We start with the following.
 2.6 “Simple” criteria for strict total positivity 69

Proposition 2.26 Under the above assumptions

   
2, . . . , n 2, . . . , n
a1j A − a1,j+1 A > 0,
1, . . . , j, . . . , n 1, . . . , j
+ 1, . . . , n
j = 2, . . . , n − 1.

Proof We use induction on the (n − 1) × (n − 1) submatrices



2, . . . , n 2, . . . , n
A , A .
1, . . . , j, . . . , n 1, . . . , j
+ 1, . . . , n
As cn−1 < cn we can apply the induction hypothesis, and thus both
matrices are strictly totally positive and also satisfy (2.13) of Assumption
A. We consider separately each of the two factors
   
2, . . . , n 2, . . . , n
A , A .
1, . . . , j, . . . , n 1, . . . , j
+ 1, . . . , n

By the Hadamard Inequality (Theorem 1.21)

   
2, . . . , n j + 2, . . . , n
A < a21 · · · aj+1,j A .
1, . . . , j
+ 1, . . . , n j + 2, . . . , n
Furthermore, from Lemma 2.19,

a1j aj+1,j+1 > cjn a1,j+1 , aj+1,j .

Thus
 
2, . . . , n
a1,j+1 A
1, . . . , j
+ 1, . . . , n
 
j + 2, . . . , n
< a21 · · · aj+1,j a1,j+1 A
j + 2, . . . , n
 
1 j + 2, . . . , n
< j a1j a21 · · · aj,j−1 aj+1,j+1 A . (2.15)
cn j + 2, . . . , n

Let us now consider

 
2, . . . , n
A .
1, . . . , j, . . . , n

As cn = 4 cos2 θn , θn = n+1
π
, and Fk (c) > 0 for c = 4 cos2 θ where θ < π
k+1
(see Lemma 2.21(i)), it follows that

Fk (cn ) > 0, k = 1, . . . , n − 1.
70 Criteria for total positivity and strict total positivity

Thus we can apply Lemma 2.25 with s = j − 2 to obtain

 
 
2, . . . , n j, j + 1, . . . , n
A > a21 · · · aj−1,j−2 Fj−2 (cn )A
1, . . . , j, . . . , n j − 1, j + 1, . . . , n
 
1 j + 1, . . . , n
− Fj−3 (cn )aj,j−1 A .
cn j + 1, . . . , n

By the induction hypothesis from (2.13)

 
j, j + 1, . . . , n
A
j − 1, j + 1, . . . , n
   
j + 1, . . . , n j + 2, . . . , n
> aj,j−1 A − aj,j+1 aj+1,j−1 A ,
j + 1, . . . , n j + 2, . . . , n

and thus
 
2, . . . , n
A
1, . . . , j, . . . , n
 

j + 1, . . . , n
> a21 · · · aj−1,j−2 aj,j−1 Fj−2 (cn )A
j + 1, . . . , n
 
j + 2, . . . , n
−aj,j+1 aj+1,j−1 Fj−2 (cn )A
j + 2, . . . , n
 
1 j + 1, . . . , n
− Fj−3 (cn ) aj,j−1 A
cn j + 1, . . . , n

  
1 j + 1, . . . , n
= a21 · · · aj−1,j−2 aj,j−1 Fj−2 (cn ) − Fj−3 (cn ) A
cn j + 1, . . . , n
 
j + 2, . . . , n
−aj,j+1 aj+1,j−1 Fj−2 (cn )A .
j + 2, . . . , n

Since

1
Fj−1 (c) = Fj−2 (c) − Fj−3 (c)
c

and

aj,j−1 aj+1,j+1 > c2n aj,j+1 aj+1,j−1

 2.6 “Simple” criteria for strict total positivity 71

we have
 
2, . . . , n
A
1, . . . , j, . . . , n
  
j + 1, . . . , n
> a21 · · · aj−1,j−2 aj,j−1 Fj−1 (cn )A
j + 1, . . . , n
 
aj,j−1 aj+1,j+1 j + 2, . . . , n
− Fj−2 (cn )A
c2n j + 2, . . . , n
  
j + 1, . . . , n
= a21 · · · aj−1,j−2 aj,j−1 Fj−1 (cn )A
j + 1, . . . , n
 
aj+1,j+1 j + 2, . . . , n
− Fj−2 (cn )A .
c2n j + 2, . . . , n

From Lemma 2.22

 
j + 1, . . . , n
A
j + 1, . . . , n

   
j + 2, . . . , n 1 j + 3, . . . , n
> aj+1,j+1 A − aj+2,j+2 A .
j + 2, . . . , n cn j + 3, . . . , n

Thus, finally,
 
2, . . . , n
A
1, . . . , j, . . . , n

> a21 · · · aj−1,j−2 aj,j−1 aj+1,j+1 Fj−1 (cn )
    
j + 2, . . . , n
1 j + 3, . . . , n
A − aj+2,j+2 A
cn
j + 2, . . . , n j + 3, . . . , n
 
1 j + 2, . . . , n
− 2 Fj−2 (cn )A
cn j + 2, . . . , n
 
1
= a21 · · · aj−1,j−2 aj,j−1 aj+1,j+1 Fj−1 (cn ) − 2 Fj−2 (cn )
cn
   
j + 2, . . . , n 1 j + 3, . . . , n
A − aj+2,j+2 Fj−1 (cn )A . (2.16)
j + 2, . . . , n cn j + 3, . . . , n
72 Criteria for total positivity and strict total positivity

We now combine (2.15) and (2.16) to obtain

   
2, . . . , n 2, . . . , n
a1j A − a1,j+1 A
1, . . . , j, . . . , n 1, . . . , j
+ 1, . . . , n
 
1
> a1j a21 · · · aj−1,j−2 aj,j−1 aj+1,j+1 Fj−1 (cn ) − 2 Fj−2 (cn )
cn
   
j + 2, . . . , n 1 j + 3, . . . , n
A − aj+2,j+2 Fj−1 (cn )A
j + 2, . . . , n cn j + 3, . . . , n
 
1 j + 2, . . . , n
− j a1j a21 · · · aj,j−1 aj+1,j+1 A
cn j + 2, . . . , n
 
1 1
= a1j a21 · · · aj−1,j−2 aj,j−1 aj+1,j+1 Fj−1 (cn ) − 2 Fj−2 (cn ) − j
cn cn
   
j + 2, . . . , n 1 j + 3, . . . , n
A − aj+2,j+2 Fj−1 (cn )A .
j + 2, . . . , n cn j + 3, . . . , n

From Lemma 2.21(ii)

1 1
Fj−1 (cn ) − Fj−2 (cn ) − j ≥ Fj (cn )
c2n cn
for n ≥ 3, j = 2, 3, . . . , n. Since for these values of j and n we have Fj (cn ) ≥
0, it follows that
   
2, . . . , n 2, . . . , n
a1j A − a1,j+1 A
1, . . . , j, . . . , n 1, . . . , j
+ 1, . . . , n
  
j + 2, . . . , n
> a1j a21 · · · aj−1,j−2 aj,j−1 aj+1,j+1 Fj (cn )A
j + 2, . . . , n
 
1 j + 3, . . . , n
− aj+2,j+2 Fj−1 (cn )A .
cn j + 3, . . . , n

From Lemma 2.24 (with k = j and r = j + 2) the above quantity satisfies

> a1j a21 · · · aj−1,j−2 aj,j−1 aj+1,j+1 aj+2,j+2 · · · ann Fn−1 (cn ).

As Fn−1 (cn ) > 0 this proves Proposition 2.26.

It now remains to advance the induction hypothesis, i.e., prove that A is

strictly totally positive and (2.13) holds. This demands proving
 
1, . . . , n
A >0 (2.17)
1, . . . , n
 2.6 “Simple” criteria for strict total positivity 73

and
     
1, . . . , n 2, . . . , n 3, . . . , n
A > a11 A − a12 a21 A . (2.18)
1, . . . , n 2, . . . , n 3, . . . , n
From Proposition 2.26 (which assumed the validity of the result for (n −
1) × (n − 1) matrices)
  n  
1, . . . , n 2, . . . , n
A = (−1)k+1 a1k A
1, . . . , n 1, . . . , k, . . . , n
k=1
   
2, . . . , n 2, 3, . . . , n
> a11 A − a12 A .
2, . . . , n 1, 3, . . . , n
By the Hadamard Inequality
   
2, . . . , n 3, . . . , n
A < a21 A
1, 3, . . . , n 3, . . . , n
and thus
     
1, . . . , n 2, . . . , n 3, . . . , n
A > a11 A − a12 a21 A .
1, . . . , n 2, . . . , n 3, . . . , n
This proves (2.18).
As (2.12) holds for r = 2, . . . , n, by the induction hypothesis, and since we
have now proven (2.18), which completes (2.13) for m = n (again applying
the induction hypothesis), we have that the results of Lemmas 2.22–2.25
hold for m = n. From Lemma 2.25 (with m = n and s = n − 2)
 
1, . . . , n
A
1, . . . , n

  
n − 1, n 1
> a11 · · · an−2,n−2 Fn−2 (cn )A − Fn−3 (cn )an−1,n−1 ann .
n − 1, n cn
Now (see the proof of Lemma 2.24)
   
n − 1, n 1
A = an−1,n−1 ann − an−1,n an,n−1 > 1 − an−1,n−1 ann .
n − 1, n cn
Thus
 
n − 1, n 1
Fn−2 (cn )A − Fn−3 (cn )an−1,n−1 ann
n − 1, n cn
 
1 1
> Fn−2 (cn ) 1 − an−1,n−1 ann − Fn−3 (cn )an−1,n−1 ann
cn cn

  
1 1
= an−1,n−1 ann Fn−2 (cn ) 1 − − Fn−3 (cn ) .
cn cn
74 Criteria for total positivity and strict total positivity

From the recurrence relation of Lemma 2.20(ii)

1
0 = Fn (cn ) = Fn−1 (cn ) − Fn−2 (cn )
cn
1 1
= Fn−2 (cn ) − Fn−3 (cn ) − Fn−2 (cn ).
cn cn
Substituting it follows that
 
1, . . . , n
A > 0.
1, . . . , n
This advances the induction and completes the proof of Theorem 2.16.

2.7 Remarks
Fekete’s Lemma 2.1 can be found in Fekete, Pólya [1912]. (Different sections
of this paper were written by the different authors.) The fact that for
strictly totally positive matrices it suffices to consider only minors with
consecutive rows and consecutive columns implies that the remaining
minors are nonnegative combinations of the minors with consecutive rows
and consecutive columns. If we detail the proof of Fekete’s Lemma 2.1 we
see that the coefficients in this combination depend upon other minors of
the same size, plus lower order minors, and in a nonlinear fashion. Such
a formula is not valid for totally positive matrices in general. Theorem
2.3 was proved in Gasca, Peña [1992]. Their proof uses factorization, a
subject to be considered in Chapter 6 (see also Metelmann [1973]). The
two proofs given here are different. Proposition 2.5 first appears in Gasca,
Peña [1992]. It is reproved in Gladwell [1998]. Theorem 2.6 is due to
Whitney [1952]. Proposition 1.9 is also to be found in Whitney [1952].
Theorem 2.8 appears in Karlin [1968], p. 85. Proposition 2.9 can be found
in Shapiro, Shapiro [1995]. Theorem 2.10 is in Cryer [1973]. Theorem 2.12
was conjectured in Cryer [1973] and proved in Cryer [1976] by very different
methods. The proof given here can be found in Micchelli, Pinkus [1991].
The example of a 4 × 4 matrix, which is not totally positive but whose
minors with consecutive rows and consecutive columns are nonnegative, is
from Cryer [1973]. Theorem 2.13 was proved in Cryer [1976]. Theorem 2.14
was conjectured in Cryer [1973] and proved in Cryer [1976]. Proposition
2.15 is from Gasca, Peña [1993].
In Proposition 1.20, as a consequence of Theorem 1.19, we proved that
a nonsingular totally positive matrix A is almost strictly totally positive if
 
i + 1, . . . , i + p
A >0
j + 1, . . . , j + p
 2.7 Remarks 75

whenever ai+k,j+k > 0, k = 1, . . . , p, for all possible i, j and p. As a further

consequence of Theorem 1.19, paralleling Proposition 2.5 and Proposition
2.9, it is not difficult to show that for a nonsingular totally positive matrix
it is hardly necessary to verify the above conditions for all i, j, and p in
order to determine if the matrix A is almost strictly totally positive. If we
know the zero entries of A then, from Theorem 1.19, we can determine
a minimal number of such conditions that must be verified. This is done
using other methods in Gasca, Peña [1995], Gladwell [2004], and Gasca,
Peña [2006].
Craven, Csordas [1998] first proved that if A = (aij ) is an n × m matrix
with strictly positive entries and
aij ai+1,j+1 ≥ c∗ ai,j+1 ai+1,j
for all i = 1, . . . , n − 1, j = 1, . . . , m − 1, where c∗ = 4.07 . . . is the unique
real root of x3 −5x2 +4x−1 = 0, then A is strictly totally positive. The fact
that there is such a universal constant was in itself a surprising result. There
was no claim made in the paper that their c∗ was best possible. On the
other hand it is not difficult to show that we must have c∗ ≥ 4. Dimitrov,
Peña [2005] conjectured that c∗ = 4, in general, and cn = 4 cos2 n+1 π
is
best possible for each n × n matrix. Unaware of Dimitrov, Peña [2005], this
conjecture was proved by Katkova, Vishnyakova [2006]. It is their proof
that is presented here. A simpler proof would be of interest.
 3
Variation diminishing

The variation diminishing property of a matrix was introduced by

I. J. Schoenberg in 1930. In Section 3.1 we prove two fundamental results
concerning variation diminishing properties of totally positive matrices.
The term “variation diminishing” was coined by Pólya to describe the
property whereby the number of sign changes of Ax is bounded above by
the number of sign changes of x. The term is somewhat of a misnomer since
it is not the “variation” that does not increase. The variation diminishing
property and the exact form in which these bounds hold are the content of
the two theorems of Section 3.1. In Section 3.2 we present one application.
Other applications can be found in subsequent chapters.

3.1 Main equivalence theorems

To explain the concept of variation diminishing we first fix some notation.
We define two counts for the number of sign changes of a vector c =
(c1 , . . . , cn ) ∈ Rn . These are

S − (c) – the number of sign changes in the sequence c1 , . . . , cn , with zero

terms discarded. Obviously S − (0) = 0.
+
S (c) – the maximum number of sign changes in the sequence c1 , . . . , cn ,
where zero terms are arbitrarily assigned values +1 or −1. For
convenience we set S + (0) = n.

Thus, for example,

S − (1, 0, 2, −3, 0, 1) = 2, and S + (1, 0, 2, −3, 0, 1) = 4 .

In our “S − count” we have a sign change from 2 to −3 and another from

−3 to 1. In our “S + count,” by giving the first (left-most) 0 the value −1,

76
 3.1 Main equivalence theorems 77

we get two additional sign changes. Giving the second 0 the value 1 or −1
does not alter the number of sign changes.
The following elementary facts will be used. Their proof is left to the
reader.

Lemma 3.1 For every c = (c1 , c2 , . . . , cn ) ∈ Rn \{0} we have

S + (c) + S − (!
c) = n − 1,

where !c = (c1 , −c2 , . . . , (−1)n−1 cn ), i.e., we replace cr in c by (−1)r−1 cr

for each r.

Lemma 3.2 If
lim ck = c
k→∞

then
lim S − (ck ) ≥ S − (c)
k→∞

and
lim S + (ck ) ≤ S + (c).
k→∞

The first main result is the following.

Theorem 3.3 Let A be an n × m strictly totally positive matrix. Then

(a) for each vector x ∈ Rm , x = 0,

S + (Ax) ≤ S − (x);

(b) if S + (Ax) = S − (x) and Ax = 0, then the sign of the first (and last)
component of Ax (if zero, the sign given in determining S + (Ax))
agrees with the sign of the first (and last) nonzero component of x.
Conversely, if (a) and (b) hold for some n×m matrix A and every x ∈ Rm ,
x = 0, then A is strictly totally positive.

Proof We first prove that if A is strictly totally positive then (a) and (b)
hold.
Assume x = 0 and S − (x) = r. The components of x may be therefore
divided into r + 1 blocks

(x1 , . . . , xs1 ), (xs1 +1 , . . . , xs2 ), . . . , (xsr +1 , . . . , xm )

78 Variation diminishing

where, without loss of generality, we assume that

xk (−1)i+1 ≥ 0, k = si−1 + 1, . . . , si ,
for i = 1, . . . , r + 1, for each i strict inequality holds for at least one k, and
where we set s0 = 0 and sr+1 = m. Let a1 , . . . , am ∈ Rn denote the column
vectors of A and set
si
yi = |xk |ak , i = 1, . . . , r + 1.
k=si−1 +1

Thus

m 
r+1
Ax = xk ak = (−1)i+1 yi .
k=1 i=1

Let Y be the n × (r + 1) matrix with columns y1 , . . . , yr+1 . Set

w = Ax = Y d,
where d = (1, −1, . . . , (−1)r ) ∈ Rr+1 . Now a simple calculation shows that
  
sj1

sjp  
i1 , . . . , ip i1 , . . . , ip
Y = ··· |xk1 | · · · |xkp |A .
j1 , . . . , jp k1 , . . . , kp
k1 =sj1 −1 +1 kp =sjp −1 +1

As  
i1 , . . . , ip
A >0
k1 , . . . , kp
for all choices of 1 ≤ i1 < · · · < ip ≤ n, 1 ≤ k1 < · · · < kp ≤ m, and
|xk1 | · · · |xkp | > 0
for some choice of admissible {k1 , . . . , kp } in the above sum, it follows that
Y is an n × (r + 1) strictly totally positive matrix.
We first prove that S + (Ax) ≤ S − (x). If n ≤ r + 1 there is nothing to
prove. For if Ax = 0, then
S + (Ax) ≤ n − 1 ≤ r = S − (x).
If n = r + 1 then Ax = 0 since Ax = Y d where Y is a nonsingular
(r + 1) × (r + 1) matrix and d = 0. Thus, if Ax = 0 then n ≤ r and
S + (Ax) = n ≤ r = S − (x).
We shall therefore assume that n > r + 1. If S + (Ax) ≥ r + 1, then there
exist indices 1 ≤ i1 < · · · < ir+2 ≤ n such that
εwij (−1)j+1 ≥ 0, j = 1, . . . , r + 2,
 3.1 Main equivalence theorems 79

for some ε ∈ {−1, 1} where w = Ax = Y d. Since Y is of rank r + 1, at

least two of the wij are not zero. Consider the determinant

wi1 y i1 1 ··· yi1 r+1

.. .. .. .
. . .
wir+2 yir+2 1 ··· yir+2 r+1
This determinant is zero since the first column is a linear combination of
the other columns. Moreover expanding by this first column gives


r+2
i1 , . . . , ij , . . . , ir+2
j+1
0= (−1) wij Y .
j=1
1, . . . , r + 1

This equality cannot possibly hold since Y is strictly totally positive and
wij (−1)j+1 is of one fixed (weak) sign and some of the wij are not zero.
This contradiction implies that S + (Ax) ≤ r = S − (x).
It remains to prove (b), namely that if S + (Ax) = S − (x) and Ax = 0,
then the sign of the first (and last) component of Ax (if zero, the sign given
in determining S + (Ax)) agrees with the sign of the first (and last) nonzero
component of x. We continue our analysis using the above notation. We
wish to prove that if S + (Ax) = S − (x) = r, Ax = 0, and
εwij (−1)j+1 ≥ 0, j = 1, . . . , r + 1,
where w = Y d = Ax, then ε = 1. Since every set of r + 1 rows of Y is
linearly independent, we can solve for any component of d based on the
values wi1 , . . . , wir+1 . Thus

wi1 y i1 2 ··· yi1 r+1

.. .. ..
. . .
wir+1 yir+1 2 · · · yir+1 r+1
1=   .
i1 , . . . , ir+1
Y
1, . . . , r + 1
Expanding by the first column gives


r+1
i1 , . . . , ij , . . . , ir+1
wij (−1)j+1 Y
j=1
2, . . . , r + 1
1=   .
i1 , . . . , ir+1
Y
1, . . . , r + 1
As
εwij (−1)j+1 ≥ 0, j = 1, . . . , r + 1,
80 Variation diminishing

and Y is strictly totally positive, we obtain ε = 1.

We now prove the converse direction. Its proof is surprisingly simple. We
assume that S + (Ax) ≤ S − (x) for every x = 0, and if S + (Ax) = S − (x)
and Ax = 0, then the sign of the first (and last) component of Ax (if zero,
the sign given in determining S + (Ax)) agrees with the sign of the first (and
last) nonzero component of x. We shall prove that A is a strictly totally
positive matrix.
Our proof is via induction on the size of the minors. We start as follows.
Take ej , the jth unit vector in Rm . As S − (ej ) = 0 and its nonzero
component is positive, we must have that Aej is a strictly positive vector.
Thus the elements of A are all strictly positive.
Now, assume that all (r − 1) × (r − 1) minors of A are strictly positive,
2 ≤ r ≤ min{n, m}, and consider the r × r submatrix


i1 , . . . , ir
B=A .
j1 , . . . , jr

If
 
i1 , . . . , ir
A =0
j1 , . . . , jr

then there exists a z = (zj1 , . . . , zjr ) ∈ Rr \{0} such that Bz = 0. Let
z ∈ Rm be the vector whose jk component is the above zjk , k = 1, . . . , r, and
whose other components are zero. Then S − (z) ≤ r − 1 while S + (Az) ≥ r.
This is a contradiction. Thus each r × r minor of A is nonsingular.
We now let x = (xj1 , . . . , xjr ) ∈ Rr satisfy Bx = d where

d = (1, −1, . . . , (−1)r+1 ).

Let x ∈ Rm be the vector whose jk component is the above xjk , k =

1, . . . , r, and whose other components are zero. Note that S − (x) ≤ r − 1.
From this construction of x we have (Ax)ik = (−1)k+1 , k = 1, . . . , r. Thus
S + (Ax) ≥ r − 1. Since, by assumption, S − (Ax) ≤ S − (x) we necessarily
have

S − (Ax) = S − (x) = r − 1

and the sign patterns must also agree. Thus

(−1)k+1 xjk > 0, k = 1, . . . , r.

 3.1 Main equivalence theorems 81

We now return to the equations Bx = d and solve for xj1 to obtain

r
i1 , . . . , i
, . . . , ir
A
j2 , . . . , jr
0 < xj1 =
=1   .
i1 , . . . , ir
A
j1 , . . . , jr
From the induction hypothesis the numerator is positive. Thus
 
i1 , . . . , ir
A > 0.
j1 , . . . , jr

For totally positive matrices the result is slightly weaker, but of the same
form.

Theorem 3.4 Let A be an n × m totally positive matrix. Then

(a) for each vector x ∈ Rm ,

S − (Ax) ≤ S − (x);

(b) if S − (Ax) = S − (x) and Ax = 0, then the sign of the first (and last)
nonzero component of Ax agrees with the sign of the first (and last)
nonzero component of x.
Conversely, if (a) and (b) hold for some n×m matrix A and every x ∈ Rm ,
then A is totally positive.

Proof Assume A is an n × m totally positive matrix. From Theorem 2.6

there exists a sequence of n × m strictly totally positive matrices (Ak ) for
which
lim Ak = A.
k→∞

Assume x = 0. Then from Theorem 3.3

S + (Ak x) ≤ S − (x).

From Lemma 3.2 it follows that

S − (Ax) ≤ lim S − (Ak x).

k→∞

As S − (Ak x) ≤ S + (Ak x) always holds this implies

S − (Ax) ≤ S − (x),

which is (a).
82 Variation diminishing

If S − (Ax) = S − (x) = r and Ax = 0, then for all k sufficiently large we

necessarily have, using Lemma 3.2 and Theorem 3.3,

r = S − (Ax) ≤ S − (Ak x) ≤ S + (Ak x) ≤ S − (x) = r,

and thus

r = S − (Ax) = S − (Ak x) = S + (Ak x) = S − (x) = r.

Since, for k sufficiently large, S − (Ak x) = S + (Ak x) the sign patterns of Ak x

do not depend upon any zero entries. From Theorem 3.3 the sign patterns
of Ak x agree with those of x. From a limiting argument these nonzero sign
patterns of Ax and Ak x agree. This proves (b).
The proof of the converse direction is essentially the same as the
corresponding proof in Theorem 3.3. We assume that S − (Ax) ≤ S − (x)
for every x, and if S − (Ax) = S − (x) then the sign of the first (and last)
component of Ax agrees with the sign of the first (and last) nonzero
component of x. We shall prove that A is a totally positive matrix.
Our proof is by induction on the size of the minors. We start as follows.
Take ej , the jth unit vector in Rm . As S − (ej ) = 0 and its nonzero
component is positive we then have that Aej is a nonnegative vector. Thus
the elements of A are all nonnegative. Now assume that all (r − 1) × (r − 1)
minors are nonnegative, 2 ≤ r ≤ min{n, m}, and consider the r × r
submatrix


i1 , . . . , ir
B=A .
j1 , . . . , jr
If B is singular, there is nothing to prove. Assume B is nonsingular. We
now repeat the argument of the previous theorem.
Let x = (xj1 , . . . , xjr ) ∈ Rr \{0} satisfy Bx = d where

d = (1, −1, . . . , (−1)r+1 ).

