Operator Theory

Advances and Applications

267
Carlos André
M. Amélia Bastos
Alexei Yu. Karlovich
Bernd Silbermann
Ion Zaballa
Editors

Operator Theory,
Operator Algebras,
and Matrix Theory
Operator Theory: Advances and Applications
Volume 267

Founded in 1979 by Israel Gohberg

Editors:
Joseph A. Ball (Blacksburg, VA, USA)
Albrecht Böttcher (Chemnitz, Germany)
Harry Dym (Rehovot, Israel)
Heinz Langer (Wien, Austria)
Christiane Tretter (Bern, Switzerland)

Associate Editors: Honorary and Advisory Editorial Board:

Vadim Adamyan (Odessa, Ukraine) Lewis A. Coburn (Buffalo, NY, USA)
Wolfgang Arendt (Ulm, Germany) Ciprian Foias (College Station, TX, USA)
B. Malcolm Brown (Cardiff, UK) J.William Helton (San Diego, CA, USA)
Raul Curto (Iowa, IA, USA) Marinus A. Kaashoek (Amsterdam, NL)
Kenneth R. Davidson (Waterloo, ON, Canada) Thomas Kailath (Stanford, CA, USA)
Fritz Gesztesy (Waco, TX, USA) Peter Lancaster (Calgary, Canada)
Pavel Kurasov (Stockholm, Sweden) Peter D. Lax (New York, NY, USA)
Vern Paulsen (Houston, TX, USA) Bernd Silbermann (Chemnitz, Germany)
Mihai Putinar (Santa Barbara, CA, USA) Harold Widom (Santa Cruz, CA, USA)
Ilya Spitkovsky (Abu Dhabi, UAE)

Subseries
Linear Operators and Linear Systems
Subseries editors:
Daniel Alpay (Orange, CA, USA)
Birgit Jacob (Wuppertal, Germany)
André C.M. Ran (Amsterdam, The Netherlands)

Subseries
Advances in Partial Differential Equations
Subseries editors:
Bert-Wolfgang Schulze (Potsdam, Germany)
Michael Demuth (Clausthal, Germany)
Jerome A. Goldstein (Memphis, TN, USA)
Nobuyuki Tose (Yokohama, Japan)
Ingo Witt (Göttingen, Germany)

More information about this series at http://www.springer.com/series/4850

Carlos André • M. Amélia Bastos
Alexei Yu. Karlovich • Bernd Silbermann
Ion Zaballa
Editors

Operator Theory, Operator

Algebras, and Matrix Theory
Editors
Carlos André M. Amélia Bastos
Departamento de Matemática Departamento de Matemática
Faculdade de Ciências Instituto Superior Técnico
Universidade de Lisboa Universidade de Lisboa
Lisboa Portugal Lisboa Portugal

Alexei Yu. Karlovich Bernd Silbermann

Departamento de Matemática Fakultät für Mathematik
Faculdade de Ciências e Tecnologia Technische Universität Chemnitz
Universidade Nova de Lisboa Chemnitz, Germany
Lisboa, Portugal

Ion Zaballa
Departamento de Matemática Aplicada y EIO
Universidad del País Vasco (UPV/EHU)
Leioa, Spain

ISSN 0255-0156 ISSN 2296-4878 (electronic)

Operator Theory: Advances and Applications
ISBN 978-3-319-72448-5 ISBN 978-3-319-72449-2 (eBook)
https://doi.org/10.1007/978-3-319-72449-2

Library of Congress Control Number: 2018949649

Mathematics Subject Classification (2010): 47-xx, 15-xx, 46-xx

© Springer International Publishing AG, part of Springer Nature 2018

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

This book is published under the imprint Birkhäuser, www.birkhauser-science.com by the registered company
Springer Nature Switzerland AG part of Springer Nature.
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Contents
Editorial Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii

C.A.M. André, F. Gomes and J. Lochon

Indecomposable Supercharacters of the Infinite Unitriangular
Group . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

M.A. Bastos, C.A. Fernandes and Yu.I. Karlovich

A C ∗ -algebra of Singular Integral Operators with Shifts and
Piecewise Quasicontinuous Coefficients . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

N. Bebiano, J. da Providência and J.P. da Providência

Non-Hermitian Quantum Mechanics of Bosonic Operators . . . . . . . . . . . 65

C. Carvalho, V. Nistor and Y. Qiao

Fredholm Conditions on Non-compact Manifolds: Theory
and Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79

K. Demirci and S. Orhan

Statistical e-Convergence of Bögel-Type Continuous Functions . . . . . . 123

K. Demirci, S. Orhan and B. Kolay

Weighted Statistical Relative Approximation by Positive
Linear Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131

M. Dodig
Descriptor Systems Under Feedback and Output Injection . . . . . . . . . . . 141

R.G. Douglas and R. Yang

Hermitian Geometry on Resolvent Set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167

F. Gómez-Cubillo and S. Villullas

Spectral Algorithms for MRA Orthonormal Wavelets . . . . . . . . . . . . . . . . 185

C.R. Johnson, C. Marijuán, P. Paparella and M. Pisonero

The NIEP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
vi Contents

A.Yu. Karlovich, Yu.I. Karlovich and A.B. Lebre

Semi-Fredholmness of Weighted Singular Integral Operators
with Shifts and Slowly Oscillating Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221

A.B. Lebre and J.S. Rodrı́guez

Factorization of Singular Integral Operators with a Carleman
Backward Shift: The Vector Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247

S. Roch
Extension-Restriction Theorems for Algebras of Approximation
Sequences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 261

S. Roch and B. Silbermann

Toeplitz and Hankel Algebras – Axiomatic and Asymptotic
Aspects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 285

P.A. Santos
More Than 40 Years of Algebraic Techniques in Numerical
Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 317

K.V. Sklyar, G.M. Sklyar and S.Yu. Ignatovich

Linearizability of Multi-Control Systems of the Class C 1 by
Additive Change of Controls . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 359

I.M. Spitkovsky
A Distance Formula Related to a Family of Projections
Orthogonal to Their Symmetries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 371
Operator Theory:
Advances and Applications, Vol. 267, vii–viii

c Springer International Publishing AG, part of Springer Nature 2018

Editorial Introduction

This book is dedicated to the International Workshop on Operator Theory

and Operator Algebras – WOAT 2016, which took place in Lisbon at Instituto
Superior Técnico, from July 5 to July 8 of 2016. This workshop continued a series
of conferences organized in Lisboa IST since 2006 and aimed to promote research
exchanges among Operator Theory, Operator Algebras and Matrix Theory areas.
The book consists of 17 chapters that cover research fields in Operator Theory
and Operator Algebras as well as in Matrix Theory and Representation Theory.
The research field in Operator Theory and Operator Algebras is mainly rep-
resented in chapters that cover the following different topics:
• Fredholm theory for non-local C ∗ -algebras of singular integral operators with
piecewise quasicontinuous coefficients and the local trajectory method. Ex-
plicit Fredholm conditions for classes of pseudodifferential operators on sin-
gular and non compact spaces and the new concept of Fredholm groupoid.
• A Korovkin type approximation theorem. Weighted statistical relative ap-
proximation by positive linear operators.
• Study of general, possibly non-commuting, tuples using geometric ideas based
on the newly emerged concept of projective joint spectrum.
• Factorization of singular integral operators with a Carleman backward shift
of linear fractional type, on Lnp . Sufficient conditions for semiFredholmness
on Lp (R+ ) of weighted singular integral operators with shifts and slowly
oscillating data with discontinuities at 0 and ∞.
• The structure of every separable C ∗ -algebra of approximation sequences and
the C ∗ -algebra of the finite sections discretization for Toeplitz operators with
continuous generating functions. Banach algebras of Toeplitz like operators
defined in an axiomatic way and the classical Toeplitz and Hankel operators.
Overview of the historical development of more than 40 years of algebraic
techniques in Numerical Analysis.
• A formula for the distance from an orthogonal projection, on some Hilbert
space, to a set of orthogonal projections.
The research field in Matrix Theory is covered in several chapters devoted to
the following topics:
• Spectral analysis of non-Hermitian operators appearing in quantum physics.
The diagonalization of such operators and their adjoints is proved.
viii Editorial Introduction

• Simultaneous feedback and output injection on descriptor linear system de-

scribed by a quadruple of matrices. Description of the possible Kronecker
invariants of the resulting pencil. A constructive and explicit solution is given
over algebraically closed fields.
• Spectral representations of the dilation and translation operators used to
construct wavelets acting on the Hilbert space of square integrable functions
on the real line. The spectral analysis concerns some particular orthonormal
bases: the Haar basis, the Walsh–Paley basis and the trigonometric basis.
• A survey on the nonnegative inverse eigenvalue problem (NIEP), and its
several variants, with an emphasis on recent results.
• Linearization of nonlinear control systems. Necessary and sufficient conditions
are found in a particular case (local A-linearizability) of the more general case
(local feedback linearization).
Finally, in the research field of Representation Theory, the notion of an in-
decomposable (extreme) supercharacter is defined for infinite algebra groups and
a description of these is given for the infinite unitriangular group in terms of su-
percharacters of the finite unitriangular groups.
The editors of the volume are grateful for the support of the Portuguese
Foundation for Science and Technology and the Center for Function Analysis,
Linear Structures and Applications. They also deeply thank the Birhäuser’s edito-
rial team, Dorothy Mazlum and Sabrina Höcklin, for their availability during the
preparation of the volume.

The editors December 2017

Operator Theory:
Advances and Applications, Vol. 267, 1–24

c Springer International Publishing AG, part of Springer Nature 2018

Indecomposable Supercharacters
of the Infinite Unitriangular Group
Carlos A.M. André, Filipe Gomes and Jocelyn Lochon

Abstract. Let U∞ (k) be the locally finite unitriangular group defined over
a finite field k with q elements. We define the notion of an indecomposable
supercharacter and describe these indecomposable supercharacters in terms
of the supercharacters of the finite unitriangular groups Un (k).
Mathematics Subject Classification (2010). Primary 20C15, 20G40; Secondary
05E10, 43A35.
Keywords. Unitriangular group; indecomposable supercharacter; branching
graph; multiplicative graph; multiplicative linear function.

1. Introduction
Let G be a group. A complex-valued function φ : G → C is said to be positive
definite if
1. φ(g −1 ) = φ(g) for all g ∈ G, and
2. For
 any finite
 collection g1 , . . . , gm of elements of G, the Hermitian matrix
φ(gi gj−1 ) 1≤i,j≤m is nonnegative.
A function φ : G → C is said to be central (or a class function) if it is constant
on conjugacy classes, that is, if φ(ghg −1 ) = φ(h) for all g, h ∈ G; and it is said
to be normalized if φ(1) = 1. We will denote by Ch(G) the set consisting of all
central, positive definite, normalized functions on G (if G is a topological group,
we additionally require the functions to be continuous). If φ, ψ ∈ Ch(G), then

This research was made within the activities of the Group for Linear, Algebraic and Combi-
natorial Structures of the Center for Functional Analysis, Linear Structures and Applications
(University of Lisbon, Portugal), and was partially supported by the Portuguese Science Foun-
dation (FCT) through the Strategic Project UID/MAT/04721/2013.
The second and third authors were partially supported by the Lisbon Mathematics PhD program
(funded by the Portuguese Science Foundation). The first half of the work is part of the third
author Ph.D. thesis, whereas the second half is part of the second author Ph.D. thesis.
2 C.A.M. André, F. Gomes and J. Lochon

for every real number 0 ≤ t ≤ 1 the function (1 − t)φ + tψ is also an element

of Ch(G), which means that Ch(G) is a convex set. A point of a convex set is
called an extreme point if it is not contained in the interior of any interval entirely
contained in the set; we denote by ExCh(G) the subset of Ch(G) consisting of
extreme points of Ch(G). We refer to an element of Ch(G) as a character of G
and to an element of ExCh(G) as an indecomposable character of G. (Indeed,
Ch(G) is a Choquet simplex which means that every character of G is uniquely
representable as a convex mixture of indecomposable characters; see, for example,
[18] or [31].)
If G is a compact group (in particular, a finite group), then the extreme points
of Ch(G) are exactly the normalized irreducible characters of G; the character of
an irreducible (complex) representation π of the group G is the function χ : G → C
given by the trace χ(g) = tr(π(g)) for all g ∈ G, and we define the normalized
irreducible character χ by
χ(g)
(g) =
χ
χ(1)
for all g ∈ G. This definition makes sense because every irreducible represen-
tation of a compact group is finite-dimensional, but it does not make sense in
general for non-compact groups. A classical construction associates indecompos-
able characters with factor representations of finite von Neumann type, and the
representation theory is then mainly focused on the classification of indecompos-
able characters and on the decomposition of a generic character as a convex sum
of indecomposable characters (the latter being as a counterpart of the decompo-
sition of a representation into the irreducible ones). In fact, the classification of
finite factor representations of the group G reduces, up to quasi-equivalence, to
the classification of ExCh(G) (see, for example, [19]).
In [34], Thoma obtained an explicit description of the indecomposable char-
acters of the infinite symmetric group S∞ , and his description of ExCh(S∞ ) was
later explained in terms of infinite paths on Young diagrams by Kerov–Vershik
[28]; these works opened up a new field of research in asymptotic representation
theory (see [26]). There are a few more examples of groups whose indecomposable
characters have been explicitly classified (see, for example, [9], [15], [16], [20], [21],
[22], [32]); all these examples are restricted to inductive limits of compact groups
(and their completion in some topology), and they depend on the knowledge of
the indecomposable characters of those compact groups. On the other hand, in
many examples the indecomposable characters may be complicated functions, and
the set of indecomposable characters may be too large; it is therefore of interest
to study smaller and more tractable families of reducible characters (for instance,
those which have some kind of symmetry or depend on some simple statistic on the
group). In this paper, we consider the standard supercharacter theory for algebra
groups (defined over finite fields), and describe all the indecomposable superchar-
acters of the locally finite unitriangular group U∞ (k) defined over a finite field k.
Indeed, the group U∞ (k) can be naturally realized as the inductive limit of the
 Indecomposable Supercharacters 3
 ∞
inductive family Un (k) n=1 of all finite unitriangular groups over k, and thus the
Kerov–Vershik’s ergodic method (as introduced in [28]) allows for the description
of the indecomposable characters of U∞ (k) in terms of the irreducible characters
of the finite groups Un (k) for all n ∈ N. However, it is well known that the clas-
sification of the irreducible characters of the finite unitriangular group is a wild
problem (and probably impossible to accomplish), and thus it is worthy to approx-
imate irreducible characters (and conjugacy classes) by the more tractable family
of supercharacters (and superclasses). The notion of a supercharacter theory will
be defined in Section 2 below where we also describe the (standard) superclasses
and supercharacters of the finite unitriangular groups and introduce the notion of
a supercharacter and of an indecomposable supercharacter of U∞ (k). (We observe
that this notion may be naturally extended to the larger family of locally finite
algebra groups; however, for the purposes of this paper, we prefer to skip this more
general notion.)

2. Superclasses and supercharacters of unitriangular groups

The notion of a supercharacter theory of a finite group was introduced by Diaconis–
Isaacs [14] to generalise an approach used by André (e.g., [2, 4, 5]) and Yan [35]
to study the irreducible characters of the finite unitriangular groups. The basic
idea is to coarsen the usual irreducible character theory of a group by replacing
irreducible characters with integer linear combinations of irreducible characters
that are constant on a set of clumped conjugacy classes.
By a supercharacter theory of a finite group G we mean a pair (X, Y) where
X is set partition of G, and Y an orthogonal set of characters of G (not necessarily
irreducible), such that:
1. |X| = |Y|,
2. every character χ ∈ Y takes a constant value on each member K ∈ X, and
3. each irreducible character is a constituent of one of the characters χ ∈ Y.
We refer to the members K ∈ X as superclasses and to the characters χ ∈ Y as
supercharacters of G. We note that the superclasses of G are always unions of
conjugacy classes; moreover, {1} forms a superclass and the principal character
1G is always a supercharacter of G.
Let p be an arbitrary (but fixed) prime number, let q = pe (e ≥ 1) be a power
of p, and let k be a finite field with q elements. Let A be a nilpotent associative
k-algebra (which is not assumed to be finite-dimensional), and let G = 1 + A be
the set of all formal expressions of the form 1 + a for a ∈ A. Then, G is a group
under the operation (1 + a)(1 + b) = 1 + (a + b + ab) for all a, b ∈ A; following [23],
we say that G = 1 + A is an algebra group over k (or simply an k-algebra group).
A prototype example is the unitriangular group Un (k) consisting of all unipotent
uppertriangular n × n matrices over k; in this case, we have Un (k) = 1 + un (k)
where un (k) is the nilpotent k-algebra consisting of all nilpotent uppertriangular
n × n matrices over k.
4 C.A.M. André, F. Gomes and J. Lochon

Throughout the paper, we set Un = Un (k) and un = un (k). Moreover, we

consider the group Bn = Bn (k) consisting of all invertible uppertriangular n × n
matrices over k; then, Un is a normal subgroup of Bn , and Bn is indeed the
semidirect product Bn = Tn Un where Tn = Tn (k) is the subgroup of Bn consisting
of all diagonal matrices. Left multiplication naturally defines an action of Bn on
the left of the k-algebra un , whereas right multiplication defines an action of Bn
on the right of un . Since the left action of Bn on un obviously commutes with the
right action, we can define an action of the group Bn × Bn on the left of un by
the rule (g, h) · a = gah−1 for all g, h ∈ Bn and all a ∈ un . Thus, the k-algebra
un is partitioned into “two-sided” orbits Bn aBn for a ∈ un , and this determines a
partition of the unitriangular group Un into subsets 1 + Bn aBn for a ∈ un ; these
are precisely what we define as the superclasses of Un . Notice that every element
of un may be transformed (or normalized) under the action of Bn × Bn to such an
element a ∈ un that:
1. for every 1 ≤ i < j ≤ n, the (i, j)th coefficient ai,j is either 0 or 1;
2. if ai,j = 1 for 1 ≤ i < j ≤ n, then ai,l = 0 for all i < l ≤ n, l = j, and
ak,j = 0 for all 1 ≤ k < j, k = i.
 The superclasses
 of the group Un are parametrised by set partitions of [n] =
1, 2, . . . , n ; we denote by SP(n) the set consisting
 of all set partitions
 of [n]. For
simplicity of writing, we also define [[n]] = (i, j) : 1 ≤ i < j ≤ n . If π ∈ SP(n),
then a pair (i, j) ∈ [[n]] is said to be an arc of π if i and j occur in the same
block B of π and there is no k ∈ B with i < k < j; we denote by D(π) the set
consisting of all arcs of π. Every set partition of [n] is uniquely determined by its
set of arcs, and hence by a subset of [[n]]. We say that D ⊆ [[n]] is a basic subset
if D = D(π) for some π ∈ SP(n); notice that the empty subset ∅ ⊆ [[n]] is basic,
and that it corresponds to the set partition π = 1/2/ . . . /n of [n]. (As usual, we
write π = B1 /B2 / . . . /Bt where B1 , B2 , . . . , Bt are the blocks of a set partition
π ∈ SP(n) which we agree to be ordered by increasing value of the smallest element
in the block.)
With every set partition π ∈ SP(n) we associate the matrix

eπ = ei,j ∈ un
(i,j)∈D(π)
 
where ei,j : (i, j) ∈ [[n]] is the standard k-basis of un ; as usual, for every (i, j) ∈
[[n]], we denote by ei,j the matrix in un with 1 in the (i, j)th entry and zeroes
elsewhere. It is obvious that every two-sided orbit of Bn on un contains a unique
matrix of the form eπ for some π ∈ SP(n), and thus the superclasses of Un are
parametrised by set partitions of [n]. For every π ∈ SP(n), we write Kπ = 1 +
Bn e π 
 Bn ; in the case where π is the unique set partition of [n] such that D(π) =
(i, j) for 1 ≤ i < j ≤ n, we simplify the notation and write Ki,j instead of Kπ .
It can be proved that

Kπ = Ki,j
(i,j)∈D(π)
 Indecomposable Supercharacters 5

for all π ∈ SP(n); this follows easily from [5, Proposition 2.4] (see also [7, Theo-
rem 3.1]) because
Kπ = Kπ (t)
t∈Tn
−1
where Kπ (t) = 1 + Un (t eπ t)Un for all t ∈ Tn . We observe that the double
Un -orbits Kπ (t), for π ∈ SP(n) and t ∈ Tn , are precisely the superclasses in the
standard supercharacter theory of Un as defined in [14]. Although our results are
valid for this standard supercharacter theory, we prefer to use the above coarsening
(which is natural and avoids considering k× -coloured set partitions); as we will see
below, this supercharacter theory has integer-valued supercharacters, and hence
its use might be combinatorially more natural and advantageous. (This slightly
coarser supercharacter theory has been somewhat explored and used in several
papers; see for example [8], [10], [33], or [13]; it is also worth mentioning that
by a result of J. Keller [24], for each finite group G, there exists a unique finest
supercharacter theory with integer values.)
We next define the set of supercharacters of Un ; our construction is based on
the results of [14], and the details can be found in [7]. Let u◦n be the dual group
◦
of the additive group u+ n of un ; hence, un consists of linear (additive) characters
of un . For each of the natural actions of Bn on un , there is a corresponding action
of Bn on u◦n : given ϑ ∈ u◦n and g ∈ Bn , we define gϑ, ϑg ∈ u◦n by the formulas
(gϑ)(a) = ϑ(g −1 a) and (ϑg)(a) = ϑ(ag −1 ) for all a ∈ un . It is routine to check
that these actions commute, and thus we have left orbits Bn ϑ, right orbits ϑBn ,
and “two-sided” orbits Bn ϑBn for any ϑ ∈ u◦n .
For notational simplicity, we introduce the (bijective) map ν : Un → un given
by ν(g) = g − 1 for all g ∈ Un , and observe that ν(gh) = gh − 1 = gh − g + g − 1 =
gν(h) + ν(g) for all g, h ∈ Un . Then, Un acts on left of un by g · a = ga + ν(g) for all
g ∈ Un and all a ∈ un ; we note that g · ν(h) = ν(gh) for all g, h ∈ Un . This action
can be extended by linearity to the group algebra Cun of un , and so Cun becomes
a left CUn -module. Moreover, it is clear that the map ν : Un → un extends linearly
to an isomorphism of CUn -modules ν : CUn → Cun , and thus the CUn -module Cun
affords the regular character ρUn of Un , that is, the character afforded by the left
(regular) CUn -module CUn (where Un acts naturally by left multiplication); recall
that the degree ρUn (1) = dimC CUn equals the order q n(n−1)/2 of Un , whereas
ρUn (g) = 0 for all g ∈ Un , g = 1. 
 Besides its natural C-basis, the group-algebra Cun also has a C-basis εϑ : ϑ ∈
u◦n where
1 
εϑ = ϑ(a) a
|un | a∈u
n

is the central primitive idempotent of Cun which corresponds to the linear char-
acter ϑ ∈ u◦n . For every g ∈ Un and every ϑ ∈ u◦n , we have g · εϑ = ϑ(ν(g −1 )) εgϑ ,
and thus 
Lϑ = Cετ
τ ∈Bn ϑ
6 C.A.M. André, F. Gomes and J. Lochon

is a CUn -submodule of Cun ; in fact, we have


Lϑ = Lϑ (t)
t∈Tn

where Lϑ (t) = CUn · εtϑ for all t ∈ Tn . Then, we define the supercharacter ξϑ of Un
associated with ϑ ∈ u◦n to be the character of Un afforded by Lϑ ; in particular, ξϑ
has degree ξϑ (1) = |Bn ϑ|. We observe that, if χϑ is the character afforded by the
left CUn -module Lϑ (1) = CUn ·εϑ , then for every t ∈ Tn the left CUn -module Lϑ (t)
affords the Tn -conjugate character (χϑ )t ; moreover, we have (χϑ )t = χtϑ for all
t ∈ Tn . It is not difficult to justify that the characters (χϑ )t , for ϑ ∈ u◦n and t ∈ Tn ,
are precisely the standard supercharacters of Un (as defined by Diaconis–Isaacs in
[14]); in particular, it follows from [14, Theorem 5.6] that
|Bn ϑ| 
ξϑ (g) = τ (g − 1)
|Bn ϑBn |
τ ∈Bn ϑBn

for all ϑ ∈ u◦n and all g ∈ Un . As a consequence, we see that, for all ϑ, ϑ ∈ u◦n ,
the supercharacters ξϑ and ξϑ have a common irreducible constituent if and only
if Bn ϑBn = Bnϑ Bn , in which
 case we must have ξϑ = ξϑ .


If X(n) = Ka : a ∈ un and Y(n) = ξϑ : ϑ ∈ u◦n , then the pair X(n), Y(n)

is a supercharacter theory of Un . We know that
 
X(n) = Kπ : π ∈ SP(n)
(in fact, the mapping π → Kπ defines a bijection between SP(n) and X(n)), and
thus |Y(n)| = |X(n)| = |SP(n)|; indeed, there is also a natural bijection between
SP(n) and Y(n). To describe this, it is convenient to consider the k-bilinear form
on un defined by a · b = tr(aT b) for all a, b ∈ un . On the other hand, we fix a non-
trivial linear character ϑ : k → C× of the additive group of k, and for every a ∈ un
we define the map ϑa : un → C× by  · b) for all b ∈ un . It is routine to
 ϑa (b) = ϑ(a
check that ϑa ∈ u◦n ; in fact, u◦n = ϑa : a ∈ un . For every set partition π ∈ SP(n),
we set ϑπ = ϑeπ and denote by ξπ the supercharacter of Un which corresponds
to ϑπ . Since every two-sided Bn -orbit contains ϑπ for a unique π ∈ SP(n), we
conclude that
 
Y(n) = ξπ : π ∈ SP(n) ;
in fact, the mapping π → ξπ defines a bijection between SP(n) and Y(n).
Finally, we mention the formula below for the supercharacter values; a proof
can be found in [7, Theorem 3.9] (in fact, it is a consequence of the correspond-
ing formula for the standard supercharacter values; see [5, Theorem 5.1] and [6,
Theorem 3]). For every set partition π ∈ SP(n), we say that (k, l) ∈ [[n]] is a π-
singular arc if i ≤ k < l ≤ j for some arc (i, j) ∈ D(π) and if these two arcs share
one and only one common vertex (that is, either k = i, or l = j, but not both);
otherwise, we say that (k, l) is a π-regular arc. We denote by Sing(π) the subset of
[[n]] consisting of all π-singular arcs, and by Reg(π) the subset of [[n]] consisting
 Indecomposable Supercharacters 7

of all π-regular arcs; in particular, we have D(π) ⊆ Reg(π). Now, if π, σ ∈ SP(n)

and g ∈ Kσ , then ξπ (g) = 0 unless D(σ) ⊆ Reg(π), in which case we have
(−1)|D(π)∩D(σ)| q dim(π) (q − 1)|D(π)|
ξπ (g) = (2.1)
q nstπ (σ) q |D(π)| (q − 1)|D(π)∩D(σ)|
where 
dim(π) = (j − i)
(i,j)∈D(π)
and   
nstπ (σ) = (k, l) ∈ D(σ) : i < k < l < j .
(i,j)∈D(π)
(There is at least one different way of defining the dimension dim(π). Our definition
is slightly less natural in light of the papers [11, 12] on set partitions statistics where
the sum is considered over j − i − 1 instead of j − i. However, this has no essential
effect other than simplifying some of the formulas. We note that our definition is
also used in other papers; see for example [8] or [13].)  
In the case where π ∈ SP(n) is such that D(π) = (i, j) , we simplify the
notation and write ξi,j instead of ξπ . It follows easily from the above formula that

ξπ = ξi,j , π ∈ SP(n); (2.2)
(i,j)∈D(π)

indeed, this is an easy consequence of [7, Theorem 3.6]. We note that, if (i, j) ∈ [[n]],
σ ∈ SP(n) and g ∈ Kσ , then
⎧
⎪ q j−i−1
⎪−
⎪ , if (i, j) ∈ D(σ),
⎪
⎨ q nsti,j (σ)
ξi,j (g) = q
j−i−1
(q − 1)   (2.3)
⎪
⎪ , if D(σ) ∩ (i, j) ∪ Sing(i, j) = ∅,
⎪
⎪ q nst i,j (σ)
⎩
0, if D(σ) ∩ Sing(i, j) = ∅.
where we define nsti,j(σ) = nstπ (σ) and Sing(i, j) = Sing(π) whenever π ∈ SP(n)
is such that D(π) = (i, j) .
Next, we introduce the notions of superclasses and supercharacters of the lo-
cally finite unitriangular group U∞ (k); for simplicity, we shall write U∞ = U∞ (k).
For every n ∈ N, we naturally identify Bn with the subgroup
  
 g 0
Bn = : g ∈ Bn
0 1
of Bn+1 ; hence, we consider Un as a subgroup of Un+1 and Tn as a subgroup of
Tn+1 . Then,
U∞ = Un ;
n∈N
moreover, if we set
B∞ = Bn and T∞ = Tn ,
n∈N n∈N
8 C.A.M. André, F. Gomes and J. Lochon

then U∞ is a normal subgroup of B∞ , and B∞ is the semidirect product  B∞ =

T∞ U∞ . Moreover, we consider the locally finite nilpotent k-algebra u∞ = n∈N un
(where we also agree that un ⊆ un+1 for all n ∈ N); hence, U∞ is the k-algebra
group U∞ = 1 + u∞ .
As in the finite case, for every a ∈ u∞ , we may consider the two-sided B∞ -
orbit B∞ aB∞ ⊆ u∞ , and define Ka = 1 + B∞ aB∞ to be a superclass of U∞ . We
note that, since a ∈ Un for some n ∈ N, there exists a unique set partition π ∈
SP(n) such that eπ ∈ Bn aBn , and thus we must have B∞ aB∞ = B∞ eπB∞ . There-
fore, the superclasses of U∞ are parametrised by set partitions π ∈ n∈N SP(n);
throughout the paper, we set

SP = SP(n).
n∈N

For every π ∈ SP, we write Kπ (∞) = 1 + B∞ eπ B∞ ; thus, Kπ (∞) is the superclass

of U∞ which contains the element gπ = 1 + eπ ∈ U∞ .
A function φ : U∞ → C is said to be supercentral (or a superclass function) if
it is constant on superclasses, that is, if φ(1 + gah) = φ(1 + a) for all g, h ∈ B∞ and
all a ∈ u∞ . We will denote by SCh(U∞ ) the set consisting of all supercentral, posi-
tive definite and normalized functions on U∞ ; it is clear that SCh(U∞ ) ⊆ Ch(U∞ ).
As in the case of characters, SCh(U∞ ) is a convex set; we denote by ExSCh(U∞ )
the subset of SCh(U∞ ) consisting of all extreme points of SCh(U∞ ). We refer to an
element of SCh(U∞ ) as a supercharacter of U∞ , and to an element of ExSCh(∞) as
an indecomposable supercharacter of U∞ . (As in the case of characters, SCh(U∞ ) is
a Choquet simplex, and thus every supercharacter of U∞ is uniquely representable
as a convex mixture of indecomposable characters.)
Similarly, we may also define supercharacters and indecomposable super-
characters of the finite unitriangular group. For every n ∈ N, the indecomposable
supercharacters are precisely the normalized supercharacters ξπ = ξπ (1)−1 ξπ for
π ∈ SP(n), whereas a supercharacter is a positive integer linear combination of the
indecomposable supercharacters. We should note that we use the word “indecom-
posable” in the sense that an indecomposable supercharacter cannot be expressed
as a sum of two other supercharacters (hence, an indecomposable supercharacter
is not necessarily an indecomposable character). On the other hand, if m ∈ N
is such that n ≤ m, then a set partition π ∈ SP(n) can be naturally identified
with the (unique) set partition π  ∈ SP(m) such that D(π  ) = D(π), and it is
easy to see that the supercharacter ξπ ∈ Y(n) can be naturally extended to the
supercharacter ξπ ∈ Y(m); hence, without loss of generality, we may assume that
Y(n) ⊆ Y(m). More generally, every normalized supercharacter ξπ , for π ∈ SP(n),
can be naturally extended to an indecomposable supercharacter of U∞ , and hence
we may assume that ξπ ∈ ExSCh(U∞ ) for all n ∈ N and all π ∈ SP(n).
  
The set ExSCh(U∞ ) is larger than the union n∈N ξπ : π ∈ SP(n) . As a first
example, let m ∈ N be fixed, and define the supercentral function ξm : U∞ → C
 Indecomposable Supercharacters 9

on an element gπ = 1 + eπ ∈ U∞ , for π ∈ SP, by the rule

  
q −nstm (π) , if D(π) ∩ (m, n) : n ∈ N, m < n = ∅,
ξm (gπ ) = (2.4)
0, otherwise,
 
where nstm (π) = (i, j) ∈ D(π) : m < i . It is clear that ξm is a supercharacter
of U∞ ; in fact, we have the following.
Proposition 2.1. In the notation as above, ξm ∈ ExSCh(U∞ ) for all m ∈ N.
Proof. Let m ∈ N be arbitrary and suppose that ξm ∈ / ExSCh(U∞ ). Then, ξm =
tζ + (1 − t)ζ  for some t ∈ R, 0 < t < 1, and some ζ, ζ  ∈ SCh(U∞ ). Let n ∈ N be
such that m < n. Then,
1
Un (ξm ) = n−m 1 + ξm,m+1 + · · · + ξm,n ,
ResU∞

q
in fact, we have
−1 U
ResU∞
Un (ξm ) = ξm,n+1 (1) ResUn+1
n
(ξm,n+1 )
1 U
= ResUn+1 (ξm,n+1 ),
(q − 1)q n−m n

whereas
U
ResUn+1
n
(ξm,n+1 ) = (q − 1) 1 + ξm,m+1 + · · · + ξm,n .
Therefore, the restriction ResU∞
Un (ζ) decomposes as a nonnegative linear combi-
nation (with rational coeficients) of some of the supercharacters 1 and ξm,k for
m < k ≤ n; moreover, if ξm,n+1 is a constituent of ResU∞
Un+1 (ζ), then
U
ResU∞ n+1 U∞
Un (ζ) = ResUn (ξm,n+1 ) = ResUn (ξm ).

Un (ζ) = ResUn (ξm ), then ξm,n+1 ∈ SCh(Un+1 ) must be a

It follows that, if ResU∞ U∞


constituent of the restriction ResU∞
Un+1 (ζ ), and thus

ResU∞ U∞
Un (ζ ) = ResUn (ξm )
U
(because ResUn+1
n
(ξm,n+1 ) must be a constituent of ResU∞
Un (ξm )). Therefore, we
conclude that
U∞ 
Un (ξm ) = tResUn (ζ) + (1 − t)ResUn (ζ )
ResU∞ U∞

Un (ζ) + (1 − t)ResUn (ξm ),

= tResU∞ U∞

Un (ζ) = ResUn (ξm ), a contradiction. It follows that ξm ∈

which implies that ResU∞ U∞

ExSCh(U∞ ), as required. 
We observe that equation (2.3) implies that, for every m ∈ N and every
 ∞
g ∈ U∞ , the sequence ξm,n (g) n=m+1 is convergent and that
ξm,n (g)
ξm (g) = lim ξm,n (g) = lim .
n→∞ n→∞ ξm,n (1)
10 C.A.M. André, F. Gomes and J. Lochon

3. The superbranching graph

A Bratteli diagram is an infinite directed graph Γ such that the vertex set V(Γ)
and the edge set E(Γ) can be partitioned into finite sets
V(Γ) = V0 (Γ) V1 (Γ) V2 (Γ) ··· and E(Γ) = E1 (Γ) E2 (Γ) ···
satisfying the following properties:
1. V0 (Γ) consists of a single vertex which we denote by ∅.
2. If t : E(Γ) → V(Γ) is the associated target map and s : E(Γ) → V(Γ) is the
associated source map, then
(a) t En (Γ) ⊆ Vn (Γ) and s En (Γ) ⊆ Vn−1 (Γ) for all n ∈ N;
(b) the set s−1 (λ) is nonempty for all λ ∈ V(Γ), and the set t−1 (λ) is
nonempty for all λ ∈ V(Γ) \ V0 (Γ).
To simplify the notation, we identify the graph Γ with its vertex set V(Γ), and
write Γn = Vn (Γ) for all n ∈ N0 1 . The level of a vertex λ ∈ Γ is the unique n ∈ N0
such that λ ∈ Γn ; we write n = |λ|. If e ∈ E(Γ) is an edge with λ = s(e) and
μ = t(e), then we represent e as λ  μ; notice that we must have |μ| = |λ| + 1.
Notice that we are implicitly assuming that a Bratelli diagram Γ is oriented so
that any vertex of level n is always “below” some vertex of level n + 1, and any
vertex of level n + 1 is “above” some vertex of level n. We also assume that there
is an edge multiplicity function κ : E(Γ) → R+ which assigns to an edge λ  μ,
for λ, μ ∈ Γ, a strictly positive real number κ(λ, μ); we refer to κ(λ, μ) as the
multiplicity of the edge λ  μ, and we emphasise that κ(λ, μ) is not necessarily a
positive integer. Following [25], we refer to the pair (Γ, κ) as a branching graph.
 Our main example occurs when we consider the locally finite group U∞ =
n∈N Un . We define the superbranching graph of U∞ to be the Bratelli diagram
Γ = SP with set of vertices
SP = SP(0) SP(1) SP(2) ··· SP(n) ···
where SP(0) consists of a single vertex ∅ (the unique partition of the empty set
∅). For simplicity, we set U0 = {1} (the trivial group); notice that U0 is the trivial
k-algebra group U0 = 1 + u0 where we agree that u0 = {0}, and the supercharacter
ξ∅ is the trivial character of U0 . In order to define the edges of Γ, we first prove
the following result (see [14, Theorem 6.4], or [7, Theorem 2.21], for the case of
the standard supercharacter theory of finite algebra groups).
Proposition 3.1. For each n ∈ N, the restriction of any supercharacter of Un is a
nonnegative integer linear combination of supercharacters of Un−1 .
Proof. It follows easily from the results of [2] (namely, Lemma 10 and Lemma 7,
respectively) that the following formulae holds for every 1 ≤ i < j < k ≤ n:
⎧
⎨ξi,j ,   if j < n,
• ResUUn−1 (ξi,j ) = ⎩
n

(q − 1) 1Un−1 + i<j  <n ξi,j  , if j = n.

1 Throughout the paper, we set N0 = N  {0}.

 Indecomposable Supercharacters 11
 

• ξi,j ξk,j = (q − 1) ξi,j + k<j  <j ξi,j ξk,j  .

The result now follows using the factorisation equation (2.2) and applying succes-
sively these formulae; notice that, if π ∈ SP(n) and σ ∈ SP(n − 1) are such that
the supercharacter ξσ of Un−1 occurs in the restriction of the supercharacter
 ξπ of
Un , then D(σ) ⊆ (i, j) ∈ [[n]] : (i, k) ∈ D(π) for some i < k ≤ n . 

It follows from the previous proposition that, for every n ∈ N0 and every
π  ∈ SP(n + 1) there exist nonnegative integers κ(π, π  ), for π ∈ SP(n), such that

κ(π, π  )ξπ .
U
ResUn+1
n
(ξπ ) =

π∈SP(n)

Then, two vertices π, π ∈ SP are joined by a directed edge π  π  if and only



if there exists n ∈ N0 such that π ∈ SP(n), π  ∈ SP(n + 1) and κ(π, π  ) = 0;

moreover, we obtain the edge multiplicity function κ : E(SP) → N which assigns
to each edge π  π  the (positive) multiplicity κ(π, π  ) of ξπ in the decomposition
U
of ResUn+1
n
(ξπ ). By the way of example, consider the set partition π = 16/24/35
of [6] and the supercharacter ξπ = ξ1,6 ξ2,4 ξ3,5 of U6 . Using the formulae above, we
deduce that

U5 (ξπ ) = (q − 1)ξ2,4 ξ3,5 + (q − 1)ξ1,2 ξ2,4 ξ3,5 + (q − 1) ξ1,4 ξ3,5

ResU6 2

+ (q − 1)2 ξ1,4 ξ2,3 ξ3,5 + q(q − 1)2 ξ1,5 ξ2,4 ,

and thus we have the following connections in the superbranching graph
16/24/35

ii i iiis4ss9 O eKKKjUUUU
KK UUUUU q(q−1)2
q−1 ii i i s UUUU
i iii s ss q−1 KK
KK UUUU
iiiii s ss q−1 (q−1)2 KK UUUU
iii ss KK UUU
ii s
1/24/35 124/35 14/2/35 14/235 15/24/3

where we label the edges according to their multiplicities.

A complex-valued function ϕ : Γ → C defined on the vertices of a branching
graph (Γ, κ) is said to be harmonic if it satisfies the condition

ϕ(λ) = κ(λ, μ)ϕ(μ) (3.1)
μ∈Γ, λ μ

for all λ ∈ Γ. We denote by H(Γ) the set of all harmonic functions defined on Γ,
and consider H(Γ) as a topological space with respect to the topology of pointwise
convergence. We denote by H+ (Γ) the subset of H(Γ) consisting of nonnegative
harmonic functions (that is, harmonic functions with nonnegative real values),
and by H1+ (Γ) the subset consisting of all functions ϕ ∈ H+ (Γ) which satisfy the
normalising condition ϕ(∅) = 1. It is clear that H1+ (Γ) is a convex subset of H(Γ),
and it is also a compact measurable space; we refer to [31] for the general theorems
about convex compact measurable sets. Let Ex(Γ) denote the set of extreme points
of H1+ (Γ); this is a set of type Gδ , and hence it is a Borel measurable set.
12 C.A.M. André, F. Gomes and J. Lochon

In our running example, every supercharacter ξ of U∞ determines uniquely

a harmonic function ϕ ∈ H1+ (SP) on its superbranching graph SP by means of the
formula 
ResU ∞
Un (ξ) = ϕ(π)ξπ ;
π∈SP(n)

in fact, it is clear that ResU∞

Un (ξ)
is a superclass function of Un (hence, it is a
C-linear combination of indecomposable supercharacters of Un ). Conversely, every
harmonic function ϕ ∈ H1+ (SP) determines uniquely a supercharacter of U∞ .
Indeed, the following result is true.
Theorem 3.2. Let (SP, κ) be the superbranching graph of U∞ . Then, for every har-
monic function ϕ ∈ H1+ (SP), there exists a unique supercharacter ξϕ ∈ SCh(U∞ )
such that 
ξϕ (g) = ϕ(π)ξπ (g)
π∈SP(n)
for all g ∈ Un and all n ∈ N. Moreover, the mapping ϕ → ξϕ establishes an affine
homeomorphism between H1+ (SP) and SCh(U∞ ).
Proof. Let ϕ ∈ H1+ (SP). Then, the harmonicity of ϕ shows that ξϕ is well defined:
indeed, for every g ∈ Un for n ∈ N, we have
    
ξϕ (g) = ϕ(π)ξπ (g) = κ(π, π  )ϕ(π  ) ξπ (g)
π∈SP(n) π∈SP(n) π  ∈SP(n+1)
    
 
= ϕ(π ) κ(π, π )ξπ (g) = ϕ(π  )ξπ (g).
π  ∈SP(n) π∈SP(n) π  ∈SP(n)

It is clear that ξϕ is a superclass function of U∞ ; it is a supercharacter because ϕ is

nonnegative and normalized (hence, ξϕ (1) = ϕ(∅) = 1). The remaining assertions
are clear. 
Let (Γ, κ) be a branching graph. If τ = (λ1 , . . . , λk ) ∈ Γk is a path in Γ
(hence, λi  λi+1 is an edge of Γ for all 1 ≤ i < k), then we define the weight of
τ to be the product
w(τ ) = κ(λ1 , λ2 ) · · · κ(λk−1 , λk ).
On the other hand, if λ, μ ∈ Γ are such that |λ| < |μ|, then we define the dimension
of the interval [λ, μ] ⊆ Γ to be the sum

dim(λ, μ) = w(τ )
τ ∈P(λ,μ)

where P(λ, μ) denotes the set of all paths joining λ and μ; we extend this definition
to all vertices λ, μ ∈ Γ by agreeing that dim(λ, λ) = 1 and that dim(λ, μ) = 0 if
λ = μ and P(λ, μ) is empty. Finally, we define the dimension of a vertex λ ∈ Γ to
be the dimension of the interval [∅, λ] ⊆ Γ, that is,
dim(λ) = dim(∅, λ), λ ∈ Γ.
 Indecomposable Supercharacters 13

We note that, for every λ, μ ∈ Γ with |λ| ≤ |μ|, the dimension dim(λ, μ) may be
computed recursively by the formulae
 
dim(λ, μ) = dim(λ, ν)κ(ν, μ) = κ(λ, ν) dim(ν, μ).
ν∈Γ, ν μ ν∈Γ, λ ν

If (SP, κ) is the superbranching graph of U∞ and if m, n ∈ N are such that

m < n, then

ResU
Um (ξπ ) =
n
dim(π  , π)ξπ
π  ∈SP(m)

for all π ∈ SP(n); in particular, since U0 = {1} and ξ∅ is the trivial character of
U0 , we have ResU U0 (ξπ ) = ξπ (1)ξ∅ , and thus dim(π) = ξπ (1) for all π ∈ SP.
n

An infinite path (λ1 , . . . , λn , . . .) in a branching graph (Γ, κ) is said to be

regular if the limit
dim(λ, λn )
lim
n→∞ dim(λn )

exists for all λ ∈ Γ. We note that, for every n ∈ N, the function ϕn : Γn → R+ ,

defined by
dim(λ, λn )
ϕn (λ) = , λ ∈ Γ,
dim(λn )
is almost harmonic: the harmonicity condition holds for all λ ∈ Γ, λ = λn . There-
fore, the limit above (if it exists) defines a normalized nonnegative harmonic func-
tion on Γ, and thus there exists an harmonic function ϕτ associated with every
regular path τ in Γ. The following result asserts that every extreme harmonic
function ϕ ∈ Ex(Γ) can be obtained in this way; it is an easy consequence of a
(slightly) variation of Birkhoff’s pointwise ergodic theorem, and a proof can be
found in [30, Appendix 3] (see also [27]).
Theorem 3.3. Let (Γ, κ) be a branching graph. For every ϕ ∈ Ex(Γ), there exists
a regular path (λ1 , . . . , λn , . . .) in Γ such that
dim(λ, λn )
ϕ(λ) = lim
n→∞ dim(λn )
for all λ ∈ Γ.
As a consequence, we obtain the following result.
Theorem 3.4. If ξ is an indecomposable supercharacter of U∞ , then there exists
a sequence ξ1 , ξ2 , . . . , ξn , . . . of supercharacters
 ∞ of U1 , U2 , . . . , Un , . . ., respectively,
such that the corresponding sequence ξn n=1 of normalized supercharacters is
pointwise convergent and
ξn (g)
ξ(g) = lim
n→∞ ξn (1)

for all g ∈ U∞ .
14 C.A.M. André, F. Gomes and J. Lochon

Proof. Let ξ ∈ SCh(U∞ ) be arbitrary, and consider the superbranching graph

(SP, κ) of U∞ . By Theorem 3.2, there exists ϕ ∈ H1+ (SP) such that ξ = ξϕ , and
the extremality of ξ implies that ϕ ∈ Ex(SP). By the previous theorem, there
exists a regular path (π1 , . . . , πn , . . .) in SP such that
dim(π, πn )
ϕ(π) = lim
n→∞ dim(πn )
for all π ∈ SP. Then, for every m ∈ N and every g ∈ Um , we deduce that
   dim(π, πn )

ξ(g) = ξϕ (g) = ϕ(π)ξπ (g) = lim ξπ (g)
n→∞ dim(πn )
π∈SP(m) π∈SP(m)
  
1 ξπn (g)
= lim dim(π, πn )ξπ (g) = lim ,
n→∞ dim(πn ) n→∞ ξπ (1)
n
π∈SP(m)

because ResU
Um (ξπn ) =
n
π∈SP(m) dim(π, πn )ξπ for all n ∈ N, n ≥ m. 

Example 1. As a first example, let I be an arbitrary  subset of N. For every n ∈ N,

let 1 ≤ i1 < · · · < ir ≤ n be such that I ∩ [n] = i1 , . . . , ir , and let πn ∈ SP(n)
be unique partition of [n] such that
 
D(πn ) = (is , n − s + 1) : 1 ≤ s ≤ r ∩ [[n]].
 ∞
Then, the sequence ξπn n=1 of normalized supercharacters is pointwise conver-
gent; indeed, if n ∈ N for σ ∈ SP(n), then
⎧
⎨ (−1)|D(πn )∩D(σ)|
, if D(σ) ⊆ Reg(πn ),
ξπn (g) = q nstπn (σ) (q − 1)|D(πn )∩D(σ)|
⎩
0, otherwise,
and thus ⎧
⎨ 1
, if D(σ) ⊆ I c × N,
lim ξπn (g) = q nstI (σ)
n→∞ ⎩0, otherwise,

where I = N \ I and nstI (σ) = i∈I nsti (σ) (notice that this sum is finite). We
c

denote by ξI the supercharacter of U∞ defined by

ξI (g) = lim ξπn (g)
n→∞

for all g ∈ U∞ , and observe that

  
ξI (g) = ξi (g) = lim ξi,n (g) ;
n→∞
i∈I i∈I

notice that, for every g ∈ U∞ , the product i∈I ξi (g) is finite. We will prove later
that ξI is an indecomposable supercharacter of U∞ .
 Indecomposable Supercharacters 15

Example 2. As a second example, let π be an arbitrary set partition of N and define

the set of arcs D(π) as in the finite case; we will denote by SP(N) the set consisting
of all set partitions of N. With π we associate a supercharacter ξπ ∈ SCh(U∞ )
as follows. For every n ∈ N, let πn ∈ SP(n) be the unique set partition such
 ∞
that D(πn ) = D(π) ∩ [[n]]. Then, the sequence ξπn n=1 is pointwise convergent;
 ∞
indeed, for every g ∈ U∞ , the sequence ξπn (g) n=1 is eventually constant. We
thus define
ξπ (g) = lim ξπn (g)
n→∞
for all g ∈ U∞ . We also note that ξπ admits the representation the infinite product

ξπ = ξi,j .
(i,j)∈D(π)

As in the previous example, we will prove that ξπ is an indecomposable superchar-

acter of U∞ .
We are now able to prove the following result.
Theorem 3.5. Forevery∞ n ∈ N, let ξn be a supercharacter of Un , and assume
that the sequence ξn n=1 is pointwise convergent. If ξ ∈ SCh(U∞ ) is given by
ξ(g) = limn→∞ ξn (g) for all g ∈ U∞ , then ξ = ξI ξπ for some subset I ⊆ N and
some partition π ∈ SP(N) such that D(π) ∩ (I × N) = ∅.
Proof. For every n ∈ N, let πn ∈ SP(n) be such that ξn = ξπn ; moreover, for
(i) (i)
every i ∈ N, let πn ∈ SP(n) be defined by D(πn ) = D(πn ) ∩ {i} × [n] , let
 (i)
D(i) = n∈N D(πn ) and, in the case where D(i) is finite and nonempty, let ji =
 
max j ∈ N : (i, j) ∈ D(i) . It is easy to see that, for every i ∈ N, the sequence
 ∞
ξπ(i) n=1 is pointwise convergent; indeed, for every g ∈ U∞ , we have
n
⎧
⎪
⎪ ξi (g), if D(i) is infinite,
⎪
⎪
⎨ξ (g), if D(i) is finite and
lim ξπ(i) (g) =  
i,ji
(i)
n→∞ n ⎪
⎪ n ∈ N : D(πn ) = ∅ is infinite,
⎪
⎪  
⎩ (i)
1, if n ∈ N : D(πn ) = ∅ is finite.
Since  
ξπn = ξi,j = ξπ(i)
n
(i,j)∈D(πn ) i∈N

for all n ∈ N, we conclude that

   
ξ(g) = lim 
ξπ(i) (g) = 
lim ξπ(i) (g)
n→∞ n n→∞ n
i∈N i∈N
  
= ξI (g) lim 
ξπ(i) (g)
n→∞ n
i∈I c
16 C.A.M. André, F. Gomes and J. Lochon
 
where I = i ∈ N : D(i) is infinite . Let ξ  ∈ SCh(U∞ ) be defined by

ξ  (g) = lim ξπ(i) (g)
n→∞ n
i∈I c

for all g ∈ U∞ ; by the above, it is clear that, if g ∈ U∞ , then


ξ  (g) = ξi,j (g) i
i∈I 

where I  is the subset of I c consisting of all i ∈ I c for which D(i) is finite and
 (i) 
n ∈ N : D(πn ) = ∅ is infinite.
 We claim that there exists a set partition π ∈ SP(N) such that D(π) =
(i, ji ) : i ∈ I  . Indeed, assume that this is not the case; hence, there must exist
i, i ∈ N with i < i and such that ji = ji . Then, by the definition, there exists
an increasing sequence n1 , n2 , n3 , . . . ∈ N such that (i, ji ) ∈ D(πnk ) for all k ∈ N;
similarly, there exists an increasing sequence n1 , n2 , n3 , . . . ∈ N such that (i , ji ) ∈
D(πnk ) for all k ∈ N. By equation (2.3), we have ξπnk (1 + ei,ji ) = −1 and ξπn (1 +
k
ei,ji ) = 1 for all k ∈ N, and thus
lim ξπnk (1 + ei,ji ) = −1 = 1 = lim ξπn (1 + ei,ji ).
k→∞ k→∞ k
 ∞
This is impossible because the sequence ξπn (1 + ei,ji ) n=1 is convergent, and so
our claim is true.
The proof is complete. 

Corollary 3.6. If ξ is an indecomposable supercharacter of U∞ , then there exist a

subset I ⊆ N and a partition π ∈ SP(N) such that D(π)∩(I ×N) = ∅ and ξ = ξI ξπ .
Proof. It is enough to apply Theorem 3.4 (and the previous theorem). 

4. Multiplicative graphs
A branching graph (Γ, κ) is said to be multiplicative if there is a unital associative
graded R-algebra

A= An
n∈N0

and an (injective) map ι : Γ → A such that the following properties are satisfied:
1. ι(∅) = 1.  
2. For every n ∈ N0 , the
set ι(λ) : λ ∈ Γn is an R-basis of An .
3. The element ι() = λ∈Γ1 κ(∅, λ)ι(λ) ∈ A1 satisfies

ι(λ)ι() = κ(λ, μ)ι(μ)
μ∈Γn+1

for all λ ∈ Γn and all n ∈ N.

 Indecomposable Supercharacters 17

If this is the situation, we let2


K= R+
0 ι(λ)
λ∈Γ
 
be the cone in A spanned by ι(λ) : λ ∈ Γ , and let Mult+ 1 (A) be the subset of
the dual vector space A∗ consisting of all multiplicative linear functions φ : A → R
which are non-negative on K and satisfy φ(ι()) = 1. We note that the inclusion
ι : Γ → A induces naturally a bijective map ι∗ : A∗ → F(Γ, R) where F(Γ, R)
denotes the set consisting of all functions Γ → R; indeed, ι∗ (φ)(λ) = φ(ι(λ)) for
all φ ∈ A∗ and all λ ∈ Γ.

Proposition 4.1. The superbranching graph (SP, κ) of U∞ is multiplicative.

Proof. For every n ∈ N, let SCn denote the vector space over R spanned by the
set Y(n) = ξπ : π ∈ SP(n) ; hence, Y(n) is an R-basis of SCn . Then, we define

SC = SCn
n∈N0

where we agree that SC0 = R. We define an (associative) multiplication on SC

as follows. (We should mention that, if we define SCn , for n ∈ N, to be the vector
space over R spanned by the set of the  (standard) supercharacters of the algebra
group Un , then the vector space SC = n∈N0 SCn has a Hopf algebra structure
studied in [1] (see also [7]); the algebra structure considered in the present work
is closely related to the algebra structure of the Hopf dual of the one studied in
that paper.)
For every m, n ∈ N, we naturally identify the direct product Um × Un with
the subgroup
  
m,n = g 0
U : g ∈ Um , h ∈ Un
0 h
of Um+n . On the other hand, let ξ ∈ SCm and ζ ∈ SCn be arbitrary, let ξ ×
ζ : Um × Un → C be the function defined by (ξ × ζ)(g, h) = ξ(g)ζ(h) for all g ∈ Um
and all h ∈ Un , and let (ξ × ζ)◦ : Um+n → C be the extension by zero of ξ × ζ
to Um+n . Notice that, if ξ = ξσ is a supercharacter of Um (corresponding to a set
partition σ ∈ SP(m)) and ζ = ξτ is a supercharacter of Un (corresponding to a set
partition τ ∈ SP(n)), then ξ × ζ = ξσ × ξτ is a supercharacter of Um × Un which
corresponds (under the similar natural construction) to the (Bm × Bn )-orbit of
the element ϑσ × ϑτ ∈ (um × un )◦ . As in [14], we define the superinduced function
U
SIndUm+n
m ×Un
(ξ × ζ) : Um+n → C by the formula
1 
(ξ × ζ)◦ (1 + h(g − 1)k)
U
SIndUm+n
m ×Un
(ξ × ζ)(g) =
|Bm+n ||Bm ||Bn |
h,k∈Bm+n

2 Throughout the paper, we define R+ +

0 = R  {0}.
18 C.A.M. André, F. Gomes and J. Lochon

U
for all g ∈ Um+n . By the definition, it is clear that SIndUm+n
m ×Un
(ξ ×ζ) is a superclass
function of Um+n , and thus
U

SIndUm+n
m ×Un
(ξ × ζ) = κξ,ζ (π) ξπ
π∈SP(m+n)

where κξ,ζ (π) ∈ C for all π ∈ SP(m + n). The multiplicities κξ,ζ (π) may be calcu-
lated using a version for supercharacters of the well-known Frobenius reciprocity.
Indeed, as in [14, Lemma 6.7], it is not difficult to check that
 U   U 
SIndUm+n
m ×Un
(ξ × ζ), ξπ = ξ × ζ, ResUm+n
m ×Un
(ξπ )
for all π ∈ SP(m + n); as usual, for any finite group G, we denote by ·, · the
Frobenius scalar product of class functions of G. Therefore, since Y(m + n) is
as orhogonal basis of SCm+n (with respect to the Frobenius scalar product), we
deduce that
   U   U 
κξ,ζ (π) ξπ , ξπ = SIndUm+n
m ×Un
(ξ × ζ), ξπ = ξ × ζ, ResUm+n
m ×Un
(ξπ ) ,
and thus Proposition 3.1) implies that κξ,ζ (π) is a nonnegative real number for
all π ∈ SP(m + n); in fact, if both ξ and ζ are supercharacters of Um and Un ,
respectively, then κξ,ζ (π) is a nonnegative rational number for all π ∈ SP(m + n).
U
In particular, we conclude that SIndUm+nm ×Un
(ξ × ζ) ∈ SCm+n , and this allows us to
define
U
ξ · ζ = SIndUm+n
m ×Un
(ξ × ζ).
to check that the mapping (ξ, ζ) → ξ · ζ defines an associative product
It is routine 
on SC+ = n∈N SCn ; we extend this associative product
 to SC by imposing
1 ∈ R = SC0 to be the identity of SC. Thus, SC = SCn becomes an
 n∈N0 
associative graded R-algebra with a countable R-basis ξπ : π ∈ SP ; we recall
that ξ∅ = 1U0 is the unit character of the trivial group U0 = {1}.
We now define the (injective) map ι : Γ → SC by the rule
 −1
ι(π) = ξπ , ξπ ξπ
 
for all π ∈ SP; notice that ι(∅) = ξ∅ = 1 ∈ SC0 , that ι(π) : π ∈ SP is an R-basis
of SC, and that ι({1}) = ξ{1} = 1U1 is the unique supercharacter of U1 = {1}
(indeed, π = {1} is the unique set partition of {1}). Let n ∈ N be arbitrary,
and recall that Un = Un × U1 (as subgroups of Un+1 ). Thus, it is obvious that
ξ × ξ{1} = ξ, and thus
U
ξ · ξ{1} = SIndUn+1
n
(ξ)
for all ξ ∈ SCn . Since
     
(ξπ ) = κ(π, π  ) ξπ , ξπ
U U
SIndUn+1
n
(ξπ ), ξπ = ξπ , ResUn+1
n

for all π ∈ SP(n) and all π  ∈ SP(n + 1), we conclude that

 
 κ(π, π  ) ξπ , ξπ
U
ξπ · ξ{1} = SIndUn+1 (ξπ ) =   ξπ
n

ξπ , ξπ
π ∈SP(n+1)
 Indecomposable Supercharacters 19

for all π ∈ SP(n). Therefore, if we define ι() = ξ{1} , then


ι(π) · ι() = κ(π, π  )ι(π  )
π  ∈SP(n+1)

for all π ∈ SP(n), and this completes the proof. 

The main goal of this section is to obtain a parametrisation of the indecom-
posable supercharacters of U∞ in terms of certain ring homomorphisms φ : SC →
R. To this end, we will refer to the following Kerov–Vershik “ring theorem”; a
proof can be found in [17, Proposition 8.4] (we should mention that the proof does
not depend on the commutativity of the R-algebra A (as it is required there)).
Theorem 4.2 (Kerov–Vershik). Let A be a unital associative algebra over R and
let K ⊆ A be a convex cone satisfying the following conditions 3 :
1. K − K = A (that is, K generates A).
2. KK ⊆ K (that is, K is stable under multiplication).
3. K is spanned by a countable set of elements.
4. For every a ∈ A, there exists ε ∈ R+ such that 1 − εa ∈ K (in particular,
1 ∈ K).
If L denotes the convex set consisting of all linear forms φ ∈ A∗ which are non-
negative on K and satisfy φ(1) = 1, then a linear form φ ∈ L is an extreme point
of L if and only if φ is multiplicative (which means that φ(ab) = φ(a)φ(b) for all
a, b ∈ A).
The following result is a non-commutative version of a theorem proved by
Kerov–Vershik in [28]; for convenience of the reader, we give here a detailed proof
(see also [29, Theorem 2.5]).
Theorem 4.3. Let (Γ, κ)  be a multiplicative branching graph associated with a
graded R-algebra A = A via an injective map ι : Γ → A. Let K denote
 0 n
n∈N 
the cone in A spanned by ι(λ) : λ ∈ Γ , and assume that K is stable under mul-
tiplication. Then, the natural map ι∗ : A∗ → F(Γ, R) induces an homeomorphism
between Mult+ 1 (A) and the subset Ex(Γ) consisting of all extreme harmonic func-
tions defined on Γ.
Proof. Let I be the ideal of A generated by the relation ι() − 1, let A  = A/I be

the quotient R-algebra, and let ν : A → A be the natural epimorphism; for every
a ∈ A, we will denote by  a the image ν(a) ∈ A.
 = ν(K), and note that K
Let K  is a cone in A
 satisfying 1 = ν(1) = ν(ι()) ∈
   
K. Since KK ⊆ K, it is clear that K K ⊆ K; moreover, since K spans A, it is
also obvious that K spans A.  On the other hand, since Γ has a countable set of

vertices, K has a countable set of generators; indeed, K  is spanned by the set
ι(Γ) = ν(ι(Γ)). Finally, let 
a∈A  be arbitrary; we claim that there exists ε ∈ R+

3 For
  
arbitrary subsets X, Y ⊆ A, we define X − Y = x − y : x ∈ X, y ∈ Y and XY = xy : x ∈

X, y ∈ Y
20 C.A.M. André, F. Gomes and J. Lochon

such that 1 − ε  Without loss of generality, we may assume that 

a ∈ K. 
a ∈ K;
   
indeed, since K − K = A, there exist b,  
c ∈ K such that  
a = b− c, and thus, if
ε ∈ R+ is such that 1 − εb ∈ K, a = 1 − εb + ε
 then 1 − ε  In fact, since K
c ∈ K.  is
spanned by ι(Γ), it is enough to show that, for every λ ∈ Γ, there exists ε ∈ R+
 Let n ∈ N, and let λ ∈ Γn . Then,
such that 1 − ει(λ) ∈ K.

ι()n = dim(μ)ι(μ),
μ∈Γn

and thus ι()n − dim(λ)ι(λ) ∈ K. Since ν(ι()n ) = ν(ι())n = 1, we conclude that


1 − dim(λ)ι(λ) = ν ι()n − dim(λ)ι(λ) ∈ ν(K) = K.

Therefore, the pair (A,  K)

 satisfies the conditions of Theorem 4.2, and so, if L 
 ∗
denotes the convex set consisting of all linear forms φ ∈ A which are nonnegative
on K 
 and satisfy φ(1) = 1, then a linear function φ ∈ L is extreme if and only if

φ is multiplicative.
∗
Now, consider the subset Mult+ 1 (A) of A , and let φ ∈ Mult1 (A) be arbitrary.
+

Then, φ(ι()) = 1; since φ(1) = 1 (because φ is multiplicative), we conclude that

φ(ι() − 1) = 0. Since φ is multiplicative, it follows that φ(a) = 0 for all a ∈ I,
and thus φ induces naturally a linear map φ ∈ A 
∗ (by the rule φ(ν(a)) = φ(a)

for all a ∈ A). It is clear that φ is multiplicative, nonnegative on K  and satisfies

φ(1) = 1. Therefore, φ ∈ Ex(L), and thus the mapping φ → φ defines a map

Λ : Mult1 (A) → Ex(L) which is clearly a homeomorphism.
+

Finally, we observe that there is a homeomorphism Λ : Ex(L)  → Ex(Γ). For

 ∗ ∗ 
every φ ∈ A , we define the function ι (φ) : Γ → R by


ι∗ (φ)(λ) = φ ι(λ)

for all λ ∈ Γ. Since ι() = 1, we deduce that

  

ι∗ (φ)(λ) = φ ι(λ) = φ ι()ι(λ) = φ κ(λ, μ)ι(μ)
μ∈Γn+1
 
= κ(λ, μ)φ ι(μ) = 
κ(λ, μ)ι∗ (φ)(μ)
μ∈Γn+1 μ∈Γn+1

for all λ ∈ Γn and all n ∈ N, which means that ι∗ (φ)  ∈ H(Γ). Therefore, the
 
 ∗  ∗ ∗
mapping φ → ι (φ) defines a map ι : A → H(Γ); since ι(λ) : λ ∈ Γ is an
R-basis of A, this map is bijective, and it is clearly a homeomorphism. It is also
 = H+ (Γ) and that ι∗ (Ex(L))
clear that ι∗ (L)  = Ex(H+ (Γ)) = Ex(Γ), Therefore,
1 1
ι induces the required homeomorphism Λ : Ex(L)
∗  → Ex(Γ), and this completes
the proof. 
 Indecomposable Supercharacters 21

By Theorem 3.2, a supercharacter ξ ∈ SCh(U∞ ) corresponds uniquely to the

harmonic function ϕ ∈ H1+ (SP) defined by the rule

ξ(g) = ϕ(π)ξπ (g)
π∈SP(n)
 −1
for all g ∈ Un and all n ∈ N; hence, if we set ξπ = ξπ , ξπ ξπ , then we have
 
U∞
ResUn (ξ), ξπ  
ϕ(π) =   = ResU ∞ 
Un (ξ), ξπ
ξπ , ξπ
for all π ∈ SP(n) and all n ∈ N. On the other hand, by the above, the harmonic
function
 ϕ corresponds
 uniquely to the linear function φ ∈ SC∗ defined on the
basis ξπ : π ∈ SP of SC by the rule
 
φ(ξπ ) = ϕ(π) = ResU ∞ 
Un (ξ), ξπ
for all π ∈ SP(n) and all n ∈ N. Therefore, if L(SC) denotes the convex subset
of SC∗ consisting
 of all normalized linear functions which are nonnegative on the
cone K = π∈SP R+ 0 ξπ in SC, then we obtain a homeomorphism Λ : SCh(U∞ ) →
L(SC) where, for every ξ ∈ SCh(U∞ ), the linear map Λ(ξ) ∈ SC∗ is defined by
  1 
Λ(ξ)(ξπ ) = ResU ∞
Un (ξ), ξπ = ξ(g)ξπ (g)
|Un |
g∈Un

for all π ∈ SP(n) and all n ∈ N.

Besides
 the supercharacter
 basis, the R-algebra SC also has a distinguished
basis κπ : π ∈ SP consisting of all superclass characteristic functions; for every
π ∈ SP, we define κπ : U∞ → C by

1, if g ∈ Kπ (∞),
κπ (g) =
0, otherwise,
for all g ∈ U∞ . Then, for every ξ ∈ SCh(U∞ ), we get
1  |Kπ |
Λ(ξ)(κπ ) = ξ(g)κπ (g) = ξ(gπ )
|Un | |Un |
g∈Un

for all π ∈ SP(n) and all n ∈ N. In particular, if we define

|Un |

κπ = κπ
|Kπ |
for all π ∈ SP(n) and all n ∈ N, then we obtain
Λ(κπ )(
κπ ) = δπ,π

 
for all π, π ∈ SP. Therefore, Λ(κπ ) : π ∈ SP is a basis of the graded
 dual vector

space SC∗gr = n∈N SC∗n , which is dual to the basis  κπ : π ∈ Γ of SC; we note
that SC∗gr should not be confused with the dual vector space SC∗ .
The multiplication in SC is easy to describe in terms of the superclass char-
acteristic functions; indeed, if we define π/π  ∈ SP to the unique set partition such
22 C.A.M. André, F. Gomes and J. Lochon
 
that D(π/π  ) = D(π) (i + n, j + n) : (i, j) ∈ D(π  ) , then it follows easily that
from the definition of superinduction (see also [7, Lemma 4.7]) that
π · 
κ π/π
κπ  = κ
for all π, π ∈ SP. Therefore, a linear map φ ∈ SC∗ is multiplicative if and only if


φ(
κπ/π ) = φ(
κπ )φ(
κπ  )
for all π, π  ∈ SP, and thus ξ ∈ SCh(U∞ ) is an indecomposable supercharacter if
and only if
ξ(gπ/π ) = ξ(gπ )ξ(gπ ) (4.1)
for all π, π  ∈ SP. (Recall that, for every π ∈ SP, gπ ∈ U∞ is the “normalized”
element gπ = 1 + eπ in the superclass Kπ .)
In particular, we conclude that every elementary supercharacter of U∞ is
indecomposable; by an elementary supercharacter of U∞ we mean, either a super-
character of the form ξi for i ∈ N, or a supercharacter of the form ξi,j for i, j ∈ N
with i < j.
Theorem 4.4. A supercharacter ξ ∈ SCh(U∞ ) is indecomposable if and only if
ξ = ξI ξπ for some subset I ⊆ N and some partition π ∈ SP(N) such that D(π) ∩
(I × N) = ∅.
Proof. By virtue of Corollary 3.6, it remains to prove that, for every I ⊆ N and
every π ∈ SP(N) such that D(π)∩(I × N) = ∅, the supercharacter ξ = ξI ξπ is inde-
 
composable. Since ξI = i∈I ξi and ξπ = (i,j)∈D(π) ξi,j , this follows immediately
because every elementary supercharacter is indecomposable. 
Acknowledgment
Many thanks to Eric Marberg for the very useful discussions we had during his
2015 visit to the Department of Mathematics of the Faculty of Sciences, University
of Lisbon. The authors also thank the anonymous referees for the very useful
suggestions (and corrections) which certainly helped us to improve this final version
of the paper.

References
[1] M. Aguiar, C.A.M. André, C. Benedetti, N. Bergeron, Z. Chen, P. Diaconis, A.
Hendrickson, S. Hsiao, I.M. Isaacs, A. Jedwab, K. Johnson, G. Karaali, A. Lauve,
T. Le, S. Lewis, H. Li, K. Magaard, E. Marberg, J.-C. Novelli, A. Pang, F. Saliola,
L. Tevlin, J.-Y. Thibon, N. Thiem, V. Venkateswaran, C.R. Vinroot, N. Yan, M.
Zabrocki, Supercharacters, symmetric functions in noncommuting variables, and
related Hopf algebras. Adv. Math. 229 (2012), no. 4, 2310–2337.
[2] C.A.M. André, Basic characters of the unitriangular group. J. Algebra 175 (1995),
287–319.
[3] C.A.M. André, Basic sums of coadjoint orbits of the unitriangular group. J. Algebra
176 (1995), 959–1000.
 Indecomposable Supercharacters 23

[4] C.A.M. André, Irreducible characters of finite algebra groups. In Matrices and
Group Representations, Coimbra, 1998. Textos Mat. Sér. B 19 (1999), 65–80.
[5] C.A.M. André, The basic character table of the unitriangular group. J. Algebra 241
(2001), 437–471.
[6] C.A.M. André, Basic characters of the unitriangular group (for arbitrary primes).
Proc. Amer. Math Soc. 130 (2002), 1943–1954.
[7] C.A.M. André, Supercharacters of unitriangular groups and set partition combi-
natorics. CIMPA school: “Modern Methods in Combinatorics ECOS2013”, 2da
Escuela Puntana de Combinatoria: Escuela de Combinatoria del Sur, Universidad
Nacional de San Luis, Argentina, 2013.
[8] N. Bergeron and N. Thiem, A supercharacter table decomposition via power-sum
symmetric functions. Internat. J. Algebra Comput. 23 (2013), 763–778.
[9] R.P. Boyer, Characters and factor representations of the unitary group of the CAR-
algebra. J. Operator Theory 30 (1993), no. 2, 315–328.
[10] D. Bragg and N. Thiem, Restrictions of rainbow supercharacters. J. Algebra 451
(2016), 357–400.
[11] B. Chern, P. Diaconis, D.M. Kane, and R.C. Rhoades, Closed expressions for aver-
ages of set partition statistics. Res. Math. Sci. 1 (2014), no. 1, 1–32.
[12] B. Chern, P. Diaconis, D.M. Kane, and R.C. Rhoades, Central limit theorems for
some set partition statistics. Adv. in Appl. Math. 70 (2015), 92–105.
[13] D. De Stavola, A Plancherel measure associated to set partitions and its limit. Adv.
in Appl. Math. 92 (2018), 73–98.
[14] P. Diaconis and I.M. Isaacs, Supercharacters and Superclasses for Algebra Groups.
Trans. Amer. Math. Soc. 360 (2007), no. 5, 2359–2392.
[15] A. Dudko, Characters on the full group of an ergodic hyperfinite equivalence relation.
J. Funct. Anal. 261 (2011), 1401–1414.
[16] A. Dudko and K. Medynets, On characters of inductive limits of symmetric groups.
Available online at https://arxiv.org/pdf/1105.6325.pdf.
[17] A. Gnedin and G. Olshanski, Coherent permutations with descent statistic and the
boundary problem for the graph of zigzag diagrams. Int. Math. Res. Not. 2006, Art.
ID 51968.
[18] K.R. Goodearl, Partially ordered abelian groups with interpolation. Mathematical
Surveys and Monographs 20, Amer. Math. Soc., Providence, RI, 1986.
[19] E. Hirai and T. Hirai, Positive definite class functions on a topological group and
characters of factor representations. J. Math. Kyoto Univ. 45 (2005), no. 2, 355–379.
[20] E. Hirai and T. Hirai, Characters of wreath products of finite groups with the infinite
symmetric group. J. Math. Kyoto Univ. 45 (2005), no. 3, 547–597.
[21] E. Hirai, T. Hirai, and A. Hora, Limits of characters of wreath products Sn (T ) of
a compact group T with the symmetric groups and characters of S∞ (T ). I. Nagoya
Math. J. 193 (2009), 1–93.
[22] E. Hirai, T. Hirai, and A. Hora, Limits of characters of wreath products Sn (T ) of
a compact group T with the symmetric groups and characters of S∞ (T ). II. From
a viewpoint of probability theory. J. Math. Soc. Japan 60 (2008), no. 4, 1187–1217.
24 C.A.M. André, F. Gomes and J. Lochon

[23] I.M. Isaacs, Characters of groups associated with finite algebras. J. Algebra 177
(1995), 708–730.
[24] J. Keller, Generalized Supercharacter Theories and Schur Rings for Hopf Algebras.
Ph.D. thesis, University of Colorado at Boulder, 2014.
[25] S.V. Kerov, The boundary of Young lattice and random Young tableaux. Formal
Power Series and Algebraic Combinatorics (New Brunswick, NJ, 1994). DIMACS
Ser. Discrete Math. Theoret. Comput. Sc., vol. 24, Amer. Math. Soc., Providence,
RI, 1996, pp. 133–158.
[26] S.V. Kerov, Asymptotic representation theory of the symmetric group and its ap-
plications in analysis. Translated from the Russian manuscript by N.V. Tsilevich.
With a foreword by A. Vershik and comments by G. Olshanski. Translations of
Mathematical Monographs 219, Amer. Math. Soc., Providence, RI, 2003.
[27] S.V. Kerov and A.M. Vershik, Asymptotic character theory of the symmetric group.
Funct. Analysis and its Applications 15 (1981), 15–27.
[28] S.V. Kerov and A.M. Vershik, Characters and factor representations of the infinite
symmetric group. Dokl. Akad. Nauk SSSR 257 (1981), no. 5, 1037–1040. English
translation: Soviet Math. Dokl. 23 (1981), no. 2, 389–392.
[29] C. Lecouvey and P. Tarrago, Central measures on multiplicative graphs, represen-
tations of Lie algebras and weight polytopes.
Available online at https://arxiv.org/pdf/1609.00138.pdf.
[30] G. Olshanski and A.M. Vershik, Ergodic unitarily invariant measures on the space
of infinite Hermitian matrices. Contemporary Mathematical Physics: F.A. Berezin
Volume, Amer. Math. Soc. Transl. 175(2) (1996), 137–175.
[31] R.R. Phelps, Lectures on Choquet’s theorem. Van Nostrand Mathematical Studies,
Vol. 7, Princeton, NJ, 1966.
[32] H.L. Skudlarek, Die unzerlegbaren Charaktere einiger diskreter Gruppen. Math.
Annal. 233 (1976), 213–231.
[33] N. Thiem, Supercharacter theories of type A unipotent radicals and unipotent poly-
topes. Available online at https://arxiv.org/pdf/1512.03446.pdf.
[34] E. Thoma, Die unzerlegbaren, positiv-definiten Klassenfunktionen der abzahlbar un-
endlichen, symmetrischen Gruppe. Math. Z. 85 (1964), 40–61.
[35] N. Yan, Representation theory of the finite unipotent linear groups. Unpublished
Ph.D. Thesis, Department of Mathematics, University of Pennsylvania, 2001.

Carlos A.M. André, Filipe Gomes and Jocelyn Lochon

Centro de Análise Funcional, Estruturas Lineares e Aplicações (CEAFEL)
Departamento de Matemática
Faculdade de Ciências da Universidade de Lisboa
Campo Grande, Edifı́cio C6, Piso 2
1749-016 Lisboa, Portugal
e-mail: [email protected]
filipe13− [email protected]
[email protected]
Operator Theory:
Advances and Applications, Vol. 267, 25–64

c Springer International Publishing AG, part of Springer Nature 2018

A C ∗ -algebra of Singular Integral

Operators with Shifts and Piecewise
Quasicontinuous Coefficients
M. Amélia Bastos, Cláudio A. Fernandes and Yuri I. Karlovich

Abstract. The C ∗ -algebra B of bounded linear operators on the space L2 (T),

which is generated by all multiplication operators by piecewise quasicontin-
uous functions, by the Cauchy singular integral operator and by the range
of a unitary representation of a group G of orientation-preserving diffeomor-
phisms of T onto itself that have the same finite set of fixed points for all
g ∈ G \ {e}, is studied. A Fredholm symbol calculus for the C ∗ -algebra B
and a Fredholm criterion for the operators B ∈ B are established by using
spectral measures and the local-trajectory method for studying C ∗ -algebras
associated with C ∗ -dynamical systems.

Mathematics Subject Classification (2010). 45E05, 47A53, 47A67, 47B33,

47G10, 47L15.
Keywords. Singular integral operator with shifts, piecewise quasicontinuous
function, C ∗ -algebra, amenable group, local-trajectory method, spectral mea-
sure, representation of a C ∗ -algebra, symbol calculus, Fredholmness.

1. Introduction
Given a Hilbert space H, we denote by B(H) the C ∗ -algebra of all bounded linear
operators on H, by K(H) the ideal of all compact operators in B(H), and by
I ∈ B(H) the identity operator on H. If S, T ∈ B(H) and S − T ∈ K(H), we
say that the operators S and T are equivalent and write S  T . For an operator

This work was partially supported by the Fundação para a Ciência e a Tecnologia (Portuguese
Foundation for Science and Technology) through the projects UID/MAT/04721/2013 (Centro
de Análise Funcional, Estruturas Lineares e Aplicações) and UID/MAT/00297/2013 (Centro de
Matemática e Aplicações). The third author was also supported by the SEP-CONACYT Project
No. 168104 (México).
26 M.A. Bastos, C.A. Fernandes and Yu.I. Karlovich

A ∈ B(H), we denote by Aπ := A + K(H) the coset of A in the Calkin algebra

B π (H) := B(H)/K(H) and by |A| the essential norm of A, that is,
|A| = Aπ  = inf{A + K : K ∈ K(H)}.
An operator A ∈ B(H) is Fredholm on a Hilbert space H if and only if the coset
Aπ is invertible in the C ∗ -algebra B π (H). Given two C ∗ -algebras A and B, we
write A ∼
= B if these C ∗ -algebras are ∗ -isomorphic and hence isometric.
Let K := K(L2 (T)) be the ideal of compact operators acting on the Lebesgue
space L2 (T), where T is the unit circle in C with the length measure and anticlock-
wise orientation, and let P QC(T) be the C ∗ -subalgebra of L∞ (T) generated by
the quasicontinuous functions, QC(T), and by the piecewise continuous functions,
P C(T), on T. Consider the C ∗ -algebra
A := alg(P QC(T), ST ) ⊂ B(L2 (T)) (1.1)
generated by all the multiplication operators aI with a ∈ P QC(T) and by the
Cauchy singular integral operator ST defined by
!
1 ϕ(τ )
(ST ϕ)(t) := lim dτ,
ε→0 πi T\T(t,ε) τ − t

where T(t, ε) = {τ ∈ T : |τ − t| < ε}, t ∈ T.

Let G be a discrete group of orientation-preserving diffeomorphisms of T onto
itself that have the same finite set Λ ⊂ T of fixed points for all g ∈ G \ {e}, where
e is the unit of G and the group operation is given by (gh)(t) = h(g(t)) for all
g, h ∈ G and all t ∈ T. In that case the group G is commutative (see [24, p. 284]
and [10, § 2]) and therefore amenable [17]. Then the group G acts on the contour
T topologically freely [1], that is, for each finite set F ⊂ G and each open arc γ ⊂ T
there exists a point t ∈ γ such that the points g(t) for g ∈ F are pairwise distinct.
The aim of this paper is to construct a Fredholm symbol calculus (in other
words, a faithful representation of the quotient C ∗ -algebra Bπ := B/K in a Hilbert
space) and to establish a Fredholm criterion for the operators in the nonlocal C ∗ -
subalgebra
B := alg(A, UG ) = alg(P QC(T), ST , UG ) (1.2)
of B(L2 (T)) generated by all singular integral operators A ∈ A and by all unitary
weighted shift operators in the group UG := {Ug : g ∈ G}, where
(Ug ϕ)(t) := |g  (t)|1/2 ϕ(g(t)) for t ∈ T. (1.3)
∗
To study the C -algebra B, we apply the local-trajectory method combined with
using suitable spectral measures (see [18], [19] and [5]). Other approaches for study-
ing nonlocal operator algebras see in [1]–[3].
The C ∗ -algebra B in the case of piecewise slowly oscillating coefficients, a sub-
set of the P QC(T) coefficients, was studied in [5] (see also [6]–[9] for other classes
of groups G with more complicate actions). The C ∗ -algebra B with P QC(T) coef-
ficients was studied in [11] in the case of groups G acting freely on T, making use of
the local-trajectory method. In the present paper we assume that the shifts g ∈ G
 A C ∗ -algebra of Singular Integral Operators with Shifts 27

have the same finite set Λ of fixed points, which essentially complicates the study
of the C ∗ -algebra (1.2) because the action of the group G on the maximal ideal
space of a central subalgebra Z π of the quotient C ∗ -algebra Aπ := A/K of A given
by (1.1) stops to be topologically free. To overcome arising difficulties, we need
to apply the technique of spectral measures combined with the local-trajectory
method and to use deep properties of quasicontinuous functions investigated in
[25] and [26].
The paper is organized as follows. In Section 2, following [25]–[26], we con-
sider the C ∗ -algebra P QC(T) of piecewise quasicontinuous functions, collect their
properties and describe the fibers and the Gelfand topology on the maximal ideal
space M (P QC(T)) of P QC(T). Section 3 contains main results of the paper: the
Fredholm symbol calculus for the C ∗ -algebra B and a Fredholm criterion for the
operators B ∈ B. In Section 4 we recall the local-trajectory method and its gen-
eralization on the basis of spectral measures. In Section 5, following [26] and [14],
we describe the Fredholm symbol calculus for the C ∗ -algebra (1.1) and study the
central C ∗ -algebra Z π of Aπ generated by the cosets (aI)π with a ∈ QC(T) and
π
by the cosets HP,t related to singular integral operators with fixed singularities
at points t ∈ T. We describe here the maximal ideal space M (Z π ) of Z π and its
Gelfand transform, and calculate symbols of important operators Ug Vt ∈ A for
t ∈ T, where the operators Ug are related to shifts g ∈ G with a fixed point at t
and the operator Vt has a fixed singularity at t.
In Section 6 we construct a spectral decomposition of the quotient C ∗ -algebra
B into the orthogonal sum of three operator C ∗ -subalgebras Barc , B◦Λ and BΛ
π

invariant with respect to the action of the group G on M (Z π ), and present an

abstract Fredholm criterion for the operators B ∈ B in terms of invertibility of
their images Barc , BΛ◦ and BΛ in the C ∗ -algebras Barc , B◦Λ and BΛ , respectively.
In Section 7 we study the invertibility in the C ∗ -algebra Aarc ⊂ Barc . Applying
the local-trajectory method and the results of Section 7, in Section 8 we obtain the
invertibility criterion for the operators Barc ∈ Barc . In Section 9 we elaborate the
invertibility criterion for the operators BΛ◦ ∈ B◦Λ . Section 10 is devoted to studying
the invertibility of functional operators with piecewise quasicontinuous coefficients
and shifts in the group G. Finally, in Section 11 we prove that the invertibility
of the operators Barc ∈ Barc implies the invertibility of the operators BΛ ∈ BΛ .
Combining the results of Sections 8–11, we obtain the Fredholm criterion for the
operators B ∈ B in terms of their Fredholm symbols (see Section 3).

2. The C ∗ -algebra P QC(T)

Let L∞ (T) be the C ∗ -algebra of all bounded measurable functions on the unit
circle T := {z ∈ C : |z| = 1}. Let C(T) and P C(T) denote the C ∗ -subalgebras of
L∞ (T) consisting, respectively, of all continuous functions on T and all piecewise
continuous functions on T, that is, the functions having finite one-sided limits at
each point t ∈ T.
28 M.A. Bastos, C.A. Fernandes and Yu.I. Karlovich

" a Lebesgue measurable set E ⊂ T, we will denote by |E| its measure,

For
|E| := E dm. For each subarc I of T and each function f ∈ L1 (T), the average of
f over I is given by !
1
I(f ) := f dm.
|I| I
Taking δ > 0 and considering arcs I ⊂ T, we let
!
1
Mδ (f ) := sup |f (t) − I(f )| dm(t).
I⊂T, |I|≤δ |I| I

Obviously, 0 ≤ Mδ1 (f ) ≤ Mδ2 (f ) if 0 < δ1 ≤ δ2 ≤ 2π. Let

M0 (f ) := lim Mδ (f ).
δ→0

According to [25], a function f ∈ L (T) is said to have vanishing mean oscillation

1

on T if M0 (f ) = 0. The set of functions of vanishing mean oscillation on T is

denoted by V M O(T).
Let H ∞ (D) denote the Hardy space of bounded analytic functions on the
unit disc D := {z ∈ C : |z| < 1}, and let H ∞ denote the set of all functions in
L∞ (T) that are non-tangential limits on T of functions in H ∞ (D). Equivalently,
H ∞ consists of the functions f ∈ L∞ (T) for which the Fourier coefficients fn
defined for n ∈ Z by
! 2π
1
fn = f (eiθ )e−inθ dθ
2π 0
equal zero for all n < 0.
The C ∗ -subalgebra of L∞ (T) defined by
QC(T) := (C(T) + H ∞ ) ∩ (C(T) + H ∞ )
is referred to as the C ∗ -algebra of quasicontinuous functions. By [25],
QC(T) = V M O(T) ∩ L∞ (T).
Let P QC(T) := alg(QC(T), P C(T)) be the C ∗ -subalgebra of L∞ (T) gener-
ated by the C ∗ -algebras QC(T) and P C(T). The functions in P QC(T) are referred
to as the piecewise quasicontinuous functions.
It is known that the maximal ideal space of C(T) and P C(T) can be identified,
respectively, with T and T × {0, 1},
M (C(T)) = T, M (P C(T)) = T × {0, 1},
where the points t ∈ T are identified with the evaluation functionals δt given for
f ∈ C(T) by δt (f ) = f (t), and the pairs (t, 0) and (t, 1) are the multiplicative linear
functionals defined for a ∈ P C(T) by (t, 0)a = a(t−0) and (t, 1)a = a(t+0), where
a(t − 0) and a(t + 0) are the left and right one-sided limits of a at the point t ∈ T.
The base of open sets on T × {0, 1} consists of all sets of the form
(t, τ ) × {0, 1}, ((t, τ ] × {0}) ∪ ((t, τ ) × {1}), ((t, τ ) × {0}) ∪ ([t, τ ) × {1}),
 A C ∗ -algebra of Singular Integral Operators with Shifts 29

where t, τ ∈ T. Since C(T) ⊂ QC(T) ⊂ P QC(T), it follows from [26] that

M (QC(T)) = Mt (QC(T)), M (P QC(T)) = Mξ (P QC(T)), (2.1)
t∈T ξ∈M(QC(T))

where the corresponding fibers are given for t ∈ T and ξ ∈ M (QC(T)) by

 
Mt (QC(T)) = ξ ∈ M (QC(T)) : ξ|C(T) = t ,
 
Mξ (P QC(T)) = y ∈ M (P QC(T)) : y|QC(T) = ξ .
For a function f ∈ L1 (T) and points t = eiθ ∈ T and λ ∈ (1, ∞), put
! θ+ πλ
λ
δ(λ,t) (f ) = f (eix )dx.
2π θ− πλ
Observe that, for each (λ, t) ∈ (1, ∞) × T, the map
δ(λ,t) : QC(T) → C, a → δ(λ,t) (a),
defines a linear functional in QC(T)∗ which allows one to identify (λ, t) ∈ (1, ∞) ×
T with a subset of QC(T)∗ . Let Mt0 (QC(T)) denote the set of functionals in
Mt (QC(T)) that lie in the weak-star closure of (1, ∞) × {t},
Mt0 (QC(T)) := Mt (QC(T)) ∩ closQC(T)∗ ((1, ∞) × {t}) .
For each t ∈ T, we also consider the sets

Mt+ (QC(T)) := ξ ∈ Mt (QC(T)) : ξ(f ) = 0 whenever f ∈ QC and

lim sup |f (z)| = 0 ,
z→t+
−

Mt (QC(T)) := ξ ∈ Mt (QC(T)) : ξ(f ) = 0 whenever f ∈ QC and

lim sup |f (z)| = 0 .
z→t−
According to [26, Lemma 8], we have the following.
Lemma 2.1. For each t ∈ T,
Mt+ (QC(T)) ∩ Mt− (QC(T)) = Mt0 (QC(T)),
Mt+ (QC(T)) ∪ Mt− (QC(T)) = Mt (QC(T)).
By Lemma 2.1, for each t ∈ T, we can split the fiber Mt (QC(T)) into the
three disjoint sets: Mt0 (QC(T)) and
#t+ (QC(T)) := Mt+ (QC(T)) \ M 0 (QC(T)),
M t
(2.2)
#t− (QC(T)) := Mt− (QC(T)) \ Mt0 (QC(T)).
M
Hence, letting
M ± (QC(T)) := Mt± (QC(T)), M 0 (QC(T)) := Mt0 (QC(T)),
t∈T t∈T
(2.3)
#± (QC(T)) :=
M #t± (QC(T)),
M
t∈T
30 M.A. Bastos, C.A. Fernandes and Yu.I. Karlovich

we obtain from (2.1) the following partition of M (QC(T)):

#− (QC(T)) ∪ M 0 (QC(T)) ∪ M
M (QC(T)) = M #+ (QC(T)). (2.4)
There is a natural one-to-one mapping
w : M (P QC(T)) → M (QC(T)) × {0, 1}
into a subset of M (QC(T)) × {0, 1}, which is given as follows: for every y ∈
M (P QC(T)), defining ξ = y|QC(T) , t = y|C(T) and v = y|P C(T) , we conclude that
w(y) = (ξ, 0) if v = (t, 0) and w(y) = (ξ, 1) if v = (t, 1).
We get the following characterization of fibers of M (P QC(T)) from [26, Lem-
mas 8, 13].
Lemma 2.2. Let t ∈ T and ξ ∈ Mt (QC(T)). Then
(i) Mξ (P QC(T)) = {(ξ, 1)} whenever ξ ∈ M #+ (QC(T));
t
(ii) Mξ (P QC(T)) = {(ξ, 0)} whenever ξ ∈ M #t− (QC(T));
(iii) Mξ (P QC(T)) = {(ξ, 0), (ξ, 1)} whenever ξ ∈ Mt0 (QC(T)). In this case, if
{λn } ⊂ (1, ∞) is such that (λn , t) → ξ in the weak-star topology on QC(T)∗ ,
then
! π !
λn θ+ λn λn θ
(ξ, 1)f = lim f (eix )dx, (ξ, 0)f = lim f (eix )dx
n→∞ π θ n→∞ π θ− λπ
n

for every f ∈ P QC(T), where t = e . iθ

The Gelfand topology on M (P QC(T)) can be described as follows. If ξ ∈

Mt (QC(T)) with t ∈ T, then a base of neighborhoods for (ξ, μ) ∈ M (P QC(T))
consists of all open sets of the form
 −

(Uξ,t × {0}) ∪ (Uξ,t × {0, 1}) ∩ M (P QC(T)) if μ = 0,
W(ξ,μ) =   (2.5)
(Uξ,t × {1}) ∪ (Uξ,t
+
× {0, 1}) ∩ M (P QC(T)) if μ = 1,
where Uξ,t = Uξ ∩ Mt (QC(T)), Uξ is an open neighborhood of ξ ∈ M (QC(T)),
−
and Uξ,t +
, Uξ,t consists of all ζ ∈ Uξ such that τ = ζ|C(T) belong, respectively, to
the open arcs (te−iε , t) and (t, teiε ) of T for some ε ∈ (0, 2π).
In order to study the C ∗ -algebra B defined in (1.2), we need to know how
the operators of multiplication aI (a ∈ P QC(T)), the Cauchy singular integral
operator ST and the unitary shift operators Ug (g ∈ G) interact modulo compact
operators in B(L2 (T)).
It follows from the Hartman theorem (see, e.g., [13, Theorem 2.18]) that
aST  ST aI for all a ∈ QC(T). (2.6)
As all shifts g ∈ G are orientation-preserving diffeomorphisms, we get
U g ST  ST U g for all g ∈ G,
which allows us to conclude that
Ug ST Ug∗  ST for all g ∈ G, (2.7)
 A C ∗ -algebra of Singular Integral Operators with Shifts 31

(see, e.g., [22, Theorem 4.1]). An easy computation shows that

Ug aUg∗ = (a ◦ g)I for all a ∈ L∞ (T). (2.8)

Obviously, for every a ∈ L∞ (T) the function a◦ g is also in L∞ (T). It was observed
in [14] that a ◦ g ∈ QC(T) whenever a ∈ QC(T). Consequently, we have the
following fact.

Lemma 2.3. If g is a diffeomorphism of T onto itself, a ∈ QC(T) and b ∈ P QC(T),

then a ◦ g ∈ QC(T) and b ◦ g ∈ P QC(T).

As a consequence of (2.7)–(2.8) and Lemma 2.3, for every shift g ∈ G the

mapping
αg : Aπ → Ugπ Aπ (Ugπ )−1 (2.9)

defines a ∗ -automorphism of the C ∗ -algebra Aπ = A/K, where K = K(L2 (T)) is

the ideal of all compact operators on L2 (T).

3. Main results: Fredholmness in the C ∗ -algebra B

Let G be a group of orientation-preserving diffeomorphisms of T onto itself such
that all g ∈ G \ {e} have the same finite set Λ of fixed points on T. Consider the
C ∗ -algebra
B = alg(P QC(T), ST , UG ) ⊂ B(L2 (T))
generated by all multiplication operators by the piecewise quasicontinuous func-
tions on T, by the Cauchy singular integral operator ST and by all shift operators
Ug (g ∈ G). Let

Narc := Mτ (QC(T)) × R, NΛ := Mτ0 (QC(T)) × R, (3.1)

τ ∈Oarc τ ∈Λ

where Oarc is a subset of T \ Λ that contains exactly one point in each G-orbit
{g(τ ) : g ∈ G} defined for τ ∈ T \ Λ by the group of shifts G. Thus,
G(τ ) = 
T \ Λ = τ ∈Oarc G(τ ). Along with Narc , we define the sets

N±
arc :=
#τ± (QC(T)) × R,
M N0arc := Mτ0 (QC(T)) × R,
τ ∈Oarc τ ∈Oarc

#± (QC(T)), M 0 (QC(T)) are given by (2.3) and Lemma 2.1.

where the sets M τ t
For each (ξ, x) ∈ Narc , we introduce the representation

Φξ,x : B → B(l2 (G, C2 )), B → Φξ,x (B), (3.2)

32 M.A. Bastos, C.A. Fernandes and Yu.I. Karlovich

given on the generators of the C ∗ -algebra B for all g ∈ G by

⎧
⎪
⎨diag{(a ◦ g)(ξ, 0), (a ◦ g)(ξ, 0)}f (g) if (ξ, x) ∈ N−arc ,
[Φξ,x (aI)f ](g) = diag{(a ◦ g)(ξ, 1), (a ◦ g)(ξ, 0)}f (g) if (ξ, x) ∈ N0arc ,
⎪
⎩
diag{(a ◦ g)(ξ, 1), (a ◦ g)(ξ, 1)}f (g) if (ξ, x) ∈ N+arc ,
  (3.3)
tanh(πx) 1/ cosh(πx)
[Φξ,x (ST )f ](g) = f (g),
1/ cosh(πx) − tanh(πx)
[Φξ,x (Uh )f ](g) = f (gh),
where a ∈ P QC(T), a(ξ, μ) is the value of the Gelfand transform of a at the point
(ξ, μ) ∈ M (P QC(T)), h ∈ G, and f ∈ l2 (G, C2 ).
For each (ξ, x) ∈ NΛ , we introduce the representation
Φξ,x : B → B(C2 ), B → Φξ,x (B), (3.4)
given on the generators of B by
 
[Φξ,x (aI)]f = diag a(ξ, 1), a(ξ, 0) f,
 
tanh(πx) 1/ cosh(πx)
[Φξ,x (ST )]f = f, (3.5)
1/ cosh(πx) − tanh(πx)
   
[Φξ,x (Uh )]f = diag eix ln h (ξ) , eix ln h (ξ) f,
where a ∈ P QC(T), a(ξ, μ) is the value of the Gelfand transform of a at the point
(ξ, μ) ∈ M (P QC(T)), h ∈ G, and f ∈ C2 .
We establish below the following Fredholm criterion for the operators B in
the C ∗ -algebra B.
Theorem 3.1. An operator B ∈ B is Fredholm on the space L2 (T) if and only if
the following two conditions are satisfied:
(i) for all (ξ, x) ∈ Narc the operators Φξ,x (B) are invertible on the space l2 (G, C2 )
and
$ $
sup $(Φξ,x (B))−1 $ < ∞;
(ξ,x)∈Narc

(ii) for all (ξ, x) ∈ NΛ the operators Φξ,x (B) are invertible on the space C2 and
$ $
inf $(Φξ,x (B))−1 $ < ∞
(ξ,x)∈NΛ

or, equivalently,
inf | det(Φξ,x (B))| > 0,
(ξ,x)∈NΛ

where the operators Φξ,x (B) are identified with their 2 × 2 matrices.
The Fredholm criterion presented in Theorem 3.1 does not depend on the
concrete choice of Narc because, for every operator B ∈ B, its representations
associated with different points of the same orbit G(τ ) are unitarily equivalent,
and therefore it suffices to use representations related to only one point in each
 A C ∗ -algebra of Singular Integral Operators with Shifts 33

orbit. Moreover, one can easily see that assertion (i) of Theorem 3.1 remains true
with Narc replaced by
NT\Λ := Mτ (QC(T)) × R.
τ ∈T\Λ

Consider the Hilbert space

     
H := l2 (G, C2 ) ⊕ C2 (3.6)
(ξ,x)∈Narc (ξ,x)∈NΛ

and the operator function Φ(B) : (ξ, x) → Φξ,x (B) defined on Narc ∪ NΛ by (3.2)–
(3.5) and equipped with the norm
% $ $ $ $ &
Φ(B)B(H) = max sup $Φξ,x (B)$B(l2 (G,C2 )) , sup $Φξ,x (B)$B(C2 ) .
(ξ,x)∈Narc (ξ,x)∈NΛ

The operator function Φ(B) is referred to as the Fredholm symbol of an operator

B ∈ B. Clearly, the set Φ(B) := {Φ(B) : B ∈ B} is a C ∗ -algebra, and the mapping
Φ : B → B(H), B → Φ(B) (3.7)
∗ ∗ ∗
is a C -algebra homomorphism of the C -algebra B onto the C -algebra Φ(B)
with kernel Ker Φ = K. Hence, for the quotient C ∗ -algebra Bπ = B/K, it follows
that Bπ ∼= Φ(B), that is, the map Bπ → B(H) is a faithful representation in
the Hilbert space (3.6). Making use of this symbol calculus, Theorem 3.1 can be
rewritten in the following form.
Theorem 3.2. An operator B ∈ B is Fredholm on the space L2 (T) if and only if
its symbol Φ(B) is invertible.

4. The local-trajectory method and spectral measures

In order to proceed to the study of the C ∗ -algebra B of singular integral operators
with P QC(T) coefficients and shifts g ∈ G, we recall here the local-trajectory
method and its generalization on the basis of spectral measures (cf. [18], [19], [5]).
Let A be a unital C ∗ -algebra, Z a central C ∗ -subalgebra of A with the same
unit I, G a discrete group with unit e, U : g → Ug a homomorphism of the group
G onto a group UG = {Ug : g ∈ G} of unitary elements such that Ug1 g2 = Ug1 Ug2
and Ue = I. Suppose that A and UG are contained in a C ∗ -algebra D. Let
B := alg(A, UG ) (4.1)
∗ ∗
be the minimal C -subalgebra of D containing the C -algebra A and the group
UG . Assume that
(A1) for every g ∈ G, the mappings αg : a → Ug a Ug∗ are ∗ -automorphisms of the
C ∗ -algebras A and Z;
(A2) G is an amenable discrete group.
According to [17, § 1.2], a discrete group G is called amenable if the C ∗ -
algebra l∞ (G) of all bounded complex-valued functions on G with sup-norm has
34 M.A. Bastos, C.A. Fernandes and Yu.I. Karlovich

an invariant mean, that is, a positive linear functional ρ of norm 1 such that
ρ(f ) = ρ(s f ) = ρ(fs ) for all s ∈ G and all f ∈ l∞ (G),
where (s f )(g) = f (s−1 g), (fs )(g) = f (gs), g ∈ G.
∗
(A1), the C -algebra B = alg(A, UG ) is the closure of the set B
0
By virtue of
of elements b = ag Ug , where ag ∈ A and g runs through finite subsets of G.
Let M (Z) be the maximal ideal space of the commutative C ∗ -algebra Z.
By the Gelfand– Naimark theorem [23, § 16], Z ∼ = C(M (Z)), where C(M (Z))
is the C ∗ -algebra of all continuous complex-valued functions on M (Z). Under
assumption (A1), identifying the characters ϕm of the commutative C ∗ -algebra Z
and the maximal ideals m = Ker ϕm ∈ M (Z), we see that each ∗ -automorphism
αg : Z → Z induces a homeomorphism βg : M (Z) → M (Z) given by the rule
z(βg (m)) = (αg (z))(m), z ∈ Z, m ∈ M (Z), g ∈ G, (4.2)
where z(·) ∈ C(M (Z)) is the Gelfand transform of the element z ∈ Z. The set
G(m) := {βg (m) : g ∈ G} is called the G-orbit of a point m ∈ M (Z).
Let PA denote the set of all pure states of the C ∗ -algebra A with the induced
weak* topology, and let Jm be the closed two-sided ideal of A generated by the
maximal ideal m ∈ M (Z) of the central C ∗ -algebra Z ⊂ A. If μ ∈ PA , then
Ker μ ⊃
 Jm where m := Z ∩ Ker μ ∈ M (Z) (see, e.g., [12, Lemma 4.1]), and hence
PA = m∈M {ν ∈ PA : Ker ν ⊃ Jm }.
Let the following version of topologically free action of the group G hold:
(A3) there is a set M0 ⊂ M (Z) such that for every finite set G0 ⊂ G \ {e} and
every nonempty open set W ⊂ PA there exists a state ν ∈ W such that
βg (mν ) = mν for all g ∈ G0 , where mν = Z ∩ Ker ν belongs to the G-orbit
G(M0 ) := {βg (m) : g ∈ G, m ∈ M0 } of the set M0 .
For every m ∈ M (Z), let π m be an isometric representation
m : A/Jm → B(Hm )
π (4.3)
of the quotient C ∗ -algebra A/Jm in a Hilbert space Hm . Consider the canonical
∗
-homomorphism m : A → A/Jm and the representation

πm : A → B(Hm ), A → (
πm ◦ m )(A). (4.4)
Since αg (Jβg (m) ) = Jm for all g ∈ G and all m ∈ M (Z) in view of (A1), it follows
that the quotient C ∗ -algebras A/Jβg (m) and A/Jm are ∗ -isomorphic, and therefore
the spaces Hβg (m) can be chosen equal for all g ∈ G.
Let Ω(M0 ) be the set of G-orbits of all points m ∈ M0 . For each G-orbit
ω ∈ Ω(M0 ), fix a point m = mω ∈ ω, put Hω = Hm , and let l2 (G, Hω ) be the
Hilbert space of all functions f : G→ Hω such that f (g) = 0 for at most countable
set of points g ∈ G and f  := ( g f (g)2 )1/2 < ∞. For every ω ∈ Ω(M0 ) we
consider the representation πω : B → B(l2 (G, Hω )) defined by

[πω (a)f ](g) = πm ω
(αg (a))f (g), [πω (Uh )f ](g) = f (gh) (4.5)
for all a ∈ A, all g, h ∈ G and all f ∈ l (G, Hω ).
2
 A C ∗ -algebra of Singular Integral Operators with Shifts 35

Under the above three conditions we have the following result ([19, Theo-
rems 4.1, 4.12], [5, Theorem 3.1]).

