You are on page 1of 37

Bol. Soc. Mat. Mex.

DOI 10.1007/s40590-016-0088-8

ORIGINAL ARTICLE

The theory of locally Toeplitz sequences: a review, an


extension, and a few representative applications

Carlo Garoni1 · Stefano Serra-Capizzano2

Received: 7 August 2015 / Revised: 2 November 2015 / Accepted: 30 January 2016


© Sociedad Matemática Mexicana 2016

Abstract The theory of locally Toeplitz (LT) sequences is a powerful apparatus for
computing the asymptotic singular value and eigenvalue distribution of the discretiza-
tion matrices An arising from the numerical approximation of partial differential
equations (PDEs). Indeed, when the discretization parameter n tends to infinity, the
matrices An give rise to a sequence {An }n , which often can be expressed as a finite
sum of LT sequences. In this work, we review and extend the theory of LT sequences,
which dates back to the pioneering work by Tilli in 1998 and was partially developed
by the second author during the last decade. We also present some applications of the
theory to the finite difference and finite element approximation of PDEs.

Dedicated to Sergei M. Grudsky on his 60th birthday.

This work was supported by the Italian MIUR Program FIR 2013 through the Project DREAMS, by the
INdAM GNCS (Gruppo Nazionale per il Calcolo Scientifico), and by the Donation KAW 2013.0341 from
the Knut & Alice Wallenberg Foundation in collaboration with the Royal Swedish Academy of Sciences.

B Carlo Garoni
carlogaroni@hotmail.it
Stefano Serra-Capizzano
stefano.serrac@uninsubria.it

1 Department of Mathematics, University of Rome ‘Tor Vergata’, Via della Ricerca Scientifica 1,
00133 Rome, Italy
2 Department of Science and High Technology, University of Insubria, Via Valleggio 11,
22100 Como, Italy
C. Garoni, S. Serra-Capizzano

Keywords Locally Toeplitz sequence · Singular value and eigenvalue distribution ·


Discretization of partial differential equations · Approximating class of sequences

Mathematics Subject Classification 47B06 · 15B05 · 65N06 · 65N30

1 Introduction

We review and extend the theory of locally Toeplitz (LT) sequences, which stems
from Tilli’s work [26] and from the theory of classical Toeplitz operators [3–6,16,25,
27–29], and was partially developed in [21]. Besides being interesting in itself, the
theory of LT sequences is the fundamental cornerstone of the theory of generalized
locally Toeplitz (GLT) sequences [12,21,22]. In Sect. 1.1, we mention some of the
main applications of the theory of LT sequences. In Sect. 1.2, we summarize its main
features. In Sect. 1.3, we describe the main contributions of this work.

1.1 Applications of the theory of LT sequences

The theory of LT sequences is a powerful apparatus for computing the asymptotic


spectral distribution of the discretization matrices arising from the numerical approxi-
mation of partial differential equations (PDEs). Let us explain this point in more detail.
When discretizing a linear PDE by means of a linear numerical method, the compu-
tation of the numerical solution un reduces to solving a linear system An un = bn .
The size dn of this linear system grows when the discretization parameter n increases,
and what is often satisfied in practice is that the sequence of discretization matrices
{An }n enjoys an asymptotic spectral distribution as n → ∞. More precisely, it often
happens that, for a large set of test functions F (usually, for all continuous functions
with bounded support), the following limit relation holds:


1 
dn
1
lim F(λ j (An )) = F( f (x))dx, (1)
n→∞ dn μk (D) D
j=1

where λ j (An ), j = 1, . . . , dn , are the eigenvalues of An , μk is the Lebesgue measure


in Rk , and f : D ⊂ Rk → C is a measurable function. In this situation, f is referred
to as the spectral symbol of the sequence {An }n . The theory of LT sequences, in
combination with the result of [14, Theorem 3.4], is one of the most successful tools for
computing the spectral symbol f. Indeed, the sequence of discretization matrices {An }n
turns out to be a finite sum of LT sequences for many classes of PDEs and numerical
methods, especially when the numerical method belongs to the family of the so-called
‘local methods’. Local methods are, for example, finite difference (FD) methods,
collocation methods, and finite element (FE) methods with ‘locally supported’ basis
functions: in short, all standard numerical methods for the approximation of PDEs.
We refer the reader to [21] for applications of the theory of LT sequences in the context
of FD discretizations of PDEs, and to [8,10,11] for recent applications to the case of
B-spline isogeometric analysis (IgA) approximations of PDEs, both in the collocation
The theory of locally Toeplitz sequences

and Galerkin frameworks.1 It is worth pointing out that the cited papers invoke the
theory of GLT sequences more than the theory of LT sequences. However, a careful
analysis in the light of the results obtained herein shows that, actually, the applications
considered in [8,10,11,21] could be almost entirely handled by relying on the theory
of LT sequences alone, without even introducing GLT sequences.

1.2 Summary of the theory of LT sequences

We provide in this section a brief summary of the theory of LT sequences. Although


at this stage we cannot be rigorous, we nevertheless think that the following can give
an idea of what we are going to deal with.
Informally speaking, an LT sequence {An }n is a sequence of matrices with increas-
(1) ( p)
ing size, equipped with a function κ (the symbol). If {A n }n , . . . , {An }n are LT
p
sequences with symbols κ1 , . . . , κ p , then the function i=1 κi characterizes the
p (i)
asymptotic singular value distribution of { i=1 An }n ; in the case where the matri-
(i)
ces An are Hermitian, it also characterizes the asymptotic eigenvalue distribution of
{An }n . Three fundamental examples of LT sequences are:
(i) the sequence of multilevel Toeplitz matrices generated by a multivariate function
f in L 1 ;
(ii) the sequence of multilevel diagonal sampling matrices containing the evaluations
of a multivariate Riemann-integrable function a over uniform grids;
(iii) any zero-distributed sequence, i.e., any sequence of matrices with an asymptotic
singular value distribution characterized by the identically zero function.
The symbol of the LT sequence (i) is f , the symbol of the LT sequence (ii) is a,
and the symbol of any LT sequence of the form (iii) is 0. The set of LT sequences
possesses good algebraic properties. The most important of them asserts that, if {An }n
and { Ãn }n are LT sequences with symbols κ and κ̃, respectively, then {An Ãn }n is an
LT sequence with symbol κ κ̃ under mild assumptions on κ and κ̃. Finally, {An }n is an
LT sequence with symbol κ if and only if there exist LT sequences {{An(m) }n }m with
(m)
symbols {κm }m , such that κm converges to κ in a suitable way and {An }n ‘converges’
(m)
to {An }n (in the sense that {{An }n }m is an approximating class of sequences for
{An }n ; see Definition 2).

1.3 Contributions of this work

In this work, we focus on the mathematical foundations of the theory of LT sequences.


We first propose a slight (but relevant) modification of the original definition of sepa-
rable locally Toeplitz (sLT) sequences appearing in [21,22], and we introduce for the
first time the notion of LT sequences in the multivariate/multilevel setting. With the
new definitions, which are based on the notion of approximating classes of sequences
[13,20], we are able to enlarge the applicability of the theory of LT sequences, by gen-
eralizing and/or simplifying a lot of key results. In particular, we show that the product

1 IgA is a modern paradigm for analyzing problems governed by PDEs; see [7].
C. Garoni, S. Serra-Capizzano

of two LT sequences with symbols κ(x, θ ) = a(x) f (θ) and κ̃(x, θ ) = ã(x) f˜(θ) is
an LT sequence with symbol κ(x, θ )κ̃(x, θ ) = a(x)ã(x) f (θ ) f˜(θ), under the only
assumption that the two ‘generating’ functions f, f˜ are conjugate (i.e., one in L p and
the other in L q , where p, q are conjugate exponents, that is, 1/ p + 1/q = 1). In this
way, we remove from [21, Theorem 5.3] both the assumption that f, f˜ are in L ∞ and
the technical condition in [21, eq. (41)]. In addition, we provide a formal proof of the
fact that the sequences of matrices mentioned in items (i)–(iii) of Sect. 1.2 fall in the
class of LT sequences; different (incomplete) versions of this result appeared in many
previous papers, but only partial proofs were given. Finally, we also provide some
illustrative applications of the theory of LT sequences in the context of the numerical
approximation of PDEs.
The paper is organized as follows. In Sect. 2 we collect all the necessary preliminar-
ies. Section 3 focuses on the notion of approximating classes of sequences. In Sect. 4,
we develop the theory of LT sequences. In Sect. 5, we present some applications. We
draw conclusions in Sect. 6.

2 Mathematical background

2.1 Notation and terminology

– Om and Im denote, respectively, the m × m zero matrix and the m × m identity


matrix. Sometimes, when the size m can be inferred from the context, O and I are
used instead of Om and Im .
– Given a m × m matrix X ∈ Cm×m , the eigenvalues and the singular values of X
are denoted by λ j (X ), j = 1, . . . , m, and σ j (X ), j = 1, . . . , m, respectively.
– If 1 ≤ p ≤ ∞, |·| p denotes both the p-norm of vectors and the associated operator
norm for matrices. The norm | · |2 is the spectral norm and will be denoted by  · .
– Given X ∈ Cm×m and 1 ≤ p ≤ ∞, X  p denotes the Schatten p-norm of X ,
which is defined as the p-norm of the vector (σ1 (X ), . . . , σm (X )); see [2]. The
Schatten 1-norm is also called the trace-norm.

– (X ) is the real part of the (square) matrix X , i.e., (X ) = X +X 2 , where X is

the conjugate transpose of X .


– If wi : Di → C, i = 1, . . . , d, are arbitrary functions, we denote by w1 ⊗ · · · ⊗
wd : D1 × · · · × Dd → C the tensor-product function

(w1 ⊗ · · · ⊗wd )(ξ1 , . . . , ξd ) = w1 (ξ1 ) · · · wd (ξd ), (ξ1 , . . . , ξd ) ∈ D1 × · · · × Dd .

– A function a : [0, 1]d → C is said to be Riemann-integrable if its real and


imaginary parts (a), (a) : [0, 1]d → R are Riemann-integrable in the classical
sense. Recall that any Riemann-integrable function is bounded.
– If g : D → C, we set g∞ = supξ ∈D |g(ξ )|. Clearly, g∞ < ∞ if and only if
g is bounded over D.
– μk is the Lebesgue measure in Rk . Throughout this work, all the terminology
coming from measure theory (such as ‘measure’, ‘measurable’, ‘a.e.’, ‘in L p ’,
etc.) is always referred to as the Lebesgue measure.
The theory of locally Toeplitz sequences

– If f : D ⊆ Rk → C is in L p (D) and the domain D is clear  from the context, we


denote by  f  L p the L p -norm of f ; that is,  f  L p = ( D | f | p )1/ p for 1 ≤ p <
∞, and  f  L ∞ = ess sup D | f | for p = ∞.
– We use a notation borrowed from probability theory to indicate sets. For example,
if f, g : D ⊆ Rk → C, then { f = 1} = {x ∈ D : f (x) = 1}, μk { f > 0, g < 0}
is the measure of the set {x ∈ D : f (x) > 0, g(x) < 0}, etc.
– If E 1 , . . . , E d ⊆ R are measurable sets and f : E 1 × · · · × E d → C, we say that
f is separable if there exist measurable functions f i : E i → C, i = 1, . . . , d,
such that f = f 1 ⊗ · · · ⊗ f d .
– A d-variate trigonometric polynomial is a finite linear combination of the Fourier
frequencies ei j ·θ , j ∈ Zd . Here, θ = (θ1 , . . . , θd ) ∈ Rd and j · θ = i=1 d
ji θi .

Multi-index notation A multi-index i ∈ Zd , also called a d-index, is just a vector


in Zd ; its components are denoted by i 1 , . . . , i d . 0 and 1 are the vectors of all zeros
and all ones (their size will be clear from the context). For any d-index m, we set
N (m) = dj=1 m j and write m → ∞ to indicate that min(m) → ∞. If h, k are
d-indices, h ≤ k means that h r ≤ kr for all r = 1, . . . , d. In this case, we define the
multi-index range h, . . . , k as the set { j ∈ Zd : h ≤ j ≤ k}. We assume for h, . . . , k
the lexicographic ordering:
 

. . . [ ( j1 , . . . , jd ) ] jd =h d ,...,kd j =h ,...,k ... . (2)


d−1 d−1 d−1 j1 =h 1 ,...,k1

For instance, if d = 2 the ordering (2) is

(h 1 , h 2 ), (h 1 , h 2 + 1), . . . , (h 1 , k2 ), (h 1 + 1, h 2 ), (h 1 + 1, h 2 + 1), . . . ,
(h 1 + 1, k2 ), . . . . . . . . . , (k1 , h 2 ), (k1 , h 2 + 1), . . . , (k1 , k2 ).