Let x ∈ Rm be the vector whose jk component is the above xjk , k =

1, . . . , r, and whose other components are zero. Note that S − (x) ≤ r − 1.
From this construction of x we have (Ax)ik = (−1)k+1 , k = 1, . . . , r. Thus
S − (Ax) ≥ r − 1. Since, by assumption, S − (Ax) ≤ S − (x) we necessarily
have
S − (Ax) = S − (x) = r − 1

and the sign patterns must also agree. Thus

(−1)k+1 xjk > 0, k = 1, . . . , r.

 3.2 Intervals of strict total positivity 83

We now return to the equations Bx = d and solve for xj1 to obtain

r
i1 , . . . , i
, . . . , ir
A
j2 , . . . , jr
0 < xj1 =
=1   .
i1 , . . . , ir
A
j1 , . . . , jr
From the induction hypothesis the numerator is positive. Thus
 
i1 , . . . , ir
A > 0.
j1 , . . . , jr

We also have the following if A is both totally positive and nonsingular.

Corollary 3.5 Let A be an n × n nonsingular totally positive matrix. Then

for each x ∈ Rn
S + (Ax) ≤ S + (x).

Proof Let D denote the n × n diagonal matrix with diagonal entries

alternately 1 and −1. From Proposition 1.6 DA−1 D is a nonsingular totally
positive matrix. As such

S − (DA−1 Dy) ≤ S − (y)

for all y ∈ Rn . In addition, we recall from Lemma 3.1 that

S + (c) + S − (Dc) = n − 1

for all c ∈ Rn \{0}.

We may assume that x = 0, as there is nothing to prove if x = 0. Setting
y = DAx we have x = A−1 Dy and from the above

S + (Ax) = S + (Dy) = (n − 1) − S − (y) ≤ (n − 1) − S − (DA−1 Dy)

= S + (A−1 Dy) = S + (x).

3.2 Intervals of strict total positivity

As shall be seen in subsequent chapters, there are numerous applications of
the variation diminishing property. We include here one such application.
Let B = (bij ) and C = (cij ) be n × n matrices. We apply an ordering
to these matrices. We say that B ≤∗ C if (−1)i+j bij ≤ (−1)i+j cij for all
i, j. Equivalently, if D is the n × n diagonal matrix with diagonal entries
84 Variation diminishing

alternately 1 and −1, then DBD ≤ DCD where ≤ is the usual entrywise
inequality.

Theorem 3.6 Assume B = (bij ) and C = (cij ) are strictly totally positive
n × n matrices satisfying
B ≤∗ C.
If A is an n × n matrix and
B ≤∗ A ≤∗ C,
then A is strictly totally positive.

Proof We prove, under the above assumptions, that A is necessarily

nonsingular. By continuity it then follows that the determinant of A is
positive. Assume, for the moment, that we have proved this fact and let us
see how it implies our theorem.
Recall, from Theorem 2.2 (or Theorem 2.3), that to prove that A is
strictly totally positive it suffices to show that
 
i + 1, . . . , i + k
A >0
j + 1, . . . , j + k
for i = 0, . . . , n − k, j = 0, . . . , n − k, and k = 1, . . . , n. For each such i, j,
and k consider the three k × k submatrices


i + 1, . . . , i + k
B ij = B ,
j + 1, . . . , j + k


i + 1, . . . , i + k
Aij = A ,
j + 1, . . . , j + k


i + 1, . . . , i + k
C ij = C .
j + 1, . . . , j + k
By assumption
B ij ≤∗ Aij ≤∗ C ij
for i + j even, while
C ij ≤∗ Aij ≤∗ B ij
for i + j odd. This implies, using an induction assumption, that
 
i + 1, . . . , i + k
A > 0,
j + 1, . . . , j + k
and the theorem is proved. Thus it suffices to prove the nonsingularity
of A.
 3.3 Remarks 85

We therefore assume that A is singular, and all k × k minors of A are

strictly positive for k = 1, . . . , n − 1. Let x = 0 satisfy Ax = 0. As A is of
rank n − 1 and
 
2, . . . , n
A >0
1, . . . , j, . . . , n
for j = 1, . . . , n, it follows that
xj xj+1 < 0, j = 1, . . . , n − 1.
Assume, without loss of generality, that
(−1)j xj > 0, j = 1, . . . , n. (3.1)
Since
(−1)i+j bij ≤ (−1)i+j aij , i, j = 1, . . . , n,
we have

n 
n 
n
(−1) i
bij xj = i+j
(−1) bij (−1) xj ≤
j
(−1)i+j aij (−1)j xj
j=1 j=1 j=1

n
= (−1)i aij xj = 0.
j=1

That is,
(−1)i (Bx)i ≤ (−1)i (Ax)i = 0, i = 1, . . . , n. (3.2)

As B is nonsingular we have Bx = 0. Thus

n − 1 = S + (Bx) = S − (x) = n − 1.
However from (3.1) and (3.2) we see that the sign of the first (and last)
component of Bx (if zero, the sign given in determining S + (Bx)) does not
agree with the sign of the first (and last) nonzero component of x. This
contradicts Theorem 3.3. Thus A is nonsingular.

3.3 Remarks
According to Schoenberg [1930] the term variation diminishing
(variationsvermindernd in German) was coined by Pólya. To obtain the
variation diminishing property of a matrix, i.e., S + (Ax) ≤ S − (x) or
S − (Ax) ≤ S − (x) one does not need that the matrix A be strictly totally
positive or totally positive, respectively. If A is an n × m matrix and n ≥ m
then A satisfies S + (Ax) ≤ S − (x) for all x ∈ Rm \{0} if and only if A is
86 Variation diminishing

strictly sign regular (see e.g., Karlin [1968], p. 219, or Ando [1987], p. 29).
An n × m matrix A is strictly sign regular if all its minors of order k are of
one fixed sign εk ∈ {−1, 1} for each k = 1, . . . , m. If we permit minors to
also vanish, then A is said to be sign regular. An n × m matrix A of rank
m satisfies S − (Ax) ≤ S − (x) for all x ∈ Rm if and only if A is sign regular.
This was first proven in Schoenberg [1930] with this rank condition (see also
Motzkin [1936] and Schoenberg, Whitney [1951] for the more general case).
The inequality n ≥ m and the rank condition, or something similar, is truly
needed in these results. However, for strictly totally positive and totally
positive matrices we have property (b) in the statements of Theorems 3.3
and 3.4. This property obviates the need for the rank condition and the
inequality n ≥ m. This fact seems to be generally overlooked. In this regard
see Brown, Johnstone, MacGibbon [1981], Lemma A.1, and Fan [1966],
Theorems 5 and 6.
For a generalization of sorts of variation diminishing, see Carnicer,
Goodman, Peña [1995].
Garloff [1982] proved Theorem 3.6 by different methods. This theorem
cannot hold for arbitrary totally positive matrices B and C. (Let B and C
be matrices all of whose entries are 0 except for a single 1, suitably chosen.)
In Garloff [1982] it is conjectured that the theorem holds if B and C are
totally positive and nonsingular. In Garloff [1982] and Garloff [2003] the
conjecture is proved for specific subclasses of such matrices.
 4
Examples

In this chapter we present various examples of totally positive matrices.

Some are specific examples, e.g., exponentials, powers, the Cauchy matrix,
and the binomial coefficients. Others are examples of more general classes
of matrices where we present conditions determining total positivity
within the class, e.g., Green, Jacobi, Hankel, Toeplitz and Hurwitz
matrices. Totally positive matrices are not closed under the operation of
Hadamard products. In the last section we consider subclasses of totally
positive matrices that are closed under this operation. We start, however,
by discussing the connection between totally positive matrices, on the
one hand, and totally positive kernels and Descartes systems, on the
other hand.

4.1 Totally positive kernels and Descartes systems

There are two types of continuous extensions of totally positive or strictly
totally positive matrices or, from the opposite perspective, totally positive
or strictly totally positive matrices are discrete counterparts of two
continuous forms.
A real-valued kernel K ∈ C(X × Y ) is said to be a totally positive or
strictly totally positive kernel if
n
(K(xi , yj ))i=1 m
j=1

is, respectively, a totally positive or strictly totally positive matrix for every
choice of x1 < · · · < xn in X, y1 < · · · < ym in Y and all possible n and m.
If X and Y are finite sets then this is simply a definition of a totally positive
or strictly totally positive matrix. As we shall see in the next two sections
K(x, y) = exy is a strictly totally positive kernel on R × R, K(x, y) = xy

87
88 Examples

is a strictly totally positive kernel on R+ × R and K(x, y) = 1/(x + y) is a

strictly totally positive kernel on R+ × R+ .
The semicontinuous analogue of a strictly totally positive matrix, or
semidiscrete analogue of a strictly totally positive kernel is given by a
Descartes system. To define a Descartes system we start with a Chebyshev
system. A Chebyshev system (generally abbreviated as a T -system) is a
system of functions {u1 , . . . , ur } defined on a set X for which
r
det (ui (xj ))i,j=1 = 0
for every choice of distinct x1 , . . . , xr in X. It is called a T+ -system if
r
det (ui (xj ))i,j=1 > 0
for every choice of x1 < · · · < xr in X. A system of functions {u1 , . . . , ur }
is said to be a Descartes system if every ordered subset of the u1 , . . . , ur is
a T+ -system. Equivalently
r
(ui (xj ))i=1 m
j=1

is a strictly totally positive matrix for all x1 < · · · < xm in X.

Descartes systems possess a variation diminishing property that follows
easily from Theorem 3.3, often termed a Generalized Descartes Rule of
Signs. It is that if {u1 , . . . , ur } is a Descartes system on an ordered set X,
then for each nontrivial “polynomial”

r
u(x) = ai ui (x)
i=1

the number of zeros of u on X is bounded above by S − (a1 , . . . , ar ). As will

be seen in the next section {xcj }rj=1 is a Descartes system on R+ for any
c1 < · · · < cr .

4.2 Exponentials and powers

Here are some elementary but instructive examples of strictly totally
positive matrices. Certain specific cases have already been considered (e.g.,
see the proof of Theorem 2.6).
Let
A = (ebi cj )ni,j=1
where b1 < · · · < bn and c1 < · · · < cn . We claim the A is a strictly totally
positive matrix. (In other words we are proving that K(x, y) = exy is a
strictly totally positive kernel on R × R.) We present a two-step proof. We
 4.2 Exponentials and powers 89

first prove that the determinant of A is nonzero, and then we prove that it
is positive.

Proof If A is singular there exists a (λ1 , . . . , λn ) ∈ Rn \{0} for which


n
λj ebi cj = 0, i = 1, . . . , n.
j=1

Set

n
f (x) = λj excj .
j=1

The function f is a nontrivial exponential polynomial with at least n

distinct zeros b1 , . . . , bn . It is not identically zero since, for example, if
m is the largest index for which λm = 0, then

lim f (x)e−xcm = λm = 0.
x→∞

We claim that no such function exists, i.e., every exponential polynomial

of this form has at most n − 1 distinct zeros, hence a contradiction, and A
is nonsingular. We prove this claim by induction on n. It obviously holds
for n = 1. Assume, as above, that f has at least n distinct zeros, n ≥ 2.
Multiply f by e−xcn and differentiate to obtain

n−1
g(x) = (f (x)e−xcn ) = λj (cj − cn )ex(cj −cn ) .
j=1

The function g is an exponential polynomial with, by Rolle’s Theorem, at

least n − 1 zeros. Thus, by our induction hypothesis, g is identically zero,
i.e., λj = 0, j = 1, . . . , n − 1. Therefore f (x) = λn excn . As f has a zero we
have λn = 0. Thus A is nonsingular.
The (bi )ni=1 and (cj )nj=1 are arbitrary increasing sequences. As such,
proving A nonsingular also proves the nonsingularity of all minors of A.
Moreover, by continuity, the sign of the determinant of A is independent
of the choice of the arbitrary increasing sequences (bi )ni=1 and (cj )nj=1 , but
may depend upon n. We prove that it is positive for each and every n.
Here are two proofs of the positivity of the determinant of A.
Assume that all k × k determinants of the form of A are strictly positive
for k = 1, . . . , n − 1. (The case k = 1 is obvious.) Let

n
f (x) = λj excj
j=1
90 Examples

satisfy f (bi ) = 0, i = 1, . . . , n − 1 and f (bn ) = 1. Such an f exists and is

unique by the nonsingularity of the associated matrix. In addition,

det (ebi cj )n−1

i,j=1
λn = .
det (ebi cj )ni,j=1

We proved earlier that f has at most n−1 zeros. Since f (bn ) = 1 it therefore
follows that f is strictly positive for all x > bn−1 . As such

λn = lim f (x)e−xcn > 0.

x→∞

By our induction hypothesis

det (ebi cj )n−1

i,j=1 > 0

and therefore
det (ebi cj )ni,j=1 > 0.

A second proof uses the fact that the sign of the determinant of A is
independent of the choice of the arbitrary increasing sequences (bi )ni=1 and
(cj )nj=1 , but may depend upon n. As such it suffices to explicitly calculate
the determinant of A for some specific choice of the (bi )ni=1 and (cj )nj=1 .
Let cj = j − 1 and set ebi = xi . Then we have

A = (xji )ni=1 n−1

j=0

where 0 < x1 < · · · < xn . This is the well-known Vandermonde matrix

whose determinant is given by

det A = (xk − x
) > 0.
1≤
<k≤n

Note that with the above substitution ebi = xi we also have that
c
A = (xi j )ni,j=1

is strictly totally positive for any 0 < x1 < · · · < xn and c1 < · · · < cn .
There is another method of proof of the total positivity of the (ebi cj ) (for
bi , cj > 0) based on the following idea. Let

m
G(x, y) = λk xk y k ,
k=0

with λk ≥ 0, k = 0, 1, . . . , m. (We assume here that m is finite, but that

 4.2 Exponentials and powers 91

need not be the case, and is not the case for G(x, y) = exy .) Then for any
0 < b1 < · · · < bn and 0 < c1 < · · · < cn the matrix

A = (G(bi , cj ))ni,j=1

is totally positive. This follows easily from the fact that

 0   
b1 . . . bm1 λ0 . . . 0 c01 . . . c0n
 ..   .. . . .  ..  .
A =  ... . . . .  . . ..  
..
.
..
. . 
b0n . . . bm
n 0 . . . λn cm
1
m
. . . cn
Set
 
b01 . . . bm
1
 .. .. ..  ,
B =  . . . 
b0n . . . bm
n
 
λ0 ... 0
 .. .. . 
Λ =  . . ..  ,
0 . . . λn
 0 
c1 . . . c0n
 .. .. ..  .
C =  . . . 
cm
1
m
. . . cn
Thus
A = BΛC.

As B and C are strictly totally positive (they are Vandermonde matrices)

and Λ is a diagonal totally positive matrix, the result follows.
In this way it can be proven that
∞
 x2k y 2k
exy + e−xy
G1 (x, y) = cosh xy = =
2 (2k)!
k=0

and
∞
 x2k+1 y 2k+1
exy − e−xy
G2 (x, y) = sinh xy = =
2 (2k + 1)!
k=0

are strictly totally positive on R+ × R+ . These facts can also be proven

directly as a consequence of the possible number of zeros of any linear
combination of the exponentials exc1 , e−xc1 , . . . , excm , e−xcm and the parity
of cosh xy and sinh xy.
92 Examples

4.3 Cauchy matrix

The matrix
 n
1
A=
xi + yj i,j=1

is called a Cauchy matrix. For 0 < x1 < · · · < xn , 0 < y1 < · · · < yn it is
strictly totally positive.
We provide two proofs of this result.

Proof I There is an explicit formula for the determinant of the above matrix
and it is given by
 n # #
1 1≤
<k≤n (xk − x
) 1≤
<k≤n (yk − y
)
det = #n .
xi + yj i,j=1 i,j=1 (xi + yj )

Here is a proof of this result. Obviously the result holds for n = 1. Now
1 1 1 x1 +y1 x1 +y1
x1 +y1 x1 +y2 ··· x1 +yn 1 x1 +y2 ··· x1 +yn
1 1 1 1 1 1
x2 +y1 x2 +y2 ··· x2 +yn 1 x2 +y1 x2 +y2 ··· x2 +yn
.. .. .. .. = .. .. .. .. .
. . . . x1 + y1 . . . .
1 1 1 1 1 1
xn +y1 xn +y2 ··· xn +yn xn +y1 xn +y2 ··· xn +yn

1
Multiply the first row by xi +y1 and subtract from the ith row to obtain
x1 +y1 x1 +y1
1 x1 +y2 ··· x1 +yn
(x2 −x1 )(y2 −y1 ) (x2 −x1 )(yn −y1 )
1 0 (x2 +y2 )(x1 +y2 )(x2 +y1 ) ··· (x2 +yn )(x1 +yn )(x2 +y1 )
= .. .. .. ..
x1 + y1 . . . .
(xn −x1 )(y2 −y1 ) (xn −x1 )(yn −y1 )
0 (xn +y2 )(x1 +y2 )(xn +y1 ) ··· (xn +yn )(x1 +yn )(xn +y1 )

1 xx11 +y
+y2
1 +y1
· · · xx11+y n
#n 0 x2 +y1
· · · 1
(x − x1 )(yk − y1 ) x2 +yn
= #n k
k=1 #n ..
2
.. .. ..
(x1 + y1 ) i=2 (xi + y1 ) j=2 (x1 + yj ) . . . .
1 1
0 xn +y2 ··· xn +yn

and now apply an induction argument.

Proof II We first prove that A is nonsingular.

Assume that det A = 0. There then exist a1 , . . . , an , not all zero, for
which
n
aj
= 0, x = x1 , . . . , xn .
j=1
x + yj
 4.3 Cauchy matrix 93

Thus # 
"n n
a
j=1 j k=1 (x + yk )
#n k=j = 0, x = x1 , . . . , xn ,
j=1 (x + yj )

and therefore
 

n
 
n

aj  (x + yk ) = 0, x = x1 , . . . , xn .
j=1 k=1
k=j

The above is a polynomial of degree at most n−1 that vanishes at n distinct

points and thus is identically zero. In addition the
 n

 n 

(x + yk )

 k=1 

k=j
j=1

are linearly independent polynomials since at the value −yr all but the rth
polynomial vanishes. Thus a1 = · · · = an = 0, which is a contradiction,
and A is nonsingular.
It remains to prove that each of these determinants is positive. By
continuity det A is of one sign for all choices of 0 < x1 < · · · < xn ,
0 < y1 < · · · < yn and a fixed n. For n = 1 it is positive by inspection. Let
us apply an induction argument.
Multiply the first row of
1 1 1
x1 +y1 x1 +y2 ··· x1 +yn
1 1 1
x2 +y1 x2 +y2 ··· x2 +yn
.. .. .. ..
. . . .
1 1 1
xn +y1 xn +y2 ··· xn +yn

by x1 + y1 to obtain
x1 +y1 x1 +y1
1 x1 +y2 ··· x1 +yn
1 1 1
x2 +y1 x2 +y2 ··· x2 +yn
.. .. .. .. .
. . . .
1 1 1
xn +y1 xn +y2 ··· xn +yn

Letting x1 , y1 → 0+ the above determinant converges to

1 1
x2 +y2 ··· x2 +yn
.. .. ..
. . .
1 1
xn +y2 ··· xn +yn

which, by induction, is positive.

94 Examples

4.4 Green’s matrices

We start not with a Green’s matrix but with a simpler related matrix.

Proposition 4.1 Let b1 , . . . , bn be n distinct numbers. Set

aij = bmin{i,j} , i, j = 1, . . . , n.

Then A = (aij )ni,j=1 is a totally positive matrix if and only if

0 ≤ b1 < · · · < bn .

Remark Note that we can always assume that the {bi } are distinct. For
if bj = bj+1 then the jth and (j + 1)st rows and columns are identical.

Proof Assume A is a totally positive matrix. Then bk ≥ 0, k = 1, . . . , n.

Now
 
i, i + 1 bi bi
0≤A = = bi (bi+1 − bi ).
i, i + 1 bi bi+1

Thus if bi > 0, then bi+1 > bi since the b1 , . . . , bn are distinct. If bi = 0,

then bi+1 > 0 = bi . Therefore

0 ≤ b1 < · · · < bn .

If b1 = 0 then the first row and the first column of A are identically zero
and we may simply disregard them. As such let us assume that

0 < b1 < · · · < bn .

Consider
 
i1 , . . . , ir
A
j1 , . . . , jr

for 1 ≤ i1 < · · · < ir ≤ n, 1 ≤ j1 < · · · < jr ≤ n. We first claim that if

i1 , j1 < i2 , j2 < · · · < ir , jr

does not hold, then

 
i1 , . . . , ir
A = 0.
j1 , . . . , jr
 4.4 Green’s matrices 95

To see this, let us assume, without loss of generality, that i
≥ j
+1 for
some  ∈ {1, . . . , r − 1}. As j1 < · · · < j
+1 ≤ i
< i
+1 < · · · < ir we have

ais jt = bjt

for s = , +1, . . . , r; t = 1, . . . , +1. This implies that the (r−+1)×(+1)

submatrix


i
, i
+1 , . . . , ir
A
j1 , j2 , . . . , j
+1
is of rank 1. Thus


i1 , . . . , ir
A
j1 , . . . , j
+1
is of rank at most , i.e., the  + 1 columns thereof are linearly dependent,
and therefore
 
i1 , . . . , ir
A = 0.
j1 , . . . , jr
Let us now assume that

i1 , j1 < i2 , j2 < · · · < ir , jr .

Set αk = min{ik , jk } and βk = max{ik , jk }, k = 1, . . . , r. By assumption

we have αk+1 > βk . We will prove that
  
i1 , . . . , ir
r−1

A = bα1 bαk+1 − bβk > 0.
j1 , . . . , jr
k=1

To this end, note the pattern of this minor, namely

bα1 bi1 b i1 ··· b i1
  bj1 bα2 b i2 ··· b i2
i1 , . . . , ir bj1 bj2 bα3 ··· b i3
A = .
j1 , . . . , jr .. .. .. .. ..
. . . . .
bj1 bj2 bj3 ··· bαr
Multiply the first row by bj1 /bα1 and subtract it from each of the other
rows to obtain
bα1 bi1 bi1 · · · bi1
0 cα2 ci2 · · · ci2
= 0 cj2 cα3 · · · ci3
.. .. .. .. ..
. . . . .
0 cj2 cj3 ··· cαr
96 Examples

where
bi1 bj1
ck = bk − .
bα1
For k > i1 , j1 , the {ck } is a strictly increasing sequence of positive numbers
and we can apply an induction hypothesis (the case r = 1 is trivial). Thus
  
i1 , . . . , ir
r−1

A = bα1 cα2 cαk+1 − cβk .
j1 , . . . , jr
k=2

As
bi1 bj1 bα bβ
ck = bk − = bk − 1 1 = bk − bβ1 ,
bα1 bα1
it therefore follows that
   
i1 , . . . , ir
r−1
 r−1

A = bα1 (bα2 − bβ1 ) bαk+1 − bβk = bα1 bαk+1 − bβk .
j1 , . . . , jr
k=2 k=1

Let us now assume that we are given two sequences c1 , . . . , cn and

d1 , . . . , dn of nonzero values. Set

aij = cmin{i,j} dmax{i,j} , i, j = 1, . . . , n.

The matrix A = (aij )ni,j=1 is called a Green’s matrix. We prove the following
result.

Theorem 4.2 The matrix A, as defined above, is totally positive if and

only if the c1 , . . . , cn and d1 , . . . , dn are all of one strict sign and
c1 cn
0< ≤ ··· ≤ .
d1 dn

Proof Assume A is totally positive, n ≥ 2. Then from the positivity of the

elements of A we must have that the c1 , . . . , cn and d1 , . . . , dn are either
all positive or all negative. (We are assuming they are all nonzero.) Let us
assume, without loss of generality, that they are all positive. Consider
   
i, i + 1 ci di ci di+1 ci+1 ci
0≤A = = ci di d2i+1 − .
i, i + 1 ci di+1 ci+1 di+1 di+1 di
Since the c1 , . . . , cn and d1 , . . . , dn are all positive we have
ci ci+1
≤ , i = 1, . . . , n − 1.
di di+1
 4.5 Jacobi matrices 97

To prove the converse direction note that since

ci ci+1
≤ , i = 1, . . . , n − 1,
di di+1
we have
 +
ci cj
aij = cmin{i,j} dmax{i,j} = min , di d j .
di dj
Let
ci
bi = .
di
Then 0 < b1 ≤ · · · ≤ bn . Setting
 n
B = bmin{i,j} i,j=1

we have
   
i1 , . . . , ir i1 , . . . , ir
A = di1 · · · dir dj1 · · · djr B .
j1 , . . . , jr j1 , . . . , jr
From Proposition 4.1 B is totally positive if 0 < b1 < · · · < bn . By
continuity B is totally positive for 0 ≤ b1 ≤ · · · ≤ bn . Thus A is totally
positive. In fact, from the previous analysis it follows that
 
i1 , . . . , ir
A =0
j1 , . . . , jr
unless
i1 , j1 < i2 , j2 < · · · < ir , jr

in which case we have

  r−1  
i1 , . . . , ir cα1  cαk+1 cβ
A = di1 · · · dir dj1 · · · djr − k
j1 , . . . , jr dα1 dαk+1 dβk
k=1

r−1
cαk+1 cβk
= cα1 dβr ,
dαk+1 dβk
k=1

where the αk and βk are as defined in Proposition 4.1.