Theorem 4.1. If assumptions (A1)–(A3) are satisfied, then an element b ∈ B is

invertible in B if and only if for every orbit ω ∈ Ω(M0 ) the operator πω (b) is
invertible on the space l2 (G, Hω ) and, in the case of infinite set Ω(M0 ),
 
sup (πω (b))−1 B(l2 (G,Hω )) : ω ∈ Ω(M0 ) < ∞.

Note that Theorem 4.1 remains true with M0 = M (Z), and then Ω(M0 ) =
Ω(M (Z)) is the set of all G-orbits on M (Z).
We will present now a generalization of the local-trajectory method for the
case when condition (A3) is not fulfilled. Such generalization based on the notion
of spectral measures was developed in [19] and [5].
Let M be a compact Hausdorff space and H a Hilbert space. By [23, p. 249],
a spectral measure P (·) is a map from the σ-algebra of all Borel sets of M into
the set of orthogonal projections in B(H) such that for every ξ ∈ H the function
Δ → (P (Δ)ξ, ξ) is the restriction to Borel sets of a measure on M defined by some
integral on C(M ). Hence, P (∅) = 0, P (M ) = I, P (Δ1 ∩ Δ2 ) = P (Δ1 )P (Δ2 ) for
all Borel sets Δ1 , Δ2 ⊂ M , and P (Δ1 ∪ Δ2 ) = P (Δ1 ) + P (Δ2 ) if Δ1 ∩ Δ2 = ∅.
Suppose the C ∗ -algebra B = alg(A, UG ) satisfies only the conditions (A1)–
(A2) of the local-trajectory method. Let R(M (Z)) denote the σ-algebra of all
Borel subsets of M (Z), and let
 
RG (M (Z)) := Δ ∈ R(M (Z)) : βg (Δ) = Δ for all g ∈ G , (4.6)
where the homeomorphisms βg are given by (4.2). As is well known (see, e.g., [15,
Theorem 2.6.1]), there exists an isometric representation π : B → B(H) of the
C ∗ -algebra B in a Hilbert space H. According to [23, § 17], for the representation
π|Z : Z → B(H) of the unital commutative C ∗ -algebra Z, there is a unique
spectral measure Pπ : R(M (Z)) → B(H) that commutes with all operators in the
C ∗ -algebra π(Z) and in its commutant π(Z) , and such that
!
π(z) = z(m) dPπ (m) for all z ∈ Z, (4.7)
M(Z)

where z(·) ∈ C(M (Z)) is the Gelfand transform of the element z ∈ Z.

The integral in (4.7) is defined as follows. Since
! ! !
z(m) dPπ (m) = Re z(m) dPπ (m) + i Im z(m) dPπ (m),
M(Z) M(Z) M(Z)

it is sufficient to define such integral for real-valued functions f ∈ C(M (Z)).

Any real-valued function f ∈ C(M (Z)) is the uniform limit inL∞ (M (Z)) of
n
a sequence {fn }n∈N of simple Borel real-valued functions fn = k=1 ck,n χΔk,n ,
where Δ1,n , . . . , Δn,n are pairwise disjoint Borel subsets of M (Z) for every n ∈ N,
ck,n are real constants that can be chosen equal to values of f at some points
36 M.A. Bastos, C.A. Fernandes and Yu.I. Karlovich

mk,n ∈ Δk,n , and χΔk,n are the characteristic functions of Δk,n . Then
! n
f (m) dPπ (m) = lim f (mk,n )Pπ (Δk,n ) ∈ B(H),
M(Z) n→∞
k=1

where the convergence is uniform in B(H) because f − fn L∞ (M(Z)) → 0 as

n → ∞. Hence, for every ξ ∈ H,
'' ! ( ( !
f (m) dPπ (m) ξ, ξ = f (m) dμξ (m),
M(Z) M(Z)

where μξ (·) := (Pπ (·)ξ, ξ) is a Borel measure, and

!
μξ (Δ) = (Pπ (Δ)ξ, ξ) = χΔ (m)dμξ (m) for all Δ ∈ R(M (Z)).
M(Z)

Since (A1) holds and since az = za for all a ∈ A and all z ∈ Z, we deduce
from [19, Lemma 4.6] and (4.6)–(4.7) that
π(b)Pπ (Δ) = Pπ (Δ)π(b) for all b ∈ B and all Δ ∈ RG (M (Z)). (4.8)
Given Δ ∈ RG (M (Z)) such that Pπ (Δ) = 0, let us define the Hilbert space
HΔ := Pπ (Δ)H = {Pπ (Δ)ξ : ξ ∈ H} and introduce the following three C ∗ -
subalgebras of B(HΔ ):
BΔ :={Pπ (Δ)π(b) : b ∈ B},
AΔ :={Pπ (Δ)π(a) : a ∈ A}, (4.9)
ZΔ :={Pπ (Δ)π(z) : z ∈ Z}.
Since Z is a central C ∗ -subalgebra of A, it follows from (4.8) that ZΔ is a central
C ∗ -subalgebra of AΔ , where AΔ ⊂ BΔ .
For each Borel set Δ ∈ R(M (Z)), let Int Δ and Δ denote the interior and
 be the closed subset of Δ given by
the closure of Δ, respectively, and let Δ
 := {m ∈ M (Z) : Pπ (Wm ∩ Δ) = 0 for each open neighborhood Wm of m}.
Δ

The next lemma summarizes some important properties of sets Δ.
Lemma 4.2. [19, Lemmas 5.1–5.2] If Δ ∈ R(M (Z)) and Int Δ = 0, then
(i) Pπ (Δ) = 0,
(ii) ZΔ ∼ 
= C(Δ),

(iii) Δ = Int Δ in case Pπ (Δ \ Int Δ) = 0.
Fix Δ ∈ RG (M (Z)). For each g ∈ G, consider the unitary operator Ug,Δ :=
Pπ (Δ)π(Ug ) on HΔ . As condition (A1) holds, the mappings
∗
αg,Δ : Pπ (Δ)π(a) → Ug,Δ Pπ (Δ)π(a)Ug,Δ = Pπ (Δ)π(Ug aUg∗ ) (g ∈ G),
are ∗ -automorphisms of the C ∗ -algebras ZΔ and AΔ defined by (4.9). Since ZΔ ∼=
 where Δ
C(Δ),  ∈ RG (M (Z)) and the isomorphism ZΔ → C(Δ)  is given by
Pπ (Δ)π(z) → z(·)|Δ  of the Gelfand transform of z ∈ Z,
 , with restriction z(·)|Δ
 A C ∗ -algebra of Singular Integral Operators with Shifts 37

we conclude that each ∗ -automorphism αg,Δ induces on Δ  the homeomorphism

βg,Δ := βg |Δ
 , where βg is defined by (4.2).
We will need below the following decomposition result.

Theorem 4.3. [6, Proposition 3.3] Let π : B → B(H) be an isometric representation

of the C ∗ -algebra B = alg(A, UG ) in a Hilbert space H and let {Δi } be an at most
countable family ofdisjoint Borel sets in RG (M (Z)) such that Pπ (Δi ) = 0 for all
i and Pπ (M (Z) \ i Δi ) = 0. If condition (A1) holds, then the mapping
 
Θ: B→ BΔi , b → Pπ (Δi )π(b),
i i

is an isometric C -algebra homomorphism from the C ∗ -algebra B into the C ∗ -

∗

algebra B := i BΔi . Then an element b ∈ B is invertible if and only if for each i
the operator Pπ (Δi )π(b) is invertible on the Hilbert space HΔi and
$ $
sup $(Pπ (Δi )π(b))−1 $ < ∞ in case {Δi } is countable.
i

It should be noted that, for the C ∗ -algebra B = alg(A, UG ), a G-invariant

decomposition considered in Theorem 4.3 does not always exist. On the other
hand, such decompositions always exist for the C ∗ -algebras B = alg(A, UG ) with
groups G containing nonempty sets of common fixed points for all g ∈ G (see
Theorem 6.3 below).

5. The C ∗ -algebra A
Consider the C ∗ -algebra

A = alg(P QC(T), ST ) ⊂ B(L2 (T))

of singular integral operators on L2 (T) with P QC(T) coefficients. With the C ∗ -

algebra A we associate the set

M := M (QC(T)) × R, (5.1)

where R = [−∞, +∞] is the two-point compactification of the real line R =

(−∞, +∞). Let B(M, C2×2 ) stand for the C ∗ -algebra of all bounded matrix func-
tions f : M → C2×2 . Put
#± (QC(T)) × R,
M± := M M0 := M 0 (QC(T)) × R, (5.2)
#± (QC(T)) and M 0 (QC(T)) are defined by (2.3).
where the sets M
According to [14, Section 7.4] and [5, Theorem 5.1], we obtain the following
symbol calculus for the C ∗ -algebra A.
38 M.A. Bastos, C.A. Fernandes and Yu.I. Karlovich

Theorem 5.1. The map Sym : {aI : a ∈ P QC(T)} ∪ {ST } → B(M, C2×2 ) given by
the matrix functions
⎧
⎪ −
⎨diag{a(ξ, 0), a(ξ, 0)} for all (ξ, x) ∈ M ,
(Sym aI)(ξ, x) := diag{a(ξ, 1), a(ξ, 0)} for all (ξ, x) ∈ M0 ,
⎪
⎩
diag{a(ξ, 1), a(ξ, 1)} for all (ξ, x) ∈ M+ , (5.3)
) *
u(x) w(x)
(Sym ST )(ξ, x) := for all (ξ, x) ∈ M,
w(x) −u(x)
where a(ξ, μ) is the Gelfand transform of a function a ∈ P QC(T) at the point
(ξ, μ) ∈ M (P QC(T)), the sets M± and M0 are given by (5.2), and
u(x) := tanh(πx), w(x) := 1/ cosh(πx) for all x ∈ R, (5.4)
extends to a C ∗ -algebra homomorphism
Sym : A → B(M, C2×2 ) (5.5)
whose kernel consists of all compact operators on L2 (T). An operator A ∈ A is
Fredholm on the space L2 (T) if and only if
det((Sym A)(ξ, x)) = 0 for all (ξ, x) ∈ M.

Proof. Let M := t∈T Mt , where, with M#t± (QC(T)) given by (2.2),

#− (QC(T)) ∪ M 0 (QC(T)) × [0, 1] ∪ M

Mt := M #+ (QC(T)), t ∈ T.
t t t

For the C ∗ -algebra A, we know from [14, Section 7] that the mapping
Sym : {aI : a ∈ P QC(T)} ∪ {ST } → B(M, C2×2 ),
defined by the matrix functions
(Sym ST )(ζ) = diag{1, −1} for all ζ ∈ M,
⎧
⎪
⎪ diag{a(ξ, 0), a(ξ, 0)} for all ζ = ξ ∈ M #− (QC(T)),
⎪
⎪ ) *
⎪
⎪
⎪
⎨ a(ξ, 1)μ + a(ξ, 0)(1 − μ) [a(ξ, 1) − a(ξ, 0)](μ)
(Sym aI)(ζ) = [a(ξ, 1) − a(ξ, 0)](μ) a(ξ, 1)(1 − μ) + a(ξ, 0)μ
⎪
⎪
⎪
⎪ for all ζ = (ξ, μ) ∈ M 0 (QC(T)) × [0, 1],
⎪
⎪
⎪
⎩
diag{a(ξ, 1), a(ξ, 1)} for all ζ = ξ ∈ M #+ (QC(T)),
+
where (μ) = μ(1 − μ) for all μ ∈ [0, 1] and the sets M #± (QC(T)) and M 0 (QC(T))
in the partition (2.4) of M (QC(T)) are defined by (2.3), extends to a C ∗ -algebra
homomorphism Sym : A → B(M, C2×2 ) whose kernel consists of all compact
operators in the C ∗ -algebra B(L2 (T)), and an operator A ∈ A is Fredholm on
space L2 (T) if and only if
det((Sym A)(ζ)) = 0 for all ζ ∈ M.
 A C ∗ -algebra of Singular Integral Operators with Shifts 39

Let M := M (QC(T)) × [0, 1] and consider the mapping

Sym[0,1] : A → B(M, C2×2 )
defined for all operator A ∈ A by

(Sym A)(ξ, μ) if (ξ, μ) ∈ M 0 (QC(T)) × [0, 1],
(Sym[0,1] A)(ξ, μ) :=
(Sym A)(ξ) if (ξ, μ) ∈ M \ (M 0 (QC(T)) × [0, 1]).

Clearly, Sym[0,1] is a C ∗ -algebra homomorphism with the same kernel as Sym, and
such that an operator A ∈ A is Fredholm on space L2 (T) if and only if
det((Sym[0,1] A)(ξ, μ)) = 0 for all (ξ, μ) ∈ M.
Since μ(x) = (1 + u(x))/2 ∈ [0, 1] for all x ∈ R, it remains to see that
(Sym A)(ξ, x) =
) + + * ) + + *−1
μ(x) 1 − μ(x) μ(x) 1 − μ(x)
+ + (Sym[0,1] A)(ξ, μ(x)) + +
1 − μ(x) − μ(x) 1 − μ(x) − μ(x)
for all (ξ, x) ∈ M. 
By Theorem 5.1,
|A| = (Sym A)IB(l2 (M,C2 )) for all A ∈ A.
For each point t ∈ T, we consider the operator Vt ∈ B(L2 (T)), with fixed
singularity at t, given by
! !
χ+t (z) ϕ(y)χ+t (y) χ−
t (z) ϕ(y)χ−t (y)
(Vt ϕ)(z) := dy − dy for z ∈ T, (5.6)
πi T y + z − 2t πi T y + z − 2t
where χ± ± −
t are the characteristic functions of arcs γt such that γt := γt ∪ γt is a
+
−
neighborhood of t, γt ∩ γt = {t}, γt is separated from −t, and γt ∩ (−t, t) = ∅,
+ +

γt− ∩ (t, −t) = ∅. Let

v(x) := −iw(x) = −i/ cosh(πx) for all x ∈ R. (5.7)
Lemma 5.2. For each t ∈ T, the operator Vt given by (5.6) belongs to the C ∗ -algebra
A and its symbol is given by

diag{v(x), v(x)} if (ξ, x) ∈ Mt0 (QC(T)) × R,
(SymVt )(ξ, x) := (5.8)
02×2 if (ξ, x) ∈ M \ (Mt0 (QC(T)) × R),
where the function v and the set M are defined by (5.7) and (5.1), respectively.
Proof. For every t ∈ T, the operator Vt belongs to the C ∗ -algebra
S := alg(P C(T), ST ) ⊂ A
n k
(cf. [5, Lemma 5.3]). Fix a sequence of polynomials Pn (u) = k=0 ak,n √ u with
coefficients ak,n ∈ C that uniformly converges on [−1, 1] to the function u − 1 ∈
2
40 M.A. Bastos, C.A. Fernandes and Yu.I. Karlovich

C([−1, 1]). Hence, because ST2 = I and therefore the spectrum of the operator ST
consists of the points ±1, we conclude that
,
lim Pn (ST ) = ST2 − I = 0. (5.9)
n→∞

By the proof of [5, Lemma 5.3],

− − π
t ST χt I − χt ST χt I)
(Vt )π = lim χt,n Pn (χ+ +
, (5.10)
n→∞

where χt,n ∈ P C(T) are characteristic

- functions of open neighborhoods of the
point t, supp χt,n+1 ⊂ supp χt,n and n∈N supp χt,n = {t}. Since
− −
t ST χt I − χt ST χt I) (ξ, x) = diag{u(x), u(x)}
Sym(χ+ +
(5.11)
for all (ξ, x) ∈ Mt0 (SO(T)) × R, and since
+ the sequence {Pn (u(x))} converges
uniformly on R to the function v(x) = u2 (x) − 1, where the functions u and
v are given by (5.4) and (5.7), respectively, we conclude from (5.10) and (5.11)
similarly to [5, Lemma 5.3] that
(SymVt )(ξ, x) = diag{v(x), v(x)} for all (ξ, x) ∈ Mt0 (QC(T)) × R. (5.12)
#t± (QC(T)) × R, then, respectively,
On the other hand, if (ξ, x) ∈ M
− −
t ST χt I − χt ST χt I) (ξ, x) = Sym(±ST ) (ξ, x).
Sym(χ+ +

Hence, taking into account (5.10), (5.3) and (5.9), we infer that
(SymVt )(ξ, x) = lim Sym(χt,n Pn (±ST )) (ξ, x) = 0 (5.13)
n→∞
-
#t± (QC(T)) × R. Finally, since
for all (ξ, x) ∈ M supp χt,n = {t}, we obtain
n∈N
(5.8) from (5.12), (5.13) and (5.3). 

We now study the product Ug Vt of the shift operator Ug given by (1.3) and
the operator Vt with a fixed singularity at a fixed point t ∈ T of g.
Lemma 5.3. Let g be an orientation-preserving diffeomorphism of T onto itself and
let t ∈ T be a fixed point of g. Then Ug Vt ∈ A, Ug Vt  Vt Ug and
(Sym(Ug Vt ))(ξ, x)
  
diag{eix ln g (t) v(x), eix ln g (t) v(x)} if (ξ, x) ∈ Mt0 (QC(T)) × R,
:=
02×2 if (ξ, x) ∈ M \ (Mt0 (QC(T)) × R).
(5.14)
Proof. Let t ∈ T be a fixed point of a g ∈ G. By the proof of [5, Lemma 5.4],
Ug Vt ∈ A, Ug Vt  Vt Ug and
− − π
(Ug Vt )π = lim χt,n Pn (χ+t ST χt I − χt ST χt I)
+
, (5.15)
n→∞
n
where the sequence {Pn (u(x))} of functions Pn (u(x)) = k=0 ak,n uk (x) with co-
efficients ak,n ∈ C converges uniformly on R to the continuous function x →
 A C ∗ -algebra of Singular Integral Operators with Shifts 41


eix ln g (t) v(x), u(x) = tanh(πx) and v(x) = −i/ cosh(πx). Repeating now the ar-
guments used in the proof of Lemma 5.2, we establish (5.14) on the basis of (5.15)
and Theorem 5.1. 

We know from [5, (4.10)] that

aVt  Vt aI, ST Vt  Vt ST for all a ∈ P C(T) and all t ∈ T. (5.16)
Since Vt ∈ S = alg(P C(T), ST ), we infer from (2.6) that
aVt  Vt aI for all a ∈ QC(T) and all t ∈ T. (5.17)
Hence, relations (5.16)–(5.17) imply
aVt  Vt aI for all a ∈ P QC(T) and all t ∈ T. (5.18)
We now introduce the C ∗ -algebra
 
Z := alg aI, HP,t : a ∈ QC(T), P ∈ P, t ∈ T ⊂ A ⊂ B(L2 (T)) (5.19)
generated by all multiplication operators aI (a ∈ QC(T)) and by all operators
− −
t ST χt I − χt ST χt I)Vt ∈ S ⊂ A (P ∈ P, t ∈ T),
HP,t := P (χ+ +
(5.20)
where
 
n 
P := Pn (u) = ak uk : ak ∈ C, n ∈ N . (5.21)
k=0
By (5.16) and (5.18), we also obtain the relations
aHP,t  HP,t aI, ST HP,t  HP,t ST for all a ∈ P QC(T), P ∈ P, t ∈ T. (5.22)
Moreover, due to (2.6) and (5.22), Z π := (Z + K)/K is a commutative central
C ∗ -subalgebra of the C ∗ -algebra Aπ := A/K.
Applying Lemma 5.2, we infer the following result by analogy with [5, Lemma 6.1].
Lemma 5.4. For each t ∈ T and each polynomial P ∈ P, the symbol of the operator
HP,t ∈ A ⊂ B(L2 (T)) is given by
(Sym HP,t )(ξ, x)

diag {P (u(x))v(x), P (u(x))v(x)} if (ξ, x) ∈ Mt0 (QC(T)) × R,
=
02×2 if (ξ, x) ∈ M \ (Mt0 (QC(T)) × R),
where the functions u, v and the set M are defined by (5.4), (5.7) and (5.1),
respectively.
Let A0 be the non-closed subalgebra of A generated by the operators aI
(a ∈ P QC 0 (T)) and ST , where 0
n P QC (T) is the dense subalgebra of P0 QC(T)
consisting of the functions k=1 ak pk , where ak ∈ QC(T) and pk ∈ P C (T) for
k = 1, 2, . . . , n, and P C 0 (T) consists of all functions in P C(T) with finite sets of
discontinuities.
42 M.A. Bastos, C.A. Fernandes and Yu.I. Karlovich

Theorem 5.5. Every operator A ∈ A is uniquely represented in the form

A = a+ PT+ + a− PT− + HA , (5.23)

where a± ∈ P QC(T), PT± = (I ± ST )/2 and (Sym HA )(ξ, x) = 02×2 for all
(ξ, x) ∈ M \ (M 0 (QC(T)) × R). Moreover, the mappings A → a± are C ∗ -algebra
homomorphisms of the C ∗ -algebra A onto the C ∗ -algebra P QC(T), and
a± L∞ (T) ≤ |A| := inf A + K. (5.24)
K∈K

Proof. Obviously, every operator A ∈ A0 is of the form (5.23), where a± ∈

P QC 0 (T) and (Sym HA )(ξ, ±∞) = 02×2 for all ξ ∈ M (QC(T)).
#± (QC(T))×R. Therefore,
Moreover, (Sym HA )(ξ, x) = 02×2 for all (ξ, x) ∈ M
by (5.3) and (5.4), (Sym A)(ξ, ±∞) is a diagonal matrix for every ξ ∈ M (QC(T))
and its diagonal entries (Sym A)11 (ξ, ±∞) and (Sym A)22 (ξ, ±∞) satisfy the rela-
tions 
a+ (ξ, 1) if ξ ∈ M + (QC(T)),
(Sym A)11 (ξ, +∞) =
#− (QC(T)),
a+ (ξ, 0) if ξ ∈ M

#+ (QC(T)),
a+ (ξ, 1) if ξ ∈ M
(Sym A)22 (ξ, −∞) =
a+ (ξ, 0) if ξ ∈ M − (QC(T)),
 (5.25)
a− (ξ, 1) if ξ ∈ M + (QC(T)),
(Sym A)11 (ξ, −∞) =
#− (QC(T)),
a− (ξ, 0) if ξ ∈ M

#+ (QC(T)),
a− (ξ, 1) if ξ ∈ M
(Sym A)22 (ξ, +∞) =
a− (ξ, 0) if ξ ∈ M − (QC(T)).
Since Aπ ∼= Sym A according to Theorem 5.1, we deduce from (5.25) in view of
topology (2.5) that
% &
a± L∞ (T) = max max |a ± (ξ, 1)|, max |a ± (ξ, 0)|
ξ∈M + (QC(T)) ξ∈M − (QC(T))
 
≤ max max |(Sym A)11 (ξ, ±∞)|, |(Sym A)22 (ξ, ∓∞)| ≤ |A|,
ξ∈M(QC(T))

respectively. Consequently, the mappings A → a± extends by continuity to C ∗ -

algebra homomorphisms of the C ∗ -algebra A onto P QC(T) that preserve (5.24).
This implies the unique representation (5.23) for every A ∈ A. 

Let Ṙ := R ∪ {∞}. Using reasonings similar to those in the proof of [5,

Theorem 6.3], we easily infer from (5.19) that

Zπ ∼
= C(Ṁ), Ṁ := M (QC(T)) × Ṙ, (5.26)

where Ṁ is the compact Hausdorff space equipped with the Gelfand topology
whose neighborhood base of a point (ξ, x) ∈ Ṁ consists of all open sets of the
 A C ∗ -algebra of Singular Integral Operators with Shifts 43

form
⎧
⎪
⎪Uξ,t × (x − ε, x + ε)
0
if (ξ, x) ∈ M 0 (QC(T)) × R,
⎨
Wξ,x = (Uξ × Ṙ) \ (Uξ,t
0
× [−ε, ε]) if (ξ, x) ∈ M 0 (QC(T)) × {∞}, (5.27)
⎪
⎪
⎩ ± ± #± (QC(T)) × Ṙ,
Uξ,t ∪ Uξ,t × Ṙ if (ξ, x) ∈ M

where ε > 0, Uξ is an open neighborhood of a point ξ ∈ M (QC(T)) and, for

t = ξ|C(T) ∈ T, Uξ,t 0
:= Uξ ∩ Mt0 (QC(T)), U  ± := Uξ ∩ M
#t± (QC(T)) and U − , U +
ξ,t ξ,t ξ,t
consists of all ζ ∈ Uξ such that τ = ζ|C(T) belong, respectively, to the open arcs
(te−iε , t) and (t, teiε ) of T for some ε ∈ (0, 2π).
Theorem 5.1 and Lemma 5.4 imply the following.

Theorem 5.6. The maximal ideal space M (Z π ) of the C ∗ -algebra Z π is homeomor-

phic to the compact Hausdorff set Ṁ given by (5.26), and the Gelfand transform
of Z π is defined by
Γ : Z π → C(Ṁ), Z π → z(ξ, x),
where

(Sym Z)11 (ξ, x) if (ξ, x) ∈ M (QC(T)) × R,
z(ξ, x) =
(Sym Z)11 (ξ, ±x) if (ξ, x) ∈ M (QC(T)) × {∞},

and (Sym Z)(ξ, x) given by (5.3)–(5.5) for each operator Z ∈ Z has the form
(Sym Z)(ξ, x) = diag{(Sym Z)11 (ξ, x), (Sym Z)22 (ξ, x)} for all (ξ, x) ∈ M
and possesses the properties
(Sym Z)11 (ξ, x) = (Sym Z)22 (ξ, x), (Sym Z)(ξ, +∞) = (Sym Z)(ξ, −∞).

It follows from [21, (5.24)] that

Ug Vt  Vg−1 (t) Ug for t ∈ T, g ∈ G. (5.28)
Taking into account (5.20), (2.7)–(2.8) and Lemma 2.3, we infer that
Ug HP,t  HP,g−1 (t) Ug for all t ∈ T, g ∈ G, P ∈ P.

Hence, for every g ∈ G, the mapping αg : Aπ → Ugπ Aπ (Ugπ )−1 introduced in (2.9)
is also a ∗ -automorphism of the central C ∗ -subalgebra Z π of Aπ that induces on
the set M (Z π ) = Ṁ given by (5.26) the homeomorphism

βg : Ṁ → Ṁ, (ξ, x) → (g(ξ), x), (5.29)

where ξ → g(ξ) is the homeomorphism on M (QC(T)) given by
a(g(ξ)) = (a ◦ g)(ξ) for all a ∈ QC(T) and all ξ ∈ M (QC(T)) (5.30)
(as usual a(ξ) := ξ(a)).
44 M.A. Bastos, C.A. Fernandes and Yu.I. Karlovich

6. Representations of the C ∗ -algebra Bπ

Consider the C ∗ -algebra B given by (1.2) and fix an isometric representation
ϕ : Bπ → B(Hϕ ) (6.1)
∗
of the C -algebra B := B/K in an abstract Hilbert space Hϕ .
π

For each point t ∈ T, we consider the following subsets of Ṁ and M (see

(5.26) and (5.1)):
M◦t := Mt0 (QC(T)) × R, M◦t := Mt0 (QC(T)) × R,
Ṁ∞
t := Mt (QC(T)) × {∞}, M∞
t := Mt (QC(T)) × {±∞}, (6.2)
Ṁt := Mt (QC(T)) × Ṙ, Mt := Mt (QC(T)) × R.
Let Λ ⊂ T be the finite set of all fixed points for all shifts g ∈ G \ {e}. We also
introduce the sets
Ṁ∞
Λ := Ṁ∞ ∞
t , MΛ := M∞
t , MΛ := Mt , M◦Λ := M◦t ,
t∈Λ t∈Λ t∈Λ t∈Λ
(6.3)
Ṁarc := Ṁt , Marc := Mt , MΛ := (Ṁt \ M◦t ),
t∈T\Λ t∈T\Λ t∈Λ

where the sets Ṁarc and M◦Λ

are open in Ṁ, while the sets Ṁ∞
Λ and MΛ are closed
in Ṁ. As a result, we get the following partition of Ṁ:
Ṁ = Ṁarc ∪ M◦Λ ∪ MΛ . (6.4)
The following simple assertion contains properties of sets (6.2)–(6.3).
Lemma 6.1. For each g ∈ G and each t ∈ T, the homeomorphism βg : Ṁ → Ṁ
defined by (5.29)–(5.30) sends the sets Ṁt and M◦t onto the sets Ṁg(t) and M◦g(t) ,
respectively, whence the sets Ṁarc , M◦Λ , MΛ and Ṁ∞
Λ belong to the set
 
RG (Ṁ) := Δ ∈ R(Ṁ) : βg (Δ) = Δ for all g ∈ G , (6.5)
where R(Ṁ) denotes the σ-algebra of all Borel subsets of Ṁ = M (Z π ).
Let Hφ denote the Hilbert space
Hφ := l2 (Marc , C2 ) ⊕ l2 (MΛ0 (QC(T)), L22 (R)), (6.6)
where
MΛ0 (QC(T)) := Mt0 (QC(T)), (6.7)
t∈Λ
Consider the C ∗ -subalgebra φ(Aπ ) of B(Hφ ) consisting of the operators
     
π
φ(A ) := (Sym A)(ξ, x)I ⊕ (Sym A)(ξ, ·)I , (6.8)
(ξ,x)∈Marc 0 (QC(T))
ξ∈MΛ

where A ∈ A and (Sym A)(ξ, ·) for every ξ ∈ MΛ0 (QC(T)) is the matrix function
x → (Sym A)(ξ, x) defined for all x ∈ R.
 A C ∗ -algebra of Singular Integral Operators with Shifts 45

Consider the C ∗ -algebras (Sym A)|Marc and (Sym A)|Marc consisting, respec-
tively, of the restrictions (Sym A)|Marc and (Sym A)|Marc of Sym A for all operators
A ∈ A to the sets Marc and
Marc := M \ M◦Λ = Marc ∪ (MΛ \ M◦Λ ), (6.9)
where the sets Marc , MΛ and M◦Λ are given by (6.2)–(6.3).
Lemma 6.2. (Sym A)|Marc ∼
= (Sym A)|Marc .
Proof. By (5.3) and (5.4), (Sym A)(ξ, ±∞) is a diagonal matrix for every ξ ∈
M (QC(T)) and every A ∈ A, and its entries
(Sym A)11 (ξ, ±∞) and (Sym A)22 (ξ, ±∞)
for all ξ ∈ Mt (QC(T)) with t ∈ Λ are values of two P QC(T) functions due to
(5.25). Since the action of the group G on the maximal ideal space M (QC(T))
is topologically free according to [11, Lemma 4.2], the mentioned entries for ξ ∈
Mt (QC(T)) with t ∈ Λ can be approximated, in view of (5.25) and the Gelfand
topology (2.5) on M (P QC(T)), by the corresponding entries of the matrices
(Sym A)(ζ, ±∞),

where ζ ∈ τ ∈T\Λ Mτ (QC(T)) and, for (Sym A)11 (ξ, ±∞), τ belong to the right
semi-neighborhood of t if ξ ∈ M + (QC(T)) and τ are in the left semi-neighborhood
#− (QC(T)), while for (Sym A)22 (ξ, ±∞), τ belong to the right semi-
of t if ξ ∈ M
#+ (QC(T)) and τ belong to the left semi-neighborhood
neighborhood of t if ξ ∈ M
−
of t if ξ ∈ M (QC(T)). Hence, for every A ∈ A,
sup (Sym A)(ξ, x)sp = sup (Sym A)(ξ, x)sp , (6.10)
(ξ,x)∈Marc (ξ,x)∈Marc ∪M∞
Λ

where  · sp is the spectral norm, and the set M∞

Λ is given by (6.2)–(6.3).
On the other hand, the singular values of the matrices (Sym A)(ξ, x) for all
(ξ, x) ∈ M− ∪ M+ coincide with the singular values
|(Sym A)11 (ξ, ±∞)| and |(Sym A)22 (ξ, ±∞)|
of the matrices (Sym A)(ξ, ±∞), respectively, where
(Sym A)11 (ξ, ±∞) = (Sym A)22 (ξ, ∓∞)
and the sets M± are given by (5.2). Hence,
sup (Sym A)(ξ, x)sp = sup (Sym A)(ξ, x)sp .
(ξ,x)∈M∞
Λ (ξ,x)∈MΛ \M◦
Λ

This together with (6.9) and (6.10) implies the equality

sup (Sym A)(ξ, x)sp = sup (Sym A)(ξ, x)sp ,
(ξ,x)∈Marc (ξ,x)∈Marc

which completes the proof. 

46 M.A. Bastos, C.A. Fernandes and Yu.I. Karlovich

Since Sym A ∼= (Sym A)|Marc ∪M◦Λ according to Lemma 6.2, we infer from
(6.6)–(6.8) and Theorem 5.1 that the homomorphism
φ : Aπ → B(Hφ ), Aπ → φ(Aπ ), (6.11)
is an isometric representation of A in the Hilbert space Hφ . Let
π

Pϕ : R(Ṁ) → B(Hϕ ), Pφ : R(Ṁ) → B(Hφ ) (6.12)

be the unique spectral measures associated to the representations (6.1) and (6.11)
in the Hilbert spaces Hϕ and Hφ , respectively, which are restricted to the unital
commutative C ∗ -algebra Z π .
For the representation (6.11) given by (6.6)–(6.8) we get the equalities
Pφ (Ṁarc ) = Iarc ⊕ OΛ , Pφ (ṀΛ ) = Oarc ⊕ IΛ , (6.13)
where Oarc and Iarc are, respectively, the zero and the identity operators on the
Hilbert space l2 (Marc , C2 ), while OΛ and IΛ are the zero and identity operators
on Hilbert space l2 (MΛ0 (QC(T)), L22 (R)).
With partition (6.4) we associate the following C ∗ -subalgebras of ϕ(Bπ ):
Barc := alg{Pϕ (Ṁarc )ϕ(Aπ ), Pϕ (Ṁarc )ϕ(Ugπ ) : A ∈ A, g ∈ G}; (6.14)
B◦Λ := alg{Pϕ (M◦Λ )ϕ(Aπ ), Pϕ (M◦Λ )ϕ(Ugπ ) : A ∈ A, g ∈ G}; (6.15)
BΛ := alg{Pϕ (MΛ )ϕ(Aπ ), Pϕ (MΛ )ϕ(Ugπ ) : A ∈ A, g ∈ G}. (6.16)
Since the sets Ṁarc , M◦Λ and MΛ belong to RG (Ṁ), we infer from Theo-
rem 4.3 and partition (6.4) the following abstract Fredholm criterion.
Theorem 6.3. An operator B ∈ B is Fredholm on the Lebesgue space L2 (T) if and
only if
(i) the operator Barc := Pϕ (Ṁarc )ϕ(B π ) is invertible on the Hilbert space
Pϕ (Ṁarc )Hϕ ;
(ii) the operator BΛ◦ := Pϕ (M◦Λ )ϕ(B π ) is invertible on the Hilbert space
Pϕ (M◦Λ )Hϕ ;
(iii) the operator BΛ := Pϕ (MΛ )ϕ(B π ) is invertible on the Hilbert space
Pϕ (MΛ )Hϕ .

7. An invertibility symbol calculus for the C ∗ -algebra Aarc

Let us study the invertibility in the C ∗ -subalgebra
Aarc := Pϕ (Ṁarc )ϕ(Aπ ) ⊂ B(Pϕ (Ṁarc )Hϕ ) (7.1)
consisting of the operators Pϕ (Ṁarc )ϕ(Aπ ) for all A ∈ A. The C ∗ -algebra Zarc :=
Pϕ (Ṁarc )ϕ(Z π ), which consists of the operators Pϕ (Ṁarc )ϕ(Z π ) with Z ∈ Z, is
a central C ∗ -subalgebra of Aarc . Since the set Ṁarc given in (6.3) is open in the
 A C ∗ -algebra of Singular Integral Operators with Shifts 47

Gelfand topology of Ṁ (see (5.27)), we infer from Lemma 4.2 that Zarc ∼ #̇
= C(Marc ),
#̇
where Marc is the closure of Ṁarc in Ṁ,
#̇
Marc := Ṁ \ M◦Λ = Ṁarc ∪ MΛ . (7.2)
We now consider the Hilbert space Hφ given by (6.6) and its subspace
Pφ (Ṁarc )Hφ , which is isometrically isomorphic to the Hilbert space l2 (Marc , C2 )
by (6.13). Along with the C ∗ -algebra Aarc , we consider the C ∗ -algebra
 arc := Pφ (Ṁarc )φ(Aπ ) ⊂ B(l2 (Marc , C2 )),
A
which consists of the operators Pφ (Ṁarc )φ(Aπ ) with A ∈ A. Comparing the images
of spectral measures (6.12), we obtain the following.
Theorem 7.1. The mapping given by
Pϕ (Ṁarc )ϕ(Aπ ) → Pφ (Ṁarc )φ(Aπ ) for all A ∈ A (7.3)
 arc .
is a C ∗ -algebra isomorphism of the C ∗ -algebra Aarc onto the C ∗ -algebra A
Proof. According to [5, Lemma 3.5], for the open Borel set Ṁarc ⊂ Ṁ and each
operator A ∈ A, we have the equalities
Pϕ (Ṁarc )ϕ(Aπ )B(Hϕ ) = sup ϕ(Z π Aπ )B(Hϕ ) , (7.4)
Z∈Z(Ṁarc )

Pφ (Ṁarc )φ(Aπ )B(Hφ ) = sup φ(Z π Aπ )B(Hφ ) , (7.5)

Z∈Z(Ṁarc )

where the set Z(Ṁarc ) consists of the operators Z ∈ Z for which the Gelfand
transform of the coset Z π is a function z(·, ·) ∈ C(Ṁ) with values in [0, 1] and
with support contained in M#̇ . Since ϕ and φ are isometric representations of
arc
the C ∗ -algebra Aπ , we conclude that the right parts of (7.4) and (7.5) are equal,
and therefore
Pϕ (Ṁarc )ϕ(Aπ )B(Hϕ ) = Pφ (Ṁarc )φ(Aπ )B(Hφ ) for all A ∈ A (7.6)
or, equivalently, for all A ∈ A,
Pϕ (Ṁarc )ϕ(Aπ )B(Pϕ (Ṁarc )Hϕ ) = Pφ (Ṁarc )φ(Aπ )B(Pφ (Ṁarc )Hφ )

because Ṁarc ∈ R(Ṁ). This implies that the mapping (7.3) is a well-defined
 arc .
isometric ∗ -isomorphism of the C ∗ -algebra Aarc onto the C ∗ -algebra A 
Combining Theorem 7.1 and Lemma 6.2, we obtain an invertibility symbol
calculus for the abstract C ∗ -algebra Aarc given by (7.1).
Theorem 7.2. The mapping
Symarc : Aarc → B(Marc , C2×2 ), Pϕ (Ṁarc )ϕ(Aπ ) → (Sym A)|Marc , (7.7)
∗
is an isometric C -algebra homomorphism.
48 M.A. Bastos, C.A. Fernandes and Yu.I. Karlovich

For every A ∈ A the operator Pϕ (Ṁarc )ϕ(Aπ ) is invertible on the space

Pϕ (Ṁarc )Hϕ if and only if
det((Sym A)(ξ, x)) = 0 for all (ξ, x) ∈ Marc .
 arc is isometrically ∗ -isomorphic to the
Proof. By (6.8) and (6.13), the C ∗ -algebra A
∗
C -algebra of all matrix functions (Sym A)|Marc , which, in its turn, is isometrically
∗
-isomorphic to the C ∗ -algebra (Sym A)|Marc according to Lemma 6.2. Hence, the
mapping
 arc → B(Marc , C2×2 ), Pφ (Ṁarc )φ(Aπ ) → (Sym A)|
A (7.8)
Marc

is an isometric C ∗ -algebra homomorphism. Since Aarc ∼

=A  arc in virtue of Theo-
rem 7.1, we conclude from (7.8) that the map (7.7) taken in the form
Pϕ (Ṁarc )ϕ(Aπ ) → Pφ (Ṁarc )φ(Aπ ) → (Sym A)|Marc

is an isometric C ∗ -algebra homomorphism of Aarc into B(Marc , C2×2 ), and its

image coincides with (Sym A)|Marc . Then the invertibility criterion for elements in
Aarc follows immediately. 

8. The C ∗ -algebra Barc

The local-trajectory method will allow us to establish in this section an invertibility
criterion for the C ∗ -algebra Barc given by (6.14). Observe that C ∗ -algebra Barc
can be viewed similarly to (4.1) as
Barc := alg(Aarc , (UG )arc ) ⊂ B(Pϕ (Ṁarc )Hϕ ),
the C ∗ -algebra generated by all operators Pϕ (Ṁarc )ϕ(Aπ ) ∈ Aarc and by all uni-
tary operators Ug, arc = Pϕ (Ṁarc )ϕ(Ugπ ), with g ∈ G, in the commutative (and
therefore amenable) group (UG )arc .
For each g ∈ G, the mapping
∗
αg, arc : Pϕ (Ṁarc )ϕ(Aπ ) → Ug, arc (Pϕ (Ṁarc )ϕ(Aπ ))Ug, arc

is a ∗ -automorphism of the C ∗ -algebra Aarc and its central C ∗ -subalgebra Zarc

because
∗ π ∗
Ug, arc (Pϕ (Ṁarc )ϕ(Aπ ))Ug, arc = Pϕ (Ṁarc )ϕ(Ug A (Ug ) )
π π

and mapping (2.9) is a ∗ -automorphism of the C ∗ -algebras Z π and Aπ .

Thus, conditions (A1)–(A2) of the local-trajectory method are satisfied for
the C ∗ -algebra Barc . For every g ∈ G, the ∗ -automorphism αg, arc induces on the
maximal ideal space M (Z ) = M #̇ of Z the homeomorphism
arc arc arc

#̇ #̇
βg, arc : Marc → Marc , (ξ, x) → βg (ξ, x), (8.1)
 A C ∗ -algebra of Singular Integral Operators with Shifts 49

where βg and M #̇
arc are given by (5.29) and (7.2), respectively. It follows from
(5.29)–(5.30) that for every g ∈ G \ {e} the set of fixed points of βg, arc is contained

in the set ṀΛ := t∈Λ Ṁt .
Let us verify the fulfillment of condition (A3) of the local-trajectory method
for the C ∗ -algebra B . For each (ξ, x) ∈ M #̇ , let J be the smallest closed
arc arc (ξ,x)
two-sided ideal of Aarc that contains the set
 
Pϕ (Ṁarc )ϕ(Z π ) : Z ∈ Z, (Sym Z)(ξ, x) = 02×2 .
The set Parc of all pure states of the C ∗ -algebra Aarc can be written as
Parc = P(ξ,x) , P(ξ,x) := {ρ ∈ Parc : Ker ρ ⊃ J(ξ,x) }.

(ξ,x)∈Ṁarc

According to [15, Theorem 2.11.8(i)] (see also [19]), the set P(ξ,x) can be identified
∗
 -algebra◦ Aarc /J(ξ,x) .
with the set of all pure states of the quotient C
By Theorem 7.2, for every (ξ,x) ∈ Marc := t∈T\Λ Mt , the quotient C ∗ -algebra
◦

Aarc /J(ξ,x) is isometrically ∗ -isomorphic to the C ∗ -subalgebra (Sym A)(ξ, x) of

C2×2 , and this isomorphism is given by
(ξ,x) : Pϕ (Ṁarc )ϕ(Aπ ) + J(ξ,x) → (Sym A)(ξ, x)
π for all A ∈ A. (8.2)
#̇
For each (ξ, x) ∈ Marc \ M◦arc , we infer from Theorems 5.1, 7.2 and the proof of
Lemma 6.2 that the C ∗ -algebra Aarc /J(ξ,x) is isometrically ∗ -isomorphic to the C ∗ -
 
subalgebra diag (Sym A)(ξ, +∞), (Sym A)(ξ, −∞) of C4×4 if (ξ, x) ∈ M0,∞ :=
M 0 (QC(T)) × {∞}, and to the C ∗ -algebra (Sym A)(ξ, +∞) ⊂ C2×2 if (ξ, x) ∈
Ṁ± := M #± (QC(T)) × Ṙ, where the isomorphism is defined for all A ∈ A by
⎧  
⎪
⎨diag (Sym A)(ξ, +∞), (Sym A)(ξ, −∞)
(ξ,x) : Pϕ (Ṁarc )ϕ(Aπ ) + J(ξ,x) →
π if (ξ, x) ∈ M0,∞ ,
⎪
⎩
(Sym A)(ξ, +∞) if (ξ, x) ∈ Ṁ± .
(8.3)
Since the matrices (SymA)(ξ,±∞) are diagonal for all A ∈ A and all ξ ∈ M (QC(T)),
we conclude from (8.3) that the C ∗ -algebras Aarc /J(ξ,x) are commutative, and
therefore for every (ξ, x) ∈ M #̇ \ M◦ the set P =P of pure states
arc arc (ξ,x) Aarc /J(ξ,x)
of the C ∗ -algebra Aarc /J(ξ,x) consists of four multiplicative linear functionals whose
values coincide with the diagonal entries of matrices (Sym A)(ξ, ±∞) if (ξ, x) ∈
M0,∞ , and consists of two multiplicative linear functionals whose values coincide
with the diagonal entries of the matrix (Sym A)(ξ, +∞) if (ξ, x) ∈ Ṁ± .
#̇
Hence, for every ξ ∈ M (QC(T)) and every (ξ, x) ∈ Marc \ M◦arc ,
 (1) (2) (1) (2) 
P(ξ,x) = P(ξ,∞) := ρξ,+∞ , ρξ,+∞ , ρξ,−∞ , ρξ,−∞ ,
(j)
where the pure states ρξ,±∞ for j = 1, 2 are given by
(j)
ρξ,±∞ : Aarc → C, Pϕ (Ṁarc )ϕ(Aπ ) → (Sym A)jj (ξ, ±∞), (8.4)
50 M.A. Bastos, C.A. Fernandes and Yu.I. Karlovich

(Sym A)jj (ξ, ±∞) is the (j, j)-entry of the matrix (Sym A)(ξ, ±∞), and the set
P(ξ,∞) consists of two elements for all (ξ, x) ∈ Ṁ± because in that case ρξ,+∞ =
(1)

(2) (2) (1)

ρξ,−∞ and ρξ,+∞ = ρξ,−∞ .
Since βg (ξ, x) = (ξ, x) for all (ξ, x) ∈ Ṁarc , to verify the fulfillment of condi-
tion (A3) of the local-trajectory method for the C ∗ -algebra Barc , it only remains
to prove the approximability (in the weak* topology) of the pure states ρ ∈ P(ξ,∞) ,
with ξ ∈ Ṁ∞Λ , by pure states in P(ξ,x) , with (ξ, x) ∈ Ṁarc . In that case M0 = Ṁarc
in condition (A3).
Fix ξ ∈ Mt (QC(T)) for t ∈ Λ. By (8.4) and the proof of Lemma 6.2, ev-
(1) (2)
ery open neighborhood of ρξ,±∞ and ρξ,±∞ in the weak* topology contains, re-
(1) (2) 
spectively, pure states ρζ,±∞ and ρζ,±∞ , where ζ ∈ τ ∈T\Λ Mτ (QC(T)) and τ
(1) (2)
is on the right of t for ρξ,±∞ and ρξ,±∞ if, respectively, ξ ∈ Mt+ (QC(T)) and
ξ ∈M #t+ (QC(T)), and τ is on the left of t for ρ(1) and ρ(2) if, respectively,
ξ,±∞ ξ,±∞
ξ ∈M #t− (QC(T)) and ξ ∈ Mt− (QC(T)). Thus, condition (A3) is also fulfilled for
the C ∗ -algebra Barc .
For each (ξ, x) ∈ Marc , we consider the representation
π(ξ,x) : Barc → B(l2 (G, C2 )) (8.5)
∗
given on the generators of the C -algebra Barc by
   
π(ξ,x) Pϕ (Ṁarc )ϕ((aI)π ) f (g) = (Sym ((a ◦ g)I))(ξ, x) f (g),
   
π(ξ,x) Pϕ (Ṁarc )ϕ(STπ ) f (g) = (Sym ST )(ξ, x) f (g), (8.6)
 
π(ξ,x) Pϕ (Ṁarc )ϕ(Uhπ ) f (g) = f (gh),
where a ∈ P QC(T), g, h ∈ G and f ∈ l2 (G, C2 ).
Fix a set Oarc ⊂ T \ Λ which contains exactly one point in each orbit defined
by the group of shifts G on T \ Λ, and consider the sets
Ṅarc := Mτ (QC(T)) × Ṙ, Narc = Mτ (QC(T)) × R. (8.7)
τ ∈Oarc τ ∈Oarc

Theorem 8.1. For each B ∈ B, the operator Barc = Pϕ (Ṁarc )ϕ(B π ) ∈ Barc is
invertible on the space Pϕ (Ṁarc )Hϕ if and only if for all (ξ, x) ∈ Narc the operators
π(ξ,x) (Barc ) are invertible on the space l2 (G, C2 ) and
$ $
sup $(π(ξ,x) (Barc ))−1 $ < ∞. (8.8)
(ξ,x)∈Narc

Proof. The set Ṅarc given by (8.7) contains exactly one point in each G-orbit
defined on the set Ṁarc ⊂ M#̇ by the group {β
arc g, arc : g ∈ G} of homeomorphisms
given by (8.1). Thus, for each (ξ, x) ∈ Ṅarc , starting with the representations
(8.2)–(8.3) and following (4.3)–(4.5), we obtain the family of representations (8.5)
defined by (8.6) and indexed by the points of the set Narc , where we take into
 A C ∗ -algebra of Singular Integral Operators with Shifts 51

account the fact that the Hilbert spaces l2 (G, C4 ) and l2 (G, C2 ) ⊕ l2 (G, C2 ) are
isometrically isomorphic. Since assumptions (A1)–(A3) are fulfilled for the C ∗ -
algebra Barc , Theorem 4.1 implies the assertion of the theorem. 
For each (ξ, x) ∈ Narc , we now consider the representation
B → B(l2 (G, C2 )), B → π(ξ,x) (Barc ). (8.9)
It is easily seen from (8.5)–(8.6) that representation (8.9) coincides with the rep-
resentation Φξ,x given by (3.2)–(3.3), that is,
Φξ,x (B) = π(ξ,x) (Barc ) for all B ∈ B and all (ξ, x) ∈ Narc .
Consequently, Theorem 8.1 is equivalent to part (i) of Theorem 3.1.

9. The C ∗ -algebra B◦Λ

In this section we establish an invertibility criterion for the operators in the C ∗ -
algebra B◦Λ = Pϕ (M◦Λ )ϕ(Bπ ) represented in the form (6.15).
Theorem 9.1. The mapping given by
Pϕ (M◦Λ )ϕ(Aπ ) → Pφ (M◦Λ )φ(Aπ ) for all A ∈ A (9.1)
∗ ∗
is a C -algebra isomorphism of the C -algebra A◦Λ := Pϕ (M◦Λ )ϕ(Aπ ) onto the
 ◦ := Pφ (M◦ )φ(Aπ ).
C ∗ -algebra A Λ Λ

Proof. Since M◦Λ is an open subset of Ṁ and the C ∗ -algebras ϕ(Aπ ) and φ(Aπ )
are isometrically ∗ -isomorphic, applying [5, Lemma 3.5], we infer that
Pϕ (M◦Λ )ϕ(Aπ )B(Hϕ ) = Pφ (M◦Λ )φ(Aπ )B(Hφ ) for all A ∈ A (9.2)
by analogy with (7.6). This implies that (9.1) is a well-defined isometric ∗ -iso-
◦ .
morphism of the C ∗ -algebra A◦Λ onto the C ∗ -algebra A 
Λ

Let M◦Λ := t∈Λ M◦t , where M◦t is given by (6.2). Since the Hilbert space
Pφ (M◦Λ )Hφ is isometrically isomorphic to the Hilbert space
◦
HΛ := l2 (MΛ0 (QC(T)), L22 (R)), (9.3)
where MΛ0 (QC(T)) is given by (6.7), we immediately obtain from (6.8) and (6.13)
the following.
Lemma 9.2. The mapping

 ◦ → B(H◦ ),
A Pφ (M◦Λ )φ(Aπ ) → (Sym A)(ξ, ·)I
Λ Λ
0 (QC(T))
ξ∈MΛ

is an isometric C ∗ -algebra homomorphism. An operator Pφ (M◦Λ )φ(Aπ ) for A ∈ A

is invertible on the space Pφ (M◦Λ )Hφ if and only if
det((Sym A)(ξ, x)) = 0 for all (ξ, x) ∈ M◦Λ . (9.4)
Theorem 9.1 and Lemma 9.2 imply the following.
52 M.A. Bastos, C.A. Fernandes and Yu.I. Karlovich

Theorem 9.3. The map Sym◦Λ : A◦Λ → B(HΛ

◦
), defined by

◦
Pϕ (MΛ )ϕ(A ) →
π
(Sym A)(ξ, ·)I f or A ∈ A,
0 (QC(T))
ξ∈MΛ

is an isometric C ∗ -algebra homomorphism. An operator Pϕ (M◦Λ )ϕ(Aπ ) for A ∈ A

is invertible on the space Pϕ (M◦Λ )Hϕ if and only if (9.4) holds.

Given t ∈ Λ, let H  π be the closed two-sided ideal of the C ∗ -algebra Z π

t
π
generated by all the cosets HP,t = HP,t + K with P ∈ P (see (5.20) and (5.21)).
∗
Since Z is a commutative C -algebra, we conclude that
π
 m 

Ht = clos
π
Zj HPj ,t : Zj ∈ Z, Pj ∈ P, m ∈ N ,
π π
(9.5)
j=1

where the closure is taken in the essential norm ·. Let H π be the closed two-sided
Λ
ideal of Z generated by all ideals H
π  (t ∈ Λ).
π
t
Following [5, Lemma 3.5], we define the set
% &
Z(M◦Λ ) := Z π ∈ Z π : supp z(·, ·) ⊂ Ṁ◦Λ , z(ξ, x) ∈ [0, 1] for all (ξ, x) ∈ Ṁ ,
(9.6)
where z(·, ·) ∈ C(Ṁ) is the Gelfand transform of the coset Z π , supp z(·, ·) is the

support of z(·, ·), and Ṁ◦Λ := t∈Λ Mt0 (QC(T)) × Ṙ is the closure in Ṁ of the set
M◦Λ given by (6.3).
 π , the following assertions hold:
Lemma 9.4. For the ideal H Λ
(i) for every B π ∈ Bπ and every H π ∈ H π ,
Λ Λ
B π
HΛπ = π π
HΛ B ∈ Aπ ; (9.7)
 π = Z(M◦ ).
(ii) H Λ Λ

Proof. (i) For B π ∈ Aπ , the property (9.7) follows from the fact that Z π is a central
subalgebra of Aπ . Since every coset H π ∈ H  π is of the form H π =  π
Λ t∈Λ Ht , where
Htπ ∈ H π , we can see that relation (9.7) for B π = U π follows from Lemma 5.3 and
t g
(5.20). Hence, part (i) is proved for every B π ∈ Bπ .
(ii) Fix t ∈ Λ. The symbols of operators HP,t (P ∈ P) are equal to 02×2
for all (ξ, x) ∈ M \ M◦t (see Lemma 5.4). Hence, by Theorem 5.6, the Gelfand
transform of the coset HP,t π
is a continuous function on Ṁ with support contained
in the closure Ṁt = Mt (QC(T)) × Ṙ of the set M◦t in Ṁ. By (9.5), this property
◦ 0

also holds for the Gelfand transform of any element Htπ of the ideal H  π . Thus,
t
 ◦
Ht ⊂ Z(Mt ), where
π
% &
Z(M◦t ) := Z π ∈ Z π : supp z(·, ·) ⊂ Ṁ◦t , z(ξ, x) ∈ [0, 1] for all (ξ, x) ∈ Ṁ .
On the other hand, we infer from [5, Lemma 6.2] and (9.5) that any element
 π , which
Z π ∈ Z(M◦t ) can be approximated in the C ∗ -algebra Z π by cosets Htπ ∈ H t
 A C ∗ -algebra of Singular Integral Operators with Shifts 53

 π . Moreover, Theorem 5.6 and [5, Lemma 6.2] imply that

implies that Z(M◦t ) ⊂ H t

the ideal Ht is isomorphic to the ideal of all continuous functions on the compact
π

 π = Z(M◦ ), which together with (9.6) implies

Ṁ that vanish on Ṁ \ M◦t . Hence, H t t
part (ii). 
◦
Consider the Hilbert space HΛ given by (9.3), with MΛ0 (QC(T)) given by
∗
(6.7), and introduce the C -algebra
 
Ψ(B◦Λ ) := alg Ψ◦Λ (Aπ ), Ψ◦Λ (Ugπ ) : A ∈ A, g ∈ G ⊂ B(HΛ◦
) (9.8)
generated by the operators Ψ◦Λ (Aπ ) (A ∈ A) and Ψ◦Λ (Ugπ ) (g ∈ G), where
 
Ψ◦Λ (Aπ ) := (Sym A)(ξ, ·)I, Ψ◦Λ (Ugπ ) := eln g (ξ) (·)I,
0 (QC(T))
ξ∈MΛ 0 (QC(T))
ξ∈MΛ
(9.9)
ix ln g (ξ)
and eln g (ξ) (x) := e M◦Λ .
for all (ξ, x) ∈
The mapping g → Ψ◦Λ (Ugπ ) is a unitary representation of the group G in the
◦
Hilbert space HΛ , the adjoint operator Ψ◦Λ (Ugπ )∗ equals Ψ◦Λ (Ugπ−1 ), and
Ψ◦Λ (Ugπ )Ψ◦Λ (Aπ )Ψ◦Λ (Ugπ )∗ = Ψ◦Λ (Aπ )
for all g ∈ G and all A ∈ A due to (9.9). Consequently, the C ∗ -algebra Ψ(B◦Λ )
is the closure of the algebra Ψ(B◦Λ )0 composed by the finite sums of the form
 ◦ ◦
g ΨΛ (Ag )ΨΛ (Ug ), where Ag ∈ A.
π π

Theorem 9.5. The mapping

  
◦
Pϕ (MΛ )ϕ Ag Ug →
π π
Ψ◦Λ (Aπg )Ψ◦Λ (Ugπ ), (9.10)
g∈F g∈F

where F is a finite subset of G and Ag ∈ A for g ∈ F , extends to a C ∗ -algebra

isomorphism of the C ∗ -algebra B◦Λ onto the C ∗ -algebra Ψ(B◦Λ ) given by (9.8).

Proof. Fix an operator B = g∈F Ag Ug ⊂ B, where F is a finite subset of G

and Ag ∈ A for g ∈ F , and put Ψ◦Λ (B π ) := g∈F Ψ◦Λ (Aπg )Ψ◦Λ (Ugπ ). Since the set
M◦Λ is open and since Pϕ (M◦Λ )ϕ(B π ) = ϕ(B π )Pϕ (M◦Λ ), we infer similarly to [5,
Lemma 3.5] that
Pϕ (M◦Λ )ϕ(B π )B(Hϕ ) = sup ϕ(B π Z π )B(Hϕ ) , (9.11)
Z π ∈Z(M◦
Λ)

where the set Z(M◦Λ ) is given by (9.6).

For each coset HΛπ ∈ H  π , it follows from Lemma 9.4(i) that B π H π ∈ Aπ .
Λ Λ
Hence, by (6.8), (5.3), (9.9) and by Lemmas 5.3–5.4, we obtain φΛ (B π HΛπ ) =
Ψ◦Λ (B π )φΛ (HΛπ ) where φΛ : Aπ → B(HΛ ◦
) is the restriction of the representation
◦
φ (see (6.6), (6.8) and (6.11)) to the space HΛ considered as an invariant Hilbert
subspace of Hφ . Therefore, applying (9.2) or Theorem 9.3, we get
Pϕ (M◦Λ )ϕ(B π HΛπ )B(Hϕ ) = Pφ (M◦Λ )φ(B π HΛπ )B(Hφ )
(9.12)
= Ψ◦Λ (B π )φΛ (HΛπ )B(H◦Λ ) .
54 M.A. Bastos, C.A. Fernandes and Yu.I. Karlovich

 π due to Lemma 9.4(ii) and since

Since Z(M◦Λ ) = H Λ

Pϕ (M◦Λ )ϕ(B π Z π ) = ϕ(B π Z π ) for all Z π ∈ Z(M◦Λ ),

we deduce from (9.5) and the equalities (9.11)–(9.12) that
sup Ψ◦Λ (B π )φΛ (Z π )B(H◦Λ ) = sup Pϕ (M◦Λ )ϕ(B π Z π )B(Hϕ )
Z π ∈Z(M◦
Λ) Z π ∈Z(M◦
Λ)

= sup ϕ(B π Z π )B(Hϕ ) = Pϕ (M◦Λ )ϕ(B π )B(Hϕ ) . (9.13)

Z π ∈Z(M◦
Λ)

We now consider the identical representation π of the unital C ∗ -algebra

Ψ(B◦Λ ) in the Hilbert space HΛ ◦
. By (9.9), φΛ (Z π ) is a central C ∗ -subalgebra of
Ψ(B◦Λ ) with the same unit. Clearly, the maximal ideal space of φΛ (Z π ) coincides
with Ṁ◦Λ . Since the set M◦Λ is an open subset of ṀΛ and since the correspond-
ing spectral projection Pπ (M◦Λ ) is the identity operator on Hilbert space HΛ ◦
, we
conclude from [5, Lemma 3.5] that
Ψ◦Λ (B π )B(H◦Λ ) = Pφ (M◦Λ )Ψ◦Λ (B π )B(H◦Λ )
= sup Ψ◦Λ (B π )φΛ (Z π )B(H◦Λ ) ,
Z π ∈Z(M◦
Λ)

which together with (9.13) implies that

Pϕ (M◦Λ )ϕ(B π )B(Hϕ ) = Ψ◦Λ (B π )B(H◦Λ ) (9.14)

for all finite sums B π = g∈F Aπg Ugπ ∈ Bπ with Aπg ∈ Aπ . Since the set of such
finite sums is dense in Bπ and since (9.14) holds, the mapping (9.10) uniquely
extends to a C ∗ -algebra isomorphism of B◦Λ onto Ψ(B◦Λ ). 
Every coset π
B of the C ∗ -algebra Bπ is the limit of a sequence of cosets of
the form Bn = g∈Fn Ag,n Ugπ where Aπg,n ∈ Aπ and g runs through finite subsets
π π

Fn of G (n ∈ N). Then, according to Theorem 9.5, the operator ΨΛ (B π ) in the

C ∗ -algebra Ψ(B◦Λ ) has the form

Ψ◦Λ (B π ) = lim Ψ◦Λ (Aπg,n )Ψ◦Λ (Ugπ ),
n→∞
g∈Fn
∗
where the -homomorphism Ψ◦Λ
: B → B(HΛ ◦
) is an extension of the ∗ -homo-
π
◦ ∗
morphism φΛ : A → B(HΛ ) to the C -algebra Bπ in view of (6.8) and (9.9).
π

Thus,

Ψ◦Λ (B π ) = BΛ◦ (ξ, ·)I ∈ B(HΛ ◦
),
0 (QC(T))
ξ∈MΛ
 (9.15)
BΛ◦ (ξ, ·) : R → C2×2 , x → lim (Sym Ag,n )(ξ, x) eln g (ξ) (x).
n→∞
g∈Fn

Taking into account the inverse closedness of the C ∗ -algebras B◦Λ and Ψ(B◦Λ )
in the C ∗ -algebras B(Pϕ (M◦Λ )Hϕ ) and B(HΛ ◦
), respectively, we immediately obtain
an invertibility criterion for the operators in the C ∗ -algebra B◦Λ from Theorem 9.5
and (9.15).
 A C ∗ -algebra of Singular Integral Operators with Shifts 55

Theorem 9.6. For each B ∈ B, the operator BΛ◦ = Pϕ (M◦Λ )ϕ(B π ) ∈ B◦Λ is in-
vertible on the space Pϕ (M◦Λ )Hϕ if and only if the operator Ψ◦Λ (B π ) ∈ Ψ(B◦Λ ) is
◦
invertible on the space HΛ = l2 (MΛ0 (QC(T)), L22 (R)), that is, if
min inf det(BΛ◦ (ξ, x)) > 0.
0 (QC(T)) x∈R
ξ∈MΛ

For each (ξ, x) ∈ NΛ , where NΛ is given by (3.1), we introduce the represen-

tation
B → B(C2 ), B → BΛ◦ (ξ, x)I, (9.16)
given on the generators of the C ∗ -algebra B by the following 2×2 matrix functions
BΛ◦ (·, ·) according to (5.3), (5.4) and (9.15):
 
(aI)◦Λ (ξ, x) = diag a(ξ, 1), a(ξ, 0) ,
 
tanh(πx) 1/ cosh(πx)
(ST )◦Λ (ξ, x) = , (9.17)
1/ cosh(πx) − tanh(πx)
   
(Uh )◦Λ (ξ, x) = diag eix ln h (ξ) , eix ln h (ξ) ,
where a ∈ P QC(T), a(ξ, μ) is the value of the Gelfand transform of a at the
point (ξ, μ) ∈ M (P QC(T)) and h ∈ G. It is easily seen from (9.17) that for each
(ξ, x) ∈ NΛ representation (9.16) coincides with the representation Φξ,x defined
by (3.4)–(3.5), that is,
Φξ,x (B) = BΛ◦ (ξ, x)I for all B ∈ B and all (ξ, x) ∈ NΛ .
Hence, identifying the operators Φξ,x (B) and the matrices BΛ◦ (ξ, x), we conclude
that Theorem 9.6 is equivalent to part (ii) of Theorem 3.1.

10. Invertibility of functional operators

Let us study the invertibility of functional operators in the C ∗ -algebra
A := alg(P QC(T), UG ) ⊂ B(L2 (T))
generated by the multiplication operators by piecewise quasicontinuous functions
on T and by the isometric shift operators Ug (g ∈ G).
The C ∗ -algebra A isthe closure of the algebra A0 ⊂ A consisting of the
functional operators A = g∈F ag Ug , where ag ∈ P QC 0 (T) and F runs through
the finite subsets of G. In order to obtain an invertibility criterion for the operators
A ∈ A, we will apply the local-trajectory method.
 := Z := {aI : a ∈ P QC(T)}. As Z ∼
Let A = P QC(T), we get M (Z)  =
M (P QC(T)). Let us check the fulfillment of assumptions made in Section 4.
Lemma 2.3 implies that for every g ∈ G and all a ∈ P QC(T) the mapping
g : aI → Ug aUg−1 = (a◦ g)I is a ∗ -automorphism of the commutative C ∗ -algebras
α
 = Z ⊂ B(L2 (T)). Since G is a commutative and therefore an amenable group,
A
we see that conditions (A1)–(A2) for the C ∗ -algebra A are satisfied.
56 M.A. Bastos, C.A. Fernandes and Yu.I. Karlovich

For every g ∈ G, the ∗ -automorphism α

g induces the homeomorphism
βg : M (P QC(T)) → M (P QC(T)), (ξ, μ) → (g(ξ), μ), (10.1)
where g(ξ) ∈ M (QC(T)) is given by (5.30). If the ideal ξ ∈ M (QC(T)) be-
longs to the fiber Mt (SO(T)), then g(ξ) ∈ Mg(t) (QC(T)). Moreover, if ξ be-
longs to M 0 (QC(T)) or M #± (QC(T)), respectively, then so is g(ξ). Hence, taking
into account the topologically free action of the group G on M (QC(T)) (see [11,
Lemma 4.2]) and the Gelfand topology (2.5) on M (P QC(T)), we easily conclude
that condition (A3) for theC ∗ -algebra A also holds, with M0 := MT\Λ (P QC(T)),
where MT\Λ (P QC(T)) := τ ∈T\Λ Mτ (P QC(T)).
With each maximal ideal (ξ, μ) ∈ M (P QC(T)) we associate the representa-
tion
Π(ξ,μ) : A → B(l2 (G)), A → A(ξ,μ) (10.2)

given for the operators A = g∈F ag Ug ∈ A by 0


(A(ξ,μ) f )(h) = [(ag ◦ h)(ξ, μ)]f (hg) (h ∈ G, f ∈ l2 (G)). (10.3)
g∈F

Following Section 8, we fix a set Oarc ⊂ T \ Λ containing exactly one point

in each orbit defined by the group of shifts G on T \ Λ and consider the set
Rarc := Mτ (P QC(T)). (10.4)
τ ∈Oarc

The set Rarc contains exactly one point in each G-orbit defined by the action of
the group G on MT\Λ (P QC(T)) by means of the homeomorphisms βg (g ∈ G)
given by (10.1).
Since conditions (A1)–(A3) are fulfilled, we get the following.
Theorem 10.1. A functional operator A ∈ A is invertible on the space L2 (T) if and
only if for all (ξ, μ) ∈ Rarc the operators A(ξ,μ) are invertible on the space l2 (G)
and $ $
sup $(A(ξ,μ) )−1 $ < ∞. (10.5)
(ξ,μ)∈Rarc

Proof. Take the maximal ideal J(ξ,μ) := {aI : a ∈ P QC(T), a(ξ, μ) = 0} of Z

 = Z,
associated to each functional (ξ, μ) ∈ M (P QC(T)). As A  the mapping

Π  (ξ,μ) → C,
 (ξ,μ) : A/J aI + J(ξ,μ) → a(ξ, μ),
 (ξ,μ) in C. Following (4.3)–
is an isometric representation of the C ∗ -algebra A/J
∗
(4.5), we construct representations of the C -algebra A in the Hilbert space l2 (G)
by formulas (10.2) and (10.3). Since A satisfies conditions (A1)–(A3) of the local-
trajectory method, Theorem 4.1 immediately implies the statement of the theorem.

Applying Theorems 10.1 and 8.1 we can prove the following result.
 A C ∗ -algebra of Singular Integral Operators with Shifts 57

Theorem 10.2. For any functional operator A ∈ A, its Fredholmness on the space
L2 (T) is equivalent to its invertibility on this space.
Proof. Obviously, we only need to prove that the Fredholmness of each functional
operator A implies its invertibility. Suppose that an operator A ∈ A is Fredholm
on the space L2 (T). Then the operator Aarc := Pϕ (Ṁarc )ϕ(Aπ ) is invertible on the
Hilbert space Pϕ (Ṁarc )Hϕ . Consequently, by Theorem 8.1, for all (ξ, x) ∈ Narc
the operators π(ξ,x) (Aarc ) are invertible on the space l2 (G, C2 ) and
$ $
sup $(π(ξ,x) (Aarc ))−1 $ < ∞. (10.6)
(ξ,x)∈Narc

By (8.5)–(8.6), (5.3) and (10.2)–(10.3), we infer that, for all x ∈ R,

⎧
#− (QC(T)),
⎨diag{A(ξ,0) , A(ξ,0) } if
⎪ ξ∈M T\Λ
π(ξ,x) (Aarc ) = diag{A(ξ,1) , A(ξ,0) } if ξ ∈ MT\Λ
0
(QC(T)), (10.7)
⎪
⎩
diag{A(ξ,1) , A(ξ,1) } if ξ∈M #+ (QC(T)),
T\Λ

where
#± (QC(T)) :=
M #± (QC(T)), M 0 (QC(T)) :=
M Mτ0 (QC(T)).
T\Λ τ T\Λ
τ ∈T\Λ τ ∈T\Λ
(10.8)
Comparing the sets Narc and Rarc given by (3.1) and (10.4), respectively, we con-
clude from (10.7) and (10.6) that for every (ξ, μ) ∈ Rarc the operator A(ξ,μ) is
invertible on the space l2 (G) and condition (10.5) is fulfilled. Hence, by Theo-
rem 10.1, the operator A is invertible on the space L2 (T). 

11. The C ∗ -algebra BΛ

In this section we will show that for every B ∈ B the invertibility of the operator
Barc = Pϕ (Ṁarc )ϕ(B π ) on the Hilbert space Pϕ (Ṁarc )Hϕ implies the invertibility
of the operators BΛ = Pϕ (MΛ )ϕ(B π ) on the Hilbert space Pϕ (MΛ )Hϕ . This
means that condition (iii) in Theorem 6.3 is superfluous.
Observe that the C ∗ -algebra B = alg(P QC(T), ST , UG ) can be viewed as the
C -algebra B = alg(A, ST ) generated by the C ∗ -algebra A studied in Section 10
∗

and by the Cauchy singular integral operator ST . By analogy with Theorem 5.5,
we can get another general form of operators B ∈ B.
Let B0 denote the dense non-closed
n subalgebra of the C ∗ -algebra B consist-
ing of all operators of the form i=1 Ti1 Ti2 . . . Tiji where n, ji ∈ N and Ti,k ∈
A0 ∪ {ST }. We denote by H the closed two-sided ideal of B generated by all the
commutators [aI, ST ] := aST − ST aI with a ∈ P C(T), that is, the closure of the
set
 n 
H :=
0
Bi Hi Ci : Bi , Ci ∈ B , Hi = [ai I, ST ], ai ∈ P C (T), n ∈ N .
0 0

i=1
58 M.A. Bastos, C.A. Fernandes and Yu.I. Karlovich

The ideal K of all compact operators on L2 (T) is contained in H (see, e.g., [16]).
Hence, the commutators [aI, ST ] (a ∈ QC(T)) and [Ug , ST ] (g ∈ G) belong to the
ideal H because these operators are compact on the space L2 (T) according to (2.6)
and (2.7). As a result, for every A ∈ A the commutators [A, ST ] also are in H.
Let A be the C ∗ -algebra of the 2 × 2 diagonal matrices with A-valued entries.
We now obtain an analogue of Theorem 5.5 for the C ∗ -algebra B (cf. [20], [21]).
Theorem 11.1. Every operator B ∈ B is uniquely represented in the form
B = A+ PT+ + A− PT− + HB , (11.1)
±
where A are functional operators in the C -algebra A,∗
PT±
= (I ± ST )/2 are
the orthogonal projections associated with the Cauchy singular integral operator
ST , HB ∈ H, the map B → diag{A+ , A− } is a C ∗ -algebra homomorphism of the
C ∗ -algebra B onto the C ∗ -algebra A with kernel H, and
A±  ≤ inf B + H ≤ |B| = inf B + K. (11.2)
H∈H K∈K

Proof. Obviously, every operator B ∈ B0 is represented in the form (11.1), and

the mapping B → diag{A+ , A− } is an algebraic homomorphism of the non-closed
algebra B0 into A with the kernel contained in H0 . This mapping is defined for
the generators of the algebra B by
aI → diag{aI, aI}, Ug → diag{Ug , Ug }, ST → diag{I, −I}.
Obviously, it only remains to prove (11.2) for all B ∈ B0 .
Taking into account the facts that the ideal H ⊂ B is generated by all the
commutators [aI, ST ] (a ∈ P C(T)) and that
(Sym([aI, ST ]))(ξ, ±∞) = 02×2 for all a ∈ P C(T) and all ξ ∈ M (QC(T))
according to (5.3), we infer from (8.6) that for any operator H ∈ H,
π(ξ,±∞) (Harc ) = 0 for all ξ ∈ M (QC(T)), (11.3)
where Harc := Pϕ (Ṁarc )ϕ(H π ). Analogously, by (5.3), for all ξ ∈ M (QC(T)),
(Sym PT± )(ξ, ±∞) = diag{1, 0}, (Sym PT± )(ξ, ∓∞) = diag{0, 1}. (11.4)
Further, from (10.7) it follows that for x ∈ {+∞, −∞},
⎧  ± 
± #−
⎨diagA(ξ,0) , A(ξ,0)  if ξ ∈ MT\Λ (QC(T)),
⎪
π(ξ,x) (A±
arc ) = diag A± ±
(ξ,1) , A(ξ,0) if ξ ∈ MT\Λ
0
(QC(T)), (11.5)
⎪
⎩  
diag A ±
,A ± #
if ξ ∈ M (QC(T)),
+
(ξ,1) (ξ,1) T\Λ

where the sets on the right of (11.5) are given by (10.8). Hence, for every operator
B = A+ PT+ + A− PT− + HB ∈ B0 we deduce from (11.3)–(11.5) that
⎧ ± 
∓ #−
⎨A(ξ,0) , A(ξ,0)  if ξ ∈ MT\Λ (QC(T)),
⎪
π(ξ,±∞) (Barc ) = diag A± ∓
(ξ,1) , A(ξ,0) if ξ ∈ MT\Λ
0
(QC(T)), (11.6)
⎪
⎩ ± 
A ,A ∓ #
if ξ ∈ M (QC(T)).
+
(ξ,1) (ξ,1) T\Λ
 A C ∗ -algebra of Singular Integral Operators with Shifts 59

± 
Therefore, letting MT\Λ (QC(T)) := τ ∈T\Λ Mτ± (QC(T)), we infer that
 
max sup A±(ξ,0) , sup A±
(ξ,1) 
− +
ξ∈MT\Λ (QC(T)) ξ∈MT\Λ (QC(T)) (11.7)
≤ Pϕ (Ṁarc )ϕ(B + H ) ≤ B + H
π π
for all H ∈ H.
By Theorem 10.1, the C ∗ -algebra A is isometrically ∗ -isomorphic to the C ∗ -
algebra
   
A(ξ,μ) : A ∈ A ⊂ B(l2 (G))
(ξ,μ)∈Rarc (ξ,μ)∈Rarc

equipped with the norm sup(ξ,μ)∈Rarc A(ξ,μ) . Therefore, for every A ∈ A,

A = sup A(ξ,μ) . (11.8)
(ξ,μ)∈Rarc

Finally, from (11.7) and (11.8) it follows that A±  ≤ B + H for all B ∈ B0
and all H ∈ H, which immediately implies (11.2). 

Let G(T) denote the set of the G-orbits G(t) := {g(t) : g ∈ G} ⊂ T for
all t ∈ T. For each w ∈ G(T) we denote by Hw the closed two-sided ideal of B
generated by the operator Vt given by (5.6), where t is an arbitrary point of the
G-orbit w. Observe that Vt Vτ  0 for t, τ ∈ T and t = τ . Hence, Hw Hv  0 if
w, v ∈ G(T) and w = v. Similarly to [21, Lemma 5.4] we get the following result.

Lemma 11.2. Every operator H ∈ H0 is of the form H  w∈GH Hw , where GH
is an at most finite subset of G(T) and Hw ∈ Hw .
Consider now the C ∗ -algebra BΛ = Pϕ (MΛ )ϕ(Bπ ) (see (6.16)). Given t ∈ Λ,
let Ht be the closed two-sided ideal of B generated by the operator Vt given by
(5.6). By Lemma 9.4(i), the ideal Ht is contained in the C∗ -algebra A. Clearly, the
ideal H is generated by all ideals Ht (t ∈ T).
Lemma 11.3. If H ∈ H, then
Pϕ (MΛ )ϕ(H π ) = 0. (11.9)

Proof. First, let us prove that, for the set MΛ closed in Ṁ,
Pϕ (MΛ )ϕ(Vτπ ) = 0 for all τ ∈ T. (11.10)
-
Fix τ ∈ T. Let Δn be a sequence of open sets of Ṁ such that n Δn = MΛ .
Analogously to Theorem 7.1 one can establish that
Pϕ (Δn )ϕ(Vτπ )B(Hϕ ) = Pφ (Δn )φ(Vτπ )B(Hφ ) for all n ∈ N. (11.11)
If τ ∈ T \ Λ, then for the Hilbert space Hφ we have the equality
lim Pφ (Δn )φ(Vτπ )B(Hφ ) = 0 (11.12)
n→∞
60 M.A. Bastos, C.A. Fernandes and Yu.I. Karlovich

because τ ∈
/ Δn for all sufficiently large n, where Δn is the closure of Δn in Ṁ.
Hence, we derive from (11.11) that, for the abstract Hilbert space Hϕ ,
lim Pϕ (Δn )ϕ(Vτπ )B(Hϕ ) = 0 (11.13)
n→∞
for all τ ∈ T \ Λ. Let now τ ∈ Λ. Then we conclude from Lemma 5.2 that again
(11.12) holds, which in view of (11.11) gives (11.13) for all τ ∈ Λ. Thus,
lim Pϕ (Δn )ϕ(Vτπ )B(Hϕ ) = 0 for all τ ∈ T. (11.14)
n→∞
Let ϕ|Z π be the restriction of the representation ϕ to the central C ∗ -subal-
gebra Z π . Consider the decomposition of the representation ϕ|Z π into the direct
sum of cyclic representations ϕα in the mutually orthogonal subspaces
 Hϕ,α of the

space Hϕ . Then, by [5, Lemma 3.4], for each subspace Hϕ = α∈Q Hϕ,α ⊂ Hϕ
with a finite set Q, we have
Pϕ (MΛ )ϕ(Vτπ ) = s-lim Pϕ (Δn )ϕ(Vτπ ).
n→∞
Hence,
0 ≤ Pϕ (MΛ )ϕ(Vτπ )B(Hϕ ) ≤ lim inf Pϕ (Δn )ϕ(Vτπ )B(Hϕ )
n→∞
≤ lim inf Pϕ (Δn )ϕ(Vτπ )B(Hϕ ) ,
n→∞

which implies according to (11.14) that for all considered subspaces Hϕ ⊂ Hϕ ,
Pϕ (MΛ )ϕ(Vτπ )B(Hϕ ) = 0. (11.15)
Since the set of vectors ξ = {ξα }α ∈ Hϕ with a finite number of non-zero entries
ξα ∈ Hϕ,α is dense in Hϕ , the relations (11.15) imply (11.10).
Consider now an operator Hw ∈ Hw with w ∈ G(T). It follows from (5.16),
(5.18) and (5.28) that 
Hwπ = lim π
Bn,τ Vτπ , (11.16)
n→∞
τ ∈Gn,w
where Gn,w are finite subsets of the G-orbit w and Bn,τ are operators in the C ∗ -
algebra B. Since the set MΛ belongs to the family RG (Ṁ) defined in (6.5), we
infer from (4.8) that for all B ∈ B,
Pϕ (MΛ )ϕ(B π ) = Pϕ (MΛ )ϕ(B π )Pϕ (MΛ ).
Therefore, from (11.16) and (11.10) it follows that

Pϕ (MΛ )ϕ(Hwπ ) = lim Pϕ (MΛ )ϕ(Bn,τπ
)Pϕ (MΛ )ϕ(Vτπ ) = 0. (11.17)
n→∞
τ ∈Gn,w (T)

Finally, by Lemma 11.2, for any operator H ∈ 

H0 there are a finite subset GH of
G(T) and operators Hw ∈ Hw such that H = w∈GH Hwπ . Hence, by (11.17), for
π

H ∈ H0 we obtain

Pϕ (MΛ )ϕ(H π ) = Pϕ (MΛ )ϕ(Hwπ ) = 0,
w∈GH

which implies (11.9) because the set H0 is dense in H. 

 A C ∗ -algebra of Singular Integral Operators with Shifts 61

Lemma 11.3 shows that the operators H ∈ H do not have influence on the
operators in the C ∗ -algebra BΛ and, in particular, implies the following.
Corollary 11.4. If A is a functional operator in A, then
Pϕ (MΛ )ϕ([A, ST ]π ) = 0.
Theorem 11.5. If B ∈ B and the operator Barc = Pϕ (Ṁarc )ϕ(B π ) is invertible on
the Hilbert space Pϕ (Ṁarc )Hϕ , then the operator BΛ := Pϕ (MΛ )ϕ(B π ) is invertible
on the Hilbert space Pϕ (MΛ )Hϕ .
Proof. Fix an operator B ∈ B. By Theorem 11.1, the operator B has the form
B = A+ PT+ + A− PT− + HB , with A± ∈ A and HB ∈ H. We then infer from
Lemma 11.3 that
BΛ = Pϕ (MΛ )ϕ((A+ PT+ + A− PT− )π ). (11.18)
If the operator Barc is invertible on the space Pϕ (Ṁarc )Hϕ , then the operators
(A± )Λ := Pϕ (MΛ )ϕ((A± )π ) are invertible on the space Pϕ (MΛ )Hϕ . Indeed, if the
operator Barc is invertible on the Hilbert space Pϕ (Ṁarc )Hϕ , then, by Theorem 8.1,
the operators π(ξ,x) (Barc ) for all (ξ, x) ∈ Narc are invertible on the Hilbert space
l2 (G, C2 ) and condition (8.8) is fulfilled. In particular, the operators π(ξ,±∞) (Barc )
are invertible on the space l2 (G, C2 ) for all (ξ, ±∞) ∈ Rarc , where Rarc is defined
by (10.4). Then it follows from (11.6) that all the operators A± (ξ,μ) for (ξ, μ) ∈ Rarc
2
are invertible on the space l (G) and
$ −1 $
sup $ A± (ξ,μ)
$ < ∞.
(ξ,μ)∈Rarc

Hence, by Theorem 10.1, the functional operators A± are invertible on the space
L2 (T), which implies the invertibility of both the operators (A± )Λ on the Hilbert
space Pϕ (MΛ )Hϕ .
Corollary 11.4 implies that for every functional operator A ∈ A,
Pϕ (MΛ )ϕ(Aπ )Pϕ (MΛ )ϕ((PT± )π ) = Pϕ (MΛ )ϕ((PT± )π )Pϕ (MΛ )ϕ(Aπ ).
Therefore, taking into account the invertibility of the operators (A± )Λ and the
equalities
Pϕ (MΛ )ϕ((PT+ )π )Pϕ (MΛ )ϕ((PT− )π ) = Pϕ (MΛ )ϕ((PT+ PT− )π ) = 0,
we conclude from (11.18) that the operator
(BΛ )−1 := ((A+ )Λ )−1 Pϕ (MΛ )ϕ((PT+ )π ) + ((A− )Λ )−1 Pϕ (MΛ )ϕ((PT− )π ),
where ((A± )Λ )−1 are the inverses for the operators (A± )Λ , is the inverse of the
operator BΛ on the space Pϕ (MΛ )Hϕ . 
Finally, to complete the proof of Theorem 3.1, which gives the Fredholm
criterion for the operators B in the C ∗ -algebra B, it remains to combine Theo-
rems 6.3, 8.1, 9.6 and 11.5. This also gives the Fredholm symbol calculus for the
C ∗ -algebra B (see (3.7)).
62 M.A. Bastos, C.A. Fernandes and Yu.I. Karlovich

Acknowledgment
We are grateful to the referee for the useful comments and suggestions.

References
[1] A. Antonevich, Linear Functional Equations. Operator Approach. Operator Theory:
Advances and Applications, vol. 83. Birkhäuser, Basel, 1996; Russian original: Uni-
versity Press, Minsk, 1988.
[2] A. Antonevich, M. Belousov, and A. Lebedev, Functional Differential Equations:
II. C ∗ -Applications. Part 2 Equations with Discontinuous Coefficients and Boundary
Value Problems. Pitman Monogr. Surveys Pure Appl. Math., vol. 95, Longman,
Harlow, 1998.
[3] A. Antonevich and A. Lebedev, Functional Differential Equations: I. C ∗ -Theory.
Pitman Monogr. Surveys Pure Appl. Math., vol. 70. Longman, Harlow, 1994.
[4] M.A. Bastos, C.A. Fernandes, and Yu.I. Karlovich, C ∗ -algebras of integral operators
with piecewise slowly oscillating coefficients and shifts acting freely. Integr. Equ.
Oper. Theory 55 (2006) 19–67.
[5] M.A. Bastos, C.A. Fernandes, and Yu.I. Karlovich, Spectral measures in C ∗ -algebras
of singular integral operators with shifts. J. Funct. Anal. 242 (2007) 86–126.
[6] M.A. Bastos, C.A. Fernandes, and Yu.I. Karlovich, C ∗ -algebras of singular integral
operators with shifts having the same nonempty set of fixed points. Compl. Anal.
Oper. Theory 2 (2008) 241–272.
[7] M.A. Bastos, C.A. Fernandes, and Yu.I. Karlovich, A nonlocal C ∗ -algebra of singular
integral operators with shifts having periodic points. Integr. Equ. Oper. Theory 71
(2011) 509–534.
[8] M.A. Bastos, C.A. Fernandes, and Yu.I. Karlovich, A C ∗ -algebra of singular integral
operators with shifts admitting distinct fixed points. J. Math. Anal. Appl. 432 (2014)
502–524.
[9] M.A. Bastos, C.A. Fernandes, and Yu.I. Karlovich, A C ∗ -algebras of singular inte-
gral operators with shifts similar to affine mappings. In “Operator Theory, Operator
Algebras and Applications”, eds. M.A. Bastos et al. Operator Theory: Advances and
Applications, vol. 242, 2014, pp. 53–79.
[10] L.A. Beklaryan, Groups of homeomorphisms of the line and the circle. Topological
characteristics and metric invariants. Russian Math. Surveys 59 (2004) 599–660.
[11] A. Böttcher, Yu.I. Karlovich, and B. Silbermann, Singular integral equations with
PQC coefficients and freely transformed argument. Math. Nachr. 166 (1994) 113–
133.
[12] A. Böttcher, Yu.I. Karlovich, and I.M. Spitkovsky, The C ∗ -algebra of singular integral
operators with semi-almost periodic coefficients. J. Funct. Anal. 204 (2003) 445–484.
[13] A. Böttcher, Yu.I. Karlovich, and I.M. Spitkovsky, Convolution Operators and Fac-
torization of Almost Periodic Matrix Functions. Operator Theory: Advances and
Applications, vol. 131. Birkhäuser, Basel, 2002.
[14] A. Böttcher, S. Roch, B. Silbermann, and I.M. Spitkovsky, A Gohberg–Krupnik–
Sarason symbol calculus for algebras of Toeplitz, Hankel, Cauchy, and Carleman
 A C ∗ -algebra of Singular Integral Operators with Shifts 63

operators. In “Topics in Operator Theory. Ernst D. Hellinger Memorial Volume”,

eds. L. de Branges et al. Operator Theory: Advances and Applications, vol. 48,
1990, pp. 189–234.
[15] J. Dixmier, C ∗ -Algebras. North-Holland, Amsterdam, 1977.
[16] I. Gohberg and N. Krupnik, On the algebra generated by the one-dimensional singular
integral operators with piecewise continuous coefficients. Funct. Anal. Appl. 4 (1970)
193–201.
[17] F.P. Greenleaf, Invariant Means on Topological Groups and Their Representations.
Van Nostrand-Reinhold, New York, 1969.
[18] Yu.I. Karlovich, The local-trajectory method of studying invertibility in C ∗ -algebras
of operators with discrete groups of shifts. Soviet. Math. Dokl. 37 (1988) 407–411.
[19] Yu.I. Karlovich, A local-trajectory method and isomorphism theorems for nonlocal
C ∗ -algebras. In “Modern Operator Theory and Applications. The Igor Borisovich
Simonenko Anniversary Volume”, eds. Ja.M. Erusalimsky et al. Operator Theory:
Advances and Applications, vol. 170, 2007, pp. 137–166.
[20] Yu.I. Karlovich and V.G. Kravchenko, An algebra of singular integral operators with
piecewise-continuous coefficients and piecewise-smooth shift on a composite contour.
Math. USSR-Izv. 23 (1984) 307–352.
[21] Yu.I. Karlovich and B. Silbermann, Fredholmness of singular integral operators with
discrete subexponential groups of shifts on Lebesgue spaces. Math. Nachr. 272 (2004)
55–94.
[22] P.S. Muhly and J. Xia, Calderón–Zygmund operators, local mean oscillation and
certain automorphisms of the Toeplitz algebra. Amer. J. Math. 117 (1995) 1157–
1201.
[23] M.A. Naimark, Normed Algebras. Wolters-Noordhoff Publishing, Groningen, 1972.
[24] S.P. Novikov, Topology of foliations. Trans. Mosc. Math. Soc. 14 (1965) 268–304.
[25] D. Sarason, Functions of vanishing mean oscillation. Trans. Amer. Math. Soc. 207
(1975) 391–405.
[26] D. Sarason, Toeplitz operators with piecewise qasicontinuous symbols. Indiana Univ.
Math. J. 26 (1977) 817–838.

M. Amélia Bastos
Departamento de Matemática
Instituto Superior Técnico
Universidad de Lisboa
Av. Rovisco Pais
1049–001 Lisboa, Portugal
e-mail: [email protected]
64 M.A. Bastos, C.A. Fernandes and Yu.I. Karlovich

Cláudio A. Fernandes
Centro de Matemática e Aplicações
Departamento de Matemática
Faculdade de Ciências e Tecnologia
Universidade Nova de Lisboa
Quinta da Torre
2829–516 Caparica, Portugal
e-mail: [email protected]
Yuri I. Karlovich
Centro de Investigación en Ciencias
Instituto de Investigación en Ciencias Básicas y Aplicadas
Universidad Autónoma del Estado de Morelos
Av. Universidad 1001, Col. Chamilpa
C.P. 62209 Cuernavaca, Morelos, México
e-mail: [email protected]
Operator Theory:
Advances and Applications, Vol. 267, 65–78

c Springer International Publishing AG, part of Springer Nature 2018

Non-Hermitian Quantum Mechanics

of Bosonic Operators
Natália Bebiano, João da Providência and J.P. da Providência

Abstract. The spectral analysis of a family of non-Hermitian operators ap-

pearing in quantum physics is our main concern. The properties of such oper-
ators are essentially different from those of Hermitian Hamiltonians, namely
due to spectral instabilities. We demonstrate that the considered operators
and their adjoints can be diagonalized when expressed in terms of certain con-
veniently constructed operators. We show that their eigenfunctions constitute
complete systems, but do not form Riesz bases. Attempts to overcome this
difficulty in the quantum mechanical set up are pointed out.
Mathematics Subject Classification (2010). 47A10, 81Q12, 34L10, 35A27,
47B44.
Keywords. Bosonic operator, non-Hermitian Hamiltonian, spectral analysis,
Riesz basis, metric.

1. Introduction
The main motivation for this article is the following. In conventional formula-
tions of non-relativistic quantum mechanics, the Hamiltonian operator is Hermit-
ian (synonymously, self-adjoint), and so it has real eigenvalues and an orthonormal
set of eigenfunctions. These fundamental issues are in the heart of von Neumann
quantum paradigm for physical observability and dynamical evolution. Certain
relativistic extensions of quantum mechanics lead to non-Hermitian Hamiltonian
operators, H = H ∗ , for H ∗ the adjoint of H (e.g., see [5, Chapter VIII]). Extending
the set of allowed operators by including PT -symmetric ones, P being the reflec-
tion operator, Pf (x) = f (−x) and T the time reflection operator, T f (x) = f (x),

This work was partially supported by the Centro de Matemática da Universidade de Coimbra
(CMUC), funded by the European Regional Development Fund through the program COMPETE
and by the Portuguese Government through the FCT – Fundação para a Ciência e a Tecnologia
under the project PEst-C/MAT/UI0324/2011.
66 N. Bebiano, J. da Providência and J.P. da Providência

yields spectra with reflection symmetry with respect to the real axis. In this con-
text, non-Hermitian Hamiltonians with a purely discrete spectrum have been the
object of intense research activity [1, 3, 4, 8, 11, 12, 14, 15], attempting to build
quantum mechanical theories with physical observables described by these oper-
ators. The subtleties of non self-adjoint Hamiltonians deserved the attention of
physicists and mathematicians, as they may originate new and unexpected phe-
nomena, and the non self-adjoint theory has revealed difficulties and challenging
problems.
We focus on the spectral analysis of non-Hermitian operators, a field with
applications in many areas of physics, most remarkably in quantum mechanics.
We summarise theoretical classical and recent developments on this topic, and the
theory is illustrated with a concrete example. We shall be particularly concerned
with L2 (R2 ), the Hilbert space of square-integrable complex valued functions in
two real variables x, y, equipped with the inner product
! +∞ ! +∞
Φ, Ψ = Φ(x, y)Ψ(x, y)dxdy,
−∞ −∞

and corresponding norm  · . The main topic of the present paper is the study of
the spectral properties of the operator in L2 (R2 ),
   
1 ∂2 ∂2 ∂ ∂
H := − + 2 +γ y +x + (x2 + y 2 ), γ ∈ R, (1)
4 ∂x2 ∂y ∂x ∂y
which is obviously non self-adjoint for γ = 0. We explicitly determine the eigen-
functions and eigenvalues of H, which can be rewritten as a quadratic function of
two interacting bosonic operators and their adjoints.
Let H be an infinite-dimensional separable Hilbert space endowed with the
inner product ·, ·, and corresponding norm  · . For H a non-Hermitian operator
in H with a purely discrete spectrum, the strategy of finding a Hermitian operator
H0 and an invertible operator P such that
H = P H0 P −1 , (2)
has been exploited. This idea refers that such a non-Hermitian H can be viewed
essentially as an alternative representation of a Hermitian operator with the same
eigenvalues and multiplicities, whenever P, P −1 are both bounded (see [10]).
The relation (2) is closely related to the quasi-Hermicity of H
QH = H ∗ Q, (3)
−1 ∗ −1
where Q = (P ) P is a positive definite operator called a metric operator.
(The relation (3) means that QH and H ∗ Q have the same domain and act in the
same manner.) In fact, H with property (2), is formally Hermitian with respect
to the modified inner product Q ·, ·,
QHx, y = x, H ∗ Qy = x, QHy, x, y ∈ H.
The metric is said to be non singular if it is bounded, invertible and boundedly
invertible, otherwise it is singular. The pathologies of non-Hermitian operators
 Non-Hermitian Quantum Mechanics of Bosonic Operators 67

with singular metric may be serious, since the metric can transform a basis into
a set without any reasonable basicity properties. Moreover, the spectral stability
with respect to small perturbations is not ensured and complex eigenvalues can
appear distant to the original ones [13].
The rest of this note is organized as follows. In Section 2 we recall some
issues used subsequently. In Section 3 we investigate spectral properties of H in
(1). It is shown that the obtained eigenfunctions are complete systems but do not
form Riesz bases, and so the validity of useful properties of Hermitian operators is
not guaranteed. In fact, non-Hermitian operators behave in an essentially different
way, due in particular to spectral instabilities and to the lack of basicity of the
Hamiltonian eigenfunctions. In Section 5 we discuss incidences of non-Hermiticity
in quantum physics.

2. Preliminaries
Many important operators in quantum physics are unbounded, which restricts
their domains of definition to adequate subsets of the Hilbert space where they
live. For instance, these domains are considered to be dense.
The set of eigenfunctions of an operator H, {Ψn }∞
n=1 , forms a complete family
in H if the span of Ψn is dense in this space, or, equivalently, if the orthogonal
complement of this linear span in the Hilbert space reduces to the zero function.
Eigenfunctions of non-Hermitian operators are in general not orthogonal or
even do not form a complete family. It should be stressed that the completeness
of an orthogonal set {Ψn }∞n=1 , adequately normalized, does not imply, by its own,
basicity. An orthonormal set {Ψn }∞ n=1 does not guarantee that any ψ ∈ H admits
a unique expansion of the form
∞

ψ= cn Ψ n ,
n=1

where, in this event, it should be cn = ψ, Ψn , n ≥ 1.

The concept of eigenbasis is very important in quantum mechanics. In con-
trast to the case of eigenfunctions of Hermitian operators, which are basis, for
non-Hermitian operators this property fails in general, and so the notion of Riesz
basis may be of interest. We say that {Ψn }∞ n=1 is a Riesz basis if for any ψ ∈ H,
there exists a positive constant C independent of ψ such that
∞

C −1 ψ2 ≤ |ψ, Ψn |2 ≤ Cψ2 .
n=1

The eigenfunctions of an operator H with purely discrete spectrum constitute

a Riesz basis if and only if H is quasi-Hermitian via a nonsingular bounded metric
Q (see [2]). The eigenfunctions of a non-Hermitian H, despite possibly being a
complete family, may not form a Riesz basis, as occurs frequently with several
68 N. Bebiano, J. da Providência and J.P. da Providência

models [10]. It should be noticed that Riesz basicity, like the spectrum, is not
preserved by unbounded metrics.
We introduce, as a first example of familiar operators in quantum mechanics,
the multiplication operators x and y
f (x, y) → xf (x, y), f (x, y) → yf (x, y),
defined in their maximal domains D(x) = {ψ ∈ L2 (R2 ) : xψ ∈ L2 (R2 )} and
D(y) = {ψ ∈ L2 (R2 ) : yψ ∈ L2 (R2 )}.
We consider the partial differential operators ∂/∂x and ∂/∂y defined by
∂f (x, y) ∂f (x, y)
f (x, y) → , f (x, y) → ,
∂x ∂y
with domains D(∂/∂x) = D(∂/∂y) = W 1,2 (R2 ), where W 1,2 (R2 ) stands for the
usual Lebesgue space of functions in L2 (R2 ), whose first partial derivatives be-
long to L2 (R2 ). These operators are known as momentum operators. The partial
differential operators ∂ 2 /∂x2 and ∂ 2 /∂y 2 are defined by
∂ 2 f (x, y) ∂ 2 f (x, y)
f (x, y) → , f (x, y) → ,
∂x2 ∂y 2
with domains D(∂ 2 /∂x2 ) = D(∂ 2 /∂y 2 ) = W 2,2 (R2 ), the Lebesgue space of func-
tions in L2 (R2 ) whose first and second derivatives belong to L2 (R2 ).
The (standard) bosonic operators
1 ∂ 1 ∂
a := x + , b := y + ,
2 ∂x 2 ∂y
with domains
D(a) = D(b) = {Ψ ∈ W 2,2 (R2 ) : xΨ, yΨ ∈ L2 (R2 )}
are useful in our discussion. The operators a and b are known to be densely defined,
closed, and their adjoints read
1 ∂ 1 ∂
a∗ = x − , b∗ = y − .
2 ∂x 2 ∂y
We recall that, conventionally, a, b are said to be annihilation operators, while a∗ , b∗
are creation operators. It is worth noticing that these operators are unbounded,
and they satisfy the commutation rules (CR’s),
[a, a∗ ] = [b, b∗ ] = 1, (4)
∗ ∗
where 1 is the identity operator on L (R ). (This means that aa f − a af =
2 2

bb∗ f − b∗ bf = f for any f in D(a) and D(b)). Furthermore,

[a, b∗ ] = [b, a∗ ] = [a∗ , b∗ ] = [a, b] = 0. (5)
As it is well known, the canonical commutation relations (4) and (5) characterize
an algebra of Weil–Heisenberg (W-H). Moreover, the following holds,
aΦ0 = bΦ0 = 0,
 Non-Hermitian Quantum Mechanics of Bosonic Operators 69

2
+y 2 )
for Φ0 = e−(x in D(a) and D(b), a so-called vacuum state. The set of functions
{Φm,n = a∗m b∗n Φ0 : m, n ≥ 0},
constitutes a basis of H, that is, every vector in L2 (R2 ) can be uniquely expressed
in terms of this system, which is complete, since 0 is the only vector orthogonal to
all its elements.

3. Non self-adjoint Hamiltonian H describing two interacting

bosonic operators
3.1. Model
The closedness of the operator H in (1) is the essential starting point for the study
of its spectrum. In the perspective of a convenient domain of H, aiming at defining
D(H), we regard the cofactor of γ in the non-self-adjoint term,
 
∂ ∂
V = y +x ,
∂x ∂y
as a perturbation of the 2D harmonic oscillator,
 
1 ∂2 ∂2
(H) = − + + (x2 + y 2 ).
4 ∂x2 ∂y 2
It is easily seen that this term is relatively bounded with the relative bound less
than one with respect to (H), provided that γ is less than one. Under this
condition, H is a closed operator on the domain of the harmonic oscillator:
W 2,2 (R2 ) ∩ L2 (R2 , (x4 + y 4 )dxdy).
Here, we have used the standard perturbation result that states the following. If
(H) is closed and V is relatively bounded with respect to (H), with the relative
bound smaller than 1, then (H) + λV , with λ < 1, is closed [9, Theorem 3.3].
We notice that D(H) contains the subspace S(R2 ) constituted by functions
f (x, y) such that e−2γxy f (x, y) ∈ S(R2 ) and, in turn, this one contains C0∞ (R2 ).
In terms of creation and annihilation operators, H is equivalently defined as
H = a∗ a + bb∗ + γ(a∗ b∗ − ab), γ ∈ R, (6)
and so H is quadratic in the bosonic operators a, b and in their adjoints.
3.2. Spectrum
As it is well known, the resolvent set of H, denoted by ρ(H), is constituted by
all the complex numbers for which the resolvent operator λ ∈ ρ(H) → (H − λ)−1
exists as a bounded operator on H. The complement
σ(H) = C\ρ(H)
is called the spectrum of H. The set of all eigenvalues of H is called the point
spectrum, denoted by σp (H), and formed by complex numbers λ for which H − λ :
D(H) → H is not injective. The spectrum of an operator on a finite-dimensional
70 N. Bebiano, J. da Providência and J.P. da Providência

Hilbert space is exhausted by the eigenvalues, but, in the infinite-dimensional set-

ting, there are additional parts to be considered. Those λ which are not eigenvalues
but H − λ is not bijective constitute the continuous or residual spectrum, depend-
ing on the range Ran(H − λ) being, respectively, dense or not. The spectrum σ(H)
is the union of these three disjoint spectra. The spectrum of self-adjoint operators
is nonempty, real, and the residual spectrum is empty, while the spectrum of non
self-adjoint operators can be empty or coincide with the whole complex plane (see,
e.g., refs. [13, 15]).