When a multi-index j varies over h, . . . , k (this is often written as j = h, . . . , k),


it is understood that j varies from h to k following the ordering (2). For instance,
if m ∈ Nd and x = [x i ]m i=1 , then x is a vector of size N (m), whose components
x i , i = 1, . . . , m, are ordered in accordance with (2). Similarly, if X = [x i j ]m i, j =1 ,
then X is a N (m)× N (m) matrix whose components are indexed by two d-indices i , j ,

both varying from 1 to m according to (2). If h ≤ k, kj =h denotes the summation over
all j = h, . . . , k. Operations involving d-indices that have no meaning in the vector
space Zd must always be interpreted in the componentwise sense. For instance, i j =
(i 1 j1 , . . . , i d jd ), i/ j = (i 1 /j1 , . . . , i d /jd ), n/m = (n 1 /m 1 , . . . , n d /m d ), etc.
Multilevel diagonal sampling matrices For n ∈ Nd and a : [0, 1]d → C, the d-level
diagonal sampling matrix Dn (a) is the following diagonal matrix of size N (n):

i
Dn (a) = diag a ,
i=1,...,n n

where we recall that i varies from 1 to n according to the lexicographic ordering (2).
Matrix sequences Throughout this paper, by a sequence of matrices (or matrix
sequence) we mean a sequence of the form {A n }n , where:
C. Garoni, S. Serra-Capizzano

– n varies in some infinite subset of N;


– n = n(n) is a d-index in Nd which depends on n, and n → ∞ as n → ∞;
– A n is a square matrix of size N (n).
Recall that n → ∞ means min(n) → ∞. Unless otherwise stated, the multi-index
that parameterizes a matrix sequence is always assumed to be a d-index.

2.2 Preliminaries on linear algebra and matrix analysis

If X ∈ Cm×m , we know from the singular value decomposition (SVD) that rank(X )
is the number of nonzero singular values of X and X  = σmax (X ) = X ∞ . Hence,


m
X 1 = σi (X ) ≤ rank(X )X  ≤ mX , X ∈ Cm×m . (3)
i=1

Another useful trace-norm inequality is the following:



m
X 1 ≤ |xi j |, X ∈ Cm×m . (4)
i, j=1

The proof is simple. Let X = U V ∗ be an SVD of X . Then, setting Q = V U ∗ , the


matrix Q is unitary and

X 1 = trace() = trace(U ∗ X V ) = trace(X Q)



m 
m 
m 
m 
m 
m
≤ |xik qki | ≤ max |qki | |xik | ≤ |xik |.
k=1,...,m
i=1 k=1 i=1 k=1 i=1 k=1

If 1 ≤ p, q ≤ ∞ are conjugate exponents, the following Hölder-type inequality holds


for the Schatten norms [2]:

X Y 1 ≤ X  p Y q , X, Y ∈ Cm×m . (5)

We also recall the classical spectral norm inequality in [15, Corollary 2.3.2]:

X  ≤ |X |1 |X |∞ , X ∈ Cm×m . (6)

This inequality is particularly useful to estimate the spectral norm of a matrix when
we have upper bounds for its components. Indeed, we recall that |X |1 (resp., |X |∞ ) is
the maximum among the 1-norms of the columns (resp., rows) of X .
Tensor products and direct sums If X, Y are matrices of any dimension, say X ∈
Cm 1 ×m 2 and Y ∈ C1 ×2 , the tensor (Kronecker) product of X and Y is the m 1 1 ×m 2 2
matrix defined by
The theory of locally Toeplitz sequences

⎡ ⎤
x11 Y · · · x1m 2 Y
 ⎢ ⎥
X ⊗ Y = xi j Y i=1,...,m 1 = ⎣ ... ..
. ⎦,
j=1,...,m 2
xm 1 1 Y · · · xm 1 m 2 Y

and the direct sum of X and Y is the (m 1 + 1 ) × (m 2 + 2 ) matrix defined by


 
X O
X ⊕ Y = diag(X, Y ) = .
O Y

Tensor products and direct sums possess a lot of good algebraic properties.
(i) Associativity, which allows one to write X 1 ⊗ · · · ⊗ X d and X 1 ⊕ · · · ⊕ X d ,
without parentheses.
(ii) The relations (X 1 ⊗Y1 )(X 2 ⊗Y2 ) = (X 1 X 2 )⊗(Y1 Y2 ) and (X 1 ⊕Y1 )(X 2 ⊕Y2 ) =
(X 1 X 2 ) ⊕ (Y1 Y2 ) hold whenever X 1 , X 2 can be multiplied and Y1 , Y2 can also
be multiplied.
(iii) For all matrices X, Y, (X ⊗ Y )∗ = X ∗ ⊗ Y ∗ , (X ⊕ Y )∗ = X ∗ ⊕ Y ∗ and
(X ⊗ Y )T = X T ⊗ Y T , (X ⊕ Y )T = X T ⊕ Y T .
(iv) Bilinearity (of tensor products): (α1 X 1 + α2 X 2 ) ⊗ (β1 Y1 + β2 Y2 ) = α1 β1 (X 1 ⊗
Y1 )+α1 β2 (X 1 ⊗Y2 )+α2 β1 (X 2 ⊗Y1 )+α2 β2 (X 2 ⊗Y2 ) for all α1 , α2 , β1 , β2 ∈ C
and for all matrices X 1 , X 2 , Y1 , Y2 such that X 1 , X 2 are summable and Y1 , Y2
are also summable.
From (i)–(iv), other interesting properties follow. For example, if X, Y are invertible,
then X ⊗ Y is invertible with inverse X −1 ⊗ Y −1 . If X, Y are normal (resp., Hermitian,
symmetric, unitary), then X ⊗ Y is also normal (resp., Hermitian, symmetric, unitary).
If X ∈ Cm×m and Y ∈ C× , the eigenvalues and singular values of X ⊗ Y (resp.,
X ⊕ Y ) are {λi (X )λ j (Y ) : i = 1, . . . , m, j = 1, . . . , } and {σi (X )σ j (Y ) : i =
1, . . . , m, j = 1, . . . , } (resp., {λi (X ) : i = 1, . . . , m}∪{λ j (Y ) : j = 1, . . . , } and
{σi (X ) : i = 1, . . . , m} ∪ {σ j (Y ) : j = 1, . . . , }). In particular, for all X ∈ Cm×m
and Y ∈ C× ,
 
X ⊕ Y  p = (X  p , Y  p ) p , X ⊗ Y  p = X  p Y  p , 1 ≤ p ≤ ∞. (7)

2.3 Singular value and eigenvalue distribution of a matrix sequence

If K is either R or C, we denote by Cc (K) the space of complex-valued continuous


functions defined on K and with bounded support.
Definition 1 Let {A n }n be a matrix sequence and let f : D ⊂ Rk → C be a measur-
able function defined on a set D with 0 < μk (D) < ∞.
– We say that {A n }n has an asymptotic singular value distribution described by f
if, for all F ∈ Cc (R),

N (n) 
1  1
lim F(σ j (A n )) = F(| f (x)|)dx. (8)
n→∞ N (n) μk (D) D
j=1
C. Garoni, S. Serra-Capizzano

In this case, f is referred to as the singular value symbol of {A n }n and we write


{A n }n ∼σ f .
– We say that {A n }n has an asymptotic eigenvalue (or spectral) distribution described
by f if, for all F ∈ Cc (C),

N (n) 
1  1
lim F(λ j (A n )) = F( f (x))dx. (9)
n→∞ N (n) μk (D) D
j=1

In this case, f is referred to as the eigenvalue (or spectral) symbol of {A n }n , and


we write {A n }n ∼λ f.

When we write a relation such as {A n }n ∼σ f or {A n }n ∼λ f , it is understood


that {A n }n is a matrix sequence and f is as in Definition 1; that is, f is a measurable
function defined on a subset D of some Rk with 0 < μk (D) < ∞. Sometimes, for
brevity, we will write {A n }n ∼σ, λ f to indicate that {A n }n ∼σ f and {A n }n ∼λ f.
Zero-distributed sequences A class of matrix sequences that plays an important role in
the theory of LT sequences is the class of zero-distributed sequences. A zero-distributed
sequence is a matrix sequence {Z n }n such that {Z n }n ∼σ 0.

Theorem 1 The following conditions for a matrix sequence {Z n }n are equivalent.


1. {Z n }n ∼σ 0.
#{ j ∈ {1, . . . , N (n)} : σ j (Z n ) > }
2. For every > 0, lim = 0.
n→∞ N (n)
rank(Rn )
3. For every n, we have Z n = Rn + Nn , where lim = lim Nn  = 0.
n→∞ N (n) n→∞

Proof (1 ⇒ 2) For every > 0, take F ∈ Cc (R) such that F = 1 over [0, /2],
F = 0 over [ , ∞) and 0 ≤ F ≤ 1 over [0, ∞). Since {Z n }n ∼σ 0, we have

#{ j ∈ {1, . . . , N (n)} : σ j (Z n ) > } #{ j ∈ {1, . . . , N (n)} : σ j (Z n ) ≤ }


=1−
N (n) N (n)
N (n)
1 
≤1− F (σ j (Z n )),
N (n)
j=1

which converges to 1 − F (0) = 0 as n → ∞.


(2 ⇒ 3) By assumption, for every > 0 the quantity

#{ j ∈ {1, . . . , N (n)} : σ j (Z n ) > }


qn ( ) =
N (n)

tends to 0 as n → ∞. Hence, there exists a sequence { n }n of positive numbers such


that, when n → ∞,

#{ j ∈ {1, . . . , N (n)} : σ j (Z n ) > n }


n → 0, qn ( n ) = → 0.
N (n)
The theory of locally Toeplitz sequences

Let Z n = Un n Vn∗ be an SVD of Z n . Let  ˆ n be the matrix obtained from n by


setting to 0 all the singular values of Z n that are less than or equal to n , and let
˜ n = n −  ˆ n be the matrix obtained from n by setting to 0 all the singular values
of Z n that exceed n . Then,

ˆ n Vn∗ + Un 
Z n = Un n Vn∗ = Un  ˜ n Vn∗ = Rn + Nn ,

where Rn = Un  ˆ n Vn∗ and Nn = Un  ˜ n Vn∗ . Since rank(Rn )/N (n) = qn ( n ) and


Nn  ≤ n tend to 0 as n → ∞, the thesis is proved.
(3 ⇒ 1) To give an elegant and short proof of this implication, we use some results
from Sect. 3. The assumption in item 3 ensures that {{O N (n) }n }m is an approximating
class of sequences for {Z n }n according to Definition 2. Moreover, it is clear that
{O N (n) }n ∼σ 0. Hence, {Z n }n ∼σ 0 by Theorem 4. 

With the terminology of clustering [14, p. 86], condition 2 in Theorem 1 is expressed
by saying that {Z n }n is weakly clustered at 0 in the sense of the singular values.

2.4 Multilevel Toeplitz matrices

Given n ∈ Nd , a matrix of the form [a i− j ]ni, j =1 ∈ C N (n)×N (n) , whose (i, j )th entry
depends only on the difference between the d-indices i and j , is called a multilevel
Toeplitz matrix (or, more precisely, a d-level Toeplitz matrix). If f : [−π, π ]d → C
is a function in L 1 ([−π, π ]d ), we denote its Fourier coefficients by

1
fk = f (θ)e−ik·θ dθ , k ∈ Zd . (10)
(2π )d [−π,π ]d

The nth multilevel Toeplitz matrix associated with f is defined as

Tn ( f ) = [ f i− j ]ni, j =1 . (11)

We call {Tn ( f )}n∈Nd the family of multilevel Toeplitz matrices associated with f ,
which, in turn, is called the generating function of {Tn ( f )}n∈Nd . For each fixed n ∈ Nd ,
the application Tn (·) : L 1 ([−π, π ]d ) → C N (n)×N (n) is linear:

Tn (α f + βg) = αTn ( f ) + βTn (g), α, β ∈ C, f, g ∈ L 1 ([−π, π ]d ).

Moreover, Tn (1) = I N (n) . For every f ∈ L 1 ([−π, π ]d ) and n ∈ Nd , it can be shown


that Tn ( f )∗ = Tn ( f ). Hence, if f is real (or a.e. real),2 , then all the matrices Tn ( f )
are Hermitian. If f 1 , . . . , f d ∈ L 1 ([−π, π ]) and n ∈ Nd , then f 1 ⊗ · · · ⊗ f d ∈
L 1 ([−π, π ]d ) and, by [12, Lemma 2.10],

Tn ( f 1 ⊗ · · · ⊗ f d ) = Tn 1 ( f 1 ) ⊗ · · · ⊗ Tn d ( f d ). (12)

2 Note that two functions f, g ∈ L 1 ([−π, π ]d ) which coincide a.e. give rise to the same multilevel Toeplitz
matrices Tn ( f ) = Tn (g), n ∈ Nd , because the Fourier coefficients of f and g coincide.
C. Garoni, S. Serra-Capizzano

Theorem 2 is the multilevel version of Szegő’s first limit theorem and of the Avram–
Parter theorem in the form proved by Tilli [25]. This theorem is a fundamental result
on multilevel Toeplitz matrices. For the eigenvalues, it goes back to Szegő [16], and
for the singular values it was established by Avram [1] and Parter [17]. They assumed
that d = 1 and f ∈ L ∞ . See Sections 5 and 6 of [5] and also Section 10.14 of [6]
for more on the subject in the case of L ∞ generating functions. The extension to
d > 1 and f ∈ L 1 was performed by Tyrtyshnikov and Zamarashkin [27–29] and
Tilli [25]. Tilli’s proof is remarkably lucid. We also refer the reader to [13] for a proof
of Theorem 2 based on the notion of approximating classes of sequences (see Sect. 3);
the proof in [13] is made only in the case of eigenvalues for d = 1, but the argument
is general and can be extended to singular values and to higher dimensions.