4.5 Jacobi matrices

A Jacobi matrix is the term given to a square tridiagonal matrix, i.e., any
matrix A = (aij )ni,j=1 where aij = 0 if |i − j| ≥ 2. In other words, it is an
98 Examples

n × n matrix of the form

 
a1 b1 0 ··· 0 0
 c ··· 
 1 a2 b2 0 0 
 
 0 c2 a3 ··· 0 0 
A=  .. .. .. .. .. .. 

 . . . . . . 
 
 0 0 0 ··· an−1 bn−1 
0 0 0 ··· cn−1 an
where aii = ai , i = 1, . . . , n, and ai,i+1 = bi , ai+1,i = ci , i = 1, . . . , n − 1.
From this form it is readily verified that
 
i1 , . . . , ir
A =0
j1 , . . . , jr
if |ik − jk | ≥ 2 for any k ∈ {1, . . . , r}, and if |i
− j
| = 1 then
       
i1 , . . . , ir i1 , . . . , i
−1 i
i
+1 , . . . , ir
A =A A A .
j1 , . . . , jr j1 , . . . , j
−1 j
j
+1 , . . . , jr
From these formulæ we easily see how to calculate every minor of A.
Namely, if
i1 = j1 , . . . , ik1 = jk1 , ik1 +1 = jk1 +1 , . . . , ik2 = jk2 ,
ik2 +1 = jk2 +1 , . . . , ik3 = jk3 , . . .
then
 
i1 , . . . , ir
A
j1 , . . . , jr
       
i1 , . . . , ik1 ik1 +1 ik2 ik2 +1 , . . . , ik3
= A A ···A A ··· .
j1 , . . . , jk1 jk1 +1 jk2 jk2 +1 , . . . , jk3
Furthermore, when calculating a principal minor
 
i1 , . . . , ir
A
i1 , . . . , ir
we see that if ij + 1 < ij+1 for some j ∈ {1, . . . , r − 1}, then
     
i1 , . . . , ir i1 , . . . , ij ij+1 , . . . , ir
A =A A .
i1 , . . . , ir i1 , . . . , ij ij+1 , . . . , ir
Thus we have the following result.

Theorem 4.3 A Jacobi matrix A is totally positive if and only if all its
off-diagonal elements {bi }, {ci } and all its principal minors containing
consecutive rows and columns are nonnegative.
 4.5 Jacobi matrices 99

There are two additional facts worth noting concerning totally positive
Jacobi matrices. In our explanation thereof we will use the following formula
based on the Laplace expansion by minors. Namely, for a Jacobi matrix A
and any i ∈ {1, . . . , n − 1} we have
     
1, . . . , n 1, . . . , i i + 1, . . . , n
A = A A
1, . . . , n 1, . . . , i i + 1, . . . , n
   
1, . . . , i − 1 i + 2, . . . , n
−ai,i+1 ai+1,i A A . (4.1)
1, . . . , i − 1 i + 2, . . . , n
The first fact we will explain is that the inverse of a nonsingular
symmetric Jacobi matrix is a Green’s matrix. The reason for this is the
following. Assume A is an n × n symmetric Jacobi matrix. Let {ci } denote
the off-diagonal entries of A. Assume, for convenience, that each of the ci
is nonzero. Set B = A−1 . Then

i+j 1, . . . , j, . . . , n
(−1) A
1, . . . , i, . . . , n
bij =   .
1, . . . , n
A 1, . . . , n

Assuming i < j this reduces to

   
1, . . . , i − 1 j + 1, . . . , n
(−1)i+j A 1, . . . , i − 1 ci · · · cj−1 A j + 1, . . . , n
bij =   .
1, . . . , n
A 1, . . . , n

Similar formulæ hold in the case i = j and i > j. Set

 
1, . . . , i − 1
(−1)i A 1, . . . , i − 1 ci · · · cn−1
di =  
1, . . . , n
A 1, . . . , n

and
 
j + 1, . . . , n
A j + 1, . . . , n
ej = (−1)j .
cj · · · cn−1
Then it is readily verified that

bij = dmin{i,j} emax{i,j} .

If A is a nonsingular symmetric totally positive Jacobi matrix and B =

A−1 , then the matrix
 
(−1)i+j bij
100 Examples

is a nonsingular totally positive Green’s matrix, i.e., setting d!i = (−1)i di

and e!i = (−1)i ei we have that the d!1 , . . . , d!n and e!1 , . . . , e!n are all positive,

(−1)i+j bij = d!min{i,j} e!max{i,j} ,

and
d!i d!i+1
< , i = 1, . . . , n − 1.
e!i e!i+1

From the above definitions of the {d!i } and {! ei } this latter inequality is
equivalent to proving
   
1, . . . , i − 1 1, . . . , i
A 1, . . . , i − 1 A 1, . . . , i
  c2 <  , i = 1, . . . , n − 1,
i + 1, . . . , n i i + 2, . . . , n
A i + 1, . . . , n A i + 2, . . . , n

i.e.,
       
1, . . . , i i + 1, . . . , n 2 1, . . . , i − 1 i + 2, . . . , n
0<A A −ci A A .
1, . . . , i i + 1, . . . , n 1, . . . , i − 1 i + 2, . . . , n
As ci = ai,i+1 = ai+1,i , from (4.1) we see that the right-hand-side equals
 
1, . . . , n
A
1, . . . , n
which is strictly positive.
The other fact we wish to remark upon are necessary and sufficient
conditions for when A is a nonsingular totally positive Jacobi matrix.
From Theorem 4.3 we have that A is totally positive if and only if all
its off-diagonal elements {bi }, {ci } and all its principal minors containing
consecutive rows and columns are nonnegative. This latter condition can be
further simplified. Namely a nonsingular Jacobi matrix A is totally positive
if and only if all its off-diagonal elements {bi }, {ci } are nonnegative and all
its principal minors composed of initial consecutive rows and columns are
strictly positive. This fact follows from Theorem 4.3 and from the identity
(4.1) with arbitrary r ∈ {2, . . . , n} and s ∈ {1, . . . , r − 1}. Namely
     
1, . . . , r 1, . . . , s s + 1, . . . , r
A = A A
1, . . . , r 1, . . . , s s + 1, . . . , r
   
1, . . . , s − 1 s + 2, . . . , r
−A as,s+1 as+1,s A .
1, . . . , s − 1 s + 2, . . . , r
The details of the proof based on this formula are left to the reader.
 4.6 Hankel matrices 101

4.6 Hankel matrices

Assume we are given a sequence b0 , b1 , . . . , b2n . The matrix

A = (bi+j )ni,j=0

is called a Hankel matrix, i.e.,

 
b0 b1 ... bn
 b1 b2 . . . bn+1 
 
A= . .. .. .. .
 .. . . . 
bn bn+1 . . . b2n

The following characterizes strictly totally positive Hankel matrices.

Theorem 4.4 The above Hankel matrix A is strictly totally positive if and
only if

n 
n−1
bi+j xi xj , bi+j+1 xi xj
i,j=0 i,j=0

are both strictly positive definite forms, i.e., the symmetric matrices
 
b0 b1 ... bn
 b1 b2 . . . bn+1 
 
 . .. .. .. 
 . . . . . 
bn bn+1 ... b2n

and
 
b1 b2 ... bn
 b2 b3 ... bn+1 
 
 .. .. .. .. 
 . . . . 
bn bn+1 ... b2n−1

are strictly positive definite.

Proof Recall that one equivalent definition of strict positive definiteness for
symmetric matrices is the strict positivity of the initial principal minors.
Thus the necessity of the above conditions is obvious since every symmetric
strictly totally positive matrix must be strictly positive definite. It remains
to prove the sufficiency.
Let us therefore assume that the two matrices in the statement of
102 Examples
"2n i
the theorem are strictly positive definite. Let p(t) = i=0 ai t be any
(nontrivial) polynomial of degree at most 2n that is nonnegative on [0, ∞).
We claim that we must have
2n

ai bi > 0.
i=0

It is known that every polynomial of degree at most 2n that is nonnegative

on [0, ∞) can be written in the form
p(t) = (q(t))2 + t(r(t))2
where q and r are polynomials of degree at most n and n − 1, respectively.
Let

n 
n−1
q(t) = µj tj , r(t) = σ j tj .
j=0 j=0

Then the condition

2n

ai bi > 0
i=0

whenever p(t) is a nonnegative (nontrivial) polynomial on [0, ∞) is easily

shown to be equivalent to

n
µj µk bj+k > 0
j,k=0

and

n−1
σj σk bj+k+1 > 0
j,k=0

for every choice of (nontrivial) (µ0 , . . . , µn ) and (σ0 , . . . , σn−1 ). These latter
conditions are exactly the strict positive definiteness of our two matrices.
A minor digression is now in order. The study of moments was a central
motivating theme in the development of functional analysis. An important
precursor in this development is Stieltjes [1894–95]. In this paper Stieltjes
discussed the problem of necessary and sufficient conditions on a sequence
(bi )∞ ∞
i=0 so that the (bi )i=0 can be represented in the form
, ∞
bi = ti dα(t), i = 0, 1, 2, . . .
0

for some nonnegative Borel measure dα. (We have rephrased Stieltjes’
original problem as there were then no Borel measures. In fact the Stieltjes
 4.6 Hankel matrices 103

integral was introduced in that paper explicitly to deal with this problem.)
We now know that this is essentially equivalent (via the Hahn–Banach
Theorem) to asking for the existence of a continuous nonnegative linear
functional L on C[0, ∞) satisfying

L(ti ) = bi , i = 0, 1, 2, . . .

In the finite case things are somewhat simpler. We have

, ∞
bi = ti dα(t), i = 0, 1, . . . , 2n,
0

(and the (bi )2n

i=0 are interior points of the associated convex moment cone)
if and only if
2n

ai bi > 0
i=0
"2n i
whenever p(t) = i=0 ai t is a nonnegative (nontrivial) polynomial on
[0, ∞).
It also follows from the theory of moments that each such (b0 , . . . , b2n )
can be represented in the form


n+1
bi = λj ξji , i = 0, 1, . . . , 2n, (4.2)
j=1

where λj > 0, j = 1, . . . , n + 1, and 0 < ξ1 < · · · < ξn+1 .

Let us therefore assume that the {bi } are as given in (4.2). Then, as is
readily verified,
 
b0 . . . bn
 .. . . ..  =
 . . . 
bn . . . b2n
   
ξ10 0
. . . ξn+1 λ1 ... 0 ξ10 . . . ξ1n
 .. .. ..   .. ..  
..   ...
. .. ..  .
 . . .  . . . . 
0
ξ1n n
. . . ξn+1 0 . . . λn+1 ξn+1 n
. . . ξn+1

Set
 
ξ10 0
. . . ξn+1
 .. .. .. 
X= . . . 
ξ1n n
. . . ξn+1
104 Examples

and
 
λ1 ... 0
 .. .. ..  .
Λ= . . . 
0 . . . λn+1

Thus
A = XΛX T .

As X and X T are strictly totally positive matrices (X is a Vandermonde

matrix) and Λ is a diagonal nonsingular totally positive matrix, the result
follows. That is,
 
i1 , . . . , ir
A =
j1 , . . . , jr

    
i1 , . . . , ir j1 , . . . , jr
λ k1 · · · λ kr X X > 0.
k1 , . . . , kr k1 , . . . , kr
1≤k1 <···<kr ≤n+1

4.7 Toeplitz matrices

Assuming we are given a bi-infinite sequence

. . . , a−2 , a−1 , a0 , a1 , a2 , . . . ,

the associated Toeplitz matrix is defined by

A = (aj−i )∞
i,j=1 .

If we are given a one-sided infinite sequence

a0 , a1 , a2 , . . . ,

then we understand this to mean that a−n = 0, n = 1, 2, . . ., in the above

definition. Sequences that give rise to totally positive Toeplitz matrices
have been totally characterized in terms of their generating functions, i.e.,
"∞ "∞
representations of k=−∞ ak z k or k=0 ak z k .
In the latter case, with the normalization a0 = 1, the sequence

a0 , a1 , a2 , . . . ,

gives rise to a totally positive Toeplitz matrix

A = (aj−i )∞
i,j=1
 4.7 Toeplitz matrices 105
"∞
if and only if ak z k has the form
k=0
#∞
(1 + αi z)
e #i=1
γz
∞
i=1 (1 − βi z)
"∞
where γ ≥ 0, αi ≥ 0, βi ≥ 0, and i=1 (αi + βi ) < ∞. The proof of this
and the corresponding representation for the bi-infinite sequence will not be
presented here. However, we will prove the characterization in the simpler
case where the sequence has only a finite number of nonzero terms.

Theorem 4.5 Let a0 > 0, an = 0 and ak = 0 for k < 0 and k > n. Then

A = (aj−i )∞
i,j=1

is a totally positive matrix if and only if


n
p(x) = ak xk
k=0

has n negative zeros. Furthermore, if A is totally positive then

 
i1 , . . . , ir
A >0
j1 , . . . , jr
if and only if
ajk −ik > 0, k = 1, . . . , r,

i.e., 0 ≤ jk − ik ≤ n, k = 1, . . . , r.

Proof Let us first assume that p has n negative zeros. Then, assuming
a0 = 1, we have
n
p(x) = (1 + αr x)
r=1

where αr > 0 for r = 1, . . . , n. Let

(r)
Er = (bj−i )∞
i,j=1

(r) (r) (r)

where b0 = 1, b1 = αr and bk = 0 for k = 0, 1. It follows by inspection,
since αr > 0, that each Er is totally positive. Furthermore

n
A= Er .
r=1

This latter result follows from the fact that the generating function of a
106 Examples

product of Toeplitz matrices is the product of their generating functions.

As A is a product of totally positive matrices, then A is totally positive.
Let us now assume that A is totally positive. We will prove that

n
p(x) = ak xk
k=0

has n negative zeros. We prove this by contradiction using the variation

diminishing property of totally positive matrices. From the fact that A
is totally positive (thus ak ≥ 0) and a0 > 0, it follows that p has no
nonnegative real zeros. Let us assume that p has nonreal zeros. Since the
coefficients of p are real, if p(w) = 0 then p(w) = 0. Thus we may assume
that w is a zero of p satisfying
w = reiθ
where 0 < θ < π.
Now, as

n
ak wk+
= 0, ∈Z
k=0

we have

n
ak rk+
sin(k + )θ = 0,  ∈ Z.
k=0

Let x = (x0 , x1 , x2 , . . .) where

xk = rk sin kθ.
Thus
Ax = 0.
Let A[m] denote the m × (n + m) submatrix of A obtained by choosing the
first m rows and (n + m) columns of A, i.e.,
 
a0 · · · an 0
 .. .. 
A[m] =  . . .
0 a0 ··· an
Set
x[m] = (x0 , . . . , xn+m−1 ).
Thus
A[m] x[m] = 0 ∈ Rm .
 4.7 Toeplitz matrices 107

As n ≥ 1 and because
rank A[m] = m
there exists an x∗ ∈ Rn+m satisfying
A[m] x∗ = d
where
d = (1, −1, , . . . , (−1)m−1 ) ∈ Rm .
Thus for any ε
A[m] (x[m] + εx∗ ) = εd.
From the variation diminishing properties of the totally positive matrix
A[m] (Theorem 3.4) we have
m − 1 = S − (εd) ≤ S − (x[m] + εx∗ )
for any ε = 0.
Take ε∗ =  0 and sufficiently small such that
(x[m] + εx∗ )k (x[m] )k > 0
if (x[m] )k = xk = 0. Note that if xk = 0, 0 < k < n + m − 1, then since
0 < θ < π we have
xk−1 xk+1 < 0.
Thus
S − (x[m] + εx∗ ) ≤ S − (x[m] ) + 2
(because we do have x0 = 0 and can have xn+m−1 = 0). Therefore
m − 1 ≤ S − (x[m] ) + 2,
i.e.,
m − 3 ≤ S − (x[m] )
for all m.
Now, as is easily seen,
(n + m)θ
S − (x[m] ) ≤
π
simply because sin α > 0 for 0 < α < π and sin α < 0 for π < α < 2π, and
the fact that when going from xk to xk+1 we alter the “argument” of xk
by θ. Thus
(n + m)θ
m−3≤
π
108 Examples

for every m. But

θ
< 1,
π
which implies that the above cannot hold for m sufficiently large. We have
arrived at a contradiction.
The last claim of the theorem (namely that A is necessarily an almost
strictly totally positive matrix) is fairly elementary. One direction is a
consequence of Theorem 1.13, i.e.,
 
i1 , . . . , ir
A >0
j1 , . . . , jr
implies that the diagonal elements of this minor are strictly positive. The
converse direction is a simple consequence of Theorem 1.19 and the fact
that
 
i + 1, . . . , i + r
A = ar0 > 0
i + 1, . . . , i + r
and
 
i + 1, . . . , i + r
A = arn > 0.
i + 1 + n, . . . , i + r + n

The above characterization does not necessarily mean that it is a simple

matter to determine whether a given Toeplitz matrix is totally positive
or not. It is simply equivalent to the fact that the associated generating
polynomial has all negative zeros. We return to this problem in Section 4.9.
Let us consider a few specific examples of totally positive Toeplitz
matrices and totally positive matrices derived from Toeplitz matrices.

1. Consider the polynomial

p(x) = (1 + x)n .

As its zeros are all negative the Toeplitz matrix with

 
n
ak = , k = 0, 1, . . . , n
k
is totally positive.
2. In Theorem 1.11 we proved that if the square matrix A is (strictly) totally
positive, then the matrix B defined by

b1j = a1j , for all j,

 4.7 Toeplitz matrices 109

and for i ≥ 2

bij = bi−1,k akj , for all j,
k

is also (strictly) totally positive. Applying this result to the above

Toeplitz matrix where n = 1, i.e., a0 = a1 = 1, am = 0 if m = 0, 1,
we get
b11 = b12 = 1

and for i ≥ 2

bi1 = 1, bij = bi−1,j−1 + bi−1,j , j ≥ 2,

which immediately implies that

 
i
bij = , i, j = 1, 2, . . . .
j−1
n
Letting m = 0 for n < m we have that the matrix of binomial
coefficients
 ∞
i
C=
j i,j=0

is totally positive. This is equivalent to the fact that the infinite Toeplitz
matrix
 ∞
1
D=
(i − j)! i,j=0
is totally positive.
3. Applying Theorem 1.11 to the infinite totally positive Toeplitz matrix
A = (aij )∞i,j=0 , where

1, i ≤ j
aij =
0, i > j
it easily follows that the matrix
 ∞
i+j
B=
j i,j=0

is totally positive, which is equivalent to the fact that the Hankel matrix
∞
D = ((i + j)!)i,j=0

is totally positive. The total positivity of D can also be proven by other

methods.
110 Examples

4. Consider the polynomial


n
p(x) = (1 + q i−1 x)
i=1

for some q > 0, q = 1. These polynomials are called Gaussian

polynomials. Set

n
p(x) = ak xk
k=0

and let us calculate the coefficients ak .

From the fact that

(1 + q n x)p(x) = (1 + x)p(qx)

we have

n 
n
(1 + q n x) ak xk = (1 + x) ak q k xk .
k=0 k=0

k
Equating the coefficients of x we obtain for k = 1, . . . , n

ak + q n ak−1 = ak q k + ak−1 q k−1

implying
(1 − q n−k+1 )
ak = ak−1 q k−1 .
(1 − q k )
Now a0 = 1 so that
(1 − q n−k+1 ) · · · (1 − q n )
ak = q (k−1)+(k−2)+···+1 ,
(1 − q k ) · · · (1 − q)
which we write as
-n.
ak = q k(k−1)/2 .
k q

The
-n. (1 − q n−k+1 ) · · · (1 − q n )
=
k q (1 − q k ) · · · (1 − q)
are termed q-binomial coefficients.
Thus the Toeplitz matrix with
-n.
ak = q k(k−1)/2 , k = 0, 1, . . . , n,
k q
is totally positive.
 4.8 Generalized Hurwitz matrices 111

4.8 Generalized Hurwitz matrices

Assume we are given real numbers a0 , a1 , . . . , an where n ≥ M ≥ 2. For
ease of exposition, set aj = 0 for j < 0 and j > n. Let us define

P = (pij )∞
i,j=1

where
pij = aM j−i , i, j = 1, 2, . . . .

Thus
 
aM −1 a2M −1 a3M −1 ···
 aM −2 a2M −2 a3M −2 ··· 
 
 .. .. .. 
 . . . 
 
 ··· 
 a0 aM a2M 
 
 0 aM −1 a2M −1 ··· 
P =
 0 aM −2 a2M −2

··· .
 
 .. .. .. 
 . . . 
 
 0 a0 aM ··· 
 
 0 0 aM −1 ··· 
 
.. .. ..
. . .
The matrix P is not a Toeplitz matrix, but it resembles a Toeplitz matrix.
It also contains M Toeplitz submatrices. Simply take all columns and every
M th row.
Somewhat surprisingly we prove that there are a finite number of
determinantal conditions that imply that this matrix is totally positive.

Theorem 4.6 Assume a0 > 0 and

 
k, k + 1, . . . , k + r − 1
P >0 (4.3)
1, 2, . . . , r
for k = 1, . . . , M − 1, and r = 1, . . . , n+k−1 M −1 . Then P is totally positive,
ak > 0, k = 0, 1, . . . , n, and
 
i1 , . . . , ir
P >0
j1 , . . . , jr
if and only if
pik jk = aM jk −ik > 0, k = 1, . . . , r,

i.e., 0 ≤ M jk − ik ≤ n.
112 Examples
n+k−1
Before we prove this theorem note that for r > M −1 the last column
of


k, k + 1, . . . , k + r − 1
P
1, 2, . . . , r

vanishes identically. That is, the conditions of (4.3), together with a0 > 0,
are exactly the demand that all possible nonvanishing minors composed
from consecutive rows and initial consecutive columns be strictly positive.
In addition, let k ∈ {1, . . . , M − 1} be such that
n+k−1
r=
M −1
is an integer. Such a k exists and r ≥ 2. For this k and r
   
k, . . . , k + r − 1 k, . . . , k + r − 2
P = an P .
1, . . . , r 1, . . . , r − 1

Thus, from (4.3), it immediately follows that an > 0.

Our proof is via induction on n. In this proof we use the following result.

Lemma 4.7 Let

bM j+k−1 = aM j+k , k = 1, . . . , M − 1, j = 0, 1, . . . ,

and
a0
bM j−1 = aM j − aM j+1 , j = 1, 2, . . . ,
a1
where applicable, to define b0 , . . . , bn−1 . Set bk = 0 for k < 0 and k > n − 1.
Define
Q = (qij )∞
i,j=1

where
qij = bM j−i , i, j = 1, 2, . . . .

Then Q satisfies the conditions of Theorem 4.6 (with n − 1 replacing n).

Proof Note that b0 = a1 = pM −1,1 > 0 and by the above

a0
b−1 = a0 − a1 = 0.
a1
Thus bj = 0 for j < 0 and, by definition, bj = 0 for j > n − 1.
 4.8 Generalized Hurwitz matrices 113

Recall that

 
aM −1 a2M −1 a3M −1 ···
 aM −2 a2M −2 a3M −2 ··· 
 
 .. .. .. 
 . . . 
 
 ··· 
 a0 aM a2M 
 
 0 aM −1 a2M −1 ··· 
P =
 0 aM −2 a2M −2

··· 
 
 .. .. .. 
 . . . 
 
 0 a0 aM ··· 
 
 0 0 aM −1 ··· 
 
.. .. ..
. . .

and Q has exactly this same form, i.e.,

 
bM −1 b2M −1 b3M −1 ···
 bM −2 b2M −2 b3M −2 ··· 
 
 .. .. .. 
 . . . 
 
 ··· 
 b0 bM b2M 
 
 0 bM −1 b2M −1 ··· 
Q=
 0 bM −2 b2M −2

··· .
 
 .. .. .. 
 . . . 
 
 0 b0 bM ··· 
 
 0 0 bM −1 ··· 
 
.. .. ..
. . .