4. Accretivity
It may be advantageous to estimate the spectrum in terms of the numerical range:
W (H) := {Hψ, ψ : ψ ∈ D(H), ψ = 1}.
In general W (H) is neither open nor closed, even when H is a closed operator.
However, it is always convex and for H bounded,
σ(H) ⊂ W (H).
Proposition 4.1. The numerical range of H is bounded by the hyperbola
y 2 + γ 2 (1 − x2 ) = 0, x ≥ 1.
Proof. The numerical range of H is determined as follows. Recalling that W (H) is
convex, let us consider the supporting line of W (H) perpendicular to the direction
θ. Recall that the distance of this line to the origin is the lowest eigenvalue of
(e−iθ H) = (a∗ a + b∗ b) cos θ − iγ(a∗ b∗ − ab) sin θ,
provided this operator is bounded from below, which occurs for −π/2 ≤ θ < π/2.
The eigenvalues of (e−iθ H) are readily determined by the EMM [5], and they
are found to be
Eθ (1 + n + m), n, m ≥ 0,
where +
1 − γ 2 + cos(2θ) + γ 2 cos(2θ)
Eθ = √ .
2
Thus, the supporting line under consideration is given by
x cos θ + y sin θ = Eθ . (7)
As it is well known, the equation of the boundary of W (H) is found eliminating θ
between (7) and
−x sin θ + y cos θ = Eθ ,
being easily obtained as the branch of the hyperbola
+
y = ±γ x2 − 1, x ≥ 1. 
 Non-Hermitian Quantum Mechanics of Bosonic Operators 71

Sectorial operators are defined by the property that their numerical range is
the subset of a sector
Sw,θ = {z ∈ C : | arg (z − w)| ≤ θ}
with w ∈ R and 0 ≤ θ < π/2, called, respectively, the vertex and semi-angle of H.
An operator is said to be accretive if the vertex can be chosen at the origin, i.e.,
W (H) ⊂ S0,π/2 . An operator H is m-accretive if its numerical range is contained
in the right closed half-plane and the resolvent bound for any λ with λ < 0, holds:
∀λ ∈ C, λ < 0, (H − λ)−1  ≤ 1/|λ|.
Proposition 4.2. The operator H is m-accretive.
Proof. We show that for any z ∈ C, with z < 0, the resolvent bound holds. We
have
dist(z, W (H)) ≤ |Hψ, ψ − z| = |(H − z)ψ, ψ| ≤ (H − z)ψ.
As dist(z, W (H)) ≥ |z|, the result follows, having in mind Proposition 4.1. 
Any m-accretive operator is closed and densely defined. In fact, by [2, Propo-
sition 5.2.1, p. 246], if H is not closed, then σ(H) = C.
A closed operator H in H has a compact resolvent if ρ(H) = ∅ and (H − λ)−1 ,
for some λ ∈ ρ(H), is a compact operator.
It is known that [9, Theorem IX, 2.3], if H has a compact resolvent, then
σ(H) = σp (H). The operator H has a compact resolvent, as (H) is an m-accretive
operator (since (H) is Hermitian and W ((H)) lies on the positive real axis) with
compact resolvent and, moreover, V is relatively bounded with respect to (H)
with relative bound smaller than 1. Then, (H) + λV has a compact resolvent
[9, Theorem 5.4.1].
4.1. Eigenvectors and eigenvalues of H
To obtain the eigenvalues of H, we firstly consider the selfadjoint operator in
L2 (R2 ),  
1 ∂2 ∂2
H0 = − + + (1 + γ 2 )(x2 + y 2 ), γ < 1. (8)
4 ∂x2 ∂y 2
We know a priori that the eigenfunctions of H0 (after normalization) form an
orthonormal family in L2 (R2 ) and the corresponding eigenvalues are real. The
eigenvalues of H0 are easily determined as
+
Em,n = (1 + m + n) 1 + γ 2 , m, n ≥ 0,
and the associated eigenfunctions Φm,n are
√ √
2 2 2 2
Φm,n = KHm ((1 + γ 2 )1/4 x)e− 1+γ x × Hn ((1 + γ 2 )1/4 y)e− 1+γ y ,
where K = (2(1 + γ 2 )/π)1/2 and Hn (t) is the nth Hermite polynomial in t. That
is, Φm,n is factorized as follows
Φm,n (x, y) = Φm (x)Φn (y),
72 N. Bebiano, J. da Providência and J.P. da Providência

where Φn (t) are the usual eigenfunctions of the famous harmonic oscillator [2]. For
the sake of completeness, we show how to obtain Φm,n . Indeed, let us consider the
differential operators
1 ∂ 1 ∂
g= 2 1/4
+ (1 + γ 2 )1/4 x, h= 2 1/4
+ (1 + γ 2 )1/4 y,
2(1 + γ ) ∂x 2(1 + γ ) ∂y
1 ∂ 1 ∂
g∗ = − + (1 + γ 2 )1/4 x, h∗ = − + (1 + γ 2 )1/4 y,
2(1 + γ 2 )1/4 ∂x 2(1 + γ 2 )1/4 ∂y
which satisfy the Weil–Heisenberg commutation rules,
[g, g ∗ ] = [h, h∗ ] = 1, [g, h] = [g ∗ , h∗ ] = [g ∗ , h] = [g, h∗ ] = 0.
We easily find that +
H0 = 1 + γ 2 (g ∗ g + h∗ h + 1).
The groundstate eigenfunction of H0 , which is the vacuum of the operators g, h, is
√ 2 2 2
Φ0 (x, y) = κ0,0 e− 1+γ (x +y ) ,
being the remaining eigenfunctions,
Φm,n (x, y) = κm,n g ∗m h∗n Φ0 (x, y),
where κm,n are normalization factors.
The eigenfunctions Φm,n (x, y) constitute a complete set in L2 (R2 ), as can be
easily verified by adapting the standard proof of completeness of Hermite functions.
The functions are orthogonal, and the system forms a basis of L2 (R2 ). Details can
be found, for instance, in [5, 6].
To find the eigenvalues of H we notice that H is formally similar to H0
H0 = e−2γxy He2γxy .
The word “formally” refers to the fact that the operator P = e2γxy is unbounded.
Nevertheless, the similarity relation is well defined on the eigenfunctions of H0 .
For Φ(x, y) an arbitrary differentiable function in the domain of H0 , we have
     
1 ∂2 ∂2 ∂ ∂
− + 2 +γ y +x + (x + y ) e2γxy Φ(x, y)
2 2
4 ∂x2 ∂y ∂x ∂y
   
1 ∂2 ∂2
=e 2γxy
− 2 2 2
+ 2 + (1 + γ )(x + y ) Φ(x, y),
4 ∂x2 ∂y
that is
He2γxy Φ = e2γxy H0 Φ,
and the operator equality
He2γxy = e2γxy H0
holds.
Since
H0 Φm,n = Em,n Φm,n ,
we easily get
He2γxy Φm,n = Em,n e2γxy Φm,n .
 Non-Hermitian Quantum Mechanics of Bosonic Operators 73

The eigenfunctions of H are expressed as

Ψm,n = e2γxy Φm,n ,
and the eigenvalues of H are those of H0 . Thus, the eigenvalues of H are real
positive and Ψm,n ∈ D(H), as it should be.

4.2. Eigenfunctions and eigenvalues of H ∗

Following the procedure in Subsection 4.1, it can be shown that, formally, we may
write,
H0 = e2γxy H ∗ e−2γxy ,
implying that eigenfunctions and associated eigenvalues of H ∗ are given respec-
tively by
+
Ψ̃m,n = e−2γxy Φm,n ∈ D(H ∗ ), Em,n = (m + n + 1) 1 + γ 2 , m, n ≥ 0.

4.3. Biorthogonality of the eigenfunctions

It is straightforward to see that the set {Ψm,n : m, n ≥ 0} is not constituted by
orthogonal functions. However, the vector system {Ψ̃m,n : m, n ≥ 0}, formed by the
eigenfunctions of H ∗ , is biorthogonal to the eigensystem of H, {Ψm,n : m, n ≥ 0}.
Indeed, we have

Ψm,n , Ψ̃p,q  = e2γxy Φm,n , e−2γxy Φp,q  = Φm,n , Φp,q  = m!n!δmp δnq ,
where δij = 1 for i = j, otherwise δij = 0, represents the Kronecker symbol.

4.4. Completeness of the eigenfunctions

The completeness of both eigensystems of H and H ∗ can be shown to hold. The
operator is m-accretive as its √
numerical range lies a hyperbolical region limited by
the branch of hyperbola y = x2 − 1, x ≥ 1. Moreover, it can be easily seen that
the imaginary part of the resolvent of −iH at δ < 0, is non-negative Hermitian
1
(−iH − δ)−1 − (iH ∗ − δ)−1 ≥ 0.
2i
As the resolvent is a trace class function, by the completeness theorem [7, Theorem
VII.8.1], we may conclude that the eigenfunctions of H form a complete system.
Analogous arguments are valid for the eigensystem of H ∗ .
We observe that completeness does not imply that any ψ ∈ L2 (R2 ) has a
unique expansion
∞
ψ= cm,n Ψm,n ,
m,n=0

a fundamental issue in quantum mechanics.

74 N. Bebiano, J. da Providência and J.P. da Providência

4.5. Asymptotic behavior of the eigenfunctions

We wish to discuss the asymptotic behavior of the Hamiltonian eigenfunctions. To
this end, it is convenient to introduce the Planck constant , explicitly. We also
change the notation slightly. We replace the notation x, y, used for the particle
coordinates, by x1 , x2 , and we denote the respective momenta by
∂ ∂
p1 = −i , p2 = −i .
∂x1 ∂x2
For simplicity, we also consider γ = 1, so that (6) becomes
1 2
H= (p + p22 ) + x21 + x22 + i(x1 p2 + x2 p1 ) − 1.
4 1
Next, we introduce the change of variables, which constitutes a canonical trans-
formation,
1 1
X= (x1 + x2 ), x = x1 − x2 , P = p1 + p2 , p = (p1 − p2 ).
2 2
In terms of the new variables, the Hamiltonian is the sum of two summands, each
one involving only one type of variable, and becomes
1 2 1 1
H= P + 2X 2 + iXP + p2 + x2 + ixp − 1.
8 2 2
We observe that it is equivalent to consider m, n → ∞ or  → 0. We firstly
concentrate on the summand
1
HX = P 2 + 2X 2 + iXP.
8
In order to characterize the asymptotic behavior of the eigenfunctions we use the
well-known Wentzel–Kramers–Brillouin [WKB] approximation [5, Chapter III] in
leading order. Indeed, we express the eigenfunction of the energy operator as
Ψ(X) = eiS(X)/ .
The function S(X), in leading order, is determined by the Jacobi equation
 2
1 dS dS
+ 2X 2 + iX = E,
8 dX dX
where E denotes the associated eigenvalue of the energy operator. Thus, we readily
obtain,
dS +
= 2 2E − 8X 2 − 4iX,
dX
and
+ E 2X
S = X 2E − 8X 2 + √ arctan √ − 2iX 2.
2 E − 4X 2
Similarly, the eigenfunction of H ∗ is
Ψ̃(X) = eiS̃(X)/ ,
 Non-Hermitian Quantum Mechanics of Bosonic Operators 75

with
+ E 2X
S̃ = X 2E − 8X 2 + √ arctan √ + 2iX 2 .
2 E − 4X 2

Therefore, the integral

! √E/2
√ ΨΨ dX
− E/2

approaches +∞ as  → 0, while the integral

! √E/2
√ Ψ̃Ψ̃ dX
− E/2

approaches 0 as  → 0.
On the other hand, the integral
! √E/2
√ ΨΨ̃ dX
− E/2

remains finite as  → 0.
The summand Hx = 12 p2 + 12 x2 + ixp may be similarly treated.
As a consequence, we get
lim Ψm,n  = ∞.
m,n→∞

Thus, the system of eigenvectors of {Ψm,n } does not form a Riesz basis for L2 (R2 ).
Analogous conclusion holds for {Ψ̃m,n }.

4.6. Metric operator

The existence of a positive definite Q satisfying (3) is equivalent to the fact that
the resolvent of H satisfies
Q(H − z0 )−1 = (H ∗ − z0 )−1 Q
for z0 ∈ ρ(H) ∩ ρ(H ∗ ) ∩ R2 .
As a consequence of the reality of the discrete spectrum of H and of the
completeness of the corresponding eigenfunctions, it can be shown, using the pro-
cedure in [13], that there exists a bounded metric for H. The proof relies on the
fact that the existence of bounded metric for an unbounded H can be transferred
to the same problem for its bounded resolvent.
Nevertheless, there does not exist a non singular metric operator ensuring
quasi-Hermiticity, because the eigenfunctions of H do not form a Riesz basis.
As a consequence, in the similarity of H to a self-adjoint operator, the basicity
properties of these operators may be very different (see [1] and [2]). However,
despite these negative features, in the subspace S = span{Ψm,n }, a new inner
product may be meaningfully defined with the help of a metric operator (non-
singular in S), with respect to which the restriction of H to S, say h, should
represent H.
76 N. Bebiano, J. da Providência and J.P. da Providência

4.7. Physical Hilbert space

In order to specify the physical Hilbert space associated with the Hamiltonian (1),
we chose its domain to be the function space
Dom(H) = {Ψ : e−2γxy Ψ ∈ W 2,2 (R2 ) ∩ L2 (R2 , (x4 + y 4 )dxdy)},
endowed with the inner product,
! ∞ ! ∞
Ξ1 , Ξ2  = dx dye−4γxy Ξ1 (x, y)Ξ2 (x, y), Ξ1 , Ξ2 ∈ Dom(H),
−∞ −∞

which involves the weight function e−4γxy . Since H0 in (8) is a closed operator on
the domain of the harmonic oscillator
Dom(H0 ) = W 2,2 (R2 ) ∩ L2 (R2 , (x4 + y 4 )dxdy),
and
He2γxy Φ = e2γxy H0 Φ, ∀Φ ∈ Dom(H0 ),
while
e−2γxy HΨ = H0 e−2γxy Ψ, ∀Ψ ∈ Dom(H),
it follows that H is a closed operator on Dom(H).
Let
S = span{Ψm,n }, S̃ = span{Ψ̃m,n},
and D = span{Φm,n }.
In terms of the linear operators e2γxy : D → S and e−2γxy : D → S̃ we may
write
Ψm,n = e2γxy Φm,n , Ψ̃m,n = e−2γxy Φm,n . (9)
From (9) it follows that
S = eS (D), S̃ = e−S (D).
Since
e−S Ψm,n , e−S Ψp,q  = δmp δnq , S = 2γxy,
it follows that the eigenfunctions Ψm,n are orthogonal for this inner product,
Ψm,n , Ψp,q  = δmp δnq .
Next we show that the linear space S, equipped with the inner product ·, ·,
allows the probabilistic interpretation of quantum mechanics. The symmetry is ob-
viously satisfied. It may be pointed out that, from the point of view of physics, only
the action of the metric operator exp(−2S) on S is significant and, in this event,
it is unambiguously defined. We verify that exp(−2S) is a metric. For φ, ψ ∈ S,
exp(−2S)φ, ψ is finite, and so the operator exp(−2S) is bounded in S. (Similarly,
exp(2S) is bounded in S̃.) Moreover,
ψ, ψ, ψ ∈ S,
is nonnegative. (We notice that H leaves S invariant.)
 Non-Hermitian Quantum Mechanics of Bosonic Operators 77

The vectors Ψm,n are orthogonal with respect to the new inner product and
any Ψ ∈ S may be expanded as a unique finite linear combination of Ψn,m ,

Ψ= cnm Ψn,m , with Ψ, Ψn,m  = cnm .
n,m

In order to ensure the probabilistic interpretation of quantum mechanics, we im-

pose the normalization Ψ, Ψ = 1, so that

Ψ, Ψ = |cnm |2 = 1.
n,m

The probability amplitude is given by cnm and satisfies n,m |cnm |2 = 1.

5. Conclusions
We have initially considered a non self-adjoint Hamiltonian H whose eigenvalues
and corresponding eigenfunctions have been explicitly determined. The investi-
gated Hamiltonian and its adjoint have real eigenvalues and systems of biorthogo-
nal eigenvectors. They have infinite diagonal matrix representations in the respec-
tive eigensystems, which are complete. Nevertheless, they do not form Riesz bases.
Viewing H as the Hamiltonian of a physical model, problems arise from
non Hermiticity. The original inner product defined in H is not adequate for the
physical interpretation of the model. A new Q-metric, which is appropriate for that
purpose, may be introduced (see Subsection 4.7). Following Mostafazadeh [11], one
can define a subspace of the Hilbert space, and the restriction of the Hamiltonian
operator to that subspace, so that it has the same spectrum and eigenfunctions as
the original one. The referred subspace remains invariant under the action of H.
Remarkably, stating that this Hermitian operator represents in a reasonable sense
the non-Hermitian operator may be controversial, since relevant information on
the Hamiltonian may not be captured in the mentioned subspace.
Non-Hermitian operators have typically non-trivial pseudospectra. It is
known that the relation (2) holds via a bounded and boundedly invertible positive
transformation if and only if (3) holds with a positive bounded and boundedly
invertible metric [10]. Further, if (3) holds with a positive bounded and boundedly
invertible metric, then the pseudospectrum of H is trivial. The concept of pseu-
dospectrum is of great relevance for the description of non-Hermitian operators
in the context of quantum mechanics. A non-trivial pseudospectrum ensures the
non-existence of a bounded metric.

References
[1] F. Bagarello, Construction of pseudo-bosons systems. J. Math. Phys. 51 (2010),
023531.
[2] F. Bagarello, J.-P. Gazeau, F.H. Szafraniec, and M. Znojil (eds.), Non-Selfadjoint
Operators in Quantum Physics: Mathematical Aspects. John Wiley & Sons, Hobo-
ken, NJ, 2015.
78 N. Bebiano, J. da Providência and J.P. da Providência

[3] N. Bebiano, J. da Providência, and J.P. da Providência, Mathematical aspects

of quantum systems with a pseudo-Hermitian Hamiltonian. Brazilian Journal of
Physics 46 (2016), 152–156.
[4] C.M. Bender and S. Boettcher, Real spectra in non-Hermitian Hamiltonians having
PT symmetry. Phys. Rev. Lett. 80 (1998), 5243–5246.
[5] A.S. Davydov, Quantum Mechanics. Pergamon Press, 1965.
[6] P.A.M. Dirac, The Principles of Quantum Mechanics. 4th ed. Oxford University
Press, Oxford, Oxford, 1958.
[7] I. Gohberg, S. Goldberg, and M. Kaashoek, Classes of Linear Operators. Vol. I.
Birkäuser, Basel, 1990.
[8] A. Gonzaléz Lopéz and T. Tanaka, Nonlinear pseudo-supersymmetry in the frame-
work of N -fold supersymmetry. J. Phys. A: Math. Gen. 39 (2006), 3715–3723.
[9] D. Krejcirik and P. Siegl, Elements of spectral theory without the spectral theorem.
In: “Non-Selfadjoint Operators in Quantum Physics: Mathematical Aspects”. John
Wiley & Sons, Hoboken, NJ, 2015, pp. 241–291.
[10] D. Krejcirik, P. Siegl, M. Tater, and J. Viola, Pseudospectra in non-Hermitian
quantum mechanics. J. Math. Phys. 56 (2015), 103513.
[11] A. Mostafazadeh, Pseudo-Hermitian quantum mechanics with unbounded metric
operators. Phil. Trans. R. Soc. A 371 (2013), 20120050.
[12] J. da Providência, N. Bebiano, and J.P. da Providência, Non Hermitian operators
with real spectra in quantum mechanics. Brazilian Journal of Physics 41 (2011),
78–85.
[13] P. Siegl and D. Krejcirik, The metric operator for the imaginary cubic oscillator.
Phys. Rev. D 86 (2012), 11702(R).
[14] F.G. Scholtz, H.B. Geyer, and F.J.W. Hahne, Quasi-Hermitian operators in quan-
tum mechanics and the variational principle. Ann. Phys. NY 213 (1992), 74.
[15] M. Znojil, Should PT symmetric quantum mechanics be interpreted as nonlinear?
J. Nonlin. Math. Phys. 9 (2002), 122–123.

Natália Bebiano
CMUC, Departament of Mathematics
Universidade de Coimbra
3001-454 Coimbra, Portugal
e-mail: [email protected]
João da Providência
Departamento de Fı́sica
Universidade de Coimbra
3001-454 Coimbra, Portugal
e-mail: [email protected]
J.P. da Providência
Department of Physics
University of Beira. Interior
P-6201-001 Covilhã, Portugal
e-mail: [email protected]
Operator Theory:
Advances and Applications, Vol. 267, 79–122

c Springer International Publishing AG, part of Springer Nature 2018

Fredholm Conditions on Non-compact

Manifolds: Theory and Examples
Catarina Carvalho, Victor Nistor and Yu Qiao

Abstract. We give explicit Fredholm conditions for classes of pseudodiffer-

ential operators on suitable singular and non-compact spaces. In particular,
we include a “user’s guide” to Fredholm conditions on particular classes of
manifolds including asymptotically hyperbolic manifolds, asymptotically Eu-
clidean (or conic) manifolds, and manifolds with poly-cylindrical ends. The
reader interested in applications should be able to read right away the results
related to those examples, beginning with Section 5. Our general, theoretical
results are that an operator adapted to the geometry is Fredholm if, and only
if, it is elliptic and all its limit operators (in a sense to be made precise) are
invertible. Central to our theoretical results is the concept of a “Fredholm
groupoid.” By definition, a Fredholm groupoid is one for which this character-
ization of the Fredholm condition is valid. We use the notions of exhaustive
and strictly spectral families of representations to obtain a general charac-
terization of Fredholm groupoids. In particular, we introduce the class of the
so-called groupoids with Exel’s property as the groupoids for which the regular
representations are exhaustive. We show that the class of “stratified submer-
sion groupoids” has Exel’s property, where stratified submersion groupoids are
defined by gluing fibered pull-backs of bundles of Lie groups. We prove that
a stratified submersion groupoid is Fredholm whenever its isotropy groups
are amenable. Many groupoids, and hence many pseudodifferential operators
appearing in practice, fit into this framework. This fact is exploited to yield
Fredholm conditions not only in the above-mentioned classes, but also on man-
ifolds that are obtained by desingularization or by blow-up of singular sets.

Mathematics Subject Classification (2010). 58J40 (primary), 58H05, 47L80,

46L60.
Keywords. Pseudodifferential operator, differential operator, Fredholm oper-
ator, groupoid, Lie group, Sobolev space, C ∗ -algebras, compact operators,
non-compact manifold, singular manifold.

Carvalho was partially supported by Fundação para a Ciência e Tecnologia UID/MAT/

04721/2013 (Portugal). Nistor has been partially supported by ANR-14-CE25-0012-01. Qiao
was partially supported by NSF of China (11301317, 11571211).
Manuscripts available from http://iecl.univ-lorraine.fr/˜Victor.Nistor/.
80 C. Carvalho, V. Nistor and Y. Qiao

1. Introduction
We obtain in this paper necessary and sufficient conditions for classes of operators
to be Fredholm. Our results specialize to yield Fredholm conditions for classi-
cal pseudodifferential operators on manifolds with cylindrical and poly-cylindrical
ends, on manifolds that are asymptotically Euclidean, and on manifolds that are
asymptotically hyperbolic. Other examples of non-compact manifolds covered by
our results include natural operators on spaces obtained by desingularization of
suitable singular spaces by successively blowing up the lowest-dimensional singular
strata.

1.1. Background and main result

Let M0 be a Riemannian manifold and let P : H s (M0 ; E) → H s−m (M0 ; F ) be
an order m classical pseudodifferential operator acting between Sobolev sections
of two smooth, Hermitian vector bundles E, F . Recall that P is called elliptic
if its principal symbol σm (P ) is invertible outside the zero section. We have the
following well-known, classical result [19, 96, 97].
Theorem 1.1 (Classical). Assume that M0 is compact and smooth. We have that
P : H s (M0 ; E) → H s−m (M0 ; F ) is Fredholm if, and only if, it is elliptic.
This classical Fredholm result has many applications, so a natural question to
ask is to what extent it extends to (suitable) non-compact manifolds. The example
of constant coefficient differential operators on Rn shows that this classical result
might be no longer true as stated if M0 is not compact.
On certain classes of manifolds, however, it is possible to reformulate Theorem
1.1 as follows. Let M0 be a non-compact manifold with “amenable ends” (see
Subsection 4.2 for the precise definition). Then we can associate to M0 the following
data:
(1) Smooth manifolds Mα , α ∈ I, and
(2) Lie groups Gα acting freely and properly on Mα , α ∈ I,
(for a suitable index set I), satisfying the following theorem.
Theorem 1.2. Let P be an order m classical pseudodifferential operator on M0
compatible with the geometry. Then one can associate to P certain Gα -invariant
pseudodifferential operators Pα on Mα with the following property:
P : H s (M0 ; E) → H s−m (M0 ; F ) is Fredholm ⇔ P is elliptic and
Pα : H (Mα ; E) → H
s s−m
(Mα ; F ) is invertible for every α ∈ I .
We stress here the appearance of the additional operators Pα , whose invert-
ibility is necessary for the given operator to be Fredholm. These are “limit” or
“boundary” operators and are, in a certain sense, limits at infinity of our operator
P . This is consistent with the physical intuition behind the HVZ theorem for the N -
body hamiltonian [32] and the references therein. See also [22, 31, 32, 50, 56, 82, 89]
and the references therein.
 Fredholm Conditions on Non-compact Manifolds 81

Many results similar to Theorem 1.2 were obtained for psedudodifferential

operators before. See for instance [25, 34, 55, 60, 62, 94, 95] and the references
therein for a very small sample. Similar results for differential operators were also
obtained earlier, see [24, 45, 46, 71, 80] and the references therein, again for a
very small sample. Other related results appeared in the context of localization
principles, often in relation to the N -body problem mentioned above. Inspired by
these works, we shall call the operators Pα the limit operators of P . We shall refer
to results similar to Theorem 1.2 as non-local Fredholm conditions.
The class of “manifolds with amenable ends” is introduced such that, almost
by definition, it is close to being the largest class of manifolds for which Theorem
1.2 is valid. The challenge then becomes to provide a large enough class of mani-
folds with amenable ends, which we do in this paper. The term “amenable” comes
from the fact that certain isotropy groups at infinity are assumed to be amenable.
The manifolds Mα and the groups Gα are obtained from the orbits of certain nat-
ural vector fields acting on a compactification M of M0 and from their isotropies.
One can see this very clearly in the case of manifolds with cylindrical and poly-
cylindrical ends, of manifolds that are asymptotically Euclidean, and of manifolds
that are asymptotically hyperbolic, which are all particular cases of manifolds with
amenable ends. See the last section of the paper for explicit statements covering
in detail the case of these classes of manifolds. It is interesting to notice that the
case of asymptotically hyperbolic manifolds leads to the study of certain operators
on non-commutative solvable Lie groups.
Theorem 1.2 is essentially a consequence of the results in [48] or [49], but the
results of those papers turned out to be difficult to use directly by non-specialists.
Moreover, the approach in this paper is direct, and we do not use the results of
[48, 49], but rather rederive them in the needed form. In particular, we have tried
in the last section of the paper to provide a review of the applications such that
the reader can understand the main results of this paper on Fredholm operators
without any knowledge of groupoids or C ∗ -algebras.

1.2. Exhaustive families of representations

It was realized by some of the authors mentioned above that the non-local Fred-
holm conditions of Theorem 1.2 are related to the representation theory of certain
C ∗ -algebra A (a different algebra for each M0 ). See also [18, 19, 80, 101] and the
references therein. The C ∗ -algebra A is such that they contain the operators of the
form a = (1 + Δ)(s−m)/2 P (1 + Δ)−s/2 acting on L2 (M0 ) (where Δ is the positive
Laplacian and P is an order m pseudodifferential operator as in Theorem 1.2).
The reason for considering the operator a is that P is Fredholm if, and only if, a is
Fredholm; moreover, a is bounded (unlike P , in general). In most practical appli-
cations, the C ∗ -algebra A can be chosen to be a groupoid C ∗ -algebra. Properties
of the operator P can then be inferred from the structure of A, taking advantage of
the fact that one has a well-developed theory of groupoid C ∗ -algebras, see Section
2 for more details and references. In particular, these ideas are used in the proof
of Theorem 4.12, which contains Theorem 1.2 as a particular case. The operators
82 C. Carvalho, V. Nistor and Y. Qiao

Pα in Theorem 1.2 then can be obtained as homomorphic images of the operator

P , that is Pα = πα (P ), for suitable representations {πα } of the algebra A.
The relevant families of representations in this setting are those of strictly
spectral (and strictly norming) families of representations, introduced by Roch in
[88] and recently studied in detail by Nistor and Prudhon in [76]. In that paper, the
concept of an exhaustive family of representations has emerged as a useful concept
to study strictly spectral families of representations. From a technical point of
view, these families of representations form the backbone of the theoretical results
in this paper.

1.3. Fredholm Lie groupoids

The key point of our approach is to start with a general study of Fredholm
conditions for pseudodifferential operators in the framework of “Fredholm Lie
groupoids.” All pseudodifferential operators considered in this paper will be classi-
cal. To each Lie groupoid G with units a manifold with corners M , one can associate
a class Ψm (G; E, F ) of (classical) pseudodifferential operators on G acting between
sections of vector bundles E, F → M , whose closure contains the groupoid C ∗ -
algebra Cr∗ (G) as an essential ideal if m = 0 and E = F = C (see Section 4.2 and
the references therein). Let ∂M denote the union of the proper faces of M (the
boundary of M ) and M0 := M  ∂M be the interior of M . For some results, we
shall need that M be compact, so, for simplicity, we shall assume that this is the
case, unless otherwise stated. The smooth manifold M0 and its compactification
M have the same meaning as in Subsection 1.1. Often one chooses for the algebra
A considered in Subsection 1.2 the algebra A := Cr∗ (G), the “reduced C ∗ -algebra”
of G, whose definition we recall below. Alternatively, on can consider A := Ψ(G),
the norm closure of Ψ0 (G).
A Fredholm Lie groupoid G is, by definition, a Lie groupoid for which the
Fredholm property of the associated operators is equivalent to the invertibility of
the fiberwise boundary restrictions and of the principal symbol on M (the principal
symbol has to be suitably defined at the boundary, using the extension of T M0 to
a Lie algebroid on M ). More precisely, restricting to operators on negative order
and assuming that GM0 = M0 × M0 is the pair groupoid on M0 , we say that G is
a Fredholm Lie groupoid if it has the following property:
“a ∈ 1 + Cr∗ (G) is Fredholm ⇔ πx (a) is invertible ∀x ∈ ∂M ,” (1)

for the regular representations πx of C ∗ (G), where the Fredholmness is considered

on L2 (M0 ). (See 4.3 for a more detailed formulation of the definition of a Fredholm
groupoid.) The reader will recognize in condition (1) the type of Fredholm condi-
tions that appeared in Theorem 1.2 (and which are typically used in practice). It
turns out that if the defining condition (1) is satisfied for all a ∈ 1 + Cr∗ (G), then
it will be satisfied for many other operators associated to G, in particular it will
be satisfied for suitable operators a = P ∈ Ψm (G; E, F ) (if m > 0, one has to add
the ellipticity condition and take into account the domain of definition).
 Fredholm Conditions on Non-compact Manifolds 83

The extension of the Fredholm conditions in Equation (1) from operators in

1 + Cr∗ (G) to operators in Ψm (G; E, F ) turns out to be almost automatic and is
conveniently formalized through the use of strictly spectral and exhaustive fami-
lies of representations of groupoid C ∗ -algebras (see Section 3 for definitions and
references). Thus, in our paper, we shift the study of the Fredholm conditions from
the study of a single operator to the study of suitable algebras containing it. In
particular, it will be enough to study the properties and the representations of
regularizing pseudodifferential operators. The Fredholm properties of higher-order
pseudodifferential operators will then follow simply by including the ellipticity con-
dition. Thus the Fredholm conditions established in this paper will be formulated
in terms of representations of groupoid C ∗ -algebras.
To make our results applicable, we give various characterizations of Fredholm
groupoids and provide methods to prove that individual groupoids are Fredholm.
In particular, we introduce the class of “stratified submersion groupoids.” A strat-
ified submersion groupoid is one that is built out by gluing fibered pull-backs
of bundles of Lie groups. We show that a stratified submersion groupoid with
amenable stabilizers is Fredholm and hence that the interior of the set of its units
yields a manifold with amenable ends, and hence Theorem 1.2 is valid for these
manifolds.

1.4. Examples and applications

The main significance of the class of stratified submersion Lie groupoids mentioned
in the last paragraph is that many of the groupoids that appear in practice are
naturally stratified submersion Lie groupoids. Typically, using the notations as
above, we have that M identifies with a compactification of M0 to a manifold
with corners, where the behavior at F := ∂M := M \ M0 models the behavior at
infinity (‘at the ends’). Often the corresponding isotropy groups are amenable. A
class of manifolds that fits this perspective is the class of Lie manifolds [48, 49], a
few examples of which appear in Section 5. See [33].
Moreover, stratified submersion Lie groupoids are also tailored to applica-
tions to singular spaces obtained by (iterative) “desingularization procedures,”
the desingularization being the analog in the category of groupoids of the blow-up
construction in the category of manifolds with corners. In fact, it can be seen that
desingularization preserves Fredholm groupoids, and that the class of stratified
submersion groupoids is closed under desingularization. While we do not pursue
the applications to singular spaces in full detail here, we outline the key ideas in-
volved and give the construction of such a groupoid yielding Fredholm conditions
for the edge calculus.
Our results can be extended to products of such manifolds, or to manifolds
that locally at infinity are products of such manifolds.

1.5. Contents of the paper

The paper consists of roughly two parts: the theoretical part and applications. The
applications are included in the last two sections. The most concrete applications
84 C. Carvalho, V. Nistor and Y. Qiao

are contained in the last section, which we tried to write in such a way that it
can, to a large extent, be read independently of the rest of the paper. The reader
interested only in applications can thus start immediately with Section 5. At least
the main results of that section should be understandable without any knowledge
of groupoids.
We now describe in detail the main contents of the paper. We start with
reviewing the relevant topics related to groupoids and groupoid C ∗ -algebras, so
Section 2 consists mostly of background material on locally compact groupoids G,
on their Haar systems, and on their C ∗ -algebras. We then review manifolds with
corners and tame submersions, and we recall the definitions of Lie groupoids in the
framework that we need, that is, that of manifolds with corners. Note that all our
Lie groupoids will be second countable and Hausdorff. We finish this section with
a review of some examples of Lie groupoids that will play a role in what follows.
Section 3 contains preliminaries on exhaustive families of representations
from [76] and some general results on groupoid C ∗ -algebras. We recall the no-
tions of strictly spectral and strictly norming families [88] and the main results
from [76]. We then introduce groupoids with “Exel’s property,” and, respectively,
with “strong Exel’s property.” A groupoid satisfies Exel’s property (respectively,
the strong Exel’s property) when its family of regular representations is exhaustive
for the reduced C ∗ -algebra (respectively, for the full C ∗ -algebra). We prove that
groupoids given by fibered pull-backs of bundles of amenable Lie groups always
have Exel’s strong property. Motivated by this result, we then introduce the class
of stratified submersion groupoids, given essentially by gluing fibered pull-backs
of bundles of Lie groups. We then prove that stratified submersion Lie groupoids
with amenable isotropy groups always have Exel’s strong property.
In Section 4, we define Fredholm Lie groupoids. Note that in this paper,
we always work in the setting of Lie groupoids, except where otherwise explicitly
stated. We provide a characterization of Fredholm Lie groupoids using strictly
spectral families of regular representations and show that groupoids that have
the strong Exel’s property are Fredholm. It follows that stratified submersion Lie
groupoids with amenable isotropy groups are Fredholm. We then specialize to
algebras of pseudodifferential operators on Lie groupoids and obtain the crucial
Theorems 4.12 and 4.17.
The last section of the paper, Section 5, contains examples and applications
of our results, namely of Theorem 4.17. We start with group actions, define the as-
sociated transformation groupoid, and use this construction to describe Fredholm
conditions related to several pseudodifferential calculi: the b-groupoid that models
operators on manifolds with cylindrical and poly-cylindrical ends; the scattering
groupoid that models operators on manifolds that are asymptotically Euclidean,
and also operators on asymptotically hyperbolic spaces. We then consider the
edge calculus and construct a suitable Fredholm stratified submersion groupoid
that will recover Fredholm conditions. This is a particular case of a desingulariza-
tion groupoid, whose construction is briefly recalled in the following subsection.
(“Desingularization groupoids” were introduced in [75], where more details on this
 Fredholm Conditions on Non-compact Manifolds 85

construction can be found.) We then give more explicit examples of the desingu-
larization and blow-up process. A first example of the desingularization procedure
deals with the blow-up of a smooth submanifold along another smooth, compact
manifold. The second example extends this first example to manifolds with bound-
ary. Finally, the last example deals with the iterated blow-up of a singular stratified
subset of dimension one. (See the main body of the paper for specific references
to the existing literature related to these results.)
For simplicity, in view of the applications considered in this paper, we shall
work almost exclusively with Lie groupoids (and their reductions), although some
definitions and results are valid in greater generality. For the most part, our man-
ifolds will be Hausdorff and second countable.

2. Groupoids and their C ∗ -algebras

We recall in this section some basic definitions and properties of groupoids. Al-
though our interest in applications lies mainly in Lie groupoids, we have found
it convenient to consider occasionally also the general case of locally compact
groupoids, so in the beginning we consider both cases. We refer to Mackenzie’s
books [53, 54] for more details and, in general, for a nice introduction to the
subject of Lie groups and Lie groupoids, as well as to further references and his-
torical comments. See also [12, 64, 85, 109] for the more specialized issues relating
to analytic applications. We use the approach in [75], where most of the needed
definitions and results can also be found.
2.1. Locally compact groupoids
Let us introduce groupoids as in [12, 54, 85].
Definition 2.1. A groupoid is a pair G = (G (1) , G (0) ), where G (i) , i = 0, 1, are sets,
together with structural morphisms
(1) d, r : G (1) → G (0) (the “domain” and “range”),
(2) ι : G (1) → G (1) (the “inverse”),
(3) u : G (0) → G (1) (the inclusion of units), and
(4) μ : G (2) := {(g, h) ∈ G (1) × G (1) | d(g) = r(h)} → G (1) (the product),
with the following properties (we write gh := μ(g, h), for simplicity):
(1) d(gh) = d(h), r(gh) = r(g) if d(g) = r(h);
(2) g1 (g2 g3 ) = (g1 g2 )g3 for all gi ∈ G such that d(gi ) = r(gi+1 );
(3) d(u(x)) = x = r(u(x)), for all x ∈ M .
(4) gu(d(g)) = g and u(r(g))g = g for all g ∈ G.
(5) gι(g) = u(r(g)) and ι(g)g = u(d(g)) for all g ∈ G.
Moreover, we see that G (0) can be regarded as the set of objects of a category
with morphisms G (1) . The objects of G will also be called units and the morphisms
of G will also be called arrows. Actually, a groupoid G is simply a small category
in which every morphism is invertible. (A small category is one whose objects
86 C. Carvalho, V. Nistor and Y. Qiao

and morphisms are sets.) For convenience, we shall identify G = G (1) and denote
M := G (0) . We shall write G ⇒ M for a groupoid G with units M . In this paper,
G always denotes a groupoid.
Definition 2.2. A locally compact groupoid is a groupoid G ⇒ M such that:
(1) G and M are locally compact spaces, with M Hausdorff;
(2) the structural morphisms d, r, ι, u, and μ are continuous;
(3) d is surjective and open.
See Subsection 2.4 for examples of groupoids. Let us notice that the definition
requires only the space of units M to be Hausdorff. In this paper, however, all
spaces will be assumed or proved to be Hausdorff.
In what follows, we denote by GA := d−1 (A), G A = r−1 (A) and GA B
:=
−1 −1
d (A) ∩ r (B). We call GA the reduction of G to A. If A is G-invariant, in
A

the sense that GA A

= GA = G A = r−1 (A), then GA is also a groupoid, called the
restriction of G to A. For any x ∈ M , Gxx = d−1 (x) ∩ r−1 (x) is a group, called the
isotropy group at x.
We will study several examples of groupoids in detail when we will introduce
Lie groupoids.

2.2. Haar systems and C ∗ -algebras

We now recall the definition of a Haar system of a locally compact groupoid and
we use this opportunity to fix some more notations to be used throughout the rest
of the paper. We refer to [12, 43, 85] for more information on the topics discussed
in this subsection.
If G is a locally compact groupoid, we shall denote by Cc (G) the space of
continuous, complex-valued, compactly supported functions on G.
Definition 2.3. A right Haar system for a locally compact groupoid G is a family
λ = {λx }x∈M , where λx is a Borel regular measure on G with support supp(λx ) =
d−1 (x) =: Gx for every x ∈ M = G (0) , satisfying
(i) The continuity condition:
!
M  x → λx (ϕ) := ϕ(g)dλx (g) ∈ C
Gx

is continuous for ϕ ∈ Cc (G).

(ii) The invariance condition:
! !
ϕ(hg)dλr(g) (h) = ϕ(h)dλd(g) (h)
Gr(g) Gd(g)

for all g ∈ G and ϕ ∈ Cc (G).

We shall assume from now on that all our locally compact groupoids are
endowed with a (right) Haar system. Let thus G be a locally compact groupoid
 Fredholm Conditions on Non-compact Manifolds 87

and {λx } be the Haar system associated to it. The space Cc (G) has a natural
product given by the formula
!
(ϕ1 ∗ ϕ2 )(g) := ϕ1 (gh−1 )ϕ2 (h)dλd(g) (h) .
d−1 (d(g))

This makes Cc (G) into an associative ∗-algebra with the involution

ϕ∗ (g) := ϕ(g −1 )
for all g ∈ G and ϕ ∈ Cc (G). There also exists a natural algebra norm on Cc (G)
defined by ! !
% &
ϕI := max sup |ϕ|dλx , sup |ϕ∗ |dλx .
x∈M Gx x∈M Gx
The completion of Cc (G) with respect to the norm  · I is denoted L1 (G).
Recall [28] that a C ∗ -algebra is a complex algebra A together with a conjugate
linear involution ∗ and a complete norm   such that (ab)∗ = b∗ a∗ , ab ≤ ab,
and a∗ a = a2 , for all a, b ∈ A. Let H be a Hilbert space and denote by
L(H) the space of linear, bounded operators on H. Then L(H) is a C ∗ -algebra.
A representation of a C ∗ -algebra A on the Hilbert space Hπ is a ∗-morphism
π : A → L(Hπ ).
To a locally compact groupoid G (endowed with a Haar system), there are
associated two basic C ∗ -algebras, the full and reduced C ∗ -algebras C ∗ (G) and
Cr∗ (G), whose definition we now recall.
Definition 2.4. The full C ∗ -algebra associated to G, denoted C ∗ (G), is defined as
the completion of Cc (G) with respect to the norm
ϕ := sup π(ϕ) ,
π

where π ranges over all contractive ∗-representations of Cc (G). (Recall that a ∗-

representation π of Cc (G) is called contractive if π(ϕ) ≤ ϕI , for all ϕ ∈ Cc (G).)
Let us define as usual for any x ∈ M the regular representation πx : C ∗ (G) →
L(L2 (Gx , λx )) by the formula
!
(πx (ϕ)ψ)(g) := ϕ ∗ ψ(g) := ϕ(gh−1 )ψ(h)dλd(g) (h) , φ ∈ Cc (G) .
Gd(g)

We then define similarly the reduced C ∗ -algebra Cr∗ (G) as the completion of Cc (G)
with respect to the norm
ϕr := sup πx (ϕ)
x∈M
which is well defined, as all representations πx are contractive (of norm ≤ 1).
The groupoid G is said to be metrically amenable if the canonical surjective ∗-
homomorphism C ∗ (G) → Cr∗ (G), induced by the definitions above, is also injective.
Also, for further use, we note that if G is second countable, then C ∗ (G) is a
separable C ∗ -algebra.
88 C. Carvalho, V. Nistor and Y. Qiao

Remark 2.5. For any G-invariant, locally closed subset A ⊆ M , the reduced
groupoid GA = GA A
is locally compact and has a Haar system λA obtained by
restricting the Haar system λ of G to GA . In particular, we can construct as above
the corresponding C ∗ -algebra C ∗ (GA ) and the reduced C ∗ -algebra Cr∗ (GA ). For
any closed subset A ⊆ M , the subset d−1 (A) ⊆ G is also closed, so the restriction
map Cc (G) → Cc (d−1 (A)) is well defined. If A is also G-invariant then the restric-
tion extends by continuity to both a ∗-homomorphism ρA : C ∗ (G) → C ∗ (GA ) and
a ∗-homomorphism (ρA )r : Cr∗ (G) → Cr∗ (GA ).
We have the following well-known, but important result [68, 69, 86] that we
record for further reference.
Proposition 2.6. Let G ⇒ M be a second countable, locally compact groupoid with
a Haar system. Let U ⊂ M be an open G-invariant subset, F := M  U .
(i) C ∗ (GU ) is a closed two-sided ideal of C ∗ (G) that yields the short exact se-
quence
ρF
0 → C ∗ (GU ) → C ∗ (G) −→ C ∗ (GF ) → 0 .
(ii) If GF is metrically amenable, then one has the exact sequence
(ρF )r
0 → Cr∗ (GU ) → Cr∗ (G) −−−−−→ Cr∗ (GF ) → 0 .
(iii) If the groupoids GF and GU (respectively, G) are metrically amenable, then G
(respectively, GU ) is also metrically amenable.
Proof. The first assertion is well known, see for instance [69, Lemma 2.10]. The
second statement is in [86, Remark 4.10]. The last part follows from (ii) and the
Five Lemma. 
Remark 2.7. We notice that the exact sequence of Proposition 2.6 corresponds to
a disjoint union decomposition G = GF GU .
2.3. Manifolds with corners and Lie groupoids
Even if one is interested only in analysis on smooth manifolds (that is, in mani-
folds without corners or boundary), one important class of applications will be to
spaces obtained as the blow-up with respect to suitable submanifolds, which leads
to manifolds with corners, as each blow-up increases the highest codimension of
corners (the rank of the manifolds) by one. To model the analysis on these spaces,
we need Lie groupoids. The advantage of using Lie groupoids is that they have a
distinguished class of Haar systems defined using the smooth structure.
We begin with some background material following [75]. It will be important
to distinguish between smooth manifolds without corners (or boundaries) and
(smooth) manifolds with corners. The former will be smooth manifolds, while the
later will be simply manifolds. Thus, in this paper, a manifold M is a second
countable, Hausdorff topological space locally modeled by open subsets of [−1, 1]n
with smooth coordinate changes. In particular, our manifolds may have corners.
By contrast, a smooth manifold will not have corners (or boundary). A point
 Fredholm Conditions on Non-compact Manifolds 89

p ∈ M of a manifold (with corners) is called of depth k if it has a neighborhood

Vp diffeomorphic to [0, a)k × (−a, a)n−k , a > 0, by a diffeomorphism φp : Vp →
[0, a)k × (−a, a)n−k with φp (p) = 0. The maximal depth of points in M will be
called the rank of M . Thus, a manifold is smooth if it has rank 0.
A function f : M → M1 between two manifolds with corners will be called
smooth if its components are smooth in all coordinate charts. A little bit of extra
care is needed here in defining the derivatives. This is illustrated clearly in one
dimension, as one may have to consider side derivatives: if f : [0, 1] → R, then
f  (0) := limh→0,h>0 h−1 (f (h) − f (0)), whereas f  (1) := limh→0,h>0 h−1 (f (1) −
f (1−h)). Since limh→0 h−1 (f (x+h)−f (x)) = limh→0 h−1 (f (x)−f (x−h)) =: f  (x),
for 0 < x < 1, this defines unambiguously f  (x). A similar comment is in order in
higher dimensions as well. In local coordinates, a smooth function [0, 1]n → [0, 1]k
is one that is the restriction of a smooth function Rn → Rk .
For x ∈ M , the set of inward pointing tangent vectors in Tx (M ) defines a
closed cone denoted Tx+ (M ). Of course, if x ∈ M0 := M  ∂M , then Tx+ (M ) =
Tx (M ). Let M and M1 be manifolds with corners and f : M1 → M be a
smooth map. Then f induces a vector bundle map df : T M1 → T M such that
df (Tx+ (M1 )) ⊂ Tf+(x) M , x ∈ ∂M . If the smooth map f : M1 → M is injective, has
injective differential df , and has locally closed range, then we say that f (M1 ) is
a submanifold of M . We stress the condition that submanifolds be locally closed
(that is, the intersection of an open and a closed subset, or, equivalently, every
point x of a submanifold has an open neighborhood Ux that intersects the subman-
ifold in a closed subset of Ux ). We hence rule out embedded submanifolds. Other
than these conditions, our concept of submanifold is the most general possible.

Definition 2.8. A tame submersion h between two manifolds with corners M1

and M is a smooth map h : M1 → M such that its differential dh is surjective
everywhere and
(dhx )−1 (Th(x)
+
M ) = Tx+ M1 , x ∈ M.
(That is, dh(v) is an inward pointing vector of M if, and only if, v is an inward
pointing vector of M1 .)

For instance, the maps h1 , h2 : [0, ∞) × [0, ∞) → [0, ∞) defined by h1 (x, y) =

x + y, h2 (x, y) = x, are not tame submersions. Clearly, if h : M1 → M is a tame
submersion of manifolds with corners, then x and h(x) will have the same depth.
In particular, the rank of M1 is at most the rank of M . We have the following
well-known lemma (see for instance [75]).

Lemma 2.9. Let h : M1 → M be a tame submersion of manifolds with corners.

(i) For m1 ∈ M1 , there exists an open neighborhood U of m1 in M1 such that
h(U ) is open and the restriction of h to U is a C ∞ fibration with basis h(U ).
(ii) Let L ⊂ M be a submanifold, then L1 := h−1 (L) is a submanifold of M1 of
rank ≤ the rank of L.
90 C. Carvalho, V. Nistor and Y. Qiao

We now define Lie groupoids roughly by replacing our spaces with manifolds
with corners and the continuity conditions with smoothness conditions. The reason
why Lie groupoids play an important role in applications is that they model the
distribution kernels of many interesting classes of pseudodifferential operators.
Definition 2.10. A Lie groupoid is a groupoid G ⇒ M such that
(i) G and M are manifolds with corners,
(ii) the structural morphisms d, r, ι, and u are smooth,
(iii) d is a tame submersion of manifolds with corners, and
(iv) μ is smooth.
In particular, the Lie groupoids used in this paper are second countable,
Hausdorff, locally compact groupoids. In different contexts in other papers, it
is sometimes useful not to assume the set of arrows G of a Lie groupoid to be
Hausdorff, although the space of units M of a Lie groupoid G is always assumed
to be Hausdorff. Lie groupoids were introduced by Ehresmann. A more general
class also useful in applications is that of continuous family groupoids, where,
roughly speaking, the structural maps are just continuous along the units, but
keeping smoothness along the fibers. See [48, 79].
For the rest of this section, G will always be a Lie groupoid. In fact, most of
the groupoids that we will consider in what follows will be Lie groupoids. However,
one has to be careful since the restriction of a Lie groupoid is not necessarily a Lie
groupoid itself. Moreover, we shall assume that all our Lie groupoids are Hausdorff.
Since d and r are tame submersions, it follows from Lemma 2.9 that the
fibers Gx := d−1 (x) and G x := r−1 (x), x ∈ M , are smooth manifolds (that is, they
have no corners). Similarly, the same lemma implies that the set G (2) ⊂ G × G of
composable units is a manifold as well (but it may have corners). In particular, it
makes sense to consider smooth maps G (2) → G (1) .
Remark 2.11. Let G ⇒ M be a Lie groupoid and let A(G) := ∪x∈M Tx Gx , where
Gx := d−1 (x), as usual. Let us denote by f∗ : T M1 → T M2 the differential of a
smooth map f : M1 → M2 . Then A(G) is the restriction to units of the vector
bundle ker(d∗ ), and hence it has a natural structure of vector bundle over M . It is
called the Lie algebroid of G. The differential r∗ of the range map r : G → M then
induces a map  := r∗ |A(G) : A(G) → T M , called the anchor map of A(G). We let
Lie(G) := (Γ(M ; A(G))) denote the image by  of the space of sections of A(G).
It coincides with the image by r∗ of the space d-vertical (i.e., tangent to the fibers
Gx of d) vector fields on G that are invariant for the action of G on itself by right
multiplication. Therefore both Γ(M ; A(G)) and Lie(G) := (Γ(M ; A(G))) are Lie
algebras for the Lie bracket. Both Lie algebras will play an important role in the
analysis of differential operators.
Remark 2.12. We can use the vector bundle A(G) to define a Haar system. Let
D := |Λn A(G)|, where n is the dimension of the Lie algebroid of G. The pull-
back vector bundle r∗ (D) is the bundle of 1-densities along the fibers of d. A
 Fredholm Conditions on Non-compact Manifolds 91

trivialization of D will hence give rise to a right invariant set of measures on Gx

and hence to a right Haar system.
We conclude that a Lie groupoid always has a (right) Haar system, which we
shall always assume to be obtained as in the remark above.

2.4. Examples of Lie groupoids

We continue with various examples of constructions of Lie groupoids that will be
needed in what follows. Recall that we only consider Hausdorff groupoids in this
paper. Also, recall that a Lie groupoid G is metrically amenable if, by definition,
the canonical morphism C ∗ (G) → Cr∗ (G) is an isomorphism.
Example 2.13. Any Lie group G gives rise to a Lie groupoid with set of units
reduced to one point: the identity element of G. Thus d, r : G → M := {e} are
constant, μ : G × G → G is the group multiplication, ι(g) := g −1 is the (usual)
inverse, and u : M → G is the inclusion of the unit. We have that the Lie algebroid
A(G) of G coincides with the Lie algebra of G (the definition of a Lie algebroid
was recalled in Remark 2.11). Hence Lie(G) := Γ(A(G)) is also the Lie algebra of
G. We have that G is metrically amenable if, and only if, it is amenable in the
usual sense of groups. In particular, if G is solvable, then G is amenable.
Example 2.14. Let M be a manifold with corners (hence Hausdorff by our con-
ventions), let G = M , and d = r = u = ι = idM . Then G is a Lie groupoid with
only units. We shall call a Lie groupoid with these properties a manifold. We have
that A(M ) = M × {0}, that is, the zero vector bundle. We have Lie(G) = 0.
Example 2.15. Let Gi → Mi , i = 1, 2, be two Lie groupoids. Then G1 × G2 is a
Lie groupoid with units M1 × M2 and A(G1 × G2 ) = A(G1 ) × A(G2 ). We have that
Lie(G1 × G2 ) is a suitable topological tensor product of Lie(G1 ) and Lie(G2 ).
The product of a manifold with a Lie group is thus again a Lie groupoid. We
can “twist” this example to obtain a (smooth) “bundle of Lie groups.”
Example 2.16. Let G be a Lie group with automorphism group Aut(G) and let
P → M be a smooth, locally trivial, principal Aut(G)-bundle, with M and P
manifolds with corners (hence Hausdorff, by our conventions). Then the associated
fiber bundle G := P ×Aut(G) G with fiber G is a Lie groupoid with units M . A
Lie groupoid of this form will be called a smooth bundle of Lie groups. It satisfies
d = r. We have A(G)  P ×Aut(G) Lie(G) is a smooth bundle of Lie algebras and
 : A(G) → T M is the zero map. The Lie algebra Lie(G) identifies thus with the
family of sections of the bundle of Lie algebras P ×Aut(G) Lie(G). If G is amenable,
then G is metrically amenable. The fact that G is metrically amenable when G is
amenable will be of crucial importance to us and will be used when the group G
is a solvable Lie group.
The following simple example of the “pair groupoid” will be fundamental in
what follows.
92 C. Carvalho, V. Nistor and Y. Qiao

Example 2.17. Let M be a smooth manifold. (Thus M has no corners, according to

our terminology.) Then the pair groupoid of M is G := M ×M . It is a groupoid with
units M . The structural morphisms are as follows: d is the second projection, r is
the first projection, and the product μ is given by (m1 , m2 )(m2 , m3 ) = (m1 , m3 ).
This determines u(m) = (m, m) and ι(m, m ) = (m , m). We thus have that the
product on Cc∞ (G) = Cc∞ (M ×M ) is the product of operators with integral kernels.
We have A(M ×M ) = T M , with anchor map the identity map. One crucial feature
of the pair groupoid is that, for any x, the regular representation πx defines an
isomorphism between C ∗ (M ×M ) and the ideal of compact operators in L(L2 (M )).
In particular, all pair groupoids are metrically amenable.

Example 2.18. Let M be a smooth manifold (so without corners) and M # be it

universal covering. Let π1 (M ) be the fundamental group of M associated to some
fixed point of M . A related example to the pair groupoid is that of the path
groupoid P(M ) := (M #× M #)/π1 (M ) of M , which will have the same Lie algebroid
as the pair groupoid: A(P(M )) = T M . The analysis associated to P(M ) is that
of π1 (M )-invariant operators on the covering space M #, thus quite different to
that of the pair groupoid. This underscores the importance of choosing the right
integrating groupoid when one is interested in Analysis.

We extend the example of the pair groupoid by defining fibered pull-back

groupoids [40, 41] (we use the terminology in [75], however).

Example 2.19. Let again M and L be manifolds and f : M → L be continuous

map. Let H ⇒ L be a Lie groupoid. The fibered pull-back groupoid is then

f ↓↓ (H) := { (m, g, m ) ∈ M × H × M | f (m) = r(g), d(g) = f (m ) } ,

with units M and product (m, g, m )(m , g  , m ) = (m, gg  , m ). We shall also
sometimes write M ×f H ×f M = f ↓↓ (H) for the fibered pull-back groupoid.
Assume now that f is a tame submersion (in particular, that it is smooth). Then
f ↓↓ (H) is a Lie groupoid and its Lie algebroid is

A(f ↓↓ (H)) = {(ξ, X)| ξ ∈ A(H) , X ∈ T M , (ξ) = f∗ (X)} ,

the thick pull-back Lie algebroid f ↓↓ A(H) of A(H), see [53, 54].

The following particular example of a fibered pull-back will be useful for the
study of the b-groupoid Gb in Section 5.

Example 2.20. Let us assume that M is a smooth manifold and let B be its set
of connected components. Let G be a Lie group. We let H := B × G, the product
of a manifold and a Lie group, and f : M → B be the map that associates to a
point its connected component. Then f ↓↓ (H) is the topological disjoint union of
the groupoids (F × F ) × G (product of the pair groupoid and a Lie group) for F
ranging through the connected components of M .
 Fredholm Conditions on Non-compact Manifolds 93

3. Exhaustive families of representations and Exel’s question

We next recall some basic facts on exhaustive families of representations following
[76]. A groupoid has Exel’s property when its set of regular representations is
exhaustive. Exel [29] has asked which groupoids have this property. Establishing
that a groupoid has Exel’s property will turn out to be important in establishing
Fredholm conditions in the next section. At the end of the section, we prove that
groupoids given by fibered pull-backs of bundles of Lie groups always have Exel’s
property, which leads us to introduce, and extend this property to the more general
class of stratified submersion groupoids.
3.1. Exhaustive families of representations of C ∗ -algebras
A two-sided ideal I ⊂ A of a C ∗ -algebra A is called primitive if it is the kernel of
an irreducible representation of A. We shall denote by Prim(A) the set of primitive
ideals of A. The trivial ideal A (of A) is not considered a primitive ideal (of A),
so A ∈/ Prim(A). For any representation φ of A, we define its support supp(φ) ⊂
Prim(A) as the set of primitive ideals of A containing ker(φ). Let J be a closed,
two-sided ideal of A. Then Prim(J) = {I ∈ Prim(A)| J ⊂ I} and Prim(A/J)
identifies with {I ∈ Prim(A)| J ⊂ I} = Prim(J)c . We endow Prim(A) with the
hull-kernel topology, which is the topology whose open sets are those of the form
Prim(J), with J a two-sided ideal of A.
We shall need the following definition [76].
Definition 3.1. Let F be a setof representations of a C ∗ -algebra A. We say that
F is exhaustive if Prim(A) = φ∈F supp(φ).
For applications, we shall need the following definition [88].
Definition 3.2. Let F be a set of representations of a unital C ∗ -algebra A.
(i) We say that F is strictly norming if, for any a ∈ A, there exists φ ∈ F such
that φ(a) = a.
(ii) We say that F is strictly spectral if, for any a ∈ A, we have that a is invertible
in A if, and only if, φ(a) is invertible for all φ ∈ F .
The two definitions above can be formulated for sets of morphisms or sets of
primitive ideals. We also have that F∗ := {π ∈ F | π = 0} has the same properties
(“exhaustive,” “strictly norming,” . . . ) as F .
If A is non-unital, we modify the last definition as follows [76]. Let A+ :=
A ⊕ C and χ0 : A+ → C be the morphism defined by χ0 = 0 on A and χ0 (1) = 1.
We then replace A with A+ and F with F + := F ∪ {χ0 }. This works also for
exhaustive families since F is exhaustive for A if, and only if, F + is exhaustive for
A+ . Sometimes it is convenient to use the following alternative characterization of
strictly spectral sets of representations. The set of representations F of A is strictly
spectral if it satisfies the following property: 1 + a ∈ A+ , a ∈ A, is invertible if,
and only if, 1 + φ(a) is invertible for any φ ∈ F.
A set of representations containing a representative in each equivalence classes
of irreducible representations of a unital C ∗ -algebra is strictly norming (see [28,
94 C. Carvalho, V. Nistor and Y. Qiao

29]). Therefore any exhaustive family is strictly norming [76]. In this regard, let us
recall now the following results from [76, 88], which establish the relations between
these concepts.
Theorem 3.3. Let F be a set of non-degenerate representations of a C ∗ -algebra A.
Then F is strictly norming if, and only if, it is strictly spectral. Every exhaustive
set of representations is strictly spectral. If A is furthermore separable, then the
converse is also true.
Let A be a C ∗ -algebra and I ⊂ A be a closed two-sided ideal. Recall that
any representation π : I → L(H) extends to a representation π : A → L(H). If π is
non-degenerate, this extension is unique. (See [28, Proposition 2.10.4].) This leads
to the following results (see [76], Proposition 3.15 and Corollary 3.16).
Proposition 3.4. Let I ⊂ A be an ideal of a C ∗ -algebra. Let FI be a set of non-
degenerate representations of I and FA/I be a set of representations of A/I. Let
F := FI ∪ FA/I , regarded as a family of representations of A. If FI and FA/I
are both exhaustive, then F is also exhaustive. The same result holds by replacing
exhaustive with strictly norming.
In the following corollary, one should think of the invertibility of a in A/I as
an “ellipticity” condition.
Corollary 3.5. Let I ⊂ A be an ideal of a unital C ∗ -algebra A and let FI be a
strictly spectral set of nondegenerate representations of I. Let a ∈ A. Then a is
invertible in A if, and only if, it is invertible in A/I and φ(a) is invertible for all
φ ∈ FI .
We now address Morita equivalence [87] in a very simple form (see also [79,
85]). It is known that there exists a homeomorphism between the primitive ideal
spectra of Morita equivalent C ∗ -algebras, so exhaustive families of representations
will correspond to exhaustive families of representations under Morita equivalence.
In particular, we have the following result.
Proposition 3.6. Let F be a set of representations of a C ∗ -algebra A.
(i) If Fn := {π ⊗ 1| π ∈ F } is the corresponding family of representations of
Mn (A) = A ⊗ Mn (C), then F is exhaustive if, and only if, Fn is exhaustive.
(ii) Let I ⊂ A be a closed, two-sided ideal. If F is exhaustive, then FI := {π|I | π ∈
F } is an exhaustive family of representations of I.
(iii) Let e ∈ Mn (A) be a projection. Then Fn defines, by restriction, an exhaustive
family of representations of eMn (A)e.
Proof. The first part follows by Morita equivalence. The second part follows since
supp(φ|I ) = supp(φ) ∩ Prim(I). The last part is proved by combining the first two
parts. Indeed, Fn defines an exhaustive set of representations of Mn (A), and hence
an exhaustive set of representations of Mn (A)eMn (A). Since the algebra eMn (A)e
is Morita equivalent to Mn (A)eMn (A), by Proposition 4.27 in [35] (see also [87]),
the result follows. 
 Fredholm Conditions on Non-compact Manifolds 95

3.2. Exel’s question and properties

We will let G denote a Lie groupoid and R(G) denote its set of regular repre-
sentations. The discussion below usually makes sense in the more general setting
of locally compact groupoids, but we restrict, nevertheless to Lie groupoids, for
simplicity.
In [29], Exel has asked for which groupoids G is the set R(G) a strictly norming
set of representations of Cr∗ (G). An example due to Voiculescu, see [76], shows that
not every locally compact groupoid has this property. (The question whether all Lie
groupoids have Exel’s property is still open.) It is easier to work with exhaustive
families of representations, so we introduce the following definition.
Definition 3.7. We say that G has Exel’s property if R(G), the set of regular rep-
resentations of G, is an exhaustive set of representations for Cr∗ (G). We say that
G has Exel’s strong property if R(G) is an exhaustive set of representations for
C ∗ (G).
We have that
.
supp(φ) = Prim(A)\ Prim(ker(φ)) = Prim(A)\ Prim(ker(φ)).
φ∈F φ∈F φ∈F

Hence, if F is an exhaustive family of representations of a C ∗ -algebra A, then

-
φ∈F Prim(ker(φ)) = ∅, and hence F is a faithful family of representations of A.
In particular, G has Exel’s strong property if, and only if, G has Exel’s property
and is metrically amenable.
We shall need the following consequence of Propositions 2.6 and 3.4 of the
previous subsection.
Corollary 3.8. Let U ⊂ M be an open, G-invariant subset and assume G to be
second countable. We let F := M  U . Let us assume that GU and GF have Exel’s
strong property. Then G also has Exel’s strong property.
Proof. Let I := C ∗ (GU ) = Cr∗ (GU ), A := C ∗ (G). Then A/I  C ∗ (GF )  Cr∗ (GF )
by Proposition 2.6(ii). Since GF and GU are metrically amenable, it follows that
G is metrically amenable as well, by Proposition 2.6(iii). Proposition 3.4 then
gives that R(G) = R(GU ) ∪ R(GF ) is an exhaustive family of representations of
C ∗ (G)  Cr∗ (G), that is, G has Exel’s strong property, as claimed. 
More generally, we can consider groupoids given by a filtration.
Corollary 3.9. Let G ⇒ M be a second countable groupoid and assume that Ui are
open, G-invariant subsets of M , such that ∅ =: U−1 ⊂ U0 ⊂ · · · ⊂ Ui−1 ⊂ Ui ⊂
· · · ⊂ UN := M , and for each S := Ui  Ui−1 , GS has Exel’s strong property. Then
G also has Exel’s strong property.
Proof. Corollary 3.8 yields the result by induction: If N = 1, it reduces to Corollary
3.8. Assuming that the result holds for a filtration in N − 1 sets, then GUN −1
has Exel’s strong property. Since by assumption GM\UN −1 also has Exel’s strong
property, the result follows again by Corollary 3.8. 
96 C. Carvalho, V. Nistor and Y. Qiao

Here is a simple, but very important example of a class of Lie groupoids G for
which the set R(G) of regular representations is an exhaustive set of representations
of C ∗ (G). The idea of the proof below is that, locally, the C ∗ -algebra of G is of
the form C ∗ (G) ⊗ K, where K are the compact operators on the (typical) fiber of
f . We found it convenient to formalize the proof of this well-known result using
Morita equivalence of groupoids.
Proposition 3.10. Let H ⇒ L be a bundle of Lie groups with fiber G and let
f : M → L be a tame submersion. Define G := f ↓↓ (H). Then G is a Lie groupoid
with Exel’s property. If G is amenable, then G has Exel’s strong property.
Proof. We have already seen that G := f ↓↓ (H) is a Lie groupoid (because f is
a tame submersion). We endow then G with one of the standard Haar measures
coming from the Lie groupoid structure. Let
X = M ×L H := {(m, γ) ∈ M × H| f (m) = r(γ)} .
Then X defines a Morita equivalence between G and H, and hence a Morita equiva-
lence between their full and reduced C ∗ -algebras of these groupoids, by the classical
results of [68, 99]. See also [103]. In particular, Prim(C ∗ (G)) and Prim(C ∗ (H)) are
homeomorphic by natural homeomorphisms [87]. The same is true for Prim(Cr∗ (G))
and Prim(Cr∗ (H)).
All irreducible representations of Cr∗ (H) factor through an evaluation mor-
phism ex : Cr∗ (H) → Cr∗ (Gx )  Cr∗ (G), for some x ∈ M , where Gx is the fiber of
H → M above x, since locally H is a product of a group and a space. We have
that the regular representation πx is obtained from the regular representation of
Gx via ex , the evaluation at x. Hence G has Exel’s property.
Let us assume now that G is amenable. Hence the fibers of H → M are
amenable, and therefore H is metrically amenable. Hence G is also metrically
amenable, by their Morita equivalence. Since G was already proved to have Exel’s
property, it follows that it has also Exel’s strong property. 

We notice that if G is isomorphic to a groupoid as in the proposition above,

then H ⇒ L and f : M → L are uniquely determined, up to equivalence. Indeed,
L is diffeomorphic to the set of orbits of G (acting on M ), f becomes the quotient
map M → M/G and H is obtained from the isotropy groupoid of G.
The above results suggest the following definition.
Definition 3.11. Let H be a groupoid with units M . We say that H is a stratified
submersion groupoid with filtration (Ui ) if Ui are open, H-invariant subsets of M ,
∅ =: U−1 ⊂ U0 ⊂ · · · ⊂ Ui−1 ⊂ Ui ⊂ · · · ⊂ UN := M ,
and each S := Ui  Ui−1 is a manifold, possibly with corners, such that there exist
a Lie group bundle GS → BS and a tame submersion fS : S → BS of manifolds
with corners such that
HS  fS↓↓ (GS ) .
 Fredholm Conditions on Non-compact Manifolds 97

A groupoid H will be called a stratified submersion groupoid if it is a stratified

submersion groupoid for some filtration.
Remark 3.12. It is clear from the definition that if H is a stratified submersion
groupoid, then, with the notation of the definition, each HUk Uj , j < k, is also a
stratified submersion groupoid. We have to notice, however, that HUk Uj may not
be a Lie groupoid, even if H is a Lie groupoid. This is because Uk  Uj may not
be a manifold (it is only a union of manifolds).
As we shall see, the class of stratified submersion groupoids is large enough to
comprise many classes of groupoids and operators that arise in practice. Combining
Corollary 3.9 with Proposition 3.10, we now obtain the following result.
Theorem 3.13. Let G ⇒ M be a stratified submersion Lie groupoid and assume
that all the isotropy groups Gxx , x ∈ M , are amenable. Then the set R(G) of
regular representations of G is an exhaustive family of representations of C ∗ (G).
In particular, G is metrically amenable and has Exel’s property.
Proof. Let us use the notation of Definition 3.11. Indeed then, the isotropy group
Gxx is isomorphic to the fiber of GS → BS above fS (x), for x ∈ S. Thus, by
Proposition 3.10, all the groupoids GS are metrically amenable and have Exel’s
property. Using Corollary 3.9, we obtain that G has Exel’s strong property, that
is, that it is metrically amenable and has Exel’s property. 
In the next section, we will see how this property plays a role in establishing
Fredholm conditions for operators on groupoids.
Remark 3.14. Let H be a stratified submersion Lie groupoid as in Definition 3.11.
In view of the remark 2.7, we have then the following disjoint union decomposition:
H = S fS↓↓ (GS ) = S S ×B S ×B GS ,
where ×B is the fibered product over the base (that is, over BS ). We notice also
that two regular representations πx and πy are unitarily equivalent if there is g ∈ G
such that d(g) = x and r(g) = y. For a stratified submersion groupoid, this is the
case exactly if x, y ∈ S := Ui−1  Ui and fS (x) = fS (y) = b ∈ Bs , in which case,
these two representations act, up to a unitary equivalence, on the set of square
integrable functions on the space Zb := fS−1 (b) × (GS )b . We denote by πb the
corresponding representation on L2 (Zb ), b ∈ BS .

4. Fredholm conditions
We now study Fredholm conditions for operators in algebras Ψ containing a re-
duced groupoid C ∗ -algebra Cr∗ (G) as an essential ideal. (Recall that an ideal of an
algebra is essential if its annihilator vanishes.) The groupoids for which we obtain
the kind of Fredholm conditions that we want (the kind that are typically used in
practice) will be called “Fredholm groupoids.” They are introduced and discussed
next. Examples of Fredholm groupoids will be provided in the next section.
98 C. Carvalho, V. Nistor and Y. Qiao

Throughout the rest of this paper, G ⇒ M denote a second countable Lie

groupoid, which is hence a locally compact groupoid. Let us fix a metric on A(G)
and hence a right Haar system λx on G. Also, recall that all our Lie groupoids are
assumed to be Hausdorff.
4.1. Fredholm groupoids and their characterization
In this subsection, we introduce Fredholm groupoids and give a first character-
ization of these groupoids. Recall that πx : C ∗ (G) → L(L2 (Gx , λx )) denotes the
regular representation on Gx , given by left convolution, πx (φ)ψ := φ ∗ ψ (Defini-
tion 2.4). The following notation will be used throughout the rest of the paper.
Notations 4.1. We assume that U ⊂ M is an open, G-invariant subset with GU 
U × U (the pair groupoid, see Example 2.17). For any x0 ∈ U , the range map
then defines a bijection r : Gx0 → U and hence a measure μ on U corresponding
to λx0 . This measure does not depend on the choice of x0 ∈ U and leads to
isometries L2 (Gx0 , λx0 )  L2 (U ; μ) that commute with the action of G. We then
denote by π0 the corresponding representation of C ∗ (G) on L2 (U ; μ) and call it
the vector representation of C ∗ (G). We shall usually write L2 (U ) := L2 (U ; μ). The
representation π0 then defines an isomorphism Cr∗ (GU )  π0 (Cr∗ (GU )) = K, the
algebra of compact operators on L2 (U ).
Remark 4.2. An important remark is that, since we only consider Hausdorff Lie
groupoids, the vector representation π0 is injective, by a result of Khoshkam and
Skandalis [44]. In particular, we have that 1 + π0 (a) is invertible in π0 (Cr∗ (G))/K if,
and only if, 1 + a is invertible in Cr∗ (G)/Cr∗ (GU ). We shall thus identify Cr∗ (G) with
its image under π0 , that is, with a class of operators on L2 (U ), without further
comment.
In the following definition, we shall use that Cr∗ (GU ) ⊂ ker πx whenever x ∈
/
U , by definition, and hence that πx descends to a representation of the quotient
algebra Cr∗ (G)/Cr∗ (GU ).
Definition 4.3. We say that G is a Fredholm Lie groupoid if:
(i) There is an open, dense, G-invariant subset U ⊂ M such that GU  U × U .
(ii) Let a ∈ Cr∗ (G). We have that 1 + a is Fredholm if, and only if, 1 + πx (a) is
invertible for each x ∈ M  U (see 4.1 for notation).
The set F := M  U will be called the set of boundary units of G and a set U as
in this definition will be called a manifold with amenable ends.
For a Fredholm Lie groupoid G ⇒ M , we shall always denote by U ⊂ M the
G-invariant open subset from the definition of a Lie Fredholm groupoid. Since U
is a dense orbit, it is uniquely determined by G. Hence, F := M  U is closed and
G-invariant.
We shall need the following result for the matricial case.
Proposition 4.4. Let us assume that G ⇒ M is a Fredholm Lie groupoid, that M
is compact, and that e ∈ Mn (C(M )) is a projection.
 Fredholm Conditions on Non-compact Manifolds 99

(i) The set of representations πx , x ∈ F defines a strictly spectral family of

representations of eMn Cr∗ (G)/Cr∗ (GU ) e.
(ii) For any a ∈ eMn (Cr∗ (G))e, we have that 1 + a is Fredholm if, and only if,
1 + πx (a) is invertible for each x ∈ F := M  U .
Proof. By Atkinson’s theorem (which states that an operator is Fredholm if, and
only if, it is invertible modulo the compacts) and by the definition of strictly spec-
tral families of representations in the non-unital case, one can see that condition
(ii) in the definition of Fredhom groupoid is equivalent to the fact that R(GF )
defines a strictly spectral family of representations of the quotient Cr∗ (G)/K 
Cr∗ (G)/Cr∗ (GU ). Since G is second countable, its associated C ∗ -algebras are sepa-
rable, hence the set {πx | x ∈ F } is strictly spectral if, and only if, it is exhaustive,
by Theorem 3.3. The first part is then a consequence of Proposition 3.6.
The second part follows again from the definition of strictly spectral families
of representations in the non-unital case, from Cr∗ (GU )  K(L2 (U )), and from
Atkinson’s theorem. 
In the next results, we give an abstract characterization of Lie Fredholm
groupoids, using some of the ideas of the previous section. We obtain in particular
conditions that will make it simpler to check whether a given Lie groupoid is
Fredholm. Recall that, in this paper, all groupoids are Hausdorff.
Theorem 4.5. Assume G is a Lie groupoid. If G is Fredholm, then the following
two conditions are satisfied:
(i) The canonical projection Cr∗ (G) → Cr∗ (GF ) induces an isomorphism
Cr∗ (G)/Cr∗ (GU )  Cr∗ (GF ) , F := M  U .
(ii) GF has Exel’s property.
(Recall that condition (ii) means that R(GF ) is an exhaustive set of repre-
sentations of Cr∗ (GF ).)
Proof. Let us assume that G is a Fredholm groupoid and check the two conditions
of the statement.
Let p : Cr∗ (G) → Cr∗ (GF ) be the natural projection induced by restriction. It
is known that Cr∗ (GU ) ⊆ ker(p) (see the proof of [85, Prop.II.4.5 (a)]). To prove
(i), we need to show that we have equality Cr∗ (GU ) = ker(p). Let us proceed by
contradiction, that is, let us assume that Cr∗ (GU ) = ker(p). Then we can choose
a = a∗ ∈ ker(p)  Cr∗ (GU ). Note that since πx = πx ◦ p, x ∈ F = M \ U , we have
ker p ⊂ ker πx , in particular, 1 − πx (a) = 1 is invertible for all x ∈
/ U . On the other
hand, since a is self-adjoint and non-zero, in the quotient ker(p)/Cr∗ (GU ), there is
0 = λ ∈ R such that λ − a is not invertible in ker(p)/Cr∗ (GU ). By rescaling, we may
assume λ = 1 and thus 1 − a is not invertible in Cr∗ (G)/Cr∗ (GU ) ∼ = Cr∗ (G)/K. We
conclude that, in this case, R(GF ) is not a strictly spectral family of representations
of Cr∗ (G)/Cr∗ (GU ). It then follows that 1−a is not Fredholm, however 1−πx (a) = 1
is invertible for all x ∈ / U . This is a contradiction.
100 C. Carvalho, V. Nistor and Y. Qiao

The first part shows that there is an isomorphism

Cr∗ (GF )  Cr∗ (G)/Cr∗ (GU )  Cr∗ (G)/K(L2 (U )).
By Fredholmness of G, we have that R(GF ) defines a strictly spectral family of
representations of Cr∗ (G)/K, so R(GF ) is also a strictly spectral family of repre-
sentations of Cr∗ (GF ). Since G is second countable, it is exhaustive, by Theorem
3.3, and hence GF has Exel’s property. This proves (ii) and hence the direct im-
plication. 
We remark that we proved in fact that, if R(GF ) defines a strictly spectral
family of representations of Cr∗ (G)/Cr∗ (GU ), then Cr∗ (G)/Cr∗ (GU )  Cr∗ (GF ) and
GF has Exel’s property.
A stronger form of the converse of Theorem 4.5 is contained in the following
theorem. Recall that a faithful representation of an ideal in a C ∗ -algebra extends
uniquely to the whole algebra.
Theorem 4.6. Let G ⇒ M be a Lie groupoid. Let us assume that U ⊂ M is a
dense G-invariant open subset such that GU  U × U . Also, let us assume that
the two conditions of Theorem 4.5 are satisfied. Then, for any unital C ∗ -algebra
Ψ containing Mn (Cr∗ (G)) as an essential ideal and for any a ∈ Ψ, we have that
π0 (a) if Fredholm if, and only if, the image of a in Ψ/Mn (Cr∗ (G)) is invertible and
all πx (a), x ∈ M  U , are invertible.
Proof. The proof is the same for any n (using also Proposition 3.6, so we let n = 1
for notational simplicity). Let us assume that conditions (i) and (ii) of Theorem
4.5 are satisfied and let Ψ be a C ∗ -algebra containing Cr∗ (G) as an essential ideal.
Note that π0 is injective on Ψ since C ∗ (G) is an essential ideal of Ψ. By our
assumptions on G, we obtain that the algebra Ψ/Cr∗ (GU )  π0 (Ψ)/K =: B contains
B0 := π0 (Cr∗ (G))/K as an ideal and B/B0  Ψ/Cr∗ (G). Moreover, B0  Cr∗ (GF )
by (i).
Let now a ∈ Ψ be arbitrary. By Atkinson’s Theorem, we know that π0 (a) is
Fredholm if, and only if, its image in B is invertible. But by (ii) and by Corollary
3.5, c ∈ B is invertible if, and only if, the image of c in Ψ/Cr∗ (G) and all πx (c) are
invertible for all x ∈ F . This completes the proof. 
Corollary 4.7. Let G ⇒ M be a Lie groupoid and ∅ = U ⊂ M be a dense, G-
invariant, open subset such that GU  U × U , and F = M \ U . Then G is Fredholm
if, and only if, Cr∗ (G)/Cr∗ (GU )  Cr∗ (GF ) and R(GF ) is a strictly spectral set of
representations of Cr∗ (GF ) (that is, if GF has Exel’s property).
Recalling the exact sequence in Proposition 2.6, it follows that, in particular,
if G is metrically amenable, then G is Fredholdm if, and only if, GF has Exel’s
property.
Remark 4.8. It is immediate to observe that, if the set R(GF ) of regular represen-
tations πx , x ∈ F , forms a strictly spectral set of representations of C ∗ (GF ), then
GF has Exel’s strong property, and conditions (ii) and (iii) of the Theorem 4.5 are
satisfied (using Proposition 2.6).
 Fredholm Conditions on Non-compact Manifolds 101

This remark and the above results then give the following sufficient conditions
for a Lie groupoid to be Fredholm.
Corollary 4.9. Let G ⇒ M be a Lie groupoid and U ⊂ M be a dense G-invariant
open subset such that GU  U × U , and F = M \ U . Assume that R(GF ) forms a
strictly spectral set of representations of C ∗ (GF ) (equivalently, that GF has Exel’s
strong property). Then G is a Fredholm groupoid.
See also [7]. Together with Theorem 3.13, we obtain an important class of
Fredholm groupoids. Recall that we assume that all our groupoids are Hausdorff.
Theorem 4.10. Let G ⇒ M be a stratified submersion Lie groupoid with filtration
Ui such that U0 is dense and GU0  U0 × U0 . Assume that all isotropy groups Gxx
are amenable. Then G is a Fredholm groupoid.
Proof. By Remark 3.12 applied to F := M  U0 and by Theorem 3.13, we know
that the family R(GF ) of regular representations of GF is an exhaustive family
of representations of C ∗ (GF ), that is, GF has Exel’s strong property. The result
follows from the previous corollary. 

We notice that we can replace πx , x ∈ F , with πb , b ∈ BS , for the strata

corresponding to the boundary. We can further restrict to the set of b’s in a dense
set of orbits, since they give rise to a faithful family of representations of GF , again
by [44].

4.2. Pseudodifferential operators

So far we have discussed mostly groupoids and operators in their (unitized) C ∗ -
algebras. The study of differential operators on suitable non-compact spaces can
be reduced to the study of groupoid algebras using pseudodifferential operators
on Lie groupoids. In this subsection, we will explain this reduction procedure. See
[42, 72, 90, 98, 108] for introductions to pseudodifferential operators.
Let G be a Lie groupoid and consider the algebra Ψ∗ (G) of pseudodifferential
operators on the groupoid G, whose definition we now briefly recall [4, 65, 77].
Then, for m ∈ Z ∪ {±∞}, Ψm (G) consists of smooth families (Px )x∈M of classical
pseudo-differential operators Px ∈ Ψm (Gx ) of order m, that are right invariant with
respect to the action of G and have compactly supported distribution kernels. In
particular, Ψ−∞ (G) is nothing but the convolution algebra of smooth, compactly
supported function on G, that is, Ψ−∞ (G)  Cc∞ (G).
This construction extends right away to operators between smooth sections
of some vector bundles E, F → M . There are two methods of doing this. The first
method is to consider families (Px )x∈M of operators Px ∈ Ψm (Gx ; r∗ (E), r∗ (F )).
We denote by Ψm (G; E, F ) the resulting set of operators. By embedding E ⊕F into
a trivial bundle of dimension N , we can identify Ψm (G; E, F ) with a subspace of
MN (Ψm (G)), when convenient. Similar considerations yield differential operators
generated by V acting between vector bundles E and F endowed with metric
preserving connections.
102 C. Carvalho, V. Nistor and Y. Qiao

The second method to deal with operators between two vector bundles
E, F → M is as follows. Let us choose an embedding of E ⊕ F → CN into a
trivial vector bundle and denote by e and f the corresponding orthogonal projec-
tions onto the images of e and f . Then we can identify
Ψm (G; E, F ) := f MN (Ψm (G))e .
This idea applies to other types of similar operators (norm closures of operators
in Ψm (G), differential operators, . . . ).
One of the main points of considering pseudodifferential operators on a
groupoid is that, if we denote V := Lie(G) := (Γ(A(G))), then
Diff m (V; E, F ) = Ψm (G; E, F ) ∩ Diff(M ; E, F ) ,
where Diff m (V; E, F ) is generated by C ∞ (M ; End(E⊕F )) and V (using compatible
connections on E and F ) and Diff(M ; E, F ) denotes all differential operators on
M acting between sections of E and F . See also 4.14.
Another main point for introducing the algebra Ψ∗ (G) is related to para-
metrices and inverses of differential operators. To discuss this, let us denote by
Ψ(G) the C ∗ -algebra obtained as the closure of Ψ0 (G) with respect to all contrac-
tive ∗-representations of Ψ−∞ (G), as in [49]. As we will see below, this definition
extends to operators acting between vector bundles. Let us denote, as usual, by
S ∗ A the set of unit vectors in the Lie algebroid A∗ (G) associated with G (as in
Remark 2.11), with respect to some fixed metric on A∗ (G). Then Ψ(G) fits into
the following exact sequence
σ0
0 → C ∗ (G) → Ψ(G) −→ C0 (S ∗ A) → 0 ,
where σ0 is an extension of the principal symbol map (see for instance [48] and
the references therein.) Moreover, if P ∈ Ψm (G; E, F ), m ≥ 0, is such that
P : H m (M ; E) → L2 (M ; F ) is invertible, then its inverse will be in Ψ(G; F, E),
the norm closure of Ψ0 (G; F, E). Typically, Fredholm conditions are obtained for
Fredholm Lie groupoids by applying our results to the algebra Ψ = MN (Ψ(G)),
for some N . Let us see how this is done.
We first need to recall the definition of Sobolev spaces. Let us fix in this
subsection a Lie groupoid G ⇒ M . For the purpose of the next result, let us
assume that its space of units M has an open, dense, G-invariant subset U ⊂
M0 := M  ∂M such that the restriction GU is isomorphic to the pair groupoid
U × U . Let us also assume that the space of units M is compact. This is needed
in order to construct canonical Sobolev spaces on the interior of M . Indeed, there
is an essentially unique class of metrics on A(G), which, by restriction, gives rise
to a class of metrics on U , that are called compatible (with the groupoid G), see
also [3]. All these metrics are Lipschitz equivalent and complete [2, 49]. In fact, the
Sobolev spaces of all these metrics will coincide. They are given as the domains
of the powers of 1 + Δ, where Δ is the (geometer’ s, i.e., positive) Laplacian.
We shall denote by H s (U ) = H s (M ) these Sobolev spaces. The Sobolev spaces
H s (M ) are discussed in detail in [2]. See also [74] for a review. The operators in
 Fredholm Conditions on Non-compact Manifolds 103

Ψm (G) (and their vector bundle analogues) model differential operators associated
to (any) compatible metric (see also Remark 4.14 below). By considering a vector
bundle E → M and ΔE the associated Laplacian on sections of E, we obtain the
spaces H s (M ; E).
Note that the Sobolev spaces H s (M ) are not defined using the compact
manifold structure on M , but rather using the complete metric on U ⊂ M . Recall
that if the set M \ U is as in Definition 4.3, then the set U is called a manifold
with amenable ends. If P ∈ Ψm (G; E, F ), then P : H s (M ; E) → H s−m (M ; F ) is
continuous.
Let Ψm (G; E, F ) denote the pseudodifferential operators acting on the lifts
of E and F to G via d. We let Ψm (G; E) := Ψm (G; E, E). Then if P : H s (M ; E) →
H s−m (M ; F ) is an order m pseudodifferential operator, we replace the study of
its Fredholm properties with those of
⎡ ⎤
0 P 0
P1 := ⎣ P ∗ 0 0 ⎦,
0 0 Qm
where Qm is an invertible order m pseudodifferential operator acting on the com-
plement of E ⊕ F . Let G denote the complement of E ⊕ F in CN . We fix the
smooth vector bundles E, F, G → M in what follows. We denote by ΔE , ΔF , and
ΔG the associated Laplacians acting on sections of E, respectively, F and G. For
instance, Qm ∈ Ψm (G; G) could be (1 + ΔG )m/2 (or a suitable approximation it,
see [49]).
One of the main drawbacks of the algebras Ψm (G) is that they are too small
to contain resolvents. This issue is easily fixed by considering completions with
respect to suitable norms. Let us consider the norm  · m,s defined by.
P m,s := (1 + ΔF )(s−m)/2 P (1 + ΔE )−s/2 L2 →L2 (2)
We then let

s (G; E, F ) :=  · m,s -closure of Ψ (G; E, F ) .

Lm m
(3)
Recall that H (M ; E) is the domain of (1 + ΔE ) , if s ≥ 0, whenever M is
s s/2

compact (see [2, 37, 49]. Hence Lm m

s (G; E, F ) is the norm closure of Ψ (G; E, F )
in the norm topology of continuous linear maps H (M ; E) → H
s s−m
(M ; F ). We
note that, if P ∈ Ψm (G; E, F ), then
(1 + ΔF )(s−m)/2 P (1 + ΔE )−s/2 ∈ Ψ(G; E, F ) =: L00 (G; E, F )
by [49] (see also [5]). Moreover, let
W m (G) := Ψm (G) + ∩s L−∞
s (G) .
∞
Then W m (G) ⊂ Lm s (G) and W (G) is an algebra of pseudodifferential operators
that contains the inverses of its L2 -invertible operators.
Recall that we denote by π0 : C ∗ (G) → L(L2 (M )) the vector representation
and that M is compact. It is unitarily equivalent to the regular representations
104 C. Carvalho, V. Nistor and Y. Qiao

πx , x ∈ U . Moreover, since G is Hausdorff, π0 is injective. We have the following

result from [49].
Proposition 4.11. Let G ⇒ M be a Lie groupoid with GU = U × U and M compact.
Then MN (Ψ(G)) = MN (L00 (G)) contains MN (C ∗ (G)) as an essential ideal. Let
s ∈ R and P ∈ Lm s (G; E, F ) ⊃ Ψ (G; E, F ). Then
m

a := (1 + ΔF )(s−m)/2 P (1 + ΔE )−s/2 ∈ Ψ(G; E, F ), .

We have that P : H s (M ; E) → H s−m (M ; F ) is Fredholm if, and only if, a :
L2 (M ; E) → L2 (M ; F ) is Fredholm.
This proposition applies, in particular, if G is a Fredholm Lie groupoid with
M compact.

Proof. We can assume N = 1. The fact that Ψ(G) contains C ∗ (G) as an essential
ideal is a general fact – true for any Lie groupoid. This general fact is true because
it is true for any non-compact manifold, in particular, for each of the manifolds Gx .
We have, by the definitions of Sobolev spaces and of Fredholm operators, that P is
Fredholm if, and only if, a is Fredholm. The fact that a := (1 + ΔF )(s−m)/2 P (1 +
ΔE )−s/2 ∈ Ψ(G; E, F ) follows from the definition of Lm s (G; E, F ), Equation (3)
and the results in [49], as noticed already above. 

This proposition then gives right away the following result (for technical
reasons, we may have to replace P with a in certain proofs). Recall that an operator
P ∈ Ψm (G; E, F ) consists of a right invariant family P = (Px ), x ∈ M , with Px
acting between the sections of r∗ (E) and r∗ (F ) on Gx . In particular, Px is invariant
with respect to the free, proper action of Gxx on Gx by right multiplication. For
simplicity, we shall drop r∗ from the notation below, since there is no danger of
confusion. We have Px = πx (P ). The operators Px , x ∈ M \ U , will be called limit
operators (of P ). They go back to [30].
Theorem 4.12. Let G ⇒ M be a Fredholm Lie groupoid with M compact and
∅ = U ⊂ M , open such that GU = U × U . Let s ∈ R and P ∈ Lm
s (G; E, F ) ⊃
Ψm (G; E, F ). We have
P : H s (M ; E) → H s−m (M ; F ) is Fredholm ⇔ P is elliptic and each
Px : H s (Gx ; E) → H s−m (Gx ; F ) , x ∈ M  U , is invertible .
Proof. Let us use the notation of Proposition 4.11. By replacing P with a and using
Proposition 4.4, we may assume that we work with N × N matrices of operators.
The proof is the same for all N , so we let N = 1. Since G is Fredholm, Theorem
4.6, applied to Ψ := Ψ(G), gives that a ∈ Ψ(G) is Fredholm if, and only if, its
image in Ψ(G)/C ∗ (G) is invertible and all the operators πx (a) are invertible. We
then notice that πx (a) = (1 + Δx )(s−m)/2 Px (1 + Δx )−s/2 since the extension of
πx to operators affiliated to Ψ(G) is given by πx (P ) = Px since this is true for
P ∈ Ψ0 (G) and πx (Δ) = Δx , the Laplacian on Gx by [49]. 
 Fredholm Conditions on Non-compact Manifolds 105

Remark 4.13. To obtain the result as formulated in Theorem 1.2, we let α be the
set of orbits of G acting on F := M  U , M0 := U , Mα := Gx , for some x in the
orbit α, Gα := Gxx , and Pα := πx (P ). It may be useful to notice here that often
the bundle Mα → Mα /Gα is trivial. This is the case, for example, for stratified
submersion groupoids.
Remark 4.14. Let V := r∗ (Γ(A(G)) be the vector fields on M coming from the
infinitesimal action of G ⇒ M on M . We denote by Diff m (V) the set of order m
differential operators on M generated by V and multiplication with functions in
C ∞ (M ) and Diff(V) = ∪m Diff m (V). We denote by Diff m (V; E, F ) the analogous
differential operators acting between sections of vector bundles E, F → M . We
have that all geometric operators (Laplace, Dirac, Hodge, . . . ) associated to a
compatible metric on M (one that comes by restriction from A(G)) belong to
Diff m (V; E, F ) for suitable E, F → M ; moreover, πx (D) for a geometric operator
is of the same type as D. [3].
Remark 4.15. If G is d-connected (in the sense that the fibers of d : G → M are
connected), then the orbits of the vector fields V are the same as the orbits of
G. If x ∈ M and α ⊂ M is its orbit, then this orbit (with its intrinsic manifold
topology) is diffeomorphic to the set Mα /Gα . This makes more explicit the data
in Remark 4.13.
Although we shall not use this in the present paper, let us record the conse-
quence for essential spectra. Notice that we do not need the closure of the unions
for the essential spectra. We denote by σ(Q) the spectrum of an operator Q, de-
fined as the set of λ ∈ C such that Q − λ is not invertible, and by σess (Q) its
essential spectrum, defined as the set of λ ∈ C such that Q − λ is not Fredholm.
Here Q may be unbounded. If P ∈ Ψm (G; E), m > 0, we consider it to be defined
on H m .
Corollary 4.16. Consider the framework of Theorem 4.12 and P ∈ Ψm (G; E),
m ≥ 0, elliptic if m > 0. We have
σess (P ) = ∪x∈∂M σ(Px ) ∪ ∪ξ∈S ∗ A σ(σ0 (P )(ξ)) , if m = 0 ,
σess (P ) = ∪x∈∂M σ(Px ) , if m > 0 .
See also the discussion in [76] for an explanation of the differences between
the cases m = 0 and m > 0 in the above corollary. Note that, if P is not elliptic,
the statement in the case m > 0 is not true anymore, and, in fact, the essential
spectrum depends on more than just the principal symbol, in general. Thus, in the
scalar, order zero case, we have
σess (P ) = ∪x∈∂M σ(Px ) ∪ Im(σ0 (P )) .
Combining Theorems 4.12 and 4.10, we obtain the following result (recall
that all our groupoids are Hausdorff and second countable).
Theorem 4.17. Let M be compact and G ⇒ M be a stratified submersion Lie
groupoid with filtration Ui such that U0 is dense in M and GU0  U0 × U0 . Assume
106 C. Carvalho, V. Nistor and Y. Qiao

that all isotropy groups Gxx are amenable. Then, for any smooth vector bundles
E, F → M and any P ∈ Lm s (G; E, F ) ⊃ Ψ (G; E, F ), we have
m

P : H s (M ; E) → H s−m (M ; F ) is Fredholm ⇔ P is elliptic and all

Px : H (Gx ; E) → H
s s−m
(Gx ; F ) , x ∈ M  U0 , are invertible .
We actually have Px : H s (Gx ; r∗ (E)) → H s−m (Gx ; r∗ (F )), but, as we have
explained above, we drop r∗ from the notation, for the sake of simplicity, since there
is no danger of confusion. We will continue to do that in all related statements
below.
Remark 4.18. If G be as in the above theorem (a stratified submersion groupoid
with filtration Ui ), and assume that it is also d-connected, as in Remark 4.13.
Then, when P = D is a differential operator, we can identify the limit operators
Dα (and hence also Mα and Gα ) in a direct way. Indeed, let α be the V-orbit
= the G-orbit through some x ∈ M  U0 . We have Gα = Gxx , as discussed in
Remark 4.13. Then Mα = α × Gα . Let D ∈ Diff(V). By linearity, we may assume
D = aXi1 Xi2 · · · Xim , with Xi ∈ V a local basis of V near x and a ∈ C ∞ (M ). (We
may also assume that ij form a non decreasing sequence, but that is not necessary.)
Since α is an orbit, we have that this basis (Xj ) consists of vectors tangent to α.
We can choose this basis such that X1 , . . . , Xk form a local basis of Tx α and Xj ,
j > k, are zero on the orbit α (where they are defined). Consequently, the vector
fields Xj , j > k, come from a basis X̃j , j > k, of Lie(Gα ). We then have
Dα = a|α Y1 Y2 · · · Ym , near {x} × Gα ⊂ α × Gα =: Mα , (4)
ij if ij > k. We shall say that the variables
where Yj = Xij |α , if ij ≤ k, and Yj = X
Yj (or Xj ) with j > k are ghost derivatives at the considered orbit α (see Remark
4.18).
See also [8, 9, 100, 106]. It would be interesting to extend the results of this
section to Lp -spaces in view of [1, 82].

5. Examples
In this section, we use the results of the previous section to obtain Fredholm con-
ditions for operators on some standard non-compact manifolds: manifolds with
(poly)cylindrical ends, asymptotically Euclidean manifolds, asymptotically hyper-
bolic manifolds, boundary fibration structures, and others. We tried to write this
section in such a way that it can to a large extent be read independently of the
rest of the paper. The reader interested only in applications can start reading the
paper with this section.
The general setting of this section is that of a (non-compact) smooth man-
ifold M0 whose geometry is determined by a compactification M to a manifold
with corners and a Lie algebra of vector fields V on M . (Although we shall not
use this, let us mention for people familiar with the concept, that (M, V) will be
 Fredholm Conditions on Non-compact Manifolds 107

a Lie manifold with some additional properties.) The differential (and pseudodif-
ferential) operators considered and for which we obtain Fredholm conditions are
the ones generated by V and C ∞ (M ), that is, the ones in Diff(V) and its vari-
ants for vector bundles. (Recall that Diff(V) was introduced in Remark 4.14.) The
proof of the Fredholm conditions in this section are obtained as particular cases
of Theorem 4.12 by showing that appropriate groupoids are Hausdorff stratified
submersion Lie groupoids with amenable isotropy groups and hence that they are
Fredholm groupoids, in view of Theorem 3.13. Many of the results below can also
be obtained from the results in [48, 49], but the approach followed here aims to
be more convenient for non-specialists.
In this section, we continue to denote by G a Lie groupoid with units M ,
a manifold of dimension n. In the theorem yielding Fredholm conditions, M will
be assumed compact and endowed with a smooth metric h. We assume that h
is defined everywhere on M , in particular, that it extends to a smooth manifold
containing M as a submanifold. Also, usually it will be no loss of generality to
assume M connected. Our results are formulated for operators in Lm s (G; E, F ),
which is a suitable completion of Ψm (G; E, F ), see Equations (2) and (3). This
allows us to greatly enlarge the scope of our results since the completion procedure
defining the spaces Lm s leads to algebras that are closed under taking the inverses
of L2 -invertible elements.

5.1. Examples related to group actions

We include first some examples that are closely related to group actions.

5.1.1. The action of a group on a space. Let us assume that Lie group G acts
smoothly on a manifold M . This yields the transformation (or group action) Lie
groupoid G := M  G, which, as a set, consists of M × G and has units M . We
have
d(x, g) := x , r(x, g) := gx , and (hx, g)(x, h) := (x, gh) .
If M = G with G acting by translations, then M  G  G × G, the product
groupoid. Therefore, if G ⊂ M as a dense, G-invariant open set, and G acts on
itself by left translation, we are in the setting in which we can ask if the resulting
groupoid is Fredholm Lie groupoid. This groupoid was used in [32, 57, 67]. The
groupoids used in [32, 67] turn out to be stratified submersion groupoids. As
discussed in those papers, this recovers the classical HVZ-theorem [21, 84, 102] as
a particular case of Theorem 4.12 (with U = G). We note that M × G is always
Hausdorff (since M and G are Hausdorff). If V ⊂ M is an open subset, then the
reduction groupoid (M  G)VV will be called a local transformation (or action)
groupoid, and will also be Hausdorff. Many related results (including the non
Lie case), were also obtained by Böttcher, Chandler-Wilde, Karlovich, Lindner,
Rabinovich, Roch, Rozenblyum, Silberman, and many others. See [6, 10, 11, 16,
59, 81–83, 89] and the references therein.
This groupoid models pseudodifferential operators compatible with any G
invariant metric on G. Let g be the Lie algebra of G. Then A(G) = M × g, with g
108 C. Carvalho, V. Nistor and Y. Qiao

acting on M via the infinitesimal action of G. The associated Lie algebra of vector
fields is Lie(G) := (Γ(A(G))) = V := C ∞ (M )g ⊂ Γ(T Ω). We have that Diff(V) is
generated by C ∞ (M ) and g.

5.1.2. The b-groupoid. Let M be a manifold with corners. Then the b-groupoid
of [49, 61, 65, 77] is defined as a set as the disjoint union

Gb := F (F × F ) × (R∗+ )kF  F (F × F ) × RkF ⇒ FF = M ,

where F ranges through the open, connected faces of M , kF is the codimension of

the face F , F × F is the pair groupoid and (R∗+ )kF is a group for componentwise
multiplication (and hence also a Lie groupoid).
To obtain the smooth structure on this groupoid, we notice that, locally, it is
a transformation groupoid [0, ∞)k  (R∗+ )k . More precisely, let us choose g ∈ Gb .
Then g = (x, y, v) ∈ (F × F ) × RkF for some connected open face F ⊂ M . We can
choose a coordinate system V ⊂ M such that x, y ∈ V , V is compact in F , and
we have a tubular neighborhood V × [0, )kF ⊂ M . Then, a neighborhood of g in
Gb is diffeomorphic to the local transformation groupoid obtained by reducing to
V × [0, )kF the action groupoid

(RdF × [0, ∞)kF )  (RdF × (0, ∞)kF ) ,

where dF = n − kF is the dimension of F . Here RdF acts by translations on itself

and R∗+ = (0, ∞) acts by multiplication on [0, ∞). In applications, it will be, in
fact, more convenient to notice that RdF  RdF is the pair groupoid, and hence to
identify a neighborhood of g with a reduction of

(V × V ) × [0, ∞)kF  (0, ∞)kF ,

the product of the pair groupoid V ×V and the transformation groupoid [0, ∞)kF 
(0, ∞)kF .
If M has embedded faces, that is, if each hyperface H has a defining function
rH , then we can identify Gb with an open subset of Monthubert’s realization of
the b-groupoid [66, Proposition 4.5]

M := { (x, y, t) ∈ M × M × [−1, 1]H | (1 − tK )xK (x) = (1 + tH )xH (y) } ,

where H denotes the set of hyperfaces of M and H, K ∈ H. See also [13, 14, 51,
52, 104].
We see that the b-groupoid is a stratified submersion groupoid as follows. The
set Uk , k ≤ n, is defined as the union of the open faces of codimension k of M .
Then (Gb )Uk \Uk−1 is isomorphic to the topological disjoint union of the groupoids
(F × F ) × Rk , where F ranges through the set of open faces of codimension k. In
particular, (Gb )Uk \Uk−1 is isomorphic to the fibered pull-back of a bundle of Lie
groups, by Example 2.20. In particular, (Gb )xx ∼
= Rn−k is amenable, for all x ∈ M .
 Fredholm Conditions on Non-compact Manifolds 109

5.1.3. Manifolds with poly-cylindrical ends. The Lie algebroid of Gb is identified by

Γ(A(Gb ))  Vb := {X ∈ Γ(M ; T M )| X tangent to all faces of M } . (5)
Let h be an ordinary metric on M . The general form of a compatible metric on
M is then
  dxH 2
gb := h + . (6)
xH
H∈H
Manifolds with metrics of this form are called manifolds with poly-cylindrical ends,
following [63]. The Lie algebroid A(Gb ) is often denoted T b M . The groupoid Gb
models pseudodifferential operators compatible with the metric gb . Thus, by def-
inition, the metric gb comes from restriction from the Lie algebroid A(Gb ), since
A(Gb )|M0 = T M0 . Since the base M is compact, all metrics on A(Gb ) will be equiv-
alent, so the constructions will not depend significantly on the choice of the metric.
In particular, all geometric operators associated to the metric gb (Laplace, Dirac,
Hodge, . . . ) will belong to Ψm (Gb ; E, F ) (which is independent of the metric), for
suitable E and F . Moreover, it turns out that Theorem 4.17 applies to Gb and
Ψm (Gb ; E, F ), provided that M is compact. (See also Remark 4.14. Further details
can be found in [3].)
The statement of Theorem 4.5 can be (slightly) simplified in this case by
noticing the following. The representations πx and πy are unitarily equivalent if x
and y are in the same open face F , in which case, they will act on F × RkF  Gx .
Let πF be the associated representation. For Fredholm conditions, it is enough to
consider the invertibility of the operators PH := πx (P ), x ∈ H, for the faces of
maximal dimension (that is, for hyperfaces) in order to obtain Fredholm conditions,
since πx is weakly contained in πy if x is contained in the closure of the face
containing y. Let H denote the set of hyperfaces of M , as before. Recall that
Lm m
s (G; E, F ) is a suitable completion of Ψ (G; E, F ), see Equations (2) and (3).
We obtain the following result [62].
Corollary 5.1. Let P ∈ Lm
s (Gb ; E, F ) ⊃ Ψ (Gb ; E, F ), M compact. We have
m

P : H s (M ; E) → H s−m (M ; F ) is Fredholm ⇔ P is elliptic and all

PH : H (H × R; E) → H
s s−m
(H × R; F ) , H ∈ H , are invertible .
The operators PH are invariant with respect to translations by t ∈ R acting on
the R component. All geometric differential operators associated to the metric gb
belong to Diff m (Vb ; E, F ) ⊂ Ψm (Gb ; E, F ), for suitable vector bundles E, F → M .
If P ∈ Diff m (Vb ; E, F ), the limit operators Px , x ∈ ∂M , are obtained as
explained in Remark 4.18. They are invariant with respect to the action of the
isotropy group Gxx (this is always the case for operators of the form πx (a)). In this
example, only derivatives of the form xH ∂xH are ghost derivatives (see Remark
4.18).
Remark 5.2. Carvalho and Qiao have constructed in [15] a similar groupoid to
the b-groupoid in order to study layer potentials. Their groupoid, however, was
110 C. Carvalho, V. Nistor and Y. Qiao

not d-connected, in general. Nevertheless, the above corollary generalizes to their

setting, after some obvious modifications.
5.1.4. Asymptotically Euclidean spaces. Let us assume that M has a smooth
boundary ∂M with defining function x = x∂M and let Vsc := xVb . The result-
ing differential operators Diff(Vsc ) and the associated pseudodifferential operators
are the SG-operators of [20, 27, 78, 91–93] (called “scattering operators” in [63]).
They can be obtained by considering the groupoid
Gsc := T M |∂M (M0 × M0 ) ⇒ ∂M M0 = M .
To obtain a manifold structure on Gsc , let us consider first G = Rn and M = the
radial compactification of G with the induced action of G. Then Gsc = M  G is
Hausdorff and a stratified submersion groupoid. In general, Gsc is locally of this
form (and can be obtained by gluing reductions of such groupoids), and hence it is
Hausdorff. It satisfies Lie(Gsc ) := Γ(A(Gsc )) = Vsc . Thus, if gb denotes a b-metric
on M (or rather, on M0 , then the natural metric associated to Vsc is
gsc := x−2 gb := x−2 h + x−2 dx2 , (7)
where gb is as in Equation (6).
We have that the orbits of G on F := M U0 = ∂M are reduced to points and
that each stabilizer Gxx = Tx M  Rn , for x ∈ F . In particular, Gx = Gxx , for x ∈ ∂M
and all derivatives at the boundary are ghost derivatives. If P ∈ Diff m (Vsc ; E, F ),
the limit operators Px are obtained as explained in Remark 4.18. Theorem 4.17
becomes in our case:

s (Gsc ; E, F ) ⊃ Ψ (Gsc ; E, F ), M compact. We have

Corollary 5.3. Let P ∈ Lm m

P : H s (M ; E) → H s−m (M ; F ) is Fredholm ⇔ P is elliptic and all

Px : H (Tx M ; E) → H
s s−m
(Tx M ; F ) , x ∈ ∂M , are invertible .
The operators Px are constant coefficient (i.e., Px is invariant with respect to the
action of Tx M on itself by translations). All geometric differential operators asso-
ciated to the metric gsc belong to Diff m (Vsc ; E, F ) ⊂ Ψm (Gsc ; E, F ), for suitable
vector bundles E, F → M .
We have Px = πx (P ) and Px is translation invariant (i.e., constant coefficient,
in this case), and hence can be studied using the Fourier transform. In this case,
all the vector fields yield ghost derivatives.