Theorem 2 If f ∈ L 1 ([−π, π ]d ) and {Tn ( f )}n is any matrix sequence extracted from
{Tn ( f )}n∈Nd , then {Tn ( f )}n ∼σ f . If moreover f is real a.e., then {Tn ( f )}n ∼λ f .

Important inequalities involving Toeplitz matrices and Schatten p-norms can be


found in [23, Corollary 4.2]. We report them in the next theorem for future use.

Theorem 3 Let f ∈ L p ([−π, π ]d ) and n ∈ Nd . Then,

N (n)1/ p
Tn ( f ) p ≤  f  L p , 1 ≤ p ≤ ∞, (13)
(2π )d/ p

where it is understood that 1/∞ = d/∞ = 0.

Lemma 1 generalizes [9, Proposition 2] and will be used in Sect. 4.1 to study the
LT operator. By Hölder’s inequality [19], if f ∈ L p ([−π, π ]d ), g ∈ L q ([−π, π ]d )
and p, q ∈ [1, ∞] are conjugate exponents, then f g ∈ L 1 ([−π, π ]d ). In this case,
we can consider the three matrices Tn ( f ), Tn (g) and Tn ( f g).

Lemma 1 Let f ∈ L p ([−π, π ]d ) and g ∈ L q ([−π, π ]d ), where 1 ≤ p, q ≤ ∞ are


conjugate exponents. Then,

Tn ( f )Tn (g) − Tn ( f g)1


lim = 0. (14)
n→∞ N (n)

Proof If f, g are in L ∞ ([−π, π ]d ), Eq. (14) holds by [9, Proposition 2]. For the
general case where f ∈ L p ([−π, π ]d ) and g ∈ L q ([−π, π ]d ), take two sequences
{ f m }m and {gm }m such that f m , gm ∈ L ∞ ([−π, π ]d ), f m → f in L p ([−π, π ]d ) and
gm → g in L q ([−π, π ]d ). By the linearity of Tn (·) and the inequalities (5), (13), for
every m and every n ∈ Nd , we have
The theory of locally Toeplitz sequences

Tn ( f )Tn (g) − Tn ( f g)1


≤ Tn ( f − f m )Tn (g)1 + Tn ( f m )Tn (g − gm )1
+ Tn ( f m )Tn (gm ) − Tn ( f m gm )1 + Tn ( f m gm − f g)1
≤ N (n)1/ p  f − f m  L p N (n)1/q g L q + N (n)1/ p  f m  L p N (n)1/q g − gm  L q
+ Tn ( f m )Tn (gm ) − Tn ( f m gm )1 + N (n) f m gm − f g L 1

≤ N (n)  f − f m  L p g L q + sup  f i  L p g − gm  L q
i

Tn ( f m )Tn (gm ) − Tn ( f m gm )1
+ +  f m gm − f g L 1 . (15)
N (n)

Note that supi  f i  L p < ∞, because f i → f in L p ([−π, π ]d ). By [9, Proposition 2],


since f m , gm ∈ L ∞ ([−π, π ]d ) we have

Tn ( f m )Tn (gm ) − Tn ( f m gm )1


lim = 0.
n→∞ N (n)

Hence, dividing (15) by N (n) and passing to the limit as n → ∞, we obtain

Tn ( f )Tn (g) − Tn ( f g)1


lim sup
n→∞ N (n)
≤  f − f m  L p g L q + sup  f i  L p g − gm  L q +  f m gm − f g L 1 . (16)
i

This relation holds for every m. When m → ∞, f m → f in L p and gm → g in L q by


construction. Moreover, f m gm → f g in L 1 ([−π, π ]d ) by Hölder’s inequality, since

 f g − f m gm  L 1 ≤ ( f − f m )g L 1 +  f m (g − gm ) L 1
≤  f − f m  L p g L q +  f m  L p g − gm  L q
≤  f − f m  L p g L q + sup  f i  L p g − gm  L q .
i

Thus, passing to the limit as m → ∞ in (16), we get the thesis. 




3 Approximating classes of sequences

In this section, we introduce the fundamental notion on which the theory of LT


sequences is based: the notion of approximating classes of sequences, which orig-
inally appeared in [20].
Definition 2 (a.c.s.) Let {A n }n be a matrix sequence. An approximating class of
sequences (a.c.s.) for {A n }n is a sequence of matrix sequences {{Bn,m }n }m with the
following property: for every m there exists n m , such that, for n ≥ n m ,

A n = Bn,m + Rn,m + Nn,m ,


(17)
rank(Rn,m ) ≤ c(m)N (n), Nn,m  ≤ ω(m),
C. Garoni, S. Serra-Capizzano

where n m , c(m), ω(m) depend only on m, and lim c(m) = lim ω(m) = 0.
m→∞ m→∞

Roughly speaking, {{Bn,m }n }m is an a.c.s. for {A n }n if, for large m, A n is eventu-


ally equal to Bn,m plus a small-rank matrix (with respect to the matrix size N (n)),
plus a small-norm matrix. In the following, we abbreviate ‘approximating classes of
sequences’ with ‘a.c.s.es’.

3.1 A.c.s.es as a tool for computing singular value and eigenvalue distributions

The importance of a.c.s.es resides in Theorems 4 and 5, which appeared in [13,20].


These theorems provide general tools for computing the singular value (resp., eigen-
value) distribution of a ‘difficult’ matrix sequence {A n }n from the singular value
(resp., eigenvalue) distributions of ‘simpler’ matrix sequences {Bn,m }n that ‘approx-
imate’ {A n }n as m → ∞ in the sense of Definition 2. For the proof of Theorems 4
and 5, see [12, Section 3.1].
Theorem 4 Let {A n }n be a matrix sequence. Assume that:
1. {{Bn,m }n }m is an a.c.s. for {A n }n ;
1  N (n)
2. for every m and every F ∈ Cc (R), lim F(σ j (Bn,m )) = φm (F) ∈
n→∞ N (n) j=1
C;
3. for every F ∈ Cc (R), lim φm (F) = φ(F) ∈ C.
m→∞
1  N (n)
Then, for every F ∈ Cc (R), lim F(σ j (A n )) = φ(F).
n→∞ N (n) j=1

Theorem 5 Let {A n }n be a sequence of Hermitian matrices. Assume that:


1. {{Bn,m }n }m is an a.c.s. for {A n }n formed by Hermitian matrices;
1  N (n)
2. for every m and every F ∈ Cc (C), lim F(λ j (Bn,m )) = φm (F) ∈
n→∞ N (n) j=1
C;
3. for every F ∈ Cc (C), lim φm (F) = φ(F) ∈ C.
m→∞
1  N (n)
Then, for every F ∈ Cc (C), lim F(λ j (A n )) = φ(F).
n→∞ N (n) j=1

3.2 The a.c.s. algebra

In this section, we report the important algebraic properties possessed by the a.c.s.es.
The properties collected in Proposition 1 are direct consequences of Definition 2.
 } } be a.c.s.es for {A } and {A } ,
Proposition 1 Let {{Bn,m }n }m and {{Bn,m n m n n n n
respectively. Then:
∗ } } is an a.c.s. for {A∗ } ;
1. {{Bn,m n m n n
 } } is an a.c.s. for {α A + β A } , for all α, β ∈ C.
2. {{α Bn,m + β Bn,m n m n n n
The theory of locally Toeplitz sequences

Proposition 1 addresses the case of a linear combination {α A n + β An }n of two


matrix sequences {A n }n and {An }n . We would like to have an analogous result for the
product {A n An }n . This requires an additional (mild) assumption on {A n }n and {An }n ,
namely that {A n }n and {An }n are sparsely unbounded.

Definition 3 (s.u. matrix sequence) We say that {A n }n is sparsely unbounded (s.u.) if


for every M > 0 there exists n M such that, for n ≥ n M ,

#{i ∈ {1, . . . , N (n)} : σi (A n ) > M}


≤ r (M),
N (n)

where lim M→∞ r (M) = 0.

It is not difficult to see that any matrix sequence enjoying an asymptotic singular
value distribution is s.u. For a formal proof of this result, see [12, Proposition 3.3].
Proposition 2 If {A n }n ∼σ f, then {A n }n is s.u.
Proposition 3 is the analog of Proposition 1 for the case of the product of two
a.c.s.es. This important result appeared for the first time in [20]. For the proof, we
refer the reader to either [12, Proposition 3.4] or [20, Proposition 2.4].
Proposition 3 Let {{Bn,m }n }m and {{Bn,m  } } be a.c.s.es for {A } and {A } ,
n m n n n n
  } } is an a.c.s.
respectively. Assume that {A n }n and {A n }n are s.u. Then {{Bn,m Bn,m n m
for {A n An }n .

3.3 A criterion to identify a.c.s.es

In practical applications, it often happens that a matrix sequence {A n }n is given together


with a sequence of matrix sequences {{Bn,m }n }m , and one would like to show that
{{Bn,m }n }m is an a.c.s. for {A n }n , without constructing the splitting (17). In this section,
we provide a useful criterion to solve this problem.
p
Lemma 2 Let C be a square matrix of size s. Suppose that C p ≤ s for some
p ∈ [1, ∞) and ≥ 0. Then we can express C in the form
1 1
C = R + N , rank(R) ≤ p+1 s, N  ≤ p+1 . (18)
p s
Proof Since C p = i=1 σi (C)
p
≤ s, the number of singular values of C that
1 1
exceed p+1 cannot be larger than p+1 s. Let C = U V ∗ be an SVD of C and write

ˆ ∗ + U V
C = U V ∗ = U V ˜ ∗,

where ˆ is obtained from  by setting to 0 all the singular values that are less than or
1
˜ = −
equal to p+1 , while  ˆ is obtained from  by setting to 0 all the singular
1
ˆ ∗ and
values that exceed p+1 . Then, the conditions in (18) are met with R = U V
N = U V˜ ∗. 

C. Garoni, S. Serra-Capizzano

Theorem 6 Let {A n }n be a matrix sequence, {{Bn,m }n }m a sequence of matrix


sequences and 1 ≤ p < ∞. Suppose that for every m there exists n m such that,
for n ≥ n m ,
p
A n − Bn,m  p ≤ (m, n)N (n),

where lim lim sup (m, n) = 0. Then {{Bn,m }n }m is an a.c.s. for {A n }n .


m→∞ n→∞

Proof By Lemma 2, for every m and every n ≥ n m , we have

A n − Bn,m = Rn,m + Nn,m ,


1 1
rank(Rn,m ) ≤ (m, n) p+1 N (n), Nn,m  ≤ (m, n) p+1 .

Let

(m) = lim sup (m, n).


n→∞

By definition of lim sup, for every m there exists n m such that, for n ≥ n m ,

1
(m, n) ≤ (m) + .
m

Setting n̂ m = max(n m , n m ), for every m and every n ≥ n̂ m , we have

A n − Bn,m = Rn,m + Nn,m ,


1 1
1 p+1 1 p+1
rank(Rn,m ) ≤ (m) + N (n), Nn,m  ≤ (m) + .
m m

Since (m) → 0 by assumption, {{Bn,m }n }m is an a.c.s. of {A n }n . 