For k = 2, . . . , M − 1, it is easily verified that

   
k, k + 1, . . . , k + r − 1 k − 1, k, . . . , k + r − 2
Q =P >0
1, 2, . . . , r 1, 2, . . . , r

n+(k−1)−1 (n−1)+k−1
for r = 1, . . . , M −1 = M −1 . We must also consider

 
1, 2, . . . , r
Q .
1, 2, . . . , r
114 Examples
n+M −2
Now, for 2 ≤ r ≤ M −1 ,

a1 aM +1 a2m+1 ···
  a0 aM a2M ···
M − 1, M, . . . , M + r − 2
0 < P = 0 aM −1 a2M −1 ···
1, 2, . . . , r
.. .. .. ..
. . . .
aM a2M ··· aM +1 a2M +1 ···
= a1 aM −1 a2M −1 ··· − a0 aM −1 a2M −1 ···
.. .. .. .. .. ..
. . . . . .
aM − aa01 aM +1 a2M − aa01 a2M +1 ···
= a1 aM −1 a2M −1 ···
.. .. ..
. . .
 
1, 2, . . . , r − 1
= a1 Q .
1, 2, . . . , r − 1
As a1 > 0 we have
 
1, 2, . . . , r − 1
Q >0
1, 2, . . . , r − 1
for 1 ≤ r − 1 ≤ n−1
M −1 , which is exactly what we wished to prove.
We now proceed to the proof of our theorem.

Proof of Theorem 4.6 We start with the case n = M . Here

 
aM −1 0 ···
 aM −2 0 ··· 
 
 . .. 
 .. . 
 
 a ··· 
 1 0 
 
 a0 aM ··· 
P = 0 aM −1

··· .
 
 .. .. .. 
 . . . 
 
 0 a ··· 
 0 
 0 0 ··· 
 
.. ..
. .
From (4.3) we have pk1 = aM −k > 0, k = 1, . . . , M − 1, and
 
M − 1, M a1 0
0<P = = a1 aM .
1, 2 a0 aM
 4.8 Generalized Hurwitz matrices 115

Thus a1 , a1 , . . . , aM > 0. The other claims follow easily from this fact and
the geometric form of P .
We now apply induction assuming the result is valid for n − 1, where
n > M . We start with the a0 , . . . , an and P as in the statement of Theorem
4.6. From Lemma 4.7, the b0 , . . . , bn−1 and the Q therein satisfy (4.3) for
n − 1, and therefore Q is totally positive, bk > 0, k = 0, . . . , n − 1, and
 
i1 , . . . , ir
Q >0
j1 , . . . , jr
if and only if
qik jk = bM jk −ik > 0, k = 1, . . . , r,

i.e., 0 ≤ M jk − ik ≤ n − 1.
From Lemma 4.7 we have

aM j+k = bM j+k−1 , k = 1, . . . , M − 1
aM j = bM j−1 + cbM j , k = 0, (4.4)

where c > 0. (Specifically c = a0 /a1 and we are given a0 > 0 and

a1 = pM −1,1 > 0.) Thus P is obtained from Q by an addition of a
positive multiple of one row to a succeeding row and shifting (renumbering)
of the rows. Both these operations (see Proposition 1.5) preserve totally
positive. Thus P is totally positive. The fact that a0 , a1 , . . . , an > 0 also
follows easily. We are given a0 > 0, and the positivity of the other ak is a
consequence of the positivity of b1 , b1 , . . . , bn−1 and (4.4).
For any nonsingular totally positive matrix the diagonal elements are
positive (see Theorem 1.13). That is, if
 
i1 , . . . , ir
P >0
j1 , . . . , jr
then
pik jk = aM jk −ik > 0, k = 1, . . . , r,

i.e., 0 ≤ M jk − ik ≤ n. It remains to prove the converse. As such let us

assume that
0 ≤ M jk − ik ≤ n, k = 1, . . . , r.

We wish to prove that

 
i1 , . . . , ir
P > 0.
j1 , . . . , jr
116 Examples

Based on (4.4) we can write

pis jt = aM jt −is = bM jt −is −1 + c(is )bM jt −is
where

0, is is not a multiple of M
c(is ) =
c, is is a multiple of M .
Now
 
i1 , . . . , ir
P = det (pis jt ) = det (bM jt −is −1 + c(is )bM jt −is ).
j1 , . . . , jr
Thus the desired minor of P is the nonnegative sum of various minors of
Q. All the minors of Q are nonnegative. Our task is to show that at least
one of the minors in this sum is positive, and has positive coefficient. This
we do by defining
is = is + 1 if 0 < M js − is ≤ n
and
is = is if M js − is = 0.
(Note that in the latter case c(is ) = c(is ) = c > 0.)
As
0 ≤ M jk − ik ≤ n, k = 1, . . . , r,
then from the above construction
0 ≤ M jk − ik ≤ n − 1, k = 1, . . . , r.
Furthermore
i1 < · · · < ir .
To see this note that if is = is + 1 then obviously is−1 < is , while if is = is
then M js = is and
is−1 ≤ M js−1 < M js = is .
Therefore  
i1 , . . . , ir
Q > 0.
j1 , . . . , jr
This minor, with positive coefficient, is one of the minors involved in
calculating the desired minor of P . Thus
 
i1 , . . . , ir
P > 0,
j1 , . . . , jr
which proves the theorem.
 4.9 More on Toeplitz matrices 117

Remark The matrix

 
a1 a3 a5 ···
 ··· 
 a0 a2 a4 
 
 0 a1 a3 ··· 
P =
 0 a0 a2 ···


 
 0 0 a1 ··· 
 
.. .. .. ..
. . . .

is called a Hurwitz matrix. This is the matrix of Theorem 4.6 with M = 2.

The conditions (4.3) of Theorem 4.6 in this case reduce to
 
1, . . . , r
P > 0, r = 1, . . . , n.
1, . . . , r
These inequalities, together with a0 > 0, are equivalent to the fact that the
polynomial
p(x) = a0 xn + a1 xn−1 + · · · + an

is a Hurwitz polynomial, i.e., has all its zeros in the open left-hand plane
Re (z) < 0. If p has all its zeros in the closed left-hand plane Re (z) ≤ 0,
then P is totally positive. Unfortunately the converse is not true.

4.9 More on Toeplitz matrices

In Section 4.7 we considered infinite Toeplitz matrices with a finite number
of nonzero terms and proved that the matrix is totally positive if and only
if its associated generating polynomial has only negative zeros. In other
words, we replaced one problem by another one which, while providing
an insight into the theory, does not always constitute a useful criterion
for determining whether a given matrix is in fact totally positive. In this
short section we apply the theory of Hurwitz matrices in order to generate
verifiable conditions for when such a matrix is totally positive, i.e., when a
polynomial has only negative zeros.
We present most of the results here without proof. Their proofs would
take us too far afield. References can be found in the last section of this
chapter.
Let
n
p(x) = ak xn−k
k=0
118 Examples

with a0 > 0, and set

q(x) = p(x2 ) + xp (x2 ).
Then we have

Theorem 4.8 The polynomial q is a Hurwitz polynomial if and only if p

has simple negative zeros.
Let
2n

q(x) = bk x2n−k .
k=0

Then
b2k = ak , k = 0, 1, . . . , n,
and
b2k+1 = (n − k)ak , k = 1, . . . , n − 1.
Let A be the infinite Toeplitz matrix based on p, i.e.,
 
a0 a1 a2 · · · an 0 0 ···
 0 a0 a1 · · · an−1 a 0 ··· 
 n 
A =  0 0 a ··· a a a ··· 
 0 n−2 n−1 n 
.. .. .. . . .. .. .. ..
. . . . . . . .
and B be the infinite Hurwitz matrix based on q, i.e.,
   
b1 b3 b5 · · · na0 (n − 1)a1 (n − 2)a2 ···
 b b2 b4 · · ·   a ··· 
 0   0 a1 a2 
   
 0 b1 b3 · · ·   0 na0 (n − 1)a1 ··· 
B=  0
 
b0 b2 · · ·  =  0 a0 a1 ···
.

   
 0 0 b1 · · ·   0 0 na0 ··· 
   
.. .. .. . . .. .. .. ..
. . . . . . . .

Theorem 4.9 The following are equivalent:

(i) A is totally positive.
(ii) B is totally positive.
(iii) p has n negative zeros.

Proof From Theorem 4.5, conditions (i) and (iii) are equivalent.
Furthermore, if B is totally positive, then since A is a submatrix of B
it follows that A is totally positive.
 4.10 Hadamard products of totally positive matrices 119

Assume p has n negative zeros. Perturb these zeros so that they are
simple and negative. Then from Theorem 4.8 q is a Hurwitz polynomial
and thus the associated Hurwitz matrix B is totally positive. Perturb back.

From the results of Section 4.8, b0 = a0 > 0 and

 
1, . . . , k
B > 0, k = 1, . . . , 2n,
1, . . . , k
is equivalent to the fact that q is a Hurwitz polynomial. Thus we have the
following.

Theorem 4.10 The polynomial


n
p(x) = ak xn−k
k=0

with a0 > 0 has n simple negative zeros if and only if

 
1, . . . , k
B > 0, k = 1, . . . , 2n.
1, . . . , k
When considering the total positivity of the associated Toeplitz matrix,
the necessity of
ak > 0, k = 0, 1, . . . , n,
is obvious. With this a priori assumption it suffices to verify the positivity
of half of the above determinants.

Proposition 4.11 Let p be as above and assume that ak > 0, k =

0, 1, . . . , n. Then p has n simple negative zeros if and only if
 
1, . . . , 2k + 1
B > 0, k = 1, . . . , n − 1.
1, . . . , 2k + 1
When p has multiple zeros then q has zeros on the imaginary axis and
the situation is more complicated.

4.10 Hadamard products of totally positive matrices

Given n × m matrices A = (aij ) and B = (bij ) we define their Hadamard
product as the n × m matrix formed by the entrywise product of the
elements of A and B, i.e., it is the n × m matrix C = (cij ) where
cij = aij bij .
120 Examples

We use the notation

C = A ◦ B.

While certain classes of matrices are closed under the operation of

Hadamard products, the class of totally positive matrices is not one of
them. For example, the matrix
 
1 1 1
A=  1 1 1 
0 1 1

is totally positive, while

 
1 1 0
A ◦ AT =  1 1 1 
0 1 1

has determinant equal to −1. (As strictly totally positive matrices are
dense in the class of totally positive matrices, it follows that there also exist
strictly totally positive matrices whose Hadamard product is not totally
positive.)
However, there are numerous subclasses of totally positive matrices that
are closed under the operation of taking Hadamard products. In this section
we detail some of them.

1. Consider a strictly totally positive matrix satisfying the criteria in

Section 2.6. That is, let A = (aij ) be an n × n matrix, all of whose entries
are strictly positive and where
 
2 π
aij ai+1,j+1 > 4 cos ai,j+1 ai+1,j
n+1

for all i, j = 1, . . . , n − 1. Now let B = (bij ) be any n × n matrix, all of

whose entries are strictly positive, and such that
 
i, i + 1
B ≥0
j, j + 1
for all i, j = 1, . . . , n − 1. Then from Theorem 2.16 the Hadamard product

A ◦ B = (aij bij )

is strictly totally positive.

 4.10 Hadamard products of totally positive matrices 121

2. Green’s matrices. If A and B are totally positive Green’s matrices,

then A◦B is also a totally positive Green’s matrix. This immediately follows
from the definition of a Green’s matrix and the characterization of Green
matrices as given in Theorem 4.2.

3. Jacobi matrices. We recall that a Jacobi matrix is a square tridiagonal

matrix, i.e., a matrix A = (aij )ni,j=1 where aij = 0 if |i − j| ≥ 2.
Furthermore, from Theorem 4.3, a Jacobi matrix is totally positive if and
only if all its off-diagonal elements and all its principal minors containing
consecutive rows and columns are nonnegative. We prove the following.

Proposition 4.12 The Hadamard product of two totally positive Jacobi

matrices is a totally positive Jacobi matrix.

Proof Assume we are given the totally positive Jacobi matrix

 
a11 a12 0 · · · 0
 a21 a22 a23 · · · 0 
 
 a32 a33 · · · 0 
A= 0 .
 . .. .. .. .. 
 .. . . . . 
0 0 0 · · · ann

Let c ≥ d ≥ 0, c > 0. Then, assuming a11 > 0, the matrix

 
ca11 a12 0 ··· 0
 0 a22 − daca
12 a21
· · · 
 a23 0 
 11

Ac,d =  0 a32 a33 · · · 0 
 . 
 . .. .. .. .. 
 . . . . . 
0 0 0 · · · ann

is also a totally positive Jacobi matrix. To see this let

 
ca11 a12 0 ··· 0
 da21 a22 a23 ··· 0 
 
A!c,d = 
 0 a32 a33 ··· 0 
.
 . .. .. .. .. 
 .. . . . . 
0 0 0 ··· ann

The matrix A !c,d is totally positive since it is obtained from A by multiplying

the first column of A by d and adding (c − d)a11 to the (1, 1) entry. Both
122 Examples

operations preserve total positivity (see Proposition 1.5). Now consider

Ac,d . Its off-diagonal elements are nonnegative and
   
i, . . . , j ! i, . . . , j
Ac,d = Ac,d
i, . . . , j i, . . . , j
except when i = 2, in which case
   
2, . . . , j 1 ! 1, . . . , j
Ac,d = Ac,d ≥ 0.
2, . . . , j ca11 1, . . . , j
Thus Ac,d is a totally positive Jacobi matrix.
Let us now assume that A and B are n × n totally positive Jacobi
matrices. Our proof is by induction on n. The result is easily verified for
n = 1 (and n = 2). Assume the result holds for r × r matrices, r ≤ n − 1.
Now
 
a11 b11 a12 b12 0 ··· 0
 a21 b21 a22 b22 a23 b23 · · · 0 
 
 0 a32 b32 a33 b33 · · · 0 
A◦B = .
 .. .. .. .. .. 
 . . . . . 
0 0 0 ··· ann bnn
We wish to prove that
 
1, . . . , n
(A ◦ B) ≥0
1, . . . , n
to advance the induction. If a11 = 0 then a12 = 0 or a21 = 0, and the above
determinant is therefore zero. The similar result holds if b11 = 0. We may
therefore assume that a11 , b11 > 0. Multiply the first row of A ◦ B by aa21 b21
11 b11
and subtract it from the second row to obtain
 
a11 b11 a12 b12 0 ··· 0
 a22 b22 − a12ab12 a21 b21
a23 b23 · · · 
 0 0 
 11 b11

D=  0 a32 b 32 a33 b 33 · · · 0 .

 .. .. .. .. .. 
 . . . . . 
0 0 0 ··· ann bnn
Therefore
   
1, . . . , n 2, . . . , n
(A ◦ B) = a11 b11 D .
1, . . . , n 2, . . . , n
We now apply the induction hypothesis to the (n−1)×(n−1) submatrix
of D obtained by deleting its first row and column. This matrix is the
 4.10 Hadamard products of totally positive matrices 123

Hadamard product of the two (n − 1) × (n − 1) totally positive Jacobi

matrices



2, . . . , n 2, . . . , n
Ab11 b22 ,b12 b21 and B .
2, . . . , n 2, . . . , n
The former matrix is totally positive by our previous analysis and since
b11 b22 ≥ b12 b21 . This proves the proposition.

Remark The same proof shows that the Hadamard product of a totally
positive Jacobi matrix and an arbitrary totally positive matrix is a totally
positive Jacobi matrix.

4. Hankel matrices. We are given Hankel matrices A = (ai+j )ni,j=0 and

B = (bi+j )ni,j=0 . If A and B are strictly totally positive, then A ◦ B is
strictly totally positive. This follows from the fact that

A ◦ B = (ai+j bi+j )ni,j=0

is also a Hankel matrix, the characterization of strictly totally positive

Hankel matrices as detailed in Theorem 4.4 and the Schur Product Theorem
which states that if C and D are two strictly positive definite matrices, then
so is their Hadamard product C ◦ D.

5. Toeplitz matrices. In Theorem 4.5 we considered Toeplitz matrices

of the form
A = (aj−i )∞
i,j=1

where a0 > 0, an = 0 and ak = 0 for k < 0 and k > n, and showed that A
is totally positive if and only if

n
p(x) = ak xk
k=0

has n negative zeros. A special case of Maló’s Theorem states that if


n
p(x) = ak xk
k=0

has n negative zeros and


m
q(x) = bk xk
k=0
124 Examples

has m negative zeros, then the polynomial


r
ak bk xk
k=0

has r negative zeros, where r = min{n, m}. Thus the Hadamard product of
two totally positive Toeplitz matrices of the above form is again a totally
positive Toeplitz matrix.

6. Hurwitz matrices. As stated in the remark at the end of Section 4.8,

the polynomial
n
p(x) = ak xk
k=0

is a Hurwitz polynomial if and only if all its zeros lie in the open left-hand
plane Re (z) < 0. Assuming a0 > 0, we have that p is a Hurwitz polynomial
if and only if the matrix
 
a1 a3 a5 · · ·
 a a a ··· 
 0 2 4 
 
 0 a1 a3 · · · 
P =  0 a a ··· 


0 2
 
 0 0 a1 · · · 
 
.. .. .. . .
. . . .

satisfies
 
1, . . . , r
P > 0, r = 1, . . . , n,
1, . . . , r
which implies that P is also totally positive (see Theorem 4.6). It is known
that if
n
p(x) = ak xk
k=0

is a Hurwitz polynomial and


m
q(x) = bk xk
k=0

is a Hurwitz polynomial, then the polynomial


r
ak bk xk
k=0
 4.11 Remarks 125

is also a Hurwitz polynomial, where r = min{n, m}. Thus the Hadamard

product of two totally positive Hurwitz matrices of the above form is again
a totally positive Hurwitz matrix of the appropriate form.

4.11 Remarks
The main reference books on totally positive kernels and Chebyshev
systems are Gantmacher, Krein [1950], Karlin [1968], Karlin, Studden
[1966] and Krein, Nudel’man [1977]. The examples of Sections 4.2 through
4.6 can all be found, in more or less detail, in Gantmacher, Krein [1937].
The examples of Sections 4.2 and 4.3 can also be found in Pólya, Szegő
[1976] (originally published in 1925). Karlin [1968] contains many, many
additional examples of totally positive and strictly totally positive kernels
and matrices. See also Carlson, Gustafson [1983] for some further examples.
For a discussion of the theory of moments as presented in the
proof of Theorem 4.4, see Karlin, Studden [1968], Chap. V, Krein,
Nudel’man [1977], Chap. V and Shohat, Tamarkin [1943]. The form of
the representation of the generating functions for totally positive infinite
Toeplitz matrices was conjectured by Schoenberg in 1951. It was partially
solved in Aissen, Schoenberg, Whitney [1952]. The final proof of the
representations is to be found in Edrei [1952] and Edrei [1953]. The
sequences in the totally positive Toeplitz matrices are also called Pólya
Frequency Sequences; see Karlin [1968], Chap. 8, and references therein. For
many more examples of sequences satisfying Theorem 4.5, see e.g., Brenti
[1989], Brenti [1995], Pitman [1997], and Wang, Yeh [2005]. Theorem 4.6
is from Goodman, Sun [2004]. The fact that in the special case M = 2
(Hurwitz polynomials) the initial minors being strictly positive implies the
total positivity of the full matrix was first proved by Asner [1970] and then
reproved in a more transparent form by Kemperman [1982]. An example
of a polynomial p whose zeros are not all in the closed left-hand plane
Re (z) ≤ 0, but where the associated Hurwitz matrix P is totally positive
(but without the strict positivity of the appropriate principal minors),
can be found in Asner [1970]. The proof of Theorem 4.6 follows the lines
of Kemperman’s proof. See also Holtz [2003] and references therein for
another approach to the Hurwitz polynomials and the total positivity of the
associated matrix. Material on Hurwitz matrices may be found in various
texts, e.g., Gantmacher [1953], Marden [1966], and Rahman, Schmeisser
[2002]. Theorem 4.8 is a special case of a more general result; see e.g.,
Gantmacher [1953], Theorem 13, Chap. XV of the English translation.
126 Examples

Theorem 4.10 and Proposition 4.11 can be found, for example, in Rahman,
Schmeisser [2002], Cor. 10.6.13 (the notation is somewhat different).
Applying the criteria of Theorem 2.16 we obtain relatively simple
sufficient conditions for when certain of the matrices considered in this
chapter are totally positive. For example, the Hankel matrix
A = (bi+j )ni,j=0
is strictly totally positive if bk > 0, k = 0, 1, . . ., and
 
2 π
bk−1 bk+1 > 4 cos b2 , k = 1, . . . , 2n − 1,
n+2 k
while the Toeplitz matrix
A = (aj−i )∞
i,j=1

with ak = 0 for k < 0 and k > n is totally positive if ak > 0, k = 0, . . . , n,

and
a2k ≥ 4ak−1 ak+1 , k = 1, . . . , n − 1.
Thus this condition is sufficient to imply that the polynomial

n
p(x) = ak xk
k=0

has all real negative zeros. This was already proved in Kurtz [1992].
For a general discussion and survey of Hadamard products see Horn,
Johnson [1991], Chap. 5, and the many references therein. It seems that
Markham [1970] was the first to consider Hadamard products of totally
positive matrices. He showed that the class of totally positive matrices
is not closed under the operation of Hadamard products and proved
Proposition 4.12. The Schur Product Theorem concerning the Hadamard
product of positive definite matrices can be found, for example, in Horn,
Johnson [1991], p. 309. Maló’s Theorem is from Maló [1895]. In Wagner
[1992] the result concerning the Hadamard product of Toeplitz matrices
is extended to a subclass of the totally positive Toeplitz matrices with an
infinite number of nonzero coefficients. The closure of Hurwitz matrices
under the Hadamard product is in Garloff, Wagner [1996a]. Its proof is far
beyond the scope of this monograph. Some additional results can be found
in Garloff, Wagner [1996b] and Crans, Fallat, Johnson [2001].
 5
Eigenvalues and eigenvectors

In this chapter we review the spectral properties of totally positive matrices.

A strictly totally positive matrix has positive, simple eigenvalues and the
associated eigenvectors possess an intricate structure. Such is not the
case for totally positive matrices. However, there is an intermediate set
of matrices with the same spectral properties as strictly totally positive
matrices. These matrices are called oscillation matrices. They shall be
discussed in Section 5.1. In Section 5.2 we present the Gantmacher–Krein
Theorem (Theorem 5.3) and give two quite different proofs thereof. This
theorem contains the main spectral properties of oscillation matrices. In
Section 5.3 we consider eigenvalues of the principal submatrices of such
matrices and study their behaviour. We study in more detail the properties
of eigenvectors of oscillation matrices in Section 5.4. Finally, in Section 5.5,
we look at how the eigenvalues of oscillation matrices vary as functions of
the elements of the matrix.

5.1 Oscillation matrices

Oscillation matrices are a class of matrices intermediary between totally
positive and strictly totally positive matrices. They share the eigenvalue
and eigenvector structure of strictly totally positive matrices.

Definition 5.1 An n × n matrix A is said to be an oscillation matrix if A

is totally positive and some power of A is strictly totally positive.

Importantly, there are relatively simple criteria for determining if a

totally positive matrix is an oscillation matrix.

127
128 Eigenvalues and eigenvectors

Theorem 5.2 An n × n matrix A = (aij )ni,j=1 is an oscillation matrix

if and only if A is totally positive, nonsingular, and ai,i+1 , ai+1,i > 0,
i = 1, . . . , n − 1. Furthermore, if A is an oscillation matrix, then An−1 is
strictly totally positive.

Proof The necessity of the above conditions is rather simple. If A is an

oscillation matrix then A is totally positive. If A is singular then every
power of A is singular and thus no power of A can be strictly totally
positive. Assume A is totally positive and nonsingular and ai,i+1 = 0 for
some i. Then from Theorem 1.19 we have ars = 0 for all r ≤ i and s ≥ i+1.
As is readily verified, this implies that the (r, s) elements of every power of
A also vanish for all r ≤ i and s ≥ i + 1. This proves the necessity.
The sufficiency is less elementary. We first prove that
 
i1 , . . . , ir
A >0
j1 , . . . , jr
for all 1 ≤ i1 < · · · < ir ≤ n and 1 ≤ j1 < · · · < jr ≤ n satisfying |ik −jk | ≤
1, k = 1, . . . , r, and max{ik , jk } < min{ik+1 , jk+1 }, k = 1, . . . , r − 1. In
this proof we make use of the fact that since A is totally positive and
nonsingular, every principal minor of A is strictly positive (see Theorem
1.13). Our proof of this claim is via induction on r. The result holds for
r = 1 since all of the elements of A on the main and first off-diagonals are
strictly positive. Assume the result holds for r − 1, but
 
i1 , . . . , ir
A =0
j1 , . . . , jr
for some i1 , . . . , ir and j1 , . . . , jr , as above. Then we must have ik = jk for
some k. From the induction hypothesis we have
   
i1 , . . . , ir−1 i2 , . . . , ir
A ,A > 0.
j1 , . . . , jr−1 j2 , . . . , jr
Thus, from Proposition 1.17, the totally positive matrix


i1 , i1 + 1, . . . , ir
B=A
j1 , j1 + 1, . . . , jr
is of rank r − 1. However, since ik = jk for some k, and from the conditions
on the i1 , . . . , ir and j1 , . . . , jr , it follows that B contains an r × r singular
principal submatrix of A. This is a contradiction since all principal minors
of A are strictly positive.
From the above it easily follows that some power of A is strictly totally
 5.1 Oscillation matrices 129

positive. However we wish to prove a bit more; namely that An−1 is strictly
totally positive. From Proposition 2.5 it suffices to prove that
 
n−1 1, . . . , r
A >0
n − r + 1, . . . , n
and
 
n−1 n − r + 1, . . . , n
A >0
1, . . . , r
for r = 1, . . . , n. We will only consider the former inequalities. Note that
 
n−1 i1 , . . . , ir
A =
j1 , . . . , jr
    1   n−2 
i1 , . . . , ir s1 , . . . , s1r s1 , . . . , sn−2
A 1 A 2 ···A r
s1 , . . . , s1r s1 , . . . , s2r j1 , . . . , jr

where the sum is over all 1 ≤ s
1 < · · · < s
r ≤ n,  = 1, . . . , n − 2. As each of
these terms is nonnegative, for strict positivity to hold it suffices to prove
that at least one of the products is strictly positive. For example,
       
1 1 2 n−1
An−1
≥A A ···A > 0,
n 2 3 n
while
         
1, 2 1, 2 1, 3 n − 3, n − 1 n − 2, n
An−1 ≥A A ···A A
n − 1, n 1, 3 2, 4 n − 2, n n − 1, n
> 0.