5.1.5. Asymptotically hyperbolic manifolds. We continue to assume that M is a

manifold with smooth boundary ∂M . The groupoid Gah modeling asymptotically
hyperbolic spaces is chosen such that
Lie(Gah )  Γ(A(Gah )) = V0 := {X ∈ Γ(M ; T M )| X|∂M = 0} .
Let Lx := Tx (∂M )  (0, ∞) be the semi-direct product as in the previous subsec-
tion, with (0, ∞) acting by dilations on Tx (∂M ). Let L → ∂M be the bundle of
 Fredholm Conditions on Non-compact Manifolds 111

Lie groups with fiber Lx . Then

Gah := L M0 × M0 ⇒ ∂M M0 .
The topology is again locally given by a transformation groupoid. This can be
seen in the case of M = Rn−1 × [0, ∞) with the natural action of the semi-direct
product (cross-product) Gn := Rn−1 (0, ∞) obtained by recalling that, as smooth
manifolds, we have Rn−1  (0, ∞) = Rn−1 × (0, ∞). The action of (0, ∞) on Rn−1
is by dilations. This groupoid is a particular case of the edge groupoid, following
next, so the reader can consult he next section for more details. We again have
that Gx = Gxx = Lx , if x ∈ ∂M , and that all derivatives are ghost derivatives at
the boundary. The metric is
gah := x−2 h ,
where x is the distance to the boundary (close to the boundary) and h is an
everywhere smooth metric on M , as before.
Corollary 5.4. Let s ∈ R and P ∈ Lm
s (Gah ; E, F ) ⊃ Ψ (Gah ; E, F ), with M com-
m

pact. Denote Mx := Tx ∂M  R = (Gah )x . We have

P : H s (M ; E) → H s−m (M ; F ) is Fredholm ⇔ P is elliptic and all
Px : H (Mx ; E) → H
s s−m
(Mx ; F ) , x ∈ ∂M , are invertible .
The operator Px is invariant with respect to the action of Mx on itself by right
translations, for each x ∈ ∂M . All geometric differential operators associated to
the metric gah belong to Diff m (Vah ; E, F ) ⊂ Ψm (Gah ; E, F ), for suitable vector
bundles E, F → M .
We thus need to study the invertibility of certain (right) invariant operators
on the group Gn−1 := Rn−1 (0, ∞), for which standard methods of representation
theory can be used.
5.2. The edge groupoid
This example is motivated by the results in [36, 38, 58, 75, 95], where the original
Fredholm results on the edge calculus can also be found. It is a particular case of
the next example, that of a desingularization groupoid, but we nevertheless treat
it separately, for the benefit of the reader. See also [23, 47, 107]. We consider the
following framework.
First, M is a manifold with smooth boundary ∂M . We assume that we are
given a smooth fibration π : ∂M → B, where B is a smooth manifold (thus without
boundary). We fix a tubular neighborhood U of ∂M in M : U  ∂M × [0, 1). Let
H := B × B be the pair groupoid. We now construct the so-called edge groupoid
Ge that will turn out to be a groupoid with Lie algebroid given by the set Ve of
vector fields on M that are tangent to the fibers of π : ∂M → B (in particular,
these vector fields are tangent to the boundary). If B = ∂M , then we recover the
groupoid that models asymptotically hyperbolic spaces. We denote by Diff(Ve )
the algebra of differential operators generated by Ve and by multiplication with
functions in C ∞ (M ).
112 C. Carvalho, V. Nistor and Y. Qiao

Let L := T B  R∗+ → B be the bundle of Lie groups obtained from T B →

B (regarded as a bundle of commutative Lie groups) by taking the semi-direct
product with R∗+ := (0, ∞) acting by dilations on the fibers of T B → B. Its
pullback π ↓↓ (L) via π : ∂M → B is hence a Lie groupoid with units ∂M . Let
M0 := M  ∂M be the interior of M . Then, as a set, the edge groupoid Ge is the
disjoint union
Ge := π ↓↓ (L) (M0 × M0 ) ⇒ ∂M M0 .
To define the smooth structure on this groupoid, we could use either the results
in [73] or proceed directly in four steps as follows.
Step 1. We first consider the adiabatic groupoid Had of H := B × B (this is
the tangent groupoid of [17]; see [75] for more details and references). The adiabatic
groupoid Had is a Lie groupoid with units B × [0, ∞) and Lie algebroid A(Had ) =
T B × [0, ∞) → B × [0, ∞), which, as a vector bundle, is the fibered pull-back
of A(H) = T B → B to B × [0, ∞) via the projection B × [0, ∞) → B. The Lie
algebroid structure on the sections of A(Had ) is not that of a fibered pull-back Lie
algebroid, but is given by [X, Y ](t) = t[X(t), Y (t)]. As a set, Had is the disjoint
union
Had := A(H) × {0} H × (0, ∞) .
The groupoid structure of Had is such that A(H) × {0} has the Lie groupoid
structure of a bundle of Lie groups and H × (0, ∞) has the product Lie groupoid
structure with (0, ∞) the groupoid associated to a space (that is (0, ∞) has only
units, and all orbits are reduced to a single point). The smooth structure is ob-
tained using the exponential map. See also [105].
Step 2. Let π : ∂M → B be the given fibration map. We denote also by π
the resulting map ∂M × [0, ∞) → B × [0, ∞). Then we consider the pullback Lie
groupoid π ↓↓ (Had ).
Step 3. Let R∗+ = (0, ∞) act by dilations on the [0, ∞) variable on π ↓↓ (Had )
and consider the semi-direct product π ↓↓ (Had ) R∗+ [26]. As a set, it is the disjoint
union of π ↓↓ (L) and of the pair groupoid of ∂M × (0, ∞).
Step 4. Let us identify the tubular neighborhood U ⊂ M of ∂M with ∂M ×
U
[0, 1). Then we can consider the reduction H := π ↓↓ (Had )  R∗+ U . By the previ-
ous step, this reduction H is the disjoint union of π ↓↓ (L) and of the pair groupoid
of ∂M ×(0, 1), which we can view as a subset of M0 ×M0 , the pair groupoid of M0 .
U
We then glue the reduction H := π ↓↓ (Had )  R∗+ U with M0 × M0 by identifying

the reduction of H to U  ∂M (which is the pair groupoid of ∂M × (0, 1), as
we have seen), with its image in M0 × M0 . This gluing construction is, of course,
nothing but a particular case of the gluing construction in [39] and [75].
In any case, we obtain right away from the definition that the edge groupoid
Ge is a Hausdorff stratified submersion groupoid. The set of units of Ge is M and
the representations πx , x ∈ ∂M , are equivalent precisely when they map to the
same point in B and they act on π −1 (b) × Lb  π −1 (b) × Tb B × R.
 Fredholm Conditions on Non-compact Manifolds 113

If P ∈ Diff(Ve ) and b ∈ B, we can obtain the limit operators Pb := πb (P ) as

follows. The restriction of P to an infinitesimal neighborhood of π −1 (b) in M will
have some hidden (ghost) derivatives coming from the Lie algebra of the group
Lb := Tb B  R∗+ . Let us choose local coordinates (y, z) on ∂M that are compatible
with ∂M → B, in the sense that z comes from a coordinate system on B. Let
x denote the defining function of ∂M . Then locally, Ve is generated by x∂x , ∂yj ,
x∂zk , with j and k (always) ranging through a suitable index set:
 
Ve = C ∞ (M )x∂x + C ∞ (M )∂yj + C ∞ (M )x∂zk .
j k

With this notation, the Lie algebra Lie(Lb ) is generated by the vector fields
x∂x and x∂zk . These come from non-zero vector fields in Ve that restrict to 0 on
π −1 (b), they are the ghost derivatives. We let the ghost derivatives act on Lb ,
and thus we obtain a differential operator Pb on π −1 (b) × Lb . It can be obtained
by patching together metrics for which x∂x , ∂yj , x∂zk are an orthonormal set of
vectors.
Theorem 4.12 then becomes.
Corollary 5.5. Let P ∈ Lm
s (Ge ; E, F ) ⊃ Ψ (Ge ; E, F ), M compact. We have
m

P : H s (M ; E) → H s−m (M ; F ) is Fredholm ⇔ P is elliptic and all

Pb : H (Mb ; E) → H
s s−m
(Mb ; F ) , b ∈ B , are invertible ,
−1
where Mb := π (b) × Tb B  R. The operator Pb is invariant with respect to the
action of Tb BR by right translations on itself. All geometric differential operators
associated to ge belong to Diff m (Ve ; E, F ) ⊂ Ψm (Ge ; E, F ), for suitable E, F → M .
If B is reduced to a point and ∂M is connected, the groupoid Ge constructed
in the last subsection recovers the groupoid of the b-calculus: Ge = Gb . It models in
this case manifolds with cylindrical ends. If B = ∂M , the corresponding groupoid
models “asymptotically hyperbolic” spaces.
5.3. Desingularization groupoids
One of the nice features of the class of stratified submersion groupoids is that it
is invariant with respect to desingularization along suitable submanifolds. In this
subsection, we will give an ad hoc argument for this statement. Recall the fibered
pull-back π ↓↓ (B) of Example 2.19. We proceed in a slightly greater generality than
in [75], to which we refer for more details and for the unexplained arguments.
Let us assume that M is a manifold with corners, that H is a hyperface of M
and that we are given a tame submersion of manifolds with corners π : H → B.
Let U  H × [0, 1) be a tubular neighborhood of H in M . The hyperface H will
play the role played by the boundary in the previous examples. Typically, M will
be the result of a desingularization procedure, such as a blow-up (see Remark 5.6
and the following section).
We assume that we are given a Lie groupoid G on M  H and a Lie groupoid
H ⇒ B such that the following is satisfied.
114 C. Carvalho, V. Nistor and Y. Qiao

Local fibered pull-back structure assumption: The reduction of G to H × (0, 1) is

isomorphic to the pullback p↓↓ (H) via the map p := π ◦ p1 : H × (0, 1) → B.
We fix H as in the above assumption throughout this subsection. We define
then a groupoid K ⇒ M such that, as a set, it is the disjoint union
K := π ↓↓ (A(H)  R∗+ ) G⇒H M  H.
The groupoid H replaces the pair groupoid B × B of the example of the edge
groupoid. We proceed as in that example to consider the adiabatic groupoid Had
of H, which has units B×[0, ∞). We pull back this groupoid to a groupoid π ↓↓ (Had )
using the map π1 : H × [0, ∞) → B × [0, ∞), and then we consider the semi-direct
product π1↓↓ (Had )  R∗+ = π1↓↓ (Had  R∗+ ) with R∗+ acting by dilations on [0, ∞).
As in the previous example of the edge groupoid, we view U as an open
subset of H × [0, ∞) and we consider the groupoid H defined as the reduction
to U of π ↓↓ (Had )  R∗+ , with H as in the local fibered pull-back assumption. The
construction is completed as in the fourth step of the edge groupoid. That is,
we use the invariant subset H × {0} of the units of H to write the groupoid as
a disjoint union using Remark 2.7. The reduction of H to the complement of
H × {0}, that is, to H × (0, 1), is (isomorphic to) the fibered pull-back p↓↓ (H) of
H to H × (0, 1), by the local fibered pull-back structure assumption. We can view
this fibered pull-back p↓↓ (H) as a subset of G. This gives that we can glue H and
G along the common open groupoid p↓↓ (H) to obtain a groupoid K.
It can be proved that if G is Fredholm, then K is also Fredholm. Moreover, if
G is a stratified submersion groupoid, then K will also be a stratified submersion
groupoid.
Remark 5.6. Typically, we start with a Lie groupoid G  ⇒ M  and L ⊂ M  is a
submanifold with corners. Then M := [M  : L], the blow-up of M  with respect to
L and H is the hyperface corresponding to L in this blow-up, with π : H → L the
blow-down map. Finally, G is the reduction of G  to the complement of L. Then
we denote [[G  : L]] := K. This is the desingularization construction from [75].
5.4. Desingularization and singular spaces
Let us show how to use the desingularization construction in some typical exam-
ples. By Mk and Lk we will denote manifolds with corners of depths k. Thus M0
and L0 will have, in fact, no corners or boundary (hence they will be “smooth”).

5.4.1. The desingularization of a smooth submanifold. The simplest example of

a desingularization groupoid is the desingularization of the pair groupoid G0 :=
M0 × M0 with respect to a smooth submanifold L0 ⊂ M0 . Recall that M0 is also
smooth. Thus neither M0 nor L0 have corners. The example of this subsection is
a particular case of the edge groupoid, and hence it is related to the edge calculus.
[38, 58, 70, 95].
Let N be the normal bundle of L0 in M0 and denote by S ⊂ N the set of unit
vectors in N , that is, S is the unit sphere bundle of the normal bundle of L0 in M0 .
 Fredholm Conditions on Non-compact Manifolds 115

We let π : S → L0 be the natural projection. Then the blow-up M1 := [M0 : L0 ]

of M0 with respect to L0 is the disjoint union
M1 := [M0 : L0 ] := (M0  L0 ) S,
with the structure of a manifold with smooth boundary S. We let G1 to be the
associated edge groupoid introduced in Subsection 5.2. It is a Lie groupoid with
base M1 := [M0 : L0 ]. We are moreover in the framework of Remark 5.6, so
G1 = [[G0 : L0 ]]. The filtration of M1 has two sets, namely, U1 = M1 and U0 :=
M0 L0 ⊂ M1 , both of which are open and invariant for G1 (but U0 is not invariant
for G0 , in general).
It turns out that G1 is a stratified submersion Lie groupoid if we consider the
two strata U0 and S. Also, (G0 )U U0 = U0 × U0 and GS is the fibered pull-back of a
0

bundle of Lie groups L → S, as described in Subsection 5.2.

5.4.2. The desingularization of a submanifold with boundary. We now generalize

the last construction to manifolds with boundary. Let M1 be a compact manifold
with smooth boundary. We denote by F := ∂M1 its boundary and by G := M1 F
its interior. On M1 we consider the Lie algebra of vector fields Vb tangent to ∂M1 ,
as before. It is the space of sections of T b M1 , the “b-tangent bundle” [61, 63] of
M1 . Let G1 := Gb , as defined in the previous section.
Let L1 ⊂ M1 be an embedded smooth submanifold assumed to be such that
its boundary is ∂L1 = L1 ∩ ∂M1 and such that L1 intersects ∂M1 transversely.
Then L1 has a tubular neighborhood U in M1 , and hence we can consider the blow-
up M2 := [M1 : L1 ], which will be a manifold with corners of rank 2. Moreover,
the reduction of G1 = Gb to U  L1 satisfies the local fibered pull-back structure
assumption, since it is the b-groupoid of U  L1 . In view of Remark 5.6, we can
then define G2 := [[G1 : L1 ]].
Again it turns out that G2 is a stratified submersion Lie manifold; indeed,
this is seen by choosing the following filtration of M2 with three sets:
U0 := M1  (L1 ∪ ∂M1 ) ⊂ U1 := M1  L1 ⊂ U2 := M2 .
The sets Uj are open and G2 invariant (but not G1 invariant). The sets U0 and
U1 are G2 -invariant by construction. Assume they are connected, for simplicity.
The restriction (G2 )U1 coincides with the reduction (G1 )U1
U1 by the definition of
the desingularization groupoid (Remark 5.6). In particular, (G2 )U0 = (G2 )U0
U0 =
(G1 )U0 2
U0 = U0 and
1 U0 ∗
U1 U0 = (U1  U0 ) × R+ ,
(G2 )U1 U0 := (G1 )U 2

where U02 and (U1  U0 )2 are pair groupoids. In the general case, if U1  U0 is
not connected, we write U1  U0 = Vj as the disjoint union decomposition of its
connected components, then we have (G2 )U1 U0 := j Vj2 × R∗+ . Let us denote as
in the boundaryless case by S the unit sphere bundle of the normal bundle to L1
in M1 . (The set S is the last stratum U2  U1 in M2 .) Then the restriction of G1
to S := M2  U1 is the following fibered pull-back groupoid. Let π : S → L be
116 C. Carvalho, V. Nistor and Y. Qiao

the natural projection, as before. Let again GS := T b L  R∗+ → L be the group

bundle over L obtained by taking the direct product of the b-tangent bundle to L
with the action of R∗+ by dilation on the fibers of T L → L. Then
GS := π ↓↓ (GS ) .
5.4.3. Desingularization of a stratified subset of dimension one. We now deal with
a slightly more complicated example by combining the two previous examples.
We thus introduce the groupoid that is obtained from the desingularization of
a stratified subset of dimension one. Full details as well as applications will be
included in a forthcoming paper with Mihai Putinar.
Let M0 be a smooth, compact manifold (so no corners). Let
L0 := {P1 , P2 , . . . , Pk } ⊂ M0
and let us assume that we are given a subset S ⊂ M0 such that
S = L0 ∪ ∪lj=1 γj ,
where each γj is the image of a smooth map cj : [0, 1] → M0 , with the following
properties:
(i) cj (t) = 0,
(ii) cj (0), cj (1) ∈ L0 := {P1 , P2 , . . . , Pk },
(iii) cj ((0, 1)) are disjoint and do not intersect L0 and
(iv) the vectors cj (0) and cj (1), j = 1, . . . , l, are all distinct.
We now introduce the desingularization of G0 := M0 ×M0 (the pair groupoid)
with respect to S. The set L0 := {P1 , P2 , . . . , Pk } ⊂ M0 satisfies the assumptions
of 5.4.1. We can first define G1 to be the desingularization of G0 with respect to
L0 as in that example:
G1 := [[G0 : L0 ]] ,
which is a groupoid with units M1 := [M0 : L0 ]. The smooth maps cj then lift to
smooth maps
c̃j : [0, 1] → [M0 : L0 ] .
The assumption that the vectors cj (0) and cj (1), j = 1, . . . , l, are all distinct then
gives that the sets γ̃j := c̃j ([0, 1]) are all disjoint and intersect the boundary of
M1 transversely. Let L1 be the disjoint union of the embedded curves γ̃j . Then we
can perform a further desingularization along L1 , as in 5.4.2, thus obtaining
G2 := [[G1 : L1 ]] ,
which is a boundary fibration Lie groupoid. The Lie groupoid G2 is the desingular-
ization groupoid of M0 with respect to S. Its structure is given as in the previous
subsection.
This example can be extended to higher-dimensional cases by using clean
intersecting families.
Acknowledgment. We thank Vladimir Georgescu, Marius Mantoiu, Jean Renault,
and Wolfgang Schulze for useful discussions.
 Fredholm Conditions on Non-compact Manifolds 117

References
[1] B. Ammann and N. Große, Lp -spectrum of the Dirac operator on products with
hyperbolic spaces. Calc. Var. Partial Differ. Equ. 55 (2016), 127–163.
[2] B. Ammann, A.D. Ionescu, and V. Nistor, Sobolev spaces on Lie manifolds and
regularity for polyhedral domains. Doc. Math. 11 (2006), 161–206.
[3] B. Ammann, R. Lauter, and V. Nistor, On the geometry of Riemannian manifolds
with a Lie structure at infinity. Int. J. Math. Math. Sci. 2004 (2004), 161–193.
[4] B. Ammann, R. Lauter, and V. Nistor, Pseudodifferential operators on manifolds
with a Lie structure at infinity. Ann. of Math. (2) 165 (2007), 717–747.
[5] B. Ammann, R. Lauter, V. Nistor, and A. Vasy, Complex powers and non-compact
manifolds. Comm. Partial Differ. Equ. 29 (2004), 671–705.
[6] M.A. Bastos, C. Fernandes, and Yu.I. Karlovich. A C ∗ -algebra of singular integral
operators with shifts admitting distinct fixed points. J. Math. Anal. Appl. 413 (2014),
502–524.
[7] I. Beltiţă, D. Beltiţă, and M. Măntoiu, Symbol calculus of square-integrable operator-
valued maps. Rocky Mountain J. Math. 46 (2016), 1795–1851.
[8] K. Bohlen, Boutet de Monvel operators on singular manifolds. C. R. Math. Acad.
Sci. Paris 354 (2016), 239–243.
[9] K. Bohlen, Boutet de Monvel operators on Lie manifolds with boundary. Adv. Math.
312 (2017), 234–285.
[10] A. Böttcher, Yu.I. Karlovich, and I.M. Spitkovsky, The C ∗ -algebra of singular in-
tegral operators with semi-almost periodic coefficients. J. Funct. Anal. 204 (2003),
445–484.
[11] A. Böttcher and B. Silbermann, Analysis of Toeplitz Operators. 2nd ed. Springer,
Berlin, 2006.
[12] M. Buneci, Groupoid C ∗ -algebras. Surv. Math. Appl. 1 (2006), 71–98.
[13] P. Carrillo Rouse and J.-M. Lescure, Geometric obstructions for Fredholm boundary
conditions for manifolds with corners. arXiv:1703.05612 [math.DG], 2017.
[14] P. Carrillo Rouse, J.M. Lescure, and B. Monthubert, A cohomological formula
for the Atiyah–Patodi–Singer index on manifolds with boundary. J. Topol. Anal.
6 (2014), 27–74.
[15] C. Carvalho and Yu Qiao, Layer potentials C ∗ -algebras of domains with conical
points. Cent. Eur. J. Math. 11 (2013), 27–54.
[16] S. Chandler-Wilde and M. Lindner, Limit operators, collective compactness, and
the spectral theory of infinite matrices. Mem. Amer. Math. Soc. 989 (2011).
[17] A. Connes, Non commutative differential geometry. Publ. Math. IHES 62 (1985),
41–144.
[18] H. Cordes, Spectral theory of linear differential operators and comparison algebras.
Cambridge University Press, Cambridge, 1987.
[19] H.O. Cordes and E.A. Herman, Gelfand theory of pseudo differential operators.
Amer. J. Math. 90 (1968), 681–717.
[20] S. Coriasco and L. Maniccia, On the spectral asymptotics of operators on manifolds
with ends. Abstr. Appl. Anal. (2013), Art. ID 909782, 21 pages.
118 C. Carvalho, V. Nistor and Y. Qiao

[21] H.L. Cycon, R.G. Froese, W. Kirsch, and B. Simon, Schrödinger Operators, with
Application to Quantum Mechanics and Global Geometry. Springer, Berlin, 1987.
[22] M. Damak and V. Georgescu, Self-adjoint operators affiliated to C ∗ -algebras. Rev.
Math. Phys. 16 (2004), 257–280.
[23] A. Dasgupta and M.W. Wong, Spectral theory of SG pseudo-differential operators
on Lp (Rn ). Studia Math. 187 (2008), 185–197.
[24] M. Dauge, Elliptic Boundary Value Problems on Corner Domains. Smoothness and
Asymptotics of Solutions. Springer, Berlin, 1988.
[25] C. Debord, J.-M. Lescure, and F. Rochon, Pseudodifferential operators on manifolds
with fibred corners. Ann. Inst. Fourier 65 (2015), 1799–1880.
[26] C. Debord and G. Skandalis, Adiabatic groupoid, crossed product by R∗+ and pseu-
dodifferential calculus. Adv. Math. 257 (2014), 66–91.
[27] A. Degeratu and M. Stern, Witten spinors on nonspin manifolds. Comm. Math.
Phys. 324 (2013), 301–350.
[28] J. Dixmier, Les C ∗ -Algèbres et Leurs Représentations. Éditions Jacques Gabay,
Paris, 1996.
[29] R. Exel, Invertibility in groupoid C ∗ -algebras. In: “Operator Theory, Operator Al-
gebras and Applications”. Oper. Theory Adv. Appl. 242 (2014), 173–183.
[30] J. Favard, Sur les équations différentielles linéaires à coefficients presque-pério-
diques. Acta Math. 51 (1928), 31–81.
[31] V. Georgescu and A. Iftimovici, Crossed products of C ∗ -algebras and spectral anal-
ysis of quantum Hamiltonians. Comm. Math. Phys. 228 (2002), 519–560.
[32] V. Georgescu and V. Nistor, On the essential spectrum of N -body Hamiltonians with
asymptotically homogeneous interactions. J. Operator Theory 77 (2017), 333–376.
[33] C. Gérard, O. Oulghazi, and M. Wrochna, Hadamard states for the Klein-Gordon
equation on Lorentzian manifolds of bounded geometry. Comm. Math. Phys. 352
(2017), 519–583.
[34] C. Gérard and M. Wrochna, Hadamard states for the linearized Yang–Mills equation
on curved spacetime. Comm. Math. Phys. 337 (2015), 253–320.
[35] J. Gracia-Bondı́a, J. Várilly, and H. Figueroa, Elements of Noncommutative Geom-
etry. Birkhäuser Boston, Inc., Boston, MA, 2001.
[36] D. Grieser and E. Hunsicker, Pseudodifferential operator calculus for generalized
Q-rank 1 locally symmetric spaces. I. J. Funct. Anal. 257 (2009), 3748–3801.
[37] N. Große and C. Schneider, Sobolev spaces on Riemannian manifolds with bounded
geometry: general coordinates and traces. Math. Nachr. 286 (2013), 1586–1613.
[38] V.V. Grušin, A certain class of elliptic pseudodifferential operators that are degen-
erate on a submanifold. Mat. Sb. (N.S.) 84 (1971), 163–195.
[39] M. Gualtieri and Songhao Li, Symplectic groupoids of log symplectic manifolds. Int.
Math. Res. Not. IMRN 2014 (2014), 3022–3074.
[40] Ph. Higgins and K. Mackenzie, Algebraic constructions in the category of Lie alge-
broids. J. Algebra 129 (1990), 194–230.
[41] Ph. Higgins and K. Mackenzie, Fibrations and quotients of differentiable groupoids.
J. London Math. Soc. (2) 42 (1990), 101–110.
 Fredholm Conditions on Non-compact Manifolds 119

[42] L. Hörmander, The Analysis of Linear Partial Differential Operators. III Pseudo-
Differential Operators. Springer, Berlin, 2007.
[43] M. Ionescu and D. Williams, The generalized Effros–Hahn conjecture for groupoids.
Indiana Univ. Math. J. 58 (2009), 2489–2508.
[44] M. Khoshkam and G. Skandalis, Regular representation of groupoid C ∗ -algebras and
applications to inverse semigroups. J. Reine Angew. Math. 546 (2002), 47–72.
[45] V.A. Kondrat ev, Boundary value problems for elliptic equations in domains with
conical or angular points. Transl. Moscow Math. Soc. 16 (1967), 227–313.
[46] V. Kozlov, V. Maz ya, and J. Rossmann, Spectral Problems Associated with Corner
Singularities of Solutions to Elliptic Equations. American Mathematical Society,
Providence, RI, 2001.
[47] T. Krainer, A calculus of abstract edge pseudodifferential operators of type ρ, δ. In:
“Elliptic and Parabolic Equations. Selected papers based on the presentations at
the workshop, Hannover, Germany, September 10–12, 2013”. Springer Proceedings
in Mathematics & Statistics 119 (2015), 179–207.
[48] R. Lauter, B. Monthubert, and V. Nistor, Pseudodifferential analysis on continuous
family groupoids. Doc. Math. 5 (2000), 625–655.
[49] R. Lauter and V. Nistor, Analysis of geometric operators on open manifolds: a
groupoid approach. In: “Quantization of Singular Symplectic Quotients”. Progr.
Math., vol. 198, pp. 181–229. Birkhäuser, Basel, 2001.
[50] M. Lein, M. Măntoiu, and S. Richard, Magnetic pseudodifferential operators with
coefficients in C ∗ -algebras. Publ. Res. Inst. Math. Sci. 46 (2010), 755–788.
[51] M. Lesch, Operators of Fuchs Type, Conical Singularities, and Asymptotic Methods.
Teubner Verlagsgesellschaft mbH, Stuttgart, 1997.
[52] M. Lesch and B. Vertman, Regularizing infinite sums of zeta-determinants. Math.
Ann. 361 (2015), 835–862.
[53] K. Mackenzie, Lie Groupoids and Lie Algebroids in Differential Geometry. Cam-
bridge University Press, Cambridge, 1987.
[54] K. Mackenzie, General theory of Lie groupoids and Lie algebroids. Cambridge Uni-
versity Press, Cambridge, 2005.
[55] M. Măntoiu, C ∗ -algebras, dynamical systems at infinity and the essential spectrum
of generalized Schrödinger operators. J. Reine Angew. Math. 550 (2002), 211–229.
[56] M. Măntoiu, R. Purice, and S. Richard, Spectral and propagation results for mag-
netic Schrödinger operators; a C ∗ -algebraic framework. J. Funct. Anal. 250 (2007),
42–67.
[57] M. Măntoiu, Essential spectrum and Fredholm properties for operators on locally
compact groups. J. Oper. Theory 77 (2017), 481–501.
[58] R. Mazzeo, Elliptic theory of differential edge operators. I. Commun. Partial Differ.
Equations 16 (1991), 1615–1664.
[59] M. Melgaard and G. Rozenblum, Spectral estimates for magnetic operators. Math.
Scand. 79 (1996), 237–254.
[60] S.T. Melo, R. Nest, and E. Schrohe, C ∗ -structure and K-theory of Boutet de Mon-
vel’s algebra. J. Reine Angew. Math. 561 (2003), 145–175.
120 C. Carvalho, V. Nistor and Y. Qiao

[61] R. Melrose, The Atiyah–Patodi–Singer Index Theorem. A.K. Peters, Wellesley, MA.,
1993.
[62] R. Melrose and P. Piazza, Analytic K-theory on manifolds with corners. Adv. Math.
92 (1992), 1–26.
[63] R.B. Melrose, Geometric Scattering Theory. Cambridge University Press, Cam-
bridge, 1995.
[64] I. Moerdijk and J. Mrčun, Introduction to Foliations and Lie Groupoids. Cambridge
University Press, Cambridge, 2003.
[65] B. Monthubert, Pseudodifferential calculus on manifolds with corners and group-
oids. Proc. Amer. Math. Soc. 127 (1999), 2871–2881.
[66] B. Monthubert, Groupoids and pseudodifferential calculus on manifolds with cor-
ners. J. Funct. Anal. 199 (2003), 243–286.
[67] J. Mougel, V. Nistor, and N. Prudhon, A refined HVZ-theorem for asymptotically
homogeneous interactions and finitely many collision planes. Revue Romaine de
Mathématiques Pures et Appliqués 62 (2017), 287–308.
[68] P. Muhly, J. Renault, and D. Williams, Equivalence and isomorphism for groupoid
C ∗ -algebras. J. Operator Theory 17 (1987), 3–22.
[69] P.S. Muhly, J. Renault, and D. Williams, Continuous-trace groupoid C ∗ -algebras.
III. Trans. Amer. Math. Soc. 348 (1996), 3621–3641.
[70] V.E. Nazaı̆kinskiı̆, A.Yu. Savin, B.Yu. Sternin, and B.-W. Shulze, On the index of
elliptic operators on manifolds with edges. Mat. Sb. 196 (2005), 23–58.
[71] S. Nazarov and B. Plamenevsky, Elliptic Problems in Domains with Piecewise
Smooth Boundaries. Walter de Gruyter & Co., Berlin, 1994.
[72] F. Nicola and L. Rodino, Global Pseudo-Differential Calculus on Euclidean Spaces.
Birkhäuser Verlag, Basel, 2010.
[73] V. Nistor, Groupoids and the integration of Lie algebroids. J. Math. Soc. Japan 52
(2000), 847–868.
[74] V. Nistor, Analysis on singular spaces: Lie manifolds and operator algebras. J.
Geom. Phys. 105 (2016), 75–101.
[75] V. Nistor, Desingularization of Lie groupoids and pseudodifferential operators on
singular spaces. arXiv:1512.08613 [math.DG], 2015. Communications in Analysis
and Geometry, to appear.
[76] V. Nistor and N. Prudhon, Exhausting families of representations and spectra of
pseudodifferential operators. J. Oper. Theory, 78 (2017), 247–279.
[77] V. Nistor, A. Weinstein, and Ping Xu, Pseudodifferential operators on differential
groupoids. Pacific J. Math. 189 (1999), 117–152.
[78] C. Parenti. Operatori pseudodifferentiali in Rn e applicazioni. Annali Mat. Pura ed
App. 93 (1972), 391–406.
[79] A. Paterson, Groupoids, Inverse Semigroups, and Their Operator Algebras. Birk-
häuser Boston, Boston, MA, 1999.
[80] B.A. Plamenevskiı̆, Algebras of Pseudodifferential Operators. Kluwer Academic
Publishers Group, Dordrecht, 1989.
 Fredholm Conditions on Non-compact Manifolds 121

[81] V. Rabinovich and S. Roch, Essential spectrum and exponential decay estimates
of solutions of elliptic systems of partial differential equations. Applications to
Schrödinger and Dirac operators. Georgian Math. J. 15 (2008), 333–351.
[82] V. Rabinovich, S. Roch, and B. Silbermann, Limit Operators and Their Applications
in Operator Theory. Birkhäuser, Basel, 2004.
[83] V. Rabinovich, B.-W. Schulze, and N. Tarkhanov, C ∗ -algebras of singular integral
operators in domains with oscillating conical singularities. Manuscripta Math. 108
(2002), 69–90.
[84] M. Reed and B. Simon, Methods of Modern Mathematical Physics. IV. Analysis of
Operators. Academic Press, New York, 1978.
[85] J. Renault, A Groupoid Approach to C ∗ -Algebras. Springer, Berlin, 1980.
[86] J. Renault, The ideal structure of groupoid crossed product C ∗ -algebras. J. Oper.
Theory 25 (1991), 3–36.
[87] M. Rieffel, Induced representations of C ∗ -algebras. Adv. Math. 13 (1974), 176–257.
[88] S. Roch, Algebras of approximation sequences: structure of fractal algebras. In
“Singular Integral Operators, Factorization and Applications”. Oper. Theory Adv.
Appl. 142 (2014), 287–310.
[89] S. Roch, P. Santos, and B. Silbermann, Non-Commutative Gelfand Theories.
Springer-Verlag London, London, 2011.
[90] M. Ruzhansky and V. Turunen, Pseudo-Differential Operators and Symmetries.
Background Analysis and Advanced Topics. Birkhäuser, Basel, 2010.
[91] E. Schrohe, The symbols of an algebra of pseudodifferential operators. Pacific J.
Math. 125 (1986), 211–224.
[92] E. Schrohe, A Ψ∗ algebra of pseudodifferential operators on noncompact manifolds.
Arch. Math. (Basel) 51 (1988), 81–86.
[93] E. Schrohe, Fréchet algebra techniques for boundary value problems on noncompact
manifolds: Fredholm criteria and functional calculus via spectral invariance. Math.
Nachr. 199 (1999), 145–185.
[94] E. Schrohe and B.-W. Schulze, Boundary value problems in Boutet de Monvel’s
algebra for manifolds with conical singularities. I. In “Pseudo-Differential Calcu-
lus and Mathematical Physics”. Math. Top., vol. 5, pp. 97–209. Akademie Verlag,
Berlin, 1994.
[95] B.-W. Schulze, Pseudo-Differential Operators on Manifolds with Singularities.
North-Holland, Amsterdam, 1991.
[96] R.T. Seeley, Singular integrals on compact manifolds. Amer. J. Math. 81 (1959),
658–690.
[97] R.T. Seeley, The index of elliptic systems of singular integral operators. J. Math.
Anal. Appl. 7 (1963), 289–309.
[98] S.R. Simanca. Pseudo-Differential Operators. Longman Scientific & Technical, Har-
low, Essex, 1990.
[99] A. Sims and D. Williams, Renault’s equivalence theorem for reduced groupoid C ∗ -
algebras. J. Oper. Theory 68 (2012), 223–239.
[100] B.K. So, On the full calculus of pseudo-differential operators on boundary groupoids
with polynomial growth. Adv. Math. 237 (2013), 1–32.
122 C. Carvalho, V. Nistor and Y. Qiao

[101] M. Taylor, Gelfand theory of pseudo differential operators and hypoelliptic operators.
Trans. Amer. Math. Soc. 153 (1971), 495–510.
[102] G. Teschl, Mathematical Methods in Quantum Mechanics. With Applications to
Schrödinger Operators. 2nd ed. American Mathematical Society, Providence, RI,
2014.
[103] J. Tu, Non-Hausdorff groupoids, proper actions and K-theory. Doc. Math. 9 (2004),
565–597.
[104] E. Van Erp and R. Yuncken, A groupoid approach to pseudodifferential operators.
arXiv:1511.01041 [math.DG], 2015.
[105] E. Van Erp and R. Yuncken, On the tangent groupoid of a filtered manifold.
arXiv:1611.01081 [math.DG], 2016. Bull. London Mas. Soc., to appear.
[106] S. Vassout, Unbounded pseudodifferential calculus on Lie groupoids. J. Funct. Anal.
236 (2006), 161–200.
[107] B. Vertman, Heat-trace asymptotics for edge Laplacians with algebraic boundary
conditions. J. Anal. Math. 125 (2015), 285–318.
[108] M.I. Višik and V.V. Grušin, Degenerate elliptic differential and pseudodifferential
operators. Russian Math. Surv. 25 (1970), 21–50.
[109] D. Williams, Crossed Products of C ∗ -Algebras. American Mathematical Society,
Providence, RI, 2007.

Catarina Carvalho
Departamento de Matemática
Instituto Superior Técnico
University of Lisbon
Av. Rovisco Pais
1049-001 Lisbon, Portugal
e-mail: [email protected]
Victor Nistor
Université de Lorraine
UFR MIM, Ile du Saulcy
CS 50128
F-57045 Metz, France
and
Inst. Math. Romanian Acad.
PO BOX 1-764
014700 Bucharest, Romania
e-mail: [email protected]
Yu Qiao
School of Mathematics and Information Science
Shaanxi Normal University
Xi’an, 710119, China
e-mail: [email protected]
Operator Theory:
Advances and Applications, Vol. 267, 123–130

c Springer International Publishing AG, part of Springer Nature 2018

Statistical e-Convergence of Bögel-Type

Continuous Functions
Kamil Demirci and Sevda Orhan

Abstract. In the present work, using the concept of statistical e-convergence

instead of Pringsheim’s sense for double sequences, we obtain a Korovkin
type approximation theorem for double sequences of positive linear operators
defined on the space of all real-valued B-continuous functions on a compact
subset of the real two-dimensional space. Then, displaying an example, it is
shown that our new result is stronger than its classical version.
Mathematics Subject Classification (2010). 40B05; 41A36.
Keywords. Double sequences, statistical e-convergence, Korovkin type approx-
imation theorem, B-continuity, positive linear operator.

1. Introduction and Preliminaries

Let N denote the set of all natural numbers. A double sequence x = {xm,n } is
convergent in Pringsheim’s sense if, for every ε > 0, there exists N = N (ε) ∈
N such that |xm,n − L| < ε whenever m, n > N . In this case L is called the
Pringsheim limit of x and is denoted by P −lim x = L (see [14]). A double sequence
is called bounded if there exists a positive number M such that |xm,n | ≤ M for
all (m, n) ∈ N2 = N × N. Note that in contrast to the case for single sequences, a
convergent double sequence need not to be bounded. In addition to the Pringsheim
convergence, Boos et al. [5, 6] introduced and investigated the following notion of
e-convergence of double sequences, which is stronger method than Pringsheim’s:
A double sequence x = {xm,n } is said to be e-convergent to a number L if
∀ε > 0 ∃n0 ∈ N ∀n ≥ n0 ∃mn ∈ N ∀m ≥ mn : |xm,n − L| < ε.
In this case, we write e − lim xm,n = L.
m,n
For related topics about these convergence methods, see [15, 18, 19]. Recently,
the statistical e-convergence has been introduced in [16]. Before this convergence
method let us first remind of the concept of statistical convergence, which is closely
124 K. Demirci and S. Orhan

related to the concept of natural density. The notion of statistical convergence of

sequences of real numbers was given by Steinhaus [17] and Fast [11], independently.
Let K be a subset of N. Then the natural density of K is given by
1
δ (K) := lim |{k ≤ n : k ∈ K}|
n n

provided that the limit on the right-hand side exists, where |A| denotes the cardi-
nality of the set A. Then a sequence x = {xn } is said to be statistically convergent
to a number L if, for every ε > 0,
δ ({n ∈ N : |xn − L| ≥ ε}) = 0.
In this case, we write st − limxn = L. It is easy to see that every convergent
n
sequence is statistically convergent but not conversely.
Recently, this concept has been extended to the double sequences by Moricz
[13]. If E ⊂ N2 is a two-dimensional subset of positive integers, then Ej,k :=
{(m, n) ∈ E : m ≤ j, n ≤ k} and |Ej,k | denotes the cardinality of Ej,k . The double
natural density of E is given by
1
δ2 (E) := P − lim |Ej,k | ,
j,k jk

if it exists. The number sequence x = {xm,n } is statistically convergent to L

provided that for every ε > 0, the set
E := Ej,k (ε) := {m ≤ j, n ≤ k : |xm,n − L| ≥ ε}
has natural density zero; in that case we write st2 − limxm,n = L.
m,n
Clearly, a P -convergent double sequence is statistically convergent in Pring-
sheim sense to the same value but its converse is not always true. Also, note that
a statistically convergent double sequence need not to be bounded. For example,
consider the double sequence x = {xm,n } given by

mn, if m and n are squares,
xm,n =
1, otherwise.
Then, clearly st2 −limxm,n = 1. Nevertheless, x is neither convergent in Pringsheim
m,n
sense nor bounded.
Now we recall the statistical e-convergence method.
Definition 1.1 ([16]). A double sequence x = {xm,n } is said to be statistically
e-convergent to the number L if for every ε > 0,
δ ({n : δ ({m : |xm,n − L| ≥ ε}) = 0}) = 1.
In this case, we write ste − lim xm,n = L. If a double sequence x = {xm,n } is e-
m,n
convergent then it is statistically e-convergent. But, the converse of this implication
may not be true. Also, a double sequence which is statistically e-convergent need
not be statistical convergent (for more properties and details, see also [16]).
 Statistical e-Convergence of Bögel-Type Continuous Functions 125

Example. Let define s = {sm,n } by

⎧
⎨ mn, m ≤ n,
sm,n := m2 + 1, m > n and m is square,
⎩
1, m > n and m is not square.
Then, it is easy to see that
ste − limsm,n = 1, (1.1)
m,n
however, e − limsm,n , P − lim sm,n and st2 − lim sm,n do not exist.
m,n m,n m,n

In this work, using the concept of statistical e-convergence for double se-
quences, we obtain a Korovkin type approximation theorem for double sequences
of positive linear operators defined on the space of all real valued B-continuous
functions on a compact subset of the real two-dimensional space. Then, displaying
an example, it is shown that our new result is stronger than its classical version.

2. A Korovkin-type approximation theorem

The Korovkin theory has been widely studied in the literature. This theory is con-
nected with the approximation to continuous functions by means of positive linear
operators (see, for instance, [1, 12]). In order to improve the classical Korovkin
theory, the space of Bögel-type continuous (or, simply, B-continuous) functions
instead of the ordinary continuity has been used in [2–4].
The definition of B-continuity was introduced by Bögel [7–9] as follows:
Let X and Y be compact subsets of the real numbers, and let D = X × Y .
Then, a function f : D → R is called a B-continuous at a point (x, y) ∈ D if, for
every ε > 0, there exists a positive number δ = δ(ε) such that
|Δx,y [f (u, v)]| < ε,
for any (u, v) ∈ D with |u − x| < δ and |v − y| < δ, where the symbol Δx,y [f (u, v)]
denotes the mixed difference of f defined by
Δx,y [f (u, v)] = f (u, v) − f (u, y) − f (x, v) + f (x, y).
By Cb (D) we denote the space of all B-continuous functions on D. Recall that
C(D) and B(D) denote the space of all continuous (in the usual sense) functions
and the space of all bounded function on D, respectively and the usual supremum
norm on the spaces B(D) is given by
f  := sup |f (x, y)| for f ∈ B(D).
(x,y)∈D

Then, notice that C(D) ⊂ Cb (D). Moreover, one can find an unbounded B-
continuous function, which follows from the fact that, for any function of the
type f (u, v) = g(u) + h(v), we have
Δx,y [f (u, v)] = 0 for all (x, y), (u, v) ∈ D.
126 K. Demirci and S. Orhan

We recall that the following lemma for B-continuous functions was proved
by Badea et al. [2].
Lemma 2.1 ([2]). If f ∈ Cb (D), then, for every ε > 0, there are two positive
numbers A(ε) = A(ε, f ) and B(ε) = B(ε, f ) such that
ε
|Δx,y [f (u, v)]| ≤ + A(ε)(u − x)2 + B(ε)(v − y)2
3
holds for all (x, y), (u, v) ∈ D.
Let T be a linear operator from Cb (D) into B (D). Then, as usual, we say
that T is positive linear operator provided that f ≥ 0 implies T f ≥ 0. Also,
we denote the value of T (f ) at a point (x, y) ∈ D by T (f (u, v); x, y) or, briefly,
T (f ; x, y).
Throughout the paper, for fixed (x, y) ∈ D and f ∈ Cb (D), we use the
function Fx,y defined as follows:
Fx,y (u, v) = f (u, y) + f (x, v) − f (u, v) for (u, v) ∈ D. (2.1)
Since
Δx,y [Fx,y (u, v)] = −Δx,y [f (u, v)]
holds for all (x, y), (u, v) ∈ D, the B-continuity of f implies the B-continuity of
Fx,y for every fixed (x, y) ∈ D. We also use the following test functions
r0 (x, y) = 1, r1 (x, y) = x, r2 (x, y) = y and r3 (x, y) = x2 + y 2 .
The following version of the classical Korovkin type approximation theorem was
given by [2].
Theorem 2.2 ([2]). Let {Tm,n } be a sequence of positive linear operators acting from
Cb (D) into B (D). Assume that the following conditions hold, for all (x, y) ∈ D
and (m, n) ∈ N2 ,
(i) Tm,n (r0 ; x, y) = 1,
(ii) Tm,n (r1 ; x, y) = r1 (x, y) + um,n (x, y),
(iii) Tm,n (r2 ; x, y) = r2 (x, y) + vm,n (x, y),
(iv) Tm,n (r3 ; x, y) = r3 (x, y) + wm,n (x, y),
where {um,n (x, y)}, {vm,n (x, y)} and {wm,n (x, y)} converge to zero uniformly on
D as m, n → ∞ (in any manner). Then the sequence
{Tm,n (Fx,y ; x, y)}
converges uniformly to f (x, y) with respect to (x, y) ∈ D, where Fx,y is given by
(2.1).
Now we have the following main result.
Theorem 2.3. Let {Tm,n } be a sequence of positive linear operators acting from
Cb (D) into B (D). Assume that the following conditions hold:
δ ({n : δ ({m : Tm,n (r0 ; x, y) = 1 for all (x, y) ∈ D}) = 1}) = 1 (2.2)
 Statistical e-Convergence of Bögel-Type Continuous Functions 127

and
ste − lim Tm,n (ri ) − ri  = 0 for i = 1, 2, 3. (2.3)
m,n
Then, for all f ∈ Cb (D), we have
ste − lim Tm,n (Fx,y ) − f  = 0, (2.4)
m,n

where Fx,y is given by (2.1).

Proof. Let (x, y) ∈ D and f ∈ Cb (D) be fixed. Putting
S := {m : Tm,n (r0 ; x, y) = 1 for all (x, y) ∈ D} , (2.5)
and
T := {n : δ ({m : Tm,n (r0 ; x, y) = 1 for all (x, y) ∈ D}) = 1} , (2.6)
we get from (2.2) that
δ (N\S) = 0 and δ (T ) = 1. (2.7)
Using the B-continuity of the function Fx,y given by (2.1), Lemma 2.1 implies
that, for every ε > 0, there exist two positive numbers A(ε) and B(ε) such that
ε
|Δx,y [Fx,y (u, v)]| ≤ + A(ε)(u − x)2 + B(ε)(v − y)2 (2.8)
3
holds for every (u, v) ∈ D. Also, by (2.2), observe that
Tm,n (Fx,y ; x, y) − f (x, y) = Tm,n (Δx,y [Fx,y (u, v)] ; x, y) (2.9)
holds for all m ∈ S and n ∈ T. Since, Tm,n linear and positive, for all m ∈ S and
n ∈ T , it follows from (2.8) and (2.9) that
|Tm,n (Fx,y ; x, y) − f (x, y)| = |Tm,n (Δx,y [Fx,y (u, v)] ; x, y)|
≤ Tm,n (|Δx,y [Fx,y (u, v)]| ; x, y)
ε
≤ + A(ε)Tm,n (u − x)2 ; x, y
3
+ B(ε)Tm,n (v − y)2 ; x, y
ε
≤ + C(ε){x2 + y 2 + Tm,n (r3 ; x, y)
3
− 2xTm,n (r1 ; x, y) − 2yTm,n(r2 ; x, y)},
where C(ε) = max{A(ε), B(ε)} and hence,
ε 3
|Tm,n (Fx,y ; x, y) − f (x, y)| ≤ + C(ε) |Tm,n (ri ; x, y) − ri (x, y)| (2.10)
3 i=1
holds for all m ∈ S and n ∈ T.
Now, taking supremum over (x, y) ∈ D on both sides of inequality (2.10) we
have, for all m ∈ S and n ∈ T ,
ε 3
Tm,n (Fx,y ) − f  ≤ + C(ε) Tm,n (ri ) − ri  .
3 i=1
128 K. Demirci and S. Orhan

For a given ε > 0, choose ε > 0 such that ε < 3ε and define
K := {n : δ ({m : Tm,n (Fx,y ) − f  ≥ ε }) = 0} ,
   
3ε − ε
Ki := n : δ m : Tm,n (ri ) − ri  ≥ = 0 , i = 1, 2, 3.
9C(ε)
-
3
Then by hypothesis (2.3) we get δ (Ki ) = 1, i = 1, 2, 3. If we take K4 = Ki , we
i=1
have δ (K4 ) = 1. Hence, δ (K4 ∩ T ) = 1.
For each n ∈ K4 ∩ T we define
 
3ε − ε
Sin = m : Tm,n (ri ) − ri  ≥ , i = 1, 2, 3.
9C(ε)
From the inequality (2.10) for each n ∈ K4 ∩ T we get
3
{m : Tm,n (Fx,y ) − f  ≥ ε } ∩ S ⊆ (Sin ∩ S),
i=1

since

3 
3
δ ({m : Tm,n (Fx,y ) − f  ≥ ε } ∩ S) ≤ δ (Sin ∩ S) ≤ δ (Sin )
i=1 i=1

and δ (Sin ) = 0, then we obtain

δ ({m : Tm,n (Fx,y ) − f  ≥ ε } ∩ S) = 0. (2.11)
Also,
δ ({m : Tm,n (Fx,y ) − f  ≥ ε })
= δ ({m : Tm,n (Fx,y ) − f  ≥ ε } ∩ S)
+ δ ({m : Tm,n (Fx,y ) − f  ≥ ε } ∩ (N\S))
≤ δ ({m : Tm,n (Fx,y ) − f  ≥ ε } ∩ S) + δ (N\S)
and it follows from (2.7) and (2.11) that
δ ({m : Tm,n (Fx,y ) − f  ≥ ε }) = 0.
This implies that K4 ∩ T ⊆ K. So, δ (K) = 1. That is (2.4) holds, which completes
the proof. 
Remark 2.4. We know that, for some f ∈ Cb (D), the function f may be unbounded
on the compact set D. However, we can say from the conditions (2.2), (2.8) and
(2.9) that the number
sup |Tm,n (Fx,y ; x, y) − f (x, y)|
(x,y)∈D

in Theorem 2.3 is finite for each m ∈ S and n ∈ T , where S and T are given by
(2.5) and (2.6).
 Statistical e-Convergence of Bögel-Type Continuous Functions 129

Remark 2.5. We now show that our result Theorem 2.3 is stronger than its clas-
sical version Theorem 2.2. To see this first consider the following Bernstein-type
operators introduced by Badea and Cottin [3]:
m  n     
s t m n s t
Hm,n (f ; x, y) = Fx,y , x y (1 − x)m−s (1 − y)n−t ,
s=0 t=0
m n s t
where Fx,y is given by (2.1), and (x, y) ∈ I 2 = [0, 1] × [0, 1]; f ∈ Cb (I 2 ). Then, by
Theorem 2.2, we know that, for any f ∈ Cb (I 2 )
P − lim Hm,n (f ) − f  = 0. (2.12)
m,n

Now, we define the following positive linear operators on Cb (I 2 ) as follows:

Tm,n (f ; x, y) = sm,n Hm,n (f ; x, y). (2.13)
where {sm,n } given by (1.1). Then, observe that the sequence of positive linear
operators {Tm,n } satisfy all hypotheses of Theorem 2.3 defined by (2.13). So, by
(2.12) and (1.1), we have
ste − lim Tm,n (f ) − f  = 0.
m,n

However, since {sm,n } is not P -convergent, the sequence {Tm,n (f ; x, y)} given by
(2.13) does not converge uniformly to the function f ∈ Cb (I 2 ). So, we conclude that
Theorem 2.2 does not work for the operators Tm,n in (2.13) while our Theorem
2.3 still works. Also, since {sm,n } is not statistically convergent, the sequence
{Tm,n (f ; x, y)} given by (2.13) does not converge statistically uniformly. Hence,
the statistical Korovkin type approximation theorem given by Dirik, Duman and
Demirci [10] does not work.

References
[1] F. Altomare and M. Campiti, Korovkin-type approximation theory and its applica-
tions. de Gruyter Stud. Math. 17, Walter de Gruyter, Berlin, 1994.
[2] C. Badea, I. Badea, and H.H. Gonska, A test function and approximation by pseu-
dopolynomials. Bull. Austral. Math. Soc. 34 (1986), 53–64.
[3] C. Badea and C. Cottin, Korovkin-type theorems for generalized Boolean sum op-
erators. Approximation theory (Kecskemét, 1990), 51–68, Colloq. Math. Soc. János
Bolyai, 58, North-Holland, Amsterdam, 1991.
[4] I. Badea, Modulus of continuity in Bögel sense and some applications for approxi-
mation by a Bernstein-type operator. Studia Univ. Babeş-Bolyai Ser. Math.-Mech.
18 (1973), 69–78 (in Romanian).
[5] J. Boos, Classical and modern methods in summability. Oxford University Press Inc.,
New York, 2000.
[6] J. Boos, T. Leiger, and K. Zeller, Consistency theory for SM-methods. Acta. Math.
Hungar. 76 (1997), 83–116.
[7] K. Bögel, Mehrdimensionale Differentiation von Funktionen mehrerer Veränderli-
cher. J. Reine Angew. Math. 170 (1934), 197–217.
130 K. Demirci and S. Orhan

[8] K. Bögel, Über mehrdimensionale Differentiation, Integration und beschränkte Vari-

ation. J. Reine Angew. Math. 173 (1935), 5–29.
[9] K. Bögel, Über die mehrdimensionale Differentiation. Jahresber. Deutsch. Mat.-
Verein. 65 (1962), 45–71.
[10] F. Dirik, O. Duman, and K. Demirci, Approximation in statistical sense to B-
continuous functions by positive linear operators. Studia Sci. Math. Hungarica 47
(3) (2010), 289–298.
[11] H. Fast, Sur la convergence statistique. Colloq. Math. 2 (1951), 241–244.
[12] P.P. Korovkin, Linear operators and approximation theory. Hindustan Publ. Co.,
Delhi, 1960.
[13] F. Moricz, Statistical convergence of multiple sequences. Arch. Math. (Basel) 81
(2004), 82–89.
[14] A. Pringsheim, Zur Theorie der zweifach unendlichen Zahlenfolgen. Math. Ann. 53
(1900), 289–321.
[15] Y. Sever and Ö. Talo, e-Core of double sequences. Acta Math. Hungar. 144 (2014),
236–246.
[16] Y. Sever and Ö. Talo, Statistical e-convergence of double sequences and its application
to Korovkin type approximation theorem for functions of two variables. Iranian Jour-
nal of Science and Technology, Transactions A: Science (2016). DOI 10.1007/s40995-
016-0018-7.
[17] H. Steinhaus, Sur la convergence ordinaire et la convergence asymtotique. Colloq.
Math. 2 (1951), 73–74.
[18] M. Zeltser, Investigation of double sequence spaces by soft and hard analitical
methods. Dissertationes Mathematicae Universitatis Tartuensis 25, Tartu Univer-
sity Press, Univ. of Tartu, Faculty of Mathematics and Computer Science, (Tartu,
2001).
[19] M. Zeltser, On conservative and coercive SM-methods. Proc. Est. Acad. Sci. Phys.
Math., 50 (2001), 76–85.

Kamil Demirci and Sevda Orhan

Sinop University
Faculty of Arts and Sciences
Department of Mathematics
TR-57000, Sinop, Turkey
e-mail: [email protected]
[email protected]
Operator Theory:
Advances and Applications, Vol. 267, 131–139

c Springer International Publishing AG, part of Springer Nature 2018

Weighted Statistical Relative Approximation

by Positive Linear Operators
Kamil Demirci, Sevda Orhan and Burçak Kolay

Abstract. In this paper we define weighted statistical relative uniform con-

vergence by using weighted density and give a Korovkin-type approximation
theorem. Then, we construct an example such that our new approximation
result works but its weighted statistical and statistical (and classical) cases do
not work. We also compute the rates of weighted statistical relative uniform
convergence of sequences of positive linear operators.

Mathematics Subject Classification (2010). 40A35, 41A25, 41A36.

Keywords. Statistical convergence, weighted statistical convergence, weighted
statistical relative uniform convergence, Korovkin theorem.

1. Introduction and preliminaries

The Korovkin theorem provides a criterion for whether a given sequence {Lk } of
positive linear operators converges to the identity operator [11]. More recently, gen-
eral versions of the Korovkin theorem were studied, in which a more general notion
of convergence is used. In particular, the use of statistical convergence had a great
impulse in recent years. Statistical convergence was first introduced independently
by Fast [7] and Steinhaus [15]. The concept of weighted statistical convergence
was introduced and studied by Karakaya and Chishti [10] and this convergence
was modified by Mursaleen et al. in [13]. But, in [9] Ghosal showed that both
definitions of weighted statistical convergence are not well defined in general and
modified the definition of weighted statistical convergence (see, [2–4,12]). Recently,
Demirci and Orhan [5] define a new type of statistical convergence by using the
notions of the natural density and the relative uniform convergence. The main
purpose of the present paper is using this thought for defining weighted statistical
relative uniform convergence and using this convergence to prove a Korovkin type
theorem and also, giving weighted statistical relative rates.
132 K. Demirci, S. Orhan and B. Kolay

Let K be a subset of N, the set of natural numbers, then the natural density
of K, denoted by δ(K), is given by
1
δ(K) := lim |{k ≤ n : k ∈ K}|
n n
whenever the limit exists, where |B| denotes the cardinality of the set B [14]. The
sequence x = {xk } is said to be statistically convergent to L if for every ε > 0,
δ({k ∈ N : |xk − L| ≥ ε}) = 0,
i.e., for every ε > 0,
1
lim |{k ≤ n : |xk − L| ≥ ε}| = 0.
n n
We note that every convergent sequence is statistically convergent, but not con-
versely.
Let p = {pk } be a sequence n of nonnegative real numbers such that p1 > 0,
lim inf n→∞ pn > 0 and Pn = k=1 pk → ∞ as n → ∞. The weighted density of
K ⊂ N denoted by δN (K), is given by
1
δN (K) := lim |{k ≤ Pn : k ∈ K}|.
n Pn

The sequence x = {xk } is said to be weighted statistically convergent (or SN -

convergent ) to L if for every ε > 0,
1
lim |{k ≤ Pn : pk |xk − L| ≥ ε}| = 0.
n Pn

We write stN − lim x = L [9]. If pk = 1 for all k, then weighted statistical conver-
gence is reduced to statistical convergence. In [9] Ghosal showed that statistical
convergence and weighted statistical convergence do not imply each other in gen-
eral.
Let f and fk belong to C(X), which is the space of all continuous real-valued
functions on a compact subset X of the real numbers.
Definition 1.1 ( [5]). A sequence {fk } is said to be statistically relatively uniform
convergent to f on X if there exists a function σ(x) satisfying |σ(x)| = 0, called a
scale function such that for every ε > 0,
 
1 fk (x) − f (x)
lim k ≤ n : sup ≥ ε = 0.
n n x∈X σ(x)
We can easily get the following definition like the definition of weighted sta-
tistical convergence for the number sequences.
Definition 1.2. A sequence {fk } is said to be weighted statistically uniform con-
vergent to f on X if for every ε > 0,
 
1
lim k ≤ Pn : pk sup |fk (x) − f (x)| ≥ ε = 0.
n Pn x∈X

This limit is denoted by (w − st) fk ⇒ f .

 Weighted Statistical Relative Approximation 133

Now, we can give the following definition.

Definition 1.3. A sequence {fk } is said to be weighted statistically relatively uni-

form convergent to f on X if there exists a function σ(x) satisfying |σ(x)| = 0,
called a scale function such that for every ε > 0,
 
1 fk (x) − f (x)
lim k ≤ Pn : pk sup ≥ ε = 0.
n Pn x∈X σ(x)

This limit is denoted by (w − st) fk ⇒ f (X; σ). If pk = 1 for all k, then weighted
statistical relative uniform convergence is reduced to statistical relative uniform
convergence.

Using the above definitions, we get following result immediately.

Lemma 1.4. (w − st) fk ⇒ f on X implies (w − st) fk ⇒ f (X; σ) .

However, one can construct an example which guarantees that the converse
of Lemma 1.4 is not always true. Such an example is given as follows.

Example 1.5. For each k ∈ N, define gk : [0, 1] → R by

k3 x
gk (x) = (1.1)
2 + k 4 x3
Then observe that
gk (x) − g(x) k3
sup =
x∈[0,1] σ(x) 2 + k4
√
and with the choice of pk = k, we have that (w − st) gk ⇒ g = 0 ([0, 1]; σ),
 1
x2 , 0 < x ≤ 1,
σ(x) =
2, x = 0.

However {gk } is not weighted statistically uniform convergent, i.e.,

√
3
k5
sup |gk (x) − g(x)| =
x∈[0,1] 3

then
 √ 3
6
1 k 13
lim k ≤ Pn : ≥ε = 0.
n Pn 3

Also, {gk } is neither ordinary uniform nor statistically uniform convergent to the
function g = 0 on the interval [0, 1].
134 K. Demirci, S. Orhan and B. Kolay

2. Weighted statistical relative Korovkin theorem

In this section we use the notion of weighted statistical relative uniform conver-
gence to prove a Korovkin type approximation theorem. Let L be a linear operator
from C(X) into itself. Then, as usual, we say that L is a positive linear operator
provided that f ≥ 0 implies L(f ) ≥ 0. Also, we denote the value of L(f ) at a
point x ∈ X by L(f (u); x) or, briefly, L(f ; x). Throughout the paper we use the
test functions fi (x) = xi (i = 0, 1, 2) and f C(X) denotes the usual supremum
norm of f in C(X).

Theorem 2.1. Let {Lk } be a sequence of positive linear operators acting from C(X)
into itself. Then, for all f ∈ C(X),
(w − st) Lk (f ) ⇒ f (X; σ) (2.1)
if and only if
(w − st) Lk (fi ) ⇒ fi (X; σi ), i = 0, 1, 2, (2.2)
where σ(x) = max{|σi (x)| : i = 0, 1, 2}, |σi (x)| > 0 and σi (x) is unbounded,
i = 0, 1, 2.

Proof. Reasoning exactly as in the proof of Theorem 2 in [5] and because of ε is

arbitrary, we arrive at

Lk (f ; x) − f (x) Lk (f0 ; x) − f0 (x)
sup ≤ M sup
x∈X σ(x) x∈X σ0 (x)
Lk (f1 ; x) − f1 (x)
+ sup
x∈X σ1 (x)

Lk (f2 ; x) − f2 (x)
+ sup (2.3)
x∈X σ2 (x)
where
2K
M =ε+K + f2 C(X) + 2f1 C(X) + f0
δ2
and σ(x) = max{|σi (x)| : i = 0, 1, 2}. Hence

Lk (f ; x) − f (x) Lk (f0 ; x) − f0 (x)
pk sup ≤ M pk sup
x∈X σ(x) x∈X σ0 (x)
Lk (f1 ; x) − f1 (x)
+pk sup
x∈X σ1 (x)

Lk (f2 ; x) − f2 (x)
+pk sup .
x∈X σ2 (x)
Now, for a given r > 0, choose ε > 0 such that ε < r. Then,
 
Lk (f ; x) − f (x)
Sn (r) := k ≤ Pn : pk sup ≥r
x∈X σ(x)
 Weighted Statistical Relative Approximation 135

and
 
Lk (fi ; x) − fi (x) r
Si,n (r) := k ≤ Pn : pk sup ≥ , i = 0, 1, 2.
x∈X σi (x) 3M
It follows from (2.3) that Sn (r) ⊂ ∪2i=0 Si,n (r) and

|Sn (r)| 
2
|Si,n (r)|
lim ≤ lim .
n Pn i=0
n Pn

Then using the hypothesis (2.2), we get

(w − st) Lk (f ) ⇒ f (X; σ)
where σ(x) = max{|σi (x)| : i = 0, 1, 2}.
This completes the proof of the theorem. 

In the following we construct an example of a sequence of positive linear

operators satisfying the conditions of Theorem 2.1 but does not satisfy the condi-
tions of the weighted statistical and statistical (and classical) cases of the Korovkin
results introduced in [12] and [8] (and [11]), respectively.
Example 2.2. Let X = [0, 1] and consider the following Bernstein–Kantorovich
polynomials
k   !
(n+1)/(k+1)
 k n
Uk (f ; x) = (k + 1) x (1 − x)k−n f (t) dt
n=0
n
n/(k+1)

where x ∈ [0, 1], f ∈ C[0, 1] and k ∈ N. Using these polynomials, we introduce the
following positive linear operators on C[0, 1]:
Tk (f ; x) = (1 + gk (x))Uk (f ; x), x ∈ [0, 1] and f ∈ C[0, 1], (2.4)
where gk (x) is given by (1.1). Then, observe that
Tk (f0 ; x) = (1 + gk (x))f0 (x),
 
k 1
Tk (f1 ; x) = (1 + gk (x)) f1 (x) + ,
k + 1 2(k + 1)
 
k(k − 1) 2kx 1
Tk (f2 ; x) = (1 + gk (x)) f2 (x) + + .
(k + 1)2 (k + 1)2 3(k + 1)2
√
Let pk = k (k = 1, 2, . . .). Since (w − st) gk ⇒ g = 0 ([0, 1]; σ), where
 1
x2 , 0 < x ≤ 1,
σ(x) =
2, x = 0,
we conclude that
(w − st) Tk (fi ) ⇒ fi ([0, 1]; σ) for each i = 0, 1, 2.
136 K. Demirci, S. Orhan and B. Kolay

So, by Theorem 2.1, we immediately see that

(w − st) Tk (f ) ⇒ f ([0, 1]; σ) for all f ∈ C[0, 1].
However, since {gk } is not weighted statistically uniform and statistically uniform
convergent to the function g = 0 on the interval [0, 1], we can say that weighted
statistical and statistical Korovkin theorem does not work for our operators defined
by (2.4). Furthermore, since {gk } and hence {Tk } is not uniformly convergent (in
the ordinary sense) to the function g = 0 on [0, 1], the classical Korovkin theorem
does not work either.

3. Weighted statistical relative rates

In this section first we define the rate of weighted statistically relatively uniform
convergent sequence and recall that the modulus of continuity of a function f ∈
C(X). Then, we study the rates of positive linear operators defined on C(X) with
the help of this definition and modulus of continuity.
We now present the following definition.
Definition 3.1. A sequence {fn } is said to converge weighted statistical uniform
relative to the scale function σ(x), |σ(x)| > 0, to f on X with the rate of β ∈ (0, 1)
if for every ε > 0,
 
fk (x) − f (x)
k ≤ Pn : pk sup ≥ε
x∈X σ(x)
lim = 0.
n n1−β Pn
In this case, it is denoted by
(w − st) − (fn − f ) = o(n−β ) (X; σ).
Then we first need the following lemma.
Lemma 3.2. Let {fn } and {gn } be function sequences belonging to C(X). Assume
that
(w − st) − (fn − f ) = o(n−β0 ) (X; σ0 )
and
(w − st) − (gn − g) = o(n−β1 ) (X; σ1 ),
|σi (x)| > 0, i = 0, 1. Let β = min{β0 , β1 }. Then the following statements hold:
(i) (w − st) − (fn + gn ) − (f + g) = o(n−β ) (X; max{|σi | : i = 0, 1}),
(ii) (w − st) − (fn − f )(gn − g) = o(n−β ) (X; σ0 σ1 ),
−β0
+ n − f )) = o(n
(iii) (w − st) − (λ(f ) (X;+σ0 ) for any real number λ,
−β0
(iv) (w − st) − |fn − f | = o(n ) (X; |σ0 |).
Proof. (i) Assume that (w − st) − (fn − f ) = o(n−β0 ) (X; σ0 ) and that (w − st) −
(gn − g) = o(n−β1 ) (X; σ1 ). Also, for every ε > 0 define
 
(fk + gk )(x) − (f + g)(x)
ϕn (ε) := k ≤ Pn : pk sup ≥ε ,
x∈X σ(x)
 Weighted Statistical Relative Approximation 137
 
fk (x) − f (x) ε
ϕ0,n (ε) := k ≤ Pn : pk sup ≥ ,
x∈X σ0 (x) 2
 
gk (x) − g(x) ε
ϕ1,n (ε) := k ≤ Pn : pk sup ≥ ,
x∈X σ1 (x) 2
where σ(x) = max{|σi (x)| : i = 0, 1}. Then observe that
ϕn (ε) ϕ0,n (ε) + ϕ1,n (ε)
≤ . (3.1)
n1−β Pn n1−β Pn
Since β = min{β0 , β1 }, by (3.1), we get
ϕn (ε) ϕ0,n (ε) ϕ1,n (ε)
1−β
≤ 1−β0 + 1−β1 . (3.2)
n Pn n Pn n Pn
Now by taking limit as n → ∞ in (3.2) and using the hypotheses, we conclude
that
ϕn (ε)
lim 1−β = 0,
n n Pn
which completes the proof of (i). Since the proofs of (ii), (iii) and (iv) are similar,
we omit them. 

On the other hand, we recall that the modulus of continuity of a function

f ∈ C(X) is defined by
ω(f, δ) = sup |f (y) − f (x)| (δ > 0).
|y−x|≤δ, x,y∈X

Then we have the following result.

Theorem 3.3. Let X be a compact subset of the real numbers, and let {Ln } be a
sequence of positive linear operators acting from C(X) into itself. Assume that the
following conditions hold:
(a) (w − st) − (Ln (f0 ) − f0 ) = o(n−β0 ) (X; σ0 ), +
(b) (w−st)−ω(f, δn ) = o(n−β1 ) (X; σ1 ), where δn (x) = Ln (ψ 2 ; x) with ψ(y) =
y − x.
Then we have, for all f ∈ C(X),
(w − st) − (Ln (f ) − f ) = o(n−β ) (X; σ),
where β = min{β0 , β1 } and
σ(x) = max{|σ0 (x)|, |σ1 (x)|, |σ0 (x)σ1 (x)|},
|σi (x)| > 0 and σi (x) is unbounded, i = 0, 1.
Proof. Let f ∈ C(X) and x ∈ X. Then it is well known that
|Ln (f ; x) − f (x)| ≤M |Ln (f0 ; x) − f0 (x)|
' + (
+ Ln (f0 ; x) + Ln (f0 ; x) ω(f, δn )
138 K. Demirci, S. Orhan and B. Kolay

where M = f C(X) (see, for instance, [1, 6]). This yields that
Ln (f ; x) − f (x) Ln (f0 ; x) − f0 (x) ω(f, δn )
sup ≤ M sup + 2 sup
x∈X σ(x) x∈X σ 0 (x) x∈X |σ1 (x)|
ω(f, δn ) Ln (f0 ; x) − f0 (x)
+ sup sup
x∈X |σ1 (x)| x∈X σ0 (x)
4
ω(f, δn ) Ln (f0 ; x) − f0 (x)
+ sup sup .
x∈X |σ1 (x)| x∈X σ0 (x)
Now considering the above inequality, the hypotheses (a), (b), and Lemma 3.2,
the proof is completed at once. 
Remark 3.4. If we replace the conditions (a), (b) in Theorem 3.3 by the following
condition:
(w − st) − (Ln (fi ) − fi ) = o(n−βi ) (X; σi ), i = 0, 1, 2, (3.3)
Then, since
Ln (ψ 2 ; x) ≤ N {|Ln (f0 ; x) − f0 (x)|
+ |Ln (f1 ; x) − f1 (x)| + |Ln (f2 ; x) − f2 (x)|} (3.4)
where N = f2 C(X) + 2f1 C(X) + f0 . We get

Ln (ψ 2 ; x) Ln (f0 ; x) − f0 (x)
sup ≤ N sup
x∈X |σ(x)| x∈X σ0 (x)

Ln (f1 ; x) − f1 (x) Ln (f2 ; x) − f2 (x)
+ sup + sup
x∈X σ1 (x) x∈X σ2 (x)
where σ(x) = max{|σi (x)| : i = 0, 1, 2}. It follows that (3.3), (3.4) and Lemma 3.2
that +
(w − st) − δn = Ln (ψ 2 ) = o(n−β ) (X; σ)
where β = min{β0 , β1 , β2 }. So, by Theorem 3.3 we get, for all f ∈ C(X),
(w − st) − (Ln (f ) − f ) = o(n−β ) (X; σ).
Hence, if we replace the conditions (a), (b) with the condition (3.3) in Theorem 3.3,
then we get the rates of weighted statistically relatively uniform convergence of
sequence of positive linear operators in Theorem 2.1.

References
[1] F. Altomare and M. Campiti, Korovkin-Type Approximation Theory and Its Appli-
cations. de Gruyter Stud. Math. 17, Walter de Gruyter, Berlin, 1994.
[2] S. Akdağ, Weighted equi-statistical convergence of the Korovkin type approximation
theorems. Results. Math., 72 (2017), 1073–1085.
[3] C. Belen and S.A. Mohiuddine, Generalized weighted statistical convergence and ap-
plication. Appl. Math and Comp. 219 (2013), 9821–9826.
 Weighted Statistical Relative Approximation 139

[4] N.L. Braha, V. Loku, and H.M. Srivastava, Λ2 -weighted statistical convergence and
Korovkin and Voronovskaya type theorems. Appl. Math and Comp. 266 (2015), 675–
686.
[5] K. Demirci and S. Orhan, Statistically relatively uniform convergence of positive
linear operators. Results. Math. 69 (2016), 359–367.
[6] R.A. DeVore, The Approximation of Continuous Functions by Positive Linear Op-
erators. Lecture Notes in Math., vol. 293, Springer-Verlag, Berlin, 1972.
[7] H. Fast, Sur la convergence statistique. Colloq. Math. 2 (1951), 241–244.
[8] A.D. Gadjiev and C. Orhan, Some approximation theorems via statistical conver-
gence. Rocky Mountain J. Math. 32 (2002), 129–138.
[9] S. Ghosal, Weighted statistical convergence of order α and its applications. Journal
of the Egyptian Mathematical Society (2016) 24, 60–67.
[10] V. Karakaya and T.A. Chishti, Weighted statistical convergence. Iran. J. Sci. Technol.
Trans. A Sci. 33 (2009), 219–223.
[11] P.P. Korovkin, Linear Operators and Approximation Theory. Hindustan Publ. Co.,
Delhi, 1960.
[12] S.A. Mohiuddine, A. Alotaibi, and B. Hazarika, Weighted A-statistical convergence
for sequences of positive linear operators. The Scientific World J. Volume 2014, Ar-
ticle ID 437863, 8 pages.
[13] M. Mursaleen, V. Karakaya, M. Ertürk, and F. Gürsoy, Weighted statistical con-
vergence and its application to Korovkin type approximation theorem. Appl. Math.
Comput. 218 (2012), 9132–9137.
[14] I. Niven and H.S. Zuckerman, An Introduction to the Theory of Numbers. John Wiley
& Sons, Fourth Ed., New York, 1980.
[15] H. Steinhaus, Sur la convergence ordinaire et la convergence asymptotique. Colloq.
Math. 2 (1951), 73–74.

Kamil Demirci, Sevda Orhan and Burçak Kolay

Sinop University
Faculty of Sciences and Arts
Department of Mathematics
57000 Sinop, Turkey
e-mail: [email protected]
[email protected]
[email protected]
Operator Theory:
Advances and Applications, Vol. 267, 141–166

c Springer International Publishing AG, part of Springer Nature 2018

Descriptor Systems Under Feedback

and Output Injection
Marija Dodig

Abstract. In this paper we study simultaneous feedback and output injection

on descriptor linear system described by a quadruple of matrices (E, A, B, C).
We describe the possible Kronecker invariants of the resulting pencil λE −(A+
BF + KC), when F and K vary, in the case when the pencil corresponding
to the system (E, A, B, C) has no infinite elementary divisors of the second,
third and fourth type. The solution is constructive and explicit, and is given
over algebraically closed fields.
Mathematics Subject Classification (2010). Primary 15A83; Secondary 93C05,
15A22.
Keywords. Descriptor systems, matrix pencils completion, state feedback, out-
put injection, Carlson problem.

1. Introduction
Let F be a field. Let E, A ∈ Fτ ×π , B ∈ Fτ ×η and C ∈ F×π be matrices. We
consider a time-invariant linear system S with outputs
E ẋ(t) = A x(t) + B u(t),
y(t) = C x(t).
Such a system, determined by the quadruple (E, A, B, C), is called a descriptor
system (see, e.g., [22]). The corresponding matrix pencil of the system S is
 
λE − A −B
. (1.1)
−C 0
We emphasize that throughout the paper matrix E does not need to be invertible,
and not even a square matrix.
This work was done within the activities of CEAFEL and was partially supported by Fundação
para a Ciência e a Tecnologia (FCT), project ISFL-1-1431. The author was supported by the FCT
Exploratory Grant “Matrix Completions”, number IF/01232/2014. This work was also partially
supported by the Ministry of Education, Science, and Technological Development of the Republic
of Serbia, project No. 174020.
142 M. Dodig

Our aim is to study the properties of descriptor systems obtained by simulta-

neous proportional feedback and proportional output injection, i.e., we study the
possible Kronecker invariants of a pencil λE − (A + BF + KC), when F and K
vary.
We note that without loss of generality we shall assume that matrices B and
C are of the full column and row rank, respectively.
Due to generality of the problem, only some partial solutions are known,
like in the case C = 0 [8, 10] or in the regular case [26]. In addition, only some
of the possible invariants of the pencil λE − (A + BF + KC) are determined,
like the eigenvalues, i.e., poles and zeros of the corresponding system, see, e.g.,
[1, 17, 18, 21, 28].
The main difficulty of the problem lies in the fact that simultaneous feedback
and output injection on the system (E, A, B, C) imply dealing with special canon-
ical form of the pencil (1.1) (for all details see Section 2, and references therein),
which has nine different types of invariants: invariant factors, column minimal in-
dices of the first and of the second type, row minimal indices of the first and of the
second type, and infinite elementary divisors of the first, second, third and fourth
type.
In this paper we consider a system (E, A, B, C) with restriction only on the
type of infinite elementary divisors involved – we study the case when the corre-
sponding pencil has no infinite elementary divisors of the second, third and fourth
type.
Apart of theoretical Matrix theory importance, this problem has applications
in Control theory. If E = I it appears naturally in the action of the Morse group,
see [24, 25]. Also, the pencil λI − (A + BF + KC) arise in combining Kalman
filter and Linear-quadratic Regulator (LQR) control in Linear-quadratic Gaussian
(LQG) control, for details see, e.g., [5, 14, 29, 30].
The LQG control problem studies linear systems with incomplete state infor-
mation (i.e., not all the state variables are measured and available for feedback)
and undergoing control subject to quadratic costs. Finally the LQG controller is
also fundamental to the optimal control of perturbed non-linear systems. For all
details and possible applications see, e.g., [4, 5, 14, 29, 30].
As the main result of the paper, we solve the following problem over an
algebraically closed field.

Problem 1. Describe the possible Kronecker invariants of

λE − (A + BF + KC),
when F and K vary, in the case when the pencil (1.1) has no infinite elementary
divisors of the second, third and fourth type.

The obtained solution is constructive and explicit.

If E = I, as a corollary of our main result we obtain a full description for
the pencil λI − (A + BF + KC) to be regular (see also [2, 26, 27]). Note that this
 Descriptor Systems, Feedback and Output Injection 143

result has possible applications in studying transfer functions of LQG, for details
see, e.g., [29, 30].
We solve Problem 1 by presenting it as a matrix pencil completion problem
(see Section 3). To that end we use a proportional feedback and proportional
output injection equivalence on (1.1), that preserves the strict equivalence class of
the pencil λE − (A + BF + KC), and the corresponding Kronecker-like canonical
form given in [22] (see also [6, 19, 21, 23]) – see Section 2. In the solution we
use recent combinatorial results in matrix pencils completion that also involve a
solution to the Carlson problem, Young diagrams and LR sequences, obtained in
[9, 11] and given in Sections 4 and 5. In order to apply the results from [9, 11]
we need the field F to be algebraically closed. Thus, from now on we shall assume
that F is an algebraically closed field.

2. Feedback equivalence and a canonical form

Let (E, A, B, C) and (E  , A , B  , C  ) be two quadruples of matrices, both from
Fτ ×π × Fτ ×π × Fτ ×η × F×π .
Let  
λE − A −B
rank = w.
−C 0
The feedback equivalence relation obtained by proportional feedback and pro-
portional output injection is given in [22] (see also [19]), by the following definition.
Definition 2.1. Two descriptor systems given by quadruples (E, A, B, C) and (E  ,
A , B  , C  ) are feedback equivalent if there exist invertible matrices P , Q, R and
T , and matrices X and Y of appropriate sizes, such that
E  = P EQ,
A = P (A + BX + Y C)Q,
B  = P BT,
C  = RCQ.
Equivalently:
     
λE  − A −B  P PY λE − A −B Q 0
= . (2.1)
−C  0 0 R −C 0 XQ T
Lemma 2.2. Let
(E, A, B, C) ∈ Fτ ×π × Fτ ×π × Fτ ×η × F×π
and
(E  , A , B  , C  ) ∈ Fτ ×π × Fτ ×π × Fτ ×η × F×π
be two feedback equivalent descriptor systems. Let M (λ) ∈ F[λ]τ ×π be a matrix
pencil. There exist matrices F and K such that λE − (A + BF + KC) is strictly
equivalent to M (λ), if and only if, there exist matrices F  and K  such that λE  −
(A + B  F  + K  C  ) is strictly equivalent to M (λ).
144 M. Dodig

Proof. Since (E, A, B, C) and (E  , A , B  , C  ) are feedback equivalent, there exist

invertible matrices P , Q, R and T , and matrices X and Y , such that (2.1) holds.
Let F ∈ Fη×π and K ∈ Fτ × be matrices. Then we define
F  = T −1 (F − X)Q, K  = P (K − Y )R−1 .
It is straightforward to see that
λE  − (A + B  F  + K  C  ) = P (λE − (A + BF + KC))Q,
as wanted. 

The canonical form and invariants for the equivalence given in Definition 2.1
are obtained in [22]. For details and motivations see also [6, 19, 21, 23]. These
invariants are:
1. Invariant factors: ᾱ1 | · · · | ᾱw ,
2. Column minimal indices of the 1st type:
c11 + 1 ≥ · · · ≥ c1m1 + 1 > 0,
3. Column minimal indices of the 2nd type:
c21 ≥ · · · ≥ c2ρ > c2ρ+1 = · · · = c2m2 = 0,
4. Row minimal indices of the 1st type:
r11 + 1 ≥ · · · ≥ rp11 + 1 > 0,
5. Row minimal indices of the 2nd type:
r12 ≥ · · · ≥ rθ2 > rθ+1
2
= · · · = rp22 = 0,
6. Infinite elementary divisors of the 1st type:
b11 ≥ · · · ≥ b1ν > b1ν+1 = · · · = b1b1 = 1,
7. Infinite elementary divisors of the 2nd type:
b21 + 1 ≥ · · · ≥ b2b2 + 1 ≥ 1,
8. Infinite elementary divisors of the 3rd type:
b31 + 1 ≥ · · · ≥ b3b3 + 1 ≥ 1,
9. Infinite elementary divisors of the 4th type:
b41 + 2 ≥ · · · ≥ b4b4 + 2 ≥ 2.
Here we have

w 
m1 
m2 
p1 
p2
w= d(ᾱi ) + (c1i + 1) + c2i + (ri1 + 1) + ri2
i=1 i=1 i=1 i=1 i=1

b1 
b2 
b3 
b4
+ b1i + (b2i + 1) + (b3i + 1) + (b4i + 2)
i=1 i=1 i=1 i=1
 Descriptor Systems, Feedback and Output Injection 145

and
τ = w + p2 − b 3 − b 4 ,
π = w + m2 − b 2 − b 4 ,
η = m1 + b 2 + b 4 ,
 = p1 + b 3 + b 4 .
Let
n = w − m1 − p1 − b2 − b3 − 2b4 .
Thus we can define polynomials α1 | · · · |αn by
αi = ᾱi+m1 +p1 +b2 +b3 +2b4 , i = 1, . . . , n,
and we also have
n 
m1 
m2 
p1 
p2 
b1 
b2 
b3 
b4
n= d(αi ) + c1i + c2i + ri1 + ri2 + b1i + b2i + b3i + b4i .
i=1 i=1 i=1 i=1 i=1 i=1 i=1 i=1 i=1
In this paper we are interested in pencils when b2 = b3 = b4 = 0. Thus, from
now on we have
τ = n + p 1 + m1 + p 2 ,
π = n + p 1 + m1 + m2 ,
η = m1 ,
 = p1 .
Throughout the paper we shall consider only monic polynomials. For any
polynomial f , d(f ) denotes its degree. If
f = λn − an−1 λn−1 − · · · − a1 λ − a0 ,
then the matrix  T
en2 · · · enn a
is called the companion matrix of the polynomial f . Here, eni is the ith column of
 T
the identity matrix In and a = a0 · · · an−1 . Also, we denote
 n T
C(f ) := λI − e2 · · · enn a .
Further on in the paper, we shall use the following notation
N = diag(C(α1 ), . . . , C(αn )),
5) 1
* ) 1
* 6
C(λb1 ) eb11 C(λbν ) eb1ν
B1 = diag b 1 , . . . , b1 , Ib1 −ν ,
(e11 )T 0 (e1ν )T
1 1
C 1 = diag(C(λc1 +1 ), . . . , C(λcm1 +1 )),
1 1
R1 = diag(C(λr1 +1 ), . . . , C(λrp1 +1 )),
' '7 2
8 7 2
8( (
C2 = diag C(λc1 ) ec21 , . . . , C(λcρ ) ec2ρ 0 ,
146 M. Dodig

where the number of zero columns in C2 is m2 − ρ. Also,

⎛ 5) 2
* ) 2
*6 ⎞
C(λr1 ) C(λrθ )
⎜ diag r2 ,..., r2 ⎟
R2 = ⎝ (e11 )T (e1θ )T ⎠,
0

where the number of zero rows in R2 is p2 − θ. Also,

r 1 +1 T rp11 +1 T
G1 = diag((e11 ) , . . . , (e1 ) ), E1 = diag(ec11 +1 , . . . , ec1m +1 ),
1

where by ei we have denoted the ith column of the unit matrix Ii ∈ Fi×i .
The canonical form (Kronecker canonical form) for the pencil
 
λE − A −B
(2.2)
−C 0

without infinite elementary divisors of the second, third and fourth type, for feed-
back equivalence given in Definition 2.1, is the following (for details see [19, 22]):

⎡ ⎤
N
⎢ B1 ⎥
⎢ ⎥
⎢ C 1
E1 ⎥
⎢ ⎥
⎢ C2 ⎥. (2.3)
⎢ ⎥
⎢ R1 ⎥
⎢ ⎥
⎣ R2 ⎦
G1
By Lemma 2.2, without loss of generality, from now on we can assume that
the system (E, A, B, C), i.e., the corresponding pencil (2.2), is in the canonical
form (2.3). This means that from now on
⎡ ⎤
N
⎢ B1 ⎥
⎢ ⎥
⎢ C 1 ⎥
λE − A = ⎢ ⎢ ⎥,
C2 ⎥
⎢ ⎥
⎣ R 1 ⎦
R2
⎡ ⎤
0
⎢ 0 ⎥
⎢ ⎥
⎢ E1 ⎥  
−B = ⎢
⎢
⎥
⎥ and −C = 0 0 0 0 G1 0 .
⎢ 0 ⎥
⎣ 0 ⎦
0
 Descriptor Systems, Feedback and Output Injection 147

3. Re-writing the main problem as a matrix pencil

completion problem
Denote by A(λ) ∈ F[λ](n+p1 +p2 )×(n+m1 +m2 ) the subpencil of (2.3) obtained from
(2.3) after erasing rows and columns corresponding to the units in E1 and G1 , i.e.,
⎡ ⎤
N
⎢ B1 ⎥
⎢ ⎥
⎢ C ⎥
⎢
A(λ) = ⎢ 1 ⎥ (3.1)
C ⎥
⎢ 2 ⎥
⎣ R1 ⎦
R2
with
' '7 8 7 8( (
1 c1
C1 = diag C(λc1 ) ec11 , . . . , C(λ ρ ) ec1 0 ,
ρ


where ρ = > 0}, and the number of zero columns is m1 − ρ , and
max{i|c1i
⎛ 5) * ) *6 ⎞
r1 rθ1 T
(e11 )T (e )
⎜ diag 1 ,..., 1
1 ⎟
R1 = ⎝ C(λr1 ) C(λrθ ) ⎠,
0
where θ = max{i|ri1 > 0}, and the number of zero rows in R1 is p1 − θ .
The pencil A(λ) has the following invariants for the equivalence relation given
by (2.1):
1. rank A(λ) = n,
2. invariant factors: α1 | · · · | αn ,
3. column minimal indices of the 1st type:
c11 ≥ · · · ≥ c1m1 ≥ 0,
4. column minimal indices of the 2nd type:
c21 ≥ · · · ≥ c2ρ > c2ρ+1 = · · · = c2m2 = 0,
5. row minimal indices of the 1st type:
r11 ≥ · · · ≥ rp11 ≥ 0,
6. row minimal indices of the 2nd type:
r12 ≥ · · · ≥ rθ2 > rθ+1
2
= · · · = rp22 = 0,
7. infinite elementary divisors of the 1st type:
b11 ≥ · · · ≥ b1b1 ≥ 1.
By the form of the pencil (2.3), it is straightforward to see that the problem of
describing the possible Kronecker invariants of the pencil λE−(A+BF +KC) when
F and K vary, is equivalent to the following completion problem: complete A(λ)
by m1 rows and p1 columns with the fixed positions of λ’s (each added row and
column has a unique λ corresponding to each column and row minimal index of the
148 M. Dodig

second kind, respectively), up to a pencil strictly equivalent to λE−(A+BF +KC).

More precisely, let
r 1 +1 r 1 +1
E2 = diag(e11 , . . . , e1p1 )
and
E3 = diag((ec11 +1 )T , . . . , (ec1m T
+1 ) ).
1

Let

n 
b1 
m1
d(αj ) = t, b1i = b, c1i = u1 ,
j=1 i=1 i=1

m2 
p1 
p2
c2i = u2 , ri1 = v1 , ri2 = v2 .
i=1 i=1 i=1
Then Problem 1 is equivalent to the following problem.
Problem 2 (Problem 1 written as a Completion Problem). Find necessary and
sufficient conditions for the existence of matrices
X1 ∈ Ft×p1 , X2 ∈ Fb×p1 , X3 ∈ Fu1 ×p1 ,
u2 ×p1
X4 ∈ F , X5 ∈ F (v1 +p1 )×p1
, X6 ∈ F(v2 +p2 )×p1 ,
m1 ×t m1 ×b
Y1 ∈ F , Y2 ∈ F , Y3 ∈ Fm1 ×(u1 +m1 ) ,
m1 ×(u2 +m2 ) m1 ×v1
Y4 ∈ F , Y5 ∈ F , Y6 ∈ Fm1 ×v2 ,
and Y7 ∈ Fm1 ×p1 , such that the pencil
⎡ ⎤
N X1
⎢ B1 X2 ⎥
⎢ ⎥
⎢ C1 X3 ⎥
⎢ ⎥
⎢ C2 X4 ⎥ (3.2)
⎢ ⎥
⎢ R1 λE2 + X5 ⎥
⎢ ⎥
⎣ R2 X6 ⎦
Y1 Y2 λE3 + Y3 Y4 Y5 Y6 Y7
has a prescribed set of Kronecker invariants.
By a partition we mean an ordered nonincreasing sequence of nonnegative
integers. Also, if (a1 , . . . , as ) is a partition, then we assume ai = +∞, for i ≤ 0,
and ai = −∞, for i ≥ s + 1.
In the rest of the paper we shall use the following notation.
Let f̄ = (f¯1 , . . . , f¯p1 +p2 ), where f¯1 ≥ · · · ≥ f¯p1 +p2 is the nonincreasing order-
ing of
(r11 + 1, . . . , rp11 + 1) ∪ (r12 , . . . , rp22 ),
and let f = (f1 , . . . , fm1 +m2 ), where f1 ≥ · · · ≥ fm1 +m2 is the nonincreasing
ordering of
(c11 + 1, . . . , c1m1 + 1) ∪ (c21 , . . . , c2m2 ).
Now we can state our main result.
 Descriptor Systems, Feedback and Output Injection 149

Theorem 3.1. Let (E, A, B, C) be a descriptor system without infinite elementary

divisors of the second, third and fourth type. Let ᾱ1 | · · · |ᾱn+m1 +p1 be its invariant
factors; let
αi = ᾱi+m1 +p1 , i = 1, . . . , n;
let

c11 + 1 ≥ · · · ≥ c1m1 + 1 > 0 and c21 ≥ · · · ≥ c2ρ > c2ρ+1 = · · · = c2m2 = 0

be its column minimal indices of the first and of the second type, respectively; let

r11 + 1 ≥ · · · ≥ rp11 + 1 > 0 and r12 ≥ · · · ≥ rθ2 > rθ+1

2
= · · · = rp22 = 0

be its row minimal indices of the first and second type, respectively; and let
b11 ≥ · · · ≥ b1b1 be its infinite elementary divisors of the first type. Suppose that
γ1 | · · · |γn+m1 +p1 are polynomials and

b̄1 ≥ · · · ≥ b̄b1 ,
d1 ≥ · · · ≥ dρ̄ > dρ̄+1 = · · · = dm2 = 0,
r̄1 ≥ · · · ≥ r̄θ̄ > r̄θ̄+1 = · · · = r̄p2 = 0

are nonnegative integers.

There exist matrices F and K such that pencil λE − (A + BF + KC) has
γ1 | · · · |γn+m1 +p1 , b̄1 ≥ · · · ≥ b̄b1 , d1 ≥ · · · ≥ dρ̄ > dρ̄+1 = · · · = dm2 = 0 and
r̄1 ≥ · · · ≥ r̄θ̄ > r̄θ̄+1 = · · · = r̄p2 = 0 as Kronecker invariants (invariant factors,
infinite elementary divisors, column and row minimal indices, respectively), if and
only if the following conditions hold:
(i) ρ ≤ ρ̄ ≤ ρ + m1 and θ ≤ θ̄ ≤ θ + p1 ,

(ii) b̄i = b1i for all i = 1, . . . , b1 ,

(iii) γi | αi | γi+m1 +p1 for all i = 1, . . . , n,

(iv) r̄i ≥ ti for all i = 1, . . . , θ̄,

(v) di ≥ si for all i = 1, . . . , ρ̄,


n 
m1 
m2 
p1 
p2
(vi) d(αi ) + (c1i + 1) + c2i + (ri1 + 1) + ri2
i=1 i=1 i=1 i=1 i=1

n+m 1 +p1

m2 
p2
= d(γi ) + di + r̄i ,
i=1 i=1 i=1


n+m 1 +p1 +p1 −k−j
n+m1
(vii) xj + yk ≤ d(γi ) − d(lcm(αi−m1 −p1 +k+j , γi ))
i=1 i=1
for all j = 0, . . . , m1 and k = 0, . . . , p1 ,
150 M. Dodig

where
hj −j

hj

xj = fi − di , j = 0, . . . , m1 , hj = min{i|di−j+1 < fi }, h0 := 0,
i=1 i=1


h̄k k −k
h̄
yk = f¯i − r̄i , k = 0, . . . , p1 , h̄k = min{i|r̄i−k+1 < f¯i }, h̄0 := 0,
i=1 i=1

and s1 ≥ · · · ≥ sρ̄ , and t1 ≥ · · · ≥ tθ̄ are nonincreasing orderings of

(c21 , . . . , c2ρ ) ∪ (c1m1 + 1, . . . , c1m1 +ρ+1−ρ̄ + 1)
and
(r12 , . . . , rθ2 ) ∪ (rp11 + 1, . . . , rp11 +θ+1−θ̄ + 1),
respectively.
Before passing to the proof, we give some auxiliary results in Sections 4 and 5
which will be used in the course of the proof.

4. Auxiliary results
If ψ1 | · · · |ψn is a polynomial chain (e.g., the list of invariant factors of a rank n
pencil), then we make the convention that ψi = 1, for all i ≤ 0, and ψi = 0, for all
i ≥ n + 1.
In the paper, we shall use a particular case of [9, Theorem 3]. If in [9, Theorem
3], we additionally have the assumption

p 
u+y
(ri + 1) = d(γ̃i ) − xl ,
i=1 i=1

then condition (41) in [9, Theorem 3] is equal to condition (42) from the same
theorem, for k = 0. So under this assumption, instead of having two, we only have
one condition.
So, if we introduce new notation and put

u+y 
p
yk = d(γi ) − xl − (ri + 1),
i=1 i=k+1

then the additional assumption becomes y0 = 0, and we can write this particular
case of [9, Theorem 3] in the following way, by using the notation appropriate to
this paper.
Theorem 4.1. Let n, p1 and m1 be nonnegative integers. Let x0 , x1 , . . . , xm1 and
y0 , y1 , . . . , yp1 be integers with x0 = y0 = 0, and let γ1 | · · · |γn+m1 +p1 be polynomials
from F[λ], such that γn = 1.
 Descriptor Systems, Feedback and Output Injection 151

There exist homogeneous polynomials β1 | · · · |βn+p1 which satisfy βn = 1 and

1
n+p
yk ≤ d(βi ), k = 1, . . . , p1 ,
i=n+p1 −k+1

 
n+m 1 +p1

d(lcm(βn+p1 +1−i , γn+m1 +p1 −j+1−i )) ≤ d(γi ) − xj , j = 0, . . . , m1 ,

i≥1 i=1

1
n+p 
n+m 1 +p1

d(βi ) = d(γi ) − xm1 ,

i=1 i=1
γi | βi | γi+m1 , i = 1, . . . , n + p1 ,
if and only if

n+m 1 +p1

xj + yk ≤ d(γi ), j = 0, . . . , m1 , k = 0, . . . , p1 .
i=n+m1 +p1 −j−k+1

Definition 4.2. Let α : α1 | · · · |αn and γ : γ1 | · · · |γn+m be two polynomial chains.

By the minimal path between α and γ we mean the polynomial chain σ(α, γ) :
σ1 (α, γ)| · · · |σm (α, γ), where
πi 
i+n
σi (α, γ) = , i = 1, . . . , m, πi = lcm(αj−i , γj ), i = 0, . . . , m.
πi−1 j=1

For more details see, e.g., [31].

Lemma 4.3 ([11, Lemma 3]). Let α, β, γ ∈ F[λ]. Then
lcm(α, β) lcm(γ, β)
lcm(α, γ) | .
β
In [12] we have introduced the following definition of majorization that deals
with three different partitions of integers.
Definition 4.4. Consider partitions a = (a1 , . . . , as ), d = (d1 , . . . , dm ) and g =
(g1 , . . . , gm+s ). If
di ≥ gi+s , i = 1, . . . , m, (4.1)
hj −j

hj
 
j
gi − di ≤ ai , hj := min{i|di−j+1 < gi }, j = 1, . . . , s, (4.2)
i=1 i=1 i=1

m+s 
m 
s
gi = di + ai , (4.3)
i=1 i=1 i=1
then we say that g is majorized by the pair (d, a). We call this type of majorization
the generalized majorization, and we write
g ≺ (d, a).
152 M. Dodig

Lemma 4.5. Let a = (a1 , . . . , as ) (with a1 ≥ · · · ≥ as ≥ 0), d = (d1 , . . . , dm )

(with d1 ≥ · · · ≥ dρ̄ > dρ̄+1 = · · · = dm = 0) and g = (g1 , . . . , gm+s ) (with
g1 ≥ · · · ≥ gg > gg+1 = · · · = gm+s = 0) be partitions such that
g ≺ (d, a). (4.4)
Then
ρ̄ ≤ g ≤ ρ̄ + s. (4.5)
Moreover, we have
g = ρ̄ + χ, (4.6)
where χ = {i ∈ {1, . . . , s}|ghi > 0}.
Convention: h0 := 0, hs+1 := m + s + 1.
Proof. By definition of the generalized majorization, (4.4) gives (4.1), (4.2) and
(4.3). We directly obtain that (4.1) implies
g − s ≤ ρ̄.
Now, let us suppose that χ > 0. Then (4.2) and (4.3) give

m+s 
m 
s
gi ≥ di + ai . (4.7)
i=hχ +1 i=hχ −χ+1 i=χ+1

Also, we have
hχ+1 −χ−1 hχ+1 −1
 
di ≥ gi . (4.8)
i=hχ −χ+1 i=hχ +1
Indeed, if χ = s, then (4.8) follows from (4.1), and if χ < s, then (4.8) follows
from the definition of hχ+1 .
Since ghχ+1 = · · · = gm+s = 0 (from the definition of χ), equations (4.7) and
(4.8) give
hχ+1 −χ−1 hχ+1 −1
  
m 
s 
m
di ≥ gi ≥ di + ai ≥ di . (4.9)
i=hχ −χ+1 i=hχ +1 i=hχ −χ+1 i=χ+1 i=hχ −χ+1

Thus we must have equalities in (4.9), and so we have that

dhχ+1 −χ = · · · = dm = 0, aχ+1 = · · · = as = 0,
as well as
di−χ = gi , i = hχ + 1, . . . , hχ+1 − 1.
Thus, we obtain
ρ̄ = g − χ ≤ g,
as wanted.
Now, suppose that χ = 0. Then, gh1 = 0, and so we have that h1 = m + 1, as
well as di ≥ gi , i = 1, . . . , m. Since (4.3) is valid, we have that in this case gi = di
for all i = 1, . . . , m, as well as a1 = · · · = as = 0. Hence, we directly obtain ρ̄ = g.
Altogether, we have proven (4.5) and (4.6), as wanted. 
 Descriptor Systems, Feedback and Output Injection 153

By using the concept of the generalized majorization and the minimal path,
together with the notation introduced in Section 2, by [13, Remark 4] and by
Lemma 4.5, we obtain that [10, Theorem 2] is equivalent to the following.
Theorem 4.6. Consider the pencil A(λ) = diag(B, N, C1 , C2 , R1 , R2 ) as in (3.1).
Let β1 | · · · |βn+m1 be polynomials, and let
b̃1 ≥ · · · ≥ b̃b1 , r̃1 ≥ · · · ≥ r̃p1 +p2 ,
and
g1 ≥ · · · ≥ gρ̄ > gρ̄+1 = · · · = gm2 = 0
be nonnegative integers. There exist matrices Y1 , . . . , Y6 such that the pencil
⎡ ⎤
B
⎢ N ⎥
⎢ ⎥
⎢ C1 ⎥
⎢ ⎥
⎢ C2 ⎥ ∈ F[λ](n+p1 +p2 +m1 )×(n+m1 +m2 )
⎢ ⎥
⎢ R1 ⎥
⎢ ⎥
⎣ R2 ⎦
Y1 Y2 λE3 + Y3 Y4 Y5 Y6
has β1 | · · · |βn+m1 as invariant factors, b̃1 ≥ · · · ≥ b̃b1 as degrees of infinite ele-
mentary divisors, r̃1 ≥ · · · ≥ r̃p1 +p2 as row minimal indices, and
g1 ≥ · · · ≥ gρ̄ > gρ̄+1 = · · · = gm2 = 0
as column minimal indices, if and only if the following conditions are valid:
(i) b̃i = b1i for all i = 1, . . . , b1 ,
(ii) (r̃1 , . . . , r̃p1 +p2 ) = (r11 , . . . , rp11 ) ∪ (r12 , . . . , rp22 ),
(iii) gi ≥ si for all i = 1, . . . , ρ̄,
(iv) βi |αi |βi+m1 for all i = 1, . . . , n,
(v) f ≺ (g, σ(α, β)),
where f = (f1 , . . . , fm1 +m2 ) with
(f1 , . . . , fm1 +m2 ) = (c11 + 1, . . . , c1m1 + 1) ∪ (c21 , . . . , c2m2 ),
(s1 , . . . , sρ̄ ) = (c21 , . . . , c2ρ ) ∪ (c1m1 + 1, . . . , c1m1 +ρ+1−ρ̄ + 1),
and g = (g1 , . . . , gm2 ).