3.4 An extension of the concept of a.c.s.es

We provide in this section an extension of the definition of a.c.s. that will be used
to define LT sequences in Sect. 4. The extension is plain. The underlying idea is
that, in Definition 2, one could choose to approximate {A n }n by a class of sequences
{{Bn,α }n }α∈A parameterized by a not necessarily integer parameter α. For example,
one may want to use a parameter > 0 and to claim that a given class of sequences
{{Bn, }n } >0 is an a.c.s. for {A n }n when → 0. Intuitively, this assertion should have
the following meaning: for every > 0, there exists n such that, for n ≥ n ,

A n = Bn, + Rn, + Nn, ,


rank(Rn, ) ≤ c( )N (n), Nn,  ≤ ω( ),
The theory of locally Toeplitz sequences

where n , c( ), ω( ) depend only on , and lim →0 c( ) = lim →0 ω( ) = 0. Actu-


ally, this is the correct meaning, which leads to the definition of a.c.s. parameterized
by a positive tending to 0; see [12, Definition 3.3]. However, for the definition of
LT sequences, we do not need the concept of a.c.s.es parameterized by a positive
→ 0. On the contrary, we need the concept of a.c.s.es parameterized by a multi-
index m → ∞. Since any m ∈ N is a special multi-index, the definition of a.c.s.
parameterized by a multi-index m → ∞ (Definition 4) is a true extension of Defin-
ition 2. In the following, a multi-index sequence of matrix sequences is any class of
sequences of the form {{Bn,m }n }m∈M , where:
1. M ⊆ Nq for some q ≥ 1, and M ∩ {i ∈ Nq : i ≥ h} = ∅ for every h ∈ Nq .
We express the latter condition by saying that ∞ is an accumulation point for M.
This is required to ensure that m can tend to ∞ inside M;
2. for every m ∈ M, {Bn,m }n is a matrix sequence.

Definition 4 (a.c.s.) Let {A n }n be a matrix sequence. An a.c.s. for {A n }n is a multi-


index sequence of matrix sequences {{Bn,m }n }m∈M with the following property: for
every m ∈ M there exists n m such that, for n ≥ n m ,

A n = Bn,m + Rn,m + Nn,m ,


(19)
rank(Rn,m ) ≤ c(m)N (n), Nn,m  ≤ ω(m),

where n m , c(m), ω(m) depend only on m, and lim c(m) = lim ω(m) = 0.
m→∞ m→∞

As already pointed out, Definition 4 extends the classical definition of a.c.s. (Defi-
nition 2). Indeed, a classical a.c.s. {{Bn,m }n }m for {A n }n is an a.c.s. also in the sense
of  4 (take M as the infinite subset of N where m varies). Moreover,
 Definition   if
{Bn,m }n m∈M is an a.c.s. for {A n }n (in the sense of Definition 4), then {Bn,m }n m
is an a.c.s. for {A n }n (in the sense of the classical Definition 2) for all sequences of
multi-indices {m = m(m)}m ⊆ M such that m → ∞ when m → ∞.
Remark 1 Let {{Bn,m }n }m∈M , {{Bn,m  } } 
n m∈M be a.c.s.es for {A n }n , {A n }n , respec-
tively. Then, the following results hold.
∗ } }
1. {{Bn,m ∗
n m∈M is an a.c.s. for {A n }n .
2. {{α Bn,m + β Bn,m }n }m∈M is an a.c.s. for {α A n + β An }n , for all α, β ∈ C.


3. If {A n }n and {An }n are s.u., then {{Bn,m Bn,m


 } } 
n m∈M is an a.c.s. for {A n A n }n .

These results are proved in the same way as the analogous results for standard a.c.s.es
stated in Propositions 1 and 3.

4 Locally Toeplitz sequences

In this section, we develop the theory of LT sequences. We first introduce and analyze
the LT operator in Sect. 4.1. Then, in Sect. 4.2, we define the LT and sLT sequences.
In Sect. 4.3, we prove that the sequences of multilevel diagonal sampling matrices
associated with a Riemann-integrable function, the sequences of multilevel Toeplitz
matrices generated by an L 1 function, and the zero-distributed sequences that fall in
C. Garoni, S. Serra-Capizzano

the class of LT sequences. In Sects. 4.4 and 4.5, we study, respectively, the spectral
properties and the algebraic properties of LT sequences. In Sect. 4.6, we present some
characterizations of LT sequences. Finally, in Sect. 4.7, we summarize the theory of
LT sequences.

4.1 The locally Toeplitz operator LTnm (a, f )

Definition 5 – Let m, n ∈ N, a : [0, 1] → C, and f : [−π, π ] → C in


L 1 ([−π, π ]). Then, we define the n × n matrix

L Tnm (a, f ) = Dm (a) ⊗ Tn/m ( f ) ⊕ On mod m


 
j
= diag a Tn/m ( f ) ⊕ On mod m .
j=1,...,m m

It is understood that L Tnm (a, f ) = On when n < m and that the term On mod m
is not present when n is a multiple of m. Moreover, here and in the following, the
tensor product operation ⊗ is always applied before the direct sum ⊕, exactly as
in the case of numbers, where multiplication is always applied before addition.
– Let m, n ∈ Nd , a : [0, 1]d → C and f 1 , . . . , f d : [−π, π ] → C in L 1 ([−π, π ]).
Then, we define the N (n) × N (n) matrix

1 ,...,m d (a(x , . . . , x ), f ⊗ · · · ⊗ f )
L Tnm (a, f 1 ⊗ · · · ⊗ f d ) = L Tnm1 ,...,n 1 d 1 d
d

2 ,...,m d
j1
= diag Tn 1 /m 1  ( f 1 ) ⊗ L Tnm2 ,...,n d
a , x 2 , . . . , x d , f 2 ⊗ · · · ⊗ f d
j1 =1,...,m 1 m1
⊕ O(n 1 mod m 1 )n 2 ···n d .

This is a recursive definition, whose base case has been considered in the previous
item.
In this section, we investigate the properties of the LT operator L Tnm (a, f ). We
prefer to write L Tnm (a, f ) instead of L Tnm (a, f 1 ⊗ · · · ⊗ f d ), because we are going
to see that L Tnm (a, f ) is well defined (in a unique way) for any f ∈ L 1 ([−π, π ]d );
see Definition 6. The main result about L Tnm (a, f ) is Theorem 7. This result allows
us to extend the definition of the LT operator as in Definition 6 and to define (in
Sect. 4.2) the notion of LT sequences in the multilevel setting. It is worth noting that
such a notion is introduced here for the first time, because, so far, LT sequences were
considered in [26] only in the unilevel case, whereas the multilevel setting addressed
in [21,22] only deals with sLT sequences. The proof of Theorem 7 is a matter of
tensor-product/direct-sum manipulations and is rather technical. For this reason, we
decided not to include it here; the interested reader is referred to [12, Theorem 4.1].
Theorem 7 For any m, n ∈ Nd , there exists a permutation matrix nm of size N (n)
such that

L Tnm (a, f 1 ⊗ · · · ⊗ f d ) = nm Dm (a) ⊗ Tn/m ( f 1 ⊗ · · · ⊗ f d ) ⊕ O (nm )T
The theory of locally Toeplitz sequences

for every a : [0, 1]d → C and every f 1 , . . . , f d ∈ L 1 ([−π, π ]).

Definition 6 (LT operator) Let m, n ∈ Nd , a : [0, 1]d → C and f ∈ L 1 ([−π, π ]d ).


Then, we define

L Tnm (a, f ) = nm Dm (a) ⊗ Tn/m ( f ) ⊕ O (nm )T ,

where nm is the permutation matrix appearing in Theorem 7.

Remark 2 It is clear that L Tnm (a, f ) = L Tnm (a, g) whenever f = g a.e. Furthermore,
if f = f 1 ⊗ · · · ⊗ f d a.e., with f 1 , . . . , f d ∈ L 1 ([−π, π ]), then L Tnm (a, f ) is equal
to L Tnm (a, f 1 ⊗ · · · ⊗ f d ), as defined by Definition 5.

We now study in some detail the properties of L Tnm (a, f ). We first note that, for
any n, m ∈ Nd , any a : [0, 1]d → C and any f ∈ L 1 ([−π, π ]d ),
 ∗
L Tnm (a, f ) = L Tnm (a, f ). (20)

This follows from Definition 6, from the relations (X ⊗Y )∗ = X ∗ ⊗Y ∗ and (X ⊕Y )∗ =


X ∗ ⊕ Y ∗ , and from the identity Tk ( f )∗ = Tk ( f ).

Proposition 4 Let m, n ∈ Nd , a : [0, 1]d → C and f ∈ L 1 ([−π, π ]d ). Then,


  
 m 
 j 
L Tnm (a, f ) p =  a  Tn/m ( f ) p , 1 ≤ p ≤ ∞. (21)
 m j =1  p

Proof Use Definition 6, the invariance of  ·  p by unitary transformations (such as


permutations), and the Eq. (7). 


We denote by C[0,1] the vector space of all functions a : [0, 1]d → C.


d

Proposition 5 Let m, n ∈ Nd . Then, the operator

L Tnm (a, ·) : L 1 ([−π, π ]d ) → C N (n)×N (n)

is linear for any a : [0, 1]d → C, and the operator

L Tnm (·, f ) : C[0,1] → C N (n)×N (n)


d

is linear for any f ∈ L 1 ([−π, π ]d ).

Proof Use Definition 6, the linearity of the operators Dm (·) and Tn/m (·) and the
bilinearity of tensor produtcs. 


In Proposition 6, we show that L Tnm (a, f )L Tnm (ã, f˜) is ‘close’ to L Tnm (a ã, f f˜),
as long as a, ã are bounded and f, f˜ are conjugate.
C. Garoni, S. Serra-Capizzano

Proposition 6 Let a, ã : [0, 1]d → C be bounded, and let f ∈ L p ([−π, π ]d ) and


f˜ ∈ L q ([−π, π ]d ), where 1 ≤ p, q ≤ ∞ are conjugate exponents. Then, for any
n, m ∈ Nd ,
 
 m 
 L Tn (a, f )L Tnm (ã, f˜) − L Tnm (a ã, f f˜) ≤ (n/m) N (n), (22)
1

where

Tk ( f )Tk ( f˜) − Tk ( f f˜)1


(k) = a ã∞
N (k)

and lim (k) = 0 by Lemma 1. In particular, for every m ∈ Nd there exists n m ∈ Nd


k→∞
such that, for n ≥ nm ,

  N (n)
 m 
 L Tn (a, f )L Tnm (ã, f˜) − L Tnm (a ã, f f˜) ≤ (23)
1 N (m)

and

L Tnm (a, f )L Tnm (ã, f˜) = L Tnm (a ã, f f˜) + Rn,m + Nn,m ,
(24)
rank(Rn,m ) ≤ √NN(n) (m)
, Nn,m  ≤ √ N1(m) .

Proof By Definition 6 and the properties of tensor products and direct sums,

L Tnm (a, f )L Tnm (ã, f˜) − L Tnm (a ã, f f˜)


  

= nm Dm (a ã) ⊗ Tn/m ( f )Tn/m ( f˜) − Tn/m ( f f˜) ⊕ O (nm )T .

Hence,

L Tnm (a, f )L Tnm (ã, f˜) − L Tnm (a ã, f f˜)1


= Dm (a ã)1 Tn/m ( f )Tn/m ( f˜) − Tn/m ( f f˜)1
Tn/m ( f )Tn/m ( f˜) − Tn/m ( f f˜)1
≤ N (n)a ã∞ ,
N (n/m)

and (22) is proved. Since lim k→∞ (k) = 0, for every m ∈ Nd there exists n m ∈ Nd
such that, for n ≥ n m , (23) holds. (24) follows from (23) and Lemma 2. 


Theorems 8 and 9 provide information about the  pasymptotic singular value and
eigenvalue distribution of a finite sum of the form i=1 L Tnm (ai , f i ). Together with
Theorems 4 and 5, they play a central role in the computation of the singular value
and eigenvalue distribution of a finite sum of LT sequences.
The theory of locally Toeplitz sequences

Theorem 8 Let a1 , . . . , a p : [0, 1]d → C and let f 1 , . . . , f p ∈ L 1 ([−π, π ]d ). Then,


for every m ∈ Nd and every F ∈ Cc (R),

N (n)
  p 
1  
lim F σr L Tn (ai , f i )
m
n→∞ N (n)
r =1 i=1
 p 
1  1
m   j 
 
= φm (F) = F  ai f i (θ )  dθ . (25)
N (m) (2π )d [−π,π ]d  m 
j =1 i=1

Moreover, if a1 , . . . , a p are Riemann-integrable, then, for every F ∈ Cc (R),

  p 
 
1  
lim φm (F) = φ(F) = F  ai (x) f i (θ ) dxdθ . (26)
m→∞ (2π )d [0,1]d ×[−π,π ]d  
i=1

Proof By Definition 6,

 p   p 
 
(nm )T L Tnm (ai , f i ) nm = Dm (ai ) ⊗ Tn/m ( f i ) ⊕ O. (27)
i=1 i=1

Recalling that Dm (ai ) = diag j =1,...,m ai ( j /m), for every j = 1, . . . , m the j th


diagonal block of size N (n/m) of the matrix (27) is given by
 p 

p
j  j
ai Tn/m ( f i ) = Tn/m ai fi .
m m
i=1 i=1

p
It follows that the singular values of i=1 L Tnm (ai , f i ) are

  p 
 j
σk Tn/m ai fi , k = 1, . . . , N (n/m), j = 1, . . . , m,
m
i=1

plus further N (n) − N (m)N (n/m) singular values equal to 0. Therefore, by Theo-
rem 2, for any F ∈ Cc (R), we have

N (n)
  p 
1  
lim F σr L Tn (ai , f i )
m
n→∞ N (n)
r =1 i=1

N (m)N (n/m)
= lim ⎣
n→∞ N (n)
C. Garoni, S. Serra-Capizzano

N (n/m)
   p ⎤
1   
m
1 j ⎦
· F σk Tn/m ai fi
N (m) N (n/m) m
j =1 k=1 i=1
 p 
1  1
m   j 
 
= F  ai f i (θ )  dθ. (28)
N (m) (2π )d [−π,π]d  m 
j =1 i=1

This proves (25).  p


If a1 , . . . , a p are Riemann-integrable, then the function x → F  i=1 ai (x)

f i (θ) is Riemann-integrable for each fixed θ ∈ [−π, π ]d , being the composition of
a continuous function with a Riemann-integrable function. Hence,
 p    p 
1 
m  j   
   
lim F  ai f i (θ ) = F  ai (x) f i (θ ) dx.
m→∞ N (m)  m  [0,1]d  
j =1 i=1 i=1

Passing to the limit as m → ∞ in (28), and using the dominated convergence theorem,
we get (26). 