The strict positivity is a consequence of what we proved above.

Set
s
= (s
1 , . . . , s
r ),

where

s
k = min{n − r + k, k + ( + k − r)+ }, k = 1, . . . , r,  = 0, . . . , n − 1.

Here m+ = max{m, 0}. It is now readily verified that |s
k − s
+1

k | ≤ 1 and

k } = sk
max{s
k , s
+1 k+1 }.

+1
< s
k+1 = min{s
k+1 , s
+1

Thus
 
s
1 , . . . , s
r
A > 0.
s
+1
1 , . . . , sr

+1
130 Eigenvalues and eigenvectors

As s0 = (1, . . . , r) and sn−1 = (n − r + 1, . . . , n) it follows that

 
n−1 1, . . . , r
A > 0.
n − r + 1, . . . , n

Remark If Ak , k = 1, . . . , n − 1, are arbitrary n × n oscillation matrices,

then we also have that A1 · · · An−1 is strictly totally positive. The same
proof applies.

5.2 The Gantmacher–Krein theorem

The main theorem concerning spectral properties of oscillation matrices is
the following.

Theorem 5.3 (Gantmacher–Krein). Let A be an n × n oscillation matrix.

Then the n eigenvalues of A are positive and simple. In addition, if we
denote by uk a real eigenvector (unique up to multiplication by a nonzero
constant) associated with the eigenvalue λk , where λ1 > λ2 > · · · > λn > 0,
then
   
 p p
q − 1 ≤ S−  ci ui  ≤ S +  ci ui  ≤ p − 1,
i=q i=q

for each 1 ≤ q ≤ p ≤ n (and ci not all zero). In particular, S − (uk ) =

S + (uk ) = k − 1 for k = 1, . . . , n.

For the definitions of S − and S + , see Chapter 3.

We present two very different proofs of Theorem 5.3. The first is the
classic proof. In this classic proof we use two known general results. The
first of these is Perron’s Theorem.

Perron’s Theorem Let A be an n × n matrix, all of whose elements are

strictly positive. Then A has a simple, positive eigenvalue that is strictly
greater in modulus than all other eigenvalues of A. Furthermore the unique
(up to multiplication by a nonzero constant) associated eigenvector may be
chosen so that all its components are strictly positive.

Proof There are numerous proofs of this result in the literature, and
an almost uncountable number of generalizations. For completeness, we
present a proof that seems to be one of the simpler and more transparent.
For vectors x and y in Rn , we write x ≥ y if xi ≥ yi , i = 1, . . . , n, and
 5.2 The Gantmacher–Krein theorem 131

x > y if xi > yi , i = 1, . . . , n. Set

λ∗ = sup{λ : Ax ≥ λx , for some x ≥ 0} .

Since all elements of A are strictly positive, we must have λ∗ > 0. A

convergence (compactness) argument implies the existence of an x∗ ≥ 0
(x∗ = 0) such that Ax∗ ≥ λ∗ x∗ . If Ax∗ = λ∗ x∗ , then A(Ax∗ ) > λ∗ (Ax∗ )
since all entries of A are strictly positive. Setting y∗ = Ax∗ it follows that
Ay∗ > λ∗ y∗ , which contradicts our definition of λ∗ . Thus

Ax∗ = λ∗ x∗ .

Now x∗ ≥ 0 (x∗ = 0) and thus Ax∗ > 0, whence x∗ > 0. We have found a
positive eigenvalue with a strictly positive eigenvector.
Let λ be any other eigenvalue of A. Then

Ay = λy

for some y ∈ Cn \{0}. Now

|λ| |y| = |λy| = |Ay| ≤ A|y| ,

where |y| = (|y1 |, . . . , |yn |). From the definition of λ∗ , it follows that |λ| ≤
λ∗ . If |λ| = λ∗ , then we must have λ∗ |y| = A|y| (since otherwise λ∗ (A|y|) <
A(A|y|), contradicting the definition of λ∗ ). Thus, for each i = 1, . . . , n,


n 
n
|λ| |yi | = aij yj = aij |yj | .
j=1 j=1

This implies the existence of a γ ∈ C, |γ| = 1, such that γyj = |yj | for
each j = 1, . . . , n. Thus, we may in fact assume that if |λ| = λ∗ , then
λ = |λ| = λ∗ , and for the associated eigenvector y, we have y ≥ 0. Two
consequences of these facts are the following: for every eigenvalue λ, λ = λ∗ ,
we have |λ| < λ∗ ; the geometric multiplicity of the eigenvalue λ∗ is exactly
1. The latter holds because, if not, we can easily construct a real eigenvector
associated with λ∗ that is not of one sign, and this contradicts the above
analysis.
It remains to prove that the eigenvalue λ∗ is of algebraic multiplicity 1.
Assume not. There then exists a vector y∗ (linearly independent of x∗ )
such that
Ay∗ = λ∗ y∗ + αx∗

with some α = 0. This y∗ is called a “generalized eigenvector.” The

transpose of A, namely AT , has the same eigenvalues as A and is obviously
132 Eigenvalues and eigenvectors

strictly positive. As such there exists an eigenvector w∗ > 0 associated

with the eigenvalue λ∗ . Now
λ∗ (w∗ , y∗ ) = (AT w∗ , y∗ ) = (w∗ , Ay∗ ) = λ∗ (w∗ , y∗ ) + α(w∗ , x∗ ) .
Since w∗ , x∗ > 0 and α = 0 we have α(w∗ , x∗ ) = 0, a contradiction. Thus
the algebraic multiplicity of λ∗ is one.
The second general result we use in the proof of Theorem 5.3 is called
Kronecker’s Theorem. We recall from Chapter 1 that the p th compound
   
matrix, denoted by A[p] , of the n × n matrix A is defined as the np × np
matrix with entries
(A(i, j))i∈Ipn ,j∈Ipn
where the i ∈ Ipn and j ∈ Ipn are arranged in lexicographic order, i.e., for
i, j ∈ Ipn we set i ≥ j (i = j) if the first nonzero term in the sequence
i1 − j1 , . . . , ip − jp is positive.

Kronecker’s Theorem Let A be an n × n matrix with   eigenvalues

λ1 , . . . , λn listed to their algebraic multiplicity. Then the np eigenvalues
of A[p] , listed to their algebraic multiplicity, are λi1 · · · λip for 1 ≤ i1 <
· · · < ip ≤ n.

Proof We present two proofs. Every n × n matrix A may be written in the

form
A = P −1 T P
where T is an upper triangular matrix. As such, the diagonal entries of T
are simply the eigenvalues of A. Now, from the Cauchy–Binet formula,
A[p] = (P −1 )[p] T[p] P[p] = (P[p] )−1 T[p] P[p] .
The matrix T[p] is upper triangular and its diagonal entries are exactly the
products of p distinct diagonal entries of T . This proves the result.
A second proof is the following. Associated with each eigenvalue λi
is an eigenvector/generalized eigenvector ui , i = 1, . . . , n, such that the
u1 , . . . , un span Cn . Now (see Chapter 1)
 
A[p] ui1 ∧ · · · ∧ uip = Aui1 ∧ · · · ∧ Auip
from which it follows that the ui1 ∧ · · · ∧ uip are eigenvectors/generalized
eigenvectors of A[p] with associated eigenvalues λi1 · · · λip . Since these
vectors are linearly independent, we have determined all the np eigenvalues
(and generalized eigenvectors) of A[p] .
 5.2 The Gantmacher–Krein theorem 133

We now present our first proof of Theorem 5.3.

Proof of Theorem 5.3 It suffices to prove the theorem for strictly totally
positive matrices. For if A has eigenvalues λ1 , . . . , λn , listed to their
algebraic multiplicity, then Ak has the eigenvalues λk1 , . . . , λkn . If we show
that λk1 > · · · > λkn > 0 for all k sufficiently large, then obviously we must
have λ1 > · · · > λn > 0. In addition, if Au = λu then Ak u = λk u, so that
if A has n distinct eigenvalues, then the eigenvectors of A and Ak are one
and the same.
Let λ1 , . . . , λn denote the eigenvalues of A listed to their algebraic
multiplicity, and assume that |λ1 | ≥ · · · ≥ |λn | ≥ 0. As A is strictly
totally positive we have that A[p] is a matrix all of whose entries are strictly
positive. From the Perron and Kronecker Theorems applied to A[p] we have

λ1 · · · λp > |λ1 · · · λp−1 λp+1 |

for all p = 1, . . . , n. In addition, A is nonsingular and thus |λn | > 0. From

these two facts we have

λ1 > · · · > λn > 0 .

From the Perron and Kronecker Theorems it also follows that by a suitable
normalization of the associated eigenvectors u1 , . . . , un , we may assume
that
u1 ∧ · · · ∧ up > 0

for p = 1, 2, . . . , n.
It remains for us to prove " the sign change properties of the eigenvectors
1 + p
n
u , . . . , u . Assume S i=q ci u
i
≥ p for some choice of nontrivial
(cq , . . . , cp ). There then exist j0 < · · · < jp and an ε ∈ {−1, 1} such that
 
p
ε(−1)k  ci ui  ≥ 0 , k = 0, . . . , p .
i=q
jk
"p
Set u0 = i=q ci ui . Let ui = (u1,i , . . . , un,i ), i = 0, 1, . . . , n, and U denote
the matrix with columns u0 , u1 , . . . , un . Thus
 
j0 , j1 , . . . , jp p
U = det (ujk ,i )i,k=0 = 0
0, 1, . . . , p

since u0 is a linear combination of the u1 , . . . , up . On the other hand, when

134 Eigenvalues and eigenvectors

expanding this matrix by the first column we obtain

  p

j0 , j1 , . . . , jp j0 , . . . , jk , . . . , jp
U = (−1)k ujk ,0 U .
0, 1, . . . , p 1, . . . , p
k=0

Since u1 ∧ · · · ∧ up > 0, we have


j0 , . . . , jk , . . . , jp
U > 0,
1, . . . , p

for each k = 0, . . . , p. As (−1)k ujk ,0 is weakly of one sign, we must have

ujk ,0 = 0, for all k = 0, . . . , p. But this is impossible since u1 ∧ · · · ∧ up > 0
implies that all p × p minors of the n × p matrix with columns u1 , . . . , up
are nonsingular. In other words, no nontrivial linear combination of the
u1 , . . . , un vanishes at p coordinates. Thus
 
p
S+  ci ui  ≤ p − 1.
i=q

Let v1 , . . . , vn be left eigenvectors of A with associated eigenvalues

λ1 , . . . , λn , respectively. We assume, by what we have proved so far, that
we have normalized the {vj }nj=1 so that v1 ∧ · · · ∧ vp > 0 for p = 1, . . . , n.
Thus, in particular,
 
p
S+  bj v j  ≤ p − 1
j=1

for every choice of nontrivial (b1 , . . . , bp ), p = 1, . . . , n.

Let u ∈ Rn \{0} be any vector satisfying

(u, vj ) = 0 , j = 1, . . . , q − 1 .

We claim that S − (u) ≥ q − 1. If S − (u) = r ≤ q − 2, then there exist indices

1 ≤ i1 < · · · < ir < n and an ε ∈ {−1, 1} such that

ε(−1)k uj ≥ 0 , ik−1 + 1 ≤ j ≤ ik , k = 1, . . . , r + 1,

and uj = 0 for some ik−1 + 1 ≤ j ≤ ik , and each k = 1, . . . , r + 1. Here

i0 = 0 and ir+1 = n. Let

r+1
v= bj v j
j=1
 5.2 The Gantmacher–Krein theorem 135

" v = 0 and vik = 0, k = 1, . . . , r. Such a v exists. Since

satisfy
+ r+1
S j=1 bj v
j
≤ r, we must have

δ(−1)k vj > 0 , ik−1 + 1 ≤ j < ik , k = 1, . . . , r + 1,

and also δ(−1)r+1 vn > 0, where δ ∈ {−1, 1}. As r + 1 ≤ q − 1 and

(u, v) = 0 ,

this is a contradiction, implying that S − (u) ≥ q − 1.

The calculation

λi (ui , vj ) = (λi ui , vj ) = (Aui , vj ) = (ui , vj A) = λj (ui , vj ),

implies that (ui , vj ) = 0 for i = j. Thus

 
 p
 ci ui , vj  = 0 , j = 1, . . . , q − 1 ,
i=q

and therefore
 
p
S−  ci ui  ≥ q − 1 ,
i=q

which proves the theorem.

Here is an alternative proof of this lower bound. The matrix A−1 is
not a strictly totally positive matrix. However, C = DA−1 D is strictly
totally positive where D is the diagonal matrix with diagonal entries
alternately 1 and −1 (see Proposition 1.6). For u = (u1 , . . . , un ), set
! = Du = (u1 , −u2 , . . . , (−1)n+1 un ), i.e., component uj is replaced by
u
(−1)j−1 uj . From Lemma 3.1

S − (u) + S + (!
u) = n − 1

for every u ∈ Rn , u = 0. As uk is an eigenvector of A with associated

!k
eigenvalue λk , the strictly totally positive matrix C has the eigenvector u
with associated eigenvalue 1/λk . Because
1 1
> ··· > >0
λn λ1
it follows from our previous result that
 
p
S+  ! i  ≤ n − q,
ci u
i=q
136 Eigenvalues and eigenvectors

implying
 
p
S−  ci ui  ≥ q − 1 .
i=q

We now present a very different proof of Theorem 5.3. The proof of the
sign change properties of the eigenvectors will be based on the variation
diminishing properties of strictly totally positive matrices. We start by
providing an alternative proof of the fact that A has n simple and positive
eigenvalues. We also simultaneously prove that the eigenvalues of the two
principal submatrices obtained by deleting from A either the first row and
column, or the last row and column, strictly interlace the eigenvalues of A.

Proposition 5.4 Let A be an n × n strictly totally positive matrix. Then

its n eigenvalues are positive and simple. In addition, if these eigenvalues
(k) (k)
are denoted by λ1 > λ2 > · · · > λn > 0, and µ1 > · · · > µn−1 > 0 are the
eigenvalues of the principal submatrix of A obtained by deleting its kth row
and column, then
(k)
λj > µj > λj+1 , j = 1, . . . , n − 1 ,

for k = 1 and k = n.

Proof For ease of notation let Aλ = A − λI. From Sylvester’s Determinant

Identity we have
   
1, . . . , n 2, . . . , n − 1
Aλ Aλ
1, . . . , n 2, . . . , n − 1
   
1, . . . , n − 1 2, . . . , n
= Aλ Aλ
1, . . . , n − 1 2, . . . , n
   
1, . . . , n − 1 2, . . . , n
−Aλ Aλ . (5.1)
2, . . . , n 1, . . . , n − 1

It is readily verified that

   
1, . . . , n − 1 2, . . . , n
Aλ , Aλ >0 (5.2)
2, . . . , n 1, . . . , n − 1
for all λ > 0. Simply expand each determinant as a polynomial in λ and
note that all the coefficients of this polynomial are strictly positive since
all minors of A are strictly positive.
We now use an induction argument to prove our result. The case n = 2
 5.2 The Gantmacher–Krein theorem 137

is easily checked by hand. As such we assume that n > 2. For notational

ease, set
 
1, . . . , n
p(λ) = Aλ
1, . . . , n
 
2, . . . , n
q1 (λ) = Aλ ,
2, . . . , n
 
1, . . . , n − 1
q2 (λ) = Aλ
1, . . . , n − 1
 
2, . . . , n − 1
r(λ) = Aλ .
2, . . . , n − 1
Thus from (5.1) and (5.2)
p(λ)r(λ) < q1 (λ)q2 (λ) (5.3)
for all λ > 0.
We assume, by the induction hypothesis, that the n − 2 zeros of r are
positive and simple, and interlace the n − 1 positive simple zeros of both
q1 and q2 . Let
µ1 > · · · > µn−1 > 0
(1) (n)
denote the zeros of either q1 or q2 , i.e., µj = µj or µj = µj , j =
1, . . . , n − 1. From (5.3) it follows that
p(µi )r(µi ) < 0 , i = 1, . . . , n − 1 .
Furthermore, as r(0) > 0 and by the induction hypothesis,
r(µi )(−1)i+n−1 > 0 , i = 1, . . . , n − 1 .
We therefore have
p(µi )(−1)i+n > 0 , i = 1, . . . , n − 1 .
As p(0) > 0 and p(µn−1 ) < 0, the polynomial p has an additional zero in
(0, µn−1 ). Furthermore, the polynomial p has leading coefficient (−1)n and
p(µ1 )(−1)n+1 > 0. Thus p has an additional zero in (µ1 , ∞). As such, p has
n positive, simple zeros that are interlaced by the {µj }n−1
j=1 . This advances
the induction step.

Another Proof of Theorem 5.3 We assume that λ > µ > 0 are eigenvalues
of A with associated eigenvectors x and y. Then from Theorem 3.3 we have
for all (α, β) = (0, 0)
S + (λαx + µβy) = S + (A(αx + βy)) ≤ S − (αx + βy) . (5.4)
138 Eigenvalues and eigenvectors

Note that this implies, taking β = 0 and α = 0, respectively, that

S + (x) = S − (x) and S + (y) = S − (y) .
Assume that both α = 0 and β = 0. We may rewrite (5.4) as
 µ 
S + αx + βy ≤ S − (αx + βy) .
λ
Iterating this process we have
  µ k 
+
S αx + βy ≤ S − (αx + βy) ,
λ
for every positive integer k. As k → ∞, and since µ/λ < 1, we have from
Lemma 3.2   µ k 
+
lim S αx + βy ≥ S − (x).
k→∞ λ
(In fact, since S − (x) = S + (x) it follows that
  µ k 
+
lim S αx + βy = S − (x) .)
k→∞ λ
Similarly, from (5.4)
 µ  
S + (αx + βy) ≤ S − αx + βy ,
λ
and iterating this process we obtain
  
+ µ k
−
S (αx + βy) ≤ S αx + βy ,
λ
for every positive integer k. As k → ∞, and since µ/λ < 1 and S − (y) =
S + (y), it follows (see Lemma 3.2) that
  
µ k
lim S − αx + βy ≤ S + (y) = S − (y) .
k→∞ λ
Thus for every (α, β) = (0, 0),
S − (x) ≤ S − (αx + βy) ≤ S + (αx + βy) ≤ S − (y) .
If S − (x) = S − (y), then equality holds throughout our sequence of
inequalities, and in particular
S + (αx + βy) = S − (αx + βy)
for all (α, β) = (0, 0). This is impossible since there are choices of α, β for
which the first or last coefficient of the vector αx + βy vanishes, in which
case the above equality cannot possibly hold. Therefore S − (x) < S − (y).
 5.2 The Gantmacher–Krein theorem 139

We proved in Proposition 5.4 that A has n simple, positive eigenvalues

and thus n associated eigenvectors. For each of these n eigenvectors we
have S + (x) = S − (x) ∈ {0, 1, . . . , n − 1}, and if λ > µ > 0 are eigenvalues
with associated eigenvectors x and y, then S − (x) < S − (y). This proves
that if
λ1 > · · · > λn > 0 ,

are the eigenvalues of A, and uk is a real eigenvector associated with the

eigenvalue λk , then necessarily

S + (uk ) = S − (uk ) = k − 1, k = 1, . . . , n.

Now assume that 1 ≤ q ≤ p ≤ n and we are given real constants

cq , . . . , cp , not all zero. Then, following the above analysis, repeated
application of
     
p 
p 
p
S+  ci λi ui  = S +  ci Aui  ≤ S −  ci ui 
i=q i=q i=q

implies that for every positive integer m

     
p  m p p
λ
S+  ui  ≤ S −  ci ui  ≤ S +  ci ui 
i
ci
i=q
λ q i=q i=q
 
p  m
λp
≤ S−  ci ui  .
i=q
λ i

We may assume that cp , cq = 0, and apply Lemma 3.2 to obtain

 
 p  m
+ λi i
lim S ci u ≥ S − (uq ) = q − 1 ,
m→∞
i=q
λ q

while
 

p  m
λp
lim S −  ci ui  ≤ S + (up ) = p − 1 .
m→∞
i=q
λi

Thus
   

p 
p
q − 1 ≤ S−  ci ui  ≤ S +  ci ui  ≤ p − 1 ,
i=q i=q

which completes the proof.

140 Eigenvalues and eigenvectors

Remark There is one more result that is sometimes stated in connection

with the Gantmacher–Krein Theorem (Theorem 5.3). It has to do with
the “interlacing of the zeros” of uk+1 and uk , which is important in the
continuous (integral) analogue of this theorem. One form of this is the
following. Let uk (t) be the continuous function defined on [1, n] which is
linear on each [j, j + 1], j = 1, . . . , n − 1, and such that uk (j) is the jth
component of uk . Since

S − (uk ) = S + (uk ) = k − 1 ,

the function uk (t) has exactly k − 1 zeros that are each strict sign changes.
Since
k − 1 ≤ S − (αuk + βuk+1 ) ≤ S + (αuk + βuk+1 ) ≤ k ,

for every choice of (α, β) = (0, 0), it can be shown that the k − 1 zeros of
uk (t) strictly interlace the k zeros of uk+1 (t).

Since strictly totally positive matrices are dense in the set of totally
positive matrices, see Theorem 2.6, and because of the continuity of the
eigenvalues as functions of the matrix entries, we have the following:

Corollary 5.5 The eigenvalues of totally positive matrices are both real
and nonnegative.

5.3 Eigenvalues of principal submatrices

What can we say about the eigenvalues of the principal submatrices of a
strictly totally positive matrix? Let

λ1 > · · · > λn > 0

denote the eigenvalues of A, and

(k) (k)
µ1 > · · · > µn−1 > 0

the eigenvalues of the principal submatrix of A obtained by deleting its kth

row and column. As we have seen in Proposition 5.4, the n − 1 positive,
(1) (n) n−1
simple eigenvalues {µi }n−1i=1 and {µi }i=1 strictly interlace those of the
original matrix A. This naturally begs the question of whether this same
property holds for all principal submatrices? It also easily follows (see the
proof of the Perron Theorem on positive matrices) that
(k)
λ1 > µ1
 5.3 Eigenvalues of principal submatrices 141

for each k. However, the strict interlacing property does not hold for all
principal submatrices. As an example, consider
 
1 1 0
A =  2 2 1 .
2 2 1
The matrix A is totally positive but not strictly totally positive. However,
since strictly totally positive matrices are dense in the class of totally
positive matrices, there are strictly totally positive matrices whose
eigenvalues (and the eigenvalues of its principal submatrices)
√ √are arbitrarily
close to those of A. The eigenvalues of A are 2 + 3, 2 − 3, and 0. The
eigenvalues of the principal submatrix of A obtained by deleting the second
row and column are 1 and 1. Obviously we do not have interlacing.
Nevertheless there is a weaker interlacing that does hold and it is the
following.

Theorem 5.6 Let A be an n × n strictly totally positive matrix. Then for

each k, 1 < k < n,
(k)
λj−1 > µj > λj+1 , j = 1, . . . , n − 1 ,

(where λ0 = λ1 ).

Proof We first prove that for any k and j as above,

(k)
µj > λj+1 .

Let Ak denote the principal submatrix of A obtained by deleting the kth

(k)
row and column. Let v denote a real eigenvector of Ak with eigenvalue µj ,
i.e.,
(k)
Ak v = µj v .

From Theorem 5.3

S + (v) = S − (v) = j − 1 .

We construct the vector v ∈ Rn from the vector v ∈ Rn−1 by simply

inserting a 0 as its new kth coordinate. That is, we write (somewhat abusing
notation)
v = (v1 , . . . , vk−1 , vk+1 , . . . , vn )

and set
v = (v1 , . . . , vk−1 , 0, vk+1 , . . . , vn ) .
142 Eigenvalues and eigenvectors

Let v be defined by

(k)
Av = µj v .

Thus
v = (v1 , . . . , vk−1 , wk , vk+1 , . . . , vn ) ,

for some easily calculated wk . From Theorem 3.3 and the properties of v,
v and v we have,

j − 1 = S + (v) ≤ S + (v ) ≤ S − (v ) = S − (v) = j − 1 .