5. Carlson problem and combinatorial results

In the following sections, we shall use the concept of Young diagrams. A Young
diagram is a finite collection of boxes, whose rows are aligned on the left, while the
row lengths (numbers of boxes in a row) are weakly decreasing. In particular, a
weakly decreasing sequence of nonnegative integers a = (a1 , a2 , . . . , an ) bijectively
corresponds to a Young diagram, denoted Ya , which has ai boxes in the ith row.
154 M. Dodig

Definition 5.1 ([11]). Let a = (a1 , a2 , . . . , an ) and c = (c1 , c2 , . . . , cn+m ) be parti-

tions such that ci ≥ ai ≥ ci+m , i = 1, . . . , n. By the minimal path between a and
c we mean the partition s = (s1 , s2 , . . . , sm ), given by
n n
si := cn+i + max(aj , cj+i−1 ) − max(aj , cj+i ), i = 1, . . . , m,
j=1 j=1

and we denote it by s(a, c).

5.1. Littlewood–Richardson coefficients and Carlson problem
We shall recall the definition of the Littlewood–Richardson coefficients, for details
see, e.g., [15, 20].
Let a = (a1 , a2 , . . . , an ), b = (b1 , b2 , . . . , bm ), c = (c1 , c2 , . . . , cn+m ) be weakly
decreasing sequences of nonnegative integers such that

n+m 
n 
m
ci = ai + bi ,
i=1 i=1 i=1
and the Young diagram of  a is contained in that for c, i.e., ai ≤ ci . Thus, the skew
m
diagram c \ a, consists of i=1 bi boxes.
Order the boxes of c \ a by first enumerating the boxes in the top row (from
right to left), then the ones in the second row (from right to left), and so on down
c
the array. Then the Littlewood–Richardson coefficient LRa,b is the number of ways
to fill the boxes of c \ a with the integers from {1, . . . , m}, so that the following
conditions are satisfied:
(a) entries in any row are weakly increasing from left to right;
(b) entries in the each column are strictly increasing from top to bottom;
(c) the integer i occurs exactly bi times,
for i ∈ {1, . . . , m};
m
(d) for any integer p, with 1 ≤ p < i=1 bj , and any positive integer i, the
number of times i occurs in the first p boxes of the ordering is, at least, as
large as the number of times that i + 1 occurs in these first p boxes.
c
Note that the Littlewood–Richardson coefficient LRa,b is positive if and only
if there exists at least one filling of the boxes of c \ a that satisfies the conditions
(a)–(d).
Throughout the paper we will deal with (chains of) polynomials from F[λ].
Definition 5.2. Let α : α1 | · · · |αn be a polynomial chain, and let ψ ∈ F[λ] be an
irreducible polynomial. Let ai denote the degree of ψ in αn+1−i , i = 1, . . . , n, i.e.,
αn+1−i = ψ ai qi , i = 1, . . . , n,
where qi ∈ F[λ] are not divisible by ψ, for any i = 1, . . . , n. Then, a = (a1 , . . . , an )
is called the partition of the ψ-elementary divisors of α.
The following proposition is straightforward.
Proposition 5.3. Let α : α1 | · · · |αn and γ : γ1 | · · · |γn+m be two polynomial chains.
Let ψ be an irreducible polynomial. Let σ = σ(α, γ) be the minimal path between
 Descriptor Systems, Feedback and Output Injection 155

α and γ. Let a = (a1 , . . . , an ), c = (c1 , . . . , cn+m ) and s = (s1 , . . . , sm ) be the

partitions of the ψ-elementary divisors of α, γ and σ, respectively. Then we have
s = s(a, c).
Lemma 5.4 ([11, Lemma 4]). Let a = (a1 , . . . , an ) and c = (c1 , . . . , cn+m+p )
be partitions such that ci ≥ ai ≥ ci+m+p , i = 1, . . . , n. Let s = s(a, c) =
(s1 , s2 , . . . , sm+p ) be the minimal path between them. Let b = (b1 , . . . , bm ) be a
partition, such that si ≥ bi ≥ si+p , i = 1, . . . , m. Then there exists a partition
d = (d1 , d2 , . . . , dn+m ), such that
ci ≥ di ≥ ai , i = 1, . . . , n + m,
and the following conditions are valid:
s(d, c) = s(b, s), and so di ≥ ci+p , i = 1, . . . , n + m,
d
LRa,b > 0.
Carlson problem. In [15], the Littlewood–Richardson coefficient is related to ma-
trix completion problems. In particular, it appears in the necessary and sufficient
condition for a solution of the classical completion problem – the Carlson problem,
see [3]. Carlson problem consists of determining the possible invariant polynomials
of the square matrix
 
A 0
C= ∈ F(n+m)×(n+m)
X B
where A ∈ Fn×n and B ∈ Fm×m are fixed matrices, and X ∈ Fm×n varies.
In [11] we have given the following result.
Theorem 5.5 ([11, Theorem 7]). Let A(λ) ∈ F[λ]n×n , B(λ) ∈ F[λ]m×m be regular
matrix pencils, with α : α1 | · · · |αn and β : β1 | · · · |βm , as homogeneous invari-
ant factors, respectively. Let γ : γ1 | · · · |γn+m be homogeneous polynomials. Let
ψ0 , ψ1 , . . . , ψk be the irreducible factors of αn βm . For every i = 0, . . . , k, denote
by ai , bi and ci the partitions corresponding to the ψi -elementary divisors of α, β
and γ, respectively.
i
If for every i = 0, . . . , k, the Littlewood–Richardson coefficient LRac i ,bi is pos-
itive, then there exists a matrix pencil X(λ) ∈ F[λ]n×m , such that the homogeneous
invariant polynomials of the regular pencil
 
A(λ) X(λ)
C(λ) =
0 B(λ)
are γ1 | · · · |γn+m .
Now we can give a proof of Theorem 3.1. In Subsection 6.1 we prove the
necessity of conditions (i)–(vii). The proof of sufficiency of the conditions is given
in Subsection 6.2. First, in Subsection 6.2.1 we prove the sufficiency of conditions
(i)–(vii) in the case when αn = 1, and finally, in Subsection 6.2.2 we prove the
sufficiency of conditions (i)–(vii) without restrictions.
156 M. Dodig

6. Proof of Theorem 3.1

Recall that in Section 3, we have proved that Problem 1 is equivalent to a matrix
pencil completion problem (Problem 2). Thus, in order to prove Theorem 3.1,
it is enough to prove that conditions (i)–(vii) are necessary and sufficient for
the existence of matrices X1 , . . . , X6 , Y1 , . . . , Y7 such that the pencil (3.2) has
γ1 | · · · |γn+m1 +p1 as its invariant factors, b̄1 , . . . , b̄b1 as degrees of its infinite el-
ementary divisors, d1 , . . . , dm2 as its column minimal indices, and r̄1 , . . . , r̄p2 as its
row minimal indices.
First consider the column completion of the pencil diag(N, B1 , C1 , C2 , R1 , R2 )
given by
⎡ ⎤
N X1
⎢ B1 X2 ⎥
⎢ ⎥
⎢ C1 X3 ⎥
⎢ ⎥ (6.1)
⎢ C2 X4 ⎥
⎢ ⎥
⎣ R1 λE2 + X5 ⎦
R2 X6

Denote the invariant factors of (6.1) by β1 | · · · |βn+p1 , row minimal indices of

(6.1) by r̃1 ≥ · · · ≥ r̃θ̃ > r̃θ̃+1 = · · · = r̃p2 = 0, column minimal indices by
g1 ≥ · · · ≥ gm1 +m2 , and degrees of infinite elementary divisors by b̃1 ≥ · · · ≥ b̃b1 .
By applying transposed version of Theorem 4.6, we obtain that this completion
exists if and only if b̃i = b1i , i = 1, . . . , b1 , g1 ≥ · · · ≥ gm1 +m2 is nonincreasing
ordering of (c11 , . . . , c1m1 ) ∪ (c21 , . . . , c2m2 ), and the following conditions are valid

r̃i ≥ ti , i = 1, . . . , θ̃, (6.2)

βi |αi |βi+p1 , i = 1, . . . , n, (6.3)

f̄ ≺ (r̃, σ(α, β)), (6.4)

with f̄ = (f¯1 , . . . , f¯p1 +p2 ), f¯1 ≥ · · · ≥ f¯p1 +p2 is nonincreasing ordering of (r11 +
1, . . . , rp11 +1)∪(r12 , . . . , rp22 ), t1 ≥ · · · ≥ tθ̃ is nonincreasing ordering of (r12 , . . . , rθ2 )∪
(rp1 +θ+1−θ̃ + 1, . . . , rp11 + 1), and r̃ = (r̃1 , . . . , r̃p2 ).
1

Now we consider completion by rows of (6.1) to (3.2). By performing strictly

equivalent operations on (3.2) we can put (6.1) in the Kronecker canonical form.
Thus, this row completion, becomes equivalent to finding necessary and sufficient
conditions for existence of matrices Ȳ1 , . . . , Ȳ5 such that the pencil
⎡ ⎤
N̄
⎢ B1 ⎥
⎢ ⎥
⎢ C ⎥
⎢ 1 ⎥
⎢ C2 ⎥
⎢ ⎥
⎣ R̄ ⎦
Ȳ1 Y¯2 λE3 + Ȳ3 Ȳ4 Ȳ5
 Descriptor Systems, Feedback and Output Injection 157

has γ1 | · · · |γn+m1 +p1 as its invariant factors, b̄1 , . . . , b̄b1 as degrees of its infinite
elementary divisors, d1 , . . . , dm2 as its column minimal indices, and r̄1 , . . . , r̄p2 as
its row minimal indices. Here N̄ = diag(C(β1 ), . . . , C(βn+p1 )),
⎛     ⎞
C(λr̃1 ) C(λr̃θ̃ )
diag ,...,
R̄ = ⎝ (e1r̃1 )T
r̃
(e1θ̃ )T ⎠,
0
where the number of zero rows in R̄ is p2 − θ̃.
As before, by Theorem 4.6 we obtain that this completion exists if and only
if the following conditions are valid:
b̄i = b1i , i = 1, . . . , b1 , (6.5)
r̄i = r̃i , i = 1, . . . , p2 , (6.6)
di ≥ si , i = 1, . . . , ρ̄, (6.7)
γi |βi |γi+m1 , i = 1, . . . , n + p1 , (6.8)

f ≺ (d, σ(β, γ)). (6.9)
Since we have (6.6), by the definition of σ(α, β), yk , and by the definition of
the generalized majorization, we have that (6.4) is equivalent to:
r̄i ≥ f¯i+p1 , i = 1, . . . , p2 , (6.10)
1 −k
1
n+p 
n+p
yk ≤ d(βi ) − d(lcm(αi−p1 +k , βi )), k = 1, . . . , p1 , (6.11)
i=1 i=1


n 
p2 
p1 1
n+p 
p2
d(αi ) + ri2 + (ri1 + 1) = d(βi ) + r̄i . (6.12)
i=1 i=1 i=1 i=1 i=1
Moreover, by Lemma 4.5 we also have
θ ≤ θ̄ ≤ θ + p1 . (6.13)
Analogously, by the definition of σ(β, γ), xj , and by the definition of the
generalized majorization, we have that (6.9) is equivalent to:

di ≥ fi+m1 , i = 1, . . . , m2 , (6.14)

n+m 1 +p1 +m1 −j
n+p1
xj ≤ d(γi ) − d(lcm(βi−m1 +j , γi ), j = 1, . . . , m1 , (6.15)
i=1 i=1

1
n+p 
m2 
m1 
n+m 1 +p1

m2
d(βi ) + c2i + (c1i + 1) = d(γi ) + di . (6.16)
i=1 i=1 i=1 i=1 i=1
We note that conditions (6.2) and (6.7) imply conditions (6.10) and (6.14),
respectively. By Lemma 4.5, we also have
ρ ≤ ρ̄ ≤ ρ + m1 . (6.17)
158 M. Dodig

By unifying these results, we obtain that the existence of matrices X1 , . . ., X6 ,

Y1 , . . ., Y7 , such that (3.2) has γ1 | · · · |γn+m1 +p1 as its invariant factors, b̄1 , . . . , b̄b1
as degrees of its infinite elementary divisors, d1 , . . . , dm2 as its column minimal
indices, and r̄1 , . . . , r̄p2 as its row minimal indices, is equivalent to the existence of
polynomials β1 | · · · |βn+p1 which satisfy the following conditions:

b̄i = b1i , i = 1, . . . , b1 , (6.18)

r̄i ≥ ti , i = 1, . . . , θ̄, (6.19)
βi |αi |βi+p1 , i = 1, . . . , n, (6.20)
f̄ ≺ (r̄, σ(α, β)), (6.21)
di ≥ si , i = 1, . . . , ρ̄, (6.22)
γi |βi |γi+m1 , i = 1, . . . , n + p1 , (6.23)
f ≺ (d, σ(β, γ)). (6.24)

Hence, we are left with proving that conditions (i)–(vii) are necessary and
sufficient for the existence of such polynomials β1 | · · · |βn+p1 .
Now we can give a proof of Theorem 3.1. We start with proving the necessity
of conditions (i)–(vii).

6.1. Necessity of conditions (i)–(vii)

Let us suppose that there exist polynomials β1 | · · · |βn+p1 which satisfy conditions
(6.18)–(6.24).
Condition (6.18) is equal to (ii). As we have showed in Lemma 4.5, conditions
(6.21) and (6.24) imply (6.13) and (6.17), respectively, thus we have (i). Conditions
(6.20) and (6.23) together give (iii). Conditions (6.19) and (6.22) are equal to (iv)
and (v), respectively. Since (6.21) is equivalent to (6.10), (6.11) and (6.12), while
(6.24) is equivalent to (6.14), (6.15) and (6.16), we have that (6.12) and (6.16) give
(vi). So we are left with obtaining the condition (vii).
By summing (6.11) and (6.15) we have that for every j = 0, . . . , m1 , and
every k = 0, . . . , p1 , the following is valid:

n+m 1 +p1 +p1 −k−j
n+m1
xj + yk ≤ d(γi ) − d(lcm(βi−m1 +j , γi ))
i=1 i=1
1 +p1 −k
n+m
− d(lcm(βi−m1 +j , γi ))
i=n+m1 +p1 −k−j+1
1 −k
1
n+p 
n+p
+ d(βi ) − d(lcm(αi−p1 +k , βi )).
i=1 i=1

By applying Lemma 4.3, on every summand of the second sum on the right-
hand side, we obtain the following inequality
 Descriptor Systems, Feedback and Output Injection 159


n+m 1 +p1 +p1 −k−j
n+m1
xj + yk ≤ d(γi ) − d(lcm(αi−m1 −p1 +k+j , γi ))
i=1 i=1
+p1 −k−j
n+m1
− d(βi−m1 +j )
i=1
+p1 −k−j
n+m1
+ d(lcm(αi−m1 −p1 +k+j , βi−m1 +j ))
i=1
1 +p1 −j
n+m
− d(lcm(βi−m1 +j , γi ))
i=n+m1 +p1 −k−j+1
1 −k
1
n+p 
n+p
+ d(βi ) − d(lcm(αi−p1 +k , βi )).
i=1 i=1

Finally, since
1
n+p 1
n+p
d(βi ) ≤ d(lcm(βi , γi+m1 −j )),
i=n+p1 −k+1 i=n+p1 −k+1

we obtain (vii), as wanted.

6.2. Sufficiency of conditions (i)–(vii)

The proof of the sufficiency of conditions (i)–(vii) will be split in two cases. First
in Subsection 6.2.1, we consider the case of Theorem 3.1 when αn = 1, and in
Section 6.2.2 we shall consider the general case, i.e., when αn = 1.

6.2.1. Case αn = 1. Let α1 = · · · = αn = 1. Let conditions (i)–(vii) be valid.

We are left with proving the existence of polynomials β1 | · · · |βn+p1 which satisfy
conditions (6.18)–(6.24).
By (iii) we have γi = 1, i = 1, . . . , n. Conditions (iv) and (v) are equal to
(6.19) and (6.22), respectively (thus implying (6.10) and (6.14), as well). Also, con-
dition (ii) is equal to (6.18). By Theorem 4.1, condition (vii) implies the existence
of polynomials β1 | · · · |βn+p1 which satisfy βn = 1 and
1
n+p
yk ≤ d(βi ), k = 1, . . . , p2 ,
i=n+p1 −k+1

 
n+m 1 +p1

d(lcm(βn+p1 +1−i , γn+p1 +m1 −j+1−i ) ≤ d(γi ) − xj , j = 0, . . . , m1 ,

i≥1 i=1

1
n+p 
n+m 1 +p1

d(βi ) = d(γi ) − xm1 , (6.25)

i=1 i=1
γi | βi | γi+m1 , i = 1, . . . , n + p1 .
160 M. Dodig

Since we are considering the case αn = 1, in this way we have obtained

polynomials β1 | · · · |βn+p1 which satisfy conditions (6.20), (6.23), (6.11), and (6.15).
Thus by the definition of generalized majorization, in order to prove (6.21) and
(6.24), we are left with proving (6.16) and (6.12). Since (vi) is valid, it is enough
to prove the validity of one of them, e.g., (6.16).
Conditions (vi), and (vii) for j = m1 and k = p1 , give
m
1 +m2 
m2 p
1 +p2

p2
0≤ fi − di + f¯i − r̄i . (6.26)
i=hm1 +1 i=hm1 −m1 +1 i=h̄p1 +1 i=h̄p1 −p1 +1

Since (6.10) and (6.14) are valid, we have that

m
1 +m2 
m2 p
1 +p2

p2
0≥ fi − di + f¯i − r̄i . (6.27)
i=hm1 +1 i=hm1 −m1 +1 i=h̄p1 +1 i=h̄p1 −p1 +1

Therefore, one must have the equality in both (6.26) and (6.27), as well as
in (6.10) for i = hm1 − m1 + 1, . . . , m2 , and in (6.14) for i = h̄p1 − p1 + 1, . . . , p2 .
Hence
h m1 hm1 −m1
  m
1 +m2 
m2
xm1 = fi − di = fi − di ,
i=1 i=1 i=1 i=1
and
h̄p1 −p1

h̄p1
 p
1 +p2

p2
yp1 = f¯i − di = f¯i − r̄i ,
i=1 i=1 i=1 i=1

which, taking into account (6.25), finishes the proof.

Now we can pass to the case αn = 1.

6.2.2. Sufficiency of conditions (i)–(vii) if αn = 1. Consider the pencil A(λ)

in its canonical form (3.1), i.e., diag(N, B1 , C1 , C2 , R1 , R2 ). By applying the re-
sult of the previous section, when αn = 1, we have that there exists matrices
X2 , . . . , X6 , Y2 , . . . , Y7 such that the pencil
⎡ ⎤
B1 X2
⎢ C1 X3 ⎥
⎢ ⎥
⎢ C X ⎥
⎢ 2 4 ⎥
⎢ R1 λE2 + X5 ⎥
⎢ ⎥
⎣ R2 X6 ⎦
Y2 λE3 + Y3 Y4 Y5 Y6 Y7
∈ F[λ](n−t+p1 +p2 +m1 )×(n−t+m1 +m2 +p1 ) , (6.28)
has rank equal to n − t + m1 + p1 , and has d1 , . . . , dm2 , as its column minimal
indices, r̄1 , . . . , r̄p2 as its row minimal indices, b̄1 , . . . , b̄b1 as degrees of its infinite
 Descriptor Systems, Feedback and Output Injection 161

elementary divisors, and

1| · · · |1 |σ1 (α, γ)| · · · |σm1 +p1 (α, γ)
A BC D
n−t

as its invariant factors.

Indeed, the pencil diag(B1 , C1 , C2 , R1 , R2 ) has the same column, row mini-
mal indices and degrees of infinite elementary divisors as A(λ) and has only triv-
ial invariant factors – exactly n − t of them. Denote those invariant factors by
α1 | · · · |αn−t = 1. Since (i)–(vii) are valid for αi ’s, γi ’s, bi ’s, c1i ’s, c2i ’s, ri1 ’s, ri2 ’s,
di ’s, b̄i ’s and r̄i ’s, from the form of the conditions, it is straightforward to see that
they are also satisfied for αi ’s and
1| · · · |1 |σ1 (α, γ)| · · · |σm1 +p1 (α, γ)
A BC D
n−t

instead of αi ’s and γi ’s, respectively, while keeping all other invariants the same.
Therefore by Subsection 6.2.1, there exists a completion (6.28).
Next, consider the subpencil
⎡ ⎤
B1 X2
⎢ C1 X3 ⎥
⎢ ⎥
⎢ C X ⎥
⎢ 2 4 ⎥
⎣ R1 λE2 + X5 ⎦
R2 X6
∈ F[λ](n−t+p1 +p2 )×(n−t+m1 +m2 +p1 ) . (6.29)
Completely analogously as in the beginning of Section 6, by applying Theo-
rem 4.6 twice – once for the completion from diag(B, C1 , C2 , R1 , R2 ) up to (6.29),
and next for the completion from (6.29) to (6.28), we obtain that (6.29) has
(c11 , . . . , c1m1 ) ∪ (c21 , . . . , c2m2 ) as its column minimal indices, r̄1 , . . . , r̄p2 as its row
minimal indices, and b̄1 , . . . , b̄b1 as degrees of its infinite elementary divisors. Also,
since (6.29) has rank equal to n − t + p1 and it has at least n − t trivial invariant
factors, we denote its invariant factors by
1| · · · |1 |β̄1 | · · · |β̄p1 .
A BC D
n−t

Now, it is straightforward to see that for the completion from (6.29) up to (6.28), we
can consider (6.29) in the canonical form: diag(B, C1 , C2 , N (β̄), R̄), where N (β̄) =
diag(C(β̄1 ), . . . , C(β̄p1 )), and
⎛     ⎞
C(λr̄1 ) C(λr̄θ̄ )
,...,
R̄ = diag ⎝ (er̄1 )T (er̄θ̄ )T ⎠,
0
where the number of the zero rows equals p2 − θ̄.
162 M. Dodig

Thus, we have that there exist Ȳ2 , . . . , Ȳ6 , such that

⎡ ⎤
B1
⎢ C1 ⎥
⎢ ⎥
⎢ C ⎥
⎢ 2 ⎥
⎢ N (β̄) ⎥,
⎢ ⎥
⎢ R̄ ⎥
⎣ ⎦
Ȳ2 λE3 + Ȳ3 Ȳ4 Ȳ5 Ȳ6
is strictly equivalent to (6.28). Thus, by applying Theorem 4.6, we obtain that the
following conditions hold:
di ≥ si , i = 1, . . . , ρ̄, (6.30)

σi (α, γ)|β̄i |σi+m1 (α, γ), i = 1, . . . , p1 , (6.31)

f ≺ (d, σ(β̄, σ(α, γ))).
Before proceeding, we note that by applying [10, Lemmas 3–4] and [7, Lemma
7], without loss of generality, we can consider X2 = 0, X3 = 0 and X4 = 0 in the
pencil (6.29).
Let ψ1 , . . . , ψk be all distinct irreducible factors of γn+m1 +p1 . For every i =
1, . . . , k, let
ai = (ai1 , ai2 , . . . , ain ),
bi = (bi1 , bi2 , . . . , bip1 ),
ci = (ci1 , ci2 , . . . , cin+m1 +p1 ),
si = (si1 , si2 , . . . , sim1 +p1 )
be the weakly decreasing partitions corresponding to the ψi -elementary divisor of
the polynomial chains α, β̄, γ and σ, respectively. More precisely:
a1 a2 ak
αn+1−j := ψ1 j ψ2 j · · · ψk j , j = 1, . . . , n,
b1j b2j bk
β̄p1 +1−j := ψ1 ψ2 · · · ψk , j
j = 1, . . . , p1 ,
c1j c2j ck
γn+m1 +p1 +1−j := ψ1 ψ2 · · · ψk , j
j = 1, . . . , n + m1 + p1 ,
s1 s2 sk
σm1 +p1 +1−j := ψ1j ψ2j · · · ψkj , j = 1, . . . , m1 + p1 .

Note that by Proposition 5.3, we have that si = s(ai , ci ), for all i. Also for
all i, by (i) we have cij ≥ aij ≥ cij+m1 +p1 , j = 1, . . . , n, and by (6.31) we have
sij ≥ bij ≥ sij+m1 , j = 1, . . . , p1 .
Now, by Lemma 5.4, for every i = 1, . . . , k, there exists a nonincreasing
partition di = (di1 , di2 , . . . , din+p1 ) such that
cij ≥ dij ≥ aij , j = 1, . . . , n + p1 , (6.32)
 Descriptor Systems, Feedback and Output Injection 163

and
s(di , ci ) = s(bi , si ), (6.33)
dij ≥ cij+m1 , j = 1, . . . , p1 , (6.34)
i
LRadi ,bi > 0. (6.35)
Finally, we define the homogeneous polynomials φ : φ1 | · · · |φn+p1 by
d1 d2 dk
φn+p1 +1−j := ψ1 j ψ2 j · · · ψk j , j = 1, . . . , n + p1 .
From the definition of the polynomials φ1 | · · · |φn+p1 and from (6.35), by
Theorem 5.5 (Carlson problem), there exists a pencil Y such that
 
N Y
0 N (β̄)
is strictly equivalent to diag(C(φ1 ), . . . , C(φn+p1 )). Thus, since (6.29) is strictly
equivalent to diag(B1 , C1 , C2 , N (β̄), R̄), we have that there exist pencils W1 and
W2 , and X1 such that the pencil
⎡ ⎤
N 0 0 0 W1 W2 X1
⎢ B 0 ⎥
⎢ ⎥
⎢ C1 0 ⎥
⎢ ⎥ (6.36)
⎢ C2 0 ⎥
⎢ ⎥
⎣ R1 λE2 + X5 ⎦
R2 X6
has φ1 | · · · |φn+p1 as its invariant factors, b̄1 , . . . , b̄b1 as degrees of its infinite elemen-
tary divisors, r̄1 , . . . , r̄p2 as its row minimal indices, and (c11 , . . . , c1m1 )∪(c21 , . . . , c2m2 )
as its column minimal indices.
By [7, Lemma 7], without loss of generality, we can consider W1 = 0 and
W2 = 0.
From the definition of φi , by (6.32) and (6.34), we have that
γi | φi | γi+m1 , i = 1, . . . , n + p1 , (6.37)
and (6.33) gives
σ(φ, γ) = σ(β̄, σ(α, γ)).
Hence (6.24) becomes
f ≺ (d, σ(φ, γ)). (6.38)
Finally, since the subpencil diag(B1 , C1 , C2 ) of (6.36) has zeros in correspond-
ing rows and columns, the pencil (6.36) is strictly equivalent to
diag(B1 , C1 , C2 , N (φ), R̄),
with
N (φ) = diag(C(φ1 ), . . . , C(φn+p1 )),
in a way that the block diag(B1 , C1 , C2 ) remains unchanged.
164 M. Dodig

Hence, by applying Theorem 4.6, we have that conditions (6.30), (6.37) and
(6.38) imply the existence of matrices Y1 ,. . . ,Y7 with m1 rows, such that the pencil
⎡ ⎤
N X1
⎢ B1 X2 ⎥
⎢ ⎥
⎢ C1 X3 ⎥
⎢ ⎥
⎢ C2 X4 ⎥
⎢ ⎥
⎢ R1 λE2 + X5 ⎥
⎢ ⎥
⎣ R2 X6 ⎦
Y1 Y2 λE3 + Y3 Y4 Y5 Y6 Y7
has d1 , . . . , dm2 , as its column minimal indices, r̄1 , . . . , r̄p2 as its row minimal in-
dices, and γ1 | · · · |γn+m1 +p1 as invariant factors, and b̄1 , . . . , b̄b1 as degrees of its
infinite elementary divisors. This finishes our proof.

7. Corollary
If E = I then we have that the pencil λI − (A + BF + KC) is regular, and its
(finite) invariant factors form a complete set of its Kronecker invariants. Moreover,
in this case we have m2 = p2 = 0. So, we directly obtain the following result.
Corollary 7.1. For a given system (I, A, B, C), there exist matrices F and K such
that λI − (A + BF + KC) is regular with γ1 | · · · |γn+m1 +p1 as invariant factors, if
and only if the following conditions are valid:
(i) γi | αi | γi+m1 +p1 , i = 1, . . . , n,
(ii)
 n 
m1 
p1 
n+m 1 +p1

d(αi ) + (c1i + 1) + (ri1 + 1) = d(γi ),

i=1 i=1 i=1 i=1

j 
k
(iii) (c1i + 1) + (ri1 + 1)
i=1 i=1

n+m 1 +p1 +p1 −k−j
n+m1
≤ d(γi ) − d(lcm(αi−m1 −p1 +k+j , γi )),
i=1 i=1
j = 0, . . . , m1 , k = 0, . . . , p1 .
Remark 7.2. By using the classical majorization [16], conditions (ii) and (iii) can
be unified as
(c1 + 1, . . . , cm1 + 1) ∪ (r1 + 1, . . . , rp1 + 1) ≺ (d(σm1 +p1 (α, γ)), . . . , d(σ1 (α, γ))).
We note that Corollary 7.1 can also be obtained as a corollary of [11, Theorem 1]
(see also [2, 26, 27]).
Acknowledgment
We would like to thank the referee for hers/his valuable comments and suggestions,
that have significantly improved the presentation of the paper.
 Descriptor Systems, Feedback and Output Injection 165

References
[1] G. Basile and G. Marro, Controlled and Conditional Invariants in Linear Systems
Theory. Prentice-Hall, Englewood Cliffs, NJ, 1992.
[2] I. Cabral and F.C. Silva, On feedback equivalence and completion problems. Linear
Algebra Appl. 265 (1997) 231–245.
[3] D. Carlson, Inequalities relating the degrees of elementary divisors within a matrix.
Simon Stevin 44 (1970), 3–10.
[4] B.M. Chen, Robust and H∞ Control. Springer, New York, 2000.
[5] B.M. Chen, Z. Lin, and Y. Shamash, Linear Systems Theory: A Structural Decom-
position Approach. Birkhäuser, Boston, 2004.
[6] B.W. Dickinson, On the fundamental theorem of linear state variable feedback. IEEE
Trans. Automat. Control AC-19 (1974) 577–579.
[7] M. Dodig, Feedback invariants of matrices with prescribed rows. Linear Algebra Appl.
405 (2005) 121–154.
[8] M. Dodig, Pole placement problem for singular systems. Numerical Linear Algebra
Appl. 18 (2011), 283–297.
[9] M. Dodig, Completion of quasi-regular matrix pencils. Linear Algebra Appl. 501
(2016), 198–241.
[10] M. Dodig and M. Stošić, Singular systems, state feedback problem. Linear Algebra
Appl. 431 (2009), 1267–1292.
[11] M. Dodig and M. Stošić, Similarity class of a matrix with prescribed submatrix.
Linear and Multilinear Algebra 57 (2009), 217–245.
[12] M. Dodig and M. Stošić, Combinatorics of column minimal indices and matrix pencil
completion problems. SIAM J. Matrix Anal. & Appl. 31 (2010), 2318–2346.
[13] M. Dodig and M. Stošić, The rank distance problem for pairs of matrices and a
completion of quasi-regular matrix pencils. Linear Algebra Appl. 457 (2014), 313–
347.
[14] J.C. Doyle, Guaranteed margins for LQG regulators. IEEE Trans. on AC 23 (1978),
756–757.
[15] W. Fulton, Eigenvalues, invariant factors, highest weights, and Schubert calculus.
Bull. Amer. Math. Soc. 37 (2000), 209–249.
[16] G. Hardy, J.E. Littlewood, and G. Pólya, Inequalities. Cambridge University Press,
Cambridge, 1991.
[17] V. Kučera, Analysis and Design of Discrete Linear Control Systems. Academia,
Prague, 1991.
[18] V. Kučera and P. Zagalak, Fundamental theorem of state feedback for singular sys-
tems, Automatica 24 (1988), 653–658.
[19] G. Lebrete and J.J. Loiseau, Proportional and proportional-derivative canonical
forms for descriptor systems with outputs. Automatica 30 (1994), 847–864.
[20] D.E. Littlewood and A.R. Richardson, Group characters and algebra. Phil. Trans.
Roy. Soc. London A 223 (1934), 99–141.
[21] J.J. Loiseau, Some geometric considerations about the Kronecker normal form. Int.
J. Control 42 (1985), 1411–1431.
166 M. Dodig

[22] J.J. Loiseau and G. Lebrete, A new canonical form for descriptor systems with out-
puts. Analysis and Optimization of Systems, vol. 144, Springer, New York, 1990.
[23] M. Malabre, A structural approach for linear singular systems. International Mini
Symposium on Singular Systems, Atlanta, 1987.
[24] S. Mondié, P. Zagalak, and V. Kučera, State feedback in linear control theory. Linear
Algebra Appl. 317 (2000), 177–192.
[25] A.S. Morse, Structural invariants of linear multivariable systems. SIAM J. Control
11 (1973), 446–465.
[26] A. Roca and I. Zaballa, Invariant factor assignment under state feedback and output
injection. Linear Algebra Appl. 332–334 (2001), 401–436.
[27] A. Roca and I. Zaballa, A polynomial approach to the invariant factors assignment
problem via state feedback and output injection. Linear Algebra Appl. 401 (2005),
135–158.
[28] H.H. Rosenbrock, State-Space and Multivariable Theory. Thomas Nelson and Sons,
London, 1970.
[29] A. Saberi, B.M. Chen, and P. Sannuti, Loop Transfer Recovery: Analysis and Design.
Springer, London, 1993.
[30] G. Stein and M. Athans, The LQG/LTR procedure for multivariable control design.
IEEE Trans. on AC 32 (1987), 105–114.
[31] I. Zaballa, Matrices with prescribed rows and invariant factors. Linear Algebra Appl.
87 (1987), 113–146.

Marija Dodig
CEAFEL, Departamento de Matématica
Universidade de Lisboa
Edificio C6, Campo Grande
1749-016 Lisbon, Portugal
and
Mathematical Institute SANU
Knez Mihajlova 36
11000 Beograd, Serbia
e-mail: [email protected]
Operator Theory:
Advances and Applications, Vol. 267, 167–183

c Springer International Publishing AG, part of Springer Nature 2018

Hermitian Geometry on Resolvent Set

Ronald G. Douglas and Rongwei Yang

Abstract. For a tuple A = (A1 , A2 , . . . , An ) of elements in a unital Banach

algebra B, its projective joint spectrum P (A) is the collection of z ∈ Cn such
that A(z) = z1 A1 + z2 A2 + · · · + zn An is not invertible. It is known that the
B-valued 1-form ωA (z) = A−1 (z)dA(z) contains much topological information
about the joint resolvent set P c (A). This paper studies geometric properties
of P c (A) with respect to Hermitian metrics defined through the B-valued
∗
fundamental form ΩA = −ωA ∧ ωA and its coupling with faithful states φ on
B, i.e., φ(ΩA ). The connection between the tuple A and the metric is the main
subject of this paper. In particular, it shows that the Kählerness of the metric
is tied with the commutativity of the tuple, and its completeness is related to
the Fuglede–Kadison determinant.

Mathematics Subject Classification (2010). Primary 47A13; Secondary 53A35.

Keywords. Maurer–Cartan form, projective joint spectrum, Hermitian metric,
Ricci tensor, Fuglede–Kadison determinant.

0. Introduction
In [6], Cowen and the first-named author introduced geometric concepts such as
holomorphic bundle and curvature into Operator Theory. This gave rise to com-
plete and computable invariants for certain operators with rich spectral structure,
i.e., the Cowen–Douglas operators. The idea was extended to multivariable cases
in the study of Hilbert modules in analytic function spaces, where curvature invari-
ant is defined for some canonical tuples of commuting operators. We refer readers
to [8–10, 24] and the references therein for more information. This paper aims
to study general (possibly non-commuting) tuples using geometric ideas. The ap-
proach is based on the newly emerged concept of projective joint spectrum P (A)
introduced by the second-named author (cf. [29]). Some preliminary findings are
reported here. They shall help to build a foundation for subsequent and more
in-depth studies.
168 R.G. Douglas and R. Yang

Let B be a complex Banach algebra with unit I and

A = (A1 , A2 , . . . , An )
be a tuple of linearly independent elements in B. The multiparameter pencil
A(z) := z1 A1 + z2 A2 + · · · + zn An
is an important subject of study in various fields, for example in algebraic geometry
[28], mathematical physics [1, 21], PDE [2], group theory [13], etc., and more
recently in the settlement of the Kadison–Singer Conjecture ([22]). In these studies,
the primary interest was in the case when A is a tuple of self-adjoint operators.
For general tuples, the notion of projective joint spectrum is defined in [29] and
its property is investigated in a series of papers (cf. [3–5, 25]).

Definition. For a tuple A = (A1 , A2 , . . . , An ) of elements in a unital Banach algebra

B, its projective joint spectrum is defined as
P (A) = {z ∈ Cn |A(z) is not invertible}.
The projective resolvent set refers to the complement P c (A) = Cn \ P (A).

Various notions of joint spectra for (commuting) tuples have been defined
in the past, for instance the Harte spectrum ([16, 17]) and the Taylor spectrum
([11, 26, 27]), and they are key ingredients in multivariable operator theory. The
projective joint spectrum, however, has one notable distinction: it is “base free” in
the sense that, instead of using I as a base point and looking at the invertibility
of (A1 − z1 I, A2 − z2 I, . . . , An − zn I) in various constructions, it uses the much
simpler pencil A(z). This feature makes the projective joint spectrum computable
in many interesting noncommuting examples, for example the tuple of general
compact operators ([25]), the generating tuple for the free group von Neumann
algebra ([3]), the Cuntz tuple of isometries (S1 , S2 , . . . , Sn ) ([7, 19]) satisfying

n
Si∗ Sj = δij I for 1 ≤ i, j ≤ n, Si Si∗ = I,
i=1

where I is the identity. For the generating tuple (1, a, t) of the infinite dihedral
group
D∞ =< a, t | a2 = t2 = 1 >,
the projective spectrum has found convincing applications to the study of group
of intermediate growth (finitely generated group whose growth is faster than poly-
nomial growth but slower than exponential growth (cf. [14])).
For a general tuple A, it can happen that P (A) = Cn and hence P c (A) is
empty. But in this paper we shall always assume that P c (A) is non-empty. It is
known that every path-connected component of P c (A) is a domain of holomorphy.
This fact is a consequence of a result in [30], and it is also proved independently
 Hermitian Geometry on Resolvent Set 169

in [20]. From this point of view, P c (A) has good analytic properties. On P c (A),
the holomorphic Maurer–Cartan type B-valued (1, 0)-form

n
ωA (z) = A−1 (z)dA(z) = A−1 (z)Aj dzj
j=1

played an important role in [29], where it is shown to contain much information

about the topology of P c (A). For example, the de Rham cohomology H ∗ (P c (A), C)
can be computed by coupling ωA with invariant multilinear functionals or cyclic
cocycles ([4, 29]). Two simple facts about ωA are useful here.
1) By differentiating the equation A(z)A(z)−1 = I, one easily verifies that
dA−1 (z) = −A−1 (z)(dA(z))A−1 (z),
and consequently,
dωA = dA−1 (z) ∧ dA(z) = −ωA ∧ ωA . (0.1)
Here d is complex differentiation (see Section 1) and ∧ is the wedge product.
2) ωA is invariant under left multiplication on the tuple A by invertible elements.
To be precise, we let GL(B) denote the set of invertible elements in B. Then
for any L ∈ GL(B) and B = LA = (LA1 , LA2 , . . . , LAn ), we have for the
multiparameter pencils B(z) = LA(z), and hence P (B) = P (A). Moreover,
ωB = B −1 (z)dB(z) = A−1 (z)L−1 LdA(z) = ωA . (0.2)
This shows that ωA is invariant under the left action of GL(B) on the tuple A.
For simplicity, we shall use Einstein convention for summation in many places
in this paper. For example, we shall write ωA (z) = A−1 (z)Aj dzj . When B is a C ∗ -
algebra, the adjoint of ωA (z) is the B-valued (0, 1)-form
∗
ωA (z) = (A−1 (z)Aj )∗ dz j .
Now we define the fundamental form for the tuple A as
∗
ΩA = −ωA ∧ ωA = (A−1 (z)Aj )∗ A−1 (z)Ak dzk ∧ dz j . (0.3)
Here the negative sign exists in the first equality because it is conventional to write
a (1, 1)-form as linear combinations of dzj ∧ dz̄k . Often, the factor 2i is used to
make 2i ΩA a self-adjoint form (cf. (1.2)), but it is not important in this paper.
For a suitable choice of linear functional φ, such as a faithful state on B, the
evaluation
φ(ΩA ) = φ (A−1 (z)Aj )∗ A−1 (z)Ak dzk ∧ dz j
induces a positive definite bilinear form on the holomorphic tangent bundle of
P c (A), thus giving a Hermitian metric on P c (A). The connection between the
metric and the tuple A shall be the primary concern of this paper. In particular,
it shows that the Kählerness of the metric is tied with the commutativity of the
tuple (cf. Theorem 2.4). A notable feature of this metric is that it has singularities
at the joint spectrum P (A). So completeness of the metric is an important issue.
170 R.G. Douglas and R. Yang

1. B-valued differential forms

Let M be a complex manifold of dimension n. If z = (z1 , z2 , . . . , zn ) is the coordi-
∂
nate in a local chart, then ∂i stands for ∂z , and ∂¯i stands for ∂∂z̄i . As a convention,
  i
we let ∂ = i ∂i , ∂¯ = i ∂¯i , and d = ∂ + ∂. ¯ Consider a globally defined smooth
(1, 1)-form Φ(z) = gjk (z)dzj ∧ dz¯k expressed in the local chart with the Einstein
summation convention. Φ induces a bilinear form Φ̂(z) = gjk (z)dzj ⊗ dz¯k on the
holomorphic tangent bundle T (M ) over the local chart such that
Φ̂(z)(∂j , ∂k ) = gjk (z).
We say that Φ defines a Hermitian metric on M if the n × n matrix function
g(z) = (gjk (z)) is positive definite for each z ∈ M . In this case g(z) is called the
associated metric matrix, and 2i Φ(z) is called the fundamental form of the metric.
Here, the constant 2i is to normalize the fundamental form so that
(a) 2i Φ(z) is real, i.e., 2i Φ(z) = 2i Φ(z);
(b) in the one variable case z = x + iy, we have 2i dz ∧ dz̄ = dx ∧ dy.
Some geometric concepts are relevant to the study here.
1) A Hermitian metric induced by Φ(z) is said to be Kähler if dΦ = 0, i.e., Φ(z)
is closed.
2) The Ricci curvature tensor is the n × n matrix R(z) = (Rjk (z)), where
Rjk = −∂j ∂¯k log det g(z), and the Ricci form is
RicΦ (z) = Rjk (z)dzj ∧ dz¯k .
3) The metric is said to be Ricci flat if R(z) = 0 for each z ∈ M .
4) The metric is said to be Einstein if RicΦ = λΦ for some constant λ, and it is
Calabi–Yau if it is Kähler and Ricci flat (i.e., λ = 0).
One checks that a Hermitian (1, 1)-form Φ(z) = gij (z)dzi ∧ dz¯j induces a
Kähler metric if and only if
∂k gij (z) = ∂i gkj (z), z ∈ M, (1.1)
for all 1 ≤ i, j, k ≤ n. Further, Φ is Ricci flat if and only if log det g(z) is pluri-
harmonic (cf. [23]). Hence in this case log det g(z) = f (z) + f (z) holds locally for
some holomorphic function f , i.e.,

det g(z) = ef (z)+f (z) = |ef (z) |2 .

Now we consider a smooth B-valued (p, q)-form ω p,q (z) defined on a com-
plex domain R ⊂ Cn , where 0 ≤ p, q ≤ n. The set of such forms is denoted by
Λp,q (R, B). Then ω p,q (z) is said to be ∂-closed or ∂-closed if ∂ω p,q (z) = 0, or re-
spectively ∂ω p,q (z) = 0 on R. If ω p,q is both ∂-closed and ∂ closed, then dω p,q = 0,
and in this case we simply say ω p,q is closed.
For the rest of the paper, unless stated otherwise, we shall assume B is a
C ∗ -algebra. Recall that for a tuple A, the fundamental form ΩA is given by (0.3).
 Hermitian Geometry on Resolvent Set 171

Now we check that 2i ΩA is self-adjoint in the sense that

 ∗
i i
ΩA = ΩA . (1.2)
2 2
Indeed, one checks by (0.3) that
 ∗
i −i ∗
ΩA = (−ωA ∧ ωA )∗
2 2
−i ∗
= (A−1 (z)Aj )∗ A−1 (z)Ak dzk ∧ dz j
2
−i −1
= A (z)Ak )∗ A−1 (z)Aj dz k ∧ dzj
2
−i ∗ i
= (ωA ∧ ωA ) = ΩA .
2 2
Further, in the case n = 1 and A1 is invertible, one has A(z) = zA1 , P c (A) =
C \ {0}, and ωA (z) = dz
z . Writing z = x + yi we have

i i(dz ∧ dz) dx ∧ dy
ΩA = = 2 ,
2 2|z|2 x + y2
which is real. For a linear functional φ ∈ B ∗ ,
 
i i
φ ΩA = φ A−1 (z)Aj )∗ A−1 (z)Ak dzk ∧ dz j ,
2 2
which is a regular (1, 1)-form. Moreover, if φ is positive then by (1.2)
     
i −i ∗ i
φ ΩA = φ ΩA = φ ΩA ,
2 2 2
and hence φ 2i ΩA is a regular real (1, 1)-form. The following definitions are B-
valued versions of Hermitian metric and Kähler metric.

Definition. A smooth B-valued (1, 1)-form Φ(z) = Φjk (z)dzj ∧ dz k on a complex

domain R is said to be Hermitian if for every z ∈ R and any set of n complex
numbers v1 , v2 , . . . , vn the double sum vi Φjk (z)vk is a positive element in B.

Definition. A smooth B-valued Hermitian (1, 1)-form Φ(z) is said to be Kähler if

it is closed, i.e., dΦ(z) = 0.

Two nuances are worth mentioning. First, there is a difference between a self-
adjoint B-valued (1, 1)-form as defined in (1.2) and a Hermitian B-valued (1, 1)-
form defined above – the latter requires positivity of B-valued matrix (Φij (z)).
Second, in the scalar-valued case, for a Hermitian metric defined by Ω(z) =
gjk (z)dzj ∧ dz̄k , the complex metric matrix (gjk (z)) needs to be positive defi-
nite, but in the B-valued case, it is only required that vj Φjk (z)vk ≥ 0 in B. For
172 R.G. Douglas and R. Yang

the fundamental form ΩA and any set of n complex numbers v1 , v2 , . . . , vn , one

checks by (0.3) that the double sum
vj A∗k (A−1 (z))∗ A−1 (z)Aj v¯k = v¯k A∗k (A−1 (z))∗ A−1 (z)vj Aj )
= (A−1 (z)A(v))∗ (A−1 (z)A(v) ≥ 0. (1.3)
So ΩA is Hermitian.
In this paper, the extended tuple  = (I, A1 , A2 , . . . , An ) is used in some
places for normalization purposes. In this case we can regard I as A0 and P (Â) is
then clearly a subset in Cn+1 . This extension guarantees two convenient facts:
1) obviously, (1, 0, 0, . . . , 0) ∈ P c (Â);
2) the identity I is in the range of Â(z) as z varies in Cn+1 .
Now we are in position to state the first result.
Proposition 1.1. For the tuple A = (A1 , A2 , . . . , An ), the following statements are
equivalent.
(i) A is commuting.
(ii) ωA (z) is closed.
(iii) ΩÂ (z) is Kähler.
Proof. Using (0.1) and the fact that dzj ∧ dzk = −dzk ∧ dzj , one checks that
dωA = −ωA ∧ ωA
= A−1 (z)Aj A−1 (z)Ak dzj ∧ dzk

= [A−1 (z)Aj , A−1 (z)Ak ]dzj ∧ dzk , (1.4)
j<k

where [a, b] = ab − ba. So when A is a commuting tuple we have

[A−1 (z)Aj , A−1 (z)Ak ] = 0, z ∈ P c (A),
and consequently dωA = ∂ωA = 0. This shows (i) implies (ii).
∗
Recall that ΩA (z) = −ωA ∧ ωA . Since ωA (z) is holomorphic
dΩA = −(∂ + ∂)(ω ∗ ∧ ωA )
¯
A
∗ ¯ ∗ ) ∧ ωA + ω ∗ ∧ (∂ωA + ∂ω
¯ A)
= −(∂ωA + ∂ω A A

= − (∂ωA )∗ ∧ ωA + (ωA
∗
∧ ∂ωA ). (1.5)
∗
So if ωA (z) is closed then dΩA = −(0 ∧ ωA ) + (ωA ∧ 0)
= 0, i.e., ΩA is Kähler. Since
 is commuting, the forgoing arguments in particular imply that Ω is Kähler as
well. This shows that (ii) implies (iii). Now if ΩÂ is Kähler then by (1.5)
0 = dΩ = −(∂ω ¯ ∗ ) ∧ ω + ω ∗ ∧ ∂ω . (1.6)
    Â
Since the first term in (1.6) is a (1, 2)-form and the second is a (2, 1)-form, the
sum is 0 only if both terms are 0. Furthermore, by (1.4) and (1.6) we have

∗
0 = ω ∧ dω = (Â−1 (z)Ai )∗ [Â−1 (z)Aj , Â−1 (z)Ak ]dz̄i ∧ dzj ∧ dzk ,
0≤j<k
 Hermitian Geometry on Resolvent Set 173

which implies that on P c (Â),

(Â−1 (z)Ai )∗ [Â−1 (z)Aj , Â−1 (z)Ak ] = 0, ∀ 0 ≤ i, j, k ≤ n.
Setting i = 0 and using the fact Â(z) = A0 = I at z = (1, 0, 0, . . . , 0) ∈ Cn+1 , we
have
Ak Aj = Aj Ak , ∀1 ≤ j, k ≤ n,
i.e., the tuple A is commuting. So (iii) implies (i). 
To prove that (iii) implies (i), we used the fact that the tuple  includes I.
In fact, as the proof indicates that if A(z) = I for some z ∈ Cn then there is no
need to use Â. But this kind of requirement is indispensable.
Example 1.2. Consider a pair A = (A1 , A2 ), where A1 is invertible. Then writing
A(z) = A1 (z1 + z2 A−1
1 A2 ) and using (0.2), we have

ωA = (z1 I + z2 A−1
1 A2 )
−1
d(z1 I + z2 A−1
1 A2 ).

Since I commutes with A−1 1 A2 , the fundamental form ΩA is Kähler by Proposi-

tion 1.1. But clearly A1 may not commute with A2 .
For every φ ∈ B ∗ , we have dφ(ωA ) = φ(dωA ). We let H q (P c (A), C) denote
the qth de Rham cohomology of P c (A). The following is then an easy consequence
of Proposition 1.1.
Corollary 1.3. Let A be a commuting tuple. Then φ(ωA ) ∈ H 1 (P c (A), C) for every
φ ∈ B∗.
Proposition 1.1 raises a natural question: is ωA exact, i.e., is there a B-valued
smooth function f (z) on P c (A) such that ωA = df ? We take time to address
this issue now. For every φ ∈ B ∗ and a B-valued smooth (p, q)-form ω p,q (z) on a
complex domain R, the coupling φ(ω p,q (z)) is a scalar-valued smooth (p, q)-form
on R. For example by (0.3),
φ(ΩA ) = φ A−1 (z)Aj )∗ A−1 (z)Ak dzk ∧ dz j .
Moreover, ω p,q (z) = 0 if and only if φ(ω p,q (z)) = 0 for every φ ∈ B ∗ . We let the
set of B-valued (p, q)-forms on R be denoted by ω p,q (R, B). The fact
∂φ(ω p,q (z)) = φ(∂ω p,q (z))
is nicely expressed in the following commuting diagram
∂p
Λp,q (R, B) Λp+1,q (R, B)
φ φ (1.7)
∂p
Λp,q (R, C) Λp+1,q (R, C),

and the parallel diagram regarding ∂. Here the subscript p in ∂p is to indicate

the space on which ∂ acts. Since d = ∂ + ∂, we see that ω p,q (z) is closed, i.e.,
dω p,q = 0, if and only if φ(ω p,q (z)) is closed for every φ ∈ B ∗ . Since by [20, 30]
174 R.G. Douglas and R. Yang

the joint resolvent set P c (A) is a union of domains of holomorphy, we shall only
consider the case when R is a domain of holomorphy, in which case its de Rham
cohomology H p (R, C) (or H p (R, B)) can be computed through holomorphic forms.
We denote the set of B-valued holomorphic p-forms by Λp (R, B). Note again that
in this case d = ∂. We refer the readers to [23] for details on domains of holomorphy
and holomorphic forms.
Since ∂p ∂p−1 = 0, one naturally has B-valued cohomology groups
H p (R, B) = ker ∂p /im∂p−1 , p ≥ 0,
where ∂−1 is the trivial inclusion map {0} → Λ0 (R, B). It then follows from the
following commuting diagram for holomorphic forms
∂p
Λp (R, B) Λp+1 (R, B)
φ φ
∂p
Λ (R, C)
p
Λ p+1
(R, C),
that a form ω is in H p (R, B) if and only if φ(ω) is in Λp (R, C) for every φ ∈ B ∗ .
It was shown in [29] that if B is a Banach algebra with a trace φ, then for every
tuple A such that P c (A) is nonempty, the 1-form φ(ωA ) is a non-trivial element
in H 1 (P c (A), C). The above observations thus lead to the following.
Corollary 1.4. Let B be a unital Banach algebra with a trace and A be a commuting
tuple of elements in B. Then ωA is a non-trivial element in H 1 (P c (A), B).
Proof. By Proposition 1.1, we see that ωA is closed when A is a commuting tuple.
If it were exact, then there exists B-valued smooth function f (z) on P c (A) such
that ωA = df . If φ is the trace, then it follows that
φ(ωA ) = φ(df ) = dφ(f ).
Since φ(f ) is globally defined on P c (A), the 1-form φ(ωA ) is a trivial element in
H 1 (P c (A), C) contradicting with the above mentioned fact in [29] that φ(ωA ) is a
non-trivial element in H 1 (P c (A), C). 
Proposition 1.1 indicates that the commutativity of tuples A has a natural
homological interpretation.

2. Hermitian metric on P c (A)

Recall that for a C ∗ -algebra B, a smooth B-valued (1, 1)-form
Φ(z) = Φjk (z)dzj ∧ dz k
on a domain R is said to be Hermitian if for any set of n complex numbers
v1 , v2 , . . . , vn the double sum vj Φjk (z)vk is a positive element in B for every z ∈ R.
It follows that if φ ∈ B ∗ is a positive linear functional, then the matrix φ(Φjk (z))
is positive semi-definite for every z ∈ R. By (1.3), we see that the fundamental
 Hermitian Geometry on Resolvent Set 175

∗
(1, 1)-form ΩA = −ωA ∧ ωA is Hermitian. This section defines some natural Her-
mitian metrics on P (A) through ΩA and its pairing with certain states on B.
c

A bounded linear functional φ on a C ∗ -algebra B is called a state if it is

positive and φ(I) = 1. Consider a tuple A = (A1 , A2 , . . . , An ) of elements in B.
Given a state φ, then on P c (A) we compute that
∗
φ(ΩA (z)) = − φ(ωA ∧ ωA )
= φ(A∗k (A−1 (z))∗ A−1 (z)Aj )dzj ∧ dz̄k
:= gjk (z)dzj ∧ dz̄k . (2.1)
As observed in the previous paragraph, the n× n matrix g(z) = (gjk (z)) is positive
semi-definite for every z ∈ P c (A). So it becomes a natural question for which φ the
matrix g(z) is positive definite on P c (A). In other words, when does φ(ΩA ) define
a Hermitian metric on P c (A)? To answer this question, we consider the operator
space HA = span{A1 , A2 , . . . , An }. A state φ on B is said to be faithful on HA if
for every nonzero element h ∈ HA we have φ(h∗ h) > 0.
Proposition 2.1. Let φ be a state on a C ∗ -algebra B. Then for any tuple A the
(1, 1)-form φ(ΩA ) induces a Hermitian metric on P c (A) if and only if φ is faithful
on HA .
Proof. For every fixed z ∈ P c (A), since A(z) = z1 A1 + z2 A2 + · · · + zn An is
invertible, there are numbers α > β > 0 such that
βI ≤ (A−1 (z))∗ A−1 (z) ≤ αI.
Note that α and β depend on z, but it is not important here. For every nonzero
vector v = (v1 , v2 , . . . , vn ) ∈ Cn , one checks that
vj gjk (z)v¯k = vj φ(A∗k (A−1 (z))∗ A−1 (z)Aj )v¯k
= φ(v¯k A∗k (A−1 (z))∗ A−1 (z)vj Aj )
= φ (A−1 (z)A(v))∗ (A−1 (z)A(v)
= φ A∗ (v)(A−1 (z))∗ A−1 (z)A(v) .
Since
βA∗ (v)A(v) ≤ A∗ (v)(A−1 (z))∗ A−1 (z)A(v) ≤ αA∗ (v)A(v)
and φ is positive, it follows that
βφ(A∗ (v)A(v)) ≤ vj gjk (z)v¯k ≤ αφ(A∗ (v)A(v)).
Hence (gjk (z)) is positive definite for every z ∈ P c (A) if and only if φ is faithful
on HA . 
Example 2.2. A state φ on a C ∗ -algebra B is called a faithful tracial state if
φ(ab) = φ(ba) for any a, b ∈ B and φ(a∗ a) > 0 when a = 0. If φ is a faithful
tracial state, then it is clearly a faithful state on HA . In this case, we shall write
the metric matrix (gij (z)) as gA (z).
176 R.G. Douglas and R. Yang

Example 2.3. Assume B is a C ∗ -algebra of operators acting on a Hilbert space H.

Let φx be a vector state defined on B by φx (a) = ax, x, where a ∈ B and x is
fixed with x = 1. If {A1 x, A2 x, . . . , An x} is linearly independent, then for any
nonzero vector (c1 , c2 , . . . , cn ) ∈ Cn , we have
φx ((ci Ai )∗ (ci Ai )) = ci Ai x2 > 0,
i.e., φ is faithful on HA .
It follows from Propositions 1.1, 2.1 and the commuting diagram (1.7) that
for a commuting tuple A = (A1 , A2 , . . . , An ), and every state φ faithful on HÂ ,
dφ(ΩÂ ) = φ(dΩÂ ) = 0.
Hence φ(ΩÂ ) defines a Kähler metric on P c (Â). Interestingly, the converse is also
true if φ is faithful on the entire B.
Theorem 2.4. Let φ be a faithful state on B. Then  = (I, A1 , A2 , . . . , An ) is
commuting if and only if φ(ΩÂ ) is Kähler.
Proof. It only remains to check sufficiency. If φ(ΩÂ ) is Kähler, then we have 0 =
dφ(ΩÂ ) = φ(dΩÂ ). Writting I as A0 , then by (1.6), we have
∗
0 = φ(ω ∧ ∂ω )
  ' (
= φ (Â−1 (z)Ai )∗ [Â−1 (z)Aj , Â−1 (z)Ak ] dz̄i ∧ dzj ∧ dzk ,
i 0≤j<k

which implies
' (
φ (Â−1 (z)Ai )∗ [Â−1 (z)Aj , Â−1 (z)Ak ] = 0, ∀0 ≤ i, j, k ≤ n. (2.2)

Applying ∂¯m to both sides of (2.2) and using the fact ∂¯m Â−1 (z) = 0 and
∂¯m (Â−1 (z))∗ = (∂m Â−1 (z))∗ = −(Â−1 (z)Am Â−1 (z))∗
we have
' (
φ (Â−1 (z)Am Â−1 (z)Ai )∗ [Â−1 (z)Aj , Â−1 (z)Ak ] = 0, ∀0 ≤ i, j, k, m ≤ n.
(2.3)
In (2.3), setting (m, i) to (j, k), then to (k, j) and taking the difference, we have
' (
φ [Â−1 (z)Aj , Â−1 (z)Ak ]∗ [Â−1 (z)Aj , Â−1 (z)Ak ] = 0, ∀j, k.

Since φ is faithful, [Â−1 (z)Aj , Â−1 (z)Ak ] = 0 for all j, k. Setting z to (1, 0, 0, . . . , 0),
we have [Aj , Ak ] = 0 for all j, k. 
Again, as indicated by Example 1.2 and the remarks before it, the requirement
that I be in the tuple A (or in HA ) is indispensable in Theorem 2.4. We end this
section by a question motivated by Theorem 2.4.
Problem 1. Find conditions on commuting tuple A and state φ such that φ(ΩA )
defines a Ricci-flat metric (i.e., Calabi–Yau metric) on P c (A).
 Hermitian Geometry on Resolvent Set 177

3. Completeness
Once a metric is defined on a set, an immediate question is whether the set is com-
plete under the metric. In this section we will study this problem for P c (A) with
respect to the metric defined by gA as in Example 2.2. Since ωA (z) = A−1 (z)dA(z)
resembles the derivative of logarithmic function, it is natural to expect that log
shall play an important role here.
In [12], Fuglede and Kadison defined the following notion of determinant
det x = exp(φ(log |x|))

√ x in a finite von Neumann algebra B with a normalized

for invertible elements
trace φ. Here |x| = x∗ x and log |x| is defined by the functional calculus, i.e.,
!
log |x| = log λdE(λ),
σ(|x|)

where E(λ) is the associated projection-valued spectral measure. Then

!
φ(log |x|) = log λdφ(E(λ)). (3.1)
σ(|x|)

Since x is invertible, the spectrum σ(|x|) is bounded away from 0. Hence the
integral in (3.1) is greater than −∞, which means det x = 0. However, there are
non-invertible (singular) elements x for which the improper integral in (3.1) is
convergent (hence det x = 0). This is essentially due to the absolute convergence
of ! 1
log t dt.
0
An example of such element is given in [12]. For elements x such that the integral
in (3.1) is equal to −∞, det x is naturally defined to be 0. This extends det to
all elements in B, and by [12] it is continuous at non-singular elements and upper
semi-continuous at singular elements with respect to the norm topology of B.
FK-determinant has been well studied in many papers, and we refer readers
to [15] for a survey. In particular, it was generalized to C ∗ -algebras in [18] as
follows. Assume GL(B) is path-connected, x ∈ GL(B) and x(t) is a piece-wise
smooth path in GL(B) such that x(0) = I and x(1) = x, then when the quantity
! 1 
−1
det x = exp φ(x (t)dx(t))
∗ 0
is independent of path x(t), it defines a notion of determinant for invertible ele-
ments x. One sees that det∗ may take on complex values, and it is shown in [15]
that  ! 1 
| det x| = exp Re φ x−1 (t)dx(t) = det x.
∗ 0
Clearly, det I = 1, and for a fixed 0 ≤ s ≤ 1 we have
! s 
−1
det x(s) = exp φ x (t)dx(t) ,
∗ 0
178 R.G. Douglas and R. Yang

so it follows that
φ(x−1 (t)dx(t)) = d log det x(t). (3.2)
∗

Here d is the differential with respect to t. The following definition is needed for
our study.

Definition. Consider a C ∗ -algebra B with a faithful tracial state φ. Then an element

x will be called φ-singular if its FK-determinant det x = 0. Furthermore, for a tuple
A of elements in B, a point p ∈ P (A) is said to be φ-singular if A(p) is φ-singular.

Now we begin to study the completeness of the metric defined by gA on

P c (A) in Example 2.2. Let [P c (A)] denote the completion of P c (A) with respect
to the metric gA . First, consider two points p and q that lie in the same connected
component of P c (A), and let γ = {z(t) : 0 ≤ t ≤ 1} be a piecewise smooth
path such that z(0) = p and z(1) = q. The length of γ with respect to the metric
gA = (gij ) is
! 1,
L(γ) = zi (t)gij (z(t))zj (t)dt.
0

The distance from p to q with respect to the metric gA is the infimum of L(γ) over
all such paths, i.e.,
dist(p, q) = inf L(γ).
γ

Lemma 3.1. Let B be a C ∗ -algebra with a faithful tracial state φ. If p and q are in
the same connected component of P c (A) then

dist(p, q) ≥ |φ(log |A(p)|) − φ(log |A(q)|)|.

Proof. First of all, by (2.1) we have gij (z) = φ (A−1 (z)Aj )∗ A−1 (z)Ai . Let p and
q be two points in the connected component U of P c (A), and let γ = {z(t)| 0 ≤
t ≤ 1} be a piecewise smooth path such that z(0) = p and z(1) = q. Then on γ we
have A (z(t)) = zj (t)Aj , and the length of γ can be computed as
! 1 ,
L(γ) = zi (t)gij (z(t))zj (t)dt
0
! 1 ,
= zi (t)φ (A−1 (z)Aj )∗ A−1 (z)Ai zj (t)dt
0
! 1 ,
= φ (A−1 (z(t))zj (t)Aj )∗ A−1 (z(t))zi (t)Ai dt
0
! 1 ,
∗
= φ (A−1 (z(t))A (z(t)) A−1 (z(t))A (z(t)) dt. (3.3)
0
 Hermitian Geometry on Resolvent Set 179

Since φ is a state, for every a, b ∈ B one has |φ(ab)|2 ≤ φ(a∗ a)φ(b∗ b). In particular,
|φ(a)|2 ≤ φ(a∗ a). Hence
! 1 ! 1
−1 
L(γ) ≥ |φ(A (z(t))A (z(t))|dt = |φ[A−1 (z(t))dA(z(t))]|
0 0
! 1 ! 1
= |φ(ωA (z(t))| ≥ φ(ωA (z(t)) .
0 0
By (3.2) we have
! 1
L(γ) ≥ d log det A(z(t))
0 ∗

= log det A(p) − log det A(q)

∗ ∗

= log | det A(p)| − log | det A(q)|

∗ ∗

+ i Arg det A(p) − Arg det A(q)

∗ ∗

≥ log | det A(p)| − log | det A(q)|

∗ ∗
= |log det A(p) − log detA(q)| .
Since γ is arbitrary, we have
dist(p, q) = inf L(γ) ≥ |log detA(p) − log detA(q)|
γ
= |φ(log |A(p)|) − φ(log |A(q)|)| . 
The proof of Lemma 3.1 can be modified a little to accommodate the case in
which q is a boundary point of a path-connected component of P c (A). First, we
observe that in this case q ∈ P (A). Suppose there exists a piece-wise smooth path
z(t) such that z(t) ∈ P c (A), 0 ≤ t < 1 but q = z(1) ∈ P (A). Then by the proof
above we have
! s
L(γ) ≥ lim− |φ(ωA (z(t))|
s→1 0
≥ lim sup |log detA(p) − log detA(z(s))| . (3.4)
s→1−

If q is a φ-singular point, we have det A(z(1)) = det A(q) = 0. Note that 0 is the
minimum possible value for det. Hence the upper semi-continuity of det A(z(s))
implies
lim det A(z(s)) = 0.
s→1
Therefore by (3.4) we have L(γ) = ∞ for every such path γ in P c (A) connecting
p ∈ P c (A) to q ∈ P (A), and consequently
dist(p, q) = inf L(γ) = ∞.
γ

Since q has infinite distance to every p ∈ P c (A), we have the following.

180 R.G. Douglas and R. Yang

Theorem 3.2. For a tuple A in a C ∗ -algebra B with a faithful tracial state φ, if

q ∈ ∂P c (A) is φ-singular, then q ∈
/ [P c (A)].

If B is a matrix subalgebra in Mk (C), we let Tr and det stand for the ordinary
trace and respectively determinant of k × k matrices. Then for any n-tuple A, the
joint spectrum P (A) is the hypersurface {det A(z) = 0} in Cn . Clearly, ∂P c (A) =
P (A). Let φ = k1 Tr. Then φ is a tracial state on Mk (C), and in this case the FK-
determinant and the usual determinant satisfies the relation det x = | det x|1/k
(cf. [12]). So in this case every point in P (A) is φ-singular. Since the factor 1/k is
not important, we have the following.

Corollary 3.3. For every tuple A of k × k matrices, the metric on P c (A) defined
by Tr(ΩA ) is complete.

The general linear group GL(k, C) is an open subset in Mk (C), hence it is

clearly not complete with respect to the usual Euclidean metric. Corollary 3.3
implies that we can endow GL(k) with a complete metric.

Corollary 3.4. Let A1 , A2 , . . . , An be a basis for the vector space Mk (C), where
n = k 2 . Then Tr(ΩA ) defines a complete, left invariant, Ricci flat, but non-Kähler
metric on the complex general linear group GL(k).

Proof. First, since A1 , A2 , . . . , An is a basis for the matrix algebra Mk (C), every
matrix is of the form A(z) for some z. Hence z ∈ P c (A) if and only if A(z) ∈ GL(k).
And this provides an identification of P c (A) with GL(k). So in particular Tr(ΩA )
defines a metric on GL(k). Recall that
∗
Tr(ΩA ) = − Tr(ωA ∧ ωA ) = Tr (A−1 (z)Aj )∗ A−1 (z)Ai dzi ∧ dz̄j .

For 1 ≤ i ≤ n, we write Ai = (ai1 , ai2 , . . . , aik ), where aim , 1 ≤ m ≤ k, is the mth

column of Ai . Then
∗
(A−1 (z)Aj )∗ A−1 (z)Ai = A−1 (z)aj1 , A−1 (z)aj2 , . . . , A−1 (z)ajk
× A−1 (z)ai1 , A−1 (z)ai2 , . . . , A−1 (z)aik ,

and hence

Tr (A−1 (z)Aj )∗ A−1 (z)Ai = A−1 (z)aim , A−1 (z)ajm Ck .
m

Let α be the n × n-matrix aim 1≤m≤k,1≤i≤n , whose ith column (in Cn ) is

denoted by αi , and let Ik be the k×k identity matrix. Then A−1 (z)⊗Ik is the n×n
block matrix with A−1 (z) on the diagonal. So we can write the sum in (3.3) as

(A−1 (z) ⊗ Ik )αi , (A−1 (z) ⊗ Ik )αj Cn .

 Hermitian Geometry on Resolvent Set 181

Since A1 , A2 , . . . , An are linearly independent, α is invertible, and one checks

by direct computation that the metric matrix
g(z) = Tr((A−1 (z)Aj )∗ A−1 (z)Ai n×n
= (A−1 (z) ⊗ Ik )αi , (A−1 (z) ⊗ Ik )αj  n×n
∗ −1 ∗ −1
= α (A (z) ⊗ Ik ) (A (z) ⊗ Ik )α.
It follows that det g(z) = | det α|2 | det A−k (z)|2 . Since det A−k (z) is holomorphic
on P c (A), the metric defined by Tr ΩA is Ricci flat by the remark after (1.1).
Moreover, because A1 , A2 , . . . , An is a basis for Mk (C), for every fixed x ∈
GL(k), xA(z) = A(w), where w ∈ P c (A) and is uniquely determined by z. This
gives rise to a natural action L of GL(k) on P c (A) given by Lx (z) = w. Since
by (0.2), ωA is invariant under left action of GL(k), the metric defined by Tr ΩA
is invariant under the action of L. The completeness of the metric follows from
Corollary 3.3. Since I is of the form A(z) for some z ∈ Cn in this case, the non-
Kählerness follows from Theorem 2.4. 

We end the paper by posing the following.

Problem 2. Is the converse of Theorem 3.2 true?

Acknowledgment
The second-named author would like to thank Guoliang Yu and the Department
of Mathematics at Texas A&M University for their support and hospitality during
his visit.

References
[1] M. Andersson and J. Söstrand, Functional calculus for non-commuting operators
with real spectra via an iterated Cauchy formula. J. Funct. Anal. 210 (2004), 341–
375.
[2] F.V. Atkinson, Multiparameter Eigenvalue Problems. Academic Press, New York
and London, 1972.
[3] P. Cade, J. Bannon, and R. Yang, On the spectrum of operator-valued entire func-
tions. Illinois J. Math. 55 (2011), 1455–1465.
[4] P. Cade and R. Yang, Projective spectrum and cyclic cohomology. J. Funct. Anal.
265 (2013), 1916–1933.
[5] I. Chagouel, M. Stessin, and K. Zhu, Geometric spectral theory for compact opera-
tors. Trans. Amer. Soc. 368 (2016), 1559–1582.
[6] M. Cowen and R.G. Douglas, Complex geometry and operator theory. Acta Math.
141 (1978), 187–261.
[7] J. Cuntz, Simple C ∗ -algebras generated by isometries. Comm. Math. Phys. 57
(1977), 173–185.
182 R.G. Douglas and R. Yang

[8] R.G. Douglas, Invariants for Hilbert modules. In: “Operator Theory: Operator Al-
gebras and Applications”, Part 1 (Durham, NH, 1988), Proc. Sympos. Pure Math.,
51, Part 1, Amer. Math. Soc., Providence, RI, 1990, pp. 179–196.
[9] R.G. Douglas and G. Misra, Geometric invariants for resolutions of Hilbert mod-
ules. In: “Nonselfadjoint Operator Algebras, Operator Theory, and Related Topics”.
Oper. Theory Adv. Appl. 104, 1998, pp. 83–112.
[10] R.G. Douglas and V. Paulsen, Hilbert Modules Over Function Algebras. Pitman Re-
search Notes in Mathematics Series, 217. Longman Scientific & Technical, Harlow;
copublished in the United States with John Wiley & Sons, Inc., New York, 1989.
[11] A. Fainshtein, Taylor joint spectrum for family of operators generating nil-potent
Lie algebras. J. Oper. Theory 29 (1993), 3–27.
[12] B. Fuglede and R. Kadison, Determinant theory in finite factors, Annals Math. 55
(1952), 520–530.
[13] R. Grigorchuk and Z. Šunić, Schreier spectrum of the Hanoi Towers group on three
pegs. Proc. of Symposia in Pure Math. Vol. 77, 2008, pp. 183–198.
[14] R. Grigorchuk and R. Yang, Joint spectrum and infinite dihedral group. Proc.
Steklov Institute Math. 297 (2017), 145–178.
[15] P. de la Harpe, The Fuglede–Kadison determinant, theme and variations. Proc. Nat.
Academy Sci. USA 110 (2013), 15864–15877.
[16] R. Harte, Spectral mapping theorems. Proc. Royal Irish Acad. 72 (1972), 89–107.
[17] R. Harte, Spectral mapping theorems for quasi-commuting systems. Proc. Royal
Irish Acad. 73 (1973), 7–18.
[18] P. de la Harpe and G. Skandalis, Déterminant associé à une trace sur une algèbre
de Banach. Ann. Inst. Fourier (Grenoble) 34 (1984), 241–260 (in French).
[19] W. He, X. Wang, and R. Yang, Projective spectrum and kernel bundle (II). J. Oper.
Theory, to appear.
[20] W. He and R. Yang, Projective spectrum and kernel bundle. Sci. China. Math. 57
(2014), 1–10.
[21] B. Jefferies, Spectral Properties of Noncommuting Operators. Lecture Notes in
Math, 1843, Springer-Verlag, Berlin, 2004.
[22] A. Marcus, D. Spielman, and N. Srivastava, Interlacing families II: Mixed char-
acteristic polynomials and the Kadison–Singer problem. Annals Math. 182 (2015),
327–350.
[23] R.M. Range, Holomorphic Functions and Integral Representations in Several Com-
plex Variables. Graduate Texts in Mathematics 108, Springer-Verlag, New York,
1986.
[24] J. Sarkar, An introduction to Hilbert module approach to multivariable operator
theory. In: “Operator Theory”. Spinger, Basel, 2015, pp. 969–1033.
[25] M. Stessin, K. Zhu, and R. Yang, Analyticity of a joint spectrum and a multivariable
analytic Fredholm theorem. New York J. Math. 17A (2011), 39–44.
[26] J.L. Taylor, A joint spectrum for several commutative operators. J. Funct. Anal. 6
(1970), 172–191.
[27] J.L. Taylor, A general framework for multi-operator functional calculus. Adv. Math.
9 (1972), 137–182.
 Hermitian Geometry on Resolvent Set 183

[28] V. Vinnikov, Determinantal representation of algebraic curves. In: “Linear Algebra

in Signals, Systems and Control (Boston, MA, 1986)”, SIAM, Philadelphia, PA,
1988, pp. 73–99.
[29] R. Yang, Projective spectrum in Banach algebras. J. Topol. Anal. 1 (2009), 289–306.
[30] M. Zaidenberg, S. Krein, P. Kuchment, and A. Pankov, Banach bundles and linear
operators. Russian Math. Surveys 30 (1975), 115–175.

Ronald G. Douglas
Department of Mathematics
Texas A&M University
College Station, TX 77843, USA
e-mail: [email protected]
Rongwei Yang
School of Mathematical Sciences
Tianjin Normal University
Tianjin 300387, P.R. China
and
Department of Mathematics and Statistics
SUNY at Albany
Albany, NY 12222, USA
e-mail: [email protected]
Operator Theory:
Advances and Applications, Vol. 267, 185–198

c Springer International Publishing AG, part of Springer Nature 2018

Spectral Algorithms for

MRA Orthonormal Wavelets
F. Gómez-Cubillo and S. Villullas

Abstract. Operator techniques lead to spectral algorithms to compute scaling

functions and wavelets associated with multiresolution analyses (MRAs). The
spectral algorithms depend on the choice of pairs of suitable orthonormal
bases (ONBs). This work presents the spectral algorithms for three different
pairs of ONBs: Haar bases, Walsh–Paley bases and trigonometric bases. The
Walsh–Paley bases connect wavelet theory and dyadic harmonic analysis. The
results for trigonometric bases are the first viable attempt to do a discrete
Fourier analysis of the problem.
Mathematics Subject Classification (2010). Primary 42C40; Secondary 47N40.
Keywords. Orthonormal wavelets, multiresolution analysis, spectral algo-
rithms.

1. Introduction
On the Hilbert space L2 (R) of square integrable complex functions consider the
translation and dilation operators, T and D, given by
[T f ](x) := f (x − 1) , [Df ](x) := 21/2 f (2x) , (f ∈ L2 (R)) .
By an orthonormal wavelet on R we mean here a function ψ ∈ L2 (R) such that
the family of dilated-translated versions of ψ
{ψj,k (x) := [Dj T k ψ](x) = 2j/2 ψ(2j x − k)}j,k∈Z
is an orthonormal basis (shortly ONB) of L2 (R).
Certain orthonormal wavelets ψ, in particular, the compactly supported or-
thonormal wavelets, are related to multiresolution analyses (MRAs), i.e., increas-
ing sequences {Vj }j∈Z of closed subspaces of L2 (R) such that ∩j∈Z Vj = {0},
∪j∈Z Vj = L2 (R), f ∈ V0 ⇔ fj,0 = Dj f ∈ Vj , and there exists a function

This work was partially supported by research projects MTM2012-31439 and MTM2014-57129-
C2-1-P (Secretarı́a General de Ciencia, Tecnologı́a e Innovación, Ministerio de Economı́a y Com-
petitividad, Spain).
186 F. Gómez-Cubillo and S. Villullas

ϕ ∈ L2 (R), the scaling function, whose integer translates {ϕ0,k = T k ϕ}k∈Z form
an ONB of V0 . In such case, ψj,k ∈ Vj for every j, k ∈ Z, and a pair of discrete
quadrature mirror filters hk k∈Z and (−1)1−k h1−k k∈Z of l2 (Z) are involved in
the two-scale relations

ϕj,0 = hk ϕj+1,k , (j ∈ Z) , (1.1)
k∈Z

ψj,0 = (−1)1−k h1−k ϕj+1,k , (j ∈ Z) . (1.2)
k∈Z

The two-scale relations (1.1) and (1.2) are the point of departure to obtain algo-
rithms that compute the graph of MRA scaling functions and wavelets. Examples
are the Daubechies [2] cascade algorithm or the Daubechies–Lagarias [5] infinite
product matrix representation. In the Fourier domain, under suitable conditions,
from relation (1.1) one gets
∞
h(e−πiθ/2 )
a

ϕ̂(θ) = ϕ̂(0) . (1.3)

a=0
21/2

where h(e2πiθ ) := k∈Z hk e2πikθ is the transfer function of the filter hk . Con-
ditions for the infinite product in (1.3) determining the Fourier transform of a
convenient scaling function can be found in [11].
Fourier transform is closely related to translation operators. On the basis of
Fourier analysis, operator techniques have been widely used in wavelet theory, see,
for example, [1, 9, 14]. Beyond Fourier analysis, different choices of ONBs may be
related to spectral and shift representations of translation and dilation operators
on L2 (R). Such ONBs allow to describe every orthonormal wavelet and MRA in
terms of rigid operator-valued functions defined on the corresponding functional
spectral spaces. By-products of the picture are “spectral formulas” useful to design
computational algorithms for orthonormal wavelets and MRAs.
(0)
In particular, in [7] the authors propose to consider arbitrary ONBs {Li }i∈I
(0)
of L2 [0, 1) and {K±,j }j∈J of L2 [±1, ±2), and extend them to a pair of ONBs of
(0) (0)
L2 (R) by means of translations of {Li }i∈I and dilations of {K±,j }j∈J . The two
resultant ONBs of L2 (R),
(n) (0) (m) (0)
{Li := T n Li }i∈I,n∈Z , {Ks,j := Dm Ks,j }s=±,j∈J,m∈Z , (1.4)
lead to shift and spectral representations of the operators T and D, respectively.
The computational usefulness of this approach has been investigated in [6, 8] for
(0) (0)
a particular choice of the ONBs {Li } and {K±,j }, the so-called Haar bases.
(0) (0)
This work explores two new choices of the ONBs {Li } and {K±,j } and
the corresponding spectral algorithms for compactly supported MRA orthonormal
wavelets. The first new bases considered are related with the Walsh–Paley system
and the dyadic harmonic analysis [15]. The second bases are related with the
trigonometric system and the usual harmonic analysis. For the sake of completeness
 Spectral Algorithms for MRA Orthonormal Wavelets 187

we include the generic spectral formulas [7] and some known results [8] for Haar
bases, which are of interest here.

2. Spectral formulas for MRA orthonormal wavelets

As we have indicated in the Introduction, suitable ONBs of L2 (R) lead to spectral
formulas for MRA scaling functions and wavelets. One begins by considering ONBs
(n) (m)
{Li }i∈I,n∈Z and {Ks,j }s=±,j∈J,m∈Z of L2 (R) of the type given by (1.4). For each
f ∈ L2 (R) one has (in L2 -sense)
 (n) (n)  (m) (m)
f= fˆi Li and f = f˜s,j Ks,j
i,n s,j,m
 (n)   (m) 
(shortly, f = fˆi = f˜s,j ). The change of representation between both ex-
pansions is governed by the matrix αs,j,m
i,n , where
!
(n) (m)
αs,j,m
i,n := Li (x) Ks,j (x) dx .
R
For an MRA of L2 (R) with associated filter (hk ), scaling function ϕ and
(n) (m)
corresponding wavelet ψ, once the ONBs {Li } and {K±,j } have been fixed,
Theorem 2.1 below replaces the classical formulas (1.1) and (1.2) for ϕ and ψ
by the spectral formulas (2.1) and (2.2). These spectral formulas are sets of linear
 (n)   (n) 
equations for the components ϕ̂i and ψ̂i with coefficients written in terms
s,j,m
of the hk ’s and αi,n ’s.
 (n) 
Theorem 2.1. Let ϕ = ϕ̂i be the scaling function for an MRA of L2 (R) with
associated filter (hk ) satisfying (1.1). Then
 s,j,m  s,j,m−1  (n−k)
ϕ̂(q)
p = αp,q αi,n hk ϕ̂i , (p ∈ I, q ∈ Z) , (2.1)
s,j,m i,n k
 (n) 
Moreover, the function ψ = ψ̂i ∈ L2 (R) determined by the relations
 s,j,m  s,j,m−1  (n−k)
ψ̂p(q) = αp,q αi,n (−1)1−k h1−k ϕ̂i , (p ∈ I, q ∈ Z) , (2.2)
s,j,m i,n k

is an orthonormal wavelet associated with the MRA generated by ϕ.

For a proof of Theorem 2.1 and further details see Section 5.1 of [7].

3. Spectral algorithms for compactly supported MRA

orthonormal wavelets
The spectral formulas (2.1) and (2.2) of Theorem 2.1 lead to algorithms to compute
(n) (n)
the components ϕi and ψi of the scaling function ϕ and orthonormal wavelet
ψ associated to an MRA with filter (hk ). Since coefficients of the spectral formulas
188 F. Gómez-Cubillo and S. Villullas

(n)
depend on the hk ’s as well as the αs,j,m
i,n ’s, different choices of the ONBs {Li }
(m)
and {K±,j } will lead to different algorithms. Here we deal with three different
(n) (m)
families of ONBs {Li } and {K±,j }: the Haar bases, the Walsh–Paley bases and
the trigonometric bases.
For practical purposes the attention is focused on compactly supported scal-
ing functions and wavelets, so that the filter (hk ) has only a finite number of
nonzero elements hk . In what follows we will assume that there exists N ∈ N such
(n)
that supp ϕ ⊆ [0, N ]. Thus, ϕ̂i = 0 only for n = 0, 1, . . . , N − 1, and for each
i ∈ I we will consider the vector ϕ̂i ∈ CN given by
(0) (1) (N −1) T
ϕ̂i := ϕ̂i , ϕ̂i , . . . , ϕ̂i .
In such case, hk = 0 only if 0 ≤ k ≤ N (see Proposition 7.2 in [11]) and two N × N
matrices, here denoted by H0 and H1 , play a prominent role in the algorithms
deduced below for the three choices of ONBs considered. The matrices H0 and H1
are determined by the filter (hk ) as follows:
[H0 ]lk := h2l−k , [H1 ]lk := h2l−k+1 , (0 ≤ l, k < N ) . (3.1)
The following additional N × N matrices, defined from H0 and H1 , will be used
too (I denotes the imaginary unit):
   
H̃0 := 2−1/2 H0 + H1 , H̃1 := 2−1/2 H0 − H1 , (3.2)
 
H̃k,a := 2−1/2 H0 + e−πIk/2 H1 , (k ∈ Z, a ∈ N ∪ {0}) .
a
(3.3)
It is worth remembering that distinct approaches, as for example fixed point
iterative approximation methods developed by Micchelli and Prautzsch [12] and
Daubechies and Lagarias [4, 5], fall also back on the matrices H0 and H1 . For a
profuse study of such pairs of matrices see Protasov [13] and references therein.
(n)
The spectral algorithms given below determine the coordinates ϕ̂i of the
 (n) 
scaling function ϕ. Once ϕ = ϕ̂i is given, the wavelet ψ can be calculated
(n) (n)
from (1.2) or (2.2). The coordinates ϕ̂i and ψ̂i identify the functions ϕ and ψ
in L2 -sense:
 (n) (n)  (n) (n)
ϕ(x) = ϕ̂i Li (x) , ψ(x) = ψ̂i Li (x) , (in L2 -sense) . (3.4)
i,n i,n

Lawton’s conditions [10] on the filter (hk ) will be of interest here.

Theorem 3.1 (Lawton). Let N ∈ N and let ϕ be a function in L2 (R) such that
supp ϕ ⊆ [0, N ] and the two-scale relation (1.1) is satisfied for a sequence (hk )k∈Z
(thus, hk = 0 for k < 0 or k > N ). Then, ϕ is the scaling function for an MRA
of L2 (R) if and only if

(a) hk = 21/2 ;
k∈Z
 
1, if n = 0,
(b) hk hk+2n =
0, if n ∈ Z\{0};
k∈Z
 Spectral Algorithms for MRA Orthonormal Wavelets 189

(c) for the [2N − 1] × [2N − 1]-dimensional Lawton matrix L defined by


N
Llj = hk hj−2l+k , −N + 1 ≤ l, j ≤ N − 1 ,
k=0
the eigenvalue 1 is nondegenerate.
3.1. Haar bases
Let h0 and h1 be the well-known Haar scaling function and wavelet defined by
h0 := χ[0,1) , h1 := χ[0,1/2) − χ[1/2,1) ,
where χ denotes the characteristic function:

1, if x ∈ [a, b),
χ[a,b) (x) :=
0, otherwise,
For this choice both sets of indices are I = J = N∪{0} when one considers the Haar
ONBs {[LH ]i }i∈N∪{0} of L2 [0, 1) and {[K H]±,j }j∈N∪{0} of L2 [±1, ±2) given by
(0) (0)

[LH ]0 := [h0 ]0,0 , [K H ]+,0 := [h0 ]0,1 , [K H ]−,0 := [h0 ]0,−2 ,

(0) (0) (0)

and, for p ∈ N ∪ {0} and q = 0, 1, . . . , 2p − 1,

[LH ]2p +q := [h1 ]p,q , [K H ]+,2p +q := [h1 ]p,2p +q , [K H ]−,2p +q := [h1 ]p,−2p+1 +q .
(0) (0) (0)

Now, according to (1.4), extend them to ONBs of L2 (R):

{[LH]i (x)}i∈N∪{0},n∈Z and {[K H ]s,j }s=±,j∈N∪{0},m∈Z .
(n) (m)

After determining the change of representation matrix αs,j,m

i,n for these bases one
obtains the following spectral algorithm for the Haar coordinates of a compactly
supported scaling function [8, Theorem 3.1]:
Theorem 3.2. Let N ∈ N and let ϕ ∈ L2 (R) be a scaling function with supp ϕ ⊆
[0, N ] and associated MRA filter (hk )k∈Z satisfying (1.1). Consider the Haar ONB
 (n) 
{[LH ]i } given above and the corresponding expansion ϕ = [ϕ̂H ]i , i.e.,
(n)


[ϕ̂H ]i [LH ]i (x) (in L2 -sense) .
(n) (n)
ϕ(x) = (3.5)
i,n

Then the coordinates

(N −1) T
ϕ̂H H H H
(0) (1)
i := [ϕ̂ ]i , [ϕ̂ ]i , . . . , [ϕ̂ ]i
satisfy the following relations:
ϕ̂H H
0 = H̃0 ϕ̂0 , (3.6)
ϕ̂H
1 = H̃1 ϕ̂H
0 , (3.7)
r k
and, for i > 1 with i = k=0 ik 2 , ik = 0 or 1, and ir = 1,
ϕ̂H
i = Hir−1 · · · Hi2 Hi1 Hi0 ϕ̂H
1 , (3.8)
where H0 , H1 , H̃0 , H̃1 are the N × N matrices given in (3.1) and (3.2).
190 F. Gómez-Cubillo and S. Villullas

Thus, according to (3.6), ϕ̂H 0 is an eigenvector for the eigenvalue 1 of the

matrix H̃0 , which exists by conditions on the filter (hk ). Then, ϕ̂H 1 is obtained
from (3.7), and the rest of ϕ̂Hi , i > 1, using the iterative formula (3.8).
As was to be expected, due to the form of the Haar basis {[LH ]i }, iterative
(n)

formula (3.8) is closely related to Daubechies–Lagarias [5] matrix product approx-

imations at dyadic points. Note, nevertheless, that here the spectral formulas lead
to both the initial conditions (3.6)–(3.7) and the iterative formula (3.8).
The coordinates [ϕ̂H ]i determine the scaling function ϕ in L2 -sense by
(n)

means of the identities (3.5). By construction, the supports of the functions [LH ]i
(n)

form a family of sparse sets. From such sparsity of supports one can deduce, among
other things, the following sufficient conditions on H0 and H1 for the uniform con-
vergence on R of the series in (3.5) (see [8, Proposition 3.1] for details):
Proposition 3.3. Under the conditions of Theorem 3.2, suppose that one of the
following two conditions is satisfied:
(i) The matrices H0 and H1 are diagonalizable and
 
ρ := max |λ| : λ is eigenvalue of H0 or H1 < 1 .
(ii) One has
n −1 3
2 2 −1
n
ρ := max |h2l−k |, |h2l−k+1 | : n1 ≤ l < n2 < 1.
k=n1 k=n1

Then the series in (3.5) converge uniformly on R.

3.2. Walsh–Paley bases
Let r be the function defined by

1 , if x ∈ [0, 12 )
r(x) :=
−1 , if x ∈ [ 12 , 1),
extended to R by periodicity with period 1. The functions of the Rademacher
system {rn }n∈N∪{0} are of the form
rn (x) := r(2n x), (x ∈ R, n ∈ N ∪ {0}).
The Walsh–Paley basis {wn }n∈N∪{0} is built from Rademacher functions as follows:
∞
if n ∈ N ∪ {0} has binary expansion n = k=0 nk 2k , with nk ∈ {0, 1}, then
∞

wn := rnk k .
k=0
This product is always finite because nk = 0 for k large enough. Moreover, w0 = 1
and w2n = rn , for all n ∈ N ∪ {0}. In what follows we restrict the domain of the
functions wn to the interval [0, 1).
The Walsh–Paley ONBs
{[LW ]i }i∈I of L2 [0, 1) and {[K W ]±,j }j∈J of L2 [±1, ±2)
(0) (0)
 Spectral Algorithms for MRA Orthonormal Wavelets 191

are given by
[LW ]i (x) := wi (x) ,
(0)
(i ∈ N ∪ {0}) ,
[K W ]+,j (x) := T wj (x) = wj (x − 1) ,
(0)
(j ∈ N ∪ {0}) .
[K W ]−,j (x) := T −2 wj (x) = wj (x + 2) ,
(0)

Thus, also in this case I = J = N ∪ {0}. Finally, according to (1.4), extend them
to ONBs {[LW ]i (x)}i∈N∪{0},n∈Z and {[K W ]s,j }s=±,j∈N∪{0},m∈Z of L2 (R).
(n) (m)

The spectral algorithm for the Walsh–Paley coefficients (Theorem 3.5 below)
may be obtained by determining explicitly the change of representation matrix
αs,j,m
i,n for these bases, or using the fact that the usual Walsh–Paley and Haar
bases are related by the Hadamard transform [15]. The first approach can be found
in [16, Section 3.3.2]. Here we introduce the second one.
For each p ∈ N, set
p
' l (
aql := 2− 2 wq p , (0 ≤ l, q < 2p , l, q ∈ N) .
(p)
2
By a Hadamard matrix we shall mean a matrix of the form
' (2p −1
(p)
A(p) := aql , (p ∈ N) .
l,q=0

Since wl (q/2 ) = wq (l/2 ), each Hadamard matrix is a real, symmetric and or-
p p

thogonal 2p × 2p matrix. An inductive argument establishes the following recursive

formula
A(0) := 1 , 
A(p) A(p) (3.9)
A(p+1)
= , (p ∈ N ∪ {0}) .
A(p) −A(p)
The Hadamard transform ⊥ is defined on the vector space of real sequences
as follows. Given a real sequence b = (bn )n∈N , define b⊥ = (b⊥
n )n∈N by

b⊥
0 := b0 ,
2p −1 (p)
b⊥
2p +q := l=0 aql b2p +l , (0 ≤ q < 2p , p ∈ N) .

Since the Hadamard matrices are orthogonal, it is clear that (b⊥ )⊥ = b.

A proof of the following result can be found in [15, page 22].
Lemma 3.4. The Walsh–Paley and Haar systems are Hadamard transforms of each
other, i.e., for each n ∈ Z and x ∈ R,
(n)
[LW ]0 (x) = [LH ]0 (x) ,
(n)

(n) 2p −1 (p)

[LW ]2p +q (x) = l=0 aql [LH ]2p +l (x) ,
(n)
(0 ≤ q < 2p , p ∈ N) .

Now, the spectral algorithm for the Walsh–Paley coefficients can be deduced
from Theorem 3.2.
192 F. Gómez-Cubillo and S. Villullas

Theorem 3.5. Let N ∈ N and let ϕ ∈ L2 (R) be a scaling function with supp ϕ ⊆
[0, N ] and associated MRA filter (hk )k∈Z satisfying (1.1). Consider the Walsh–
 (n) 
Paley ONB {[LW ]i } given above and the corresponding expansion ϕ = [ϕ̂W ]i ,
(n)

i.e.,

[ϕ̂W ]i [LW ]i (x) (in L2 -sense) .
(n) (n)
ϕ(x) = (3.10)
i,n

Then the coordinates

(N −1) T
ϕ̂W W W W
(0) (1)
i := [ϕ̂ ]i , [ϕ̂ ]i , . . . , [ϕ̂ ]i

satisfy the following relations:

ϕ̂W W
0 = H̃0 ϕ̂0 , (3.11)
r k
and, for i > 1 with i = k=0 ik 2 , ik = 0 or 1, and ir = 1,

ϕ̂W W
i = H̃i0 H̃i1 H̃i2 · · · H̃ir ϕ̂0 , (3.12)
where H̃0 and H̃1 are the N × N matrices given in (3.2).

Proof. The definition of the matrices H0 , H1 , H̃0 and H̃1 given in (3.1) and (3.2)
and the recursive formula (3.9) imply that, for p ∈ N, 0 ≤ q, j < 2p ,
p
2 −1
(p)
H̃(2p +j)0 H̃(2p +j)1 · · · H̃(2p +j)p−1 = aql H(2p +l)p−1 H(2p +l)p−1 · · · H(2p +l)0 ,
l=0
p
p p k
where 2 + j = k=0 (2 + j)k 2 . Use these relations and Lemma 3.4 to deduce
(3.11) and (3.12) from (3.6)–(3.8). 

Once more, the coordinates [ϕ̂W ]i determine the scaling function ϕ in L2 -

(n)

sense by means of the identities (3.10). In this case, conditions (i) or (ii) in Propo-
sition 3.3 also imply the uniform convergence on R of the series in (3.10). See
[16, Proposition 47] for details.

3.3. Trigonometric bases

Let {ek (x) := exp(2πIkx)}k∈Z be the usual trigonometric basis of L2 [0, 1). Define
the trigonometric ONBs {[LT ]i }i∈I of L2 [0, 1) and {[K T ]±,j }j∈J of L2 [±1, ±2) by
(0) (0)

[LT ]i (x) := ei (x) ,

(0)
(i ∈ Z) ,
[K T ]+,j (x)
(0)
:= T ej (x) = ej (x − 1) ,
(j ∈ Z) .
[K T ]−,j (x) := T −2 ej (x) = ej (x + 2) ,
(0)

In this case I = J = Z and one obtains the ONBs

{[LT ]i }i∈Z,n∈Z and {[K T ]s,j }s=±,j∈Z,m∈Z

(n) (m)
 Spectral Algorithms for MRA Orthonormal Wavelets 193

of L2 (R) following relations (1.4). In what follows, given ϕ ∈ L2 (R), we will

study the coefficients [ϕ̂T ]i corresponding to the expansion of ϕ in the ONB
(n)

{[LT ]i }i∈Z,n∈Z :
(n)


[ϕ̂T ]i [LT ]i (x) (in L2 -sense) .
(n) (n)
ϕ(x) =
i,n

The operator norm for N × N matrices H is defined by

||Hv||2
||H|| := sup ,
||v||2 =0 ||v||2

where

N 1/2
||v||2 := |vn |2
n=1

is the Euclidean norm of v = (v1 , . . . , vN ) ∈ CN .

Theorem 3.6. Let N ∈ N and consider the two-scale equation (1.1), where hk = 0
for k < 0 or k > N and
 
h2n = h2n+1 = 2−1/2 . (3.13)
n∈Z n∈Z

Let H0 , H1 and H̃i,a be the N × N matrices given in (3.1) and (3.3). Define E1
to be the (N − 1)-dimensional subspace of CN orthogonal to e1 = (1, . . . , 1), the
common left eigenvector of H0 and H1 for the eigenvalue 2−1/2 . Assume that there
exist λ < 1 and C > 0 such that, for all m ∈ N,
 m
λ
max Hd1 Hd2 · · · Hdm |E1  ≤ C . (3.14)
dj ∈{0,1} 21/2
j=1,2,...,m

Then, there is a solution ϕ to (1.1) such that supp ϕ ⊆ [0, N ] and ϕ is Hölder
continuous,
|ϕ(x) − ϕ(y)| ≤ C |x − y|| ln(λ)|/ ln(2) . (3.15)
Moreover, for each i ∈ Z, the following limit of matrix products exists

m ∞

lim H̃i,a = lim H̃i,0 H̃i,1 · · · H̃i,m =: H̃i,a = H̃i,0 H̃i,1 H̃i,2 · · ·
m→∞ m→∞
a=0 a=0

and

∞ 
= [ϕ̂T ]i
(n)
H̃i,a , (0 ≤ n, l < N ) . (3.16)
a=0 n,l

Proof. Under the conditions of the theorem, the existence of a solution ϕ to the
two-scale equation (1.1) such that supp ϕ ⊆ [0, N ] and ϕ satisfies (3.15) is part of
Theorem 2.2 in [5], taking into account that the matrices T0 and T1 considered by
194 F. Gómez-Cubillo and S. Villullas

Daubechies and Lagarias in [5] are in correspondence with our 21/2 H0 and 21/2 H1 ,
respectively. Now, for x ∈ [0, 1), consider its binary expansion
∞

x= dj (x) 2−j , where dj (x) = 0 or 1 for all j .
j=1

It is known (see Remark 2 on page 1040 of [5]) that, under the conditions of
Theorem,
⎛ ⎞
ϕ(x) ϕ(x) ... ϕ(x)
∞ ⎜ ϕ(x + 1) ϕ(x + 1) ... ϕ(x + 1) ⎟
⎜ ⎟
21/2 Hdj (x) = ⎜ .. .. .. ⎟,
j=1
⎝ . . . ⎠
ϕ(x + N − 1) ϕ(x + N − 1) . . . ϕ(x + N − 1)
where the convergence of the product is uniform on x ∈ [0, 1). Indeed, since
e1 21/2 H0 = e1 21/2 H1 = e1 , for all x ∈ [0, 1), s > r ∈ N and v ∈ CN ,
7 
s 8
e1 21/2 Hdj (x) − Id v T = 0,
j=r+1
 s 
i.e., j=r+1 21/2 Hdj (x) − Id v T ∈ E1 , so that, by (3.14),

s 
r
r→∞
21/2 Hdj (x) − 21/2 Hdj (x) ≤ Cλr −→ 0 .
j=1 j=1

Then, given  > 0, there exists r1 ∈ N such that, for all r ≥ r1 ,

'
r ( 
21/2 Hdj (x) − ϕ(x + n − 1) < , (1 ≤ n, l < N, x ∈ [0, 1)) . (3.17)
j=1
n,l 2

On the other hand, for p ∈ Z and r ∈ N ∪ {0},

r '
H0 + e− 2a H1 ( −1

r  πpI 2r+1 
r
1 πpIm
H̃p,a = = e− 2r Hmr−k , (3.18)
a=0 a=0
21/2 2(r+1)/2 m=0 k=0
r
where m = k=0 mk 2k . For each such 0 ≤ m < 2r+1 , put

m 
r 
r+1
x(m)
r := = mk 2k−r−1 = mr+1−j 2−j ,
2r+1 j=1
k=0

so that (3.18) can be written as


r
1 −1
2r+1
(m) 
r+1
H̃p,a = e−2πpIxr 21/2 Hdj (x(m) ) . (3.19)
a=0
2r+1 m=0 j=1
r
 Spectral Algorithms for MRA Orthonormal Wavelets 195

Since ϕ is Hölder continuous and supp ϕ ⊆ [0, N ], the function e−2πpI· ϕ(·) is
Riemann integrable and one can approximate its integral on each interval [n, n+1],
(0 ≤ n < N ), by means of Riemann sums:
−1 −2πpI m ' m
2r+1 ( ! n+1
1
e−2πpIx ϕ(x) dx = [ϕ̂T ](n)
r→∞
e 2r+1 ϕ + n −→ p ,
2r+1 m=0 2r+1 n

that is, given  > 0, there exists r2 ∈ N such that for all r ≥ r2 ,

1 −1
2r+1
(m) 
e−2πpIxr ϕ x(m)
r + n − [ϕ̂T ](n)
p < . (3.20)
2r+1 m=0
2

Thus, from (3.17), (3.19) and (3.20), for r ∈ N such that r ≥ max{r1 , r2 } and
0 ≤ n, l < N ,
' r (
H̃p,a − [ϕ̂T ](n)
p
n,l
a=0

1 −1
2r+1
(m)
' r+1
 (
= e−2πpIxr 21/2 Hdj (x(m) ) − [ϕ̂T ](n)
p
2r+1 m=0 j=1
r n,l
⎡ ⎤
1
r+1
2 −1 ' r+1
 (
−2πpIx(m) ⎣
≤ e r 21/2 Hdj (x(m) ) − ϕ x(m)
r +n ⎦
2r+1 m=0 j=1
r n,l

−1
r+1
2
1 (m)
+ e−2πpIxr ϕ x(m)
r + n − [ϕ̂T ](n)
p
2r+1 m=0

−1 ' r+1 (

r+1
1
2  
< 21/2 Hdj (x(m) ) − ϕ x(m)
r +n + <
2r+1 m=0 j=1
r n,l 2

and one gets (3.16). 