Theorem 9 Let a1 , . . . , a p : [0, 1]d → C and f 1 , . . . , f p ∈ L 1 ([−π, π ]d ). Then,


for every m ∈ Nd and every F ∈ Cc (C),

N (n)
   p 
1  
lim F λr  L Tn (ai , f i )
m
n→∞ N (n)
r =1 i=1
  p 
1  1
m   j
= φm (F) = F  ai f i (θ ) dθ . (29)
N (m) (2π )d [−π,π ]d m
j =1 i=1

Moreover, if a1 , . . . , a p are Riemann-integrable, then, for every F ∈ Cc (C),


   p 
1 
lim φm (F) = φ(F) = F  ai (x) f i (θ ) dxdθ .
m→∞ (2π )d [0,1]d ×[−π,π ]d
i=1
(30)

Proof The proof is completely analogous to the proof of Theorem 8. For the sake of
brevity, we omit the details and refer the reader to [12, Theorem 4.3]. 


4.2 Definition of LT and sLT sequences

We remind that, unless otherwise specified, the multi-index that parameterizes a matrix
sequence {A n }n is always assumed to be a d-index, n = (n 1 , . . . , n d ).

Definition 7 (LT sequence) Let {A n }n be a matrix sequence, n ∈ Nd , and let a :


[0, 1]d → C be Riemann-integrable and f ∈ L 1 ([−π, π ]d ). We say that {A n }n is a
The theory of locally Toeplitz sequences

 
locally Toeplitz (LT) sequence with symbol a ⊗ f if {L Tnm (a, f )}n m∈Nd is an a.c.s.
of {A n }n . This means that, for all m ∈ Nd there is n m such that, for n ≥ n m ,

A n = L Tnm (a, f ) + Rn,m + Nn,m ,


(31)
rank(Rn,m ) ≤ c(m)N (n), Nn,m  ≤ ω(m),

where n m , c(m), ω(m) do not depend on n, and lim m→∞ c(m) = lim m→∞ ω(m) =
0. In this case, we write {A n }n ∼LT a ⊗ f . The functions a and f are, respectively,
the weight function and the generating function of {A n }n .3
Definition 8 (sLT sequence) Let {A n }n be a matrix sequence, n ∈ Nd . We say that
{A n }n is a separable locally Toeplitz (sLT) sequence if {A n }n ∼LT a ⊗ f for some
Riemann-integrable function a : [0, 1]d → C and some separable function f ∈
L 1 ([−π, π ]d ). In this case, we write {A n }n ∼sLT a ⊗ f .
It is clear from the definition that an sLT sequence is just an LT sequence with
separable generating function. From now on, if we write {A n }n ∼LT a ⊗ f (resp.,
{A n }n ∼sLT a ⊗ f ), it is understood that {A n }n is a matrix sequence, a : [0, 1]d → C
is Riemann-integrable and f ∈ L 1 ([−π, π ]d ) (resp., f ∈ L 1 ([−π, π ]d ) is separable).

4.3 Zero-distributed sequences, sequences of multilevel diagonal sampling


matrices and sequences of multilevel Toeplitz matrices

Three fundamental examples of LT sequences were indicated in items (i)–(iii) of


Sect. 1.2. Actually, the matrix sequences in (i)–(iii) may be regarded as the building
blocks of the theory of LT sequences. In this section, we prove that they are indeed
LT sequences. Recall that, according to our terminology, ∞ is an accumulation point
for a subset M ⊆ Nq if M ∩ {i ∈ Nq : i ≥ h} = ∅ for all h ∈ Nq .
Theorem 10 Let {Z n }n be a matrix sequence. Let M be a subset of some Nq such that
∞ is an accumulation point for M. Then, the following conditions are equivalent:
1. {Z n }n ∼σ 0.
2. {{O N (n) }n }m∈M is an a.c.s. for {Z n }n .
3. {{Z n }n }m∈M is an a.c.s. for {O N (n) }n .
In particular, {Z n }n is zero distributed if and only if {Z n }n ∼sLT 0.
Proof The equivalence 2 ⇔ 3 is obvious from Definition 4.
Let us prove that 1 ⇒ 2. By Theorem 1, we can write Z n = Rn + Nn for all n, where
limn→∞ rank(Rn )/N (n) = limn→∞ Nn  = 0. It follows that {{O N (n) }n }m∈M is an
a.c.s. for {Z n }n . To see this, take, in (19), Rn,m = Rn , Nn,m = Nn , c(m) and
ω(m) any two positive functions of m that converge to 0 as m → ∞ (for instance,
c(m) = ω(m) = 1/ min(m)), and n m any integer such that rank(Rn )/N (n) ≤ c(m)
and Nn  ≤ ω(m) for n ≥ n m .

3 We refer the reader to the introduction of Tilli’s paper [26] for the origin and the meaning of this termi-
nology.
C. Garoni, S. Serra-Capizzano

Let us now prove that 2 ⇒ 1. By assumption, {{O N (n) }n }m∈M is an a.c.s. for {Z n }n .
Hence, if we take any sequence {m = m(m)}m ⊆ M such that m → ∞ as m → ∞,
{{O N (n) }n }m is a (classical) a.c.s. for {Z n }n . Moreover, it is clear that {O N (n) }n ∼σ 0.
Hence, {Z n }n ∼σ 0 by Theorem 4.
The fact that {Z n }n is zero distributed if and only if {Z n }n ∼sLT 0 follows from
the equivalence 1 ⇔ 2 (applied with M = Nd ) and from the simple observation that
{{O N (n) }n }m∈Nd = {{L Tnm (0, 0)}n }m∈Nd and 0 = 0 ⊗ 0. 


Theorem 11 If a : [0, 1]d → C is Riemann-integrable and {Dn (a)}n is any matrix


sequence extracted from {Dn (a)}n∈Nd , then {Dn (a)}n ∼sLT a ⊗ 1.

Theorem 12 If f ∈ L 1 ([−π, π ]d ) and {Tn ( f )}n is any matrix sequence extracted


from {Tn ( f )}n∈Nd , then {Tn ( f )}n ∼LT 1 ⊗ f .

The proofs of Theorems 11 and 12 are rather technical. For this reason, we decided
not to report them here and to refer the reader to [12, Theorems 4.5–4.6].

4.4 Singular value and eigenvalue distribution of a finite sum of LT sequences

Theorem 13 provides the singular value distribution of a finite sum of LT sequences.


(i) p (i)
Theorem 13 Suppose that {A n }n ∼LT ai ⊗ f i , i = 1, . . . , p; then, { i=1 A n }n ∼σ
p
i=1 ai ⊗ f i .

Proof Choose a sequence {m = m(m)}m ⊆ Nd such that m → ∞ when m → ∞.


From the properties of a.c.s.es (see Proposition 1), and from the definition of LT
p p (i)
sequences, we know that {{ i=1 L Tnm (ai , f i )}n }m is an a.c.s. for { i=1 A n }n . By
Theorem 8, for every F ∈ Cc (R) we have

N (n)
  p 
1  
lim F σr L Tn (ai , f i )
m
= φm (F),
n→∞ N (n)
r =1 i=1
  p 
 
1  
lim φm ( f ) = φ(F) = F  ai (x) f i (θ )  dxdθ .
m→∞ (2π )d [0,1]d ×[−π,π ]d  
i=1

p (i) p
Hence, by Theorem 4, { i=1 A n }n ∼σ i=1 ai ⊗ fi . 


Using Theorem 13, we show in Proposition 7 that the symbol of an LT sequence


is essentially unique. Afterward, in Proposition 8, we show that the symbol of an LT
sequence formed by Hermitian matrices is real a.e.
Proposition 7 If {A n }n ∼LT a ⊗ f and {A n }n ∼LT ã ⊗ f˜, then a ⊗ f = ã ⊗ f˜ a.e.

Proof By Theorem 13, {O N (n) }n ∼σ a ⊗ f − ã ⊗ f˜. Hence, for every F ∈ Cc (R),



1
F(0) = F(|a(x) f (θ ) − ã(x) f˜(θ )|)dxdθ . (32)
(2π )d [0,1]d ×[−π,π ]d
The theory of locally Toeplitz sequences

If we assume by contradiction that a ⊗ f does not coincide a.e. with ã ⊗ f˜, then
μ2d {|a ⊗ f − ã ⊗ f˜| ≥ } > 0 for some > 0. Choose a function F ∈ Cc (R) such
that F(0) = 1, 0 ≤ F ≤ 1 and F = 0 over [ , ∞). For this F, Eq. (32) cannot hold,
because

F(|a(x) f (θ) − ã(x) f˜(θ )|)dxdθ
[0,1]d ×[−π,π ]d
≤ μ2d {|a ⊗ f − ã ⊗ f˜| < } = (2π )d − μ2d {|a ⊗ f − ã ⊗ f˜| ≥ } < (2π )d .

This contradiction shows that a ⊗ f = ã ⊗ f˜ a.e. 



Proposition 8 Let {A n }n ∼LT a ⊗ f and assume that the matrices A n are Hermitian.
Then a ⊗ f ∈ R a.e.
Proof It holds in general that {A n }n ∼LT a ⊗ f implies that {A∗n }n ∼LT a ⊗ f . This
result follows immediately from the definition of LT sequences and from (20). If the
matrices A n are Hermitian, then, by Proposition 7, we obtain a ⊗ f = a ⊗ f a.e., i.e.,
a ⊗ f ∈ R a.e. 

Theorem 14 provides the spectral distribution of a finite sum of LT sequences
(formed by Hermitian matrices).
(i) p (i)
Theorem 14 Suppose that {A n }n ∼LT ai ⊗ f i , i = 1, . . . , p; then {( i=1 A n )}n
p (i)
∼λ ( i=1 ai ⊗ f i ). In particular, if all the matrices A n are Hermitian,
p (i) p
{ i=1 A n }n ∼λ i=1 ai ⊗ f i .
Proof Choose a sequence {m = m(m)}m ⊆ Nd such that m → ∞ when
m → ∞. From the properties of a.c.s.es (see p Proposition 1), and from the defi-
nition of LT sequences, we know that {{( i=1 L Tnm (ai , f i ))}n }m is an a.c.s. for
p (i)
{( i=1 A n )}n . By Theorem 9, for every F ∈ Cc (C), we have

N (n)
   p 
1  
lim F λr  L Tn (ai , f i )
m
= φm (F),
n→∞ N (n)
r =1 i=1
   p 
1 
lim φm ( f ) = φ(F) = F  ai (x) f i (θ ) dxdθ .
m→∞ (2π )d [0,1]d ×[−π,π ]d
i=1

p p
Hence, by Theorem 5, {( i=1 A(i) n )}n ∼λ ( i=1 ai ⊗ f i ). To conclude the proof,
p p
we note that, if all the matrices A(i) (i)
n are Hermitian, then ( i=1 A n ) =
(i)
i=1 A n
p p
and ( i=1 ai ⊗ f i ) = i=1 ai ⊗ f i a.e. (by Proposition 8). 