Thus
S + (v ) = S − (v ) = j − 1 .

Let u denote a real eigenvector of A with eigenvalue λj+1 , i.e.,

Au = λj+1 u .

Then from Theorem 5.3

S + (u) = S − (u) = j .

If uk = 0, then the vector obtained by deleting uk from u is an eigenvector

of Ak , also with eigenvalue λj+1 . Since this new vector also has exactly j
sign changes, it follows, again from Theorem 5.3, that
(k) (k)
λj+1 = µj+1 < µj .

On the other hand, if wk = 0 then v = v is an eigenvector of A with j − 1

sign changes. As such
(k)
µj = λj > λj+1 .

Thus we may now assume that both uk and wk are nonzero.

The vectors u and v are defined up to multiplication by a nonzero
constant. As such we assume that uk and wk are both positive and set

c∗ = inf{c : c > 0, S − (cv + u) ≤ j − 1} .

We claim that c∗ is a well-defined positive number. This is a consequence

of the following. For all c sufficiently large, e.g., such that c|vi | > |ui | if
vi = 0, we have

S − (cv + u) ≤ S + (cv + u) ≤ S + (v ) = j − 1 .

(Here we have used the fact that cvk + uk = uk and wk are of the same
 5.3 Eigenvalues of principal submatrices 143

sign.) For all c sufficiently small and positive, e.g., such that c|vi | < |ui | if
ui = 0, we have
S − (cv + u) ≥ S − (u) = j .
From the definition of c∗ and continuity properties of S + and S − , it
follows that
S − (c∗ v + u) ≤ j − 1
and
S + (cv + u) ≥ j
for all c ≤ c∗ .
Now
(k)
A(c∗ v + u) = c∗ µj v + λj+1 u .
Let
(k)
c∗ µj
c= > 0.
λj+1
From Theorem 3.3,
S + (cv + u) ≤ S − (c∗ v + u) ≤ j − 1 .
Since cv + u and cv + u differ only in their kth coordinates, where both
are positive, it follows that
S + (cv + u) = S + (cv + u) ≤ j − 1 .
(k)
This implies, from the above, that c > c∗ . Thus µj > λj+1 .
The proof of the reverse inequality
(k)
λj−1 > µj
for any k, and j = 2, . . . , n − 1, is essentially the same. Let x be a real
eigenvector of A with eigenvalue λj−1 . That is,
Ax = λj−1 x .
Note that
S + (x) = S − (x) = j − 2 .
(k)
The vectors v, v , and v are as defined above. If wk = 0 then µj = λj <
(k) (k)
λj−1 . If xk = 0, then λj−1 = µj−1 > µj . We may thus assume that wk
and xk are both positive. We set
a∗ = inf{a : a > 0, S − (ax + v ) ≤ j − 2} ,
144 Eigenvalues and eigenvectors

and now follow the previous reasoning, essentially verbatim.

As strictly totally positive matrices are dense in the set of totally positive
matrices and because of the continuity of the eigenvalues as functions of
the matrix entries, we have the following.

Corollary 5.7 Let A be an n × n totally positive matrix with eigenvalues

λ1 ≥ · · · ≥ λn ≥ 0

(listed to their algebraic multiplicity). Let

(k) (k)
µ1 ≥ · · · ≥ µn−1 ≥ 0

denote the eigenvalues of the principal submatrix of A obtained by deleting

its kth row and column. Then for any k, 1 < k < n,
(k)
λj−1 ≥ µj ≥ λj+1 , j = 1, . . . , n − 1 ,

(where λ0 = λ1 ).

5.4 Eigenvectors
In this section we study in considerably more detail the eigenvector
structure of a strictly totally positive or oscillation matrix. We have from
Theorem 5.3 that if uk is a real eigenvector associated with the kth
eigenvalue in magnitude of an oscillation matrix, then
   
p p
q − 1 ≤ S−  ci ui  ≤ S +  ci ui  ≤ p − 1
i=q i=q

for 1 ≤ q ≤ p ≤ n and nontrivial (cq , . . . , cp ).

Given vectors u1 , . . . , un satisfying the above inequalities, are they
necessarily the eigenvectors of some oscillation and thus strictly totally
positive matrix? In other words, do the above inequalities exactly
characterize the set of eigenvectors of strictly totally positive matrices?
In addition, what do these inequalities imply about the vectors uk ? That
is, what can we say about the eigenmatrix U whose columns are the uk ?
We prove the following.

Theorem 5.8 Assume we are given vectors u1 , . . . , un in Rn . Let U be the

n × n matrix whose kth column is uk . The following are equivalent:
 5.4 Eigenvectors 145

(i) There exists an n × n strictly totally positive matrix A with

eigenvalues λ1 > · · · > λn > 0 and associated eigenvectors
u1 , . . . , un , respectively.
(ii) For every 1 ≤ q ≤ p ≤ n and nontrivial (cq , . . . , cp )
   
p 
p
q − 1 ≤ S−  ck uk  ≤ S +  ck uk  ≤ p − 1. (5.5)
k=q k=q

(iii) There exist εp ∈ {−1, 1}, p = 1, . . . , n, such that

 
i1 , . . . , ip
εp U >0 (5.6)
1, . . . , p
for all 1 ≤ i1 < · · · < ip ≤ n and p = 1, . . . , n, and δq ∈ {−1, 1},
q = 1, . . . , n, such that
 
n
iq , . . . , in
δq U (−1) k=q ik > 0 (5.7)
q, . . . , n
for all 1 ≤ iq < · · · < in ≤ n and q = 1, . . . , n. If we assume
that εn = (−1)n(n+1)/2 (δ1 = 1), then εp δp+1 = (−1)p(p+1)/2 ,
p = 1, . . . , n.

Proof (i) ⇒ (ii) is contained in Theorem 5.3.

(ii) ⇒ (iii). We first prove (5.6). The result holds for p = 1 by inspection
since S + (u1 ) = S − (u1 ) = 0. We assume, by induction, that the result
holds for p − 1 and prove it for p, where p ≥ 2. If
 
i1 , . . . , ip
U =0
1, . . . , p
for some 1 ≤ i1 < · · · < ip ≤ n then there exists a nontrivial (c1 , . . . , cp )
such that
 p

ck uk = 0, j = 1, . . . , p.
k=1 ij

But then


p
+
S ck uk ≥p
k=1

which is a contradiction. Thus

 
i1 , . . . , ip
U = 0.
1, . . . , p
146 Eigenvalues and eigenvectors

It remains to prove that all such minors are of one sign.

Choose 1 ≤ i1 < · · · < ip−1 ≤ n. Let (c1 , . . . , cp ) be a nontrivial vector
of coefficients satisfying
 p

ck uk = 0, j = 1, . . . , p − 1.
k=1 ij

This implies that



p
+
S ck uk ≥ p − 1,
k=1

and from (5.5) we must have



p
S+ ck uk = p − 1.
k=1

Thus cp = 0, and for ij−1 < r < ij , where i0 = 0, ip = n + 1,

 p

j
ε(−1) ck uk > 0, j = 1, . . . , p,
k=1 r

for some ε ∈ {−1, 1}. Solve for cp using the p equations

 p

ck uk = 0, j = 1, . . . , p − 1,
k=1 ij

and 

p
ck uk = αr .
k=1 r

For any r ∈ {1, . . . , n}\{i1 , . . . , ip−1 }, assuming ij−1 < r < ij , we have
 
i1 , . . . , ip−1
αr U
1, . . . , p − 1
cp =  .
i1 , . . . , ij−1 , r, ij , . . . , ip−1
(−1)j+p U
1, . . . , p
As sgn αr = (−1)j ε for some ε ∈ {−1, 1} it follows, using the induction
hypothesis, that for some εp ∈ {−1, 1}
 
i1 , . . . , ij−1 , r, ij , . . . , ip−1
εp U >0
1, . . . , p
for all r between ij−1 and ij and all j = 1, . . . , p, i.e., for all
r ∈ {1, . . . , n}\{i1 , . . . , ip−1 }
 5.4 Eigenvectors 147

with the rows arranged in increasing order. From a connectedness argument

this implies that
 
i1 , . . . , ip
εp U >0
1, . . . , p
for all 1 ≤ i1 < · · · < ip ≤ n.
We now prove (5.7). For u = (u1 , u2 , . . . , un ) we set

! = (u1 , −u2 , . . . , (−1)n+1 un )

u

and recall that

S − (u) + S + (!
u) = n − 1

for every u ∈ Rn , u = 0, see Lemma 3.1. Thus the inequality

 
n
S−  ck uk  ≥ q − 1,
k=q

is equivalent to
 

n
S+  ! k  ≤ n − q,
ck u
k=q

for nontrivial (cq , . . . , cn ). This is exactly the same inequality as assumed

in the previous paragraph. As such there exists a δ!q ∈ {−1, 1} such that
 
! ! iq , . . . , in
δq U >0
q, . . . , n

for all 1 ≤ iq < · · · < in ≤ n, where U ! denotes the n × n matrix with

columns u ! . This immediately translates into
k

 
n
iq , . . . , in
δ!q U (−1) k=q ik +1 > 0,
q, . . . , n
implying
 
n
iq , . . . , in k=q ik
δq U (−1) >0
q, . . . , n
for some δq ∈ {−1, 1} and all 1 ≤ iq < · · · < in ≤ n.
Multiplying any one of the uk by a nonzero constant we may and will
assume that
 
1, . . . , n
U = (−1)n(n+1)/2 ,
1, . . . , n
148 Eigenvalues and eigenvectors

i.e., εn = (−1)n(n+1)/2 and δ1 = 1. Now for any p ∈ {1, . . . , n − 1}, by the

Laplace expansion by minors,

(−1)n(n+1)/2 = det U
     

n
ik +k i1 , . . . , ip ip+1 , . . . , in
= (−1) k=p+1 U U
1, . . . , p p + 1, . . . , n
1≤i1 <···<ip ≤n

where the i1 < · · · < ip and ip+1 < · · · < in are complementary indices in
{1, . . . , n}. As
 
i1 , . . . , ip
εp U >0
1, . . . , p
and
 
ip+1 , . . . , in
n 
k=p+1 ik
δp+1 U (−1) >0
p + 1, . . . , n

it follows that εp δp+1 = (−1)p(p+1)/2 .

(iii) ⇒ (i). Set V = U −1 where, as above, we assume, without loss of
generality, that det U = (−1)n(n+1)/2 . Then
 
p
  

1, . . . , p j1 , . . . , jn−p
V = (−1) k=1 k+jk U (−1)n(n+1)/2
j1 , . . . , jp p + 1, . . . , n
where the j1 < · · · < jp and j1 < · · · < jn−p
are complementary indices in
{1, . . . , n}. As
  

n−p 
j1 , . . . , jn−p
δp+1 U (−1) k=1 jk > 0
p + 1, . . . , n
we have
 
p(p+1)/2 1, . . . , p
δp+1 (−1) V >0
j1 , . . . , jp
for all 1 ≤ j1 < · · · < jp ≤ n and all p = 1, . . . , n − 1, i.e.,
 
1, . . . , p
εp V >0
j1 , . . . , jp
for all 1 ≤ j1 < · · · < jp ≤ n and all p = 1, . . . , n. This implies that
   
i1 , . . . , ip 1, . . . , p
U V >0 (5.8)
1, . . . , p j1 , . . . , jp
for every choice of 1 ≤ i1 < · · · < ip ≤ n and 1 ≤ j1 < · · · < jp ≤ n, and
all p.
 5.5 Eigenvalues as functions of matrix elements 149

For any λ1 > · · · > λn > 0 let Λ be the n × n diagonal matrix with
diagonal entries {λ1 , . . . , λn }. Consider

A = U ΛV.

The columns of U are the (right) eigenvectors of A with associated

eigenvalues λ1 , . . . , λn . We will choose the λj dependent upon U and V .
To prove that A is strictly totally positive it suffices to show that
 
i1 , . . . , ip
A >0
j1 , . . . , jp
for all choices of 1 ≤ i1 < · · · < ip ≤ n and 1 ≤ j1 < · · · < jp ≤ n. For A as
above we have from the Cauchy–Binet formula
      
i1 , . . . , ip i1 , . . . , ip k1 , . . . , kp
A = U λk 1 · · · λ k p V .
j1 , . . . , jp k1 , . . . , kp j1 , . . . , jp
1≤k1 <···<kp ≤n

From (5.8) we see that we can choose the λ1 > · · · > λn > 0 so that A is
strictly totally positive. (From the above construction we see that for any
choice of λ1 > · · · > λn > 0 and A = U ΛV , we have that Am = U Λm V is
a strictly totally positive matrix for m sufficiently large.)

5.5 Eigenvalues as functions of matrix elements

How do the eigenvalues of an oscillation matrix vary as functions of the
elements of the matrix? This is the topic of this section.
We start with a general result.

Proposition 5.9 Let A = (aij ) be an n × n matrix with distinct simple

eigenvalues λ1 , . . . , λn and associated eigenvectors u1 , . . . , un . Let U =
(uij ) be the n × n eigenmatrix whose kth column is uk . Set V = (vij )
where V = U −1 , and let vk denote the kth row of V . Then
∂λk
= vki ujk
∂aij

for all i, j, k ∈ {1, . . . , n}.

Proof The {λk } are locally differentiable functions of the aij since they are
the distinct roots of a polynomial of degree n whose coefficients depend
algebraically upon the aij .
150 Eigenvalues and eigenvectors

From the above Auk = λk uk , and thus


n
ars usk = λk urk , r = 1, . . . , n.
s=1

Similarly vk A = λk vk , i.e.,

n
vkr ars = λk vks , s = 1, . . . , n.
r=1

Now
A = U ΛV

where Λ is the n × n diagonal matrix with diagonal entries {λ1 , . . . , λn }.

Thus
V AU = Λ,

i.e.,

n
λk = vkr ars usk , k = 1, . . . , n.
r,s=1

Differentiating the above with respect to aij , we have

∂λk  n
∂vkr ∂usk
= vki ujk + ars usk + vkr ars
∂aij r,s=1
∂a ij ∂aij
  n
 n
∂vkr 
n n  ∂usk
= vki ujk + ars usk + vkr ars .
r=1
∂aij s=1 s=1 r=1
∂aij

Substituting from above gives

∂λk 
n
∂vkr 
n
∂usk
= vki ujk + λk urk +
λk vks
∂aij r=1
∂a ij s=1
∂aij
 n
 ∂vkr n
∂usk
= vki ujk + λk urk + vks
r=1
∂aij s=1
∂aij
 n
∂ 
= vki ujk + λk vkr urk
∂aij r=1
= vki ujk
"n
since r=1 vkr urk ≡ 1 as V = U −1 .

An application of Proposition 5.9 leads to the following.

 5.5 Eigenvalues as functions of matrix elements 151

Theorem 5.10 Let A = (aij ) be an n × n oscillation matrix with

eigenvalues λ1 > · · · > λn > 0. Then
∂λ1 ∂λn
> 0, (−1)i+j > 0, i, j = 1, . . . , n,
∂aij ∂aij
and
∂λk ∂λk ∂λk ∂λk
> 0, > 0, (−1)k+1 > 0, (−1)k+1 > 0, k = 1, . . . , n.
∂a11 ∂ann ∂a1n ∂an1

Proof Let U be as in Proposition 5.9. Without loss of generality we assume

that u1k > 0, k = 1, . . . , n. This implies, from Theorem 5.3, that uk1 > 0,
k = 1, . . . , n, and (−1)k+1 ukn > 0, (−1)k+1 unk > 0, k = 1, . . . , n. Set
V = U −1 . Then the exact same inequalities hold for V , i.e., v1k , vk1 > 0,
k = 1, . . . , n, and (−1)k+1 vkn , (−1)k+1 vnk > 0, k = 1, . . . , n. Theorem 5.10
is now an immediate consequence of these inequalities and the previous
Proposition 5.9.

The inequality
∂λ1
>0
∂aij
is also a consequence of Perron’s Theorem. Another consequence of Perron’s
Theorem is the following.
Let
 
a11 · · · a1n
 .. .. 
 . . 
 
Ac =  cak1 · · · cakn 


 . . 
 . . . 
.
an1 ··· ann
i.e., Ac is obtained from A by multiplying the elements of the kth row by
c.

Proposition 5.11 Let A be an n × n oscillation matrix. For c > 0, let

λ1 (c) > · · · > λn (c) > 0 denote the eigenvalues of Ac . Then
∂λ1 (c) ∂λn (c)
> 0, >0
∂c ∂c
and
∂(λ1 (c) · · · λr (c))
> 0, r = 1, . . . , n.
∂c
152 Eigenvalues and eigenvectors

Proof All three inequalities are consequences of Perron’s Theorem. That is,
we use the characterization
λ1 = sup{λ : Ax ≥ λx, for some x ≥ 0}
valid for any strictly positive matrix. The first inequality is a direct
application thereof (and is also in Theorem 5.10). The second inequality
follows from an application of Perron’s Theorem to the oscillation matrix
DA−1 c D. For a proof of the third inequality, apply Perron’s Theorem to the
rth compound matrix A[r] .
It is not necessarily true that
∂λr (c)
>0
∂c
for all r. The 3 × 3 totally positive matrix
 
1 1 0
Ac =  2c 2c c 
2 2 1
is such that λ2 (c) is a strictly decreasing function of c, c > 0.

5.6 Remarks
The study of the spectral properties of integral equations with totally
positive continuous kernels substantially predates the study of the spectral
properties of totally positive matrices. In 1918 O. D. Kellogg proved
the main spectral properties in the case of a symmetric continuous
totally positive kernel (see Kellogg [1918]). Kellogg was an American
mathematician who obtained his doctorate from Göttingen in 1903 under
the supervision of Hilbert. He is best known for his work on potential
theory, and his book thereon Kellogg [1929] has been reprinted many times
since. Both Krein and Gantmacher were very much influenced by Kellogg’s
work on this subject. An announcement of the parallel result for continuous
non-symmetric totally positive kernels is in Gantmacher [1936].
The main results concerning spectral properties of totally positive
matrices are in Gantmacher, Krein [1937]. An announcement appears
in Gantmacher, Krein [1935]. Oscillation matrices were introduced in
Gantmacher, Krein [1937], and Theorem 5.2 can be found therein on p. 454
(see also Gantmacher, Krein [1950], Chap. II, §7, Karlin [1968], Chap. 2,
§9, and Ando [1987]). The proof of Theorem 5.3, based on the Perron
and Kronecker theorems, is from Gantmacher, Krein [1937], Theorems 10
 5.6 Remarks 153

and 14. The same proof also appears in Gantmacher [1953], Gantmacher,
Krein [1950], and Ando [1987]. The strict interlacing of the eigenvalues
of a strictly totally positive matrix with the eigenvalues of the principal
submatrices obtained by deleting the first (or last) row and column was
first proved in Gantmacher, Krein [1937]. The proof as given in Proposition
5.4 is from Koteljanskii [1955]. The second proof of Theorem 5.3, based
on variation diminishing, is from Elias, Pinkus [2002], which contains
generalizations of Theorem 5.3 to the nonlinear setting (see also Pinkus
[1985a], [1985b] and Buslaev [1990]). For more on the eigenvalues and
eigenvectors of oscillation matrices see Eveson [1996], Karlin [1965], Karlin
[1972] and Karlin, Pinkus [1974]. Fallat, Gekhtman, Johnson [2000] consider
the possible spectral structure of a subclass of totally positive matrices.
Theorem 5.6 is due to Pinkus [1998]. The example of a totally positive
matrix where interlacing of the eigenvalues of the minor fails is from Karlin,
(k)
Pinkus [1974]. Friedland [1985] had previously proved that µ1 ≥ λ2 and
(k)
µn−1 ≥ λn for k = 1, . . . , n − 1. Parts of Theorem 5.8, in a slightly different
form, appear in Gantmacher, Krein [1937], Theorem 16. Theorem 5.10 is
also from Gantmacher, Krein [1937], Theorems 18 and 19.
A survey of the spectral properties of totally positive kernels and
matrices, with extensive references, can be found in Pinkus [1996].
 6
Factorizations of totally positive matrices

In Chapter 2 (Theorem 2.12) we proved that every n × n totally positive

matrix can be factored in the form

A = LU

where L and U are totally positive matrices, L being a lower triangular

matrix and U an upper triangular matrix.
In this chapter we study factorizations of totally positive matrices in
considerably greater detail. We prove that if A is an n × n totally positive
matrix, then A can always be factored in the form

A = L1 · · · Ln−1 U 1 · · · U n−1

where each Lk is a 1-banded, lower triangular, totally positive matrix and

each U k is a 1-banded, upper triangular, totally positive matrix. In fact
there are generally an uncountable number of such factorizations, and at
least 22n−2 different factorizations with a maximum number of zero entries.
In Section 6.2, after some preliminaries, we start with one such
factorization for A strictly totally positive by essentially writing down the
factorization, and then proving its validity. We then show, using elementary
operations, how it is possible to construct many other such factorizations.
In Section 6.3 we generalize these results to totally positive matrices.

6.1 Preliminaries
We start with two definitions to set our notation.

Definition 6.1 For each i, j ∈ {1, . . . , n}, i = j, we define

Ei,j (α)

154
 6.1 Preliminaries 155

as the n × n unit diagonal matrix with α in the (i, j) position and zeroes
in the other off-diagonal entries.
We recall that right multiplication of A by Ei,j (α) is the operation
whereby α times the ith column of A is added to the jth column of A
with the other columns left unchanged. Left multiplication of A by Ei,j (α)
is the operation whereby α times the jth row of A is added to the ith row
of A with the other rows left unchanged. In addition, as is readily verified,
(Eij (α))−1 = Eij (−α).

Definition 6.2 We say that a lower triangular matrix L = (ij ) is r-banded

if ij = 0 for i − j > r.
Thus L has the form
 
1
 × · 
 
 × · · 
 
 
 × · · · 
L =  .
 × · · · · 
 
 · · · · · 
 
 · · · · · 
× × × × 1
/ 01 2
r

If L is an n × n matrix and, for whatever reason, we say that L is r-banded

and lower triangular, where r ≥ n − 1, then this statement is meaningless
and should be so understood.
We list three elementary properties of r-banded matrices for easy
reference.

Proposition 6.3 If L is an r-banded lower triangular matrix and M is

an s-banded lower triangular matrix, then LM is an (r + s)-banded lower
triangular matrix.

Proposition 6.4 A 1-banded lower triangular matrix is totally positive if

and only if all its elements are nonnegative.
We repeatedly use the following (contrast it with Proposition 6.3).

Proposition 6.5 For k = 1, . . . , n − 1, let

C k = E2,1 (α1 ) · · · Ek+1,k (αk ).
156 Factorizations of totally positive matrices

Then C k is a 1-banded unit diagonal lower triangular matrix with

 k
C i+1,i = αi , i = 1, . . . , k,
 k
C i+1,i = 0, i = k + 1, . . . , n − 1.

If
Rk = En−k+1,n−k (αn−k ) · · · En,n−1 (αn−1 ),

then Rk is a 1-banded unit diagonal lower triangular matrix with

 k
R i+1,i = 0, i = 1, . . . , n − k − 1,
 k
R i+1,i = αi , i = n − k, . . . , n − 1.

Note that when reversing the multiplication order (in C k ) we get

Ek+1,k (α1 ) · · · E2,1 (αk ),

which is a k-banded matrix, whose (i, j)th entry, 1 ≤ j < i ≤ k, is exactly

αj · · · αi−1 ,

with other off-diagonal entries being zero.

6.2 Factorizations of strictly totally positive matrices

We recall (Theorem 2.10) that an n × n strictly totally positive matrix A
has a unique factorization of the form

A = LDU

where L is a unit diagonal lower strictly totally positive matrix, U is a unit

diagonal upper strictly totally positive matrix, and D is a diagonal matrix
whose diagonal entries are strictly positive. We also recall that the unit
diagonal matrix L is lower strictly totally positive if it is lower triangular
and
 
i1 , . . . , ik
L >0
j1 , . . . , jk
whenever im ≥ jm , m = 1, . . . , k.
We first prove a fundamental result from which all other factorization
results will follow.

Theorem 6.6 Let L be an n × n unit diagonal lower triangular matrix.

 6.2 Factorizations of strictly totally positive matrices 157

Then L is lower strictly totally positive if and only if L can be factored in

the form
L = R1 · · · Rn−1 (6.1)

where

Rk = En−k+1,n−k (αk,n−k ) · · · En,n−1 (αk,n−1 ), k = 1, . . . , n − 1,

(6.2)
with αk,j > 0, j = n − k, . . . , n − 1, k = 1, . . . , n − 1.

Proof Assume L has the form (6.1) where each Rk satisfies (6.2), with
the appropriate αk,j strictly positive. As L is the product of unit diagonal
lower totally positive matrices, it is necessarily a unit diagonal lower totally
positive matrix. But we want to prove that it is a lower strictly totally
positive matrix. We recall (Theorem 2.8) that to prove that L is lower
strictly totally positive it suffices to prove that
 
i + 1, . . . , i + r
L > 0, i = 0, 1, . . . , n − r, r = 1, . . . , n.
1, . . . , r
(Since L is unit diagonal, the equations in the case i = 0 are of no interest.)
In fact (see Proposition 2.9), we need only prove the strict positivity of the
above inequalities for i = n − r. However, in the proof of the converse
direction we make use of all these inequalities, and so we also prove them
here.
Recall that
 
1
 0 · 
 
 · · 
 
 · · 
 
 
Rk =  0 · .
 