Remarks 3.7.
(1) Conditions (a) and (b) in Theorem 3.1 imply (3.13). See Remark 5 on page
144 and the beginning of Section 6.2 in [3].
(2) If λ < 1/2 in (3.14), then (3.15) implies that |ϕ(x) − ϕ(y)| ≤ C |x − y|μ with
μ > 1. In such case, ϕ must be constant and, since supp ϕ ⊆ [0, N ], one has
ϕ = 0. Thus, λ in (3.14) necessarily satisfies λ ∈ [1/2, 1).
(3) The condition (3.14) is equivalent to the following one: λm < 2−1/2 for some
m ∈ N, where
λm := max Hd1 Hd2 · · · Hdm |E1 1/m .
dj ∈{0,1}
j=1,2,...,m

This condition can also be interpreted in terms of the joint spectral radius
for H0 |E1 and H1 |E1 . See Lemmas 2.3 and 2.4 in [5] for details.
196 F. Gómez-Cubillo and S. Villullas

(4) With respect to the regularity of the solutions to the two-scale relation (1.1)
Daubechies and Lagarias proved the following result [4, Theorem 5.1]: Let
N ∈ N and consider the two-scale equation (1.1), where hk = 0 for k < 0

or k > N and N k=0 hk = 2
1/2
(condition (a) in Theorem 3.1). Define M =
(Mlk ) to be the [N − 1] × [N − 1] matrix
Mlk := h2l−k+1 , (0 ≤ l, k < N − 1) .
If there exists a non-trivial L -solution ϕ of (1.1) (thus, with supp ϕ ⊆
1

[0, N ]), which is in C m (R), then 2−1/2 , 2−(1+1/2) , . . . , 2−(m+1/2) belong to

the spectrum of M . In particular, m < N − 1.
Corollary 3.8. Under the conditions of Theorem 3.6, the coordinates
(N −1) T
ϕ̂T T T T
(0) (1)
i := [ϕ̂ ]i , [ϕ̂ ]i , . . . , [ϕ̂ ]i

satisfy the following relations: for i ∈ Z,

' ∞ (
ϕ̂T
i = H̃i,a ϕ̂T
0 . (3.21)
a=0

where H̃k,a are the matrices given in (3.3). Then, for i ∈ Z and b ∈ N,

ϕ̂T b T
2b i = H̃0 ϕ̂i , (3.22)
where H̃0 is the N × N matrix given in (3.2). In particular,
ϕ̂T T
0 = H̃0 ϕ̂0 . (3.23)
Proof. Formula (3.21) is a direct consequence of (3.16) and the usual assumption
for scaling functions 
[ϕ̂T ]0 = ϕ̂(0) = 1 .
(n)

n
(Necessarily, |ϕ̂(0)| = 1, so that the above assumption means a choice for the phase
of ϕ, see [3, Section 6.2] for details). According to (3.16) and the definition of the
matrices H̃k,a given in (3.3), for i ∈ Z and b ∈ N,

1 7 8 17 87  8
∞ b−1 ∞
ϕ̂T
2b i = H̃2b i,a eT1 = H̃2b i,a H̃2b i,a eT1
N a=0 N a=0
a=b

1 7 87  8
b−1 ∞
= H̃0 H̃i,a eT1 = H̃0b ϕ̂T
i ,
N a=0 a=0

that is, (3.22) is satisfied. Now, to get (3.23), it suffices to take i = 0 and b = 1 in
(3.22). 
Taking into account that the usual Fourier series of an MRA scaling function
and orthogonal wavelet are unuseful, because all their Fourier coefficients (except
for k = 0) are null (see [3, Section 6.2]), Corollary 3.8 can be seen as an attempt to
achieve a discrete frequencial analysis for these functions. Unfortunately, relation
 Spectral Algorithms for MRA Orthonormal Wavelets 197

(3.21) does not disengage us from the infinite products already present in the
continuous Fourier formula (1.3).
There are other ways to obtain the results of Theorem 3.6 and Corollary 3.8.
One of them is to consider the change of representation matrix αs,j,m
i,n for trigono-
metric bases. The details can be found in Appendix A.3 and Section 3.3.3 of [16].
A second method uses the additional functions e−2π 2a Ix , for a ≥ 0. Let us put
i

(N −1) T
ϕ̂Tia := [ϕ̂T ] , [ϕ̂T ] , . . . , [ϕ̂T ]
(0) (1)
i i i ,
2 2a 2a 2a

where
! n+1
[ϕ̂T ] ϕ(x) e−2π 2a Ix dx ,
(n) i
i := (0 ≤ n < N ) .
2a
n
Under the conditions of Theorem 3.6, the relationship between trigonometric and
Haar bases leads to

−2a I 7' ( ' (2 ' u−1 ( 8

∞ 
T − 2πiI H − 2a+u
H̃i,a+v ϕ̂H
πiI
ϕ̂ i = 1−e a−1
ϕ̂0 + 2 1−e
u
1 ,
2a 2πi u=0 v=0

so that
ϕ̂Tia = H̃p,a ϕ̂T i , (a ≥ 0) .
2 2a+1

A recursive use of this formula imply, in particular, (3.21). See again [16, Section
3.3.3] for details.

4. Conclusions
Operator techniques developed in [7] lead to spectral formulas for scaling functions
and orthonormal wavelets associated with MRAs of L2 (R). The approach takes
(n) (m)
advantage of good properties of pairs of suitable ONBs {Li } and {K±,j } relative
to translation and dilation operators on L2 (R). The spectral formulas are sets of
linear equations for the coordinates of the scaling functions and wavelets in their
(n)
expansions with respect to the ONB {Li } (see Theorem 2.1 and (3.4)).
(n) (m)
Once the ONBs {Li } and {K±,j } are fixed, a spectral algorithm derives
from the spectral formulas to compute scaling functions and wavelets. Here we
have presented the spectral algorithms for three different families of ONBs: Haar
bases in Section 3.1, Walsh–Paley bases in Section 3.2 and trigonometric bases in
Section 3.3.
Particular choices of ONBs focus attention on concrete properties of scaling
functions and wavelets. So, for example, Haar bases target at point properties,
like regularity, whereas trigonometric bases address a discrete frequencial analysis
(recall that Fourier series are unusable for compactly supported scaling functions
and wavelets). Other choices will be studied elsewhere.
198 F. Gómez-Cubillo and S. Villullas

References
[1] X. Dai and D.R. Larson, Wandering vectors for unitary systems and orthogonal
wavelets. Mem. Amer. Math. Soc. 134 (1998), pp. viii+68.
[2] I. Daubechies, Orthonormal bases of compactly supported wavelets. Comm. Pure
Appl. Math. 41 (1988), 909–996.
[3] I. Daubechies, Ten lectures on wavelets, vol. 61 of CBMS-NSF Regional Conference
Series in Applied Mathematics, Society for Industrial and Applied Mathematics
(SIAM), Philadelphia, PA, 1992.
[4] I. Daubechies and J.C. Lagarias, Two-scale difference equations. I. Existence and
global regularity of solutions. SIAM J. Math. Anal. 22 (1991), 1388–1410.
[5] I. Daubechies and J.C. Lagarias, Two-scale difference equations. II. Local regularity,
infinite products of matrices and fractals. SIAM J. Math. Anal. 23 (1992), 1031–
1079.
[6] F. Gómez-Cubillo and Z. Suchanecki, Inner functions and local shape of orthonor-
mal wavelets. Appl. Comput. Harmon. Anal. 30 (2011), 273–287.
[7] F. Gómez-Cubillo and Z. Suchanecki, Spectral models for orthonormal wavelets and
multiresolution analysis of L2 (R). J. Fourier Anal. Appl. 17 (2011), 191–225.
[8] F. Gómez-Cubillo, Z. Suchanecki, and S. Villullas, Orthonormal MRA wavelets:
spectral formulas and algorithms. Int. J. Wavelets Multiresolut. Inf. Process. 10
(2012), pp. 1250008, 19.
[9] T.N.T. Goodman, S.L. Lee, and W.S. Tang, Wavelets in wandering subspaces.
Trans. Amer. Math. Soc. 338 (1993), 639–654.
[10] W.M. Lawton, Necessary and sufficient conditions for constructing orthonormal
wavelet bases. J. Math. Phys. 32 (1991), 57–61.
[11] S. Mallat, A Wavelet Tour of Signal Processing. Academic Press, Inc., San Diego,
CA, 1998.
[12] C.A. Micchelli and H. Prautzsch, Uniform refinement of curves. Linear Algebra
Appl. 114/115 (1989), 841–870.
[13] V. Protasov, Refinement equations and corresponding linear operators. Int. J.
Wavelets Multiresolut. Inf. Process. 4 (2006), 461–474.
[14] A. Ron and Z. Shen, Affine systems in L2 (Rd ): the analysis of the analysis operator.
J. Funct. Anal. 148 (1997), 408–447.
[15] F. Schipp, W.R. Wade, and P. Simon, Walsh Series. An Introduction to Dyadic
Harmonic Analysis. Adam Hilger, Ltd., Bristol, 1990.
[16] S. Villullas, Fórmulas y algoritmos espectrales para wavelets ortonormales y wavelet
frames ajustados. Aplicaciones en el procesamiento digital de imágenes. PhD thesis,
IMUVa-LPI, Universidad de Valladolid, Valladolid, Spain, 2016.

F. Gómez-Cubillo and S. Villullas

Dpto de Análisis Matemático
Universidad de Valladolid, Facultad de Ciencias
E-47011 Valladolid, Spain
e-mail: [email protected]
[email protected]
Operator Theory:
Advances and Applications, Vol. 267, 199–220

c Springer International Publishing AG, part of Springer Nature 2018

The NIEP
Charles R. Johnson, Carlos Marijuán,
Pietro Paparella and Miriam Pisonero

Abstract. The nonnegative inverse eigenvalue problem (NIEP) asks which

lists of n complex numbers (counting multiplicity) occur as the eigenvalues of
some n-by-n entry-wise nonnegative matrix. The NIEP has a long history and
is a known hard (perhaps the hardest in matrix analysis?) and sought after
problem. Thus, there are many subproblems and relevant results in a variety
of directions. We survey most work on the problem and its several variants,
with an emphasis on recent results, and include 130 references. The survey
is divided into: a) the single eigenvalue problems; b) necessary conditions; c)
low-dimensional results; d) sufficient conditions; e) appending 0’s to achieve
realizability; f) the graph NIEP’s; g) Perron similarities; and h) the relevance
of Jordan structure.
Mathematics Subject Classification (2010). Primary 15A18, 15A29, 15A42.
Keywords. Nonnegative matrices, nonnegative inverse eigenvalue problem
(NIEP), symmetric realizability.

1. Introduction
The Nonnegative Inverse Eigenvalue Problem (NIEP) asks which collections of
n complex numbers (counting multiplicities) occur as the eigenvalues of an n-by-n
matrix, all of whose entries are nonnegative real numbers. This is a long stand-
ing problem that is very difficult and, perhaps, the most prominent problem in
matrix analysis. Unlike many other inverse eigenvalue problems, tools are limited,
and seemingly small results are very welcome. Nonetheless, the problem is very
attractive and has been attacked by many excellent researchers (see references).
There have been prior surveys of work on the NIEP [4, 7, 20, 27, 92, 117, 129],
but none recently. With a great deal of important recent activity, a new survey,
emphasizing this activity, will be welcome to anyone considering inquiry in this
area. Our intent here is to be broad and informal.
Partially supported by MTM2015-365764-C-1-P(MINECO/FEDER).
200 C.R. Johnson, C. Marijuán, P. Paparella and M. Pisonero

Another intriguing aspect of the NIEP is how many avenues can be taken
to gain insight into it. Thought about it spawns endless challenging and worthy
specific questions. An example is how many interesting variations there are. The
real NIEP (R-NIEP) restricts the question to real spectra, and the symmetric
NIEP (S-NIEP) further restricts to real spectra that are realizable by symmetric
nonnegative matrices. It is known [48] that the row stochastic NIEP (restriction
to nonnegative matrices with row sums 1) is equivalent to the general NIEP, but
the doubly stochastic NIEP (DS-NIEP), restriction to matrices with both row and
column sums 1, is properly more restrictive, but no less difficult. Of course, there
are variations on the DS-NIEP, as well: the real DS-NIEP and the symmetric DS-
NIEP. The Jordan NIEP (J-NIEP) asks about possible Jordan canonical forms,
when there are repeated eigenvalues, and the diagonalizable NIEP (D-NIEP) is
the special case in which the realizing matrix is diagonalizable. Finally, there are
graph NIEP’s when we consider only nonnegative matrices with a given (directed
or undirected) graph G (G-NIEP), or only matrices subordinate to same. Further
particularizations may also be imagined, but these are the primary ones, so far. We
should also mention that the NIEP and each variation also has a trace 0 version.
This considers only spectra, the sum of whose elements is 0, or, equivalently,
nonnegative matrices, each of whose diagonal entries is 0. In some cases, this is
sufficiently restrictive, so as to make the problem easier.
We denote by NIEPn the set of all spectra enjoyed by some n-by-n nonneg-
ative matrix. Such a spectrum is called realizable and a nonnegative matrix with
the spectrum is called a realization or realizing matrix. The notation and termi-
nology are similar for NIEP variants. A solution to the NIEP (or variants), for a
particular n, is an “explicit” description of NIEPn , viewed as a subset of vectors
in Cn . NIEPn is a closed set that is connected, and even star-shaped from the
origin (or from the all 1’s vector e). The set is also semi-algebraic, but is not gen-
erally convex. If Λ = {λ1 , . . . , λn } is a proposed spectrum, we denote by Tr(Λ)
 the
sum of its components, counting multiplies. Other power sums sk (Λ) = ni=1 λki ,
called kth moments, are also important, and we usually reserve λ1 as the spectral
radius of Λ.
The Perron–Frobenius theory of nonnegative matrices (e.g., [43]) provides
several simple, but important, necessary conditions for the NIEP (and variations).
These basic necessary conditions are presented in Section 3, followed by several
more subtle necessary conditions. The “JLL” conditions are now essential to work
on the problem.
The NIEP has a long history since its proposal by Kolmogorov [62], and there
are many substantial results. We begin at the natural starting point of the single
eigenvalue problem (Section 2) for both the row stochastic NIEP and the DS-
NIEP: which individual complex numbers occur in the spectra of row stochastic
and of doubly stochastic matrices. The former is solved, but there have been
important advances in the description of the solution and in realizations. The latter
is unsolved, beyond n = 4, though there have been important recent developments.
 The NIEP 201

Section 4 summarizes low-dimensional complete results for certain NIEP vari-

ants. General sufficient conditions for the realizability of spectra are given in Sec-
tion 5. It has long been known that n-fold spectra that meet simple necessary
conditions, but are not realizable, may be made realizable by appending of 0
eigenvalues. Information about this phenomenon is given in Section 6. We turn to
what is known about the graph-NIEP’s in Section 7. The new idea of Perron simi-
larities – studying the diagonalizable NIEP’s via the diagonalizing similarities – is
discussed in Section 8, and the role of Jordan structure in the NIEP and R-NIEP
in Section 9.

2. The Single Eigenvalue Problems

In [62], Kolmogorov posed the problem, denoted by SISEP (Stochastic Inverse
Single Eigenvalue Problem), of characterizing the subset of the complex plane,
denoted by Θn , consisting of the individual eigenvalues of all n-by-n stochastic
matrices.
It can easily be verified that for each n ≥ 2, the region Θn is closed, inscribed
in the unit-disc, star-convex (with star-centers at zero and one), and symmetric
with respect to the real-axis. Furthermore, it is clear that Θn ⊆ Θn+1 , ∀n ∈ N.
In view of these properties, each region Θn is determined by its boundary, which
consists of so-called extremal numbers, i.e., ∂Θn = {λ ∈ Θn : αλ ∈ Θn , ∀α > 1}.
In pursuit of characterizing Θn , Dmitriev and Dynkin [24] (and independently
Karpelevich [57]) showed that if λ = a + bi ∈ Θn and b = 0, then
'π(
a + |b| tan ≤ 1.
n
Let A be a nonnegative matrix of order n with spectral radius ρ and associated
directed graph G. Let m be the length of the longest simple circuit of G. Kellogg
and Stephens [59, Theorem 1] showed that if m = 2, all eigenvalues of A are real.
If 2 < m ≤ n, and if λ = a + bi is an eigenvalue of A, then
'π(
a + |b| tan ≤ ρ.
m
For further results on this topic, see [49].
Karpelevich [57, Theorem B], expanding on the work of Dmitriev and Dynkin
in [25], resolved SISEP by showing that the boundary of Θn consists of curvilinear
arcs (called K-arcs), whose points satisfy a polynomial equation that is determined
by the endpoints of the arc (which are consecutive roots of unity). The statement
of this theorem is unwieldy, but Ito [46, Theorem 2] provides a useful simplification
of the result.
Noticeably absent in the Karpelevich Theorem (and other works) are realiz-
ing-matrices (i.e., a matrix whose spectrum contains a given eigenvalue) for points
on the K-arcs. Dmitriev and Dynkin [25, Basic Theorem] give a schematic descrip-
tion of such matrices for points on the boundary of Θn \Θn−1 and Swift [129, §2.2.2]
202 C.R. Johnson, C. Marijuán, P. Paparella and M. Pisonero

provides such matrices for 3 ≤ n ≤ 5. Recently, Johnson and Paparella [55] pro-
vide, for every n and for each K-arc, a single parametric stochastic matrix that
realizes the entire K-arc as the parameter runs from 0 to 1.
For n ∈ N, denote by Πn the convex-hull of the nth roots-of-unity, i.e.,
n−1 3
 
n−1
Πn = αk exp (2πik/n) ∈ C : αk ≥ 0, αk = 1 .
k=0 k=0

Denote by Ωn the subset of the complex-plane containing all single eigenvalues of

all n-by-n
n doubly stochastic matrices. Perfect and Mirsky [102] conjectured that
Ωn = k=1 Πk and proved their conjecture when 1 ≤ n ≤ 3. Levick et al. [75]
proved the Perfect–Mirsky conjecture when n = 4 but a counterexample when
n = 5 was given by Mashreghi and Rivard [87]. Recently, Levick et al. conjectured
that Ωn = Θn−1 ∪ Πn ([75, Conjecture 1]), but there is computational evidence
that suggests that the n = 5 case is either a rare exception, or the only exception,
to Perfect–Mirsky.

3. Necessary conditions for the NIEP’s

For a proposed spectrum Λ = {λ1 , λ2 , . . . , λn }, repeats allowed, to be (NIEP-)
realizable, a number of necessary conditions are known. The most basic of these
follow from the fact that a nonnegative matrix has real entries and nonnegative
trace, and from the Perron–Frobenius theory of nonnegative matrices. (We assume
the reader is familiar with the Perron–Frobenius theory, which is recounted, for
example in [43].) For simplicity, we label these according to simple titles. Since a
nonnegative matrix is real, it must have the eigenvalues of a real matrix, i.e., the
characteristic polynomial must have real coefficients, or, as a list
(Reality) Λ̄ = Λ.

Since the trace of a matrix is the sum of the eigenvalues and the trace of a non-
negative matrix is nonnegative, we also have

n
(Trace) Tr Λ = λi ≥ 0.
i=1

Of course, it follows that if Λ is realizable and Tr (Λ) = 0, then all diagonal entries
of any realizing matrix must be 0. Higher power sums are also nonnegative as they
are traces of positive integral powers of nonnegative matrices:

(kth moment) sk (Λ) = λk1 + · · · + λkn ≥ 0, k ≥ 1.

According to the Perron–Frobenius theory, the spectral radius of a nonnega-

tive matrix must, itself, be an eigenvalue, so among the eigenvalues Λ must be a
 The NIEP 203

nonnegative one, at least as big in absolute value as any others. Without loss of
generality, this one may be taken to be λ1 . So, we have

(Perron) λ1 ≥ |λi |, i = 2, . . . , n,

and if any of the other eigenvalues are the same as λ1 , any realizing matrix must
be reducible. Moreover, if all the inequalities are strict, any irreducible realizing
matrix must be primitive.
The above conditions are necessary for all the NIEP’s, but there are some
necessary conditions for particular NIEP’s.
Further necessary conditions are more subtle, but many have been noticed.
The most ubiquitous of these was noticed independently in [48] and in [77] and
is usually referred to as the J-LL condition. They generalize the trace condition
and follow from the fact that every power of a nonnegative matrix is nonnegative
and that positive diagonal entries must contribute to positive diagonal entries in
powers (e.g., sp (Λ) > 0 implies spq (Λ) > 0 for integers p ≥ 1 and q ≥ 1). The
general (quantitative) version is

(JLL) (sk (Λ))m ≤ nm−1 skm (Λ), k, m = 1, 2, . . . .

Let A be an n-by-n real matrix with spectrum Λ = {λ1 , . . . , λn }. Denote the

principal submatrix of A lying in the rows and columns given by the index set
α ⊆ {1, . . . , n} by A[α]. Define the kth elementary symmetric function

Ek (Λ) = λi1 · · · λik
1≤i1 <···<ik ≤n

and the kth Newton coefficient

Ek (Λ)
ck (Λ) = n , k = 1, . . . , n, with c0 ≡ 1.
k

Since Ek (Λ) = |α|=k det A[α], as well, ck (Λ) may be viewed as the average value
of the k-by-k principal minors of A. The spectrum Λ is called Newton if
ck (Λ)2 ≥ ck−1 (Λ)ck+1 (Λ), k = 1, . . . , n − 1,
and these inequalities are referred to as the Newton inequalities [42, 51].
They hold for Λ ≥ 0 [96], Λ ⊂ R [80], and they are valid for real diagonal
matrices, diagonalizable matrices with real spectra (the ck are invariant under
similarity), and matrices with real spectra. In [42] it was proved that the Newton
inequalities also hold for M -matrices and, thus, inverse M -matrices and it was
observed that if A is a nonnegative matrix with spectral radius ρ(A), then ρ(A)I−A
is an M-matrix and, therefore, must satisfy the Newton inequalities. If we denote
its spectrum {ρ(A) − λ1 , . . . , ρ(A) − λn } by ρ(A) − Λ, then we have new necessary
conditions:
(H) ck (ρ(A) − Λ)2 ≥ ck−1 (ρ(A) − Λ)ck+1 (ρ(A) − Λ), k = 1, . . . , n − 1.
204 C.R. Johnson, C. Marijuán, P. Paparella and M. Pisonero

In [130] the authors did not focus the attention directly on the spectrum but
on the coefficients of the characteristic polynomial. Thus, the NIEP that they con-
sider is: given real numbers k1 , k2 , . . . , kn , find necessary and sufficient conditions
for the existence of a nonnegative matrix of order n with characteristic polynomial
xn + k1 xn−1 + k2 xn−2 + · · · + kn .
The coefficients of the characteristic polynomial are closely related to the
cyclic structure of the weighted digraph associated with the matrix A, as estab-
lished by the Coefficients Theorem [22, Theorem 1.3*]. The authors [130] introduce
graphic tools to study the NIEP from the characteristic polynomial and use the
following method: if P (x) is a realizable polynomial, in the sense that there exists
a nonnegative matrix with characteristic polynomial P (x), we try to maximize
each coefficient kj as a function of the previous coefficients, preserving the real-
izability for a polynomial of degree n with the same previous coefficients. Note
that kj is a continuous function (sum of determinants) of the entries of a nonnega-
tive matrix A realizing the polynomial P (x) and that, as the previous coefficients
are bounded above, then the entries of A involved in the expression of kj are
also bounded above (Coefficients Theorem); therefore, this maximum is attained.
In this way, new necessary conditions on the three first coefficients are obtained
[130, Theorem 3]:
k1 ≤ 0;
n−1 2
(TAAMP) k2 ≤ k ;
2n 1
⎧  ' ( 3/2 
⎪
⎨ n−2 k1 k2 + n−1 k 2
− 2nk2
− k 3
if (n−1)(n−4) 2
n 3n 1 n−1 1 2(n−2)2 k1 < k2 ,
k3 ≤
⎪
⎩ k k − (n−1)(n−3) k 3
1 2 3(n−2)2 1 if k2 ≤ (n−1)(n−4) 2
2(n−2)2 k1 .

Another necessary condition in terms of the kth moments sk = Tr (Ak ),

k = 1, 2, 3, is obtained in [21, Theorem 3]:
(CL) Φ := n2 s3 − 3ns1 s2 + 2s31 + √n−2 (ns2
n−1
− s21 )3/2 ≥ 0.

The necessary conditions previous to (JLL) are not independent. In fact, (kth
moment) implies (Perron) [33] and also (Reality) [77]. For k = 1, (JLL) is reduced
to (s1 (Λ))m ≤ nm−1 sm (Λ), and so, if s1 (Λ) ≥ 0, then (JLL) implies (kth moment).
On the other hand, if we denote the two bounds given for the coefficient k3
in (TAAMP) by k3max 1 and k3max 2 , we can use the Newton identities
sm + k1 sm−1 + · · · + km−1 s1 + km m = 0, k, m = 1, 2, 3, (1)
to rewrite the condition (CL) in the form
Φ = 3n2 (k3max 1 − k3 ).
In the first case, i.e., if k2 > (n−1)(n−4)
2(n−2)2 k1 , we have that Φ ≥ 0 implies k3
2 max 1
≥ k3 .
In the contrary case, Φ = 3n (k32 max 1
− k3 ) ≥ 3n (k3
2 max 2
− k3 ) ≥ 0. So, in any case,
the condition on the coefficient k3 is stronger than the condition (CL).
 The NIEP 205

In [42] it was proved that (kth moment), (JLL) and (H) are mutually indepen-
dent. In [83, Theorem 11] that (JLL) for k = 1 and m = 2, s1 (Λ)2 ≤ ns2 (Λ), the
first (H), c1 (ρ−Λ)2 ≥ c2 (ρ−Λ), and (TAAMP) over the second coefficient, k2 (Λ) ≤
n−1 2
2n k1 (Λ) , are equivalent. But, in general, at least (JLL) and (TAAMP) are inde-
pendent of the others,
√ and√this tandem implies (kth moment). Some examples: the
spectra {20, −18, 5 2 ± 5 2i} and {1, 1, 1, 0, 0} satisfy (JLL), (H) and (CL), but
not (TAAMP); the spectra {2, −2, −2, 1 ± i} and {3, 1, 1, 1, 1, 1, −2, −2, −2, −2}
satisfy (H) and (TAAMP), but not (JLL). That the necessary conditions (JLL),
(H) and (TAAMP)
√ are not sufficient for the NIEP is proved by the non-realizable
list {3, 3, − 3 ± i}.
Conjecture. (JLL) and (TAAMP) imply all known necessary conditions. (It is
enough to prove that (JLL) and (TAAMP) imply (H).)
In [68] a necessary condition for trace 0 and n, odd was obtained, given by
(LM) (s2 (Λ))2 ≤ (n − 1)s4 (Λ).
This condition has been generalized in terms of the coefficients by the follow-
ing result [130, Lemma 37]: if xn + kp xn−p + · · · + k2p xn−2p + · · · + kn , kp = 0, is
the characteristic polynomial of a nonnegative matrix, then
 
1 1
k2p ≤ 1− kp2 . (2)
2 %n/p&
We can use the Newton identities (1) to express the inequality (2) in terms of the
kth moments: if s1 = · · · = sp−1 = 0, we have the more general condition
EnF n
s2p ≤ p s2p , p = 1, . . . ,
p 2
that coincides with (LM) in the particular case p = 2 and n odd.
Note also that the (JLL) condition for m = 2 is s2p ≤ ns2p , and that
E F
p np ≤ n. Then the expression (2) is a restricted refinement of the (JLL) condi-
tions.

4. Low-dimensional results
The NIEP for n ≤ 3 was solved independently by Oliveira [97, Theorem (6.2)] and
Loewy–London [77]. For the non-real case:
{λ, z, z̄} is NIEP-realizable ⇐⇒ z ∈ λΠ3 = {λz  : z  ∈ Π3 }
Meehan [89] solved the NIEP for n = 4 in terms of the kth moments, and Torre–
Mayo et al. [130] in terms of the coefficients of the characteristic polynomial using
the necessary conditions (TAAMP). For n ≥ 5 it remains unsolved.
The R-NIEP for n ≤ 4 was solved independently by Perfect [99], Oliveira
[97, §9] and Loewy–London [77]; in these cases the necessary conditions (Trace)
and (Perron) are also sufficient.
206 C.R. Johnson, C. Marijuán, P. Paparella and M. Pisonero

Perfect [101, Theorem 4] solved the R-NIEP for n = 3 with fixed diagonal
entries: {d1 , d2 , d3 } is the diagonal of a 3-by-3 nonnegative matrix with spectra
{λ1 , λ2 , λ3 }, where λ1 ≥ λ2 ≥ λ3 , if and only if

3 
3  
0 ≤ di ≤ λ1 ; di = λi ; di dj ≥ λi λj ; max{di } ≥ λ2 .
i=1 i=1 i=j i=j

The S-NIEP and the R-NIEP are equivalent for n ≤ 4, and remain unsolved
for n ≥ 5. Fiedler [32, Theorem 4.8] solved the S-NIEP in the case n = 3 with fixed
diagonal entries: {d1 , d2 , d3 }, with d1 ≥ d2 ≥ d3 ≥ 0, is the diagonal of a 3-by-3
symmetric nonnegative matrix with spectra {λ1 , λ2 , λ3 }, where λ1 ≥ λ2 ≥ λ3 , if
and only if {λ1 , λ2 , λ3 } majorizes {d1 , d2 , d3 } and d1 ≥ λ2 .
Johnson–Laffey–Loewy [50] showed that the R-NIEP and the S-NIEP are
different, and Egleston–Lenker–Narayan [27] proved that they are different for
n ≥ 5.
The S-NIEP for n = 5 has been widely studied [27, 78, 88], but not fully
resolved. It is common to study it considering the number of positive eigenval-
ues. When there are 1, 4 or 5 positive eigenvalues the answer for the S-NIEP is
straightforward. Recently, Johnson–Marijuán–Pisonero [53] resolved all cases with
2 positive eigenvalues, and they give a method, based upon the eigenvalue inter-
lacing inequalities for symmetric matrices, to rule out many unresolved spectra
with 3 positive eigenvalues. In particular, this method shows that the nonnegative
realizable spectrum {6, 3, 3, −5, −5} is not symmetrically realizable.
Also recently, Loewy–Spector [79, Theorem 4] characterize the case n = 5 in a
particular case: Λ = {λ1 , λ2 , . . . , λ5 }, where λ1 ≥ λ2 ≥ · · · ≥ λ5 and 2s1 (Λ) ≥ λ1 ,
is (S-NIEP) realizable if and only if (Perron), λ2 + λ5 ≤ Tr (Λ), λ3 ≤ Tr (Λ). This
last condition is implied by the constraint 2s1 (Λ) ≥ λ1 .
The trace 0 NIEP has also been extensively studied. Reams [106] solved
the case n = 4: {λ1 , λ2 , λ3 , λ4 } is trace 0 (NIEP)-realizable if and only if s1 =
0, s2 , s3 ≥ 0 and s22 ≤ 4s4 . The case n = 5 was first studied by Reams [106] and
he gave a sufficient condition. The case n = 5 was finally solved by Laffey–Meehan
[69]: {λ1 , . . . , λ5 } is+
trace 0 (NIEP)-realizable if and only if s1 = 0, s2 , s3 ≥ 0, s22 ≤
4s4 and 12s5 + 5s3 4s4 − s22 ≥ 5s2 s3 .
Torre–Mayo et al. [130] generalize these solutions in terms of the coefficients
of the characteristic polynomial: the polynomial xn + kp xn−p + · · · + kn−1 x + kn ,
with 2 ≤ p ≤ n ≤ 2p + 1, is (NIEP)-realizable if and only if
k2
kp , . . . , k2p−1 ≤ 0; k2p ≤ 4p ;
⎧
⎨ kp kp+1 if k2p ≤ 0,
 , 2 
k2p+1 ≤
⎩ kp+1 k2p − k4p − k2p if k2p > 0.

Spector [127] characterized trace 0 S-NIEP realizability for n = 5 by the

conditions λ2 + λ5 ≤ 0 and s3 ≥ 0.
 The NIEP 207

5. Sufficient conditions
The first known sufficient condition for the NIEP, that in fact is for the R-NIEP,
was announced by Suleı̌manova [128] in 1949 and proved by Perfect [100] in 1953:
3
Λ = {λ1 , . . . , λn } real, λ1 ≥ |λ| for λ ∈ Λ
 =⇒ Λ is realizable.
and λ1 + λi <0 λi ≥ 0
Several other proofs have been given, e.g., [98]. There are several sufficient condi-
tions for the R-NIEP that are checkable in a straightforward way, that is, one only
needs to check a few algebraic inequalities, perhaps after ordering the spectrum.
The authors of them are: Ciarlet in 1968 [19], Kellogg in 1971 [58], Salzmann in
1972 [111] and Fiedler in 1974 [32]. Borobia in 1995 [6] extended Kellogg’s condi-
tion by grouping negative eigenvalues.
Other sufficient conditions for the R-NIEP involve partitions of the spectra
considered, such as an immediate piece-wise extension of the Suleı̌manova condi-
tion. We will name this condition Suleı̌manova–Perfect [100, 128]:
⎫
Λ = {λ1 , λ11 , . . . , λ1t1 , . . . , λr , λr1 , . . . , λrtr } real, ⎪
⎪
⎬
λ1 ≥ |λ| for λ ∈ Λ =⇒ Λ is realizable.

and λj + λji ≥ 0 for j = 1, . . . , r ⎪ ⎪
⎭
λji <0

Other sufficient conditions of this type are more elaborate and we will name them
by their authors: Perfect 1 in 1953 [100], Soto 2 in 2003 [114] or its extension Soto
p in 2013 [118].
Some of the sufficient conditions besides partitions involve the knowledge of
the diagonal entries of a realization of part of the spectrum. The first condition of
this type is due to Perfect in 1955 [101]:
⎫
Λ = {λ1 , . . . , λr } ⎪
⎪
∪{λ11 , . . . , λ1t1 } ∪ · · · ∪ {λr1 , . . . , λrtr } ⎪
⎪
⎪
⎪
⎪
{λ1 , . . . , λr } the spectrum of a nonnegative ⎪ ⎪
⎪
⎪
⎪
matrix with diagonal d1 , . . . , dr , ⎪ ⎬
⇒ Λ is realizable.
λji ≤ 0 for j = 1, . . . , r and i = 1, . . . , tj , ⎪ ⎪
 ⎪
⎪
λ1 ≥ |λ| for λ ∈ Λ , λ≥0, ⎪ ⎪
⎪
⎪
⎪

λ∈Λ ⎪
⎪
and dj + λji ≥ 0 , j = 1, . . . , r ⎪ ⎪
⎭
1≤i≤tj

When λj ≥ 0, for j = 1, . . . , r, we call this condition Perfect 2+ (see [85]). There

are two equivalent conditions that extend this condition: Soto–Rojo in 2006 [122]
and Soto–Rojo–Manzaneda in 2011 [123].
Other well-known sufficient conditions manipulate certain spectra to get a
new realizable spectrum: Guo in 1997 [40, Theorems 2.1 and 3.1] or C-realizability
[11] that we name the game condition [84].
208 C.R. Johnson, C. Marijuán, P. Paparella and M. Pisonero

In order to construct a map of sufficient conditions for the R-NIEP, Marijuán–

Pisonero–Soto compared these conditions and established inclusion relations or
independence relations between them, [84, 85]:

TSN  TS
P2+ [6] Borobia = B
[19] Ciarlet = C
Sp [32] Fiedler = F
K
[58] Kellogg = K
F [100] Perfect 1 = P 1
Sa B [101] Perfect 2+ = P 2+
[111] Salzmann = Sa
Su C [118] Soto p = Sp
SP P1 [122] Soto–Rojo = SR
[123] Soto–Rojo–Manzaneda = SRM
[128] Suleı̌manova = Su
[100] Suleı̌manova–Perfect = SP

Q 4,
icog
[6] Borobia = B
Uqvqu [101] Perfect 2+ = P 2+
A [118] Soto p = Sp
C Tq Tq,3
[118] Soto p + 1 = Sp+1
[118] Sotos = Soto p
p≥2

The first known sufficient condition for the S-NIEP is due to Perfect–Mirsky
in 1965 [102] for doubly stochastic matrices, and Fiedler in 1974 [32] gave the first
one for symmetric nonnegative matrices. Several sufficient conditions which were
first obtained for the R-NIEP have later been shown to be valid also for the S-
NIEP as well. Fiedler [32], Radwan [105] and Soto [116] showed, respectively, that
Kellogg [58], Borobia [6] and Soto 2 [114] are also symmetric sufficient conditions.
Soules in 1983 [126] gave two constructive sufficient conditions for symmetric
realization. The inequalities that appear in these conditions are obtained by re-
quiring the diagonal entries of the matrix Rdiag(λ1 , . . . , λn )RT to be nonnegative,
in which R is an orthogonal matrix with a certain pattern. For a particular R, this
condition is the Perfect–Mirsky condition.
Soto–Rojo–Moro–Borobia gave in 2007 [124] a symmetric version of the Soto–
Rojo condition. Soto p [118] and Soto–Rojo–Manzaneda [123] have also symmetric
 The NIEP 209

versions. Laffey–Šmigoc in 2007 [71] gave the symmetric realizability of a spectrum

by manipulating two spectra.
Again in order to construct a map of sufficient conditions for the S-NIEP,
Marijuán–Pisonero–Soto [86] compared these conditions and established inclusion
relations or independence relations between them:

[6] Borobia = B
[19] Ciarlet = C
[32] Fiedler = F
SRM  SRMB [58] Kellogg = K
Sotos [71] Laffey–Šmigoc = LS
[100] Perfect 1 = P 1
B K
Sp [102] Perfect–Mirsky = 
LS F [111] Salzmann = Sa
**
[118] Soto p = Sp
* Sa
[118] Sotos = Soto p
p≥2
Su C [123] Soto–Rojo–Manzaneda
SP = SRM
P1
[124] Soto–Rojo–Moro–Borobia
= SRM B
[126] Soules 1 = 
[126] Soules 2 = ∗∗
[126] Soules 2 corollary =∗
[128] Suleı̌manova = Su
[100] Suleı̌manova–Perfect = SP

The discontinuous line for Soules 2 in the map means that we only conjecture this
position for this sufficient condition.
Recently, Ellard–Šmigoc [29] have modified the Laffey–Šmigoc condition and
the Soules 2 condition and they have proved the following equivalence:

Soules 2 modified ⇐⇒ Laffey–Šmigoc modified ⇐⇒ game ⇐⇒ Sotos.

This implies the symmetric realizably of the game condition and resolves the ques-
tion on the RNIEP diagram about whether there exists something between Sotos
and game.
There are also some sufficient conditions for the NIEP. Guo in 1997 [40,
Theorems 2.1 and 3.1] gave results about changing a realizable spectrum to obtain
a realizable spectrum. Šmigoc in 2004 [112, Theorems 10 and 14] gave other such
results. Finally, let us mention some generalizations of the Suleı̌manova condition:
Laffey–Šmigoc in 2006 [70, Theorems 1 and 3] and Borobia–Moro–Soto in 2008
[11, Theorem 3.3].
210 C.R. Johnson, C. Marijuán, P. Paparella and M. Pisonero

6. Embedding spectra, by adding 0’s, to achieve realizability

Consider now spectra that meet the most basic necessary conditions for NIEP-
realizability: (Perron), (Trace) and (Reality). Even when these conditions √ are
strictly met, the proposed spectrum, σ, need not be realizable (e.g., {3,3,− 3 ± i}).
This raises the natural question of whether the spectrum may be embedded in
a larger one, that is realizable, by appending some additional eigenvalues, e.g.,
σ → σ ∪ τ . If we are too liberal about what eigenvalues may be appended, this
question becomes trivial. For example, realizability may always be achieved by ap-
pending a single, sufficiently large, positive eigenvalue; the Perron and trace condi-
tions may be arbitrarily improved. Thus, some condition must be placed upon the
appended eigenvalues. A natural one is that only 0 eigenvalues may be appended;
now, the Perron and trace conditions are not enhanced – but, the dimension is
increased. (Intermediate restrictions seem not yet to have been considered.) The
increase in dimension does improve the possibility of meeting the JLL conditions.
The notion of appending 0’s to “repair” a nonrealizable spectrum may and
has often been viewed another way: what collections of complex numbers occur as
the “nonzero part” of the spectrum of an entry-wise nonnegative matrix. It turns
out that this question is quite different from and more tractable than the classical
NIEP.
The first to show that appending 0’s can help was [48], in which it was shown
that the spectrum J J
3 3
1, i, − i
8 8
is not realizable in dimension 3 (because the single eigenvalue conditions [57] are
not met, or the JLL conditions are not met), but the spectrum
J J
3 3
1, i, − i, 0
8 8
is realizable in dimension 4 by the matrix
⎛ ⎞
1/4 0 3/4 0
⎜ 1/4 1/4 0 1/2 ⎟
⎜ ⎟.
⎝ 0 3/4 1/4 0 ⎠
7/24 0 11/24 1/4
Later in [12], it was shown, remarkably, that if a spectrum (with no 0’s)
meets the basic necessary conditions strictly, then it is the nonzero part of the
spectrum of a nonnegative matrix. Sufficiently many 0’s may be appended to
achieve realizability. Of course, the number of 0’s that need be added may be very
large (not uniformly bounded in terms of n) because of JLL. Interestingly, the
newish method of symbolic dynamics was used in an import way, though many
easily proven matricial lemmas were needed as well. Specifically the result is
Theorem 1. The list of nonzero complex numbers Λ = {λ1 , λ2 , . . . , λn } is the
nonzero spectrum of a primitive matrix if and only if
 The NIEP 211

1. strict Perron condition: Λ contains a positive eigenvalue of multiplicity one

that is greater in absolute value than all other λi s;

n
2. reality condition: p(t) = (t − λi ) has real coefficients;
i=1
and

n 
n
3. extended trace condition: λmk
i ≥ 0 for all k and λki > 0 implies
i=1 i=1


n
λmk
i > 0 for all m.
i=1

The necessity of these conditions is easily verified (the last one via JLL, for
example), and the interesting point is that necessary conditions become sufficient
when primitivity is the goal and the dimension may be arbitrarily increased. When
the Perron condition is not strict it may not be possible to “save” a spectrum
meeting obvious necessary conditions. The familiar example 3, 3, −2, −2, −2 is not
only not realizable, but is never the nonzero part of the spectrum of a nonnegative
matrix. However, both 3 + , 3, −2, −2, −2 and 3 + , 3 − , −2, −2, −2 are both
the nonzero parts of the spectra of primitive matrices for arbitrarily small  > 0
(though they are not realizable for  small).
More recently there have been further developments about realizability after
appending 0’s to a spectrum. In [65] there is matricial proof of key results from [12],
which is much more explicit. Though the number of 0’s needed to make a spectrum
realizable may be very large (and not easy to estimate from [12]), estimates have
recently been given under some circumstances by bringing s2 (Λ), as well as the
trace, into play [70].

7. The graph NIEP

Not surprisingly, realizable spectra that are, in some way extremal, are often re-
alizable by nonnegative matrices with many 0 entries, as are many non-extremal
spectra. This raises the question that, if we fix the 0-pattern of a nonnegative
matrix, how is the NIEP restricted, i.e., which realizable spectra occur? A natural
way to describe a particular 0-pattern is via a graph, which could be directed or
undirected. For a particular graph G on n vertices, consider the set of nonnegative
n-by-n matrices N (G) for which A = (aij ) satisfies aij > 0 if and only if (i, j)
is a directed edge, i = j, of G. (Here, we consider an undirected edge {i, j}, if
G is undirected, to consist of two directed edges (i, j) and (j, i).) For simplicity
we consider graphs without loops, and no restriction is placed by the graph upon
the diagonal entries, other than nonnegativity. The G-NIEP then just asks which
spectra occur among matrices in N (G)? (if we wish to emphasize the dimension,
which we generally take to be implicit, we may write Nn (G).) Of course the NIEP
is just the union of the solutions to the G-NIEP’s, over all directed graphs G.
The same is true for various variations upon the NIEP. For example, the solution
212 C.R. Johnson, C. Marijuán, P. Paparella and M. Pisonero

to the S-NIEP is just the solution to the G-NIEP, restricted to symmetric matri-
ces, over all undirected graphs. Thus far, the G-NIEP has been considered only
for undirected graphs G, symmetric matrices and real eigenvalues. This seems a
fertile area for future work.
A prototype of the G-NIEP, though not presented in graph terms, appeared
in [34], in which tridiagonal matrices were considered. Notice that the spectrum of
a nonnegative tridiagonal matrix is not only necessarily real but also the spectrum
of a symmetric tridiagonal nonnegative matrix. Also, tridiagonal matrices are the
case in which G is a path (and an edge only requires nonnegativity of the entry);
off-diagonal 0 entries are important. We say that a matrix A is subordinate to a
graph G if G(A) has the same vertices as G and the edges of G(A) are contained
among those of G; equivalently, for A = (aij ), aij = 0 implies {i, j} is an edge of
G. The main result of [34] is for the path P on n vertices. Then, for λ1 ≥ λ2 ≥
· · · ≥ λn , there is an n-by-n nonnegative matrix AT = A subordinate to P and
with eigenvalues λ1 , λ2 , . . . , λn if and only if
λi + λn−i+1 ≥ 0
i = 1, 2, . . . , n. Additional conditions, when the graph is precisely P , are also
discussed, but the results are incomplete.
In [48, 74], the observation of [34] is dramatically generalized. Recall that a
path is a tree and that all trees are bipartite (but many non-trees are bipartite as
well). All bipartite graphs are considered in [74]. The main result is that λ1 ≥ · · · ≥
λn are the eigenvalues of an n-by-n nonnegative symmetric matrix A, subordinate
to a given bipartite graph G on n vertices, if and only if
λ1 + λn ≥ 0
λ2 + λn−1 ≥ 0
..
.
λm + λn−m+1 ≥ 0
λm+1 ≥ 0
..
.
λn−m ≥ 0
in which m is the matching number of G. (Note that % n2 & is the matching number of
a path on n vertices, so that the result of [34] is a special case.) Further observations
about the S-NIEP for matrices subordinate to a given graph G are also made.

8. Perron similarities
If S is an invertible matrix and there is a real, diagonal, nonscalar matrix D
with SDS −1 ≥ 0, then S is called a Perron similarity. Perron similarities were
 The NIEP 213

introduced by Johnson and Paparella to study the diagonalizable R-NIEP and the
S-NIEP.
Perron similarities were characterized in several ways in [54] and it was shown
that C(S) := {x ∈ Rn : SDx S −1 ≥ 0} is a polyhedral cone, i.e., a convex cone in
Rn with finitely-many extremals. This was called the (Perron) spectracone of S,
and a certain cross-section, a polytope called the (Perron) spectratope of S, was
also discussed.
These polyhedral sets were used to verify the known necessary and sufficient
conditions for the R-NIEP and the S-NIEP for orders up to four. For orders 1, 2,
and 4, it is shown that a finite number of Perron similarities are required to cover
the realizable region, whereas when n = 3, it is shown, via the relative gain array
(see, e.g., [44]), that an uncountable number of Perron similarities is required to
cover the realizable region.
For every n ≥ 1, the spectracone and spectratope of Hn were characterized,
where Hn denotes the canonical Hadamard matrix of order 2n. More specifically,
the spectracone of Hn is the conical hull of its rows and the spectratope of Hn is the
convex hull of its rows. The spectratope of a general, normalized Hadamard matrix
was used to give a constructive proof, for Hadamard orders, of a result by Fiedler
[32], that every Suleı̌manova spectrum is the spectrum of a symmetric nonnegative
matrix (in fact, this result was strengthened to show that the constructed realizing
matrix is also doubly stochastic). It is still an open problem to find a constructive
proof that every Suleı̌manova spectrum is SNIEP-realizable. A constructive proof
of the Boyle–Handelman theorem for Suleı̌manova spectrum: augmenting any such
spectrum with zeros up to a Hadamard order yields a spectrum realizable by a
nonnegative matrix that is symmetric and doubly stochastic.

9. Jordan structure and the NIEP’s

When there are repeated eigenvalues in a proposed spectrum, a natural question
is whether Jordan structure for the repeated eigenvalues can play a role in re-
alizability. The J-NIEP asks which particular Jordan canonical forms occur for
n-by-n nonnegative matrices and the D-NIEP is the special case of which spectra
occur among diagonalizable n-by-n nonnegative matrices. Of course, the D-NIEP
and the NIEP are the same for spectra with distinct eigenvalues. It is also a sim-
ple exercise that any spectrum that is D-NIEP realizable by a positive matrix is
also J-NIEP realizable for any Jordan canonical form possible for its eigenvalues.
However, whether D-NIEP realizability always implies J-NIEP realizability for any
possible Jordan form is unclear. This is so for Suleı̌manova spectra [23, 120, 121].
There are spectra that are realizable (in fact, R-NIEP realizable), but are
not diagonalizably realizable. The smallest dimension in which this occurs is 5.
For n ≤ 4, any realizable spectrum with (non-real) complex eigenvalues is diag-
onalizably realizable (the only difficulty could come from a multiple Perron root
and that is easily handled by a reducibility argument). For n ≤ 4, R-NIEP and
214 C.R. Johnson, C. Marijuán, P. Paparella and M. Pisonero

S-NIEP realizability are the same, which settles the matter. However, for n = 5,
the spectrum
3 + t, 3 − t, −2, −2, −2
√ 1/2
is realizable for t > (16 6) − 39 ≈ 0.437 . . ., [69]. However, it is diagonalizably
realizable iff t ≥ 1 (in which case it is also symmetrically realizable). Thus, for
0.437 . . . < t < 1, this 5-spectrum is realizable, but not diagonalizable so. This
suggests that this phenomenon is fairly common.

References
[1] R.B. Bapat and T.E.S. Raghavan, Nonnegative Matrices and Applications. Cam-
bridge University Press, Cambridge, 1997.
[2] L. Benvenuti, A geometrical representation of the spectra of four dimensional non-
negative matrices. Linear Algebra Appl. 445 (2014), 162–180.
[3] L. Benvenuti, A note on eigenvalues location for trace zero doubly stochastic ma-
trices. Electron. J. Linear Algebra 30 (2015), 599–604.
[4] A. Berman and R.J. Plemmons, Nonnegative Matrices in the Mathematical Sci-
ences. SIAM Publications, Philadelphia, 1994.
[5] G. Bharali and O. Holtz, Functions preserving nonnegativity of matrices. SIAM J.
Matrix Anal. Appl. 30 (2008), 84–101.
[6] A. Borobia, On the nonnegative eigenvalue problem. Linear Algebra Appl. 223–224
(1995), 131–140.
[7] A. Borobia, Inverse eigenvalue problems. In: “Handbook of Linear Algebra”, chapter
20. 20-1–20-14, Chapman & Hall/CRC, USA 2007.
[8] A. Borobia and J. Moro, On non-negative matrices similar to positive matrices.
Linear Algebra Appl. 266 (1997) 365–379.
[9] A. Borobia and J. Moro, On the boundary of the set of real spectra of non-negative
matrices. Linear Algebra Appl. 278 (1998), 287–293.
[10] A. Borobia, J. Moro, and R.L. Soto, Negativity compensation in the nonnegative
inverse eigenvalue problem. Linear Algebra Appl. 393 (2004), 73–89.
[11] A. Borobia, J. Moro, and R.L. Soto, A unified view on compensation criteria in
the real nonnegative inverse eigenvalue problem. Linear Algebra Appl. 428 (2008),
2574–2584.
[12] M. Boyle and D. Handelman, The spectra of nonnegative matrices via symbolic
dynamics. Ann. Math. (2) 133 (1991), 249–316.
[13] A. Brauer, Limits for the characteristic roots of a matrix, IV: Applications to sto-
chastic matrices. Duke Math. J. 19 (1952), 75–91.
[14] R.A. Brualdi, Spectra of digraphs. Linear Algebra Appl. 432 (2010), 2181–2213.
[15] J. Ccapa and R.L. Soto, On spectra perturbation and elementary divisors of positive
matrices. Electron. J. Linear Algebra 18 (2009), 462–481.
[16] J. Ccapa and R.L. Soto, On elementary divisors perturbation of nonnegative ma-
trices. Linear Algebra Appl. 432 (2010), 546–555.
 The NIEP 215

[17] M.T. Chu and K.R. Driessel, Constructing symmetric nonnegative matrices with
prescribed eigenvalues by differential equations. SIAM J. Math. Anal. 22 (1991),
1372–1387.
[18] M.T. Chu and G.H. Golub, Inverse Eigenvalue Problems: Theory, Algorithms and
Applications. Oxford University Press, New York, 2005.
[19] P.G. Ciarlet, Some results in the theory of nonnegative matrices. Linear Algebra
Appl. 1 (1968), 139–152.
[20] A.G. Cronin, Characterizing the spectra of nonnegative matrices. PhD thesis, School
of Mathematical Sciences, University College Dublin, 2012.
[21] A.G. Cronin and T.J. Laffey, An inequality for the spectra of nonnegative matrices.
Linear Algebra Appl. 436 (2012), 3225–3238.
[22] D.M. Cvetković, M. Doob, and H. Sachs, Spectra of Graphs. Johann Ambrosius
Barth, Heidelberg, 1995.
[23] R.C. Dı́az and R.L. Soto, Nonnegative inverse elementary divisors problem in the
left half plane. Linear Multilinear Algebra 64 (2016), 258–268.
[24] N. Dmitriev and E. Dynkin, On the characteristic numbers of a stochastic matrix.
Dokl. Akad. Nauk SSSR 49 (1945), 159–162 (in Russian).
[25] N. Dmitriev and E. Dynkin, Eigenvalues of a stochastic matrix. Izv. Akad. Nauk
SSSR Ser. Mat. 10 (1946), 167–184 (in Russian).
[26] D.Ž. Dokovič, Cyclic polygons, roots of polynomials with decreasing nonnegative
coefficients, and eigenvalues of stochastic matrices. Linear Algebra Appl. 142 (1990),
173–193.
[27] P.D. Egleston, T.D. Lenker, and S.K. Narayan, The nonnegative inverse eigenvalue
problem. Linear Algebra Appl. 379 (2004), 475–490.
[28] R. Ellard and H. Šmigoc, Constructing new realisable lists from old in the NIEP.
Linear Algebra Appl. 440 (2014), 218–232.
[29] R. Ellard and H. Šmigoc, Connecting sufficient conditions for the symmetric non-
negative inverse eigenvalue problem. Linear Algebra Appl. 498 (2016), 521–552.
[30] L. Elsner, R. Nabben, and M. Neumann, Orthogonal bases which lead to symmetric
nonnegative basis. Linear Algebra Appl. 271 (1998), 323–343.
[31] G.M. Engel, H. Schneider, and S. Sergeev, On sets of eigenvalues of matrices with
prescribed row sums and prescribed graph. Linear Algebra Appl. 455 (2014), 187–
209.
[32] M. Fiedler, Eigenvalues of nonnegative symmetric matrices. Linear Algebra Appl.
9 (1974), 119–142.
[33] S. Friedland, On an inverse problem for nonnegative and eventually nonnegative
matrices. Israel J. Math. 29 (1978), 43–60.
[34] S. Friedland and A.A. Melkman, On the eigenvalues of nonnegative Jacobi matrices.
Linear Algebra Appl. 25 (1979), 239–253.
[35] G. Frobenius, Über Matrizen aus positiven Elementen. S.-B. K. Preuss. Akad. Wiss.
Berlin, 1908 p. 471–476, 1909 p. 514–518.
[36] G. Frobenius, Über Matrizen aus nicht negativen Elementen. S.-B. K. Preuss. Akad.
Wiss. Berlin, 1912 p. 456–477.
216 C.R. Johnson, C. Marijuán, P. Paparella and M. Pisonero

[37] R. Gómez and J.M. Salazar, Spanning tree invariants, loop systems and doubly
stochastic matrices. Linear Algebra Appl. 432 (2010), 556–565.
[38] S. Guo and W. Guo, Perturbing non-real eigenvalues of nonnegative matrices. Lin-
ear Algebra Appl. 426 (2007), 199–203.
[39] W. Guo, An inverse eigenvalue problem for nonnegative matrices. Linear Algebra
Appl. 249 (1996), 67–78.
[40] W. Guo, Eigenvalues of nonnegative matrices. Linear Algebra Appl. 266 (1997),
261–270.
[41] D. Hershkowits, Existence of Matrices Satisfying Prescribed Conditions. M. S. The-
sis, Technion-Israel Inst. of Technology, Haifa, Israel, 1978.
[42] O. Holtz, M-matrices satisfy Newton’s inequalities Proc. Amer. Math. Soc. 133
(2005), 711–717.
[43] R.A. Horn and C.R. Johnson, Matrix Analysis. Cambridge University Press, Cam-
bridge 1990. Corrected reprint of the 1985 original.
[44] R.A. Horn and C.R. Johnson, Topics in Matrix Analysis. Cambridge University
Press, Cambridge 1991.
[45] S.-G. Hwang and S.-S. Pyo, The inverse eigenvalue problem for symmetric doubly
stochastic matrices. Linear Algebra Appl. 379 (2004), 77–83.
[46] H. Ito, A new statement about the theorem determining the region of eigenvalues of
stochastic matrices. Linear Algebra Appl. 267 (1997), 241–246.
[47] C.R. Johnson, An inclusion for the field of values of a doubly stochastic matrix
based on its graph. Aequationes Mathematicae 17 (1978), 305–310.
[48] C.R. Johnson, Row stochastic matrices similar to doubly stochastic matrices. Linear
Multilinear Algebra 10 (1981), 113–130.
[49] C.R. Johnson, R.B. Kellogg, and A.B. Stephens, Complex eigenvalues of a nonnega-
tive matrix with a specified graph. II. Linear Multilinear Algebra 7 (1979), 129–143.
[50] C.R. Johnson, T.J. Laffey, and R. Loewy, The real and the symmetric nonnegative
inverse eigenvalue problems are different. Proc. Amer. Math. Soc. 124 (1996), 3647–
3651.
[51] C.R. Johnson, C. Marijuán, and M. Pisonero, Matrices and spectra satisfying the
Newton inequalities. Linear Algebra Appl. 430 (2009), 3030–3046.
[52] C.R. Johnson, C. Marijuán, and M. Pisonero, Symmetric nonnegative realizability
via partitioned majorization. Linear Multilinear Algebra 65 (2017), 1417–1426.
[53] C.R. Johnson, C. Marijuán, and M. Pisonero, Ruling out certain 5-spectra for
the symmetric nonnegative inverse eigenvalue problem. Linear Algebra Appl. 512
(2017), 129–135.
[54] C.R. Johnson and P. Paparella, Perron spectratopes and the real nonnegative inverse
eigenvalue problem. Linear Algebra Appl. 493 (2016), 281–300.
[55] C.R. Johnson and P. Paparella, A matricial view of the Karpelevič theorem. Linear
Algebra Appl. 520 (2017), 1–15.
[56] A.I. Julio and R.L. Soto, Persymmetric and bisymmetric nonnegative inverse eigen-
value problem. Linear Algebra Appl. 469 (2015), 130–152.
[57] F. Karpelevich, On the eigenvalues of a matrix with nonnegative elements. Izv.
Akad. Nauk SSR Ser. Mat. 15 (1951), 361–383 (in Russian).
 The NIEP 217

[58] R. Kellogg, Matrices similar to a positive or essentially positive matrix. Linear

Algebra Appl. 4 (1971), 191–204.
[59] R.B. Kellogg and A.B. Stephens, Complex eigenvalues of a non-negative matrix
with a specified graph. Linear Algebra Appl. 20 (1978), 179–187.
[60] K.H. Kim, N.S. Ormes, and F.W. Roush, The spectra of nonnegative integer ma-
trices via formal power series. J. Amer. Math. Soc. 13 (2000), 773–806.
[61] J. Knudsen and J.J. McDonald, A note on the convexity of the realizable set of eigen-
values for nonnegative symmetric matrices. Electron. J. Linear Algebra 8 (2001),
110–114.
[62] A.N. Kolmogorov, Markov chains with a countable number of possible states. Bull.
Moskow Gos. Univ. Mat. Meh. 1 (3)(1937), 1–16.
[63] T.J. Laffey, Realizing matrices in the nonnegative inverse eigenvalue problem. Ma-
trices and group representations (Coimbra, 1998), vol. 19 of Textos Mat. Sér. B,
21–32. Univ. Coimbra, Coimbra, 1999.
[64] T.J. Laffey, Perturbing non-real eigenvalues of nonnegative real matrices. Electron.
J. Linear Algebra 12 (2005), 73–76.
[65] T.J. Laffey, A constructive version of the Boyle–Handelman theorem on the spectra
of nonnegative matrices. Linear Algebra Appl. 436 (2012), 1701–1709.
[66] T.J. Laffey, R. Loewy, and H. Šmigoc, Nonnegative matrices that are similar to
positive matrices. SIAM J. Matrix Anal. Appl. 31 (2009), 629–649.
[67] T.J. Laffey, R. Loewy, and H. Šmigoc, Power series with positive coefficients arising
from the characteristic polynomials of positive matrices. Math. Ann. 364 (2016),
687–707.
[68] T.J. Laffey and E. Meehan, A refinament of an inequality of Johnson, Loewy and
London on nonnegative matrices and some applications. Electron. J. Linear Algebra
3 (1998), 119–128.
[69] T.J. Laffey and E. Meehan, A characterization of trace zero nonnegative 5 × 5
matrices. Linear Algebra Appl. 302/303 (1999), 295–302.
[70] T.J. Laffey and H. Šmigoc, Nonnegative realization of spectra having negative real
parts. Linear Algebra Appl. 416 (2006), 148–159.
[71] T.J. Laffey and H. Šmigoc, Construction of nonnegative symmetric matrices with
given spectrum. Linear Algebra Appl. 421 (2007), 97–109.
[72] T.J. Laffey and H. Šmigoc, On a classic example in the nonnegative inverse eigen-
value problem. Electron. J. Linear Algebra. 17 (2008), 333–342.
[73] T.J. Laffey and H. Šmigoc, Nonnegatively realizable spectra with two positive eigen-
values. Linear Multilinear Algebra 58 (2010), 1053–1069.
[74] A. Leal-Duarte and C.R. Johnson, Resolution of the symmetric nonnegative inverse
eigenvalue problem for matrices subordinate to a bipartite graph. Positivity 8 (2004),
209–213.
[75] J. Levick, R. Pereira, and D.W. Kribs, The four-dimensional Perfect–Mirsky con-
jecture. Proc. Amer. Math. Soc. 143 (2014), 1951–1956.
[76] Y.J. Ley, W.R. Xu, Y. Lu, and Y.R. Niu, On the symmetric doubly stochastic
inevrese eigenvalue problem. Linear Algebra Appl. 445 (2014), 181–205.
218 C.R. Johnson, C. Marijuán, P. Paparella and M. Pisonero

[77] R. Loewy and D. London, A note on the inverse problem for nonnegative matrices.
Linear Multilinear Algebra 6 (1978), 83–90.
[78] R. Loewy and J.J. McDonald, The symmetric nonnegative inverse eigenvalue prob-
lem for 5 × 5 matrices. Linear Algebra Appl. 393 (2004), 275–298.
[79] R. Loewy and O. Spector, A necessary condition for the spectrum of nonnegative
symmetric 5 × 5 matrices. Linear Algebra Appl. 528 (2017), 206–272.
[80] C. Maclaurin, A second letter to Martin Folkes, Esq.; concerning the roots of eqnar-
rays, with the demonstration of other rules in algebra. Phil. Transactions 36 (1729),
59–96.
[81] M. Marcus and H. Minc, A Survey of Matrix Theory and Matrix Inequalities. Allyn
and Bacon, Boston, 1964.
[82] C. Marijuán and M. Pisonero, Relaciones entre condiciones suficientes en el prob-
lema espectral real inverso no negativo. Proceedings V Jornadas de Matemática
Discreta y Algorı́tmica (2006), 335–342.
[83] C. Marijuán and M. Pisonero, Polinomios caracterı́sticos de digrafos pesados y
desigualdades de Newton. Proceedings VII Jornadas de Matemática Discreta y Al-
gorı́tmica (2010), 447–458.
[84] C. Marijuán and M. Pisonero, On sufficient conditions for the RNIEP and their
margins of realizability. Electronic Notes in Discrete Mathematics 46 (2014), 201–
208.
[85] C. Marijuán, M. Pisonero, and R.L. Soto, A map of sufficient conditions for the real
nonnegative inverse eigenvalue problem. Linear Algebra Appl. 426 (2007), 690–705.
[86] C. Marijuán, M. Pisonero, and R.L. Soto, A map of sufficient conditions for the sym-
metric nonnegative inverse eigenvalue problem. Linear Algebra Appl. 530 (2017),
344–365.
[87] J. Mashreghi and R. Rivard, On a conjecture about the eigenvalues of doubly sto-
chastic matrices. Linear Multilinear Algebra 55 (2007), 491–498.
[88] J.J. McDonald and M. Neumann, The Soules approach to the inverse eigenvalue
problem for nonnegative symmetric matrices of order n ≤ 5. Contemp. Math. 259
(2000), 387–407.
[89] E. Meehan, Some Results On Matrix Spectra. PhD thesis, National University of
Ireland, Dublin, 1998.
[90] H. Minc, Inverse elememtary divisor problem for nonnegative matrices. Proc. Amer.
Math. Soc. 83 (1981), 665–669.
[91] H. Minc, Inverse elememtary divisor problem for doubly stochastic matrices. Linear
Multilinear Algebra 11 (1982), 121–131.
[92] H. Minc, Non-Negative Matrices. John Wiley and Sons, New York, 1988.
[93] L. Mirsky, Results and problems in the theory of doubly-stochastic matrices. Z.
Wahrscheinlichkeitstheorie 1 (1963), 319–334.
[94] B. Mourad, An inverse problem for n × n symmetric doubly stochastic matrices.
Inverse Problems 19 (2003), 821–831.
[95] B. Mourad, A note on the boundary of the set where the decreasingly ordered spectra
of symmetric doubly stochastic matrices lie. Linear Algebra Appl. 416 (2006), 546–
558.
 The NIEP 219

[96] I. Newton, Arithmetica universalis: sive de compositione et resolutione arithmetica

liber. 1707.
[97] G.N. de Oliveira, Sobre matrizes estocásticas e duplamente estocásticas. PhD thesis,
Coimbra, 1968.
[98] P. Paparella, Realizing Suleimanova-type spectra via permutative matrices. Electron.
J. Linear Algebra 31 (2016), 306–312.
[99] H. Perfect, On positive stochastic matrices with real characteristic roots. Proc. Cam-
bridge Phil. Soc. 48 (1952), 271–276.
[100] H. Perfect, Methods of constructing certain stochastic matrices. Duke Math. J. 20
(1953), 395–404.
[101] H. Perfect, Methods of constructing certain stochastic matrices II. Duke Math. J.
22 (1955), 305–311.
[102] H. Perfect and L. Mirsky, Spectral properties of doubly stochastic matrices. Monatsh.
Math. 69 (1965), 35–57.
[103] H. Perfect and L. Mirsky, The distribution of positive elements in doubly-stochastic
matrices. J. London Math. Soc. 40 (1965), 689–698.
[104] O. Perron, Zur Theorie der Matrizen. Math. Ann. 64 (1907), 248–263.
[105] N. Radwan, An inverse eigenvalue problem for symmetric and normal matrices.
Linear Algebra Appl. 248 (1996), 101–109.
[106] R. Reams, An inequality for nonnegative matrices and inverse eigenvalue problem.
Linear Multilinear Algebra 41 (1996), 367–375.
[107] R. Reams, Construction of trace zero symmetric stochastic matrices for the inverse
eigenvalue problem. Electron. J. Linear Algebra 9 (2002), 270–275.
[108] O. Rojo and R.L. Soto, Existence and construction of nonnegative matrices with
complex spectrum. Linear Algebra Appl. 368 (2003), 53–69.
[109] O. Rojo and R.L. Soto, Guo perturbations for symmetric nonnegative circulant
matrices. Linear Algebra Appl. 431 (2009), 594–607.
[110] V. Romanovsky, Recherches sur les chaı̂nes de Markoff (Premier Mémoire). Acta
Math. 66 (1936), 147–251.
[111] F. Salzmann, A note on the eigenvalues of nonnegative matrices. Linear Algebra
Appl. 5 (1972), 329–338.
[112] H. Šmigoc, The inverse eigenvalue problem for nonnegative matrices. Linear Alge-
bra Appl. 393 (2004), 365–374.
[113] H. Šmigoc, Construction of nonnegative matrices and the inverse eigenvalue prob-
lem. Linear Multilinear Algebra 53 (2005), 85–96.
[114] R.L. Soto, Existence and construction of nonnegative matrices with prescribed spec-
trum. Linear Algebra Appl. 369 (2003), 169–184.
[115] R.L. Soto, Realizability by symmetric nonnegative matrices. Proyecciones 24 (2005),
65–78.
[116] R.L. Soto, Realizability criterion for the symmetric nonnegative inverse eigenvalue
problem. Linear Algebra Appl. 416 (2006), 783–794.
[117] R.L. Soto, Nonnegative Inverse Eigenvalue Problem, Linear Algebra – Theo-
rems and Applications. Prof. Hassan Yasser (ed.), InTech, DOI: 10.5772/48279.
220 C.R. Johnson, C. Marijuán, P. Paparella and M. Pisonero

Available from: http://www.intechopen.com/books/linear-algebra-theorems-and-

applications/nonnegative-inverse-eigenvalue-problem, 2012.
[118] R.L. Soto, A family of realizability criteria for the real and symmetric nonnegative
inverse eigenvalue problem. Numer. Linear Algebra Appl. 20 (2013), 336–348.
[119] R.L. Soto, A. Borobia, and J. Moro, On the comparison of some realizability crite-
ria for the real nonnegative inverse eigenvalue problem. Linear Algebra Appl. 396
(2005), 223–241.
[120] R.L. Soto and J. Ccapa, Nonnegative matrices with prescribed elementary divisors.
Electron. J. Linear Algebra 17 (2008), 287–303.
[121] R.L. Soto, R.C. Dı́az, H. Nina, and M. Salas, Nonnegative matrices with prescribed
spectrum and elementary divisors. Linear Algebra Appl. 439 (2013), 3591–3604.
[122] R.L. Soto and O. Rojo, Applications of a Brauer theorem in the nonnegative inverse
eigenvalue problem. Linear Algebra Appl. 416 (2006), 844–856.
[123] R.L. Soto, O. Rojo, and C.B. Manzaneda, On nonnegative realization of partitioned
spectra Electron. J. Linear Algebra 22 (2011), 557–572.
[124] R.L. Soto, O. Rojo, J. Moro, and A. Borobia, Symmetric nonnegative realization of
spectra. Electron. J. Linear Algebra 16 (2007), 1–18.
[125] R.L. Soto, M. Salas, and C.B. Manzaneda, On nonnegative realization of complex
spectra. Electron. J. Linear Algebra 20 (2010), 595–609.
[126] G. Soules, Constructing symmetric nonnegative matrices. Linear Multilinear Alge-
bra 13 (1983), 241–251.
[127] O. Spector, A characterization of trace zero symmetric nonnegative 5x5 matrices.
Linear Algebra Appl. 434 (2011), 1000–1017.
[128] H.R. Suleı̌manova, Stochastic matrices with real characteristic values. Dokl. Akad.
Nauk SSSR. 66 (1949), 343–345 (in Russian).
[129] J. Swift, The Location of Characteristic Roots of Stochastic Matrices. M.Sc. thesis,
McGill University, Montreal, 1972.
[130] J. Torre–Mayo, M.R. Abril-Raymundo, E. Alarcia-Estévez, C. Marijuán, and M.
Pisonero, The nonnegative inverse eigenvalue problem from the coefficients of the
characteristic polynomial. EBL digraphs. Linear Algebra Appl. 426 (2007), 729–773.

Charles R. Johnson Pietro Paparella

Department of Mathematics Div. of Engineering & Mathematics
College of William and Mary University of Washington Bothell
Williamsburg, VA 23187-8795, USA Bothell, WA 98011-8246, USA
e-mail: [email protected] e-mail: [email protected]

Carlos Marijuán Miriam Pisonero

Dpto. Matemática Aplicada Dpto. Matemática Aplicada
E.I. Informática E.T.S. de Arquitectura
Paseo de Belén 15 Avenida de Salamanca 18
E-47011 Valladolid, Spain E-47014 Valladolid, Spain
e-mail: [email protected] e-mail: [email protected]
Operator Theory:
Advances and Applications, Vol. 267, 221–246

c Springer International Publishing AG, part of Springer Nature 2018

Semi-Fredholmness of Weighted
Singular Integral Operators with
Shifts and Slowly Oscillating Data
Alexei Yu. Karlovich, Yuri I. Karlovich and Amarino B. Lebre

Abstract. Let α, β be orientation-preserving homeomorphisms of [0, ∞] onto

itself, which have only two fixed points at 0 and ∞, and whose restrictions
to R+ = (0, ∞) are diffeomorphisms, and let Uα , Uβ be the corresponding
isometric shift operators on the space Lp (R+ ) given by Uμ f = (μ )1/p (f ◦μ) for
μ ∈ {α, β}. We prove sufficient conditions for the right and left Fredholmness
on Lp (R+ ) of singular integral operators of the form A+ Pγ+ + A− Pγ− , where
Pγ± = (I ± Sγ )/2, Sγ is a weighted Cauchy singular integral operator, A+ =
 k
 k
k∈Z ak Uα and A− = k∈Z bk Uβ are operators in the Wiener algebras of
functional operators with shifts. We assume that the coefficients ak , bk for
k ∈ Z and the derivatives of the shifts α , β  are bounded continuous functions
on R+ which may have slowly oscillating discontinuities at 0 and ∞.
Mathematics Subject Classification (2010). 45E05, 47A53, 47G10, 47G30.
Keywords. Right Fredholmness, left Fredholmness, slowly oscillating shift,
Wiener algebra of functional operators, weighted singular integral operator,
Mellin pseudodifferential operator.

1. Introduction
Let B(X) denote the Banach algebra of all bounded linear operators acting on a
Banach space X. Recall that an operator A ∈ B(X) is said to be left invertible
(resp. right invertible) if there exists an operator B ∈ B(X) such that BA = I
(resp. AB = I) where I ∈ B(X) is the identity operator on X. The operator B
is called a left (resp. right) inverse of A. An operator A ∈ B(X) is said to be

This work was partially supported by the Fundação para a Ciência e a Tecnologia (Portuguese
Foundation for Science and Technology) through the projects UID/MAT/00297/2013 (Centro de
Matemática e Aplicações) and UID/MAT/04721/2013 (Centro de Análise Funcional, Estruturas
Lineares e Aplicações). The second author was also supported by the SEP-CONACYT Project
No. 168104 (México).
222 A.Yu. Karlovich, Yu.I. Karlovich and A.B. Lebre

invertible if it is left invertible and right invertible simultaneously. We say that A

is strictly left (resp. right) invertible if it is left (resp. right) invertible, but not
invertible. If the operator A is invertible only from one side, then the corresponding
inverse is not uniquely defined. We refer to [10, Section 2.5] for further properties
of one-sided invertible operators acting on Banach spaces.
Let K(X) be the closed two-sided ideal of all compact operators in B(X), and
let B π (X) := B(X)/K(X) be the Calkin algebra of the cosets Aπ := A + K(X)
where A ∈ B(X). Following [4, Chap. XI, Definition 2.3], an operator A ∈ B(X)
is said to be left Fredholm (resp. right Fredholm) if the coset Aπ is left invertible
(resp. right invertible) in the Calkin algebra B π (X). An operator A ∈ B(X) is
said to be semi-Fredholm if it is left or right Fredholm. We will write A  B if
A − B ∈ K(X).
Let Cb (R+ ) denote the C ∗ -algebra of all bounded continuous functions on
the positive half-line R+ := (0, +∞). Following Sarason [31, p. 820], a function
f ∈ Cb (R+ ) is called slowly oscillating (at 0 and ∞) if
lim sup |f (t) − f (τ )| = 0 for s ∈ {0, ∞}.
r→s t,τ ∈[r,2r]

The set SO(R+ ) of all slowly oscillating functions is a C ∗ -algebra. This algebra
properly contains C(R+ ), the C ∗ -algebra of all continuous functions on the two-
point compactification R+ := [0, +∞] of R+ .
Suppose α is an orientation-preserving homeomorphism of R+ onto itself,
which has only two fixed points 0 and ∞, and whose restriction to R+ is a dif-
feomorphism. We say that α is a slowly oscillating shift if log α is bounded and
α ∈ SO(R+ ). The set of all slowly oscillating shifts is denoted by SOS(R+ ). By
[15, Lemma 2.2], an orientation-preserving diffeomorphism α of R+ onto itself be-
longs to SOS(R+ ) if and only if α(t) = teω(t) for t ∈ R+ and a real-valued function
ω ∈ SO(R+ ) ∩ C 1 (R+ ) is such that the function ψ given by ψ(t) := tω  (t) also
belongs to SO(R+ ) and inf t∈R+ 1 + tω (t) > 0. The real-valued slowly oscillating
function
ω(t) := log[α(t)/t], t ∈ R+ ,
is called the exponent function of α ∈ SOS(R+ ).
Through the paper, we will suppose that 1 < p < ∞ and will use the following
notation:
B := B(Lp (R+ )), K := K(Lp (R+ )).
It is easily seen that if α ∈ SOS(R+ ), then the weighted shift operator defined by
Uα f := (α )1/p (f ◦ α)
is an isometric isomorphism of the Lebesgue space Lp (R+ ) onto itself. It is clear
that Uα−1 = Uα−1 , where α−1 is the inverse function to α. For k ∈ N, we denote by
Uα−k the operator (Uα−1 )k . Let Wα,p
SO
denote the collection of all operators of the
form 
A= ak Uαk (1.1)
k∈Z
 Semi-Fredholmness of Singular Integral Operators with Shifts 223

where ak ∈ SO(R+ ) for all k ∈ Z and


AWα,pSO := ak Cb (R+ ) < +∞. (1.2)
k∈Z
SO
The set Wα,p is, actually, a Banach algebra with respect to the usual operations
and the norm (1.2). By analogy with the Wiener algebra of absolutely convergent
SO
Fourier series, we will call Wα,p the Wiener algebra.
Let γ and *γ denote the real and imaginary part of γ ∈ C, respectively. If
γ ∈ C satisfies
0 < 1/p + γ < 1, (1.3)
then the operator
!  γ
1 t f (τ )
(Sγ f )(t) := dτ, (1.4)
πi R+ τ τ −t
where the integral is understood in the principal value sense, is bounded on the
Lebesgue space Lp (R+ ) (see, e.g., [30, Proposition 4.2.11]). Put
Pγ± := (I ± Sγ )/2.
This paper is a continuation of our recent works [8, 9, 19, 20] (see also ref-
erences therein). Let α, β belong to SOS(R+ ) and ak , bk ∈ SO(R+ ) for all k ∈ Z.
In [19, 20] we found criteria for the Fredholmness and a formula permitting to
calculate the index of the weighted singular integral operator of the form
M := (a0 I + a1 Uα )Pγ+ + (b0 I + b1 Uβ )Pγ− .
In this paper we assume that
 
A+ := ak Uαk ∈ Wα,p
SO
, A− := bk Uβk ∈ Wβ,p
SO
(1.5)
k∈Z k∈Z

and consider the weighted singular integral operator of the form

N := A+ Pγ+ + A− Pγ− . (1.6)
Criteria for the Fredholmness of the operator N in the particular case of α = β
and γ = 0 were obtained in [9]. The proof of the sufficiency portion is based on
the Allan–Douglas local principle and follows ideas of [15]. In this paper we will
show that the localization technique is flexible enough to treat also the case of
the left and right Fredholmness for arbitrary shifts α, β and arbitrary γ satisfying
(1.3), provided that there are one-sided inverses of A+ and A− belonging to the
SO SO
Wiener algebras Wα,p and Wβ,p , respectively. We show that the required result
on one-sided inverses can be obtained from [8].
By M (A) we denote the maximal ideal space of a unital commutative Banach
algebra A. Identifying the points t ∈ R+ with the evaluation functionals t(f ) = f (t)
for f ∈ C(R+ ), we get M (C(R+ )) = R+ . Consider the fibers
 
Ms (SO(R+ )) := ξ ∈ M (SO(R+ )) : ξ|C(R+ ) = s
224 A.Yu. Karlovich, Yu.I. Karlovich and A.B. Lebre

of the maximal ideal space M (SO(R+ )) over the points s ∈ {0, ∞}. By [23, Propo-
sition 2.1], the set
Δ := M0 (SO(R+ )) ∪ M∞ (SO(R+ ))
coincides with closSO∗ R+ \ R+ , where closSO∗ R+ is the weak-star closure of R+
in the dual space of SO(R+ ). Then M (SO(R+ )) = Δ ∪ R+ . In what follows we
write a(ξ) := ξ(a) for every a ∈ SO(R+ ) and every ξ ∈ Δ.
With the operators A± defined by (1.5), we associate the functions a± defined
on R+ × R by
 
a+ (t, x) := ak (t)eikω(t)x , a− (t, x) := bk (t)eikη(t)x , (1.7)
k∈Z k∈Z

where ω, η ∈ SO(R+ ) are the exponent functions of α, β, respectively. Since the

series in (1.5) converge absolutely, we have a± (·, x) ∈ SO(R+ ) for all x ∈ R. With
the operator N we associate the function n defined on R+ × R by
−
n(t, x) = a+ (t, x)p+
γ (x) + a− (t, x)pγ (x),
where
p±
γ (x) := (1 ± sγ (x))/2, sγ (x) := coth[π(x + i/p + iγ)], x ∈ R. (1.8)
Since a± (·, x), n(·, x) ∈ SO(R+ ) for every x ∈ R, taking the Gelfand transform of
n(·, x), we obtain for (ξ, x) ∈ (Δ ∪ R+ ) × R,
−
n(ξ, x) := a+ (ξ, x)p+
γ (x) + a− (ξ, x)pγ (x), (1.9)
which gives extensions of the functions n(·, x) to M (SO(R+ )).
Theorem 1.1 (Main result). Let 1 < p < ∞ and let γ ∈ C satisfy (1.3). Suppose
ak , bk ∈ SO(R+ ) for all k ∈ Z and α, β ∈ SOS(R+ ). If
(i) the functional operators
 
A+ := ak Uαk ∈ Wα,p
SO
, A− := bk Uβk ∈ Wβ,p
SO

k∈Z k∈Z

are left (resp. right) invertible on the space Lp (R+ );

(ii) for every ξ ∈ Δ, the function n defined by (1.7)–(1.9) satisfies the inequality
inf |n(ξ, x)| > 0; (1.10)
x∈R

then the operator N = A+ Pγ+ + A− Pγ− is left (resp. right) Fredholm on the space
Lp (R+ ).
We conjecture that conditions (i) and (ii) of Theorem 1.1 are also necessary
for the one-sided Fredholmness of the operator N .
The paper is organized as follows. Section 2 contains some auxiliary results.
In Section 3, on the basis of recent results from [8], we show that if an operator
A ∈ Wα,pSO
is left (resp. right) invertible, then at least one of its left (resp. right)
SO
inverses belongs to the same algebra Wα,p .
Section 4 is devoted to the algebra A generated by the identity operator
I and the operator S0 . We recall that A is the smallest Banach subalgebra of B
 Semi-Fredholmness of Singular Integral Operators with Shifts 225

that contains all operators similar to Mellin convolution operators with continuous
symbols. In particular, the algebra A contains the operator Sγ and the operator
Rγ with fixed singularities defined by
!  γ
1 t f (τ )
(Rγ f )(t) := dτ, (1.11)
πi R+ τ τ +t
where the integral is understood in the principal value sense. We also recall the
description of the maximal ideal space of the algebra Aπ := (A + K)/K.
In Section 5, we recall a version of the Allan–Douglas local principle suitable
for the study of one-sided invertibility in subalgebras of the Calkin algebra B π =
B/K (see [3, Theorem 1.35(a)]). Following [15, Section 6], we consider the algebra
Z generated by A ∪ K and the operators of the form cR0 , where c ∈ SO(R+ ).
We recall that the maximal ideal space of Z π := Z/K is homeomorphic to the set
{−∞, +∞} ∪(Δ× R). Further, we consider the algebra Λ of operators of local type
that consists of all operators A ∈ B such that AC − CA ∈ K for all C ∈ Z. Since
Z π is a commutative central subalgebra of Λπ := Λ/K, we can apply the Allan–
Douglas local principle to Λπ ⊂ B π and its central subalgebra Z π . In particular, an
operator T ∈ Λ is left (resp. right) Fredholm if certain cosets T π + Jξ,x
π
, T π + J+∞
π
,
and T + J−∞ are left (resp. right) invertible in the corresponding local algebras
π π

Λπξ,x , Λπ+∞ , and Λπ−∞ . Here (ξ, x) runs through Δ × R. This result is applicable to
the operator N because it belongs to the algebra Fα,β generated by the operators
S0 , Uα±1 , Uβ±1 and the multiplication operators cI with c ∈ SO(R+ ). In turn, this
algebra is contained in the algebra Λ of operators of local type.
In Section 6, we recall the definition of the algebra E(R+ , V (R)) of slowly
oscillating functions on R+ with values in the algebra V (R) of all absolutely con-
tinuous functions of finite total variation. This algebra is important for our pur-
poses because Mellin pseudodifferential operators with symbols in E(R+ , V (R))
commute modulo compact operators. Moreover, if α ∈ SOS(R+ ), then Uα Rγ is
similar to a Mellin pseudodifferential operator with symbol in E(R+ , V (R)) up
to a compact operator. These results are important ingredients of the proof of
two-sided invertibility of the cosets N π + Jξ,x π
in the quotient algebras Λπξ,x for
(ξ, x) ∈ Δ × R under condition (ii) of Theorem 1.1.
In Section 7, we prove Theorem 1.1. Since, according to Section 3, there are
left/right inverses of A+ (resp. A− ) belonging to Wα,p SO
⊂ Λ (resp. Wβ,p SO
⊂ Λ),
the left/right invertibility of A± implies the left/right invertibility of the coset
Aπ± + J±∞ π
= N π + J±∞π
in the local algebra Λπ±∞ . Finally, with the aid of the
results of Section 6, we show that condition (ii) of Theorem 1.1 is sufficient for
the two-sided invertibility of the cosets N π + Jξ,x π
in the local algebras Λπξ,x for
all (ξ, x) ∈ Δ × R. To complete the proof of Theorem 1.1, it remains to apply the
Allan–Douglas local principle (see Section 5).
Finally, in Section 8 we formulate criteria for the two-sided and one-sided
invertibility of a binomial functional operator with shift in the form A = aI − bUα ,
which were obtained in [18]. These results together with Theorem 1.1 imply more
226 A.Yu. Karlovich, Yu.I. Karlovich and A.B. Lebre

effective sufficient conditions for the left, right, and two-sided Fredholmness of
the operator (aI − bUα )Pγ+ + (cI − dUβ )Pγ− with a, b, c, d ∈ SO(R+ ) and α, β ∈
SOS(R+ ).

2. Auxiliary results
2.1. One-sided invertibility of operators on Hilbert spaces
Lemma 2.1. Let H be a Hilbert space and A ∈ B(H).
(a) The operator A is left invertible on the space H if and only if the operator
A∗ A is invertible on the space H. In this case, one of the left inverses of A
is given by AL = (A∗ A)−1 A∗ .
(b) The operator A is right invertible on the space H if and only if the operator
AA∗ is invertible on the space H. In this case, one of the right inverses of A
is given by AR = A∗ (AA∗ )−1 .

This statement is known, although we are not able to provide a precise ref-
erence. The proof of the sufficiency portion of part (a) is a trivial computation.
Now assume that ·, · is the inner product of H and AL ∈ B(H) is a left inverse
of A. Then for every f ∈ H,
f 2 ≤ AL 2 Af 2 = AL 2 |Af, Af | = AL 2 |A∗ Af, f |.
In view of the previous inequality, the invertibility of the operator A∗ A follows from
the Lax–Milgram theorem (see, e.g., [34, Chap. III, Section 7]). This completes the
proof of part (a). The proof of part (b) is reduced to the previous one by passing
to adjoint operators.
Another proof of the above lemma can be obtained from general results for
C ∗ -algebras contained in [25, § 23, Corollaries 2–3].
Lemma 2.1 can also be deduced from more general results on the Moore–
Penrose invertibility of operators on a Hilbert space (see [13, Example 2.16] or
[2, Theorem 4.24]). Notice that the operator AL (resp. AR ) is the Moore–Penrose
inverse of the operator A.

2.2. Fundamental property of slowly oscillating functions

Lemma 2.2 ([23, Proposition 2.2]). Let {ak }∞
k=1 be a countable subset of SO(R+ )
and s ∈ {0, ∞}. For each ξ ∈ Ms (SO(R+ )) there exists a sequence {tn }n∈N ⊂ R+
such that tn → s as n → ∞ and
ak (ξ) = lim ak (tn ) for all k ∈ N. (2.1)
n→∞

Conversely, if {tn }n∈N ⊂ R+ is a sequence such that tn → s as n → ∞ and

the limits limn→∞ ak (tn ) exist for all k ∈ N, then there exists a functional ξ ∈
Ms (SO(R+ )) such that (2.1) holds.
 Semi-Fredholmness of Singular Integral Operators with Shifts 227

2.3. Properties of iterations of slowly oscillating shifts

In this subsection we collect some properties of iterations of slowly oscillating
shifts. For t ∈ R+ and k ∈ N, let
α0 (t) := t, αk (t) := α[αk−1 (t)], α−k (t) := α−1 [α−k+1 (t)].
Lemma 2.3 ([16, Corollary 2.5]). If α ∈ SOS(R+ ), then αk ∈ SOS(R+ ) for every
k ∈ Z.
Lemma 2.4 ([15, Lemma 2.3]). If c ∈ SO(R+ ) and α ∈ SOS(R+ ), then c ◦ α
belongs to SO(R+ ) and
lim (c(t) − c[α(t)]) = 0 for s ∈ {0, ∞}.
t→s

Lemma 2.5 ([17, Lemma 2.6]). Let α ∈ SOS(R+ ) and α−1 be the inverse function
to α. If ω and ω−1 are the exponent functions of α and α−1 , respectively, then
ω(ξ) = −ω−1 (ξ) for all ξ ∈ Δ.
Lemma 2.6. Let α ∈ SOS(R+ ) and let ω be its exponent function. If k ∈ Z and
ωk is the exponent function of αk , then ωk (ξ) = kω(ξ) for every ξ ∈ Δ.
Proof. For k = 0, 1, the statement is trivial. If k > 1, then
⎛ ⎞
αk (t)  α[αj (t)]
k−1 
k−1
ωk (t) = log = log ⎝ ⎠= ω[αj (t)], t ∈ R+ . (2.2)
t j=0
αj (t) j=0

Since ω ∈ SO(R+ ), we deduce from Lemmas 2.3–2.4 that for every integer j ∈
{0, . . . , k − 1}, the function ω ◦ αj belongs to SO(R+ ) and
lim (ω(t) − ω[αj (t)]) = 0, s ∈ {0, ∞}. (2.3)
t→s

Fix s ∈ {0, ∞} and ξ ∈ Ms (SO(R+ )). By Lemma 2.2, there is a sequence

{tn }n∈N ⊂ R+ such that tn → s as n → ∞ and
ω(ξ) = lim ω(tn ), (ω ◦ αj )(ξ) = lim ω[αj (tn )]. (2.4)
n→∞ n→∞

Equalities (2.3)–(2.4) imply that for j ∈ {0, . . . , k − 1},

(ω ◦ αj )(ξ) − ω(ξ) = lim (ω[αj (tn )] − ω(tn )) = 0.
n→∞

We derive from (2.2) and the above equalities that


k−1
ωk (ξ) = (ω ◦ αj )(ξ) = kω(ξ),
j=0

which completes the proof for k > 1.

If k < 0, then we have ω−k (ξ) = −kω(ξ) by the statement just proved.
On the other hand, we deduce from Lemma 2.5 that ωk (ξ) = −ω−k (ξ). Thus,
ωk (ξ) = −(−k)ω(ξ) = kω(ξ) for all ξ ∈ Δ. 
228 A.Yu. Karlovich, Yu.I. Karlovich and A.B. Lebre

SO
3. Weak one-sided inverse closedness of the algebra Wα,p
SO
3.1. Inverse closedness of the algebra Wα,p in the algebra B
Let A ⊂ B be two Banach algebras with the same unit element. Recall that the
algebra A is said to be inverse closed in the algebra B if for every element a ∈ A
invertible in the algebra B its inverse a−1 belongs to the algebra A.
We say that the algebra A is weakly left (resp. right) inverse closed in the
algebra B if for every element a ∈ A, which is left (resp. right) invertible in the
algebra B, there exists at least one its left (resp. right) inverse a(−1) that belongs
to the algebra A.
Theorem 3.1 ([8, Theorem 7.4]). For every p ∈ (1, ∞), the algebra Wp,α
SO
is inverse
closed in the algebra B.
S
3.2. One-sided inverses belonging to the algebra Wα,p
A function f ∈ L∞ (R+ ) is said to be essentially slowly oscillating (at 0 and ∞) if
for each (equivalently, for some) λ ∈ (0, 1),
lim ess sup |f (t) − f (τ )| = 0, s ∈ {0, ∞}.
r→s t,τ ∈[λr,r]

Fix α ∈ SOS(R+ ) and τ ∈ R+ . Consider the semi-segment γ ⊂ R+ with the

endpoints τ and α(τ ) such that τ ∈ γ and α(τ ) ∈ / γ. Following [8, Section 3.2],
let S denote the C ∗ -subalgebra of L∞ (R+ ) consisting of all functions on R+ that
are continuous on every semi-segment αn (γ) with n ∈ Z, have one-sided limits at
the points αn (τ ) for n ∈ Z, and are essentially slowly oscillating at 0 and ∞. Let
S
Wα,p be the unital Banach algebra of operators of the form (1.1) with ak ∈ S for
all k ∈ Z and the norm

AWα,p
S = ak L∞ (R+ ) < ∞.
k∈Z

From [8, Theorems 6.3–6.4] we get the following.

Theorem 3.2. Let 1 < p < ∞, α ∈ SOS(R+ ), ak ∈ SO(R+ ) for all k ∈ Z, and

A= ak Uαk ∈ Wα,p
SO
.
k∈Z

If A is left (resp. right) invertible on Lp (R+ ), then there exists a left inverse AL
S S
(resp. right inverse AR ) of A such that AL ∈ Wα,p (resp. AR ∈ Wα,p ).
SO
3.3. Weak one-sided inverse closedness of the algebra Wα,p in B
SO
We will show that the algebra Wα,p is weakly left and right inverse closed in the
algebra B. For every operator A ∈ Wα,pSO
of the form (1.1), define its formally
adjoint A by

A := (ak ◦ α−k )Uα−k ∈ Wα,p
SO
.
k∈Z
 Semi-Fredholmness of Singular Integral Operators with Shifts 229

Theorem 3.3. Let 1 < p < ∞, α ∈ SOS(R+ ), ak ∈ SO(R+ ) for all k ∈ Z, and

A= ak Uαk ∈ Wα,p
SO
.
k∈Z
p
(a) If A is left invertible on L (R+ ), then the operator A A is invertible on the
space Lp (R+ ), the operator AL := (A A)−1 A is a left inverse of A, and
AL ∈ Wα,pSO
.
(b) If A is right invertible on Lp (R+ ), then the operator AA is invertible on the
space Lp (R+ ), the operator AR := A (AA )−1 is a right inverse of A, and
AR ∈ Wα,p SO
.
Proof. Along with the operator Uα acting on Lp (R+ ), consider the operator Uα,2
acting on L2 (R+ ) and defined by the same rule Uα,2 f = (α )1/2 (f ◦ α). Then we
can define the canonical isometric isomorphisms of Banach algebras
S S
ΨS : Wα,p → Wα,2 , SO
ΨSO : Wα,p → Wα,2
SO

by the formulas
5 6 5 6
   
k k
ΨS ck U α = ck Uα,2 , ΨSO ck Uαk = k
ck Uα,2 ,
k∈Z k∈Z k∈Z k∈Z

respectively.
If A ∈ Wα,p
SO
is left invertible on the space Lp (R+ ), then by Theorem 3.2,
there exists an operator A L ∈ Wα,p S
such that AL A = I on Lp (R+ ). Hence
ΨS (A )ΨS (A) = I on L (R+ ). Therefore, the operator
L 2

A2 := ΨSO (A) = ΨS (A) ∈ Wα,2

SO

is left invertible on L2 (R+ ). Hence, in view of Lemma 2.1, the operator A∗2 A2
is invertible on L2 (R+ ). Observe that A∗2 ∈ Wα,2
SO
and A = Φ−1 ∗
SO (A2 ) ∈ Wα,p .
SO
∗
Since A2 A2 ∈ Wα,2 , we deduce from the inverse closedness of the algebra Wα,2 in
SO SO

the algebra B (see Theorem 3.1) that (A∗2 A2 )−1 ∈ Wα,2

SO
. Then Ψ−1 ∗
SO ((A2 A2 )
−1
)∈
SO
Wα,p . Now it is easy to check that
(A A)−1 = Ψ−1 ∗
SO ((A2 A2 )
−1
).
Hence AL = (A A)−1 A is a left inverse to A. Part (a) is proved.
(b) The proof of part (b) is reduced to the previous one by passing to adjoint
operators. 

4. Algebra A of singular integral operators

4.1. Fourier and Mellin convolution operators
Let F : L2 (R) → L2 (R) denote the Fourier transform,
!
(F f )(x) := f (y)e−ixy dy, x ∈ R,
R
230 A.Yu. Karlovich, Yu.I. Karlovich and A.B. Lebre

and let F −1 : L2 (R) → L2 (R) be the inverse of F . A function a ∈ L∞ (R) is called

a Fourier multiplier on Lp (R) if the mapping f → F −1 aF f maps L2 (R) ∩ Lp (R)
into itself and extends to a bounded operator on Lp (R). The latter operator is
then denoted by W 0 (a). We let Mp (R) stand for the set of all Fourier multipliers
on Lp (R). One can show that Mp (R) is a Banach algebra under the norm
aMp (R) := W 0 (a)B(Lp (R)) .
Let dμ(t) = dt/t be the (normalized) invariant measure on R+ . Consider
the Fourier transform on L2 (R+ , dμ), which is usually referred to as the Mellin
transform and is defined by
!
dt
M : L (R+ , dμ) → L (R), (Mf )(x) :=
2 2
f (t)t−ix .
R+ t
This operator is invertible, with inverse given by
!
1
M−1 : L2 (R) → L2 (R+ , dμ), (M−1 g)(t) = g(x)tix dx.
2π R
Let E be the isometric isomorphism
E : Lp (R+ , dμ) → Lp (R), (Ef )(x) := f (ex ), x ∈ R. (4.1)
Then the map
A → E −1 AE (4.2)
transforms the Fourier convolution operator W 0 (a) = F −1 aF to the Mellin con-
volution operator
Co(a) := M−1 aM
with the same symbol a. Hence the class of Fourier multipliers on Lp (R) coincides
with the class of Mellin multipliers on Lp (R+ , dμ).

4.2. Continuous and piecewise continuous multipliers

We denote by P C the C ∗ -algebra of all bounded piecewise continuous functions on
Ṙ = R ∪ {∞}. By definition, a ∈ P C if and only if a ∈ L∞ (R) and the one-sided
limits
a(x0 − 0) := lim a(x), a(x0 + 0) := lim a(x)
x→x0 −0 x→x0 +0

exist for each x0 ∈ Ṙ. If a function a is given everywhere on R, then its total
variation is defined by

n
V (a) := sup |a(xk ) − a(xk−1 )|,
k=1

where the supremum is taken over all n ∈ N and

−∞ < x0 < x1 < · · · < xn < +∞.
If a has a finite total variation, then it has finite one-sided limits a(x − 0) and
a(x + 0) for all x ∈ Ṙ, that is, a ∈ P C (see, e.g., [26, Chap. VIII, Sections 3
 Semi-Fredholmness of Singular Integral Operators with Shifts 231

and 9]). The following theorem gives an important subset of Mp (R). Its proof can
be found, e.g., in [1, Theorem 17.1] or [6, Theorem 2.11].

Theorem 4.1 (Stechkin’s inequality). If a ∈ P C has finite total variation V (a),

then a ∈ Mp (R) and
aMp (R) ≤ SR B(Lp (R)) aL∞ (R) + V (a) ,
where SR is the Cauchy singular integral operator on R.

According to [6] or [1, p. 325], let P Cp be the closure in Mp (R) of the set of
all functions a ∈ P C with finite total variation on R. Following [1, p. 331], put
Cp (R) := P Cp ∩ C(R),

where R := [−∞, +∞]. It is easy to see that P Cp and Cp (R) are Banach algebras.

4.3. Maximal ideal space of the algebras A and Aπ

Let A be a Banach algebra and C be a subset of A. Following [3, Section 3.45], we
denote by algA C the smallest closed subalgebra of A containing C and by idA C the
smallest closed two-sided ideal of A containing C.
Put A := algB {I, S0 }. Obviously, the algebra A is commutative. Consider
the isometric isomorphism
Φ : Lp (R+ ) → Lp (R+ , dμ), (Φf )(t) := t1/p f (t), t ∈ R+ .
The following result is well known. It is essentially due to Duduchava [6, 7] and
Simonenko, Chin Ngok Minh [32]. It can be found with a proof in [5, Section 1.10.2],
[12, Section 2.1.2], [30, Sections 4.2.2–4.2.3].

Theorem 4.2.
(a) The algebra A is the smallest closed subalgebra of B that contains the opera-
tors Φ−1 Co(a)Φ with a ∈ Cp (R).
(b) The maximal ideal space of the commutative Banach algebra A is homeomor-
phic to R. In particular, the operator Φ−1 Co(a)Φ with a ∈ Cp (R) is invertible
if and only if a(x) = 0 for all x ∈ R. Thus A is an inverse closed subalgebra
of B.
(c) The operator Φ−1 Co(a)Φ with a ∈ Cp (R) belongs to idA {R0 } if and only if
a(−∞) = a(+∞) = 0.
(d) If γ ∈ C satisfies (1.3), then the function sγ given by (1.8) and the function
rγ defined by
rγ (x) := 1/ sinh[π(x + i/p + iγ)], x ∈ R,

belong to Cp (R) and

Sγ = Φ−1 Co(sγ )Φ, Rγ = Φ−1 Co(rγ )Φ.
232 A.Yu. Karlovich, Yu.I. Karlovich and A.B. Lebre

Let us describe the quotient algebra

Aπ = (A + K)/K.
By [30, Proposition 4.2.14], a Mellin convolution operator is Fredholm on the space
Lp (R+ , dμ) if and only if it is invertible on this space. Hence, Theorem 4.2 implies
the following.
Corollary 4.3. (a) The algebra Aπ is commutative and its maximal ideal space is
homeomorphic to R.
(b) The Gelfand transform of the coset (Φ−1 Co(a)Φ)π ∈ Aπ for a ∈ Cp (R) is
given by
 −1 
(Φ Co(a)Φ)π (x) = a(x) for x ∈ R.
4.4. Some operator relations
Lemma 4.4 ([14, Lemma 2.4], [20, Lemma 4.2]). Let 1 < p < ∞ and γ, δ ∈ C be
such that 0 < 1/p + γ < 1 and 0 < 1/p + δ < 1. Then
1 eiπ(δ−γ)
Pδ+ − Pγ+ = Pγ− − Pδ− = sinh[πi(γ − δ)]Rγ Rδ , Pγ− Pδ+ = − Rγ Rδ .
2 4
4.5. Compactness of commutators of singular integral operators and
functional operators
Fix α, β ∈ SOS(R+ ) and consider the Banach algebra of functional operators with
shifts and slowly oscillating data defined by
F Oα,β := algB {Uα , Uα−1 , Uβ , Uβ−1 , cI : c ∈ SO(R+ )}.

Lemma 4.5 ([17, Lemma 2.8]). Let α, β ∈ SOS(R+ ). If A ∈ F Oα,β and B ∈ A,

then AB  BA.

5. Allan–Douglas localization
5.1. The Allan–Douglas local principle
Let A be a Banach algebra with identity. A subalgebra Z of A is said to be a
central subalgebra of A if za = az for all z ∈ Z and all a ∈ A.
The proof of the following result is contained, e.g., in [3, Theorem 1.35(a)].
Theorem 5.1 (Allan–Douglas). Let A be a Banach algebra with identity e and let
Z be a closed central subalgebra of A containing e. Let M (Z) be the maximal ideal
space of Z, and for ω ∈ M (Z), let Jω refer to the smallest closed two-sided ideal
of A containing the ideal ω. Then an element a is left (resp. right, two-sided)
invertible in A if and only if a + Jω is left (resp. right, two-sided) invertible in the
quotient algebra A/Jω for all ω ∈ M (Z).

The algebra A/Jω is referred to as the local algebra of A at ω ∈ M (Z).

 Semi-Fredholmness of Singular Integral Operators with Shifts 233

5.2. Algebras of singular integral operators with shifts and algebras

of operators of local type
Following [15, Section 6.3], we consider the following sets:
Z := algB {I, S0 , cR0 , K : c ∈ SO(R+ ), K ∈ K},
Λ := {A ∈ B : AC − CA ∈ K for all C ∈ Z}.
By [15, Lemma 6.7(a)], the set Λ is a closed unital subalgebra of the algebra B,
which is usually called the algebra of operators of local type.
For α, β ∈ SOS(R+ ), put
Fα,β := algB ({S0 } ∪ F O α,β ).
By a minor modification of the proof of [15, Theorem 6.8] with the aid of
Lemma 4.5, we get the following.
Theorem 5.2. We have K ⊂ Z ⊂ Fα,β ⊂ Λ.
5.3. Maximal ideal space of the algebra Z π
It follows from Theorem 5.2 that the quotient algebras Z π := Z/K and Λπ := Λ/K
are well defined. Clearly, Z π lies in the center of Λπ .
Theorem 5.3 ([15, Theorem 6.11]). For the commutative Banach algebra Z π the
following statements hold:
(a) the maximal ideal space M (Z π ) of Z π is homeomorphic to the set
{−∞, +∞} ∪ (Δ × R);
(b) any coset Z ∈ Z is of the form
π π


mn
Z π = (c+ P0+ )π + (c− P0− )π + lim (cn,k Hn,k )π (5.1)
n→∞
k=1
where c± ∈ C, cn,k ∈ SO(R+ ), Hn,k ∈ idA {R0 }, and mn ∈ N;
(c) the Gelfand transform of the coset Z π ∈ Z π defined by (5.1) is given for a
point (ξ, x) ∈ Δ × R by

mn
−
(Z π )(ξ, x) = c+ p+
0 (x) + c p
− 0 (x) + lim π
cn,k (ξ)(Hn,k )(x),
n→∞
k=1
where π
(Hn,k )(x) π
is the Gelfand transform of a coset Hn,k ∈ Aπ , which is
calculated in Corollary 4.3(b).
5.4. Semi-Fredholmness of operators of local type
Let J+∞
π
, J−∞
π
and Jξ,xπ
for (ξ, x) ∈ Δ × R be the closed two-sided ideals of the
π
Banach algebra Λ generated, respectively, by the maximal ideals
 
I+∞
π
:= idZ π (P0− )π , (gR0 )π : g ∈ SO(R+ ) ,
 
I−∞
π
:= idZ π (P0+ )π , (gR0 )π : g ∈ SO(R+ ) ,
 
Iξ,x
π
:= Z π ∈ Z π : (Z π )(ξ, x) = 0
234 A.Yu. Karlovich, Yu.I. Karlovich and A.B. Lebre

of the algebra Z π , and let

Λπ+∞ := Λπ /J+∞
π
, Λπ−∞ := Λπ /J−∞
π
, Λπξ,x := Λπ /Jξ,x
π

be the corresponding quotient algebras (see also [15, Section 6.6]).

Obviously, an operator T ∈ Λ is left Fredholm (resp. right Fredholm) on the
space Lp (R+ ) if the coset T π := A + K is left invertible (resp. right invertible) in
the quotient Banach algebra Λπ ⊂ B π . Applying now Theorem 5.1 with A = Λπ
and Z = Z π , we immediately obtain the following.
Theorem 5.4. An operator T ∈ Λ is left (resp. right) Fredholm on the space Lp (R+ )
if the following three conditions are fulfilled:
(i) the coset T π +J+∞
π
is left (resp. right) invertible in the quotient algebra Λπ+∞ ;
(ii) the coset T +J−∞ is left (resp. right) invertible in the quotient algebra Λπ−∞ ;
π π

(iii) for every (ξ, x) ∈ Δ × R, the coset T π + Jξ,xπ

is left (resp. right) invertible in
π
the quotient algebra Λξ,x .