4.5 Algebraic properties of LT sequences

Propositions 9 and 10 collect the most elementary algebraic properties of LT sequences,


which follow from Definition 2, Eq. (20), and the linearity of the LT operator (Propo-
sition 5).
C. Garoni, S. Serra-Capizzano

Proposition 9 Suppose that {A n }n ∼LT a ⊗ f . Then {α A n }n ∼LT αa ⊗ f for all


α ∈ C and {A∗n }n ∼LT a ⊗ f .
(i)  (i)
Proposition 10 If {A n }n ∼LT a ⊗ f i , i = 1, . . . , r , then { ri=1 A n }n ∼LT a ⊗
r 
(i=1 f i ). Similarly, if {A(i) r (i)
n }n ∼LT ai ⊗ f , i = 1, . . . , r , then { i=1 A n }n ∼LT
r
( i=1 ai ) ⊗ f .
In Theorem 15, we show, under mild assumptions, that the product of LT sequences
is again an LT sequence with symbol given by the product of the symbols. Theorem 15
extends the result obtained in [21, Theorem 5.3], by removing both the assumptions
that f, f˜ are in L ∞ and the technical hypothesis in [21, eq. (41)].
Theorem 15 Let {A n }n ∼LT a ⊗ f and { Ã n }n ∼LT ã ⊗ f˜, where f ∈ L p ([−π, π ]d ),
f˜ ∈ L q ([−π, π ]d ), and 1 ≤ p, q ≤ ∞ are conjugate exponents. Then, {A n à n }n ∼LT
a ã ⊗ f f˜.
Proof By Theorem 13 and Proposition 2, every LT sequence is s.u., so {A n }n and { Ã n }n
are s.u. Since {{L Tnm (a, f )}n }m∈Nd and {{L Tnm (ã, f˜)}n }m∈Nd are a.c.s.es for {A n }n
and { à n }n , respectively, {{L Tnm (a, f )L Tnm (ã, f˜)}n }m∈Nd is an a.c.s. for {A n à n }n
(Remark 1). The thesis follows from Definition 7 and Proposition 6 (see (24)). 

As a consequence of Theorem 15 and Theorems 11 and 12, we immediately obtain
the following result.
Theorem 16 Let a : [0, 1]d → C be Riemann-integrable, f ∈ L 1 ([−π, π ]d ) and
{Dn (a)Tn ( f )}n be any matrix sequence extracted from {Dn (a)Tn ( f )}n∈Nd ; then,
{Dn (a)Tn ( f )}n ∼LT a ⊗ f .

4.6 Characterizations of LT sequences

Theorem 16 shows that, for any a, f as in Definition 7, there always exists an LT


sequence {A n }n ∼LT a ⊗ f (take A n = Dn (a)Tn ( f )). Theorem 17 shows that the
sequences of the form {Dn (a)Tn ( f )}n play a central role in the world of LT sequences.
Indeed, {A n }n ∼LT a ⊗ f, if and only if A n equals Dn (a)Tn ( f ) up to a small-rank
plus small-norm correction. In fact, any LT sequence {A n }n ∼LT a ⊗ f admits the
fixed matrix sequence {Dn (a)Tn ( f )}n as an a.c.s., and, vice versa, any sequence {A n }n
admitting {Dn (a)Tn ( f )}n as an a.c.s. is an LT sequence with symbol a ⊗ f . For the
proof of Theorem 17, see [12, Theorem 4.9].
Theorem 17 Let {A n }n be a matrix sequence, a : [0, 1]d → C be a Riemann-
integrable function, and f ∈ L 1 ([−π, π ]d ). Then, the following conditions are
equivalent.
1. {A n }n ∼LT a ⊗ f .
2. For all sequences {am }m , { f m }m , {{A(m)
n }n }m with the following properties:
– am : [0, 1]d → C is Riemann-integrable and am → a in L 1 ([0, 1]d );
– f m ∈ L 1 ([−π, π ]d ) and f m → f in L 1 ([−π, π ]d );
(m)
– {A n }n ∼LT am ⊗ f m ;
The theory of locally Toeplitz sequences

(m)
it holds that {{A n }n }m is an a.c.s. for {A n }n .
3. There exist sequences {am }m , { f m }m such that:
– am : [0, 1]d → C is continuous, am ∞ ≤ a L ∞ for all m and am → a a.e.;
– f m : [−π, π ]d → C is a d-variate trigonometric polynomial and f m → f
a.e. and in L 1 ([−π, π ]d );
– {{Dn (am )Tn ( f m )}n }m is an a.c.s. for {A n }n .
(m)
4. There exist sequences {am }m , { f m }m , {{A n }n }m such that:
– am : [0, 1] → C is Riemann-integrable and am → a in L 1 ([0, 1]d );
d

– f m ∈ L 1 ([−π, π ]d ) and f m → f in L 1 ([−π, π ]d );


(m) (m)
– {A n }n ∼LT am ⊗ f m and {{A n }n }m is an a.c.s. for {A n }n .
5. {{Dn (a)Tn ( f )}n }m is an a.c.s. for {A n }n .
6. For every n, we have A n = Dn (a)Tn ( f ) + Z n , where {Z n }n is zero distributed.

Remark 3 Theorem 17 continues to hold if f is assumed to be separable, item 1 is


replaced by ‘{A n }n ∼sLT a ⊗ f ’, and we add in item 3 the requirement that each f m
is separable; see [12, Remark 4.3].

4.7 Summary of the theory of LT sequences

After developing the theory of LT sequences, it is worth emphasizing its main features
in view of the applications. These main features are collected in the following items.
p p
LT 1 If {A(i) (i)
n }n ∼LT ai ⊗ f i , i = 1, . . . , p, then { i=1 A n }n ∼σ i=1 ai ⊗ f i . If
(i)  p (i) p
moreover the matrices A n are Hermitian, then { i=1 A n }n ∼λ i=1 ai ⊗ f i .
LT 2 {Tn ( f )}n ∼LT 1 ⊗ f for every f ∈ L 1 ([−π, π ]d ).
LT 3 {Dn (a)}n ∼LT a ⊗ 1 for every Riemann-integrable function a : [0, 1]d → C.
LT 4 {Z n }n ∼LT 0 if and only if {Z n }n ∼σ 0.
LT 5 If {A n }n ∼LT a ⊗ f , { Ã n }n ∼LT ã ⊗ f˜ and f ∈ L p ([−π, π ]d ), f˜ ∈
L q ([−π, π ]d ), with 1 ≤ p, q ≤ ∞ conjugate exponents, then {A n à n }n ∼LT
a ã ⊗ f f˜.
LT 6 Let a : [0, 1]d → C Riemann-integrable and f ∈ L 1 ([−π, π ]d ). Then,
(m)
{A n }n ∼LT a ⊗ f if and only if there are LT sequences {A n }n ∼LT am ⊗ f m
such that am → a in L 1 ([0, 1]d ), f m → f in L 1 ([−π, π ]d ), and {{A(m) n }n }m
is an a.c.s. for {A n }n .
We emphasize that LT 6 implies {Bn + Z n }n ∼LT a ⊗ f whenever {Bn }n ∼LT a ⊗ f
and {Z n }n is zero distributed. To see this, apply LT 6 with A n = Bn + Z n , A(m)
n = Bn ,
am = a and f m = f , and observe that {Bn }n is an a.c.s. for {Bn + Z n }n (by the
equivalence 1 ⇔ 3 in Theorem 1).

5 Applications

In this section, we outline a few applications of the theory of LT sequences, to give a


flavor of its applicative interest. The idea is to show that items LT 1–LT 6 are powerful
tools for computing the singular value and spectral distribution of PDE discretization
C. Garoni, S. Serra-Capizzano

matrices. For more applications, we refer the reader to Sect. 1.1, where specific pointers
to the literature were provided.

5.1 FD discretization of one-dimensional elliptic PDEs

Consider the following second-order elliptic PDE with Dirichlet boundary conditions:

−(a(x)u  (x)) = f (x), x ∈ (0, 1),


(33)
u(0) = u(1) = 0,

where f : [0, 1] → R is continuous and a : [0, 1] → R is a function in C 1 ([0, 1])


such that a(x) > 0 for every x ∈ [0, 1]. We consider the discretization of (33) by the
classical second-order central FD scheme. In the case where a(x) is constant, this is
also known as the (−1, 2, −1) FD scheme. Let us describe it shortly; for more details
on FD methods, we refer the reader to the available literature (see, e.g., [24] or any
good book on FDs). We choose a discretization parameter n ∈ N, set h = n+1 1
and
x j = j h for all j ∈ [0, n +1], and we note that, for j = 1, . . . , n, we can approximate
(a(x)u  (x)) |x=x j by the following FD formula:

a(x j+ 1 )u  (x j+ 1 ) − a(x j− 1 )u  (x j− 1 )
− (a(x)u  (x)) |x=x j ≈ − 2 2 2 2
h
u(x j+1 ) − u(x j ) u(x j ) − u(x j−1 )
a(x j+ 1 ) − a(x j− 1 )
≈−
2 h 2 h
 h 
−a(x j+ 1 )u(x j+1 ) + a(x j+ 1 ) + a(x j− 1 ) u(x j ) − a(x j− 1 )u(x j−1 )
= 2 2 2 2
. (34)
h2
Then, we approximate the solution of (33) by the piecewise linear function that takes
the values u j in x j for j = 0, . . . , n +1, where u 0 = u n+1 = 0 and u = (u 1 , . . . , u n )T
is the solution of the linear system
 
−a(x j+ 1 )u j+1 + a(x j+ 1 ) + a(x j− 1 ) u j − a(x j− 1 )u j−1
2 2 2 2

= h f (x j ),
2
j = 1, . . . , n. (35)

The matrix of the linear system (35) is the tridiagonal symmetric matrix
⎡ ⎤
a(x 1 ) + a(x 3 ) −a(x 3 )
⎢ −a(x
2
)
2 2
3 ) + a(x 5 ) −a(x 5 )

⎢ 3 a(x ⎥
⎢ 2 2 2 2 ⎥
⎢ .. .. ⎥
⎢ −a(x ) . . ⎥
An = ⎢ 5 ⎥.
⎢ 2 ⎥
⎢ .. .. ⎥
⎢ . . −a(xn− 1 ) ⎥
⎣ 2

−a(xn− 1 ) a(xn− 1 ) + a(xn+ 1 )
2 2 2
(36)
The theory of locally Toeplitz sequences

In this example, we will see that the theory of LT sequences allows us to compute the
singular value and eigenvalue distribution of the sequence of discretization matrices
{An }n . Actually, this is the fundamental example that led to the birth of the theory
of LT sequences [26] and, subsequently, of GLT sequences [12,21,22]. Given the
importance, we will compute the singular value and eigenvalue distribution of {An }n
by two different methods, both of them paradigmatic.
Method 1 If a(x) is constant, say a(x) = 1 identically, we have An = Tn (2 − 2 cos θ ).
Using Theorem 2, we immediately get {An }n ∼σ, λ 2 − 2 cos θ . Actually, this can also
be obtained by direct computation, considering that the singular values and eigen-

values of Tn (2 − 2 cos θ ) are given explicitly by 2 − 2 cos n+1 , j = 1, . . . , n; see
[4, p. 35] or [24, p. 154]. If a(x) is not constant, the expression of An is given by
(36) and the Toeplitzness seems to be completely lost. In reality, we find it again
‘in an approximated sense’ and ‘at a local scale’. Indeed, we note that a(x) varies
smoothly from a(0) to a(1). Therefore, assuming that n is large with respect to k,
any k × k leading principal submatrix of An shows an approximate Toeplitz struc-
ture. Let us be more quantitative. Fix a large m ∈ N and assume n > m. Then,
n is large with respect to n/m and, so, according to the previous reasoning, any
n/m × n/m leading principal submatrix of An shows an approximate Toeplitz
structure. In fact, the evaluations of a(x) appearing in the first n/m × n/m lead-
ing principal submatrix of An are approximately equal to a( m1 ); the evaluations of
a(x) appearing in the second n/m × n/m leading principal submatrix of An are
approximately equal to a( m2 ) and so on until the evaluations of a(x) appearing in the
mth n/m × n/m leading principal submatrix of An , which are approximately
equal to a(1). If, for all j = 1, . . . , m, we replace by a( mj ) the evaluations of a(x) in
the jth n/m × n/m leading principal submatrix of An , this submatrix becomes
a( mj )Tn/m (2−2 cos θ ). In conclusion, the matrix An is approximated by the LT oper-

ator L Tnm (a(x), 2 − 2 cos θ ) = diag j=1,...,m a( mj )Tn/m (2 − 2 cos θ ) ⊕ On mod m .
In fact, {{L Tnm (a(x), 2 − 2 cos θ )}n }m is an a.c.s. for {An }n , because it can be shown
that

An = L Tnm (a(x), 2 − 2 cos θ ) + Rn,m + Nn,m



1 m+1
rank(Rn,m ) ≤ 3m, Nn,m  ≤ ωa + ,
m n+1

with ωa (·) being the modulus of continuity of a. Thus, by definition,

{An }n ∼LT a(x)(2 − 2 cos θ ), (37)

and so

{An }n ∼σ, λ a(x)(2 − 2 cos θ ). (38)

Method 2 As already pointed out, the example we are dealing with led to the birth
of the theory of LT sequences. In particular, the procedure followed in Method 1 to
obtain (38) motivated the definition of LT sequences, as well as the introduction of the
C. Garoni, S. Serra-Capizzano

LT operator. However, now that we have fully developed the theory of LT sequences,
we should say that Method 1 is not the most effective way to obtain (38). The method
we are going to see is by far more effective.
Let x̂ j = nj , j = 1, . . . , n and note that |x j − x̂ j | ≤ n+1
1
= h for all j = 1, . . . , n.
Consider the matrix
⎡ ⎤
2a(x̂1 ) −a(x̂1 )
⎢−a(x̂2 ) 2a(x̂2 ) −a(x̂2 ) ⎥
⎢ ⎥
⎢ . . . . ⎥
Dn (a)Tn (2 − 2 cos θ ) = ⎢ ⎢ −a(x̂3 ) . . ⎥ . (39)

⎢ .. .. ⎥
⎣ . . −a(x̂n−1 )⎦
−a(x̂n ) 2a(x̂n )

By comparing (39) and (36), we see that the modulus of each entry of the matrix
Z n = An − Dn (a)Tn (2 − 2 cos θ ) is bounded by 2ωa (3h/2). It follows that the 1-
norm and the ∞-norm of Z n are bounded by 6ωa (3h/2), and so Z n  ≤ 6ωa (3h/2)
by (6). In particular, Z n  → 0 and {Z n }n ∼σ 0 (Theorem 1). By LT 2, LT 3 and
LT 5 we have {Dn (a)Tn (2 − 2 cos θ )}n ∼LT a(x)(2 − 2 cos θ ). Since

An = Dn (a)Tn (2 − 2 cos θ ) + Z n , (40)

by LT 6 we get (37) and by LT 1 we get (38).