 αk,n−k · 
 
 · · 
 
 · · 
αk,n−1 1

Because of the format (zero elements) of the Rk it easily follows that

(L)i+1,1 = (Rn−i )i+1,i · · · (Rn−1 )2,1 , i = 1, . . . , n − 1.

As
(Rn−i )i+1,i > 0, i = 1, . . . , n − 1,
158 Factorizations of totally positive matrices

we have
(L)i+1,1 > 0, i = 1, . . . , n − 1.

Now consider
 
i + 1, i + 2
L .
1, 2
Again, from the form of the R1 , . . . , Rn−1 , it follows that
     
i + 1, i + 2 i + 1, i + 2 2, 3
L = Rn−i · · · Rn−1 , i = 1, . . . , n − 2.
1, 2 i, i + 1 1, 2
By assumption,
 
i + 1, i + 2
R n−i
= (Rn−i )i+1,i (Rn−i )i+2,i+1 > 0, i = 1, . . . , n − 2.
i, i + 1
Thus
 
i + 1, i + 2
L > 0, i = 1, . . . , n − 2.
1, 2
We progress in this fashion to obtain
 
i + 1, . . . , i + r
L > 0, i = 1, . . . , n − r, r = 1, . . . , n − 1.
1, . . . , r
Thus L is lower strictly totally positive.
Let us now assume that L is lower strictly totally positive. For k =
1, . . . , n − 1 we define the unit diagonal, 1-banded matrix Rk with
 k
R i+1,i = 0, i = 1, . . . , n − k − 1,
 k
R i+1,i > 0, i = n − k, . . . , n − 1

in the following manner. We will determine the (Rk )i+1,i , i = n − k, . . . ,

n − 1, k = 1, . . . , n − 1, from the previous sets of equations. That is, the
equations

(L)i+1,1 = (Rn−i )i+1,i · · · (Rn−1 )2,1 , i = 1, . . . , n − 1,

and the fact that (L)i+1,i > 0 define for us the strictly positive values

(Rn−i )i+1,i , i = 1, . . . , n − 1.

The equations
     
i + 1, i + 2 i + 1, i + 2 2, 3
L = Rn−i · · · Rn−1 , i = 1, . . . , n − 2,
1, 2 i, i + 1 1, 2
 6.2 Factorizations of strictly totally positive matrices 159

and the strict positivity of the left-hand side of the equations define for us
the strictly positive values
 
i + 1, i + 2
Rn−i = (Rn−i )i+1,i (Rn−i )i+2,i+1 , i = 1, . . . , n − 2,
i, i + 1
(since we set (Rn−i )i+2,i = 0), which in turn define for us the strictly
positive values
(Rn−i )i+2,i+1 , i = 1, . . . , n − 2.
We continue in this fashion to define all the desired (Rk )i+1,i .
We claim that
L = R1 · · · Rn−1 .
That is, we have chosen the values of the Rk so that the formulæ for the
 
i + 1, . . . , i + r
L , i = 1, . . . , n − r, r = 1, . . . , n − 1,
1, . . . , r
are valid. But do these formulæ necessarily guarantee that
L = R1 · · · Rn−1 ?
The answer is, of course, yes. The above set of data, and their strict
positivity, determine explicitly all the entries of L. That is, the first column
of L is given (r = 1). Given the first column (and the fact that its entries
are not zero) and
 
i + 1, i + 2
L , i = 1, . . . , n − 2,
1, 2
determines the entries of the second column of L, etc. . . Thus we do have
L = R1 · · · Rn−1 .
This proves the theorem.

Remark From the above follow very explicit formulæ for the (Rk )i+1,i ,
i = n − k, . . . , n − 1, k = 1, . . . , n − 1, in terms of minors of L. But we shall
not list them here.
Let us now consider how we can construct a different factorization for L.

Corollary 6.7 Let L be an n × n unit diagonal lower triangular matrix.

Then L is lower strictly totally positive if and only if L can be factored in
the form
L = C n−1 · · · C 1 (6.3)
160 Factorizations of totally positive matrices

where
C k = E2,1 (βk,1 ) · · · Ek+1,k (βk,k ) (6.4)

with βk,j > 0, j = 1, . . . , k, k = 1, . . . , n − 1.

Proof There are two simple ways of justifying this corollary. The first is to
take the proof of Theorem 6.6 and make the simple obvious modifications
therein.
The second explanation is more satisfying. Recall from Propositions 1.2
and 1.3 that a matrix A is totally positive (strictly totally positive) if and
only if AT is totally positive (strictly totally positive) if and only if QAQ
is totally positive (strictly totally positive), where
 
0 ··· 0 1
 0 ··· 1 0 
 
Q= . . .
 .. . . ... ... 
1 ··· 0 0

Note that QAQ is the matrix A with the order of its rows and columns
reversed. Given a lower triangular matrix L = (ij )ni,j=1 , then

(QLQ)T = QLT Q = (n−j+1,n−i+1 )ni,j=1

is also lower triangular. In addition, L is lower strictly totally positive if

and only if (QLQ)T is lower strictly totally positive.
Applying Theorem 6.6 to (QLQ)T gives us

(QLQ)T = R1 · · · Rn−1

for R1 , . . . , Rn−1 as in the statement of Theorem 6.6. Thus

L = (QRn−1 Q)T · · · (QR1 Q)T .

Set
C k = (QRk Q)T , k = 1, . . . , n − 1.

The C k are as in the statement of Corollary 6.7.

We can and should look at the factorizations of Theorem 6.6 and

Corollary 6.7 from a somewhat different (and historically more valid)
perspective. To explain, let

S k = Rk · · · Rn−1 , k = 1, . . . , n − 1.
 6.2 Factorizations of strictly totally positive matrices 161

Thus
L = R1 · · · Rk−1 S k .

As may be seen, S k is a full (n − k)-banded matrix (more on the structure

of the S k anon) and
S k = Rk S k+1 .

To explain what is happening recall that (Ei,j (α))−1 = Ei,j (−α) and

Rk = En−k+1,n−k (αk,n−k ) · · · En,n−1 (αk,n−1 ).

Thus
(Rk )−1 = En,n−1 (−αk,n−1 ) · · · En−k+1,n−k (−αk,n−k ),

and
S k+1 = En,n−1 (−αk,n−1 ) · · · En−k+1,n−k (−αk,n−k )S k .

As previously noted, left multiplication of a matrix A by Er+1,r (−α) is the

operation whereby α times the rth row of A is subtracted from the r + 1st
row of A, with the other rows left unchanged. Thus S k+1 is obtained from
S k by successively eliminating (making zero) the k strictly positive elements
along the (n − k)-band of S k (i.e., elements in positions (n − k + j, j),
j = 1, . . . , k). It does this by successive row subtraction. It first eliminates
the element in the (n − k + 1, 1) position by subtracting αk,n−k times the
(n − k)th row from the (n − k + 1)st row. It then eliminates the element in
the (n−k+2, 2) position by subtracting αk,n−k+1 times the (n−k+1)st row
from the (n − k + 2)nd row. This does not effect the previous “elimination,”
etc.
Similarly, from
L = C n−1 · · · C 1

as in Corollary 6.7, let

T k = C n−1 · · · C k , k = 1, . . . , n − 1.

Thus
L = T k C k−1 · · · C 1

and
T k = T k+1 C k .

The matrix T k is a full (n − k)-banded matrix. The difference here is

that (C k )−1 uses column subtraction (rather than the row subtraction of
162 Factorizations of totally positive matrices

(Rk )−1 ) to take the (n − k)-banded matrix T k to the (n − k − 1)-banded

matrix T k+1 .
Starting from a unit diagonal lower strictly totally positive matrix it is
readily verified that both the S k and T k are unit diagonal lower totally
positive matrices and are as strictly totally positive as they can be,
considering their geometric structure, i.e., considering the fact that they
are (n − k)-banded. The matrix S k satisfies
 
i1 , . . . , ir
Sk >0
j1 , . . . , jr
if and only if

js ≤ is ≤ js + (n − k), s = 1, . . . , r,

and the exact same inequalities hold for T k . Thus the S k and T k have
exactly the same total positivity structure.
We can state this procedure formally as follows.

Proposition 6.8 Let P be an n × n unit diagonal lower triangular (n − k)-

banded totally positive matrix for which
 
i1 , . . . , ir
P >0
j1 , . . . , jr

if and only if js ≤ is ≤ js + (n − k), s = 1, . . . , r. Then there exist Rk and

C k where
Rk = En−k+1,n−k (αn−k ) · · · En,n−1 (αn−1 )

with αj > 0, j = n − k, . . . , n − 1, and

C k = E2,1 (β1 ) · · · Ek+1,k (βk )

with βj > 0, j = 1, . . . , k, such that

P = Rk S = T C k

where S and T are n × n unit diagonal lower triangular (n − k − 1)-banded

totally positive matrices and
   
i1 , . . . , ir i1 , . . . , ir
S , T >0
j1 , . . . , jr j1 , . . . , jr

if and only if js ≤ is ≤ js + (n − k − 1), s = 1, . . . , r.

 6.2 Factorizations of strictly totally positive matrices 163

We sequentially did the same thing at each step in Theorem 6.6 and 6.7,
i.e., we decomposed
S k = Rk S k+1

for k = 1, . . . , n − 1, in Theorem 6.6 and decomposed

T k = T k+1 C k

for k = 1, . . . , n − 1, in Corollary 6.7. However we could just as easily have

decomposed S k via
S k = S k+1 C k

and T k via
T k = Rk T k+1

where C k has the structure of C k , Rk has the structure of Rk , and S k+1 ,

T k+1 , S k+1 and T k+1 all have the same structure. This is the content of
Proposition 6.8. Since we can make the choice of using row subtraction (Rk )
or column subtraction (C k ) at each step of the procedure we see that we
actually have 2n−1 possible factorizations. (Note that there is an ordering
to these 2n−1 factorizations. Consider R1 · · · Rn−1 C n−1 · · · C 1 and for each
k delete one and only one of the Rk or C k . These are the possible forms of
the factorizations where it is understood that in every distinct factorization
the entries of the Rk or C k depend on the choice of the factorization.)
What characterizes these 2n−1 factorizations is that the Rk and C k
have exactly k positive off-diagonal entries, the last k entries or the first
k entries, respectively, and this is minimal. On the other hand, without
the minimality of nonzero terms there can be (and will be) a continuum
of factorizations of a lower strictly totally positive n × n matrix into the
product of (n − 1) 1-banded totally positive factors. For example, consider
the matrix
 
1 0 0
A =  1 1 0 .
1 2 1

This matrix can be factored into a continuum of a product of two 1-banded

unit diagonal lower totally positive matrices, namely for every x ∈ [1/2, 1]
    
1 0 0 1 0 0 1 0 0
A= 1 1 0 = 1−x 1 0  x 1 0 .
1 2 1 0 1/x 1 0 2 − 1/x 1
164 Factorizations of totally positive matrices

(Every 3 × 3 lower strictly totally positive matrix can be factored into a

continuum of products of two 1-banded lower totally positive matrices.)
What was done for a lower strictly totally positive matrix is also valid for
an upper strictly totally positive matrix, with the obvious modifications.
Thus, taking into consideration the above results together with the LDU
factorization of a strictly totally positive matrix as detailed in Theorem
2.10, we can summarize as follows.

Theorem 6.9 Let A be an n × n strictly totally positive matrix. Then A

can be factored in many ways in the form

A = L1 · · · Ln−1 DU 1 · · · U n−1

where D is a diagonal matrix whose diagonal entries are strictly positive,

the Lk are 1-banded, unit diagonal, lower totally positive matrices, and the
U k are 1-banded, unit diagonal, upper totally positive matrices. There are
also many (in general 22n−2 ) different factorizations of this form where
among the L1 , . . . , Ln−1 there is exactly one matrix with k strictly positive
entries on its off-diagonal, k = 1, . . . , n − 1, and among the U 1 , . . . , U n−1
there is exactly one matrix with k strictly positive entries on its off-diagonal,
k = 1, . . . , n − 1.

We should also recall that every strictly totally positive (totally positive
matrix) also has a factorization of the form A = U DL, where U is upper
strictly totally positive (totally positive) and L is lower strictly totally
positive (totally positive). Thus there are many, many factorizations of A
as a product of 2n − 1 1-banded totally positive matrices.

6.3 Factorizations of totally positive matrices

What happens when L is only a lower totally positive matrix? Do we still
have the above factorizations? If L is nonsingular (let us assume that it
is unit diagonal) then the answer is yes, although we may lose uniqueness
of the Rk and C k . For example, if (L)n,1 = (L)n−1,1 = 0, then in the
factorization
L = R1 · · · Rn−1

not only is R1 = En,n−1 (α1,n−1 ) superfluous, but in fact it may be

arbitrarily chosen with sufficiently small nonnegative values for α1,n−1 .
(This choice will effect the entries of R2 , . . . , Rn−1 .) Nonetheless the proof
 6.3 Factorizations of totally positive matrices 165

of Theorem 6.6 remains valid, modulo this type of proviso, because L is

nonsingular and as we have seen in Proposition 1.15 and Theorem 1.19 if
 
i + 1, . . . , i + r
L =0
1, . . . , r
then necessarily
 
j + 1, . . . , j + s
L =0
1, . . . , s
for all j ≥ i and all s ≥ r, so that the equations in the proof of Theorem
6.6 are solvable.
However, if L is singular then we cannot apply the previous reasoning.
Nevertheless, from density, and paralleling the analysis of Theorem 2.12,
we have the following.

Theorem 6.10 An n×n lower triangular matrix L is lower totally positive

if and only if it can be factored in the form

L = R1 · · · Rn−1

where Rk is a 1-banded lower totally positive matrix with (Rk )i+1,i = 0,

i = 1, . . . , n − k − 1. Similarly, L is lower totally positive if and only if it
can be factored in the form

L = C n−1 · · · C 1

where C k is a 1-banded lower totally positive matrix with (C k )i+1,i = 0,

i = k + 1, . . . , n − 1.

Proof If L has either of the above factorizations, then L is a lower totally

positive matrix as it is a product of the lower totally positive matrices Rk
(or C k ). It remains to prove the converse claims.
A simple variation of Theorem 2.6 proves that there exist a sequence
of lower strictly totally positive matrices (Lm ) which approximate L, i.e.,
such that
lim Lm = L.
m→∞

From Theorem 6.6 each Lm can be factored in the form

1
Lm = Rm · · · Rm
n−1

k
where the Rm are nonsingular 1-banded lower totally positive matrices with
166 Factorizations of totally positive matrices

(Rmk
)i+1,i = 0, i = 1, . . . , n − k − 1. By pre- and postmultiplying by positive
diagonal matrices we can factor Lm in the form
! mR
Lm = D !m1 !n−1
···R m

!k are stochastic (rows sums 1). As a consequence we also have

where the Rm

ij (m) = d!ii (m)
j

for all i, where the d!ii (m) is the (i, i) entry of D

! m . Thus all entries in each
of the factors are uniformly bounded. Now that we have bounds on all
the entries of the matrices in the factorization we may take a subsequence
along which each of the entries of these matrices converges. Denote their
respective limits by D, R1 , . . . , Rn−1 and redefine R1 as DR1 . Then

L = R1 · · · Rn−1

where Rk is a 1-banded lower totally positive matrix with (Rk )i+1,i = 0,

i = 1, . . . , n − k − 1. Similarly it follows that L can be factored in the form

L = C n−1 · · · C 1

where each C k is a 1-banded lower totally positive matrix that satisfies

(C k )i+1,i = 0, i = k + 1, . . . , n − 1.

As was seen in Section 6.2, these are just two of at least 2n−1 (generally
distinct) possible factorizations. That is, we can mix and match the factors
where for each k we choose to obtain a C k or an Rk .
Note that if L is nonsingular then each of the Rk (or C k ) must be
nonsingular. Thus if the L of Theorem 6.10 is unit diagonal, then we can
also make this same demand on the Rk (or C k ).
Paralleling Theorem 6.9, we can record the following.

Theorem 6.11 Let A be an n × n totally positive matrix. Then A can be

factored in many ways in the form

A = L1 · · · Ln−1 U 1 · · · U n−1

where the Lk are 1-banded, unit diagonal, lower totally positive matrices,
and the U k are 1-banded, unit diagonal, upper totally positive matrices.

We also know that we can restrict the Lk and U k as in Theorem 6.10

so that they have a fairly large number of zero entries on their nonzero
off-diagonal.
 6.4 Remarks 167

6.4 Remarks

The first factorization theorem for totally positive matrices is often called
the Whitney Theorem. The name was given by Loewner [1955], although
there is no evidence to indicate that Whitney [1951] ever thought of her
result in this way. Whitney only talks about a “reduction theorem” and
proves what we have listed as Proposition 1.9. Loewner gave this name to
his factorization theorem because it is proved by repeated applications of
Proposition 1.9. In fact the Whitney/Loewner method of factorization is
not, from our perspective, a good factorization. Rather than eliminating
successive off-diagonals, i.e., going from a (k + 1)-banded lower (upper)
totally positive matrix to a k-banded lower (upper) totally positive matrix
via multiplication by one 1-banded totally positive matrix, the procedure
of Whitney/Loewner eliminates (makes zero) successive columns in a lower
totally positive matrix. This implies the need for n(n − 1)/2 1-banded
factors when factoring a lower totally positive matrix (rather than just
the n − 1 factors). In the Whitney Theorem one assumes nonsingularity of
the matrix. Cryer [1976] obtained 1-banded factorizations for an arbitrary
totally positive matrix. But again the number of such factors can be very
large. In fact Metelmann [1973] (the paper was almost unnoticed) was the
first to state a factorization theorem for totally positive matrices with the
correct number of factors. In Cavaretta, Dahmen, Micchelli, Smith [1981]
can be found a factorization theorem (with the correct number of factors)
for infinite strictly m-banded totally positive matrices. This was generalized
in de Boor, Pinkus [1982] to an arbitrary totally positive nonsingular matrix
(banded if the matrix is infinite). All three of these last mentioned papers,
i.e., Metelmann [1973], Cavaretta, Dahmen, Micchelli, Smith [1981] and
de Boor, Pinkus [1982] also consider totally positive matrices A = (aij )
that are (r, s)-banded, i.e., for which aij = 0 only if −s ≤ i − j ≤ r,
and in this case show that one can make do with r + s factors (aside
from the diagonal matrix). This also follows from the method of proof of
the results in this chapter and Proposition 6.8. The proofs in some of the
papers mentioned above tend to be rather different. The method of proof
of Theorem 6.6 as presented here is basically to be found in Micchelli,
Pinkus [1991]. In that paper a factorization is also given for n × m totally
positive matrices. If n > m then aside from m − 1 lower and m − 1 upper
triangular m × m 1-banded totally positive factors, there are also n − m
totally positive factors that are k × (k − 1) matrices, k = m + 1, . . . , n, all of
whose entries are zero except for those on the two main diagonals (i.e., the
(i, i)- and (i + 1, i)-entries, i = 1, . . . , k − 1). This follows by expanding the
168 Factorizations of totally positive matrices

n × m totally positive matrix to an n × n totally positive matrix (just add

columns of zeros), choosing the suitable factorization, and then truncating
the result.
In a series of papers (see Gasca, Peña [1996] and references therein),
factorizations of totally positive matrices are studied in depth using Neville
elimination. These authors also use factorizations to obtain many of the
characterizations of totally positive matrices as presented in Chapter 2;
see e.g., Gasca, Peña [1992], [1993]. In this monograph we have chosen
a different approach. Factorizations are also used in the papers of Koev
[2005], Demmel, Koev [2005] and Koev [2007] where their concern is with
computing using totally positive matrices. For the use of factorizations in
connection with weighted planar networks, see Fomin, Zelevinsky [1999]
and Fomin, Zelevinsky [2000].
 Afterword

It is very difficult, if not well nigh impossible, to give an exact history of the
development of any set of ideas. Nonetheless, there are four persons whose
contributions “stand out” when considering the history of total positivity.
They are I. J. Schoenberg, M. G. Krein, F. R. Gantmacher, and S. Karlin.
Of course they did not work in a vacuum and numerous influences are very
evident in their research.
It was Schoenberg who initiated the study of the variation diminishing
properties of totally positive matrices in 1930 in Schoenberg [1930], and
the study of Pólya frequency functions in the late 1940s and early 1950s.
Independently, and unaware of Schoenberg’s work, Krein was developing
the theory of total positivity as it related to ordinary differential equations
whose Green’s functions are totally positive. Furthermore, in the mid-
1930s Krein, together with Gantmacher, proved the spectral properties
of totally positive kernels and matrices, and many other properties (see
Gantmacher, Krein [1935], Gantmacher [1936], Gantmacher, Krein [1937],
and their influential Gantmacher, Krein [1941], which was later reissued as
Gantmacher, Krein [1950], and its translations in German in 1960 and in
English in 1961 and 2002). These topics are the foundations upon which has
been constructed the theory of total positivity. Karlin’s role was somewhat
different. His books Karlin, Studden [1966] and Karlin [1968], the latter
titled Total Positivity. Volume 1 (but there is no Volume 2), presented
many new results and ideas and also synthesized and popularized many of
these ideas.
As the reader has hopefully noted, each chapter of this monograph
ends with remarks that include bibliographical references and explanations.
However I wanted to take this opportunity to write a “few words” in
memory of each of these gentlemen.
I. J. (Iso) Schoenberg (1903–1990) was born in Galatz, Romania and

169
170 Afterword

I. J. Schoenberg, 1903–1990

died in Madison, Wisconsin. His family moved to Jassy in 1910, and

he studied mathematics at the university there. He spent three years in
Germany studying with both Edmund Landau and Issai Schur. Influenced
by Schur, he wrote a thesis in Analytic Number Theory. (From Landau
he also took Landau’s daughter Charlotte (Dolli) as his first wife. He is
one of many exemplars of the well-known fact that mathematical talent is
an inherited trait: from father to son-in-law.) Schoenberg was interested
in the problem of estimating the number of real zeros of a polynomial,
and this led him to his work on variation diminishing transformations
and Pólya frequency functions and kernels, which are two major topics in
the theory of total positivity. A Rockefeller fellowship fortunately brought
him to the United States in 1930. He spent time at the University of
Chicago, was a Fellow at the newly established Princeton Institute of
Advanced Studies, was on the faculty of Colby College from 1936 to 1941,
and was then at the University of Pennsylvania from 1941 to 1965. In
1965 Schoenberg moved to the University of Wisconsin and joined the
Mathematics Research Center and the Department of Mathematics. He
retired in 1973 but remained mathematically active until his death. He
contributed to many areas of mathematics, in particular total positivity
and splines. For further information, see de Boor [1988] (and especially
Schoenberg’s autobiographical “A brief account of my life and work” at
the beginning of the first volume), Askey, de Boor [1990], MacTutor, and
references therein.
Mark Grigorievich Krein (1907–1989) was born in Kiev and died in
Odessa. He was a truly eminent and exceedingly prolific mathematician
 Afterword 171

M. G. Krein, 1907–1989

who contributed significantly to and had a tremendous impact upon many

different areas of mathematics (see e.g., Gohberg [1989] and Gohberg [1990],
MacTutor, and references therein). The story of M. G. Krein, and the
mathematical schools he built, is fundamentally marred by the tyranny
and antisemitism to which he was constantly exposed. Krein was dismissed
from his position at the University of Odessa in 1944 and from his part-
time position at the Mathematical Institute of the Ukrainian Academy
of Sciences in Kiev in 1952. We can only speculate on what might have
been if he had been treated with the respect and dignity that were his
due. In 1939 he was elected a corresponding member of the Ukrainian
Academy of Sciences. He was never elected a full member. (This prompted
the famous mathematical joke. Ques: How do you know that the Ukrainian
Academy of Sciences is the best academy in the world? Ans: Because Krein
is only a corresponding member.) From 1944 to 1954 Krein held the chair
in theoretical mechanics at the Odessa Naval Engineering Institute, and
from 1954 until his retirement he held the chair in theoretical mechanics
at the Odessa Civil Engineering Institute. Despite being persecuted Krein
received international recognition. He was elected an honorary member of
the American Academy of Arts and Sciences in 1968, a foreign member
of the National Academy of Sciences (of the USA) in 1979, and in 1982
he was awarded the Wolf Prize. The citation for this prize states: “Krein
brought the full force of mathematical analysis to bear on problems of
function theory, operator theory, probability and mathematical physics.
His contributions led to important developments in the applications of
mathematics to different fields ranging from theoretical mechanics to
172 Afterword

electrical engineering. His style in mathematics and his personal leadership

and integrity have set standards of excellence.”