It follows from Theorems 4.2(d) and 5.2 that N ∈ Fα,β ⊂ Λ. Thus, Theo-
rem 5.4 is applicable to N . Hence, our next aim is to study one-sided invertibility
of the cosets N π + J+∞π
, N π + J−∞
π
and N π + Jξ,x
π
in the corresponding local
algebras Λ+∞ , Λ−∞ and Λξ,x for all (ξ, x) ∈ Δ × R.
π π π

6. Mellin pseudodifferential operators and their symbols

6.1. Mellin PDO’s: overview
Mellin pseudodifferential operators are generalizations of Mellin convolution op-
erators. Let a be a sufficiently smooth function defined on R+ × R. The Mellin
pseudodifferential operator (shortly, Mellin PDO) with symbol a is initially defined
for smooth functions f of compact support by the iterated integral
[Op(a)f ](t) = [M−1 a(t, ·)Mf ](t)
! !  ix
1 t dτ
= dx a(t, x) f (τ ) for t ∈ R+ . (6.1)
2π R R+ τ τ
Obviously, if a(t, x) = a(x) for all (t, x) ∈ R+ × R, then the Mellin pseudodiffer-
ential operator Op(a) becomes the Mellin convolution operator
Op(a) = Co(a).
In 1991 Rabinovich [27] (see also [28]) proposed to use Mellin pseudodif-
ferential operators with C ∞ slowly oscillating symbols to study singular integral
operators with slowly oscillating coefficients on Lp spaces. Namely, he considered
symbols a ∈ C ∞ (R+ × R) such that
sup (t∂t )j ∂xk a(t, x) (1 + x2 )k/2 < ∞ for all j, k ∈ Z+ (6.2)
(t,x)∈R+ ×R
 Semi-Fredholmness of Singular Integral Operators with Shifts 235

and
lim sup (t∂t )j ∂xk a(t, x) (1 + x2 )k/2 = 0 for all j ∈ N, k ∈ Z+ , (6.3)
t→s x∈R

where Z+ = N ∪ {0} and s ∈ {0, ∞}. Here and in what follows ∂t and ∂x denote
the operators of partial differentiation with respect to t and to x. Notice that (6.2)
0
defines nothing but the Mellin version of the Hörmander class S1,0 (R) (see, e.g.,
m
[24, Chap. 2, Section 1] for the definition of the Hörmander classes S,δ (Rn )). If a
satisfies (6.2), then the Mellin PDO Op(a) is bounded on the spaces Lp (R+ , dμ) for
1 < p < ∞ (see, e.g., [33, Chap. VI, Proposition 4] for the corresponding Fourier
PDO’s). Condition (6.3) is the Mellin version of Grushin’s definition of symbols
slowly varying in the first variable (see, e.g., [11], [24, Chap. 3, Definition 5.11]).
The idea of application of Mellin PDO’s with considered class of symbols was
exploited in a series of papers by Rabinovich and coauthors (see [29, Sections 4.6–
4.7] for the complete history up to 2004). On the other hand, the smoothness
conditions imposed on slowly oscillating symbols are very strong. In particular,
they are not applicable directly to the problem we are dealing with in the present
paper.
In 2005 the second author [22] developed a Fredholm theory for the Fourier
pseudodifferential operators with slowly oscillating symbols of limited smoothness
in the spirit of Sarason’s definition [31, p. 820] of slow oscillation adopted in the
present paper (much less restrictive than in [28] and in the works mentioned in
[29]). Necessary for our purposes results from [22] were translated to the Mellin
setting, for instance, in [16] with the aid of the transformation defined by (4.1)–
(4.2). For the convenience of readers, we reproduce the results required in what
follows exactly in the same form as they were stated in [16], where more details
on their proofs can be found.

6.2. Boundedness of Mellin PDO’s

If a is an absolutely continuous function of finite total variation on R, then its
derivative belongs to L1 (R) and
!
V (a) = |a (x)| dx
R

(see, e.g., [26, Chap. VIII, Sections 3 and 9; Chap. XI, Section 4]). The set V (R)
of all absolutely continuous functions of finite total variation on R forms a Banach
algebra when equipped with the norm
!
aV := aL∞ (R) + V (a) = aL∞ (R) + |a (x)| dx. (6.4)
R

Let Cb (R+ , V (R)) denote the Banach algebra of all bounded continuous
V (R)-valued functions on R+ with the norm
a(·, ·)Cb (R+ ,V (R)) = sup a(t, ·)V .
t∈R+
236 A.Yu. Karlovich, Yu.I. Karlovich and A.B. Lebre

As usual, let C0∞ (R+ ) be the set of all infinitely differentiable functions of compact
support on R+ .
Theorem 6.1 ([16, Theorem 3.1]). If a ∈ Cb (R+ , V (R)), then the Mellin pseudodif-
ferential operator Op(a), defined for functions f ∈ C0∞ (R+ ) by the iterated integral
(6.1), extends to a bounded linear operator on the space Lp (R+ , dμ) and there is a
number Cp ∈ (0, ∞) depending only on p such that
Op(a)B(Lp (R+ ,dμ)) ≤ Cp aCb (R+ ,V (R)) .
6.3. Products of Mellin PDO’s
Consider the Banach subalgebra SO(R+ , V (R)) of the algebra Cb (R+ , V (R)) con-
sisting of all V (R)-valued functions a on R+ that slowly oscillate at 0 and ∞,
that is,
lim max a(t, ·) − a(τ, ·)L∞ (R) = 0, s ∈ {0, ∞}.
r→s t,τ ∈[r,2r]

Let E(R+ , V (R)) be the Banach algebra of all V (R)-valued functions a in the
algebra SO(R+ , V (R)) such that
$ $
lim sup $a(t, ·) − ah (t, ·)$ = 0 V
|h|→0 t∈R+

where ah (t, x) := a(t, x + h) for all (t, x) ∈ R+ × R.

Theorem 6.2 ([16, Theorem 3.3]). If a, b ∈ E(R+ , V (R)), then
Op(a)Op(b)  Op(ab).
Lemma 6.3 ([16, Lemma 3.4]). If a, b, c ∈ E(R+ , V (R)) are such that a depends
only on the first variable and c depends only on the second variable, then
Op(a)Op(b)Op(c) = Op(abc).
6.4. Applications of Mellin pseudodifferential operators
We immediately deduce the following assertion from [14, Lemma 4.1].
Lemma 6.4. Suppose f ∈ SO(R+ ) and γ ∈ C satisfies (1.3). Then the functions
f(t, x) := f (t), sγ (t, x) := sγ (x), rγ (t, x) := rγ (x), (t, x) ∈ R+ × R,
belong to the Banach algebra E(R+ , V (R)).
Lemma 6.5. If b ∈ E(R+ , V (R)), then the operator Φ−1 Op(b)Φ belongs to the
algebra Λ.
Proof. Let c ∈ SO(R+ ). It follows from Lemma 6.4 that the functions
c0 (t, x) := c(t)r0 (x), s0 (t, x) := s0 (x), (t, x) ∈ R+ × R,
belong to the algebra E(R+ , V (R)). Since E(R+ , V (R)) ⊂ Cb (R+ , V (R)), Theo-
rem 6.1 implies that B := Φ−1 Op(b)Φ ∈ B. We infer from Theorem 6.2 that
Op(s0 )Op(b) − Op(b)Op(s0 ) ∈ K(Lp (R+ , dμ)), (6.5)
Op(c0 )Op(b) − Op(b)Op(c0 ) ∈ K(L (R+ , dμ)). p
(6.6)
 Semi-Fredholmness of Singular Integral Operators with Shifts 237

On the other hand, by Theorem 4.2(d) and Lemma 6.3,

S0 = Φ−1 Co(s0 )Φ = Φ−1 Op(s0 )Φ, (6.7)
−1 −1
cR0 = cΦ Co(r0 )Φ = Φ Op(c0 )Φ. (6.8)
Combining (6.5)–(6.8), we conclude that S0 B − BS0 , (cR0 )B − B(cR0 ) ∈ K. Hence
B ∈ Λ. 

Applying [14, Lemma 4.4] and making minor modifications in the proof of
[16, Lemma 4.5], we get the following.

Lemma 6.6. Let γ ∈ C satisfy (1.3). Suppose α ∈ SOS(R+ ), ω is its exponent

function, and Uα is the associated isometric shift operator on Lp (R+ ). Then the
operator Uα Rγ can be realized as the Mellin pseudodifferential operator up to a
compact operator:
Uα Rγ  Φ−1 Op(c)Φ,
where the function c, given by
c(t, x) := eiω(t)x rγ (x) for (t, x) ∈ R+ × R,
belongs to the Banach algebra E(R+ , V (R)).

7. Sufficient conditions for the semi-Fredholmness

7.1. One-sided invertibility in the quotient algebras Λπ π
+∞ and Λ−∞
The next lemma shows that the operator N can be written as a paired operator
with respect to the pair (P0+ , P0− ).

Lemma 7.1. Let 1 < p < ∞ and let γ ∈ C satisfy (1.3). Suppose ak , bk belong to
SO(R+ ) for all k ∈ Z, α, β belong to SOS(R+ ), the operators A+ ∈ Wα,p SO
and
A− ∈ Wβ,p are given by (1.5), and the operator N is given by (1.6). Then the
SO

operator N can be represented in each of the forms

N = A+ P0+ + C− P0− = C+ P0+ + A− P0− , (7.1)
where
C+ := A+ + 2 sinh(πiγ)eπiγ (A+ − A− )Pγ− ,
C− := A− + 2 sinh(πiγ)e−πiγ (A+ − A− )Pγ+ .

Moreover, all operators DP0+ − P0+ D and DP0− − P0− D, where D belongs to the
set {A+ , A− , C+ , C− }, are compact on the space Lp (R+ ).

Proof. Both representations follow from Lemma 4.4. The compactness of the com-
mutators DP0± − P0± D is a consequence of parts (a) and (d) of Theorem 4.2 and
Lemma 4.5. 
238 A.Yu. Karlovich, Yu.I. Karlovich and A.B. Lebre

The following statement generalizes [15, Theorem 8.1] from the case of two-
sided invertible binomial functional operators A+ and A− to the case of one-sided
invertible operators A+ ∈ Wα,pSO
and A− ∈ Wβ,pSO
. This generalization is possible
thanks to Theorem 3.3, although the proof follows the same lines as in [15].
Theorem 7.2. Let 1 < p < ∞ and let γ ∈ C satisfy (1.3). Suppose ak , bk belong
to SO(R+ ) for all k ∈ Z, α, β belong to SOS(R+ ), the operators A+ ∈ Wα,p
SO
and
A− ∈ Wβ,p are given by (1.5), and the operator N is given by (1.6).
SO

(a) If the operator A+ is left (resp. right) invertible on the space Lp (R+ ), then the
coset N π + J+∞π
is left (resp. right) invertible in the quotient algebra Λπ+∞ .
(b) If the operator A− is left (resp. right) invertible on the space Lp (R+ ), then the
coset N π + J−∞π
is left (resp. right) invertible in the quotient algebra Λπ−∞ .
Proof. Recall that F Oα,β ⊂ Fα,β ⊂ Λ in view of Theorem 5.2. By Lemma 7.1, the
operator N is represented in each of the forms (7.1), where
A+ ∈ Wα,p
SO
⊂ F Oα,β ⊂ Fα,β ⊂ Λ, A− ∈ Wβ,p
SO
⊂ F Oα,β ⊂ Fα,β ⊂ Λ,
and C+ , C− ∈ Fα,β ⊂ Λ.
(a) Take A+ ∈ Wα,p
SO
. If A+ is left (resp. right) invertible in B, then it follows
(−1)
from Theorem 3.3 that there exists a left (resp. right) inverse A+ of A+ such
(−1)
that A+ ∈ Wα,pSO
⊂ Fα,β ⊂ Λ. Hence the coset Aπ+ = A+ + K is left (resp.
right) invertible in the quotient algebra Λπ , which implies the left (resp. right)
invertibility of the coset Aπ+ + J+∞
π
in the quotient algebra Λπ+∞ . Hence we infer
from (7.1) that
N π + J+∞
π
= (A+ P0+ + C− P0− )π + J+∞
π

= Aπ+ + [(C− − A+ )P0− ]π + J+∞

π
= Aπ+ + J+∞
π

because (P0− )π ∈ I+∞π

⊂ J+∞
π
. Thus, the left (resp. right) invertibility of the
operator A+ in B implies the left (resp. right) invertibility of the coset N π + J+∞
π
π
in the quotient algebra Λ+∞ . Part (a) is proved.
(b) The proof is analogous. 

7.2. Invertibility in the quotient algebras Λπ

ξ,x with (ξ, x) ∈ Δ × R
By a literal repetition with minor modifications of the proof of [15, Lemma 7.4],
we get the following.
Lemma 7.3. Let γ ∈ C satisfy (1.3). Suppose α ∈ SOS(R+ ) and ω is its exponent
function. If ξ ∈ Δ and
a(t, x) := eiω(t)x (rγ (x))2 for (t, x) ∈ (Δ ∪ R+ ) × R,
then there exists a function bξ ∈ E(R+ , V (R)) such that
a(t, x) − a(ξ, x) = (ω(t) − ω(ξ))bξ (t, x)rγ (x) for (t, x) ∈ R+ × R.
 Semi-Fredholmness of Singular Integral Operators with Shifts 239

Lemma 7.4. Let γ ∈ C satisfy (1.3). Suppose α is a slowly oscillating shift, ω is its
exponent function, and Uα is the associated isometric shift operator on Lp (R+ ). If
(ξ, x) ∈ Δ × R, then
(Uα Rγ2 )π − eiω(ξ)x (rγ (x))2 I π ∈ Jξ,x
π
.
Proof. The proof is developed by analogy with [15, Lemma 8.3]. In view of Lem-
ma 6.6,
Uα Rγ  Φ−1 Op(cω,γ )Φ, (7.2)
where cω,γ ∈ E(R+ , V (R)) is given by
cω,γ (t, y) := eiω(t)y rγ (y) for (t, y) ∈ R+ × R.
On the other hand, in view of Theorem 4.2(d) and Lemma 6.4,
Rγ = Φ−1 Co(rγ )Φ = Φ−1 Op(rγ )Φ, (7.3)
where rγ ∈ E(R+ , V (R)) is given by
rγ (t, y) = rγ (y) for (t, y) ∈ R+ × R.
It follows from (7.2)–(7.3) and Lemma 6.3 that
Uα Rγ2  Φ−1 Op(a)Φ, (7.4)
where
a(t, y) = cω,γ (t, y)rγ (t, y) = eiω(t)y (rγ (y))2 for (t, y) ∈ R+ × R.
Since a(·, y) ∈ SO(R+ ) for every y ∈ R, taking the Gelfand transform of a(·, y),
we obtain
a(t, y) = eiω(t)y (rγ (y))2 for (t, y) ∈ (Δ ∪ R+ ) × R.
Fix (ξ, x) ∈ Δ × R. Let us represent the function a in the form
a(t, y) = a(t, y) − a(ξ, y) + cω,γ (ξ, y)rγ (y)
= [a(t, y) − a(ξ, y)] + [cω,γ (ξ, y) − cω,γ (ξ, x)]rγ (t, y)
+ cω,γ (ξ, x)[rγ (t, y) − rγ (t, x)] + a(ξ, x), (7.5)
where (t, y) ∈ R+ × R. Further, we deduce from Lemma 7.3 that there exists a
function bξ ∈ E(R+ , V (R)) such that
a − a(ξ, ·) = (ω − ω(ξ))bξ rγ .
Hence, we infer from the above equality and Lemmas 6.4 and 6.3 that
Φ−1 Op(a − a(ξ, ·))Φ = Φ−1 Op((ω − ω(ξ))bξ rγ )Φ
= (Φ−1 Op(ω − ω(ξ))Φ) · (Φ−1 Op(bξ )Φ) · (Φ−1 Op(rγ )Φ).
The latter equality, Theorem 6.2 and equality (7.3) imply that
Φ−1 Op(a − a(ξ, ·))Φ  (ω − ω(ξ))Rγ (Φ−1 Op(bξ )Φ). (7.6)
Applying Theorem 5.3(c) and Corollary 4.3(b), we obtain
([(ω − ω(ξ))Rγ ]π )(ξ, x) = (ω(ξ) − ω(ξ))rγ (x) = 0.
240 A.Yu. Karlovich, Yu.I. Karlovich and A.B. Lebre

Therefore [(ω − ω(ξ))Rγ ]π ∈ Iξ,x

π
. On the other hand, since bξ ∈ E(R+ , V (R)), we
conclude from Lemma 6.5 that Φ−1 Op(bξ )Φ ∈ Λ. Then, taking into account (7.6)
and the definition of the ideal Jξ,x
π
, we infer that
[Φ−1 Op(a − a(ξ, ·))Φ]π = [(ω − ω(ξ))Rγ (Φ−1 Op(bξ )Φ)]π ∈ Jξ,x
π
. (7.7)
Taking into account the definition of the norm (6.4) in the algebra V (R), it is easy
to see that the function cω,γ (ξ, ·) belongs to V (R), where
cω,γ (ξ, y) = eiω(ξ)y rγ (y) for y ∈ R+ .
Then, by Stechkin’s inequality (Theorem 4.1), cω,γ (ξ, ·) ∈ Cp (R). Hence, it fol-
lows from Theorem 4.2(a) that Φ−1 Co(cω,γ (ξ, ·))Φ ∈ A. By Theorem 5.3(c) and
Corollary 4.3(b),
([Φ−1 Op(cω,γ (ξ, ·) − cω,γ (ξ, x))Φ]π )(ξ, x)
= ([Φ−1 Co(cω,γ (ξ, ·))Φ]π )(ξ, x) − ([cω,γ (ξ, x)I]π )(ξ, x)
= eiω(ξ)x rγ (x) − eiω(ξ)x rγ (x) = 0.
Therefore
[Φ−1 Op(cω,γ (ξ, ·) − cω,γ (ξ, x))Φ]π ∈ Iξ,x
π
.
By this observation, Lemma 6.3 and equality (7.3), we obtain
[Φ−1 Op((cω,γ (ξ, ·) − cω,γ (ξ, x))rγ )Φ]π
= [Φ−1 Op(cω,γ (ξ, ·) − cω,γ (ξ, x))Φ]π [Φ−1 Op(rγ )Φ]π
= [Φ−1 Op(cω,γ (ξ, ·) − cω,γ (ξ, x))Φ]π Rγπ ∈ Jξ,x
π
. (7.8)
Finally, in view of (7.3), Theorem 5.3(c) and Corollary 4.3(b), we deduce that
([Φ−1 Op(rγ − rγ (x))Φ]π )(ξ, x) = ([Rγ − rγ (x)I]π )(ξ, x) = rγ (x) − rγ (x) = 0.
Hence
cω,γ (ξ, x)[Φ−1 Op(rγ − rγ (x))Φ]π ∈ Iξ,x
π
⊂ Jξ,x
π
. (7.9)
Combining (7.4)–(7.5) with (7.7)–(7.9), we arrive at the relation
(Uα Rγ2 )π − eiω(ξ)x (rγ (x))2 I π = [Φ−1 Op(a)Φ − a(ξ, x)I]π ∈ Jξ,x
π
,
which completes the proof. 

Now we are in a position to prove that condition (ii) of Theorem 1.1 is

sufficient for the invertibility of the coset N π + Jξ,x
π
in the quotient algebra Λπξ,x .
Theorem 7.5. Let 1 < p < ∞ and let γ ∈ C satisfy (1.3). Suppose ak , bk belong
to SO(R+ ) for all k ∈ Z, α, β belong to SOS(R+ ), the operators A+ ∈ Wα,pSO
and
A− ∈ Wβ,p are given by (1.5), and the operator N is given by (1.6). If n(ξ, x) = 0
SO

for some (ξ, x) ∈ Δ × R, where the function n is defined by (1.7)–(1.9), then the
coset N π + Jξ,x
π
is two-sided invertible in the quotient algebra Λπξ,x .
 Semi-Fredholmness of Singular Integral Operators with Shifts 241

Proof. We follow the main lines of the proof of [15, Theorem 8.4].
Fix (ξ, x) ∈ Δ × R and consider the operators
p±
γ (x)
H± := R2 . (7.10)
[rγ (x)]2 γ
Then it follows from Theorem 5.3(c) and Corollary 4.3(b) that
π
(H± )(ξ, x) = p±
γ (x). (7.11)
Therefore, taking into account Corollary 4.3(b) once again, we get
(Pγ± − H± )π ∈ Iξ,x
π

and  π
A+ (Pγ+ − H+ ) + A− (Pγ− − H− ) ∈ Jξ,x
π
,
whence
N π + Jξ,x
π
= (A+ H+ + A− H− )π + Jξ,x
π
. (7.12)
We know from Theorem 4.2(d),(a) that H± ∈ A. Hence, we infer from Lemma 4.5
that for all k ∈ Z,
(Uαk H+ )π = (H+ Uαk )π , (Uβk H− )π = (H− Uβk )π . (7.13)
Taking into account (7.11), it is easy to see that for all k ∈ Z,
   
(ak H+ )π − (ak (ξ)H+ )π (ξ, x) = 0, (bk H− )π − (bk (ξ)H− )π (ξ, x) = 0.
Hence
(ak H+ )π − (ak (ξ)H+ )π , (bk H− )π − (bk (ξ)H− )π ∈ Iξ,x
π
⊂ Jξ,x
π
. (7.14)
Applying (7.13) and (7.14), we obtain for all k ∈ Z,
(ak Uαk H+ )π − (ak (ξ)Uαk H+ )π = [(ak H+ − ak (ξ)H+ )Uαk ]π ∈ Jξ,x
π
, (7.15)
(bk Uβk H− )π − (bk (ξ)Uβk H− )π = [(bk H− − bk (ξ)H− )Uβk ]π ∈ Jξ,x
π
. (7.16)
Then it follows from (7.12) and (7.15)–(7.16) that
 π
N π + Jξ,x
π
= ak (ξ)Uαk H+ + bk (ξ)Uβk H− + Jξ,x
π
. (7.17)
k∈Z

In view of Theorem 5.3(c) and Corollary 4.3(b), it is easy to see that

 π
(rγ (x))−2 Rγ2 − I (ξ, x) = 0.
Hence
H± π
− p±
γ (x)I ∈ Iξ,x ⊂ Jξ,x .
π π π
(7.18)
By Lemmas 2.3, 2.6, and 7.4, we deduce for all k ∈ Z that
(Uαk Rγ2 )π − eikω(ξ)x (rγ (x))2 I π , (Uβk Rγ2 )π − eikη(ξ)x (rγ (x))2 I π ∈ Jξ,x
π
.
The above inclusions together with (7.10) imply for every k ∈ Z that
(Uαk H+ )π − eikω(ξ)x p+
γ (x) I ∈ Jξ,x ,
π π
(7.19)
(Uβk H− )π − eikη(ξ)x p−
γ (x) I
π
∈ Jξ,x
π
. (7.20)
242 A.Yu. Karlovich, Yu.I. Karlovich and A.B. Lebre

Combining (7.18)–(7.20), we arrive at the equality

N π + Jξ,x
π
= n(ξ, x)I π + Jξ,x
π
,
where n(ξ, x) is given by (1.8)–(1.9). If n(ξ, x) = 0, then one can check straight-
forwardly that (1/n(ξ, x))I π + Jξ,x
π
is the inverse of the coset N π + Jξ,x
π
in the
quotient algebra Λπξ,x . 

7.3. Proof of Theorem 1.1

The proof is analogous to that of [15, Theorem 1.2]. We know from Theorem 4.2(d)
and Theorem 5.2 that N ∈ Λ. If condition (i) of Theorem 1.1 is fulfilled, that is, if
the operators A+ and A− are left (resp. right) invertible, then by Theorem 7.2 the
coset N π + J+∞
π
is left (resp. right) invertible in the quotient algebra Λπ+∞ and the
coset N + J−∞ is left (resp. right) invertible in the quotient algebra Λπ−∞ . On the
π π

other hand, if condition (ii) of Theorem 1.1 holds, then in view of Theorem 7.5,
the coset N π + Jξ,x
π
is two-sided invertible in the quotient algebra Λπξ,x for every
pair (ξ, x) ∈ Δ × R. Then, by Theorem 5.4, the operator N ∈ Λ is left (resp. right)
Fredholm. 

8. Semi-Fredholmness of weighted singular integral operators with

coefficients being binomial functional operators
8.1. Criteria for the two-sided and strict one-sided invertibility of aI − bUα
Suppose a, b ∈ SO(R+ ) and α ∈ SOS(R+ ). For s ∈ {0, ∞}, put
L∗ (s; a, b) := lim inf (|a(t)| − |b(t)|), L∗ (s; a, b) := lim sup(|a(t)| − |b(t)|).
t→s t→s

Fix a point τ ∈ R+ and put

τ−,α := lim αn (τ ), τ+,α := lim αn (τ ).
n→−∞ n→+∞

Then either τ−,α = 0 and τ+,α = ∞, or τ−,α = ∞ and τ+,α = 0. The points τ+,α
and τ−,α are called attracting and repelling points of α, respectively.
We say that the triple {α, a, b} satisfies conditions (I1), (I2), (LI), (RI) if
(I1) L∗ (τ−,α ; a, b) > 0 and L∗ (τ+,α ; a, b) > 0 and inf |a(t)| > 0;
t∈R+
(I2) L∗ (τ−,α ; a, b) < 0 and L∗ (τ+,α ; a, b) < 0 and inf |b(t)| > 0;
t∈R+
(LI) L∗ (τ−,α ; a, b) < 0 < L∗ (τ+,α ; a, b) and for every t ∈ R+ there is an integer kt
such that b[αk (t)] = 0 for k < kt and a[αk (t)] = 0 for k > kt .
(RI) L∗ (τ+,α ; a, b) < 0 < L∗ (τ−,α ; a, b) and for every t ∈ R+ there is an integer kt
such that b[αk (t)] = 0 for k ≥ kt and a[αk (t)] = 0 for k < kt .
Theorem 8.1 ([18, Theorems 1.1–1.2]). Let a, b ∈ SO(R+ ), α ∈ SOS(R+ ), and let
the binomial functional operator A be given by
A := aI − bUα .
 Semi-Fredholmness of Singular Integral Operators with Shifts 243

(a) The operator A is invertible on the Lebesgue space Lp (R+ ) if and only if the
triple {α, a, b} satisfies either condition (I1), or condition (I2).
(b) The operator A is strictly left invertible on the space Lp (R+ ) if and only if
the triple {α, a, b} satisfies condition (LI).
(c) The operator A is strictly right invertible on the space Lp (R+ ) if and only if
the triple {α, a, b} satisfies condition (RI).
8.2. Sufficient conditions for the semi-Fredholmness
Combining Theorem 1.1 and Theorem 8.1, we arrive at the following.
Corollary 8.2. Let 1 < p < ∞ and let γ ∈ C satisfy (1.3). Suppose a, b, c, d belong to
SO(R+ ), α, β belong to SOS(R+ ), and ω, η ∈ SO(R+ ) are the exponent functions
of the shifts α, β, respectively. Consider the operator
M := (aI − bUα )Pγ+ + (cI − dUβ )Pγ−
and the corresponding function m defined for (ξ, x) ∈ (R+ ∪ Δ) × R by
iη(ξ)x −
γ (x) + (c(ξ) − d(ξ)e
m(ξ, x) := (a(ξ) − b(ξ)eiω(ξ)x )p+ )pγ (x),
where the functions p±γ are defined by (1.8).
(a) If each of the triples {α, a, b} and {β, c, d} satisfies either condition (I1) or
condition (I2) (but not necessarily the same condition), and
inf |m(ξ, x)| > 0 for every ξ ∈ Δ, (8.1)
x∈R

then the operator M is Fredholm on the space Lp (R+ ).

(b) If each of the triples {α, a, b} and {β, c, d} satisfies only one of conditions
(I1), (I2) and (LI) (but not necessarily the same condition), and condition
(8.1) is fulfilled, then the operator M is left Fredholm on the space Lp (R+ ).
(c) If each of the triples {α, a, b} and {β, c, d} satisfies only one of conditions
(I1), (I2) and (RI) (but not necessarily the same condition), and condition
(8.1) is fulfilled, then the operator M is right Fredholm on the space Lp (R+ ).
Another (more involved) proof of Corollary 8.2(a), relying on criteria for
the Fredholmness of Mellin pseudodifferential operators (see [22] and [16, Theo-
rem 3.6]), is given in [19, Theorem 1.3]. The converse statement to Corollary 8.2(a)
is proved in [20, Theorem 1.2]. The statements of parts (b) and (c) in Corollary 8.2
are new.
Acknowledgment
We would like to thank the anonymous referee for pointing out that Lemma 2.1
can be obtained from [13, Example 2.16].
Added in proof
We proved recently in [21, Theorem 1.1] that conditions (i) and (ii) of Theorem
1.1 are also necessary for the left (resp. right) Fredholmness of the operator N .
244 A.Yu. Karlovich, Yu.I. Karlovich and A.B. Lebre

References
[1] A. Böttcher, Yu.I. Karlovich, and I.M. Spitkovsky, Convolution Operators and Fac-
torization of Almost Periodic Matrix Functions. Operator Theory: Advances and
Applications, vol. 131. Birkhäuser, Basel, 2002.
[2] A. Böttcher and B. Silbermann, Introduction to Large Truncated Toeplitz Matrices.
Springer, New York, 1999.
[3] A. Böttcher and B. Silbermann, Analysis of Toeplitz Operators. Springer, Berlin,
2006.
[4] J.B. Conway, A Course in Functional Analysis. Springer, Berlin, 1990.
[5] V.D. Didenko and B. Silbermann, Approximation of Additive Convolution-Like Op-
erators. Real C*-Algebra Approach. Birkhäuser, Basel, 2008.
[6] R. Duduchava, Integral Equations with Fixed Singularities. Teubner Verlagsge-
sellschaft, Leipzig, 1979.
[7] R. Duduchava, On algebras generated by convolutions and discontinuous functions.
Integr. Equ. Oper. Theory 10 (1987) 505–530.
[8] G. Fernández-Torres and Yu.I. Karlovich, Two-sided and one-sided invertibility of
the Wiener type functional operators with a shift and slowly oscillating data. Banach
J. Math. Analysis 11 (2017), 554–590.
[9] G. Fernández-Torres and Yu.I. Karlovich, Fredholmness of nonlocal singular integral
operators with slowly oscillating data. In: “Integral Methods in Science and Engi-
neering, vol. 1”, eds. C. Constanda et al. Birkhäuser, Cham, 2017, pp. 95–105.
[10] I. Gohberg and N. Krupnik, One-Dimensional Linear Singular Integral Equations.
I. Introduction. Operator Theory: Advances and Applications, vol. 53. Birkhäuser,
Basel, 1992.
[11] V.V. Grushin, Pseudodifferential operators on Rn with bounded symbols. Funct. Anal.
Appl. 4 (1970) 202–212.
[12] R. Hagen, S. Roch, and B. Silbermann, Spectral Theory of Approximation Methods
for Convolution Equations. Operator Theory: Advances and Applications, vol. 74.
Birkhäuser, Basel, 1994.
[13] R. Hagen, S. Roch, and B. Silbermann, C*-Algebras and Numerical Analysis. Pure
and Applied Mathematics, vol. 236. Marcel Dekker, New York, 2001.
[14] A.Yu. Karlovich, Fredholmness and index of simplest weighted singular integral op-
erators with two slowly oscillating shifts. Banach J. Math. Anal. 9 (2015) 24–42.
[15] A.Yu. Karlovich, Yu.I. Karlovich, and A.B. Lebre, Sufficient conditions for Fred-
holmness of singular integral operators with shifts and slowly oscillating data. Integr.
Equ. Oper. Theory 70 (2011) 451–483.
[16] A.Yu. Karlovich, Yu.I. Karlovich, and A.B. Lebre, Fredholmness and index of sim-
plest singular integral operators with two slowly oscillating shifts. Operators and
Matrices 8 (2014) 935–955.
[17] A.Yu. Karlovich, Yu.I. Karlovich, and A.B. Lebre, On a weighted singular integral
operator with shifts and slowly oscillating data. Complex Anal. Oper. Theory 10
(2016) 1101–1131.
 Semi-Fredholmness of Singular Integral Operators with Shifts 245

[18] A.Yu. Karlovich, Yu.I. Karlovich, and A.B. Lebre, One-sided invertibility criteria
for binomial functional operators with shift and slowly oscillating data. Mediterr. J.
Math. 13 (2016) 4413–4435.
[19] A.Yu. Karlovich, Yu.I. Karlovich, and A.B. Lebre, The index of weighted singular
integral operators with shifts and slowly oscillating data. J. Math. Anal. Appl. 450
(2017), 606–630.
[20] A.Yu. Karlovich, Yu.I. Karlovich, and A.B. Lebre, Necessary Fredholm conditions for
weighted singular integral operators with shifts and slowly oscillating data. J. Integr.
Equ. Appl. 29 (2017), 365–399.
[21] A.Yu. Karlovich, Yu.I. Karlovich, and A.B. Lebre, Criteria for n(d)-normality of
weighted singular integral operators with shifts and slowly oscillating data. Proc.
London Math. Soc. 116 (2018) 997–1027.
[22] Yu.I. Karlovich, An algebra of pseudodifferential operators with slowly oscillating
symbols. Proc. London Math. Soc. 92 (2006) 713–761.
[23] Yu.I. Karlovich, Nonlocal singular integral operators with slowly oscillating data. In
“Operator Algebras, Operator Theory and Applications”, eds. M.A. Bastos et al.
Operator Theory: Advances and Applications, vol. 181, 2008, pp. 229–261.
[24] H. Kumano-go, Pseudo-Differential Operators. The MIT Press. Cambridge, MA,
1982.
[25] M.A. Naimark, Normed Algebras. Wolters-Noordhoff Publishing, Groningen, 1972.
[26] I.P. Natanson, Theory of Functions of a Real Variable. Frederick Ungar Publishing
Co., New York, 1955.
[27] V.S. Rabinovich, Singular integral operators on a composed contour with oscillating
tangent and pseudodifferential Mellin operators. Soviet Math. Dokl. 44 (1992) 791–
796.
[28] V.S. Rabinovich, Mellin pseudodifferential operators techniques in the theory of sin-
gular integral operators on some Carleson curves. In “Differential and Integral Op-
erators (Regensburg, 1995)”, eds. I. Gohberg et al. Operator Theory: Advances and
Applications, vol. 102, 1998, pp. 201–218.
[29] V.S. Rabinovich, S. Roch, and B. Silbermann, Limit Operators and Their Applica-
tions in Operator Theory. Operator Theory: Advances and Applications, vol. 150.
Birkhäuser, Basel, 2004.
[30] S. Roch, P.A. Santos, and B. Silbermann, Non-Commutative Gelfand Theories. A
Tool-kit for Operator Theorists and Numerical Analysts. Springer, Berlin, 2011.
[31] D. Sarason, Toeplitz operators with piecewise quasicontinuous symbols. Indiana Univ.
Math. J. 26 (1977) 817–838.
[32] I.B. Simonenko and Chin Ngok Minh, Local Approach to the Theory of One-
Dimensional Singular Integral Equations with Piecewise Continuous Coefficients.
Noethericity. Rostov University Press, Rostov-on-Don, 1986 (in Russian).
[33] E.M. Stein, Harmonic Analysis: Real-Variable Methods, Orthogonality, and Oscilla-
tory Integrals. Princeton Univ. Press, Princeton, NJ, 1993.
[34] K. Yosida, Functional Analysis. 6th ed. Springer, Berlin, 1980.
246 A.Yu. Karlovich, Yu.I. Karlovich and A.B. Lebre

Alexei Yu. Karlovich

Centro de Matemática e Aplicações
Departamento de Matemática
Faculdade de Ciências e Tecnologia
Universidade Nova de Lisboa
Quinta da Torre
2829–516 Caparica, Portugal
e-mail: [email protected]
Yuri I. Karlovich
Centro de Investigación en Ciencias
Instituto de Investigación en Ciencias Básicas y Aplicadas
Universidad Autónoma del Estado de Morelos
Av. Universidad 1001, Col. Chamilpa
C.P. 62209 Cuernavaca, Morelos, México
e-mail: [email protected]
Amarino B. Lebre
Centro de Análise Funcional, Estruturas Lineares e Aplicações
Departamento de Matemática
Instituto Superior Técnico
Universidade de Lisboa
Av. Rovisco Pais
1049–001 Lisboa, Portugal
e-mail: [email protected]
Operator Theory:
Advances and Applications, Vol. 267, 247–259

c Springer International Publishing AG, part of Springer Nature 2018

Factorization of Singular Integral

Operators with a Carleman
Backward Shift: The Vector Case
Amarino B. Lebre and Juan S. Rodrı́guez

p (p ∈ (1, ∞), n ∈
Abstract. On the vector Lebesgue space on the unit circle Ln
N) we consider singular integral operators with a Carleman backward shift of
linear fractional type, of the form
TA,B = A P+ + B P−
with
A = aI + bU, B = cI + dU,
where a, b, c, d ∈ Ln×n
∞ , P± = 12 (I ± S) are the Cauchy projectors in Ln p de-
fined componentwise, and U is an involutory shift operator associated with
the given Carleman backward shift also defined componentwise. By general-
ization to the vector case (n > 1) of the previously obtained results for the
scalar case (n = 1), it is shown that whenever a certain 2n × 2n matrix func-
tion, associated with the original singular integral operator, admits a bounded
factorization in L2n
p the Fredholm characteristics of the paired operator TA,B
can be obtained in terms of that factorization, in particular the dimensions
of the kernel and of the cokernel.
Mathematics Subject Classification (2010). Primary 47G10; Secondary 47A68.
Keywords. Singular integral operators, singular integral operators with shifts,
backward linear fractional shift, Carleman shift.

1. Introduction
Let T denote the unit circle in the complex plane, denote by L∞ := L∞ (T) the
algebra of essentially bounded functions on T, and for p ∈ [1, ∞) let Lp := Lp (T)
denote the Banach space of functions which are absolutely integrable to the pth
power on T.

This work was partially supported by the Fundação para a Ciência e a Tecnologia (Portuguese
Foundation for Science and Technology) through the project UID/MAT/04721/2013 (Centro de
Análise Funcional, Estruturas Lineares e Aplicações).
248 A.B. Lebre and J.S. Rodrı́guez

Two previous papers [4, 6] were devoted to the study of scalar singular in-
tegral operators with a linear fractional backward Carleman shift acting on the
Lebesgue space Lp := Lp (T), p ∈ (1, ∞), of the general form
TA,B = A P+ + B P− . (1.1)
with
A = aI + bU, B = cI + dU, (1.2)
where a, b, c, d ∈ L∞ , P± = 1
2 (I± S) are the Cauchy projectors in Lp associated
with the singular integral operator with Cauchy kernel S, which is defined by
!
1 1
Sϕ(t) = P.V. ϕ(τ ) dτ , (1.3)
πi T τ − t
and U is an involutive operator on Lp (U 2 = I) which anticommutes with S,
U S = −SU. (1.4)
As a model of the operator U we considered a weighted shift operator associated
with a linear fractional backward Carleman shift (see Section 2 for details).
The main results of [6] are the following: (1) under certain conditions a sin-
gular integral operator TA,B of the form (1.1) admits a factorization as a product
of three operators with the outer factors in that representation being invertible;
(2) such a representation can be obtained from a factorization of a matrix function
C ∈ L2×2
∞ associated with the original operator; (3) one can characterize both the
kernel and the cokernel of the operator, which depend on the partial indices of the
factorization of C as well as on two numbers in the set {−1, 1}, which can also be
obtained from the factorization of C.
From the results obtained in [6] it is clear that, even in the case where the
coefficients are scalar functions, the operator TA,B has a vector behaviour, in
the sense that this operator is equivalent, after extension, to a singular integral
operator with (2 × 2) matrix coefficients (cf. relation (3.2)). Therefore, it is rather
natural to ask what happens if we consider matrix coefficients from the beginning.
Although the question is natural the answer to this question took some time and
required the deduction of a special representation, presented in Proposition 3.2,
which proved to be the missing piece of the puzzle.
In recent years a series of papers by V. Didenko and B. Silbermann were pub-
lished devoted to the characterization of the kernel and cokernel of some Toeplitz
plus Hankel operators (see [2] and the references therein). The classes considered
are characterized by a special relation satisfied by the original coefficients and their
actions under the shift. It is shown that, in the case of Fredholmness, the defect
numbers can be obtained in terms of the generalized factorizations of two scalar
functions in Lp . We do not pursue this line of research here, though in [6, Sec-
tion 7] some Toeplitz plus Hankel operators were considered as an application of
the results obtained.
Supposing that the coefficients a, b, c, d ∈ Ln×n
∞ and the operators P± and U
are defined componentwise, we consider the operator TA,B given by (1.1) defined
 Factorization of SIO’s with a Carleman Backward Shift 249

on Lnp . We will show that the above-mentioned results from [6] are valid in the
new context, with the matrix function C now belonging to L∞ 2n×2n
. Also, apart
from to the partial indices of C we now need 2n numbers in the set {−1, 1}. This
question is addressed in Section 3. In the next section we give some background to
the subject, recalling in particular the main tool used throughout the paper, the
notion of factorization of a matrix function.

2. Preliminaries
The study of singular integral operators and its connection with the factorization
of matrix functions has a very long and rich history. Throughout that development
and referring only to the case of operators without shift under consideration (i.e.,
b = d = 0 in (1.2)) acting on the Lebesgue space Lnp , the concept of factorization
of matrix functions evolved in a manner that tied it to the Fredholm property of
the original operator. Let us recall the basic results in this context, which can be
found in [1] or [9].
−
For p ∈ (1, ∞) set q = p/(p − 1) and let L+ p = im P+ , Lp = im P− ⊕ C. Let
m ∈ N. An invertible matrix function C ∈ Lm×m ∞ is said to admit a generalized
factorization in (or relative to) Lm
p if
(a) it can be represented in the form
C = C+ Λ C− ,
−1 −1
where C+ ∈ (L+
p)
m×m
, C+ ∈ (L+ q )
m×m
, C− ∈ (L−q )
m×m
, C− ∈ (L−p)
m×m
, and
Λ = diag{t , . . . , t }, where κ1 ≥ · · · ≥ κm are integers,
κ1 κm
−1
(b) The operator B acting according to the rule Bϕ = C+ P− C+ ϕ is either
m
a bounded linear operator on Lp or a densely defined operator with a continuous
extension to Lmp .
The numbers κj , j = 1, . . . , m, which are uniquely determined by the matrix
function C (when ordered as above), are called the partial indices of C.
In what follows we shall restrict our analysis to the case where the generalized
factorization of C is bounded, i.e., when the factors C± belong to Lm×m∞ , in which
case the operator B above is automatically bounded. This also allows us to give
meaning to certain operators written in terms of the factors of the factorization of
C (see below).
We have,
Theorem 2.1. Let A, B ∈ Lm×m
∞ and consider in Lm
p the operator

SA,B = A P+ + B P− .
Then SA,B is a Fredholm operator on Lm
p if and only if the following conditions
hold:
(i) A, B are invertible in Lm×m
∞ ,
(ii) C = A−1 B admits a generalized factorization in Lm
p .
250 A.B. Lebre and J.S. Rodrı́guez

Moreover, when SA,B is a Fredholm operator on Lm p it can be factorized, by

p of the matrix function C, in the form
means of the factorization in Lm
SA,B = (AC+ ) S I,Λ SC −1 ,C− . (2.1)
+

The operator on the left of the right-hand side of the above representation is just
a multiplication operator, corresponding to a singular integral operator with equal
coefficients AC+ = SAC+ ,AC+ . Observe that this way of writing SA,B is meaningless
in the general case, but if we suppose that C has a bounded factorization, as we do,
then both of the outer operators in the above representation, AC+ and SC −1 ,C− ,
+
are bounded and invertible in Lp . Hence, the Fredholm characteristics of SA,B
coincide with those of the middle factor SI,Λ . In fact, the following general result
holds.
Theorem 2.2. Let A and B be invertible in Lm×m∞ and suppose that C = A−1 B
admits a factorization in Lp , say C = C+ Λ C− . Then
m

 
dim ker SA,B = κj , codim im SA,B = − κj .
κj >0 κj <0

A very special case occurs if all the partial indices are equal to zero (thus
Λ = I), in which case the factorization C is said to be canonical, and corresponds
to the invertibility of the singular integral operator SA,B .

On the unit circle we consider a Carleman backward shift α of linear fractional

type, i.e., the function α is either of the form
t−β
α(t) = , |β| > 1, t ∈ T, (2.2)
βt−1
or
τ
α(t) = , t, τ ∈ T. (2.3)
t
The latter form with τ = 1 can be seen as the limit of the first expression as |β|
tends to infinity. The shift function (2.2) can be factorized as follows:
α = α+ t−1 α− (2.4)
with
t−β λt
α+ (t) = , α− (t) = , (2.5)
λ βt−1
+
where λ = i |β|2 − 1. The following property holds for arbitrary t ∈ T:
α± (α(t)) = α∓ (t).
The shift α has two fixed points t± on T, given by t± = 1±λ β
, and we have
= ±t± . When α√is of the form (2.3) the factorization (2.4) holds
α+ (t± ) = α− (t± )√
with α+ = α− ≡ τ and t± = ± τ , for any choice of the square root.
 Factorization of SIO’s with a Carleman Backward Shift 251

Associated with the shift function (2.2) or (2.3) we consider a weighted shift
operator U : Lp → Lp , defined by

(U ϕ) (t) = (χ ϕ)(α(t)) = χ −1 (t) ϕ(α(t)), (2.6)

where
χ (t) = tα−1
− (t), t ∈ T. (2.7)
The following properties of this function will be used frequently below:
(a) indT χ = 1, (b) χ χ α ≡ 1, (c) χ (t± ) = ±1. (2.8)
Here we have used the following notational convention:
xα = x ◦ α
for any function x defined on T, a convention that we adopt throughout the pa-
per. Taking into account the above-mentioned properties of the shift and related
functions it is now straightforward to verify that the weighted shift operator U
given by (2.6) is an involutive operator (U 2 = I) and that the anticommutation
relation (1.4) holds.
For general information about singular integral operators with shift we refer
the reader to [8] and [9].

3. Generalization to the vector case

In this section we generalize the results obtained in our previous work [4, 6] to the
case of matrix-valued coefficients, that is we suppose that the coefficients a, b, c, d
of the functional operators A and B in (1.2) belong to Ln×n ∞ , n > 1. Supposing
that the operators P± , U are defined componentwise in Lnp all the notation used in
those papers keeps its meaning with obvious changes. For instance, the constant
matrix E there must be replaced by
 
0 In
E= , (3.1)
In 0
where In is the identity matrix of order n.
As in the scalar case the starting point is the following: based on the well-
known operator relation
     
1 I I TA,B 0 I U
= SA,B , (3.2)
2 U −U 0 TA,
B I −U
 = a I − b U, B
where A  = c I − d U , and
 
a d
A= , B = EAα E, (3.3)
b α cα
252 A.B. Lebre and J.S. Rodrı́guez

to each singular integral operator with shift TA,B acting on Lnp of the form (1.1)
with coefficients in Ln×n
∞ we associate a singular integral operator without shift
SA,B acting on L2np and we write

SA,B = Φ(TA,B ) (3.4)

to express that correspondence. Roughly speaking, doubling the dimension of the
space is the price to pay for the elimination of the dependence on the shift. Notice
that the map Φ is well defined in the class of singular integral operators TA,B with
Ln×n
∞ coefficients, is injective, but is not a homomorphism. However, it enjoys a
useful property: if SÂ,B̂ = Φ(TÂ,B̂ ), SĂ,B̆ = Φ(TĂ,B̆ ) and, moreover, SÂ,B̂ SĂ,B̆ =
SA,B with SA,B = Φ(TA,B ), then TÂ,B̂ TĂ,B̆ = TA,B .
Moreover, notice that the coefficients A and B of the singular integral oper-
ator (without shift) SA,B = Φ(TA,B ) are connected in a special manner (cf. (3.3)),
and this connection is responsible for all further results that we will derive, as we
shall see.
Taking into account that the invertibility of A is necessary for the Fredholm
property of SA,B , recalling Theorem 2.1, the analysis of such a singular integral
operator SA,B is closely related with the factorization properties of the matrix
function C ∈ L2n×2n
∞ defined by
C = A−1 B = A−1 EAα E. (3.5)
Observe that the matrix function C does not uniquely determine the matrix func-
tion A, and that is the reason why we will refer to A as a representative of C.
The set of all C ∈ L2n×2n
∞ for which there exists an invertible matrix function
A ∈ L∞ 2n×2n
such that (3.5) holds will be denoted by L2n×2n
∞,α .
A very important feature of matrix functions in L2n×2n
∞,α is that, whenever it
is factorable in L2n
p , there exists a factorization of it such that the outer factors
are related in a particular way. This will be highlighted in Proposition 3.2 below,
but first we introduce some notation and derive an auxiliary result.
When dealing with a factorization in L2n p of a matrix function C ∈ L∞
2n×2n

it is sometimes convenient to consider the number  ≤ 2n of partial indices which

are pairwise distinct and we will write the central factor in the form
Λ = diag{tκ1 Id1 , . . . , tκ Id }, (κ1 > · · · > κ ) (3.6)
where di is the multiplicity of the partial index κi , i = 1, . . . , . The following result
is the vector analogue of [6, Proposition 2.3] and its proof can be established sim-
ilarly to the scalar case, bearing in mind well-known results about the connection
between two different factorizations of the same matrix function (see [9]).
Proposition 3.1. Let C ∈ L∞,α 2n×2n
, n ∈ N, and suppose that C = C+ ΛC− is a
generalized factorization of C in Lp . Write Λ as in (3.6) and set
2n



Λ± = diag{ακ±1 Id1 , . . . , ακ± Id }, κ1 > · · · > κ , dj = 2n.
j=1
 Factorization of SIO’s with a Carleman Backward Shift 253

Then
H = Λ+ (C− )α EC+
is a polynomial matrix function satisfying
Hα = Λ− Λ−1 H−1 ΛΛ−1
− (3.7)
with the following block structure
⎛ ⎞
H1 ∗ ∗
⎜ .. ⎟
H=⎝ 0 . ∗ ⎠, (3.8)
0 0 Hl
where each block Hi , i = 1, . . . , , in the main diagonal is a non-singular constant
matrix of order di , the multiplicity of the partial index κi , such that Hi2 = Idi .
It should be noted that the blocks Hi , i = 1, . . . , , are uniquely determined
by the matrix function C up to similarity.
Now let us prepare the result that allows us to relate the outer factors of a
factorization of C given by (3.5). This result is based on an idea of T. Ehrhardt
presented in [3] and already used in [5]. Suppose that C satisfies the hypotheses of
the above proposition, let H be the matrix function given there and let
L = diag{H1 , . . . , H }ΛΛ−1
− = diag{
χ κ1 H1 , . . . , χ κ H }.

Taking into account that Hj = Hj−1 , j = 1, . . . , , due to (3.8), we may write

H−1 ΛΛ−1
− = (I + K)L, (3.9)
where K is a nilpotent polynomial matrix function. Further, each Hj is similar to
a diagonal matrix with the numbers ±1 in the diagonal, and therefore there exists
a non-singular 2n × 2n matrix X such that
D = X −1 LX = diag{ε1 χ κ1 , . . . , ε2n χ κ2n } (3.10)
with εj ∈ {−1, 1}, j = 1, . . . , 2n. As the blocks Hj are uniquely determined by
the matrix function C up to similarity, the numbers εj , j = 1, . . . , 2n, are uniquely
determined by C up to reordering within the set of coefficients of the same power
of χ .
Proposition 3.2. Let C be given by (3.5), suppose that it satisfies the conditions
of Proposition 3.1 and let H be the matrix function given there. Then C admits a
representation of the form
C = X+ DE X− , (3.11)
1
where X+ = E(X−−1 )α E ∈ (L+p)
2n×2n
is given by X+ = C+ (I + K) 2 X , for X and
K as introduced above, and DE = DE with D given by (3.10).
Proof. We adapt the proof of [3, Theorem 3.2] to our case. Starting as in the scalar
p , C = C+ ΛC− , we establish that
case, from the given factorization of C in L2n
C = C+ H−1 ΛΛ−1 −1
− (C+ )α E, (3.12)
254 A.B. Lebre and J.S. Rodrı́guez

where H is the polynomial matrix function referred to in Proposition 3.1. Note

that, due to (3.12) and the properties of the factorization of the shift α, we have
H−1 ΛΛ−1− = (H
−1
ΛΛ−1 −1
− )α . Consequently, together with H
−1
ΛΛ−1
− = (I + K)L,
we also have
H−1 ΛΛ−1
− = L(I + Kα )
−1
, (3.13)
because LLα = I2n , due to χ χ α ≡ 1. Then KL = LK  with K = (I + Kα )−1 − I.
Using the method of induction one establishes that K L = LK
m  m for every m ∈ N.
The matrix functions (I + K) 1/2 
and (I + K) 1/2
= (I + Kα )−1/2
are well defined
by series expansions, which are finite due to nilpotency of K and K. Using these
expansions, it follows that

(I + K)1/2 L = L(I + Kα )−1/2 .

This, together with (3.9) and (3.13), implies that

H−1 ΛΛ−1
− = (I + K)
1/2
L (I + Kα )−1/2

and, using the diagonalizing matrix X for L, introduced before this proposition,
we also get
H−1 ΛΛ−1
− = (I + K)
1/2
X D X −1 (I + Kα )−1/2 ,
where D is the diagonal matrix function (3.10). Thus, substituting this relation in
(3.12), we obtain the representation of C as stated in the proposition. 

Proposition 3.2 constitutes the breakthrough needed to generalize our previ-

ous results. In fact, once this result is established, it is easy to see that the main
theorem about the factorization of the operator TA,B , [6, Theorem 4.2], maintains
its validity in the new context with the same type of proof. This allows us to rep-
resent the given operator TA,B as a product of three operators in the same class
where the outer operators in this representation are invertible.

Theorem 3.3. Let α be the linear fractional backward shift (2.2), let U be the
corresponding shift operator given by (2.6), and let TA,B be the singular integral
operator with shift as defined in (1.1)–(1.2) with coefficients a, b, c, d belonging to
∞ . Suppose that the matrix functions A and B, given by (3.3),
the algebra Ln×n
are such that A is invertible in L∞2n×2n
and that C = A−1 B admits a bounded
2n
factorization in Lp . Then TA,B admits the factorization

TA,B = TA1 ,B1 TA0 ,B0 TA2 ,B2 (3.14)

where each operator on the right-hand side is a singular integral operator with
backward shift α of the form (1.1), the outer operators TAj ,Bj , j = 1, 2, are in-
vertible in Lnp , and the central operator TA0 ,B0 has simple coefficients which can be
obtained from a representative of the anti-block-diagonal matrix function DE for
D given by (3.10).
 Factorization of SIO’s with a Carleman Backward Shift 255

It comes out from the proof of this theorem that the central operator TA0 ,B0 ,
which is therefore responsible for the Fredholm properties of the original operator
TA,B , is of a very special form. Indeed,
Φ(TA0 ,B0 ) = RP+ + ERα EP− = R(P+ + DE P− ),
so that R is a representative of the anti-block diagonal matrix function DE with
the diagonal matrix function D given by (3.10). Thus, as in the scalar case, we are
left with the problem of finding a representative of the matrix function DE and to
use it to characterize the central operator in the operator factorization (3.14). Let
us now analyze these problems in the new context, but first we recall the result
obtained in [6] for the scalar case.
Theorem 3.4 ([6, Theorem 6.2]). Let n = 1. Under the conditions of Theorem 3.3,
let D be the matrix function (3.10), D = diag{ε1 χ κ1 , ε2 χ κ2 }. Then
⎧
⎪
⎨ 0 if κ1 ≤ 0,
dim ker TA,B = κ1
[2 + 4 ]1−ε1
if κ2 ≤ 0 and κ1 > 0,
⎪
⎩ [ κ1 + 1−ε1 ] + [ κ2 + 1−ε2 ] if κ > 0,
2 4 2 4 2

and
⎧
⎪
⎨ 0 if κ2 ≥ 0,
codim im TA,B = [− κ22 + 1+ε2
4 ] if κ1 ≥ 0 and κ2 < 0,
⎪
⎩ [− κ2
2 + 1+ε 2 κ1
4 ] + [− 2 +
1+ε1
4 ] if κ1 < 0.

We recall that the above numbers κ1 and κ2 (κ1 ≥ κ2 ) are nothing but the
partial indices of C in L2p and that ε1 and ε2 are the numbers on the main diagonal
of H in (3.7) for the scalar case.
To proceed we will need to write some formulas where matrices of dimension
2n × 2n are written in terms of blocks of matrices with dimension 2 × 2, which
have a meaning in the context of the scalar case (n = 1). We then make the
following notational convention: we put the subscript 0 in every term of dimension
2 × 2, which corresponds to the scalar case. A similar convention will be used for
operators. According to this convention, for instance, we will denote by E0 the
matrix corresponding to E in (3.1) for the scalar case.
Let us denote by E the 2n × 2n diagonal matrix with n 2 × 2 blocks all equal
to E0 . Clearly, the matrices E and E are similar and so there exists a non-singular
matrix M such that
E = M−1 EM.
Now, suppose we are given a 2n × 2n diagonal matrix function D  in the form
of n blocks of 2 × 2 diagonal matrix functions,
 = diag{D(1) , . . . , D(n) },
D 0 0
(j)
and suppose in addition that each block D0 is of the form (3.10) for n = 1. Using
(j) (j)
the results of [6, Section 5], we know how to find a representative R0 of (D0 )E0
256 A.B. Lebre and J.S. Rodrı́guez

(j) (j)
with (D0 )E0 = D0 E0 , yielding

D0 = (R0 )−1 E0 (R0 )α ,

(j) (j) (j)
j = 1, . . . , n.
 be the 2n × 2n matrix function with the n-block form
Let R
 = diag{R(1) , . . . , R(n) }.
R 0 0

Then
D  −1 ER
=R α (3.15)
and, consequently,
 = (MR)
D  −1 E(MR)  α, (3.16)
 
which means that MR is a representative of DE .
In general, the problem of finding a representative of the matrix function DE
with D given by (3.10) can be solved by this procedure. It amounts to setting:
 = M−1 DM.
D
In fact, from (3.16) it follows that
 −1 )−1 E(MRM
D = (MRM  −1 )α = R−1 ERα (3.17)
and therefore R = MRM  −1 is a representative of DE .
In terms of operators we have
SR,ERα E = R(P+ + DE P− )
 EM
= R(P+ + MD  −1 P− )
 EP
= RM(P+ + D  − )M−1
 + +D
= MR(P  EP
 − )M−1 = MS    M−1 .
R,E Rα E

Taking (3.15) into account, the operator SR,  

 E = R(P+ + DEP− ) is a direct sum
 ER
of n scalar operators, namely

n 
n
SR,  
 E = R(P+ + DEP− ) =
 ER
(j) (j)
R0 (P+ + (D0 )E0 P− ) = SR(j) ,E0 R(j) E0 .
0 0
j=1 j=1

Recalling that SR,ERE = Φ(TA0 ,B0 ), we conclude that TA0 ,B0 is equivalent to the
direct sum of n scalar operators

n
TA0 ,B0 ∼ TA(j) ,B (j)
0 0
j=1

where Φ0 (TA(j) ,B (j) ) = SR(j) ,E0 R(j) E0 , j = 1, . . . , n, Φ0 denoting the map (3.4) for
0 0 0 0
the scalar case. Consequently, its Fredholm properties can be derived from those
of these n scalar operators. Therefore we have proved the main result concerning
the properties of the operator TA,B in the vector case.
 Factorization of SIO’s with a Carleman Backward Shift 257

Theorem 3.5. Under the conditions of Theorem 3.3, we have


n 
n
dim ker TA,B = dim ker TA(j) ,B (j) , codim im TA,B = codim im TA(j) ,B (j) ,
0 0 0 0
j=1 j=1

where the dimension of the kernel and the codimension of the image of each
TA(j) ,B (j) , j = 1, . . . , n, are given by Theorem 3.4.
0 0

Finally, whenever needed, the above results can also be used to completely
characterize the kernel and the cokernel of the operator TA,B , since that charac-
terization for the scalar case is available in [6, Section 6].
In [6] we gave several examples of invertible scalar singular integral operators
with shift, of the class under consideration, which differ in the regularity of the
coefficients. These examples were constructed based on the knowledge of the repre-
sentatives of anti-diagonal 2 × 2 matrix functions, which can have different nature
depending on the parity of the partial indices of C ∈ L2×2 ∞ and on the sign of the
determinant of the corresponding polynomial 2 × 2 matrix function det H = ε1 ε2 ,
as follows:
– a rational matrix function, if the partial indices of C are of the same parity
and det H = −1;
– a continuous non-rational matrix function on T except at one point (one of
the fixed points of the shift), if the partial indices of C are of distinct parity;
– a piecewise rational matrix function on T with two points of discontinuity,
the fixed points of the shift α, if the partial indices of C are of the same parity
and det H = 1.
We finishing by giving an example of an invertible singular integral operator
with shift on L2p . According to the previous results, the defect numbers of such
an operator are determined by the central factor DE of the factorization given in
(3.11). So, we concentrate our attention on the case where the matrix function C
reduces to the central factor DE = DE. More precisely, suppose we want to identify
the operator TA,B on L2p , p ∈ (1, ∞), such that
Φ(TA,B ) = R(P+ + DE P− ),
where R, is a representative of DE with
⎛ ⎞
0 0 1 0
⎜ 0 0 0 1 ⎟
D = diag{χ , χ , χ , −χ −1 }, E=⎜
⎝ 1
⎟.
0 0 0 ⎠
0 1 0 0
Note that the (decreasing) partial indices of D are the tuple (1, 1, 1, −1). Using
the method described above, we have
⎛ ⎞ ⎛ ⎞
0 1 0 0 ..
⎜ 1 0 0 0 ⎟ ⎜ E0 . 0 ⎟
E = MEM = ⎜ ⎝ 0 0 0 1 ⎠=⎝
⎟ ··· ··· ⎠
..
0 0 1 0 0 . E0
258 A.B. Lebre and J.S. Rodrı́guez

with ⎛ ⎞
1 0 0 0
⎜ 0 0 1 0 ⎟
M = M−1 =⎜
⎝ 0
⎟,
1 0 0 ⎠
0 0 0 1
and
 = MDM = diag{D
D 2 }
1 , D
with
1 = diag{χ , χ },
D 2 = diag{χ , −χ −1 }.
D
Using the results of [6, Section 5], we know representatives for the 2 × 2 anti-
j E0 , j = 1, 2. In fact, setting
diagonal matrices D
   
 γ γα  1 −χ
R1 = χ α γα χ α γ , R2 = χα ,
1
where γ is the characteristic function of one of the two arcs of T connecting the
two fixed points of the shift α, it is straightforward to verify that, for each j = 1, 2,
the equation (R j )−1 E0 (R
 j )α = D
j holds, which means that R  j is a representative

of Dj E0 . From (3.17), the representative R of DE is R = Mdiag{R  1, R
 2 }M.
In terms of operators, we have,
Φ(TA,B ) = R(P+ + DEP− ) = RP+ + E(R)α EP−
 
R 1 (P+ + D1 E0 P− ) 0
=M  2 (P+ + D 2 E0 P− ) M
0 R
5 6
 1 P+ + E0 (R
R  1 )α E0 P− ) 0
=M  2 P+ + E0 (R  2 )α E0 P− ) M
0 R
which means that TA,Bj is (as expected) the direct sum of two scalar operators:
TA,B = TA1 ,B1 ⊕ TA2 ,B2
with
TA1 ,B1 = γ(I + χ U )P+ + γα (χ I + U )P−
and
TA2 ,B2 = (I + χ U )P+ + (I − χ U )P− .
Moreover, it follows from Theorem 3.4 that each of the scalar operators TA1 ,B1 and
TA2 ,B2 is invertible in Lp , and thus TA,B is invertible in L2p . In fact, we have for
the case of TA1 ,B1 : κ1 = κ2 = ε1 = ε2 = 1, and for the case of TA2 ,B2 : κ1 = ε1 = 1
and κ2 = ε2 = −1.
Observe then that in spite of the fact that the partial indices of D are all
nonzero the resulting singular integral operator is invertible, being the direct sum
of two scalar operators with different regularity of the coefficients: TA1 ,B1 with
discontinuous coefficients (the points of discontinuity are the fixed points of the
shift α) and TA2 ,B2 with polynomial and simple coefficients.
 Factorization of SIO’s with a Carleman Backward Shift 259

Acknowledgment
We would like to thank the anonymous referee for useful comments and remarks
about some inaccuracies in the original version, which helped to improve the final
version of this paper. We are also deeply grateful to our colleague Roger Picken,
who took the unfortunate task of correcting our bad sentences in English, his
mother tongue.

References
[1] K. Clancey and I. Gohberg, Factorization of Matrix Functions and Singular Integral
Operators. Operator Theory: Advances and Applications, vol. 3. Birkhäuser, Basel,
1981.
[2] V.D. Didenko and B. Silbermann, Generalized Toeplitz plus Hankel operators: kernel
structure and defect numbers. Complex Anal. Oper. Theory 10 2016, 1351–1381.
[3] T. Ehrhardt, Invertibility theory for Toeplitz plus Hankel operators and singular in-
tegral operators with flip. J. Funct. Anal. 208 (2004), 64–106.
[4] V.G. Kravchenko, A.B. Lebre, and J.S. Rodrı́guez, Factorization of singular integral
operators with a Carleman shift via factorization of matrix functions: the anticom-
mutative case. Math. Nachr. 280 (2007), 1157–1175.
[5] V.G. Kravchenko, A.B. Lebre, and J.S. Rodrı́guez, Matrix functions consimilar to
the identity and singular integral operators. Complex Anal. Oper. Theory 2 (2008),
593–615.
[6] V.G. Kravchenko, A.B. Lebre, and J.S. Rodrı́guez, Factorization of singular integral
operators with a Carleman backward shift: the case of bounded measurable coeffi-
cients. J. Anal. Math. 107 (2009), 1–37.
[7] V.G. Kravchenko and G.S. Litvinchuk, Introduction to the Theory of Singular Inte-
gral Operators with Shift. Kluwer Academic Publishers, London, 1994.
[8] G.S. Litvinchuk, Solvability Theory of Boundary Value Problems and Singular Inte-
gral Equations with Shift. Kluwer Academic Publishers, London, 2000.
[9] G.S. Litvinchuk and I.M. Spitkovskii, Factorization of Measurable Matrix Functions.
Operator Theory: Advances and Applications, vol. 25, Birkhäuser, Basel, 1987.

Amarino B. Lebre Juan S. Rodrı́guez

Centro de Análise Funcional, Centro de Análise Funcional,
Estruturas Lineares e Aplicações Estruturas Lineares e Aplicações
and and
Departamento de Matemática Departamento de Matemática, F.C.T.
Instituto Superior Técnico Universidade do Algarve
Universidade de Lisboa Campus de Gambelas
Av. Rovisco Pais 8005-139 Faro, Portugal
1049–001 Lisboa, Portugal e-mail: [email protected]
e-mail: [email protected]
Operator Theory:
Advances and Applications, Vol. 267, 261–284

c Springer International Publishing AG, part of Springer Nature 2018

Extension-Restriction Theorems for

Algebras of Approximation Sequences
Steffen Roch
Dedicated to Bernd Silbermann on the
occassion of his 75th birthday.

Abstract. The C ∗ -algebra S(T(C)) of the finite sections discretization for

Toeplitz operators with continuous generating functions is fairly well under-
stood. Since its description in [3], this algebra serves both as a source of
inspiration and as an archetypal example of an algebra generated by an dis-
cretization procedure. Moreover, it turns out that every separable C ∗ -algebra
of approximation sequences has a common structure with S(T(C)) after an
extension by compact sequences and a suitable fractal restriction. We explain
what this statement means and give a proof.
Mathematics Subject Classification (2010). Primary 65J10; Secondary 46L99,
47N40.
Keywords. Finite sections discretization, Toeplitz operators, fractal restric-
tion, compact extension, Silbermann algebras.

1. Introduction
Let H be a Hilbert space and P = (Pn )n≥1 a filtration on H, i.e., a sequence
of orthogonal projections of finite rank that converges strongly to the identity
operator on H. By F P we denote the set of all bounded sequences (An )n≥1 of
operators An ∈ L(im Pn ), and by G P the set of all sequences (An ) ∈ F P with
An  → 0. Provided with the operations
(An ) + (Bn ) := (An + Bn ), (An )(Bn ) := (An Bn ), (An )∗ := (A∗n )
and the norm (An ) := sup An Pn , F P becomes a unital C ∗ -algebra and G P is
a closed ideal of F P . The importance of the quotient algebra F P /G P in numerical
analysis stems from the fact that a coset (An ) + G P is invertible in F P /G P if and
only if the operators An are invertible for all sufficiently large n and if the norms
262 S. Roch

of the inverses are uniformly bounded. This is equivalent to saying that (An ) is a
stable sequence. Note in that connection that
(An ) + G P F P /G P = lim sup An . (1.1)
With every non-empty subset A of L(H), we associate the smallest closed and
symmetric subalgebra S P (A) of F P that contains all sequences (Pn APn )n≥1 with
A ∈ A. The algebra S(T(C)) of the finite sections discretization for Toeplitz oper-
ators with continuous generating function arises exactly in this way. Here, H is the
Hilbert space l2 (Z+ ), Pn is the projection which sends the sequence (x0 , x1 , . . .) to
(x0 , . . . , xn−1 , 0, 0, . . .), and A is the C ∗ -algebra T(C) generated by all Toeplitz
operators T (a) with a being a continuous function on the complex unit circle T.
Recall that T (a) is given by the matrix representation (ai−j )i,j≥0 with respect to
the standard basis of l2 (Z+ ), where ak denotes the kth Fourier coefficient of a.
(We agree to omit the superscript P when the filtration is specified in this way.)
The sequences in S(T(C)) are completely described in the following theorem,
discovered by A. Böttcher and B. Silbermann and first published in 1983 in their
paper [3] on the convergence of the finite sections method for quarter plane Toeplitz
operators (see also [4] and [6], Section 1.4.2). To state their results, let Rn stand
for the operator on l2 (Z+ ) which sends (x0 , x1 , . . .) to (xn−1 , . . . , x0 , 0, 0, . . .). It
is not hard to see that for each sequence A = (An ) ∈ S(T(C)), the strong limits
W (A) := s-lim An Pn and W # (A) := s-lim Rn An Rn Pn exist and that W and W # are
∗
unital -homomorphisms from S(T(C)) to L(l (Z )) (actually, to T(C)).
2 +

Theorem 1.1.
(a) The algebra S(T(C)) consists of all sequences (An )n≥1 of the form
(An ) = (Pn T (a)Pn + Pn KPn + Rn LRn + Gn ) (1.2)
where a ∈ C(T), K and L are compact operators on l2 (Z+ ), and (Gn ) ∈ G.
The representation of a sequence (An ) ∈ S(T(C)) in this form is unique.
(b) For every sequence A ∈ S(T(C)), the coset A + G is invertible in the quotient
algebra S(T(C))/G (equivalently, in F /G) if and only if the operators W (A)
and W# (A) are invertible.
Corollary 1.2. The quotient algebra S(T(C))/G is ∗ -isomorphic to the C ∗ -algebra
of all pairs
(T (a) + K, T (ã) + L) ∈ L(l2 (Z+ )) × L(l2 (Z+ )) (1.3)
with a ∈ C(T) and K, L compact. In particular, the mapping which sends the se-
quence (1.2) to the pair (1.3) is a ∗ -homomorphism from S(T(C)) onto S(T(C))/G
with kernel G.
Now we can give a first idea of what the statement in the abstract that “an
algebra has a common structure with S(T(C))” means. It is not hard to check
that the set J of all sequences (Pn KPn + Rn LRn + Gn ) with K, L compact
and (Gn ) ∈ G forms a closed two-sided ideal of S(T(C)). By Corollary 1.2, the
quotient J /G is naturally isomorphic to the product K(l2 (Z+ )) × K(l2 (Z+ )).
Thus, this quotient has two natural irreducible representations (K, L) → K and
 Extension-restriction Theorems 263

(K, L) → L which extend uniquely to irreducible representations of S(T(C))/G.

These extensions coincide (up to unitary equivalence) with the quotients W G and
# G of the homomorphisms W and W
W # by G (note that G lies in the kernel of both
W and W # ), and these quotients have the property described in assertion (b) of
Theorem 1.1.
To fix the latter property formally, we introduce the following notions. Given
a unital C ∗ -algebra A and a family W of unital ∗ -homomorphisms W from A to
certain unital C ∗ -algebras, we say that W is strictly spectral for A if an element
a ∈ A is invertible if and only if W (a) is invertible for every W ∈ W. We say that
W is spectral for A if an element a ∈ A is invertible if and only if W (a) is invertible
for every W ∈ W and if supW ∈W (W (a))−1  < ∞. Clearly, both notions coincide
if W is a finite family.
We recall further that a C ∗ -algebra A is called elementary if it is isomorphic
to the algebra of the compact operators on some Hilbert space, and that it is a
dual algebra if it is isomorphic to a direct sum of elementary algebras. If J a closed
ideal of A which is elementary (respective dual) when considered as a C ∗ -algebra,
then we call J an elementary (respective a dual) ideal of A. It is easy to see that
every dual ideal J is generated (as a C ∗ -algebra) by its elementary ideals, Kt with
t ∈ T , say. Since every closed ideal of J is also a closed ideal of A, J can be
identified with the smallest closed ideal of A which contains all elementary ideals
Kt . See [2] for an overview on dual algebras.
Now we can make the statement that a C ∗ -subalgebra A of F has a “common
structure structure with S(T(C))” more precise: It means that A contains G, that
the quotient A/G contains an ideal J which is dual, and that the extensions of
the irreducible representations of J /G to A/G form a spectral family for A/G.
The motivation for this paper came from the observation made again and
again over the years since Silbermann’s paper [19] that many concrete algebras
of approximation sequences have a “common structure with S(T(C))”. This ob-
servation appeared (at least to the author) as a big miracle, since neither were
the discretized operators very close to Toeplitz operators, nor had the used dis-
cretization procedures something in common with the fairly simple idea of taking
finite sections. To get an impression, here is a very incomplete list of papers from
different fields where this phenomenon occurs: [3, 5, 7–9, 11]. The explanation of
that fact proposed in the present paper is that at least every separable subalgebra
of F has a “common structure with S(T(C))”, after suitable extension by compact
sequences and suitable fractal restriction.
It should be mentioned that the proposed solution is by no means the only
way to explain the “miracle”: In [18], Santos and Silbermann singled out properties
of the operator algebra to be discretized (e.g., generated by a shift) and of the
discretization mapping (e.g., its positivity, a nice behavior of quasicommutators)
which ensure that the resulting algebras have a common structure with S(T(C)).
The paper is organized as follows. The phrases “fractal restriction” and “ex-
tension by compact sequences” are explained in the following two sections, followed
264 S. Roch

by a section which studies “fractal algebras of compact sequences”. The goal of

these sections is to introduce the language and to provide some facts for later
reference without proofs. The heart of the paper is Section 5 where two versions
of extension-restriction theorems are derived.

2. Fractal restriction
The idea behind the notion of a fractal algebra comes from a remarkable property
of the algebra S(T(C))/G: the structure of this algebra is determined by two
representations W and W # , and these representations are defined by certain strong
limits. A consequence of this “limit form” is that the operators W (A) and W # (A)
can be determined from each subsequence of the sequence A ∈ S(T(C)). This
observation implies that whenever a subsequence of a sequence A ∈ S(T(C)) is
stable, then the operators W (A) and W # (A) are already invertible and, hence, the
full sequence A is stable by Theorem 1.1.
One can state this observation in a slightly different way: every sequence in
S(T(C)) can be recovered from each its subsequence up to a sequence tending
to zero in the norm. In that sense, the essential information on a sequence in
S(T(C)) is stored in each of its subsequences. Subalgebras of F with this property
were called fractal in [17] in order to emphasize exactly this self-similarity aspect.
We will recall some basic properties of fractal algebras that will be needed in what
follows and start with the official definition of a fractal algebra. (Note that this
definition also makes sense in a more general context when F is the direct product
and G the direct sum of a sequence of C ∗ -algebras.)
Let SN denote the set of all strictly increasing sequences η : N → N. Given
η ∈ SN , we let Fη denote the set of all subsequences (Aη(n) ) of sequences (An ) in
F . One can make Fη to a C ∗ -algebra in a natural way. The mapping Rη : F →
Fη , (An ) → (Aη(n) ) is called the restriction of F onto Fη . For every subset S of
F , we abbreviate Rη S by Sη . It is easy to see that Gη coincides with the ideal of
the sequences in Fη which tend to zero in the norm.
Let A be a C ∗ -subalgebra of F . A ∗ -homomorphism W from A into a C ∗ -
algebra B is called fractal if, for every η ∈ SN , there is a mapping Wη : Aη → B
such that W = Wη Rη |A . Thus, the image of a sequence in A under a fractal
homomorphism can be reconstructed from each of its (infinite) subsequences. It is
not hard to check that Wη is a ∗ -homomorphism again. The homomorphisms W
and W # defined in Section 1 are archetypal examples of a fractal homomorphism.

Definition 2.1.
(a) A C ∗ -subalgebra A of F is called fractal if the canonical homomorphism
π : A → A/(A ∩ G), A → A + (A ∩ G) is fractal.
(b) A sequence A ∈ F is called fractal if the smallest C ∗ -subalgebra of F which
contains the sequence A and the identity sequence is fractal.
Here are some equivalent characterizations of fractal algebras.
 Extension-restriction Theorems 265

Theorem 2.2.
(a) A C ∗ -subalgebra A of F is fractal if and only if the implication
Rη (A) ∈ Gη ⇒ A ∈ G (2.1)
holds for every sequence A ∈ A and every η ∈ SN .
(b) If A is a fractal C ∗ -subalgebra of F , then Aη ∩ Gη = (A ∩ G)η for every
η ∈ SN .
(c) A C ∗ -subalgebra A of F is fractal if and only if the algebra A + G is fractal.
In many instances, the following theorem offers a comfortable way to check
the fractality of a specific subalgebra of F (for example, the fractality of S(T(C)),
where the corresponding homomorphism W is the product of fractal homomor-
phisms W and W # ).

Theorem 2.3. A unital C ∗ -subalgebra A of F is fractal if and only if there is a

unital and fractal ∗ -homomorphism W from A into a unital C ∗ -algebra B such
that, for every sequence A ∈ A, the coset A + (A ∩ G) is invertible in A/(A ∩ G)
if and only if W (A) is invertible in B.

The following results from [17] give a first impression of the power of fractality.

Proposition 2.4. Let A be a unital fractal C ∗ -subalgebra of F and A = (An ) ∈ A.

Then,
(a) the sequence A is stable if and only if it possesses a stable subsequence.
(b) the limit limn→∞ An  exists and is equal to A + G.

Thus, the upper limit in (1.1) is in fact a limit if the sequence (An ) belongs
to a fractal algebra. A similar improvement can be observed for the convergence
of certain spectral quantities. We will need the following notions.
Let (Mn )n∈N be a sequence of non-empty compact subsets of the complex
plane. The upper limit lim sup Mn (also called the partial limiting set) resp. the
lower limit lim inf Mn (or the uniform limiting set) of the sequence (Mn ) consists
of all points x ∈ C which are partial limits resp. the limit of a sequence (mn ) of
points mn ∈ Mn . The upper and lower limits of a sequence (Mn ) coincide if and
only if this sequence converges with respect to the Hausdorff metric
h(L, M ) := max{max dist (l, M ), max dist (m, L)}.
l∈L m∈M

Recall in this connection that non-empty compact subsets of C form a com-

plete metric space with respect to the Hausdorff distance and that the relatively
compact subsets of this space are precisely its bounded subsets.
We denote the spectrum of an element a of a unital algebra A by σA (a) or
σ(a). Recall also that an element a of a C ∗ -algebra is called normal if aa∗ = a∗ a.
Assertion (a) of the following result is the analog of the limsup-formula (1.1) for
norms; assertion (b) its improvement in the presence of fractality.
266 S. Roch

Proposition 2.5.
(a) If (An ) ∈ F is a normal sequence, then
lim sup σ(An ) = σF /G ((An ) + G).
(b) If (An ) is a normal sequence in a unital and fractal C ∗ -subalgebra of F , then
lim sup σ(An ) = lim inf σ(An ). (2.2)
(c) A normal sequence (An ) ∈ F is fractal if and only if (2.2) holds.
Similar results hold for other spectral quantities, for example for the se-
quences of the condition numbers, the sets of the singular values, the -pseudo-
spectra, and the numerical ranges of the An (see Chapter 3 in [6]). These results
indicate that, given an (in general non-fractal) sequence in F , it is of vital impor-
tance to single out (one of) its fractal subsequences. That this is indeed possible
is a consequence of the following fractal restriction theorem first proved in [12].
The proof given there was based on Proposition 2.5 (c) and is rather involved. A
much simpler proof employs the converse of Proposition 2.4 (b), which on its hand
follows easily from Theorem 2.2 (a).
Theorem 2.6. Let A be a separable C ∗ -subalgebra of F . Then there is an η ∈ SN
such that the restricted algebra Aη is fractal. In particular, every sequence in F
possesses a fractal subsequence.
One cannot expect that Theorem 2.6 holds for arbitrary C ∗ -subalgebras of
F ; for example it is certainly not true for the algebra l∞ . On the other hand, non-
separable fractal algebras exist: the finite sections algebra for Toeplitz operators
with piecewise continuous generating function can serve as an example.

3. Compactness and essential fractality

The ideal J of the finite sections algebra S(T(C)) is evidently related with compact
operators. We will make this relation precise by introducing an ideal K of F
consisting of sequences of compact type. The role of this ideal in numerical analysis
is comparable with the role of the ideal of the compact operators in operator theory.
Throughout this section we assume that F is a direct product of matrix algebras
Cδ(n)×δ(n) with dimension function δ : N → N tending to infinity.
Definition 3.1. Let K denote the smallest closed ideal of F which contains all
sequences (Kn ) ∈ F with rank Kn ≤ 1 for every n. We refer to the elements of K
as compact sequences and call a sequence in F a Fredholm sequence if it is invertible
modulo the ideal K.
Thus, a sequence (An ) ∈ F belongs to K if and only if, for every ε > 0, there
is a sequence (Kn ) ∈ F such that
sup An − Kn  < ε and sup rank Kn < ∞. (3.1)
n n
It is a simple consequence of this fact that G ⊆ K.
 Extension-restriction Theorems 267

We say that a sequence A ∈ F has finite essential rank if it is the sum of a

sequence in G and a sequence (Kn ) ∈ F with supn rank Kn < ∞. If A is of finite
essential rank, then there is a smallest integer r ≥ 0 such that A can be written as
(Gn )+(Kn ) with (Gn ) ∈ G and supn rank Kn = r. We call this integer the essential
rank of A and write ess rank A = r. If A is not of finite essential rank, then we put
ess rank A = ∞. Thus, the sequences of essential rank 0 are just the sequences in G.
In our running example, the intersection S(T(C)) ∩ K is just the ideal J , and
the essential rank of the sequence (Pn KPn +Rn LRn +Gn ) equals rank K +rank L.
Both the compactness and the Fredholm property of a sequence (An ) ∈ F
can be characterized in terms of the asymptotic behavior of the singular values of
the An ; see Sections 4.2 and 5.1 in [14] for the following results. We denote the
decreasingly ordered singular values of an n × n matrix A by
A = Σ1 (A) ≥ Σ2 (A) ≥ · · · ≥ Σn (A) ≥ 0 (3.2)
and set σk (A) := Σn−k+1 (A).
Theorem 3.2. The following conditions are equivalent for a sequence (Kn ) ∈ F :
(a) limk→∞ supn≥k Σk (Kn ) = 0;
(b) limk→∞ lim supn→∞ Σk (Kn ) = 0;
(c) (Kn ) is compact.
Corollary 3.3.
(a) A sequence (Kn ) ∈ F is of essential rank r if and only if
lim sup Σr (Kn ) > 0 and lim Σr+1 (Kn ) = 0.
n→∞ n→∞

(b) If (Kn ) ∈ K, then limn→∞ σk (Kn ) = 0 for every k.

Theorem 3.4. The following conditions are equivalent for a sequence (An ) ∈ F :
(a) (An ) is a Fredholm sequence.
(b) There are sequences (Bn ) ∈ F and (Jn ) ∈ K with supn rank Jn < ∞ such
that Bn An = In + Jn for all n ∈ N.
(c) There is a k ∈ N such that lim inf n→∞ σk+1 (An ) > 0.
The smallest non-negative integer k which satisfies condition (c) in the previ-
ous theorem is called the α-number α(A) of the Fredholm sequence A = (An ). It
corresponds to the kernel dimension of a Fredholm operator. Equivalently, α(A) is
the smallest non-negative integer k for which there exist a sequence (Bn ) ∈ F and
a sequence (Jn ) ∈ K of essential rank k such that Bn A∗n An = In + Jn for all n ∈ N.
Next we extend the concept of fractality and define fractality with respect to
the ideal of the compact sequences. For details see [16].
Definition 3.5. A C ∗ -subalgebra A of F is said to be essentially fractal (or K-
fractal) if the canonical homomorphism π K : A → A/(A ∩ K) is fractal and if
(A ∩ K)η = Aη ∩ Kη for each η ∈ SN .
In the same way one defines J -fractality with respect to an arbitrary closed
ideal J of F . If J = G, then the second condition in Definition 3.5 is automatically
268 S. Roch

satisfied by Theorem 2.2; thus, G-fractality coincides with fractality in the sense
of Definition 2.1.
The following result shows that essential fractality implies what one expects.
Theorem 3.6. Let A be an essentially fractal and unital C ∗ -subalgebra of F . A
sequence in A is compact (Fredholm) if and only if one of its subsequences is
compact (Fredholm).
Corollary 3.7. Let A be a fractal C ∗ -subalgebra of F . If (A ∩ K)η = Aη ∩ Kη for
each η ∈ SN , then A is essentially fractal.
Essential fractality has striking consequences for the behavior of the smallest
singular values of a Fredholm sequence.
Theorem 3.8. Let A be an essentially fractal and unital C ∗ -subalgebra of F . A
sequence (An ) ∈ A is Fredholm if and only if there is a k ∈ N such that
lim sup σk (An ) > 0. (3.3)
n→∞

Consequently, if a sequence (An ) in an essentially fractal and unital C ∗ -

subalgebra of F fails to be Fredholm, then
lim σk (An ) = 0 for each k ∈ N. (3.4)
n→∞
We call a sequence with that property not normally solvable, in analogy with the
corresponding notion from operator theory.
Corollary 3.9. Let A be an essentially fractal and unital C ∗ -subalgebra of F . Then
a sequence in A is either Fredholm or not normally solvable.
Example. The finite sections algebra S(T(C)) for Toeplitz operators is essen-
tially fractal, as follows easily from the description of S(T(C)) in Theorem 1.1
in combination with Corollary 3.7. The finite sections algebra S(BDO(N)) for
band-dominated operators examined in [14] is an example of an algebra which is
essentially fractal but not fractal. Finally, the sequence (An ) where

diag (0, 0, . . . , 0, 1) if n is even
An :=
diag (0, 1, . . . , 1, 1) if n is odd,
is fractal, but not essentially fractal: its subsequence (A2n ) is compact, but (A2n+1 )
is Fredholm.
As for fractality, there is an essential fractal restriction theorem (see [16]).
Its proof is based on the fact that, for every separable C ∗ -subalgebra A of F ,
there is a sequence η such that not only the sequence of the norms (Aη(n) )n≥1 =
(Σ1 (Aη(n) ))n≥1 converges for every (An ) ∈ A (which was basic in the proof of
the fractal restriction theorem), but also every sequence (Σk (Aη(n) ))n≥1 of the
singular values, for every k ∈ N.
Theorem 3.10. Let A be a separable C ∗ -subalgebra of F . Then there is an η ∈ SN
such that the restricted algebra Aη is essentially fractal.
 Extension-restriction Theorems 269

4. Fractal algebras of compact sequences

Compact sequences in fractal algebras behave particularly well. To state these
results, we need some more notions.
A non-zero element k of a C ∗ -algebra A is said to be of algebraic rank one if,
for each a ∈ A, there is a complex number μ such that kak = μk. We let C(A) stand
for the smallest closed subalgebra of A which contains all elements of algebraic
rank one. If such elements do not exist, we set C(A) = {0}. In any case, C(A) is
a closed ideal of A, the elements of which we call compact. The following theorem
summarizes some well-known equivalent descriptions of C(A) (see Theorem 1.4.5
in [1], and recall the notion of a dual algebra from the introduction).

Theorem 4.1. Let A be a unital C ∗ -algebra and J a closed ideal of A. The following
assertions are equivalent:
(a) J = C(J ).
(b) J is a dual algebra.
(c) The spectrum of every self-adjoint element of J is at most countable and
has 0 as only possible accumulation point.

Every dual ideal of a C ∗ -algebra comes with an associated lifting theorem

(see [19] for a first version of that theorem and [6] for a proof).

Theorem 4.2 (Lifting theorem for dual ideals). Let A be a unital C ∗ -algebra. For
every element t of a set T , let Jt be an elementary ideal of A such that Js Jt = {0}
whenever s = t, and let Wt : A → L(Ht ) denote the irreducible representation of
A which extends the (unique up to unitary equivalence) irreducible representation
of Jt . Let further J stand for the smallest closed ideal of A which contains all
ideals Jt .
(a) An element a ∈ A is invertible if and only if the coset a + J is invertible in
A/J and if every operator Wt (a) is invertible in L(Ht ).
(b) The separation property holds, i.e., Ws (Jt ) = {0} whenever s = t.
(c) If j ∈ J , then Wt (j) is compact for every t ∈ T .
(d) If a + J is invertible in A/J , then all operators Wt (a) are Fredholm and all
but a finite number of them is invertible.

A basic observation of [13] is that a compact sequence in a fractal algebra has

the spectral property of Theorem 4.1 (c). The following result is thus an immediate
consequence of that theorem. It implies in particular that every unital and fractal
C ∗ -subalgebra of F which contains non-trivial compact sequences is subject to the
lifting theorem.

Corollary 4.3. Let A be a unital and fractal C ∗ -subalgebra of F which contains the
ideal G. Then the ideal (A ∩ K)/G of A/G is a dual algebra.
270 S. Roch

5. Restriction-extension theorems
5.1. Weights of elementary algebras of sequences
A projection in a C ∗ -algebra is a self-adjoint element p with p2 = p. A closed
ideal J of a C ∗ -algebra A lifts projections, if every projection in A/J contains a
representative which is a projection in A. Closed ideals of C ∗ -algebras do not lift
projections in general (take A = C([0, 1]) and J = {f ∈ A : f (0) = f (1) = 0}).
The following proposition states that elementary ideals of F /G lift projections.
More general, every dual ideal of a C ∗ -algebra owns the projection lifting property.
Proposition 5.1. Let B be an elementary C ∗ -subalgebra of F /G.
(a) Every projection p ∈ B lifts to a sequence (Πn ) ∈ F of orthogonal projections,
i.e., (Πn ) + G = p.
(b) If p and q are rank one projections in B which lift to projections (Πpn ) and
(Πqn ) in F , respectively, then dim im Πpn = dim im Πqn for all sufficiently
large n.
Thus, for large n, the numbers dim im Πpn are uniquely determined by the
algebra B; they do neither depend on the rank one projection p nor on the choice of
its lifting. For a precise formulation of that property, define an equivalence relation
∼ on the set of all sequences of non-negative integers by calling two sequences
(αn ), (βn ) equivalent if αn = βn for all sufficiently large n. Then Proposition 5.1
states that the equivalence class which contains the sequence (dim im Πpn )n≥1 is
uniquely determined by the algebra B. We denote this equivalence class by αB
and call it the weight of the elementary algebra B. We say that B is of positive
weight if αB contains a sequence of positive numbers, and B is of weight one if αB
contains the constant sequence of ones. Note that the weight is bounded if B is in
K/G; in this case (Πpn ) is a compact sequence of projections and therefore of finite
essential rank.
5.2. Silbermann pairs and J -Fredholm sequences
Next we are going to examine the consequences of the lifting theorem in the context
of Silbermann pairs.
A Silbermann pair (A, J ) consists of a unital C ∗ -subalgebra A of F and a
closed ideal J of A such that G ⊆ J ⊆ K, G = J and J /G is a dual algebra. Every
sequence in A which is invertible modulo J is called a J -Fredholm sequence. Note
that every J -Fredholm sequence is a Fredholm sequence in sense of Definition
3.1 (but a Fredholm sequence in A need not be J -Fredholm because J may be
properly contained in A ∩ K).
Under the hypotheses of Corollary 4.3, (A, A ∩ K) is a Silbermann pair, and
a sequence in A is (A ∩ K)-Fredholm if and only if it is Fredholm. The study of
Silbermann pairs (in the special case when J /G is an elementary subalgebra of
K/G) was initiated by Silbermann in [20].
Let (A, J ) be a Silbermann pair. Being dual by definition, the algebra J /G is
the direct sum of a family (It )t∈T of elementary algebras with associated bijective
 Extension-restriction Theorems 271

representations Wt : It → K(Ht ). These representations extent to irreducible

representations A → L(Ht ) which we denote by Wt again. In this context, the
Lifting theorem 4.2 specifies as follows.
Theorem 5.2. Let (A, J ) be a Silbermann pair.
(a) A sequence A ∈ A is stable if and only if it is J -Fredholm and if the operators
Wt (A) are invertible for each t ∈ T .
(b) Ws (It ) = {0} whenever s = t.
(c) If J ∈ J , then Wt (J) is a compact operator for every t ∈ T .
(d) If A ∈ A is J -Fredholm, then all operators Wt (A) are Fredholm and all but
a finite number of them is invertible.
For every elementary ideal It , let (αtn ) be a representative of the weight αIt .
Assertion (d) of Theorem 5.2 implies that for every J -Fredholm sequence A ∈ A,
the sum 
αn (A) := αtn dim ker Wt (A) (5.1)
t∈T
is finite. This definition obviously depends on the choice of the representatives
of the weights. But the equivalence class of the sequence (αn (A)) modulo ∼ is
uniquely determined, since only a finite number of items in the sum (5.1) is not
zero.
A basic phenomenon of a J -Fredholm sequence (An ) is the following splitting
property of the (increasingly ordered) singular values σk (An ) of An .
Theorem 5.3. Let (A, J ) be a Silbermann pair and A = (An ) ∈ A a J -Fredholm
sequence. Then A is a Fredholm sequence, and
lim σαn (A) (An ) = 0 whereas lim inf σαn (A)+1 (An ) > 0. (5.2)
n→∞ n→∞

The proof makes use of results on lifting of families of mutually orthogo-

nal projections and on generalized (or Moore–Penrose) invertibility. For details
see [13].
Theorem 5.3 has some remarkable consequences. First note that the number
α(A) := lim sup αn (A) (5.3)
n→∞

is well defined and finite for every J -Fredholm sequence A ∈ A. Since (αn (A)) is
a sequence of non-negative integers, it has a constant subsequence the entries of
which are equal to α(A). Together with (5.2), this shows that
lim inf σα(A) (An ) = 0 and lim inf σα(A)+1 (An ) > 0. (5.4)
n→∞ n→∞

Corollary 5.4. Let (A, J ) be a Silbermann pair and A ∈ A a J -Fredholm sequence.

Then A is a Fredholm sequence, and its α-number is given by (5.3).
Many subalgebras of F which arise from concrete approximation methods
have the property that every rank one projection in J /G lifts to a sequence of
projections of rank one (equivalently, that every elementary algebra It has weight
272 S. Roch

one). In this case we call (A, J ) a Silbermann pair of local weight one. For Sil-
bermann pairs with this property, Theorem 5.3 and its Corollary 5.4 specify as
follows.
Corollary 5.5. Let (A, J ) be a Silbermann pair of local weight one and A = (An ) ∈
A a J -Fredholm sequence. Then

α(A) = dim ker Wt (A), (5.5)
t∈T

and the sequence A has the α(A)-splitting property, i.e., the number of the singular
values of An which tend to zero is α(A).
5.3. Spectral Silbermann pairs
A Silbermann pair (A, J ) is called spectral or strictly spectral if the family {Wt }t∈T
of the lifting homomorphisms of (A, J ) is spectral or strictly spectral for the
algebra A/G, respectively. For strictly spectral Silbermann pairs, the assertions of
the lifting theorem can be completed as follows.
Theorem 5.6. Let (A, J ) be a strictly spectral Silbermann pair and A ∈ A. Then
(a) A is stable if and only if all operators Wt (A) are invertible;
(b) A + GF /G = maxt∈T Wt (A).
(c) A is J -Fredholm if and only if all operators Wt (A) are Fredholm and if there
are only finitely many of them which are not invertible;
(d) A ∈ J if and only if all operators Wt (A) are compact and if, for each ε > 0,
there are only finitely many of them with Wt (A) > ε.
Proof. Assertion (a) is a re-formulation of the strict spectral condition. Assertion
(b) is a consequence of (a) and of general properties of strictly spectral families
of homomorphisms; see Theorem 5.39 in [6] (where strictly spectral families were
called sufficient).
(c) The ‘only if’ part follows from the Lifting theorem 5.2 (d). Conversely, let
A ∈ A be a sequence for which all operators Wt (A) are Fredholm and for which
there is a finite subset T0 of T which consists of all t such that Wt (A) is not
invertible. Then all operators Wt (A∗ A) are Fredholm, and they are invertible if
/ T0 . Let t ∈ T0 . Then Wt (A∗ A) is a Fredholm operator of index 0. Hence,
t ∈
there is a compact operator Kt such that Wt (A∗ A) + Kt is invertible. Choose a
sequence Kt ∈ J with Wt (Kt ) = Kt and Ws (Kt ) = 0 for s = t (which ispossible
by the separation property in Theorem 5.2). Then the sequence K := t∈T0 Kt
belongs to the ideal J , and all operators Wt (A∗ A+K) are invertible. By assertion
(a), the sequence A∗ A + K is stable. Similarly, one finds a sequence L ∈ J such
that AA∗ + L is stable. Thus, the sequence A is invertible modulo J , whence its
J -Fredholm property.
(d) The ‘only if’ part follows again from the Lifting theorem 5.2 (c). For the ‘if’
part, let K ∈ A be a sequence such that, for every ε > 0, there are only finitely
many t ∈ T with Wt (K) > ε. For n ∈ N, let Tn stand for the (finite) subset of
 Extension-restriction Theorems 273

T which collects all t with Wt (K) > 1/n. For each t ∈ Tn , choose a sequence
Kt ∈ J with Wt (Kt ) = Wt (K) and Ws (Kt ) = 0for s = t (employ the separation
property in Theorem 5.2 again), and set Kn := t∈Tn Kt . Then Wt (K − Kn ) = 0
for t ∈ Tn and Wt (Kn ) = 0 for t ∈ / Tn . Hence, supt∈T Wt (K − Kn ) ≤ 1/n
for every n ∈ N. By Theorem 5.6 (b), the supremum in this estimate coincides
with K − Kn + GF /G . Thus, K is the norm limit of a sequence in J , whence
K ∈ J. 
An example for a strictly spectral Silbermann pair is (S(T(C)), J ), with J
specified as in the introduction. A sequence in S(T(C)) is Fredholm if and only
if its strong limit is a Fredholm operator (note that T (a) and T (ã) are Fredholm
operators only simultaneously). Equivalently, the sequence A := (Pn T (a)Pn +
Pn KPn + Rn LRn + Gn ) with a ∈ C(T), K, L compact and (Gn ) ∈ G is Fredholm
if and only if T (a) is a Fredholm operator. In this case,
α(A) = dim ker(T (a) + K) + dim ker(T (ã) + L). (5.6)
In particular, if K = L = 0 and if a is not the zero function, then
α(A) = dim ker T (a) + dim ker T (ã)
= max{dim ker T (a), dim ker T (ã)}.
The second equality holds by a theorem of Simonenko and Coburn which states
that one of the quantities dim ker T (a) or dim ker T (ã) is zero for each non-zero
Toeplitz operator.
5.4. Silbermann algebras
We call a unital fractal C ∗ -subalgebra A of F a Silbermann algebra if (A, A ∩ K)
is a spectral Silbermann pair; a Silbermann algebra A is called strict if the pair
(A, A ∩ K) is strictly spectral. The following result is an immediate consequence
of Theorem 5.6.
Theorem 5.7.
(a) Let A be a Silbermann algebra. A sequence A ∈ A is compact if and only if
all operators Wt (A) are compact and if, for each ε > 0, only a finite number
of them has a norm greater than ε.
Let A be a strict Silbermann algebra. A sequence A ∈ A is
(b) Fredholm if and only if all operators Wt (A) are Fredholm and if only a finite
number of them is not invertible;
(c) stable if and only if all operators Wt (A) are invertible.
The goal of this section is fractality and essential fractality properties of
Silbermann algebras, which we prepare by two technical results. The first one
describes the fractal algebras among the elementary subalgebras of K.
Proposition 5.8. Let J be a C ∗ -subalgebra of K which contains G properly and for
which the quotient algebra J /G is elementary. Then J is fractal if and only if
J /G is of positive weight.
274 S. Roch

Proof. First let J /G be of positive weight and let η ∈ SN . The mapping Wη : J →

Fη /Gη , J → Rη J + Gη is a ∗ -homomorphism which has the ideal G in its kernel.
Since J /G is elementary (thus, simple), either ker Wη = G or ker Wη = J . The
latter is impossible: the positivity of the weight implies that there is a sequence
in J which consists of non-zero projections, and no restriction of that sequence
can tend to zero in the norm. Thus, ker Wη = G, and the quotient homomorphism
Wηπ : J /G → Jη /Gη which sends J + G to Wη (J) is a ∗ -isomorphism between these
algebras. Define
πη : Jη → J /G, Rη J → (Wηπ )−1 (Rη J + Gη ).
Then πη Rη is the canonical homomorphism from J onto J /G, whence the frac-
tality of J follows.
Let now J be a fractal algebra and assume that J /G is not of positive weight.
Then the weight (αn ) of J /G contains infinitely many zeros, say αη(n) = 0 for a
certain η ∈ SN . Thus, every algebraic rank one projection p in J /G lifts to a
sequence (Πn ) of projections with Πη(n) = 0 for all n. Then (Πn ) ∈ G by Theorem
2.2, i.e., p = 0. This is impossible, since J contains G properly. 
The following lemma is immediate from the definition of a fractal homomorphism.
Lemma 5.9. Let A be a C ∗ -subalgebra of F , J a closed ideal of A with G ⊆ J ,
W : J → L(H) an irreducible representation of J with G ⊆ ker W , and W  : A →
L(H) the (unique) irreducible extension of W . If W is fractal, then W  is fractal.
We say that a Silbermann pair (A, J ) has positive or constant local weight
if every elementary component of J /G has a positive or constant weight, respec-
tively. We call a Silbermann algebra A of positive or constant local weight if the
Silbermann pair (A, A ∩ K) has the corresponding property. If A is a Silbermann
algebra of constant local weight, then the numbers αn (A) in (5.1) are indepen-
dent of n.
Theorem 5.10. Let (A, J ) be a strictly spectral Silbermann pair of positive local
weight. Then
(a) A is a fractal algebra,
(b) J = A ∩ K.
Proof. (a) Let Wt be the lifting of an elementary component Jt of J . We use
the notation Wt also for the extensions of Wt to irreducible representations of J
and A. Since (A, J ) is a strictly spectral Silbermann pair, the family {Wt }t∈T
of the lifting homomorphisms of J /G is strictly spectral for A/G. Moreover, all
homomorphisms Wt are fractal. Indeed, since every elementary component of J is
fractal by Proposition 5.8 and G ⊆ ker Wt , the representations Wt of Jt are fractal.
Lemma 5.9 implies that the Wt are also fractal as representations of the algebra
A. Now the fractality of A comes from Theorem 2.3 (consider the homomorphism
W : A → Πt∈T L(Ht ), a → (t → Wt (a))
which is unital and fractal and satisfies the invertibility condition in Theorem 2.3).
 Extension-restriction Theorems 275

(b) A is fractal by assertion (a), so (A ∩ K)/G is a dual algebra by Corollary 4.3.

Let {Wt }t∈T refer to the family of the lifting homomorphisms associated with that
algebra. Since J /G is a closed ideal of (A ∩ K)/G, the lifting homomorphisms of
J /G form a subset {Wt }t∈S of {Wt }t∈T , with S a non-empty subset of T .
Let A ∈ A be invertible modulo A ∩ K. By Theorem 5.6 (c), all operators
Wt (A) are Fredholm, and only a finite number of them is not invertible. Hence,
all operators Wt (A) with t ∈ S are Fredholm, and all but finitely many of them
are invertible. Again by By Theorem 5.6 (c), the sequence A is invertible modulo
J . Hence, invertibility modulo A ∩ K is equivalent to invertibility modulo J for
sequences in A, which implies A ∩ K = J . 
Corollary 5.11. If (A, J ) is a strictly spectral Silbermann pair of positive local
weight, then A is a strict Silbermann algebra.
Corollary 5.12. Let J be a C ∗ -subalgebra of F with the following properties:
(a) G ⊂ J ⊂ K,
(b) J /G is a dual algebra,
(c) J is of positive local weight, and
(d) at least one elementary component of J /G has infinite dimension.
Then the minimal unitization CI + J of J is fractal.
Indeed, by Theorem 5.10, we have to show that (CI + J , J ) is a strictly
spectral Silbermann pair. But this follows easily from the lifting theorem. 
Next we turn our attention to relations between a strict Silbermann algebra
A and its restriction Rη A = Aη .
Theorem 5.13. Let A be a unital and fractal C ∗ -subalgebra of F which contains
the ideal G and let η ∈ SN . Then the following assertions are equivalent:
(a) A is a strict Silbermann algebra.
(b) (Aη , (A ∩ K)η ) is a strictly spectral Silbermann pair of a positive local weight.
Proof. First we show that if A is fractal, then (Aη , (A ∩ K)η ) is a Silbermann
pair of a positive local weight. If A is fractal, its restriction Aη is fractal. Then,
by Corollary 4.3, (Aη ∩ Kη )/Gη is a dual algebra. Being a closed ideal of a dual
algebra, the algebra (A ∩ K)η is dual itself, and this algebra consists of compact
sequences in Fη only. Hence, (Aη , (A ∩ K)η ) is a Silbermann pair.
Let p be a minimal projection in (A ∩ K)η /Gη , and let (Pη(n) ) be a lifting of p
to a sequence of projections. Since Aη is fractal, this sequence is fractal. Theorem
2.2 (a) then implies that Pη(n) = 0 for all sufficiently large n. Hence, the weight of
the elementary component containing p is positive.
Next we show that for every fractal algebra A and every η ∈ SN , there is a
natural ∗ -isomorphism
ξη : A/G → Aη /Gη , A + G → (Rη A) + Gη .
Indeed, ξη is well defined since A − B ∈ G implies that (Rη A) − (Rη B) ∈ Gη ,
and ξη has a trivial kernel since Rη A ∈ Gη implies A ∈ G via Theorem 2.2. The
276 S. Roch

isomorphism ξη maps (A ∩ K)/G onto (A ∩ K)η /Gη . Hence, both ideals are canon-
ically isomorphic, which implies that if W is one of the lifting homomorphisms of
the Silbermann pair (A, A ∩ K), then W ξη−1 is a lifting homomorphism of the pair
(Aη , (A ∩ K)η ) and, conversely, if Wη is a lifting homomorphism of (Aη , (A ∩ K)η ),
then Wη ξη is a lifting homomorphism of (A, A ∩ K).
Now it is easy to see that the strictly spectral hypotheses in (a) and (b)
imply each other. For example, we verify that (b) ⇒ (a); the reverse implication
follows similarly. Let the Silbermann pair (Aη , (A ∩ K)η ) be strictly spectral, and
let {Wt }t∈T refer to the family of the lifting homomorphisms of the Silbermann
pair (A, A ∩ K). Let A ∈ A be a sequence for which all operators Wt (A + G) are
invertible. Then all operators Wt ξη−1 ξη (A + G) with t ∈ T are invertible. Since the
Wt ξη−1 run through the lifting homomorphisms of the pair (Aη , (A ∩ K)η ), and
since this pair is strictly spectral, the restricted coset ξη (A + G) is invertible in
Aη /Gη . Then A + G is invertible in A/G, since ξη is an isomorphism. 
Corollary 5.14. Any restriction of a strict Silbermann algebra is again a strict
Silbermann algebra.
Proof. Let A be a strict Silbermann algebra and η ∈ SN . Then (Aη , (A ∩ K)η ) is a
strictly spectral Silbermann pair, and every elementary component of (A∩K)η )/Gη
has positive weight by Theorem 5.13. By Corollary 5.11, Aη is a strict Silbermann
algebra. 
Corollary 5.15. Any strict Silbermann algebra is essentially fractal.
Indeed, the canonical homomorphism A → A/(A ∩ K) is fractal since A is
fractal and G is in the kernel of that homomorphism. Further, (A ∩ K)η = Aη ∩ Kη
by Theorem 5.10 (b). 
Corollary 5.16. Let A be a Fredholm sequence in a strict Silbermann algebra of con-
stant local weight. Then every restriction Aη of A is Fredholm, and the sequences
A and Aη have the same α-number.
5.5. Forcing the spectral property
Again, let F := F δ be an algebra of matrix sequences with dimension function δ
and G := G δ the associated ideal of zero sequences. We say that a C ∗ -subalgebra
Aext of F is an extension of a C ∗ -subalgebra A of F by compact sequences if there
is a subset K of the ideal K of the compact sequences in F such that Aext is the
smallest C ∗ -subalgebra of F which contains A and K . The goal of this section is
to prove the following result.
Theorem 5.17. Let A be a unital separable C ∗ -subalgebra of F . Then there is an
extension Aext of A by compact sequences and an η ∈ SN such that the restriction
Aext
η is a Silbermann algebra.
In other words, after extending A by a suitable set of compact sequences
and then passing to a suitable restriction, we arrive at a spectral Silbermann pair
η , Aη ∩ Kη ).
(Aext ext
 Extension-restriction Theorems 277

Proof. Let A0 be a countable dense subset of A containing the identity sequence.