Remark 4 Problem (33) can be rewritten in the form

−a(x)u  (x) − a  (x)u  (x) = f (x), x ∈ (0, 1),


(41)
u(0) = u(1) = 0.

From this reformulation, it appears more clearly that the symbol a(x)(2 − 2 cos θ )
consists of two ‘ingredients’:
– the coefficient a(x) of the higher-order differential operator in (41);
– the trigonometric polynomial 2 − 2 cos θ = −eiθ + 2 − e−iθ associated with the
FD formula (−1, 2, −1) used to approximate the higher-order derivative −u  (x).
In particular, the term −a  (x)u  (x), which only depends on lower-order derivatives of
u(x), does not enter the expression of the symbol.

Remark 5 Suppose that we add to the diffusion equation (33) a convection and a
reaction term. In this way, we obtain the following reaction–convection–diffusion
PDE:

−(a(x)u  (x)) + β(x)u  (x) + γ (x)u(x) = f (x), x ∈ (0, 1),


(42)
u(0) = u(1) = 0,

where we assume that β, γ : [0, 1] → R are bounded. Based on the discussion in


Remark 4, we expect that the term β(x)u  (x) + γ (x)u(x), which only involves lower-
order derivatives of u(x), does not enter the expression of the symbol. In other words,
The theory of locally Toeplitz sequences

if we discretize the higher-order term −(a(x)u  (x)) as in (34), the symbol of the FD
discretization matrices Bn resulting from (42) should be again a(x)(2 − 2 cos θ ). This
is in fact the case. Let us provide the proof. Consider the discretization of (42) by the
FD scheme defined as follows:

– to approximate −(a(x)u  (x)) , use again the FD formula (34);


– to approximate β(x)u  (x), use any (consistent) FD formula; a natural choice is the
centered formula

u(x j+1 ) − u(x j−1 )


β(x)u  (x)|x=x j ≈ β(x j ) ;
2h

– to approximate γ (x)u(x), use the obvious equation

γ (x)u(x)|x=x j = γ (x j )u(x j ).

Let Bn be the resulting discretization matrix. Then,

Bn = An + Yn , (43)

where An is given by (36) and Yn is the matrix coming from the approximation of the
term β(x)u  (x) + γ (x)u(x). Since β(x)u  (x) + γ (x)u(x) only involves lower-order
derivatives of u(x), it is not difficult to see that Yn  ≤ C/n for some constant C.
Hence, {Yn }n ∼σ 0 by Theorem 1, and in view of (37), (43) and LT 6, we get

{Bn }n ∼LT a(x)(2 − 2 cos θ ). (44)

Now, if the convection term is not present, i.e., β(x) = 0 identically, then Bn is
symmetric and so (44) and LT 1 yield

{Bn }n ∼σ, λ a(x)(2 − 2 cos θ ). (45)

If β(x) is not identically 0, Bn is not symmetric and thus (44) and LT 1 only imply
that {Bn }n ∼σ a(x)(2 − 2 cos θ ). However, in view of the decomposition (43), since
Yn  ≤ C/n, Yn 1 = o(n) by (3) and the previous inequality, An  ≤ 4a∞ by
(6), and {An }n ∼λ a(x)(2 − 2 cos θ ) by (38), all the hypotheses of [14, Theorem 3.4]
are met and the relation {Bn }n ∼λ a(x)(2 − 2 cos θ ) holds even if b(x) is an arbitrary
(bounded) function. In short, the relations (45) are always satisfied.

Remark 6 Based on the discussion in Remark 4, if we change the FD scheme to


approximate (33) or (42), the symbol becomes a(x) p(θ ), where p(θ ) is the trigono-
metric polynomial associated with the new FD formula used to approximate the second
derivative −u  (x) (the higher-order differential operator).
C. Garoni, S. Serra-Capizzano

5.2 FD discretization of d-dimensional elliptic PDEs

We briefly outline in this section an extension to the d-dimensional setting of the


analysis carried out in the previous section. However, to avoid technicalities, we will
not enter into the details. For more details, we refer the reader to [21, Section 6]. Let
us consider the following second-order d-dimensional elliptic problem:
⎧ ⎛ ⎞

⎪ d
∂ ⎝
d
∂u

−∇ · A(x)∇u(x) ≡ − ai j (x) (x)⎠ = f (x), x ∈ (0, 1)d ,
∂ xi ∂x j (46)

⎪ i=1 j=1

u(x) = 0, x ∈ ∂((0, 1)d ),

where f : [0, 1]d → R is continuous and A : [0, 1]d → Rd×d is a symmetric positive
definite matrix of functions ai j ∈ C 1 ([0, 1]d ). For simplicity, we only consider the
case of a square domain, but the analysis can be extended to the case of an arbitrary
domain , provided that  can be exactly described by a global (and regular) geometry
map G : [0, 1]d → ; note that this is precisely what happens in the IgA framework
[7,8,10,11]. Problem (46) can be rewritten in the form

⎨ −  a (x) ∂ u (x) −  ∂ai j (x) ∂u (x) = f (x),
d d
⎪ 2
ij x ∈ (0, 1)d ,
∂ xi ∂ x j ∂ xi ∂x j (47)

⎩ i, j=1 i, j=1
u(x) = 0, x∈ ∂((0, 1)d ),

From (47) we see that the higher-order operator in (46)–(47) is


d
∂ 2u
− ai j (x) (x) = −1 ( A(x) ◦ H u(x)) 1T , (48)
∂ xi ∂ x j
i, j=1

where 1 = (1, . . . , 1) ∈ Nd , H u(x) is the Hessian of u(x), i.e.,

∂ 2u
(H u(x))i j = (x),
∂ xi ∂ x j

and ◦ denotes the componentwise Hadamard product of matrices.


Similarly to the one-dimensional case, the second-order central FD scheme analo-
gous to the one considered in Sect. 5.1 leads to a sequence of discretization matrices
{A n }n , where the parameters n = (n 1 , . . . , n d ), which are related to the discretization
steps in each direction via the equation h i = n i 1+1 , are chosen as a function of n ∈ N,
in such a way that n → ∞ when n → ∞. Usually, one gives to every direction
the same attention by choosing n j = n for all j = 1, . . . , d. The matrices A n are
symmetric and can be written in the form


d
An = Dn (ai j )Tn ( pi j ) + Z n ,
i, j=1
The theory of locally Toeplitz sequences

where {Z n }n ∼σ 0 and pi j is a (separable) d-variate trigonometric polynomial which


only depends on the FD formula used to approximate the second-order partial deriv-
ative ∂ x∂i ∂ x j . By LT 1, LT 2, LT 3, LT 4, LT 5 it follows that
2


d
{A n }n ∼σ, λ ai j (x) pi j (θ ) = 1 ( A(x) ◦ H (θ )) 1T , (49)
i, j=1

where
 d
H (θ ) = pi j (θ ) i, j=1 .

The matrix H (θ) collects the trigonometric polynomials associated with the FD for-
mulas used to approximate the second-order partial derivatives. Noting the formal
analogy between the symbol in (49) and the higher-order differential operator (48),
the matrix H (θ) is sometimes referred to as the ‘symbol of the (negative) Hessian
operator’, although this terminology is not rigorous from the mathematical viewpoint.
As in the one-dimensional case, if we add to the diffusion problem (46) a convection
term β(x) · ∇u(x) and a reaction term γ (x)u(x), obtaining a reaction–convection–
diffusion equation, the symbol remains the same. Indeed, the symbol is not affected
by terms with lower-order derivatives (see Remark 5).

5.3 FE approximation of one-dimensional second-order PDEs

Consider the following second-order differential problem:

−(a(x)u  (x)) = f (x), x ∈ (0, 1),


(50)
u(0) = u(1) = 0,

where f ∈ L 2 ([0, 1]) and a : [0, 1] → R is only assumed to be Riemann-integrable.


We consider the approximation of (50) by linear FEs on the uniform mesh in [0, 1]
with step size h = n+11
. Let us briefly describe this approximation technique; for more
details, we refer the reader to [18, Chapter 4]. The weak form of (50) reads as follows:
find u ∈ H01 ([0, 1]) such that, for all w ∈ H01 ([0, 1]),
 1  1
a(x)u  (x)w  (x)dx = f (x)w(x)dx. (51)
0 0

Set x j = j h, j = 0, . . . , n + 1, and fix the subspace Wn = span(ϕ1 , . . . , ϕn ) ⊂


H01 ([0, 1]), where ϕ1 , . . . , ϕn are the so-called ‘hat functions’. For every i = 1, . . . , n,
the function ϕi is given explicitly by

x − xi−1 xi+1 − x
ϕi (x) = χ[xi−1 ,xi ) (x) + χ[x ,x ) (x), (52)
xi − xi−1 xi+1 − xi i i+1
C. Garoni, S. Serra-Capizzano

where χ E is the characteristic function of the set E. In the FE approach, we look for an
approximation u Wn of u by solving the following (Galerkin) problem: find u Wn ∈ Wn
such that, for all w ∈ Wn ,

 1  1
a(x)u Wn (x)w  (x)dx = f (x)w(x)dx. (53)
0 0


Since {ϕ1 , . . . , ϕn } is a basis of Wn , we can write u Wn = nj=1 u j ϕ j for a unique
vector u = (u 1 , . . . , u n )T . By linearity, the computation of u Wn (i.e., of u) reduces to
solving the linear system

An (a)u = f, (54)


n
1
where f = 0 f (x)ϕi (x)dx and An (a) is the stiffness matrix,
i=1

 1 n
An (a) = a(x)ϕ j (x)ϕi (x)dx . (55)
0 i, j=1

Note that An (a) is symmetric. In the following, we compute the spectral and singular
value distribution of the sequence of normalized stiffness matrices { n+1
1
An (a)}n using
the theory of LT sequences. More precisely, we show that

)
1
An (a) ∼LT a(x)(2 − 2 cos θ ), (56)
n+1 n

which, by LT 1, implies that

)
1
An (a) ∼σ, λ a(x)(2 − 2 cos θ ). (57)
n+1 n

The proof of (56) consists of the following three steps.