F. R. Gantmacher, 1908–1964

Feliks Ruvimovich Gantmacher (1908–1964) was born in Odessa and he

studied there. In 1934 he moved to Moscow, where he resided until his
death. Gantmacher is, of course, known for his excellent and influential
book The Theory of Matrices (Gantmacher [1953]), and his book with
Krein; Gantmacher, Krein [1941], [1950]. He was also one of the organizers
and editors of the journal Uspekhi Mat. Nauk (Russian Math. Surveys).
Gantmacher was instrumental in the establishment and organization of the
well-known Moscow Physico-Technical Institute, where from 1953 until his
death he headed the Department of Theoretical Applied Mathematics. An
obituary on Gantmacher may be found in Gantmacher [1965].

S. Karlin, 1924–2007
 Afterword 173

Samuel (Sam) Karlin (1924–2007) was born in Yonova, Poland, but he

was raised in Chicago. He earned his PhD from Princeton in 1947 under the
guidance of S. Bochner. From 1956 he was a faculty member at Stanford
University. The breadth and depth of his interests and contributions in
mathematics and in science are astounding. Karlin is the author of more
than 450 papers and 10 books, and he had numerous doctoral students.
Karlin was passionate about mathematics and science. His passion showed
in his lectures, his lifestyle, and his interaction with students and colleagues.
Karlin was widely honored, and he received honorary doctorates, numerous
prizes (the John von Neumann Theory Prize in 1987, and the National
Medal of Science in 1989 to name but two), and he was elected to various
academies (see Karlin [2002] and MacTutor).
 References

Aissen M., I. J. Schoenberg and A. M. Whitney [1952], On the generating

functions of totally positive sequences I, J. d‘Anal. Math. 2, 93–103.
Ando T. [1987], Totally positive matrices, Lin. Alg. and Appl. 90, 165–219.
Askey R. and C. de Boor [1990], In memoriam : I. J. Schoenberg (1903-1990),
J. Approx. Theory 63, 1–2.
Asner B. A., Jr. [1970], On the total nonnegativity of the Hurwitz matrix, SIAM
J. Appl. Math. 18, 407–414.
Beckenbach, E. F. and R. Bellman [1961], Inequalities, Springer–Verlag,
New York.
Berenstein A., S. Fomin and A. Zelevinsky [1996], Parametrizations of canonical
bases and totally positive matrices, Adv. Math. 122, 49–149.
Boocher A. and B. Froehle [2008], On generators of bounded ratios of minors for
totally positive matrices, Lin. Alg. and Appl. 428, 1664–1684.
de Boor C. [1982], The inverse of totally positive bi-infinite band matrices,
Trans. Amer. Math. Soc. 274, 45–58.
de Boor C. [1988], I. J. Schoenberg: Selected Papers, 2 Volumes, Birkhäuser,
Basel.
de Boor C. and A. Pinkus [1977], Backward error analysis for totally positive
linear systems, Numer. Math. 27, 485–490.
de Boor C. and A. Pinkus [1982], The approximation of a totally positive band
matrix by a strictly banded totally positive one, Lin. Alg. and Appl. 42, 81–98.
Brenti F. [1989], Unimodal, log-concave, and Pólya frequency sequences in
combinatorics, Mem. Amer. Math. Soc. 413.
Brenti F. [1995], Combinatorics and total positivity, J. Comb. Theory, Series A
71, 175–218.
Brenti F. [1996], The applications of total positivity to combinatorics and
conversely, in Total Positivity and its Applications eds., C. A. Micchelli and
M. Gasca, Kluwer Acad. Publ., Dordrecht, 451–473.
Brown L. D., I. M. Johnstone and K. B. MacGibbon [1981], Variation diminishing
transformations: a direct approach to total positivity and its statistical
applications, J. Amer. Stat. Assoc. 76, 824–832.

174
 References 175

Brualdi R. A., and H. Schneider [1983], Determinantal identities: Gauss, Schur,

Cauchy, Dylvester, Kronecker, Jacobi, Binet, Laplace, Muir, and Cayley, Lin.
Alg. and Appl. 52/53, 769–791.
Buslaev A. P. [1990], A variational description of the spectra of totally positive
matrices, and extremal problems of approximation theory, Mat. Zametki 47,
39–46; English transl. in Math. Notes 47 (1990), 26–31.
Carlson B. C. and J. L. Gustafson [1983], Total positivity of mean values and
hypergeometric functions, SIAM J. Math. Anal. 14, 389–395.
Carlson D. [1968], On some determinantal inequalities, Proc. Amer. Math. Soc.
19, 462–466.
Carnicer J. M., T. N. T. Goodman and J. M. Peña [1995], A generalization of
the variation diminishing property, Adv. Comp. Math. 3, 375–394.
Cavaretta A. S. Jr., W. A. Dahmen and C. A. Micchelli [1991], Stationary
subdivision, Mem. Amer. Math. Soc. 93.
Cavaretta A. S. Jr., W. A. Dahmen, C. A. Micchelli and P. W. Smith [1981],
A factorization theorem for banded matrices, Lin. Alg. and Appl. 39, 229–245.
Crans A. S., S. M. Fallat and C. R. Johnson [2001], The Hadamard core of the
totally nonnegative matrices, Lin. Alg. and Appl. 328, 203–222.
Craven T. and G. Csordas [1998], A sufficient condition for strict total positivity
of a matrix, Linear and Multilinear Alg. 45, 19–34.
Cryer C. [1973], The LU-factorization of totally positive matrices, Lin. Alg. and
Appl. 7, 83–92.
Cryer C. [1976], Some properties of totally positive matrices, Lin. Alg. and Appl.
15, 1–25.
Dahmen W., C. A. Micchelli, and P. W. Smith [1986], On factorization of
bi-infinite totally positive block Toeplitz matrices, Rocky Mountain J. Math.
16, 335–364.
Demmel J. and P. Koev [2005], The accurate and efficient solution of a totally
positive generalized Vandermonde linear system, SIAM J. Matrix. Anal. Appl.
27, 142–152.
Dimitrov D. K. and J. M. Peña [2005], Almost strict total positivity and a class
of Hurwitz polynomials, J. Approx. Theory 132, 212–223.
Edrei A. [1952], On the generating functions of totally positive sequences II,
J. d’Anal. Math. 2, 104–109.
Edrei A. [1953], Proof of a conjecture of Schoenberg on the generating functions
of totally positive sequences, Canadian J. Math. 5, 86–94.
Elias U. and A. Pinkus [2002], Non-linear eigenvalue-eigenvector problems for
STP matrices, Proc. Royal Society Edinburgh: Section A 132, 1307–1331.
Eveson S. P. [1996], The eigenvalue distribution of oscillatory and strictly
sign-regular matrices, Lin. Alg. and Appl. 246, 17–21.
Fallat S. M., M. I. Gekhtman and C. R. Johnson [2000], Spectral structures
of irreducible totally nonnegative matrices, SIAM J. Matrix Anal. Appl. 22,
627–645.
Fallat S. M., M. I. Gekhtman and C. R. Johnson [2003], Multiplicative
principal-minor inequalities for totally nonnegative matrices, Adv. Applied
Math. 30, 442–470.
176 References

Fan K. [1966], Some matrix inequalities, Abh. Math. Sem. Univ. Hamburg 29,
185–196.
Fan K. [1967], Subadditive functions on a distributive lattice and an extension of
Szasz’s inequality, J. Math. Anal. Appl. 18, 262–268.
Fan K. [1968], An inequality for subadditive functions on a distributive lattice
with application to determinantal inequalities, Lin. Alg. and Appl. 1, 33–38.
Fekete M. and G. Pólya [1912], Über ein Problem von Laguerre, Rend. C. M.
Palermo 34, 89–120.
Fomin S. and A. Zelevinsky [1999], Double Bruhat cells and total positivity,
J. Amer. Math. Soc. 12, 335–380.
Fomin S. and A. Zelevinsky [2000], Total positivity: tests and parametrizations,
Math. Intell. 22, 23–33.
Friedland S. [1985], Weak interlacing properties of totally positive matrices,
Lin. Alg. and Appl. 71, 95–100.
Gantmacher F. [1936], Sur les noyaux de Kellogg non symétriques, Comptes
Rendus (Doklady) de l’Academie des Sciences de l’URSS 1 (10), 3–5.
Gantmacher F. R. [1953], The Theory of Matrices, Gostekhizdat, Moscow–
Leningrad; English transl. as Matrix Theory, Chelsea, New York, 2 vols., 1959.
Gantmacher F. R. [1965], Obituary, in Uspekhi Mat. Nauk 20, 149–158; English
transl. as Russian Math. Surveys, 20 (1965), 143–151.
Gantmacher F. R. and M. G. Krein [1935], Sur les matrices oscillatoires, C. R.
Acad. Sci. (Paris) 201, 577–579.
Gantmakher F. R. and M. G. Krein [1937], Sur les matrices complètement non
négatives et oscillatoires, Compositio Math. 4, 445–476.
Gantmacher F. R. and M. G. Krein [1941], Oscillation Matrices and
Small Oscillations of Mechanical Systems (Russian), Gostekhizdat, Moscow-
Leningrad.
Gantmacher F. R. and M. G. Krein [1950], Ostsillyatsionye Matritsy
i Yadra i Malye Kolebaniya Mekhanicheskikh Sistem, Gosudarstvenoe
Izdatel’stvo, Moskva-Leningrad, 1950; German transl. as Oszillationsmatrizen,
Oszillationskerne und kleine Schwingungen mechanischer Systeme, Akademie
Verlag, Berlin, 1960; English transl. as Oscillation Matrices and Kernels and
Small Vibrations of Mechanical Systems, USAEC, 1961, and also a revised
English edition from AMS Chelsea Publ., 2002.
Garloff J. [1982], Criteria for sign regularity of sets of matrices, Lin. Alg. and
Appl. 44, 153–160.
Garloff J. [2003], Intervals of almost totally positive matrices, Lin. Alg. and Appl.
363, 103–108.
Garloff J. and D. Wagner [1996a], Hadamard products of stable polynomials are
stable, J. Math. Anal. Appl. 202, 797–809.
Garloff J. and D. Wagner [1996b], Preservation of total nonnegativity under the
Hadamard products and related topics, in Total Positivity and its Applications
eds., C. A. Micchelli and M. Gasca, Kluwer Acad. Publ., Dordrecht, 97–102.
Gasca M. and C. A. Micchelli [1996], Total Positivity and its Applications, eds.,
Kluwer Acad. Publ., Dordrecht.
 References 177

Gasca M., C. A. Micchelli and J. M. Peña [1992], Almost strictly totally positive
matrices, Numerical Algorithms 2, 225–236.
Gasca M. and J. M. Peña [1992], Total positivity and Neville elimination, Lin.
Alg. and Appl. 165, 25–44.
Gasca M. and J. M. Peña [1993], Total positivity, QR factorization, and Neville
elimination, SIAM J. Matrix Anal. Appl. 14, 1132–1140.
Gasca M. and J. M. Peña [1995], On the characterization of almost strictly totally
positive matrices, Adv. Comp. Math. 3, 239–250.
Gasca M. and J. M. Peña [1996], On factorizations of totally positive matrices,
in Total Positivity and its Applications eds., C. A. Micchelli and M. Gasca,
Kluwer Acad. Publ., Dordrecht, 109–130.
Gasca M. and J. M. Peña [2006], Characterizations and decompositions of almost
strictly totally positive matrices, SIAM J. Matrix Anal. 28, 1–8.
Gladwell G. M. L. [1998], Total positivity and the QR algorithm, Lin. Alg. and
Appl. 271, 257–272.
Gladwell G. M. L. [2004], Inner totally positive matrices, Lin. Alg. and Appl.
393, 179–195.
Gohberg I. [1989], Mathematical Tales, in The Gohberg Anniversary Collection,
Eds. H. Dym, S. Goldberg, M. A. Kaashoek, P. Lancaster, pp. 17–56, Operator
Theory: Advances and Applications, Vol. 40, Birkhäuser Verlag, Basel.
Gohberg I. [1990], Mark Grigorievich Krein 1907–1989, Notices Amer. Math. Soc.
37, 284–285.
Goodman T. N. T. [1995], Total positivity and the shape of curves, in Total
Positivity and its Applications eds., C. A. Micchelli and M. Gasca, Kluwer
Acad. Publ., Dordrecht, 157–186.
Goodman T. N. T. and Q. Sun [2004], Total positivity and refinable functions
with general dilation, Appl. Comput. Harmon. Anal. 16, 69–89.
Gross K. I. and D. St. P. Richards [1995], Total positivity, finite reflection groups,
and a formula of Harish-Chandra, J. Approx. Theory 82, 60–87.
Holtz O. [2003], Hermite-Biehler, Routh-Hurwitz and total positivity, Lin. Alg.
and Appl. 372, 105–110.
Horn R. A. and C. R. Johnson [1991], Topics in Matrix Analysis, Cambridge
University Press, Cambridge.
Karlin S. [1964], The existence of eigenvalues for integral operators, Trans. Amer.
Math. Soc. 113, 1–17.
Karlin S. [1965], Oscillation properties of eigenvectors of strictly totally positive
matrices, J. D’Analyse Math. 14, 247–266.
Karlin S. [1968], Total Positivity. Volume 1, Stanford University Press, Stanford,
CA.
Karlin S. [1972], Some extremal problems for eigenvalues of certain matrix and
integral operators, Adv. in Math. 9, 93–136.
Karlin S. [2002], Interdisciplinary meandering in science. 50th anniversary issue
of Operations Research. Oper. Res. 50, 114–121.
Karlin S. and A. Pinkus [1974], Oscillation properties of generalized characteristic
polynomials for totally positive and positive definite matrices, Lin. Alg. and
Appl. 8, 281–312.
178 References

Karlin S. and W. J. Studden [1966], Tchebycheff Systems: with Applications in

Analysis and Statistics, Interscience Publishers, John Wiley, New York.
Katkova O. M. and A. M. Vishnyakova [2006], On sufficient conditions for the
total positivity and for the multiple positivity of matrices, Lin. Alg. and Appl.
416, 1083–1097.
Kellogg O. D. [1918], Orthogonal function sets arising from integral equations,
Amer. J. Math. 40, 145–154.
Kellogg O. D. [1929], Foundations of Potential Theory, Springer-Verlag, Berlin.
Kemperman J. H. B. [1982], A Hurwitz matrix is totally positive, SIAM J. Math.
Anal. 13, 331–341.
Koev P. [2005], Accurate eigenvalues and SVDs of totally nonnegative matrices,
SIAM J. Matrix. Anal. Appl. 27, 1–23.
Koev P. [2007], Accurate computations with totally nonnegative matrices, SIAM
J. Matrix. Anal. Appl. 29, 731–751.
Koteljanskii D. M. [1950], The theory of nonnegative and oscillating matrices
(Russian), Ukrain. Mat. Z. 2, 94–101; English transl. in Amer. Math. Soc.
Transl., Series 2 27 (1963), 1–8.
Koteljanskii D. M. [1955], Some sufficient conditions for reality and simplicity of
the spectrum of a matrix, Mat. Sb. N.S. 36, 163–168; English transl. in Amer.
Math. Soc. Transl., Series 2 27 (1963), 35–42.
Krein M. G. and A. A. Nudel’man [1977], The Markov Moment Problem and
Extremal Problems, Transl. Math. Monographs, Vol. 50, Amer. Math. Society,
Providence.
Kurtz D. C. [1992], A sufficient condition for all the roots of a polynomial to be
real, Amer. Math. Monthly 99, 259–263.
Loewner K. [1955], On totally positive matrices, Math. Z. 63, 338–340.
Lusztig G. [1994], Total positivity in reductive groups, in Lie Theory and
Geometry, Progress in Mathematics, 123, Birkhäuser, Boston, 531–568.
Maló E. [1895], Note sur les équations algébriques dont toutes les racines sont
réelles, Journal de Mathématiques Spéciales 4, 7–10.
Marden M. [1966], Geometry of Polynomials, Math. Surveys, Vol. 3 (2nd edition),
Amer. Math. Society, Providence.
Markham T. L. [1970], A semi-group of totally nonnegative matrices, Lin. Alg.
and Appl. 3, 157–164.
Marshall A. W. and I. Olkin [1979], Inequalities: Theory of Majorization and its
Applications, Academic Press, New York.
Metelmann K. [1973], Ein Kriterium für den Nachweis der Totalnichtnegativität
von Bandmatrizen, Lin. Alg. and Appl. 7, 163–171.
Micchelli C. A. and A. Pinkus [1991], Descartes systems from corner cutting,
Constr. Approx. 7, 161–194.
Motzkin Th. [1936], Beiträge zur Theorie der linearen Ungleichungen, Doctoral
Dissertation, Basel, 1933. Azriel Press, Jerusalem.
Mühlbach G. and M. Gasca [1985], A generalization of Sylvester’s identity on
determinants and some applications, Lin. Alg. and Appl. 66, 221–234.
Pinkus A. [1985a], Some extremal problems for strictly totally positive matrices,
Lin. Alg. and Appl. 64, 141–156.
 References 179

Pinkus A. [1985b], n-widths of Sobolev spaces in Lp , Const. Approx. 1, 15–62.

Pinkus A. [1985c], n-Widths in Approximation Theory, Springer–Verlag, Berlin.
Pinkus A. [1996], Spectral properties of totally positive kernels and matrices,
in Total Positivity and its Applications eds., C. A. Micchelli and M. Gasca,
Kluwer Acad. Publ., Dordrecht, 477–511.
Pinkus A. [1998], An interlacing property of eigenvalues of strictly totally positive
matrices, Lin. Alg. and Appl. 279, 201–206.
Pinkus A. [2008], Zero minors of totally positive matrices, Electronic J. Linear
Algebra 17, 532–542.
Pitman J. [1997], Probabilistic bounds on the coefficients of polynomials with
only real zeros, J. Comb. Theory, Ser. A 77, 279–303.
Pólya G. and G. Szegő [1976], Problems and Theorems in Analysis II, Springer–
Verlag, New York.
Rahman Q. I. and G. Schmeisser [2002], Analytic Theory of Polynomials, Oxford
University Press, Oxford.
Schoenberg I. J. [1930], Über variationsvermindernde lineare Transformationen,
Math. Z. 32, 321–328.
Schoenberg I. J. and A. Whitney [1951], A theorem on polygons in n dimensions
with application to variation-diminishing and cyclic variation-diminishing
linear transformations, Compositio Math. 9, 141–160.
Schumaker L. L. [1981], Spline Functions: Basic Theory, John Wiley & Sons, New
York.
Shapiro B. Z. and M. Z. Shapiro [1995], On the boundary of totally positive upper
triangular matrices, Lin. Alg. and Appl. 231, 105–109.
Shohat J. A. and J. D. Tamarkin [1943], The Problem of Moments,
Math. Surveys, Vol. 1, Amer. Math. Society, Providence.
Skandera M. [2004], Inequalities in products of minors of totally nonnegative
matrices, J. Alg. Comb. 20, 195–211.
Smith P. W. [1983], Truncation and factorization of bi-infinite matrices, in
Approximation Theory, IV (College Station, Tex., 1983), 257–289, Academic
Press, New York.
Stieltjes T. J. [1894–95], Recherches sur les fractions continues, Annales
Fac. Sciences Toulouse 8, 1–122; 9, 1–47.
Wagner D. [1992], Total positivity of Hadamard products, J. Math. Anal. Appl.
163, 459–483.
Wang Y. and Y.-N. Yeh [2005], Polynomials with real zeros and Pólya frequency
sequences, J. Comb. Theory, Ser. A 109, 63–74.
Whitney A. [1952], A reduction theorem for totally positive matrices, J. d’Anal.
Math. 2, 88–92.
 Author index

Aissen, M., 125, 174 Garloff, J., 86, 126, 176

Ando, T., x, 33, 86, 152, 153, 174 Gasca, M., 34, 74, 75, 168, 176–178
Askey, R., 170, 174 Gekhtman, M. I., 34, 35, 153, 175
Asner, B. A. Jr., 125, 174 Gladwell, G. M. L., 74, 75, 177
Gohberg, I., 171, 177
Beckenbach, E. F., 34, 174 Goodman, T. N. T., ix, 86, 125, 175, 177
Bellman, R., 34, 174 Gross, K. I., ix, 177
Berenstein, A., ix, 174 Gustafson, J. L., 125, 175
Boocher, A., 34, 35, 174
de Boor, C., ix, 34, 167, 170, 174
Brenti, F., ix, 125, 174 Holtz, O., 125, 177
Brown, L. D., 86, 174 Horn, R. A., 126, 177
Brualdi, R. A., 33, 175
Buslaev, A. P., 153, 175 Johnson, C. R., 34, 35, 126, 153, 175, 177
Johnstone, I. M., 86, 174
Carlson, B. C., 125, 175
Carlson, D., 34, 175
Karlin, S., ix, x, 33, 34, 74, 86, 125, 152,
Carnicer, J. M., 86, 175
153, 169, 173, 177, 178
Cavaretta, A. S. Jr., ix, 167, 175
Katkova, O. M., 75, 178
Crans, A. S., 126, 175
Kellogg, O. D., 33, 152, 178
Craven, T., 75, 175
Kemperman, J. H. B., 125, 178
Cryer, C., 74, 167, 175
Koev, P., ix, 168, 175, 178
Csordas, G., 75, 175
Koteljanskii, D. M., 34, 153, 178
Dahmen, W. A., ix, 167, 175 Krein, M. G., ix, x, 33, 34, 125, 152, 153,
Demmel, J., ix, 168, 175 169, 172, 176, 178
Dimitrov, D. K., 75, 175 Kurtz, D. C., 126, 178

Edrei, A., 125, 175 Loewner, K., 167, 178

Elias, U., 153, 175 Lusztig, G., ix, 178
Eveson, S. P., 153, 175

Fallat, S. M., 34, 35, 126, 153, 175 Mühlbach, G., 34, 178
Fan, K., 34, 86, 176 MacGibbon, K. B., 86, 174
Fekete, M., 74, 176 Maló, E., 126, 178
Fomin, S., ix, 168, 174, 176 Marden, M., 125, 178
Friedland, S., 153, 176 Markham, T. L., 126, 178
Froehle, B., 34, 35, 174 Marshall, A. W., ix, 178
Metelmann, K., 74, 167, 178
Gantmacher, F. R., ix, x, 33, 34, 125, Micchelli, C. A., ix, 34, 74, 167, 175–178
152, 153, 169, 172, 176 Motzkin, Th., 86, 178

180
 Author index 181

Nudel’man, A. A., x, 125, 178 Shohat, J. A., 125, 179

Skandera, M., 34, 35, 179
Olkin, I., ix, 178 Smith, P. W., ix, 167, 175, 179
Stieltjes, T. J., 102, 179
Pólya, G., 74, 125, 176, 179 Studden, W. J., x, 125, 169, 178
Peña, J. M., 34, 74, 75, 86, 168, 175, 177 Sun, Q., 125, 177
Pinkus, A., ix, 34, 74, 153, 167, 174, 175, Szegő, G., 125, 179
177–179
Pitman, J., 125, 179 Tamarkin, J. D., 125, 179

Rahman, Q. I., 125, 126, 179 Vishnyakova, A. M., 75, 178

Richards, D. St. P., ix, 177
Wagner, D., 126, 176, 179
Schmeisser, G., 125, 126, 179 Wang, Y., 125, 179
Schneider, H., 33, 175 Whitney, A. M., 34, 74, 86, 125, 167, 174,
Schoenberg, I. J., x, 33, 85, 86, 125, 169, 179
174, 179
Schumaker, L. L., ix, 179 Yeh, Y.-N., 125, 179
Shapiro, B. Z., 74, 179
Shapiro, M. Z., 74, 179 Zelevinsky, A., ix, 168, 174, 176
 Subject index

LDU factorization, 50 lower strictly totally positive, 47

lower totally positive, 47
almost strictly totally positive, 24, 34 lower triangular, 47

banded, 155 Maló’s Theorem, 123

Cauchy matrix, 92 oscillation matrix, 127

Cauchy–Binet formula, 2, 33
Chebyshev system, 88 Perron’s Theorem, 130
compound matrix, 2 pivot block, 4
principal minor, 5
Descartes system, 88 principal submatrix, 4
dispersion, 37
Schur Product Theorem, 123
eigenvalue interlacing, 140 shadow, 13
exponentials, 88 sign changes, 76
sign regular, 86
Fekete’s Lemma, 37 strictly sign regular, 86
strictly totally positive, 2
Gantmacher–Krein Theorem, 130 strictly totally positive kernel, 87
Gaussian polynomials, 110 Sylvester’s Determinant identity, 3
generalized Hadamard inequality, 24 Szasz’s inequality, 34
generalized Hurwitz matrix, 111
Grassman product, 3 Toeplitz matrix, 104, 123
Green’s matrix, 96, 121 totally positive, 2
totally positive kernel, 87
Hadamard inequality, 24, 34 triangular total positivity, 47
Hadamard product, 119
Hankel matrix, 101, 123 upper strictly totally positive, 47
Hurwitz matrix, 117, 124 upper totally positive, 47
Hurwitz polynomial, 117, 124 upper triangular, 47

Jacobi matrix, 97, 121 variation diminishing, 76, 85

Kronecker’s Theorem, 132 Whitney Theorem, 167

182