Step 1 We first consider the case where a(x) is constant, say a(x) = a0 identically. In
this case, a direct computation based on (52) shows that An (a0 ) = a0 Tn (2−2 cos θ ) =
Tn (a0 (2 − 2 cos θ )). Hence, (56) follows from LT 2.
Step 2 Now, we assume that a(x) is continuous. To prove (56) in this case, we first
illustrate the idea and then go into the details. The proof is based on the fact that the
hat functions (52) are ‘locally supported’. Indeed, the support [xi−1 , xi+1 ] of the ith
hat function ϕi (x) is localized near the point x̂i = ni ∈ [xi , xi+1 ], and the amplitude
of the support tends to 0 when n → ∞. Since a(x) varies continuously over [0, 1],
The theory of locally Toeplitz sequences

the (i, j)th entry of An (a) can be approximated as follows, for all i, j = 1, . . . , n:

 1  xi+1
(An (a))i j = a(x)ϕ j (x)ϕi (x)dx = a(x)ϕ j (x)ϕi (x)dx
0 xi−1
 xi+1  1
≈ a(x̂i ) ϕ j (x)ϕi (x)dx = a(x̂i ) ϕ j (x)ϕi (x)dx = a(x̂i )(An (1))i j .
xi−1 0
(58)

After normalization, we can rewrite the approximation (58) in matrix form as follows:

1 1
An (a) ≈ Dn (a)An (1). (59)
n+1 n+1

We shall see that (59) implies { n+11


An (a) − n+1 1
Dn (a)An (1)}n ∼σ 0, and so (56)
follows from the relation { n+1 An (1)}n ∼LT 2 − 2 cos θ proved in Step 1 and from
1

LT 3, LT 5, LT 6.
Now, let us go into the details. Since |ϕi (x)| ≤ n + 1, we have

 
   1 
(An (a))i j − (Dn (a)An (1))i j  =  a(x) − a(x̂i ) ϕ j (x)ϕi (x)dx 

0
 xi+1
≤ (n + 1)2 |a(x) − a(x̂i )|dx
xi−1

2
≤ 2(n + 1) ωa ,
n+1

for all i, j = 1, . . . , n, where ωa (·) is the modulus of continuity of a. It follows that the
modulus of each component of the matrix n+1 1
An (a)− n+1
1
Dn (a)An (1) is bounded by
2 ωa ( n+1 ). Moreover, n+1 An (a) − n+1 Dn (a)An (1) is banded (actually, tridiagonal),
2 1 1

because supp(ϕi ) ∩ supp(ϕ j ) = ∅ whenever |i − j| > 1. Thus, both the 1-norm and
the ∞-norm of n+1 1
An (a) − n+1
1
Dn (a)An (1) are bounded by 6 ωa ( n+1 2
), and so, by
(6),  n+1 An (a) − n+1 Dn (a)An (1) ≤ 6 ωa ( n+1 ) → 0 as n → ∞. By Theorem 1,
1 1 2

we conclude that { n+1 1


An (a) − n+1
1
Dn (a)An (1)}n ∼σ 0, and this implies (56) by the
relation { n+1 An (1)}n ∼LT 2 − 2 cos θ and LT 3, LT 5, LT 6.
1

Step 3 Finally, we consider the general case where a(x) is only assumed to be Riemann-
integrable. By the Lusin theorem [19], there exists a sequence of continuous functions
am ∈ C([0, 1]) such that μ1 {am = a} ≤ m1 and am ∞ ≤ a L ∞ . By Step 2,
{ n+1
1
An (am )}n ∼LT am (x)(2−2 cos θ ) and, moreover, am → a in L 1 ([0, 1]). In addi-
tion, {{ n+1
1
An (am )}n }m is an a.c.s. for { n+1
1
An (a)}n . Indeed, using (4) and observing
n 
that i=1 |ϕi (x)| ≤ 2(n + 1) for all x ∈ [0, 1], we obtain
C. Garoni, S. Serra-Capizzano

An (a) − An (am )1


 n  1


n
     
≤ (An (a))i j − (An (am ))i j  =  a(x) − a (x) ϕ (x)ϕ 
(x)dx 
 m j i 
i, j=1 i, j=1 0
 
n
≤ 2a L ∞ |ϕ j (x)| |ϕi (x)|dx ≤ 8(n + 1)2 a L ∞ μ1 {a = am },
{a =am } i, j=1

and so
 
 1 1  n
 
 n + 1 An (a) − n + 1 An (am ) ≤ C m
1

for some constant C independent of n and m. Thus, {{ n+1


1
An (am )}n }m is an a.c.s. for
{ n+1 An (a)}n by Theorem 6, and the relation { n+1 An (a)}n ∼LT a(x)(2 − 2 cos θ )
1 1

follows from LT 6.

Remark 7 Using the theory of GLT sequences, it can be shown that the relations (57)
continue to hold even if a(x) is only assumed to be in L ∞ ([0, 1]); see [12, Remark 6.2].
We refer the reader to [10,11] for an extension of the analysis considered in this section
to the multidimensional case.

5.4 Final remarks

In the applications considered in this paper, we have seen that the discretization of
PDEs leads to sequences of matrices whose symbol belongs to L ∞ . It is natural to
ask whether there are situations in which the symbol is not in L ∞ . The answer to
this question is affirmative: the discretization of PDEs may lead to symbols which are
not in L ∞ and, actually, not even in L 1 (they are just measurable). Let us describe a
situation where this phenomenon is encountered. Suppose we discretize an arbitrary
PDE by a local method (such as FDs, FEs or IgA) in which the grid is obtained as the
mapping of a uniform mesh through a non-regular function G (i.e., a function whose
derivative vanishes at some points). For simplicity, let us assume that G  vanishes at a
unique site. In this case, the grid points rapidly accumulate at this site, and the symbol
has a pole whose order depends on the considered discretization technique and on the
order of the zero of G  . Depending on the order of the pole, the symbol may or may
not be in L 1 . The next example will help to clarify this discussion.

Example Suppose that the differential problem (50) is approximated by Galerkin IgA
based on B-splines of degree p, as described in [10, Section 3]. Assume that the grid
is obtained as the mapping of the uniform mesh ni , i = 0, . . . , n, through the map

G : [0, 1] → [0, 1], G(x) = x q , q > 1.


The theory of locally Toeplitz sequences

 q
Then, G  (0) = 0 and the grid points, ni , i = 0, . . . , n, rapidly accumulate at x = 0.
The symbol of the resulting discretization matrices is

a(G(x))
f (x, θ ) = f p (θ ), (60)
|G  (x)|

where f p (θ ) is a cosine trigonometric polynomial of degree p (for p = 1 we have


f 1 (θ ) = 2−2 cos θ ). We see from (60) that the symbol has a pole at x = 0. Depending
on q, the symbol may or may not be in L 1 ([0, 1] × [−π, π ]). Equation (60) was
established in [10,11] in the case of a regular geometry map G, i.e., a map G ∈
C 1 ([0, 1]) such that G  (x) = 0 for all x ∈ [0, 1]. However, through a limit process
and a (not so easy) application of the theory of GLT sequences, one might extend the
validity of (60) to non-regular maps G; cf. [10, Remark 4.5]. Numerical experiments
in [10, Section 6.3] reveal that (60) is indeed the spectral symbol also in the case of a
non-regular G.

Finally, a question: why should one be interested in discretizing the PDE (50) with
a grid that rapidly accumulates at a point? The answer is that this local refinement is
necessary in some situations where the coefficient a(x) is strongly anisotropic. If a
uniform discretization were used, the associated discretization step should be chosen
to be very small, and this would result in a linear system with extremely large size:
the computational cost to solve it would be unsustainable. For this reason, one adopts
a local refinement, so that a coarse grid is used in the subregions of the domain where
a(x) is sufficiently smooth, and a finer grid is used only in the subregions where a(x)
is, say, ‘not well behaved’ (e.g., remarkably oscillatory).

6 Conclusions

In this work, we fully developed the theory of LT sequences. We made a significant


review of the original theory by ‘correcting’ all the relevant definitions and generalizing
and/or simplifying a lot of key results. Finally, we provided a precise summary of the
theory of LT sequences in Sect. 4.7, with the purpose of giving to the reader an easy-
to-use LT manual. Some hints on how to use this manual in practical applications were
given in Sect. 5.
We conclude by highlighting that the review and extension of the theory of LT
sequences presented herein is the first main novelty of our book [12]. We plan to
complete the picture started with this work by publishing in a forthcoming paper our
review and extension of the theory of GLT sequences. This is in fact the second main
novelty of [12]. It is worth noting that, as indicated in Remark 7 and Sect. 5.4, the
theory of GLT sequences is not merely an academic completion of the theory of LT
sequences.

Acknowledgments The authors express their sincere gratitude to the Editor Albrecht Böttcher and to the
referee, who helped them to improve the paper. In particular, Sect. 5.4 originated from a specific remark by
the referee.
C. Garoni, S. Serra-Capizzano

References
1. Avram, F.: On bilinear forms in Gaussian random variables and Toeplitz matrices. Probab. Theory
Relat. Fields 79, 37–45 (1988)
2. Bhatia, R.: Matrix Analysis. Springer, New York (1997)
3. Böttcher, A., Grudsky, S.M.: Toeplitz Matrices, Asymptotic Linear Algebra, and Functional Analysis.
Birkhäuser Verlag, Basel-Boston-Berlin (2000)
4. Böttcher, A., Grudsky, S.M.: Spectral Properties of Banded Toeplitz Matrices. SIAM, Philadelphia
(2005)
5. Böttcher, A., Silbermann, B.: Introduction to Large Truncated Toeplitz Matrices. Springer, New York
(1999)
6. Böttcher, A., Silbermann, B.: Analysis of Toeplitz Operators, 2nd edn. Springer, Berlin (2006)
7. Cottrell, J.A., Hughes, T.J.R., Bazilevs, Y.: Isogeometric Analysis: Toward Integration of CAD and
FEA. Wiley, Chichester (2009)
8. Donatelli, M., Garoni, C., Manni, C., Serra-Capizzano, S., Speleers, H.: Spectral analysis and spectral
symbol of matrices in isogeometric collocation methods. Math. Comput. doi:10.1090/mcom/3027
9. Donatelli, M., Garoni, C., Mazza, M., Serra-Capizzano, S., Sesana, D.: Spectral behavior of precon-
ditioned non-Hermitian multilevel block Toeplitz matrices with matrix-valued symbol. Appl. Math.
Comput. 245, 158–173 (2014)
10. Garoni, C., Manni, C., Serra-Capizzano, S., Sesana, D., Speleers, H.: Spectral analysis and spectral
symbol of matrices in isogeometric Galerkin methods. Math. Comput. (to appear)
11. Garoni, C., Manni, C., Serra-Capizzano, S., Sesana, D., Speleers, H.: Lusin theorem, GLT sequences
and matrix computations: an application to the spectral analysis of PDE discretization matrices. Tech-
nical Report 2015-012, Department of Information Technology, Uppsala University (2015). http://
www.it.uu.se/research/publications/reports/2015-012/
12. Garoni, C., Serra-Capizzano, S.: Generalized Locally Toeplitz Sequences: Theory and Applications.
Springer (to appear). The preliminary (unpublished and incomplete) version of this book is available,
under a different title, as Technical Report 2015-023 at the Department of Information Technology of
Uppsala University: http://www.it.uu.se/research/publications/reports/2015-023/
13. Garoni, C., Serra-Capizzano, S., Vassalos, P.: A general tool for determining the asymptotic spectral
distribution of Hermitian matrix-sequences. Oper. Matrices 9, 549–561 (2015)
14. Golinskii, L., Serra-Capizzano, S.: The asymptotic properties of the spectrum of nonsymmetrically
perturbed Jacobi matrix sequences. J. Approx. Theory 144, 84–102 (2007)
15. Golub, G.H., Van Loan, C.F.: Matrix Computations, 4th edn. Johns Hopkins University Press, Baltimore
(2013)
16. Grenander, U., Szegő, G.: Toeplitz Forms and Their Applications, 2nd edn. AMS Chelsea Publishing,
New York (1984)
17. Parter, S.V.: On the distribution of the singular values of Toeplitz matrices. Linear Alg. Appl. 80,
115–130 (1986)
18. Quarteroni, A.: Numerical Models for Differential Problems, 2nd edn. Springer-Verlag Italia, Milan
(2014)
19. Rudin, W.: Real and Complex Analysis, 3rd edn. McGraw-Hill, Singapore (1987)
20. Serra-Capizzano, S.: Distribution results on the algebra generated by Toeplitz sequences: a finite
dimensional approach. Linear Alg. Appl. 328, 121–130 (2001)
21. Serra-Capizzano, S.: Generalized locally Toeplitz sequences: spectral analysis and applications to
discretized partial differential equations. Linear Alg. Appl. 366, 371–402 (2003)
22. Serra-Capizzano, S.: The GLT class as a generalized Fourier Analysis and applications. Linear Alg.
Appl. 419, 180–233 (2006)
23. Serra-Capizzano, S., Tilli, P.: On unitarily invariant norms of matrix-valued linear positive operators.
J. Inequal. Appl. 7, 309–330 (2002)
24. Smith, G.D.: Numerical Solution of Partial Differential Equations: Finite Difference Methods, 3rd edn.
Oxford University Press, New York (1985)
25. Tilli, P.: A note on the spectral distribution of Toeplitz matrices. Linear Multilinear Algebra 45, 147–159
(1998)
26. Tilli, P.: Locally Toeplitz sequences: spectral properties and applications. Linear Alg. Appl. 278, 91–
120 (1998)
The theory of locally Toeplitz sequences

27. Tyrtyshnikov, E.E.: A unifying approach to some old and new theorems on distribution and clustering.
Linear Alg. Appl. 232, 1–43 (1996)
28. Tyrtyshnikov, E.E., Zamarashkin, N.L.: Spectra of multilevel Toeplitz matrices: advanced theory via
simple matrix relationships. Linear Alg. Appl. 270, 15–27 (1998)
29. Zamarashkin, N.L., Tyrtyshnikov, E.E.: Distribution of eigenvalues and singular values of Toeplitz
matrices under weakened conditions on the generating function. Sb. Math. 188, 1191–1201 (1997)

You might also like