Then the set A∗0 A0 is also countable and dense in A.
For each sequence A = (An ) in A0 , we write
A∗n An = En∗ diag (λ1 (A∗n An ), . . . , λδ(n) (A∗n An ))En (5.7)
with a unitary matrix En and increasingly ordered eigenvalues
0 ≤ λ1 (A∗n An ) ≤ · · · ≤ λδ(n) (A∗n An )
of A∗n An . For l, r ∈ N, let Kl,r,n be the δ(n) × δ(n)-matrix which is zero if
max{l, r} > δ(n) and which has a 1 at the lrth entry and zeros at all other entries
if max{l, r} ≤ δ(n). The sequence KA,l,r with entries KnA,l,r := En∗ Kl,r,n En (note
that En depends on A) is a sequence of matrices of rank at most 1, hence compact.
Let Aext stand for the smallest C ∗ -subalgebra of F which contains the algebra A,
the ideal G, and all sequences KA,l,r with A ∈ A0 and l, r ∈ N. This algebra
is also separable. Hence, by Theorem 2.6 and Theorem 3.10, there is an η ∈ SN
such that the restriction Aext
η is both fractal and essentially fractal. We claim that
r
Aη is a Silbermann algebra. Note that the sequences K
ext A,r
:= l=1 KA,l,l with
entries
KnA,r := En∗ diag (1, . . . , 1, 0, . . . , 0)En (5.8)
where r ones followed by δ(n) − r zeros belong to Aext .
To simplify notation, we will assume that η is the identity mapping (otherwise
replace δ by δ ◦ η in what follows). Let J := Aext ∩ K.
Since A is a C ∗ -subalgebra and J is a closed ideal of Aext , the algebraic sum
A + J is a C ∗ -subalgebra of Aext . This subalgebra contains A, G and all sequences
KA,l,r . Thus, Aext = A + J .
Since Aext is fractal, the ideal J /G is dual by Corollary 4.3. Let (It )t∈T
denote the set of its elementary components and let Wt : It → L(Ht ) stand for
the irreducible representation associated with It . As before, we will denote an
irreducible representation of It and its irreducible extensions to Aext /G and Aext
by the same symbol.
Let A ∈ A0 . We claim that the coset KA,1,1 + G = KA,1 + G is an algebraic
rank one projection in (Aext ∩ K)/G. Indeed, the entries KnA,1 are projection ma-
trices of rank one. Hence, for every positive sequence (Bn∗ Bn ) ∈ Aext , there is a
sequence (βn ) of complex numbers such that
KnA,1 Bn∗ Bn KnA,1 = βn KnA,1 for every n ∈ N.
The sequence (βn KnA,1 )n∈N is fractal, and βn is the largest singular value
of βn KnA,1 . By Proposition 2.4 (b), the sequence (βn ) is convergent. Since every
sequence in Aext is a linear combination of four positive sequences, we conclude
that, for every sequence C = (Cn ) ∈ Aext , there is a convergent sequence (γn ) of
complex numbers such that
KnA,1 Cn KnA,1 = γn KnA,1 for every n ∈ N.
278 S. Roch

Put γ := limn→∞ γn . Then KA,1 CKA,1 − γKA,1 ∈ G, which proves the

claim.
Since the elementary components of J /G are generated by algebraic rank
one projections, there is a t(A) ∈ T such that KA,1 + G ∈ It(A) . Since It(A) is an
ideal, the equality
KA,l,r = KA,l,1 KA,1,1 KA,1,r
implies that KA,l,r + G ∈ It(A) for every pair l, r ∈ N. In particular, all cosets
KA,r + G belong to It(A) . Since the cosets KA,l,l + G are linearly independent
algebraic rank one projections and the representation Wt(A) is irreducible, the
operators Wt(A) (KA,l,l ) form a linearly independent set of projection operators of
rank one. Hence, the Hilbert space Ht(A) has infinite dimension.
Since A0 is dense in A, the sequences A + K with A ∈ A0 and K ∈ Aext ∩ K
form a dense subset Aext
0 of Aext . Let B := A + K be a sequence of this form, for
which B B is not a Fredholm sequence (equivalently, B∗ B is not a J -Fredholm
∗

sequence, since J contains all compact sequences in Aext ). Then A∗ A = (A∗n An )

is not a Fredholm sequence, hence
lim λr (A∗n An ) = 0 for every r ∈ N (5.9)
n→∞

by (3.4) (recall (5.7) and remember that Aext is essentially fractal after restriction).
From (5.7)–(5.9) we conclude that A∗ AKA,r ∈ G for every r ∈ N, hence
Wt(A) (A∗ A)Wt(A) (KA,r ) = 0 for every r ∈ N. (5.10)

Since (Wt(A) (KA,r ))r≥1 is an increasing sequence of orthogonal projections on

Ht(A) , this sequence converges strongly, and its limit, P , is the orthogonal projec-
tion from Ht(A) onto the closure of the linear span of the union of the ranges of
the Wt(A) (KA,r ) (see, for example, Theorem 4.1.2 in [10]). So we conclude from
(5.10) that Wt(A) (A∗ A)P = 0. Thus, and by Theorem 5.2 (c),
Wt(A) (B∗ B)P = Wt(A) (A∗ A)P + Wt(A) (A∗ K + K∗ A + K∗ K)P
is a compact operator. Then Wt(A) (B∗ B) cannot be invertible: otherwise, the
projection P were compact, but the range of P has infinite dimension, which
follows by the same arguments as the infinite dimensionality of Ht(A) .
Thus, whenever B ∈ Aext 0 and B∗ B is not a Fredholm sequence, then one
of the operators Wt (B B) is not invertible. Conversely, if all operators Wt (B∗ B)
∗

with t ∈ T are invertible, then B∗ B is a Fredholm sequence. By Theorem 5.2 (a)

this implies that, whenever all operators Wt (B∗ B) with t ∈ T are invertible, then
the sequence B∗ B is a stable. This fact holds for all sequences B in the dense
subset Aext
0 of Aext , from which it is easy to conclude that the family (Wt )t∈T is
spectral for Aext . 

It is not clear if one can make Aext to a strict Silbermann algebra by a suitable
restriction. The point is that the implication, obtained at the end of the previous
 Extension-restriction Theorems 279

proof, is verified only for sequences in a dense subset of Aext . Another open ques-
tion is if there is a version of Theorem 5.17 which works without restriction if one
already starts with a fractal and essentially fractal separable subalgebra A of F .

5.6. Forcing local rank one

Our final goal is forcing the local rank one property by a suitable extension and re-
striction. The following result on restrictions of Silbermann pairs has been already
shown as part of the proof of Theorem 5.13.
Proposition 5.18. Let (A, J ) be a Silbermann pair (in F ) and η ∈ SN . Then
(Aη , Jη ) is a Silbermann pair (in Fη ).
Let J /G =∼ ⊕t∈T K(Ht ) be the representation of the dual algebra J /G as a
direct sum of elementary components. Since these components are simple, every
homomorphic image of K(Ht ) is either {0} or isomorphic to K(Ht ). Given η ∈ SN ,
let Tη denote the set of all t ∈ T which are not mapped to the zero ideal under
the homomorphism
(An ) + G → (Aη(n) ) + Gη . (5.11)
The following lemma says that the weights of J /G behave naturally under the
mapping (5.11), provided that J is fractal. Recall from Corollary 5.12 that the
latter condition holds if all elementary components of J /G have positive weight
and if at least one of them has infinite dimension.
Lemma 5.19. Let J be a fractal algebra and t ∈ Tη . Let Jt,η be the image of
the elementary component Jt ∼ = K(Ht ) of J /G under the homomorphism (5.11).
Then Jt,η ∼= K(H t ) is elementary, and the weight of Jt,η is the restriction αη of
the weight α of Jt .
Proof. Let (Aη(n) )+Gη be a projection of algebraic rank one in Jt,η . The fractality
of J ensures that the quotient homomorphism (5.11) is an isomorphism from J /G
to Jη /Gη and, thus, from Jt to Jt,η . Hence, the pre-image (An ) + G ∈ Jt of
(Aη(n) ) + Gη is uniquely determined, and it is a projection of algebraic rank one. It
is clear that if (Πn ) is a projection which lifts (An )+G, then (Πη(n) ) is a projection
which lifts (Aη(n) ) + Gη . This is the assertion. 

In the next result we will see how to force local constant weight (recall that
the latter means that every elementary component of J /G has constant weight).
Theorem 5.20. Let (A, J ) be a Silbermann pair and assume that J is separable
and fractal. Then there is an η ∈ SN such that the Silbermann pair (Aη , Jη ) has
constant local weight.
Proof. Since J is separable, the number of the elementary components of J /G is
at most countable. We enumerate these components by J (1) , J (2) , . . . and denote
the weight of J (i) by α(i) . Every weight α(i) is bounded and has, thus, a constant
subsequence. This fact is used in the following construction.
280 S. Roch

Starting with the identity mapping η0 on N, there is a subsequence η1 of

(1)
η0 such that the restriction αη1 is a positive constant. We continue in this way
(k)
and get, for every k ≥ 1, a subsequence ηk of ηk−1 such that the restriction αηk
is a positive constant. If the number of elementary components of J /G is finite,
we let η be the last of the sequences ηk obtained in this way; otherwise we set
(i)
η(n) := ηn (n) for n ∈ N. In each case, every restriction αη is a positive constant.
Then (Aη , Jη ) is a Silbermann pair by Proposition 5.18, and the local weights of
that pair are just the η-restrictions of the weights of (A, J ) by Lemma 5.19. Thus,
(Aη , Jη ) is a Silbermann pair with constant local weight. 

Let A be a separable Silbermann algebra with a constant local weight. Again,

we denote the elementary components of (A ∩ K)/G by J (1) , J (2) , . . . and write
α(i) for the weight of J (i) , which now can be identified with a non-negative integer.
For every i, let pi ∈ J (i) be a minimal (hence, algebraic rank one) projection. Given
(i)
pi , choose a projection (Πn )n≥1 ∈ A which lifts pi and which is specified such
(i) (i) α(i) (i,k)
that dim im Πn = α(i) for all n, and write Πn as a sum k=1 Πn of projections
(i,k) (i)
with dim im Πn = 1. For example, choose unitary matrices En such that
(i) ∗
Π(i) (i)
n = En diag (1, . . . , 1, 0, . . . , 0)(En ) ,

with α(i) of ones followed by δ(n) − α(i) of zeros. Then set

Π(i,k)
n := En(i) diag (0, . . . , 0, 1, 0, . . . , 0)(En(i) )∗ ,
(i,k)
with the 1 at the kth position. (Πn )n≥1 is a sequence of rank one projections
in F . Let à be the smallest closed subalgebra of F which contains the algebra A
and all sequences (Πn )n≥1 . Since à is again a separable C ∗ -algebra, the fractal
(i,k)

restriction theorem applies.

Theorem 5.21. Let A be a separable Silbermann algebra and let η ∈ SN be such that
Ãη is fractal. Then the algebra Ãη obtained in this way is a Silbermann algebra of
local weight one.
Proof. For brevity, we write A and à for Aη and Ãη , respectively, and set J :=
A ∩ K and J˜ := à ∩ K. Then à is both a fractal algebra and an extension of A
by compact sequences. In particular,
Ã/J˜ = (A + J˜)/J˜ ∼
= A/(A ∩ J˜) = A/(A ∩ K) = A/J .
We are going to relate the irreducible representations of A, associated with
the elementary components of J /G, with the irreducible representations of Ã,
associated with the elementary components of J˜/G.
Let W : J → L(K) be an irreducible representation of J associated with an
elementary component of J /G. Let x ∈ K be a cyclic vector. We extend W to an
irreducible representation W̃ of A by setting W̃ (a)W (j)x := W (aj)x for a ∈ A
and j ∈ J .
 Extension-restriction Theorems 281

Since G is in the kernel of W , the quotient of W̃ by G is well defined; we

denote it by W̃ again. Then there is an irreducible representation π : J˜ → L(H̃),
a closed subspace H of H̃ which is invariant with respect to π(J ), and a unitary
operator U : H → K such that
W (j) = U π(j)|H U ∗ for j ∈ J . (5.12)
∗
Set U x =: y ∈ H. Then y = 0; hence y is cyclic for π, and we can extend π
to an irreducible representation π̃ : Ã → L(H̃) by
π̃(a)π(j)y := π(aj)y for a ∈ Ã, j ∈ J˜.
From (5.12) we conclude that W (aj)x = U π(aj)|H U ∗ x for j ∈ J and a ∈ A;
hence,
W̃ (a)W (j)x = W (aj)x
= U π(aj)|H U ∗ x
= U π̃(a)π(j)|H y
= U π̃(a)|H U ∗ U π(j)U ∗ x
= U π̃(a)|H U ∗ W (j)x.
Since the vectors W (j)x lie dense in K, we conclude that H is an invariant
subspace for π̃(A) and W̃ (a) = U π̃(a)|H U ∗ for a ∈ A. Note that every π̃ obtained
in this way is an irreducible representation of J˜ with G in its kernel; so it is
associated with an elementary component of J˜/G.
Let (It )t∈T be the elementary components of J /G and Wt : A → L(Kt )
be the lifting homomorphism associated with It . For every t ∈ T , we construct
an extension of Wt to an irreducible representation of à as above. In general,
this extension will not be unique. For a given irreducible representation π̃ of Ã
obtained from an irreducible representation of J˜/G, we let Tπ̃ denote the set of
all t ∈ T for which the above extension of Wt leads to π̃, i.e., for which there are
a closed subspace Ht of H̃ invariant with respect to π̃(A) and a unitary operator
Ut : Ht → Kt such that
Wt (a) = Ut π̃(a)|Ht Ut∗ for a ∈ A. (5.13)
We claim that the Hilbert spaces Ht with t ∈ Tπ̃ are pairwise orthogonal. Let
s, t ∈ Tπ̃ and s = t. Inserting a ∈ Is into (5.13) we get 0 = Wt (a) = Ut π̃(a)|Ht Ut∗
whence π̃(Is )|Ht = 0. On the other hand, let 0 = xs ∈ Ks . Since Ws : Is → L(Ks )
is irreducible, xs is algebraically cyclic, i.e., Ws (Is )xs = Ks . With ys := Us xs we
obtain
Hs = Us∗ Ks = Us∗ Ws (Is )xs = π̃(Is )Us xs = π̃(Is )ys .
Let now zs ∈ Hs and zt ∈ Ht . By the previous line, there is a js ∈ Is such
that zs = π̃(js )ys . Then
zs , zt  = π̃(js )ys , zt  = ys , π̃(js )∗ zt  = ys , π̃(js∗ )zt  = 0
because of π̃(Is )|Ht = 0 as observed above. This proves the claim.
282 S. Roch

Now we can finish the proof of the theorem. Let à ∈ à be a sequence such
that all operators π̃(Ã) are invertible and the norms of their inverses are uniformly
bounded by a constant C, where π̃ runs through the irreducible representations
of à coming from the elementary components of J˜/G. Write à as A + J̃ where
A ∈ A and J̃ ∈ J˜ (recall that à is an extension of A by compact sequences).
Since J˜ is dual, there are only finitely many π with π(J̃) ≥ 1/(2C). Thus, all
operators π̃(A) are Fredholm, only a finite number of them is not invertible, and
the norms of their inverses are uniformly bounded.
From (5.13) we then conclude that all operators Wt (A) are Fredholm, that
only a finite number of them is not invertible, and that the norms of their inverses
are uniformly bounded. Then all operators Wt (A∗ A) are Fredholm with index 0,
only a finite number of them is not invertible, and the norms of their inverses
are uniformly bounded. Thus, there is a sequence J ∈ J such that all operators
Wt (A∗ A+J) are invertible and the norms of their inverses are uniformly bounded;
the (finite) sum J of the (compact) orthogonal projections onto the kernels of
Wt (A∗ A) will do the job. Because A is a Silbermann algebra, the sequence A∗ A+J
is invertible. In particular, A∗ A is a Fredholm sequence. Analogously one shows
that AA∗ is a Fredholm sequence. But then A itself is a Fredholm sequence in
A, which implies that A + J̃ = Ã is a Fredholm sequence in Ã, hence invertible
modulo J˜. By the lifting theorem, applied to the Silbermann pair (Ã, J˜), the
sequence à is stable. Thus, à is a Silbermann algebra. 
If one starts with a strict Silbermann algebra A, then this construction yields
a strict Silbermann algebra Ãη of local weight one.
5.7. A few examples
Block Toeplitz operators. As in the introduction, we consider Toeplitz operators
T (a) on l2 (Z+ ) and their finite sections with respect to the filtration (Pn ), but
now the Toeplitz operators are generated by matrix-valued continuous functions
a : T → CN ×N . We denote the related algebra of the (full) finite sections discretiza-
tion for these operators by S(T(C N ×N )). It is not hard to derive the analogues
of Theorem 1.1 and Corollary 1.2 in this setting, with the sequences (Rn LRn )
replaced by the sequences (Rn Li Rn ) when n = kN + i with compact operators Li ,
and with the homomorphism W # replaced by the family of the N homomorphisms
#i (A) := s-limn→∞ RnN +i AnN +i RnN +i PnN +i
W
with i ∈ {0, 1, . . . , N − 1}. In particular, every restriction Sηi (T(C N ×N )) with
ηi (n) := nN + i is a strict Silbermann algebra (and has, in that sense, the same
structure as S(T(C))).
One-sided almost Mathieu operators. These are the operators
Hα, λ, θ := V−1 + V1 + aI : l2 (Z+ ) → l2 (Z+ )
where V1 and V−1 stand for the forward and backward shift operator, respectively,
and where a(n) := λ cos 2π(nα + θ) with real parameters α, λ and θ. Further we
 Extension-restriction Theorems 283

write APα for the smallest closed subalgebra of L(l2 (Z+ )) which contains all oper-
ators Hα, λ, θ with arbitrary λ and θ (but with fixed α) and all compact operators.
The algebra of the finite sections discretization of APα with respect to the same
filtration (Pn ) as before is denoted by S(APα ).
We only deal with the non-periodic case when α ∈ (0, 1) is irrational. Then
we write α as a continued fraction with nth approximant pn /qn such that
|α − pn /qn | < qn−2 .
It has been shown in [11, Theorem 5.1] that the restriction Sη (APα ) with η(n) := qn
is a strict Silbermann algebra (with exactly the same structure as S(T(C))).
Operators in Cuntz algebras. For N ≥ 2, let ON denote the smallest C ∗ -subalgebra
of L(l2 (Z+ )) which contains the operators

xr if k = rN + i
Si : (xk )k≥0 → (yk )k≥0 with yk := (5.14)
0 else
with i = 0, . . . , N − 1. These operators are isometries, and they satisfy the Cuntz
axiom
S0 S0∗ + · · · + SN −1 SN
∗
−1 = I. (5.15)
Since the (abstract) Cuntz algebra ON is simple, ON is ∗ -isomorphic to the (con-
crete = represented) algebra ON . We use the same filtration (Pn ) as before and
consider the smallest closed subalgebra S(ON ) of F = F P which contains all finite
sections sequences (Pn APn ) with A ∈ ON . It turns out that the algebra S(ON )
fails to be fractal, and it is a main result of [15] that η(n) := N n is a right choice
to make the restricted algebra Sη (ON ) fractal.

References
[1] W. Arveson, An Invitation to C ∗ -Algebras. Springer, 1976.
[2] M.C.F. Berglund, Ideal C ∗ -algebras. Duke Math. J. 40 (1973), 241–257.
[3] A. Böttcher, B. Silbermann, The finite section method for Toeplitz operators on the
quarter-plane with piecewise continuous symbols. Math. Nachr. 110 (1983), 279–291.
[4] A. Böttcher, B. Silbermann, Introduction to Large Truncated Toeplitz Matrices.
Springer, 1999.
[5] T. Ehrhardt, S. Roch, B. Silbermann, Finite section method for Toeplitz and singular
integral operators of operator valued piecewise quasicontinuous functions. Singular
Integral Operators and Related Topics (eds. A. Böttcher, I. Gohberg), Operator
Theory: Adv. Appl. 90, Birkhäuser, 1996, 204–243.
[6] R. Hagen, S. Roch, B. Silbermann, C ∗ -Algebras and Numerical Analysis. Marcel
Dekker, Inc., 2001.
[7] P. Junghanns, R. Kaiser, D. Potts, Collocation quadrature methods and fast sum-
mation for Cauchy singular equations with fixed singularities. Lin. Alg. Appl. 491
(2016), 187–238.
284 S. Roch

[8] P. Junghanns, B. Silbermann, Local theory of the collocation method for the approx-
imate solution of singular integral equations, I. Integral Equations Oper. Theory 7
(1984), 791–807.
[9] H. Mascarenhas, B. Silbermann, Sequences of variable-coefficient Toeplitz matrices
and their singular values. J. Functional Anal. 270 (2016), 4, 1479–1500.
[10] G.J. Murphy, C ∗ -Algebras and Operator Theory. Academic Press, Inc., 1990.
[11] V.S. Rabinovich, S. Roch, B. Silbermann, Finite sections of band-dominated oper-
ators with almost periodic coefficients. Modern Operator Theory and Applications
(eds. Y.M. Erusalimsky et al.), Operator Theory: Adv. Appl. 170, Birkhäuser, 2006,
205–228.
[12] S. Roch, Algebras of approximation sequences: Fractality. Problems and Methods in
Mathematical Physics, Oper. Theory: Adv. Appl. 121, Birkhäuser, 2001, 471–497.
[13] S. Roch, Algebras of approximation sequences: Fredholm theory in fractal algebras.
Studia Math. 150 (2002), 1, 53–77.
[14] S. Roch, Finite sections of band-dominated operators. Memoirs AMS 191, 895, Prov-
idence, R.I., 2008.
[15] S. Roch, Spatial discretization of Cuntz algebras. Houston J. Math. 36 (2010), 4,
1097–1132.
[16] S. Roch, Arveson dichotomy and essential fractality. Operator Theory, Pseudo-
Differential Equations, and Mathematical Physics (eds. Y.I. Karlovich, L. Rodino,
B. Silbermann, I.M. Spitkovski), Oper. Theory: Adv. Appl. 228, Birkhäuser, 2013,
325–342.
[17] S. Roch, B. Silbermann, C ∗ -algebra techniques in numerical analysis. J. Oper. The-
ory 35 (1996), 2, 241–280.
[18] P.A. Santos, B. Silbermann, An approximation theory for operators generated by
shifts. Numerical Functional Anal. Optimization 27 (2006), 451–484.
[19] B. Silbermann, Lokale Theorie des Reduktionsverfahrens für Toeplitzoperatoren.
Math. Nachr. 104 (1981), 137–146.
[20] B. Silbermann, Fredholm theory and numerical linear algebra. Recent Advances in
Operator Algebra and Its Applications, Oper. Theory: Adv. Appl. 160, Birkhäuser,
2005, 403–411.

Steffen Roch
Technische Universität Darmstadt
Fachbereich Mathematik
Schlossgartenstrasse 7
D-64289 Darmstadt, Germany
e-mail: [email protected]
Operator Theory:
Advances and Applications, Vol. 267, 285–315

c Springer International Publishing AG, part of Springer Nature 2018

Toeplitz and Hankel Algebras –

Axiomatic and Asymptotic Aspects
Steffen Roch and Bernd Silbermann

Abstract. In 1983, the authors introduced a Banach algebra of – as they called

them – Toeplitz-like operators. This algebra is defined in an axiomatic way;
its elements are distinguished by the existence of four related strong limits.
The algebra is in the intersection of Barria and Halmos’ asymptotic Toeplitz
operators and of Feintuch’s asymptotic Hankel operators.
In the present paper, we start with repeating and extending this ap-
proach and introduce Toeplitz and Hankel operators in an abstract and ax-
iomatic manner. In particular, we will see that our abstract Toeplitz opera-
tors can be characterized both as shift invariant operators and as compres-
sions. Then we show that the classical Toeplitz and Hankel operators on the
spaces H p (T), lp (Z+ ), and Lp (R+ ) are concrete realizations of our abstract
Toeplitz operators. Finally we generalize some results by Didas on derivations
on Toeplitz and Hankel algebras to the axiomatic context.

Mathematics Subject Classification (2010). Primary 47L80; Secondary 47B35.

Keywords. Toeplitz-like operators, abstract Toeplitz and Hankel operators,
Toeplitz algebras, quasicommutator ideals.

1. Introduction
When an operator A on a Banach space X bears the name Toeplitz operator then
it is usually distinguished by one of the following properties:
• A is the compression of a “nice” (e.g., a normal) operator onto a non-trivial
complementable closed subspace of X, or
• A owns a kind of shift-invariance.
For a concrete example, we recall the definition of the classical Toeplitz oper-
ators (as well as of their close relatives, the Hankel operators) on the Hardy space.
Let 1 < p < ∞. For a function f ∈ Lp (T), the Lebesgue space over the unit circle
286 S. Roch and B. Silbermann

T := {z ∈ C : |z| = 1} in the complex plane, we let

! 2π
1
ˆ
fn = f (eiϕ )e−inϕ dϕ
2π 0
denote its nth Fourier coefficient. It is well known that the Riesz projection
 
P : fˆn einϕ → fˆn einϕ ,
n∈Z n∈Z+

is bounded on L (T); its range is called the Hardy space H p = H p (T). We still set
p

Q := I − P and introduce the reflection operator

J : Lp (T) → Lp (T), (Jf )(t) := t−1 f (t−1 ).
Then J 2 = I and JP J = Q.
Let a ∈ L∞ (T). We write aI for the operator f → af of multiplication by a
on L (T). Then the Toeplitz operator T (a) and the Hankel operator H(a) on H p
p

are defined as the compressions

T (a) := P aI|im P = P aI|H p and H(a) := P aJ|imP = P aJ|H p ,
respectively. In particular, T (a) is the compression of the “nice” multiplication
operator aI onto the Hardy space. The function a is called the generating function
of both T (a) and H(a). The assignment a → T (a) is one-to-one, whereas a → H(a)
is not. Toeplitz and Hankel operators possess completely different properties, but
are nevertheless closely related by
T (ab) = T (a)T (b) + H(a)H(b̃), H(ab) = T (a)H(b) + H(a)T (b̃) (1.1)
∞
where c̃(t) := c(1/t) for c ∈ L (T). Note that JcJ = c̃I.
The function system {χn }n∈Z+ , χn (t) := tn (t ∈ T), forms a Schauder basis
of H p . The matrix representations of T (a) and H(a), a ∈ L∞ (T), with respect to
this basis are given by
(âj−k )∞
j,k=0 , (âj+k+1 )∞
j,k=1 ,

respectively, from which we conclude the other of the distinguishing properties of

Toeplitz operators, as follows. For every positive integer n, define
Vn : H p → H p , f → χn f and V−n : H p → H p , f → P (χ−n f ).
Then A is a Toeplitz operator on H p if and only if it is shift-invariant in the sense
that
A = V−n AVn for all positive n ∈ Z. (1.2)
This is the starting point for several lines of research. Many work has been
done to understand algebras generated by Toeplitz (and Hankel) operators. To
mention at least one result in that direction, let A be a closed subalgebra of
L∞ (T). We denote by alg T (A) the smallest closed subalgebra of L(H p ), the Ba-
nach algebra of the bounded linear operators on H p , which contains all Toeplitz
 Toeplitz and Hankel Algebras – Axiomatic and Asymptotic Aspects 287

operators with generating function in A. Further we write qc (T ) for the quasicom-

mutator ideal of alg T (A), i.e., for the smallest closed two-sided ideal of alg T (A)
which contains all operators of the form T (ab) − T (a)T (b) with a, b ∈ A.
Theorem 1.1.
(a) The algebra alg T (A) decomposes into the direct sum
alg T (A) = T (A) ⊕ qc (T ).
(b) The sequence
{0} → qc (T ) → alg T (A) → A → {0}
is short exact. Here, the second arrow stands for the embedding and the
third one for the canonical mapping from alg T (A) onto the quotient alge-
bra alg T (A)/qc (T ), which is isomorphic to A.
Further results along these lines, also for other kinds of Toeplitz operators,
can be found in [4, Section 2.40].
Barria and Halmos [2] took (1.2) as a starting point to initiate the study of
asymptotic Toeplitz operators. They call an operator A ∈ L(H 2 ) an asymptotic
Toeplitz operator if the sequence (V−n AVn )n≥0 converges strongly1 . If the strong
limit of that sequence exists, then it is necessarily a Toeplitz operator, whence the
notation. They also proved that every operator A in alg T H(L∞ ), the smallest
closed subalgebra of L(H 2 ) which contains all Toeplitz and Hankel operators with
generating functions in L∞ , is asymptotically Toeplitz. Since the strong limit of the
operators V−n AVn is a Toeplitz operator T (ϕ) with generating function ϕ ∈ L∞ ,
they arrived at a symbol map smb : A → ϕ. It is proved in [2] that the restriction
of smb to alg T H(L∞ ) is a contractive ∗ -homomorphism from alg T H(L∞ ) onto
L∞ . This symbol map obviously fulfills
smb T (a) = a for every a ∈ L∞ .
For the Toeplitz algebra alg T (L∞ ) ⊂ L(H 2 ), this result was already discovered
by R.G. Douglas [10] using different methods.
Using a similar strong limit, Feintuch [11] introduced asymptotic Hankel op-
erators. Both “asymptotic approaches” have the disadvantage that the sets of all
asymptotic Toeplitz (resp. Hankel) operators do not form algebras (see [2], Exam-
ple 13).
In 1983, we published a preprint [19] where we independently of [2] intro-
duced a class of asymptotic Toeplitz operators (called Toeplitz-like operators in
what follows). Besides the strong limit by Barria and Halmos, we required the
existence of three other strong limits (one of which is Feintuch’s, but note that
[11] appeared only in 1989) so that the set of these operators actually forms a
Banach algebra. The motivation for introducing that algebra came from the study
of the finite section method for operators belonging to the algebra generated by
1 Other kinds of convergence, e.g., norm, weak and Cesaro convergence, are considered as well.

Our focus is on strong convergence.

288 S. Roch and B. Silbermann

Toeplitz and Hankel operators with piecewise continuous generating functions. In

1985 we published our papers [20], [21] where we presented the results and related
applications.
The approach of [19] offers moreover the possibility to introduce Toeplitz and
Hankel operators in an abstract and axiomatic manner. Especially, the analogs of
(1.1) and of Theorem 1.1 can be proved for this class of abstract Toeplitz and
Hankel operators. In the present paper, we start with repeating and extending
this approach. In particular, we will see that our abstract Toeplitz operators can
be characterized both as shift invariant operators and as compressions. Then we
show that the classical Toeplitz and Hankel operators on the spaces H p (T), lp (Z+ ),
and Lp (R+ ) are concrete realizations of our abstract Toeplitz operators. Finally
we generalize some results by Didas [8] on derivations on Toeplitz and Hankel
algebras to the axiomatic context.

2. Toeplitz-like operators
Here we introduce, in an axiomatic way, the algebra TL(X) of the Toeplitz-like
operators on a Banach space X. For full proofs and some more details we refer to
[19, 20] where this algebra was introduced and studied for the first time. A few
proofs are sketched here to make the presentation more accessible.
2.1. The axioms
Let X be a Banach space and write L(X) for the Banach algebra of the bounded
linear operators on X. The identity operator on X is denoted by I. Further let
G = {0} be a subgroup of the additive group of the real numbers and G+ the
semigroup of its non-negative elements. Our construction of the algebra of the
Toeplitz-like operators is based on two families, V = (Vt )t∈G and R = (Rt )t∈G , of
bounded linear operators on X which are specified and related by the following
axioms:
(T1 ) The mappings t → Vt and t → V−t are semigroup homomorphisms on G+ .
In particular, Vs Vt = Vs+t and V−s V−t = V−(s+t) for all s, t ∈ G+ .
(T2 ) V−t Vt = I but Vt V−t = I for all t ∈ G+ \ {0}. Thus, for positive t, the
operators Vt are invertible only from the left-hand side.
By (T1 ) and (T2 ), V0 = I. Set Qt := Vt V−t and Pt := I − Qt for t ∈ G.
(T3 ) Rt2 = Pt and Vs Rt = Rs+t Pt for s, t ∈ G.
(T4 ) V−t → 0 strongly as t → ∞ and supt∈G+ {Vt , Rt } < ∞.
The following properties are easily verified.
Lemma 2.1.
(a) Pt2 = Pt and Q2t = Qt for t ∈ G.
(b) P−t = 0 and Q−t = I for t ∈ G+ .
(c) Pt → I and Qt → 0 strongly as t → ∞.
Thus, Pt and Qt are complementary projections.
 Toeplitz and Hankel Algebras – Axiomatic and Asymptotic Aspects 289

Lemma 2.2. For all s, t ∈ G,

(a) Vs Vt = Vs+t whenever s ≤ 0 or t ≥ 0.
(b) Vs Vt = Qs Vs+t = Vs+t Q−t .
(c) Ps Vs+t = Vs+t P−t .
(d) Ps Pt = Pmin{s,t} and Qs Qt = Qmax{s,t} .
Lemma 2.3. For all s, t ∈ G,
(a) Rt Pt = Pt Rt = Rt .
(b) Rt = 0 for t ≤ 0.
(c) Rt V−s = Pt Rs+t .
(d) Rs Rt = Vs−t Pt = Ps Vs−t .
Notice also that
sup{Vt , Pt , Qt , Rt } =: M < ∞. (2.1)
t∈G

We will see several concrete examples of operators Vt and Rt satisfying axioms

(T1 )–(T4 ) in Section 5. Note also that if the operators Vt and Rt satisfy these
axioms on a Banach space X, then the n × n diagonal matrices
diag (Vt , Vt , . . . , Vt ) and diag (Rt , Rt , . . . , Rt )
satisfy these axioms in place of Vt and Rt on the direct sum X n of n copies of X.

2.2. The algebra TL(X) of the Toeplitz-like operators

Given families (Vt )t∈G and (Rt )t∈G of operators subject to axioms (T1 )–(T4 ), let
TL(X) (with TL for Toeplitz-like) stand for the set of all operators A ∈ L(X) for
which the sequences (V−t AVt ), (Rt ARt ), (V−t ARt ) and (Rt AVt ) converge strongly
as t → ∞. Their strong limits are denoted by T (A), T (A), H(A) and H(A), 
respectively.
Example.
(a) The operators Vt and Rt belong to TL(X), and
T (Vt ) = Vt , T (Vt ) = V−t , H(Vt ) = Rt ,  t ) = 0.
H(V
The operators Rt lie in the kernel of each of the four strong limits.
(b) If Vt → 0 and Rt → 0 weakly as t → ∞, then every compact operator K on
X belongs to TL(X), and
T (K) = T (K) = H(K) = H(K)
 = 0.
The following is Theorem 1.1 in [20].
Theorem 2.4.
(a) TL(X) is a norm-closed subalgebra of L(X) which contains the identity op-
erator.
290 S. Roch and B. Silbermann

(b) The following identities hold for all A, B ∈ TL(X)

T (AB) = T (A)T (B) + H(A)H(B),

T (AB) = T (A)T (B) + H(A)H(B),



H(AB) = H(A)T (B) + T (A)H(B),


H(AB) = T (A)H(B)
 
+ H(A)T (B).
Theorem 2.5.
(a) T , T , H and H
 are bounded linear operators on TL(X) which map this algebra
into itself.
(b) The composition of any two of the operators T , T , H and H  belongs to
 
{T , T , H, H, 0}. In particular,
◦ T T H H 
T T T 0 0
T T T 0 0
H H H 0 0
H H H 0 0.
Proof. The linearity of the operators in (a) is evident; their boundedness is a
consequence of (2.1). So let us verify the first column of the table for example. A
basic ingredient are the identities collected in the following lemma.
Lemma 2.6. Let A ∈ TL(X) and s ∈ G+ . Then
(a) V−s T (A)Vs = T (A),
(b) Rs T (A)Rs = Ps T (A)Ps ,
(c) V−s T (A)Rs = H(A)Ps ,

(d) Rs T (A)Vs = Ps H(A).
Now let A ∈ TL(X) and write
T (A) = V−t AVt + Ct with Ct → 0 strongly as t → ∞.
Then, for every fixed s ∈ G+ ,
V−s T (A)Vs = V−s V−t AVt Vs + V−s Ct Vs = V−s−t AVs+t + V−s Ct Vs .
Letting t go to ∞ we arrive at identity (a) in Lemma 2.6, which on its hand implies
that T (T (A)) = T (A). In other words, T ◦ T = T .
Analogously, using (T3 ) and Lemma 2.3 (c) we write
Rs T (A)Rs = Rs V−t AVt Rs = Ps Rs+t ARs+t Ps + Rs Ct Rs ,
and passage to the strong limit as t → ∞ yields
Rs T (A)Rs = Ps T (A)Ps .
Letting s go to ∞ we obtain T (T (A)) = T (A) or T ◦ T = T . Similarly,
V−s T (A)Rs = V−s−t ARs+t Ps + V−s Ct Rs
 Toeplitz and Hankel Algebras – Axiomatic and Asymptotic Aspects 291

gives V−s T (A)Rs = H(A)Ps , whence H(T (A)) = H(A) and, finally,
Rs T (A)Vs = Ps Rs+t AVs+t + Rs Ct Vs

implies Rs T (A)Vs = Ps H(A)  (A)) = H(A).
and H(T  The other entries of the table
can be checked analogously. 

Theorem 2.7.
(a) T is a bounded projection on TL(X), and im T is a closed subspace of L(X).
(b) TL(X) = im T ⊕ ker T .
(c) ker T is a closed two-sided ideal of TL(X) and a left-sided ideal of L(X).
It is simple consequence of the last two identities in Theorem 2.4 that ker H
and ker H are closed subalgebras of TL(X).

2.3. Compact operators in TL(X)

As mentioned in Example (b) in Section 2.2, the compact operators belong to
TL(X) if Vs → 0 and Rs → 0 weakly. We will see now that the converse is also
true.
Theorem 2.8. The following assertions are equivalent:
(a) all operators of rank one are in TL(X);
(b) Vs → 0 and Rs → 0 weakly;
(c) all compact operators are in TL(X).
Proof. The only implication that needs a proof is (a) ⇒ (b). We prepare this proof
by a few facts.
Fact 1. The Rs do not converge strongly. Indeed, suppose the Rs converge strongly
to some operator R ∈ L(X). Then
V−s ARs → 0AR = 0 strongly for every A ∈ L(X).
In particular, H(A) = 0 for every A ∈ TL(X). But then Rs = H(Vs ) = 0, hence
Ps = Rs2 = 0 for every s ∈ G+ , which contradicts the strong convergence of Ps to
the identity.
Fact 2. Every rank one operator in TL(X) lies in ker T . Indeed, let Kx := x, yz
with y ∈ X ∗ and z ∈ X. Then
V−s KVs x = Vs x, yV−s z → 0
since V−s → 0 strongly and sup Vs  < ∞ by axiom (T4 ).
Fact 3. There is a constant c > 0 such that
c−1 Ps x ≤ Rs x ≤ cPs x for all x ∈ X.
Indeed, with c := sup Rs  we obtain
Rs x = Rs Ps x ≤ cPs x = cRs2 x ≤ c2 Rs x.
292 S. Roch and B. Silbermann

Now to the proof of the implication. By Fact 1, there is a z ∈ X such that the Rs z
do not converge. Clearly, z = 0. Consider the rank one operators
Ky x := x, yz with y ∈ X ∗ .
They belong to TL(X) by assumption and are, hence, in ker T by Fact 2. Since
ker T = ker T by Theorem 2.5, we conclude that
Rs Ky Rs x = Rs x, yRs z → 0 for all x ∈ X, y ∈ X ∗ . (2.2)
Since Ps z → z and by Fact 3, we get
Rs z ≥ c−1 Ps z ≥ (2c)−1 z
for s sufficiently large. Hence, (2.2) implies that Rs x, y → 0 for all x ∈ X and
y ∈ X ∗ ; in other words: Rs → 0 weakly.
The proof for Vs runs similarly. Now we use the inclusion ker T ⊆ ker H by
Theorem 2.5 to conclude that
Rs Ky Vs x = Vs x, yRs z → 0 for all x ∈ X, y ∈ X ∗
from which we obtain the weak convergence Vs → 0 as before. 
It is not by accident that both sequences in assertion (b) of the previous
theorem tend weakly to zero.
Proposition 2.9. Let (T1 )–(T4 ) be satisfied. Then
Vs → 0 weakly ⇔ Rs → 0 weakly.
Proof. We need the following fact, which is easy to check: If the sequence (At ) is
uniformly bounded and if At x, y → Ax, y for all x in a dense subset of X and
for all y ∈ X ∗ , then At → A weakly. Using this fact, we can prove the assertion as
follows.
First assume that Vs → 0 weakly. Then, for every r ∈ G+ and x ∈ X,
Rt Pr x, y = Vt−r Rr x, y → 0 as t → ∞
by (T3 ). Since Pt → I strongly, the elements of the form Pr x with r ∈ G+ and
x ∈ X lie dense in X. Thus, we conclude from the fact that Rt → 0 weakly.
Conversely, assume that Rs → 0 weakly. Then, for every r ∈ G+ and x ∈ X,
Vt Pr x, y = Rt−r Rr x, y → 0 as t → ∞
by Lemma 2.3 (d). Now the fact implies that Vt → 0 weakly. 

3. Abstract Toeplitz and Hankel operators

3.1. Symbols
Let Smb denote the quotient algebra TL(X)/ ker T , which is correctly defined
by Theorem 2.7, and let smb refer to the canonical homomorphism from TL(X)
onto Smb. We call smb A the symbol of A. The symbol of the identity operator
is denoted by e. The symbol of an operator carries important information. For
 Toeplitz and Hankel Algebras – Axiomatic and Asymptotic Aspects 293

example, we mention the following inverse closedness result from [20], the proof of
which makes use of the fact that ker T is a left-sided ideal of L(X).
Theorem 3.1. Let A ∈ TL(X) be invertible in L(X). Then A is invertible in TL(X)
if and only if smb A is invertible in Smb.
By Theorem 2.7 (a), every coset q = smb A ∈ Smb contains exactly one
operator from im T , namely T (A). We call this operator the abstract Toeplitz
operator with symbol q and denote it by T(q). Similarly, we call H(q) := H(T(q))
the abstract Hankel operator with symbol q. The mappings
T : Smb → im T , q → T(q) and H : Smb → im H, q → H(q) (3.1)
are linear by construction. Whereas T is also a bijection, H fails to be injective in
general. We will see in Corollary 3.20 that these mappings are bounded.
There is a natural involution q → q̃ on the symbol algebra Smb defined by
q̃ := smb T (T(q)).
Lemma 3.2.
(a) The mapping q → q̃ is an automorphism of Smb with q̃ = q.
(b) If M = 1 in (2.1), then ∼ is an isometry and qSmb = T(q)L(X) for all
q ∈ Smb.
It is not hard to see that T (T(q)) = T(q̃) and H(T(q))
 = H(q̃). The following
lemmas show that abstract Toeplitz and Hankel operators behave as the concrete
Toeplitz and Hankel operators introduced in the introduction and deserve, hence,
their name.
Lemma 3.3. Let p, q ∈ Smb. Then
T(pq) = T(p)T(q) + H(p)H(q̃), H(pq) = H(p)T(q̃) + T(p)H(q).
Proof. Choose A and B in TL(X) such that smb A = p and smb B = q. Then
pq = smb (AB), and the operators T(p), T(q) and T(pq) coincide with T (A),
T (B) and T (AB), respectively, by definition. Since A − T(p) and B − T(q) lie
in ker T , we conclude from Theorem 2.5 (b) that H(A) = H(T(p)) = H(p) and

H(B) 
= H(T(q)) = H(q̃). Hence, the assertion of the lemma is a consequence of
the identities in Theorem 2.4 (b). 
n n
Lemma 3.4. Let q1 , . . . , qn ∈ Smb and set A := i=1 T(qi ) and q := i=1 qi .
Then
T (A) = T(q), T (A) = T(q̃), H(A) = H(q), H(A)  = H(q̃).
There are several algebras that can be associated with abstract Toeplitz and
Hankel operators. The simplest ones are the Toeplitz algebras alg T(S) where S
is a closed subalgebra of the symbol algebra Smb. By definition, alg T(S) is the
smallest closed subalgebra of L(X) which contains all abstract Toeplitz operators
T(c) with c ∈ S. Clearly, alg T(S) is a closed subalgebra of TL(X).
294 S. Roch and B. Silbermann

Similarly, given closed subalgebras S1 and S2 of the symbol algebra Smb, let
alg TH(S1 , S2 ) stand for the smallest closed subalgebra of L(X) which contains all
abstract Toeplitz operators T(c1 ) with c1 ∈ S1 and all abstract Hankel operators
H(c2 ) with c2 ∈ S2 . If S1 = S2 =: S, then we simply write alg TH(S) in place of
alg TH(S, S). Algebras of this form are often referred to as Hankel algebras. By
Theorem 2.5, alg TH(S1 , S2 ) ⊆ TL(X). We will see in Section 5.1 that, in general,
alg TH(Smb) is properly contained in TL(X).
A third natural candidate, with a strong coupling between Toeplitz and Han-
kel operators, is the Toeplitz-plus-Hankel algebra alg TH+ (S) which is the small-
est closed subalgebra of L(X) which contains all Toeplitz-plus-Hankel operators
T (c) + H(c) with c ∈ S. Clearly, alg TH+ (S) ⊆ alg TH(S).

3.2. Quasicommutators and decompositions

The goal of this section is to understand the ideal ker T as a quasicommutator ideal.
A general approach to this circle of ideas is as follows. Let A and B be Banach
algebras and D : A → B a linear and bounded mapping. We write alg D(A) for
the smallest closed subalgebra of B which contains all elements D(a) with a ∈ A.
We say that D(A) generates B (as a Banach algebra) if alg D(A) = B.
The quasicommutator ideal generated by D is a measure for the deviation of D
from being multiplicative (i.e., an algebra homomorphism). This quasicommutator
ideal, denoted by qc (D), is the smallest closed ideal of alg D(A) which contains all
quasi-commutators D(ab) − D(a)D(b) with a, b ∈ A. Clearly, if D is multiplicative
then qc (D) = {0}, whereas qc (D) = alg D(A) if this algebra is simple and D
is not multiplicative. The following lemma provides an equivalent description of
qc (D) is terms of higher quasi-commutators. The proof is an easy exercise.
Lemma 3.5. Let A, B be Banach algebras and D : A → B a bounded linear mapping
such that D(A) generates B. Then qc (D) is equal to
clos spanB {D(a1 · · · an ) − D(a1 ) · · · D(an ) : n ∈ N, a1 , . . . , an ∈ A}.
In what follows we consider bounded linear mappings D : A → B which own
a homomorphic left inverse, i.e., we suppose there is a bounded homomorphism
W : B → A such that W (D(a)) = a for all a ∈ A. Mappings D with that property
are often called discretizations or quantizations, in order to emphasize their role in
numerical analysis or operator theory, respectively (see, e.g., Section 6.5.3 in [14]).
Theorem 3.6. Let A and B be Banach algebras, D : A → B a bounded linear
mapping, and W : B → A a bounded homomorphism such that W (D(a)) = a for
every a ∈ A. Then
(a) D is injective, D(A) is a closed subspace of B, and W is surjective,
(b) the algebra B decomposes into the direct sum
B = D(A) ⊕ ker W, (3.2)
(c) if D(A) generates B, then ker W = qc (D).
 Toeplitz and Hankel Algebras – Axiomatic and Asymptotic Aspects 295

(d) if D and W are contractions, then D is an isometry, the canonical projection

P : B → D(A) associated with the decomposition B = D(A) ⊕ ker W has
norm 1, and
D(a) = min D(a) + k for every a ∈ A. (3.3)
k∈ker W

Proof. (a) First we show that D(A) is a closed subspace of B. Let D(an ) be a
sequence in D(A) which converges to b ∈ B. Then the W (D(an )) = an form a
Cauchy sequence in A which converges to an element a ∈ A. Since D is bounded,
the D(an ) converge to D(a); hence b = D(a) ∈ D(A), and D(A) is closed. The
other assertions follow immediately from W ◦ D = idA .
(b) Writing b ∈ B as b = D(W (b)) + (b − D(W (b)) and noting that
W (b − D(W (b)) = W (b) − W (D(W (b))) = 0
we obtain B = D(A) + ker W . To see that this sum is direct, assume that D(a) ∈
ker W . Then a = W (D(a)) = 0, hence D(a) = 0.
(c) Since W (D(ab) − D(a)D(b)) = W (D(ab)) − W (D(a))W (D(b)) = 0 and qc (D)
is a closed ideal of B, the inclusion qc (D) ⊆ ker W is clear (and it holds without
the additional assumption in (c)).
For the reverse inclusion,
  let k ∈ ker W . Since D(A) generates B, there is a
sequence of elements kn = i j D(aijn ) with aijn ∈ A such that k − kn  → 0.
Using Lemma 3.5, we can write kn as

kn = D( aijn ) + qn = D(W (kn )) + qn = D(W (kn − k)) + qn
i j

with qn ∈ qc (D) (note that W (k) = 0). Then

k − qn  ≤ k − kn  + kn − qn  ≤ k − kn  + D(W (kn − k))
≤ (1 + D W ) k − kn  → 0.
Thus, k can be approximated as closely as desired by elements in qc (D). Since
the quasi-commutator ideal is closed, the assertion follows.
(d) Let a ∈ A and k ∈ ker W . Then a = W (D(a)) ≤ D(a) ≤ a and
D(a) = D(W (D(a))) = D(W (D(a) + k)) ≤ D(a) + k,
which implies assertion (d). 

The conditions in Theorem 3.6 are also necessary, as the following proposition
shows.
Proposition 3.7. Let the Banach algebra B = M ⊕ J be the direct sum of a closed
subspace M and a closed ideal J . Then there are a Banach algebra A, a bounded
linear mapping D : A → M such that D(A) = M, and a bounded homomorphism
W : B → A which is a left inverse for D. Moreover, J = ker W , and J = ker W =
qc (D) if M generates B.
296 S. Roch and B. Silbermann

Proof. Let P : B → M be the canonical projection associated with the decompo-

sition B = M ⊕ J , set A := B/J , and write W for the canonical homomorphism
from B onto A. Then the mapping D : A → M, a + J → P (a), is correctly
defined, linear and surjective, and
W ◦ D : a + J → P (a) → P (a) + J = a + J
is the identity mapping on A. Since M is closed and D is inverse to the (evidently
bounded and surjective) mapping W |M : M → A, D is bounded by the bounded
inverse mapping theorem. The equality J = ker W = qc (D) is a consequence of
Theorem 3.6. 
3.3. Algebras of abstract Toeplitz and Hankel operators
Next we are going to apply these general results in the context of (abstract)
Toeplitz and Hankel algebras. We start with the Toeplitz algebra alg T(S), with
S a closed subalgebra of Smb. Slightly modifying the above notation, we write
qcT (S) (instead of qc (T|S )) for the quasicommutator ideal of the mapping T :
Smb → L(X) restricted to S. Thus, qcT (S) is the smallest closed ideal of alg T(S)
which contains all operators T(cd)−T(c)T(d) with c, d ∈ S. Then the restriction of
the symbol homomorphism smb to alg T(S) is a left inverse to the linear mapping
T, restricted to S. In this context, Theorem 3.6 specifies as follows.
Corollary 3.8.
(a) qcT (S) = ker(smb |alg T(S) ) = ker(T |alg T(S) ).
(b) The algebra alg T(S) decomposes into the direct sum
alg T(S) = T(S) ⊕ qcT (S).
(c) The sequence
{0} → qcT (S) → alg T(S) → S → {0},
with the second arrow standing for the natural embedding and the third one
for the restriction of the symbol mapping, is exact.
In general, the Hankel algebras alg TH(S1 , S2 ) do not arise from a discretiza-
tion mapping in the above sense; in particular there is no natural quasicommutator
ideal. So we deal with the kernel of T instead of the quasicommutator ideal. The-
orem 2.7 specifies as follows to this context.
Corollary 3.9.
(a) alg TH(S1 , S2 ) = T(S1 ) ⊕ ker(T |alg TH(S1 , S2 ) ).
(b) The sequence
{0} → ker(T |alg TH(S1 , S2 ) ) → alg TH(S1 , S2 ) → S1 → {0},
where the second arrow stands for the embedding and the third one for the
restriction of the symbol mapping, is exact.
In this context, there is an equivalent description of the kernel of T (see [8]
for the H 2 -setting).
 Toeplitz and Hankel Algebras – Axiomatic and Asymptotic Aspects 297

Lemma 3.10. Let S1 ⊆ S2 . Then ker(T |alg TH(S1 , S2 ) ) is the smallest closed ideal of
alg TH(S1 , S2 ) which contains all Hankel operators H(q) with q ∈ S2 .
Proof. Abbreviate ker(T |alg TH(S1 , S2 ) ) by J1 and the closed ideal generated by
H(S2 ) by J2 . Since every Hankel operator is in ker T , it is immediate that J2 ⊆ J1 .
For the reverse inclusion, let A ∈ J1 . Then A can be represented as a norm limit
⎛ ⎞

αn γn  δn
A = lim ⎝ T(dij )⎠
(n) (n) (n) (n)
Ai H(bi )Ci +
n→∞
i=1 i=1 j=1

(n) (n) (n) (n)

with symbols dij ∈ S1 and bi ∈ S2 and with operators Ai , Ci ∈ TH(S1 , S2 ).
Since both A and all Hankel operators are in ker T and T is continuous, we con-
clude from Lemma 3.4 that
⎛ ⎞
γn 
δn
0 = T (A) = lim T ⎝ dij ⎠ .
(n)
n→∞
i=1 j=1

Hence, A = A − T (A) is equal to

⎛ ⎛ ⎞⎞

αn 
γn 
δn γn 
δn
lim ⎝ T(dij ) − T ⎝ dij ⎠⎠ .
(n) (n) (n) (n) (n)
Ai H(bi )Ci +
n→∞
i=1 i=1 j=1 i=1 j=1

The first item is evidently in J2 ; the second one is in the quasicommutator ideal
of T(S1 ) by Lemma 3.5. This ideal is generated by the quasicommutators
T(a)T(b) − T(ab) = −H(a)H(b̃)
which are in J2 because of S1 ⊆ S2 . 
In contrast to the algebras alg TH(S1 , S2 ), the Toeplitz-plus-Hankel algebras
alg TH+ (S) again arise from a discretization mapping, namely from the restriction
of the mapping
T + H : Smb → LT(X), s → T(s) + H(s)
to S, again with smb as a homomorphic left inverse. We denote the related qua-
sicommutator ideal by qcT +H (S), i.e., qcT +H (S) is the smallest closed ideal of
alg TH+ (S) which contains all operators
T(cd) + H(cd) − (T(c) + H(c))(T(d) + H(d)) with c, d ∈ S.
Corollary 3.11.
(a) qcT +H (S) = ker(smb |alg TH+ (S) ) = ker(T |alg TH+ (S) ).
(b) The algebra alg TH+ (S) decomposes into the direct sum
alg TH+ (S) = {T(c) + H(c) : c ∈ S} ⊕ qcT +H (S).
(c) The sequence
{0} → qcT +H (S) → alg TH+ (S) → S → {0},
298 S. Roch and B. Silbermann

with the second arrow standing for the natural embedding and the third one
for the restriction of the symbol mapping, is exact.

We conclude by a few remarks on commutators. The commutator ideal com B

of a Banach algebra B is the smallest closed ideal of B which contains all com-
mutators ab − ba with a, b ∈ B. It is easy to see that in the context of Theorem
3.6 and if D(A) generates B, the commutator ideal of B is already generated by
commutators D(c)D(d) − D(d)D(c) with c, d ∈ A. Thus, if A is commutative,
then com B is contained in qc (A), which follows from
D(c)D(d) − D(d)D(c) = (D(c)D(d) − D(cd)) − (D(d)D(c) − D(dc)).
Note also that the commutator ideal of a closed subalgebra A of LT(X) with
commutative symbol algebra smb A is contained in ker T .
The above remarks apply in particular to Toeplitz and Toeplitz-plus-Hankel
algebras. For the commutator ideal of Hankel algebras, we have the following.

Lemma 3.12. If S1 is commutative, then com alg TH(S1 , S2 ) is contained in ker T .

Proof. Let A, B ∈ alg T H(S1 , S2 ). By Corollary 3.9, we write A = T(a) + K and

B = T(b) + L with a, b ∈ S1 and K; L ∈ ker T . Then
AB − BA = T(a)T(b) − T(b)T(a) + M
with M ∈ ker T . The assertion follows since
T (T(a)T(b) − T(b)T(a)) = T(ab − ba) = 0
by Lemma 3.4 and since S1 is commutative. 

3.4. Continuous symbols

The operators Vt belong to the algebra TL(X) as we observed in Example 2.2.
Let alg (V) stand for the smallest closed subalgebra of TL(X) which contains all
of these operators.

Lemma 3.13. Let A be a closed subalgebra of TL(X) which contains V. Then the
following sets are equal for every t0 ∈ G+ \ {0}:
(a) the smallest closed ideal of A which contains all operators Pt , t ∈ G;
(b) the smallest closed ideal of A which contains Pt0 .

Proof. Let I stand for the smallest closed ideal of A which contains Pt0 . Clearly,
I is contained in the ideal described in (a). For the reverse inclusion, we have to
show that Pt ∈ I for every t ∈ G+ . Let k ∈ Z+ and k ≥ 2. Then
Pkt0 − P(k−1)t0 = V(k−1)t0 V−(k−1)t0 − Vkt0 V−kt0
= V(k−1)t0 (I − Vt0 V−t0 )V−(k−1)t0
= V(k−1)t0 Pt0 V−(k−1)t0 ∈ I.
 Toeplitz and Hankel Algebras – Axiomatic and Asymptotic Aspects 299

Summing up we conclude that Pkt0 ∈ I for every k ∈ Z+ . Now, given t ∈ G+ ,

choose k ∈ Z+ such that t < kt0 . Then, by Lemma 2.2 (a),
V−(kt0 −t) Pkt0 Vkt0 −t = I − V−(kt0 −t) Vkt0 V−kt0 Vkt0 −t
= I − Vt V−t = Pt ,
whence Pt ∈ I. 
In particular we see that if one of the projections Pt with t ∈ G+ \ {0} is
compact, then each of these projections is compact.
Let I(P) denote the smallest closed ideal of alg (V) which contains (one or all
of) the projections Pt . Further write ϕt for the symbol of Vt and C for the smallest
closed subalgebra of the symbol algebra Smb which contains all symbols ϕt . We
call C the algebra of the continuous symbols. This notion is inspired by Gohberg
and Feldman’s text [13]. Note that ϕs ϕt = ϕs+t for all s, t ∈ G and that ϕ0 is the
identity element of Smb.
Lemma 3.14.
(a) alg (V) = alg (T(C)).
(b) I(P) = ker(T |alg (V) ) = qcT (C).
Proof. Assertion (a) is a consequence of T(ϕt ) = Vt . The second equality in (b)
comes from Corollary 3.8. The first equality is then evident since the quasicom-
mutators Pt = I − Vt V−t belong to the kernel of T , and they generate the quasi-
commutator ideal qcT (C). 
Whereas the symbol algebra Smb need not to be commutative, its subalgebra
of the continuous symbols is commutative. Moreover, the following holds.
Lemma 3.15. C is a closed subalgebra in the center of Smb.
Proof. Let q ∈ Smb and t ∈ G+ . By Lemma 2.6 (a), V−t T(q)Vt = T(q). Passing
to symbols we obtain ϕ−t qϕt = q or, equivalently, qϕt = ϕt q for t ∈ G+ . Since
ϕ−t = (ϕt )−1 , this equality holds for negative t as well. 
3.5. Abstract Toeplitz operators as compressions
We infer from Lemma 2.6 (a) that the elements of im T , hence all abstract Toeplitz
operators, are shift invariant. We will see now that abstract Toeplitz operators also
arise as compressions of abstract Laurent operators, for which we have to embed
X into a larger space. This construction will also shed new light onto the algebra
TL(X) since it identifies the four strong limits T , T , H and H as components of
a single strong limit.
Let the Banach space X, the group G and the operators Vt , Rt , Pt and Qt
with t ∈ G be as in Section 2.1. Let X 2 stand for the direct sum X ⊕ X, provided
with the norm (x1 , x2 ) = x1 X + x2 X (or with another norm making X 2 to
a Banach space into which X is isometrically embedded). Thinking of the elements
of X 2 as column vectors, we identify operators in L(X 2 ) with 2 × 2 matrices with
entries in L(X).
300 S. Roch and B. Silbermann

For t ∈ G+ , define operators U±t on X 2 by

   
Vt Rt V−t 0
Ut := and U−t := . (3.4)
0 V−t Rt Vt
Proposition 3.16. The operators Ut form a commutative group, i.e., Us Ut = Ut Us =
Us+t for all s, t ∈ G.
Proof. We prepare the proof by establishing the identities
Vs Vt + Rs R−t = Vs+t for all s, t ∈ G (3.5)
Vs Rt + Rs V−t = Rs+t for all s, t ∈ G. (3.6)
From Lemmas 2.2 (b) and 2.3 (d) we know that Rs R−t = Ps Vs+t and Vs Vt =
Qs Vs+t . Summing up gives (3.5). For (3.6), recall from Axiom (T3 ) and Lemma
2.3 (c) that Vs Rt = Rs+t Pt and Rs V−t = Ps Rs+t . Since V−t = V−t Qt , the last
identity implies that Rs V−t = Ps Rs+t Qt . It thus remains to show that
Ps Rs+t Qt = Rs+t Qt or, equivalently, Qs Rs+t Qt = 0,
which is further equivalent to V−s Rs+t Vt = 0. Now we have
V−s Rs+t Vt = Rt Ps+t Vt by Axiom (T3 )
= Rt Pt Ps+t Vt by Lemma 2.3 (a)
= Rt Ps+t Pt Vt by Lemma 2.2 (d)
=0 since Pt Vt = 0.
The identities (3.5) and (3.6) together imply that
    
Vs Rs Vt Rt Vs+t Rs+t
=
R−s V−s R−t V−t R−s−t V−s−t
for all s, t ∈ G, from which the assertion follows (note that Rt = 0 for t ≤ 0). 

The simple identity

   
A 0 V−t AVt V−t ARt
U−t Ut = ,
0 0 Rt AVt Rt ARt
holding for every A ∈ L(X) and t ∈ G+ , implies the relation between the two
approaches to Toeplitz-like operators.
Theorem 3.17. Let A ∈ L(X). Then the strong limit
 
A 0
L(A) := s-limt→∞ U−t Ut
0 0
exists if and only if A ∈ TL(X). In this case,
 
T (A) H(A)
L(A) =  . (3.7)
H(A) T (A)
 Toeplitz and Hankel Algebras – Axiomatic and Asymptotic Aspects 301

Corollary 3.18.
(a) The mapping L : TL(X) → L(X 2 ) is a bounded homomorphism.
(b) U−t L(A)Ut = L(A) for every A ∈ TL(X) and t ∈ G.
Proof. Assertion (a) comes from the definition of L and Theorem 2.4 (b); assertion
(b) can be proved in the same as Lemma 2.6 (likewise: it follows from that lemma).


Define mappings
E : X → X 2 , x → (x, 0)T , E −1 : X 2 → X, (x, y)T → x.
Then, by (3.7),
E −1 L(A)E = T (A) for every A ∈ LT(X). (3.8)
In other words, if P ∈ L(X ) stands for the projection (x, y) → (x, 0) , then
2 T T

every abstract Toeplitz operator can be viewed as the compression P A|imP of an

abstract Laurent operator A ∈ im L.
Theorem 3.19.
(a) im L is a closed subalgebra of L(X 2 ).
(b) The mapping A → smb (E −1 AE) is a bounded isomorphism from im L onto
Smb = smb TL(X).
Proof. (a) Let L(An ) be a sequence in im L which converges in L(X 2 ). Then, by
(3.8), (T (An )) is a Cauchy sequence in im T . Since im T is closed by Theorem 2.7
(a), there is a operator A ∈ LT(X) such that T (An ) → T (A) in the operator norm.
The first column in the table in Theorem 2.5 shows that L(T (B)) = L(B) for every
B ∈ TL(X). Hence, L(An ) = L(T (An )) converges in the norm to L(A) = L(T (A)).
Thus, im L is a closed subspace of L(X 2 ). It is also an algebra by Corollary 3.18.
(b) The mapping A → smb (E −1 AE) sends L(A) to smb T (A) by (3.8). This
mapping is a homomorphism because T (A) − T (A)T (B) ∈ ker T and smb
is a homomorphism. We are going to show that this mapping is injective. Let
smb (E −1 L(A)E) = smb T (A) = 0 for an operator A ∈ TL(X). Then T (A) ∈
ker T ; hence, T (A) = 0. The first column in the table in Theorem 2.5 then shows
that L(A) = L(T (A)) = 0. 

It is not hard to see that the inverse of the mapping A → smb (E −1 AE) in
Theorem 3.19 is explicitly given by
(smb TL(X) = ) Smb → im L, q → L(T(q)).
So we have two ways to think of the symbol of a Toeplitz-like operator: first, as an
element of the quotient algebra TL(X)/ ker T ; second as an operator in L(TL(X))
acting on X 2 .
Corollary 3.20. The mappings (3.1) are bounded.
302 S. Roch and B. Silbermann

Proof. We shall demonstrate this for the mapping T : Smb → im T , q → T(q).

Let μ : Smb → im L denote the inverse of the bounded isomorphism in Theorem
3.19 (b). Then μ is bounded by Banach’s bounded inverse theorem; hence, the
mapping
Smb → im T , q → T(q) = P μ(q)|X
is bounded. 

Remark 3.21. With the above notation and results, it becomes evident that The-
orem 2.7 (b) can be viewed as a special case of Theorem 3.6, with D : im L →
TL(X), A → E −1 AE as the discretization mapping and L as its homomorphic
left inverse. 

4. Laurent-like operators
4.1. The axioms
In the previous sections, we started with two families V and R of operators to
define first an algebra of Toeplitz-like operators and related abstract Toeplitz and
Hankel operators, and then a group of shifts on a larger space, together with the
related abstract Laurent operators. We will see now that one can also go the way
around.
This approach works for Banach spaces X  which are a direct sum of
two closed subspaces of equal size. Formally this means that there are operators
P, Q, J ∈ L(X  ) such that
P 2 = P, Q2 = Q, P + Q = I and J 2 = I, JP J = Q.
Then X  is equal to the direct sum im P ⊕ im Q, the mapping J|imP : im P → im Q
is a linear isomorphism and, with X := im P , there is a linear isomorphism
X  → X 2, x → (P x, JQx) = (P x, P Jx).
In what follows we simply assume that X  is already of the form X 2 for a certain
Banach space X and that P, Q and J are given by
P : (x, y) → (x, 0), Q : (x, y) → (0, y) and J : (x, y) → (y, x).
Again we write the elements of X 2 as column vectors and identify operators in
L(X 2 ) with 2 × 2 matrices with entries in L(X).
Let G = {0} and G+ be as before, and let (Ut )t∈G be a bounded family of
operators Ut ∈ L(X 2 ) subject to the following axioms:
(L1 ) The mapping t → Ut is a group isomorphism on G. In particular, Us Ut = Us+t
for all s, t ∈ G.
(L2 ) U0 = I and JUt J = U−t for every t ∈ G.
(L3 ) P Us P Ut P = P Us+t P for all s, t ∈ G+ .
(L4 ) U−t P Ut → I strongly as t → ∞.
 Toeplitz and Hankel Algebras – Axiomatic and Asymptotic Aspects 303

The first condition in axiom (L2 ) ensures that all operators Ut are invertible in
L(X 2 ) and that Ut−1 = U−t . By calling an invertible operator U ∈ L(X 2 ) J-
unitary, we can rephrase the second condition in axiom (L2 ) as the J-unitarity of
the Ut . It is easy to see that if U is J-unitary, then there are operators V± , R± ∈
L(X) such that
   
V+ R+ −1 V− R−
U= and U = .
R− V− R+ V+
In particular, there are operators Vt , Rt ∈ L(X) such that
   
Vt Rt −1 V−t R−t
Ut = and Ut = U−t =
R−t V−t Rt Vt
for all t ∈ G+ .
Lemma 4.1. The so-defined operators V±t and R±t are subject to axioms (T1 )–(T4 ).
Proof. The group property of the Ut implies that
Vs Vt + Rs R−t = Vs+t , Vs Rt + Rs V−t = Rs+t for s, t ∈ G. (4.1)
Since Vs Vt = Vs+t by (L3 ), the first identity in (4.1) implies that
Rs R−t = 0 for s, t ∈ G+ . (4.2)
Further we conclude from (L4 ) that
       
V−t R−t I 0 Vt Rt V−t Vt V−t Rt I 0
= →
Rt Vt 0 0 R−t V−t R−t Vt Rt2 0 I
as t → ∞, whence Rt2 → I as t → ∞. Together with (4.1) this implies that
0 = Rs2 R−t → R−t as s → ∞.
Thus, R−t = 0 for t ∈ G+ , and the matrix representations of Ut and U−t are upper
and lower triangular, respectively. With this information it is easy to check that
the (V−t )t∈G+ own the semigroup property (which gives (T1 ) and that V−t Vt = I
for t ∈ G+ and Vt V−t → 0 as t → ∞. Then V−t Vt V−t = V−t → 0 as t → ∞,
i.e., (T4 ) holds. Another consequence of Vt V−t → 0 is that Vt V−t = I respective
Pt := I − Vt V−t = 0 for large positive t. It follows as in the proof of Lemma 3.13
that then Pt = 0 for all positive t, i.e., Vt V−t = I for all positive t, which finishes
the proof of (T2 ).
It remains to check (T3 ). Let t ∈ G+ . The identity Vt V−t + Rt2 = I is nothing
but the south-east corner of 2 × 2-matrix identity U−t Ut = I. Further, Rt Vt = 0 by
the south-west corner of that identity. Hence, Rt Qt = 0 and Rt Pt = Pt Rt = Rt .
Now the second identity in (T3 ) follows easily by multiplying the second identity
in (4.1) from the right-hand side by Pt . 

Note that, conversely, the operators defined by (3.4) satisfy the axioms (L1 )–(L4 ).
304 S. Roch and B. Silbermann

4.2. Laurent-like operators

Let X 2 and the operators Ut be as in Section 4.1. We let LL(X 2 ) stand for the set
of all operators A ∈ L(X 2 ) for which the two strong limits
L± (A) := s-limt→±∞ U−t AUt
exist. We call the elements of LL(X 2 ) Laurent-like operators.
Theorem 4.2.
(a) LL(X 2 ) is a closed subalgebra of L(X 2 ).
(b) An operator A ∈ L(X 2 ) is in LL(X 2 ) if and only if
 
B C
A= with B, C, D, E ∈ TL(X) and C, D ∈ ker T . (4.3)
D E
(c) If A is of the form (4.3), then
   
T (B) H(B) T (E) H(E)

L+ (A) =  and L− (A) = . (4.4)
H(B) T (B) H(E) T (E)
Proof. Assertion (a) is standard. Let A be of the form (4.3). Then, for t ∈ G+ ,
 
V−t BVt V−t BRt + V−t CV−t
U−t AUt = .
Rt BVt + Vt DVt Rt BRt + Vt DRt + Rt CV−t + Vt EV−t
The operators V−t BVt , V−t BRt , Rt BVt and Rt BRt converge strongly to T (B),

H(B), H(B) and T (B) as t → ∞, respectively, because B ∈ TL(X). Since V−t → 0
strongly, the operators V−t CV−t and Vt EV−t converge strongly to 0. Finally, the
operators Vt DVt and Vt DRt converge strongly to 0 because D ∈ ker T and by
Lemma 1.11 in [20]. This proves the first assertion in (c); the second one follows
analogously.
It remains to verify the “only if” implication in (b). From (L4 ) and the J-
unitarity of the Ut we conclude that P and Q belong to LL(X 2 ). Thus, if
 
B C
A= ∈ LL(X 2 ),
D E
then        
B 0 0 C 0 0 0 0
, , , ∈ LL(X 2 ).
0 0 0 0 D 0 0 E
From    
B 0 V−t BVt V−t BRt
U−t Ut =
0 0 Rt BVt Rt BRt
we conclude that B ∈ TL(X). Further, the equality
   
0 0 0 0
U−t Ut =
D 0 Vt DVt Vt DRt
implies that the operators Vt DVt and Vt DRt converge strongly. Since V−t → 0
strongly, we conclude that
DVt = V−t Vt DVt → 0 and DRt = V−t Vt DRt → 0
 Toeplitz and Hankel Algebras – Axiomatic and Asymptotic Aspects 305

strongly as t → ∞. But then

V−t DVt → 0, Rt DVt → 0, V−t DRt → 0, Rt DRt → 0
strongly as t → ∞. Hence, D ∈ TL(X), and each of the four strong limits is
zero. In particular, D ∈ ker T . Working with L− instead of L+ , we obtain that
C, E ∈ TL(X) and C ∈ ker T . 

5. Examples of Toeplitz-like operators

Here we present a few concrete Banach spaces and examine the related algebras of
Toeplitz-like operators. A main objective is to identify the abstract Toeplitz and
Hankel operators T(a) and H(a) with classical (concrete) Toeplitz and Hankel op-
erators acting on these spaces. For a general acquaintance with (concrete) Toeplitz
and Hankel operators see, e.g., [4, 9, 16, 18].
5.1. Toeplitz-like operators on H p
Let H p , 1 < p < ∞, be the Hardy space introduced in Section 1, and let G = Z.
Let V0 := I and, for n ∈ Z+ \ {0}, define operators V±n on H p by
5 6

n−1
Vn f := χn f and V−n f := χ−n f − fˆi χi ,
i=0
respectively. Clearly, V−n Vn = I and V−n → 0 strongly as n → ∞. Note also that
 3

n−1
ker V−n = f ∈ H p : f = fˆi χi .
i=0
p
Finally, we define operators Rn on H by

n−1
Rn f := fˆn−1−i χi = JQχ−n P
i=0

if n ∈ Z+ \ {0} and Rn = 0 else. Then all axioms of Section 2.1 are satisfied, and
the corresponding algebra TL(H p ) is well defined. Moreover, Vn → 0 and Rn → 0
weakly. We let Un stand for the related operators on (Hp )2 , defined as in Section
3.5.
As already mentioned, the space H p ⊂ Lp (T) is equal to im P , with P the
classical Riesz projection. Set Q := I − P . Thus, Lp (T) = im P ⊕ im Q, whereas
(H p )2 = im P ⊕ im P , which we may write as
   
im P im P
Lp (T) = and (H p )2 = .
im Q im P
With respect to this identification, we may further identify operators from Lp (T)
to (H p )2 with 2 × 2-matrices. It is easy to see that the operators
     
P 0 im P im P
η := : →
0 JQ im Q im P
306 S. Roch and B. Silbermann

and      
−1 P 0 im P im P
η := : →
0 JP im P im Q
are inverse to each other and that η −1 Un η = χn I for all n ∈ Z. More precisely,
consider the mapping
   
im P Ph
ρ : Lp (T) → Lp2 (T) := , h→
im Q Qh
which is an isomorphism with inverse
 
−1 Ph
ρ : Lp2 (T) → L (T),
p
→ P h + Qh = h.
Qh
The norm on Lp2 (T) is chosen in such a way that ρ becomes an isometry. A simple
computation gives for h ∈ Lp (T) and n ∈ Z+ that
χn h = (P + Q)χn (P + Q)h
equals  
P χn P P χn Q
χn h = ρ−1 ρh.
Qχn P Qχn Q
Using the relations P χn P = Vn , Qχn P = 0,
P χn Q = P JJχn JP J = P Jχ−n P J = JQχ−n P J = Rn J
(holding by the definition of Rn ) and
Qχn Q = JP Jχn JP J = JP χ−n P J = JV−n J,
one gets  
−1 Vn Rn J
χn h = ρ ρh.
0 JV−n J
On the other hand,
 
Vn Rn J
ρ−1 η −1 Un ηρ = ρ−1 ρ
0 JV−n J
which, together with the previous identity, implies χn h = ρ−1 η −1 Un ηρh. In the
same manner one can check that, for h ∈ Lp (T) and n ∈ Z \ Z+ ,
χ−n h = ρ−1 η −1 U−n ηρh.
Now let A ∈ TL(H p ) and n ∈ Z. From U−n L(A)Un = L(A) we conclude that
χ−n (ρ−1 η −1 L(A)ηρ)χn I = ρ−1 η −1 L(A)ηρ.
Hence, ρ−1 η −1 L(A)ηρ is the operator of multiplication by some function q ∈
L∞ (T) (see [5]). This gives
 
−1 P qP P qQ
η L(A)η =
QqP QqQ
 Toeplitz and Hankel Algebras – Axiomatic and Asymptotic Aspects 307

which finally implies  

T (q) H(q)
L(A) = .
H(q̃) T (q̃)
In particular, T (A), T (A), H(A) and H(A)
 coincide with the familiar (concrete)
Toeplitz and Hankel operators T (q), T (q̃), H(q) and H(q̃) on H p as defined in
Section 1.
As a consequence, we observe that the corresponding symbol algebra Smb is
a subalgebra of L∞ (T). In fact, Smb = L∞ (T), which is an immediate consequence
of the identities
V−n T (a)Vn = T (a), Rn T (a)Rn = Pn T (ã), (5.1)
V−n T (a)Rn = H(a)Pn , Rn T (a)Vn = Pn H(ã) (5.2)
for all a ∈ L∞ (T) and n ∈ Z+ . Let us check the last identity of (5.2), for example.
By the first relation in (1.1) we have
I = T (χn χ−n ) = T (χn )T (χ−n ) + H(χn )H(χn ).
Hence, H(χn ) = I − Qn = Pn . Further,
2

Rn T (a)Vn = Rn T (aχn ) = Rn (T (χn )T (a) − H(χn )H(ã)).

Since Rn = H(χn ) and Rn T (χn ) = 0, we conclude that Rn T (a)Vn = Pn H(ã),
as desired. Note that one could also start with (5.1), (5.2) in order to derive that
η −1 L(T (a))η = aI.
Let C be a closed subalgebra of L∞ (T). Then, in the present context, Corol-
lary 3.9 specifies to
alg T H(C) = T (C) ⊕ ker(T |alg T H(C) ).
This result was recently proved by Didas [8, Theorem 2.4] for p = 2 and
under the assumption that C is an inner subalgebra of L∞ (T) in the sense of [17].
For example, C(T) is inner, and so is every closed subalgebra of L∞ (T) which
strictly contains H ∞ . Moreover, Didas showed that ker(T |alg T H(C) ) equals the
commutator ideal of alg T H(C) under this additional assumption.
 are subalgebras
At the end of Section 2.2 we remarked that ker H and ker H
of TL(X). In the setting of this section, we can describe these algebras more
precisely as
ker H = {T (a) : a ∈ H ∞ } ⊕ ker T ,  = {T (a) : a ∈ H ∞ } ⊕ ker T .
ker H
Remark 5.1. The construction of the algebra TL(X) works as well on weighted
Hardy spaces H p (w) and the above results remain valid in this setting provided
that the weight w is symmetric (i.e., w(t) = w(1/t)) and that the Riesz projection
P is bounded on Lp (w) (which holds if w is a Muckenhoupt weight, for example if
w is a power weight; see [3], Chapter 2). In this setting, the operators Vn and Rn
can be defined as above; they are uniformly bounded and all requirements made
in Section 2.1 are satisfied.
308 S. Roch and B. Silbermann

5.2. Toeplitz-like operators on lp

Here we are going to employ the alternative approach of Section 3.5 to define
Toeplitz-like operators on the classical sequence spaces lp (Z+ ), 1 ≤ p < ∞. Con-
sider the reflection operator J : (xn )n∈Z → (yn )n∈Z where yn = x−n−1 and the
discrete Riesz projection P : (xn )n∈Z → (yn )n∈Z where yn = xn if n ≥ 0 and
yn = 0 else. Clearly, J 2 = I and JP J = I − P =: Q.
Let a ∈ L∞ (T). On l0 (Z), the linear space of the finitely supported sequences,
we consider the Laurent operator L(a) defined by

(L(a)x)k := âk−m xm .
m∈Z

Note that only finitely many items in this sum do not vanish. The function
a is called a multiplier on lp (Z) if
L(a) := sup{L(a)xp : x ∈ l0 (Z), xp } < ∞.
If a is a multiplier, then L(a) extends to a bounded linear operator on lp (Z)
which we denote by L(a) again. The set M p of all multipliers on lp (Z) forms a
Banach algebra under the norm aM p := L(a)L(lp (Z)) (see [4, Chapter 2]). In
particular, every function a ∈ L∞ (T) with bounded total variation belongs to M p ,
and there is a constant cp independent of a such that Stechkin’s inequality
aM p ≤ cp (a∞ + Var (a))
holds. Note also that JL(a)J = L(ã) for every a ∈ M p .
Write lp for lp (Z+ ) = im P and let a ∈ M p . The operators
T (a) : lp → lp , f → P L(a)f, H(a) : lp → lp , f → P L(a)QJf
are called the (concrete) Toeplitz and Hankel operator with generating function a,
respectively. These operators are bounded if a ∈ M p . It is also easy to see that if
T (a)f = g and H(a)f = h then
∞
 ∞

gj = âj−k fk , hj = âj+k+1 fk .
k=0 k=0

Let {ej }j∈Z+ stand for the standard basis on lp , i.e., the jth entry of ej is 1
whereas all other entries are 0. The lp -version of a Theorem by Brown and Halmos
(Theorem 2.7 (b) in [4]) tells us that if A ∈ L(lp ) and if there is a sequence (an )n∈Z
in C such that
Aej , ek  = aj−k for all j, k ∈ Z+ ,
then there exists a multiplier a ∈ M p such that  A = T (a) and an is the nth
Fourier coefficient of a. Here, as usual, x, y := xn yn for x ∈ lp and y ∈ lq with
p−1 + q −1 = 1.
Next we are going to define the related algebra TL(lp ) of Toeplitz-like oper-
ators. Set G = Z and define
Vn := T (χn ) and Rn := H(χn ) for n ∈ Z (5.3)
 Toeplitz and Hankel Algebras – Axiomatic and Asymptotic Aspects 309

with χn (t) = tn as before. One easily checks that these operators satisfy axioms
(T1 )–(T4 ); hence, the algebra TL(lp ) with respect to these families is well defined.
Moreover, if 1 < p < ∞, then Vn → 0 and Rn → 0 weakly as n → ∞. Note that
this claim fails for p = 1. Indeed, consider the bounded linear functional

g : l1 (Z+ ) → C, x = (xi )i∈Z+ → xi .
i∈Z+

Then g(Vn x) = g(x) for all x ∈ l1 (Z+ ) and n ∈ Z+ ; hence, the Vn cannot converge
weakly to zero. Now the assertion for the Rs comes from Proposition 2.9.
If A ∈ TL(lp ) and T (A) = A, then V−1 AV1 = A. Thus, A satisfies the
hypotheses of the Brown/Halmos theorem, and there is a multiplier a ∈ M p such
that A = T (a). Thus, every abstract Toeplitz (or Hankel) operator on lp is a
concrete Toeplitz (or Hankel) operator as defined above. The lp -version of Theorem
2.7 implies
TL(lp )/ ker T ∼
= M p.

Remark 5.2. An equivalent treatment of Toeplitz-like operators on weighted lp -

spaces fails: the operators (5.3) are not longer uniformly bounded then. This is in
strong contrast to the H p -setting in the previous section. 

Our final goal is to provide some peculiar properties of the algebra TL(X),
its ideal ker T and its subalgebras alg T (L∞ ) and alg T H(L∞ ). In particular, we
are going to show that
(a) TL(l2 ) is not a C ∗ -algebra.
(b) ker T ⊂ TL(l2 ) is not a two-sided ideal of L(l2 ).
(c) alg T H(L∞ ) is a proper subalgebra of TL(l2 ).
(d) alg T (L∞ ) is a proper subalgebra of alg T H(L∞ ).
For (a), assume that TL(l2 ) is a C ∗ -algebra. Then ker T is a closed symmetric
ideal of TL(l2 ) and, hence, a closed symmetric left-sided ideal of L(l2 ). So we have
BC, B ∗ C ∗ ∈ ker T for every B ∈ L(l2 ) and C ∈ ker T by the left ideal property,
whence (B ∗ C ∗ )∗ = CB ∈ ker T by the symmetry. Thus, ker T turns out to be a
closed two-sided ideal of L(l2 ). Moreover, ker T contains non-compact operators
(for example, the Hankel operator which generates the Hilbert matrix), but not
every bounded linear operator on l2 is in ker T (for example, I ∈ ker T ). This
contradicts the well-known fact (first observed by Calkin [6]; see also [12]) that
the only non-trivial closed two-sided ideal of L(l2 ) is the ideal of the compact
operators.
The same arguments show (b), and (c) follows because alg T H(L∞ ) is a
symmetric closed subalgebra of TL(l2 ), whereas TL(l2 ) fails to be symmetric by
(a). Finally, (d) comes from the existence of Hankel operators which do not belong
to alg T (L∞ ) as first observed by S. Axler and proved in Power [18].
The following explicit example of an operator A ∈ TL(l2 ) for which A∗ ∈
TL(l ) is due to our former colleague Hans-Jürgen Fischer.
2
310 S. Roch and B. Silbermann

Example. Let the operator A be given by the matrix representation A = (aij )∞

i,j=0
where  1
√
2i
if 0 ≤ j ≤ 2i − 1,
aij =
0 if j ≥ 2i .
We show that A is a bounded operator on l2 . Let x = (xi )i≥0 ∈ l2 and x = 1.
Then
2
∞ 2n −1
1 
Ax =
2
xk .
n=0
2n
k=0
1
For k ∈ Z+ , set λk := (k + 1)− 4 . Then, by the Cauchy–Schwarz inequality,
n
2 −1
2 n
2 −1
2 52n −1 6 52n −1 6
  xk 2
xk = λk λ−1
k xk ≤ λ2k .
λk
k=0 k=0 k=0 k=0
The estimate n
2 −1 n
2 −1
1 √
λ2k = (k + 1)− 2 ≤ 2 2n − 1
k=0 k=0
then gives
∞ √ 52n −1 6
 2 2n − 1  xk
2
Ax2 ≤
n=0
2n λk
k=0

∞ 2−1 n
2
1 xk
≤2 √
n=0 k=0
2n λk
∞
  2
1 xk
=2 √ .
k=0 n≥log2 (k+1)
2n λk
Using
 1 1 1 1 √ √
√ = √ √ = √ 2( 2 + 1),
n≥log2 (k+1)
2 n k + 1 1 − 1/ 2 k+1
we arrive at
∞ ∞

√ √ 1 xk
2 √ √
Ax2 ≤ 2 2( 2 + 1) √ =2 2( 2 + 1) |xk |2 ,
k=0
k + 1 λk k=0
√ √
whence the boundedness of A and the estimate A ≤ 2 2( 2 + 1). From 2

AVn ek 2 = Aen+k 2 = 2− log2 (n+k) for k, n ≥ 0

and
ARn ek 2 = Aen−k 2 = 2− log2 (n−k) for n > k ≥ 0
we further conclude that AVn → 0 and ARn → 0 strongly as n → ∞. Hence, A ∈
TL(l2 ) (and A ∈ ker T ). Suppose that A∗ ∈ TL(l2 ). Then (V−n A∗ Vn ) converges
strongly to some operator T (A∗ ) by definition. On the other hand, (V−n A∗ Vn ) =
 Toeplitz and Hankel Algebras – Axiomatic and Asymptotic Aspects 311

(V−n AVn )∗ → 0 weakly because V−n AVn → 0 strongly. Hence, T (A∗ ) = 0. Then,
by Lemma 1.11 in [20], A∗ Vn → 0 strongly, which is impossible since A∗ Vn e0  = 1
for all n ∈ Z+ . 
Remark 5.3. It is an open question if alg T H(L∞ ) is a proper subalgebra of
TL∗ (l2 ) := {A ∈ TL(l2 ) : A∗ ∈ TL(l2 )}.
5.3. Toeplitz-like operators on Lp (R+ )
Throughout this section, we let Lp and Lp± stand for the classical Lebesgue spaces
on R, R+ = [0, ∞) and R− = R \ R+ , respectively. Every function a ∈ L∞ induces
a bounded operator f → af on Lp which we denote by m(a). In particular, if χ+
and χ− stand for the characteristic functions of R+ and R− , then we let P and
Q denote the operators m(χ+ ) and m(χ− ), respectively. The ranges of P and Q
can be identified with Lp+ and Lp− . We will also need the flip operator J on Lp
which is defined by (Jf )(t) = f (−t). Clearly, J 2 = I and Jm(a)J = m(ã) where
ã = a(−t) for a ∈ L∞ .
We write the Fourier transform on L1 in the form
!
1
(F f )(x) = √ f (t)eixt dt, x ∈ R,
2π R
and use the same letter F to denote the continuous extension of F to a unitary
operator on L2 . Note that F −1 = JF .
Let a ∈ L∞ . Then the operator M (a) := F −1 m(a)F : L2 → L2 is bounded
and has norm a∞ . For 1 ≤ p < ∞, we let M p (R) stand for the collection of
all functions a ∈ L∞ owning the following property: Whenever f ∈ L2 ∩ Lp , then
M (a) ∈ Lp , and there is a constant cp independent of f such that M (a)f p ≤
cp f p for all f ∈ L2 ∩ Lp . If a ∈ M p (R), then M (a) : L2 ∩ Lp → Lp extends to a
bounded operator on Lp , called the operator of convolution by a. We denote this
operator by M (a) again. The elements of M p (R) are also called Lp -multipliers.
An operator A on Lp is called translation invariant if U−s AUs = A for all
s ∈ R where (Us f )(t) := f (t − s). If a ∈ M p (R), then the convolution operator
M (a) is translation invariant. A theorem of Hörmander [15] (see also 9.2 in [4])
states that the converse is also true: every translation invariant operator on Lp is
an operator of convolution by a certain Lp -multiplier.
If a ∈ L∞ is of finite total variation V (a), then a ∈ M p (R), and the Stechkin
inequality
M (a)L(Lp) ≤ Cp (a∞ + V (a))
holds. Here, Cp is a constant independent of a which can be chosen concretely
as Cp = M (χ+ − χ− )L(Lp ) . Note also that JM (a)J = M (ã). Further basic
properties of M p (R) can be found in [4, Chapter 9], for example.
Let 1 ≤ p < ∞. With every multiplier a ∈ M p (R), we associate the Wiener–
Hopf integral operator W (a) and the Hankel integral operator HR (a) by
W (a) := P M (a)|Lp+ : Lp+ → Lp+ and HR (a) := P M (a)QJ|Lp+ : Lp+ → Lp+ ,
312 S. Roch and B. Silbermann

respectively. These operators are bounded on Lp+ , and the following analogues of
(1.1) hold for all Lp -multipliers a, b:
W (ab) = W (a)W (b) + HR (a)HR (b̃), HR (ab) = W (a)HR (b) + HR (a)W (b̃).
For s ∈ R, define ωs (x) := eisx . Then ωs ∈ M p (R) for every p ∈ [1, ∞) and M (ωs )
is nothing but the translation operator Us . Now set Vs := W (ωs ) for s ∈ G,
Rs := HR (ωs ) for s ∈ R+ and Rs = 0 for s ∈ R− . Then the axioms in Section 2.1
are satisfied, thus, the algebra TL(Lp+ ) is well defined. Moreover, the operators Vs
and Rs tend weakly to 0 as s → ∞ if 1 < p < ∞. For p = 1, neither the Vs nor the
Rs tend weakly to zero which follows by arguments similar to those for X = l1 .
As before, one can easily prove that for s ≥ 0 and a ∈ M p (R),
V−s W (a)Vs = W (a), Rs W (a)Rs = Ps W (ã)Ps ,
V−s W (a)Rs = HR (a)Ps , Rs W (a)Vs = Ps HR (ã).
What results is that every Wiener–Hopf operator W (a) belongs to TL(Lp+ ) and
that
T (W (a)) = W (a), T (W (a)) = W (ã),
H(W (a)) = HR (a), 
H(W (a)) = HR (ã).
Conversely, we are going to show that if A ∈ TL(Lp+ ) and T (A) = A, then there
exists a multiplier a ∈ M p (R) such that A = W (a). We proceed as in Section 5.1.
Slightly abusing the notation, we let Us stand for the related operators on (Lp+ )2 ,
defined as in Section 3.5. As already mentioned, the space Lp+ ⊂ Lp is equal to
im P . Thus, Lp = im P ⊕ im Q, whereas (Lp+ )2 = im P ⊕ im P , which we may write
as    
im P im P
Lp = and (Lp+ )2 = .
im Q im P
With respect to this identification, we may further identify operators from Lp to
(Lp+ )2 with 2 × 2-matrices. It is easy to see that the operators
     
P 0 im P im P
η := : →
0 JQ im Q im P
and      
P 0 im P im P
η −1 := : →
0 JP im P im Q
are inverse to each other and that η −1 Us η = M (ωs ) for all s ∈ R.
Now let A ∈ TL(Lp+ ). From U−s L(A)Us = L(A) for s ∈ R+ we conclude that
M (ω−s )(η −1 L(A)η)M (ωs ) = η −1 L(A)η.
This equality implies that η −1 L(A)η is translation invariant. By Hörmander’s the-
orem, there is a multiplier a ∈ M p (R) such that
 
−1 P M (a)P P M (a)Q
η L(A)η = M (a) = ,
QM (a)P QM (a)Q
 Toeplitz and Hankel Algebras – Axiomatic and Asymptotic Aspects 313

which implies  
W (a) HR (a)
L(A) = .
HR (ã) W (ã)
In particular, T (A), T (A), H(A) and H(A)
 coincide with the familiar (concrete)
Wiener–Hopf and Hankel operators W (a), W (q̃), HR (a) and HR (ã) as defined in
the beginning of this section.
We thus arrived at the following theorem which can be viewed as the analogue
of the Brown/Halmos theorem in the Lp+ -setting:
Theorem 5.4.
(a) If a ∈ M p (R), then W (a) belongs to TL(Lp+ ).
(b) If A ∈ TL(Lp+ ) and A = T (A), then there exists a multiplier a ∈ M p (R) such
that A = W (a).
Corollary 5.5. TL(Lp+ )/ ker T ∼
= M p (R).
In particular we conclude that M p (R) actually forms a Banach algebra.

6. Derivations
Let A and B be Banach algebras. A derivation on A is a (not necessarily bounded)
linear operator D : A → B such that D(ab) = D(a)b + aD(b) for all a, b ∈ A. Every
element c ∈ A induces a derivation a → ca − ac on A.
Chernoff [7] showed that if X is an infinite-dimensional Banach space and
A a closed subalgebra of L(X) which contains the compact operators, then every
derivation D : A → L(X) is bounded, and there is an operator C ∈ L(X) such
that D(A) = CA − AC for all A ∈ A.
Theorem 6.1. Let A be a closed subalgebra of LT(X) and suppose that the symbol
algebra smb A is commutative and semisimple. Then every bounded derivation
D : A → A maps A into ker T .
Proof. Let π denote the canonical homomorphism from A onto A/com A =: Aπ ,
with com A again referring to the commutator ideal of A. From
D(AB − BA) = (D(A)B + AD(B)) − (D(B)A + BD(A))
= (D(A)B − BD(A)) + (AD(B) − D(B)A)
we conclude that D maps com A into com A. Hence, the quotient mapping
Dπ : Aπ → Aπ , π(A) → π(D(A))
is a well-defined bounded derivation on Aπ . Since Aπ = A/com A is commutative,
Dπ maps the algebra Aπ into its Jacobson radical rad Aπ by the Singer–Wermer
theorem (see [22]).
Since smb A is commutative, the commutator ideal com A is contained in
A ∩ ker T , as we observed before Lemma 3.12. Hence, π maps A ∩ ker T onto a
closed ideal (A ∩ ker T )π of Aπ . Let ω stand for the canonical homomorphism
314 S. Roch and B. Silbermann

from Aπ onto Aπ /(A ∩ ker T )π . Then ω ◦ π is a homomorphism on A with kernel

A∩ker T . Hence, the quotient algebra Aπ /(A∩ker T )π is isomorphic to the algebra
smb A. Since smb A is semisimple by assumption, we conclude that rad Aπ ⊆
(A ∩ ker T )π . Thus, Dπ maps Aπ into (A ∩ ker T )π , which implies that D maps A
into A ∩ ker T . 

Combining this result with Chernoff’s theorem, we arrive at the following.

Corollary 6.2. Let A be as in Theorem 6.1 and suppose that A contains the compact
operators. Then every derivation D : A → A is bounded, there is an operator
T ∈ L(X) such that D(A) = T A − AT , and D maps into ker(T |A ).
Note that if A contains the compact operators then TL(X) contains the
compact operators, which happens if and only if Vs → 0 and Rs → 0 weakly as
s → ∞ by Theorem 2.8.
Didas [8] derived Corollary 6.2 in the context of (concrete) Toeplitz and
Hankel operators acting on the Hardy space H 2 on the unit disk (as defined in the
introduction). Moreover, assuming that C is an inner subalgebra of L∞ (T) which
strictly contains H ∞ , he showed that every derivation on alg T H(C) maps into the
commutator ideal of that algebra. The strict containment of H ∞ in C implies that
alg T H(C) contains the compact operators and that the commutator ideal equals
alg T H(C) ∩ ker T (see [9]).
If 1 < p < ∞ and S is a closed subalgebra of the symbol algebra L∞ (T) which
contains C(T), then the Toeplitz algebra alg T(S) ⊂ L(H p ) contains all compact
operators. An analogous result holds for concrete Toeplitz algebras on lp and
Lp (R), 1 < p < ∞, if S is a closed subalgebra of corresponding multiplier algebras
which contains all functions in the Wiener algebra over T and R, respectively.

References
[1] J. Barria, On Hankel operators not in the Toeplitz algebra. Proc. Amer. Math. Soc.
124 (1996), 1507–1511.
[2] J. Barria and P.R. Halmos, Asymptotic Toeplitz operators. Trans. Amer. Math. Soc.
273 (1982), 621–630.
[3] A. Böttcher and Yu.I. Karlovich, Carleson Curves, Muckenhoupt Weights, and
Toeplitz Operators. Birkhäuser, Basel, 1997.
[4] A. Böttcher and B. Silbermann, Analysis of Toeplitz Operators. 2nd ed., Springer,
Berlin, 2006).
[5] A. Brown and P.R. Halmos, Algebraic properties of Toeplitz operators. J. Reine
Angew. Math. 213 (1964), 89–102.
[6] J.W. Calkin, Two-sided ideals and congruences in the ring of bounded operators in
Hilbert space. Ann. Math. 42 (1941), 839–873.
[7] P.R. Chernoff, Representations, automorphisms, and derivations of some operator
algebras. J. Funct. Anal. 12 (1973), 275–289.
 Toeplitz and Hankel Algebras – Axiomatic and Asymptotic Aspects 315

[8] M. Didas, On the structure of Hankel algebras. Integr. Equ. Oper. Theory 80 (2014),
511–525.
[9] R.G. Douglas, Banach Algebra Techniques in Operator theory. Academic Press, New
York 1972.
[10] R.G. Douglas, Banach algebra techniques in the theory of Toeplitz operators. CBMS
Regional Conf. Ser. Math. 15, Amer. Math. Soc., Providence, R. I., 1973.
[11] A. Feintuch, On asymptotic Toeplitz and Hankel operators. Oper. Theory: Adv. Appl.
41, Birkhäuser 1989, 241–254.
[12] I.C. Gohberg, A.S. Markus, and I.A. Feldman, Normally solvable operators and ideals
associated with them. Amer. Math. Soc. Transl. 61 (1967), 63–84.
[13] I. Gohberg and I. Feldman, Convolution Equations and Projection Methods for Their
Solution. Nauka, Moskva 1971 (Russian; English transl.: Am. Math. Soc. Transl. of
Math. Monographs 41, Providence, R. I., 1974).
[14] R. Hagen, S. Roch, and B. Silbermann, C ∗ -Algebras and Numerical Analysis. Marcel
Dekker, Inc., 2001.
[15] L. Hörmander, Estimates for translation invariant operators in Lp spaces. Acta
Math. 104 (1960), 93–140.
[16] V. Peller, Hankel Operators and Their Applications. Springer, Berlin, 2003.
[17] S.C. Power, C ∗ -algebras generated by Hankel operators and Toeplitz operators. J.
Funct. Anal. 31 (1979), 52–68.
[18] S.C. Power, Hankel operators on Hilbert space. Bull. London Math. Soc. 12 (1980),
422–442.
[19] S. Roch and B. Silbermann, Toeplitz-like operators, quasicommutator ideals, numer-
ical analysis I. Preprint P-MATH-22/83, Institut für Mathematik der AdW, Berlin
1983.
[20] S. Roch and B. Silbermann, Toeplitz-like operators, quasicommutator ideals, numer-
ical analysis I. Math. Nachr. 120 (1985), 141–173.
[21] S. Roch and B. Silbermann, Toeplitz-like operators, quasicommutator ideals, numer-
ical analysis II. Math. Nachr. 134 (1987), 245–255.
[22] M. Singer and J. Wermer, Derivations on commutative normed algebras. Math. Ann.
129 (1955), 260–264.

Steffen Roch
Technische Universität Darmstadt
Fachbereich Mathematik
Schlossgartenstrasse 7
D-64289 Darmstadt, Germany
e-mail: [email protected]
Bernd Silbermann
Technische Universität Chemnitz
Fakultät für Mathematik
D-09107 Chemnitz, Germany
e-mail: [email protected]
Operator Theory:
Advances and Applications, Vol. 267, 317–358

c Springer International Publishing AG, part of Springer Nature 2018

More Than 40 Years of Algebraic Techniques

in Numerical Analysis
Pedro A. Santos

Abstract. Algebraic techniques in Numerical Analysis represents the use of

Banach and C∗ -algebra theory to help answer questions and solve problems
in Numerical Analysis. On the forty-fourth anniversary of the seminal work
by Kozak relating stability of an approximation sequence with invertibility
in a suitably chosen Banach algebra, the author presents a review of the
main issues, history, results and a few of the remaining challenges facing this
interdisciplinary area of research, on the frontier between Operator Theory,
Banach and C∗ -algebras, and Numerical Analysis.

Mathematics Subject Classification (2010). Primary 65-02, 65J10; Secondary

46H99, 47B35, 47L99.
Keywords. Algebraic methods, numerical analysis, history.

1. Introduction
Let X, Y be Banach spaces and denote by B(X, Y ) the Banach algebra of all
bounded linear operators on X with values in Y , with the usual norm
A := sup Ax/x.
x=0

We write B(X) for B(X, X).

Given A ∈ B(X, Y ), consider the equation
Ax = y (1)

with y ∈ Y given, and x ∈ X unknown. If the dimension of X is very large or

infinite, it is usually not possible to solve such an equation directly. To solve the
equation (1) numerically by a direct method, formally one can specify a sequence of
simpler (e.g., acting on a finite-dimensional space, or being an integral operators
318 P.A. Santos

over a compact interval) operators Aτ , with τ in some (eventually partially1 )

ordered unbounded set I, which “approximate” A, in the following way. Let Pτ ∈
B(X) and Rτ ∈ B(Y ) be projections such that Pτ , Rτ → I strongly and let Aτ be
a sequence of bounded operators Aτ : Im Pτ → Im Rτ . Put yτ = Rτ y and replace
the equation Ax = y by the sequence of the (simpler) equations Aτ xτ = yτ . The
crucial question is if this method applies, i.e., if there is a τ0 such that the equations
Aτ xτ = yτ possess unique solutions for every right-hand side y for τ ≥ τ0 , and if
the sequence (xτ )τ ≥τ0 converges to a solution x of the original equation Ax = y.
Note that the notion of applicability implies that A is surjective.
One could think that the invertibility of the operator A and strong conver-
gence of Aτ to A would be necessary and sufficient for the approximation method
to apply. But, unless the operators Aτ approximate A in the operator norm, that
is not always the case, as the following simple example shows. We will consider X
and Y as the space lZ2 of the double infinite sequences (uj )+∞ j=−∞ which are square
summable:
⎛ ⎞1/2

+∞ 
+∞
|uj |2 < ∞, u := ⎝ |uj |2 ⎠
j=−∞ j=−∞

and the respective canonical basis. Two simple operators on lZ2 are the forward
shift defined as (Su)j := uj−1 and its inverse. They are unitary operators. Another
operator acting on lZ2 is the finite section projection Pn , n ≥ 0, defined by

uj if − n ≤ j ≤ n
(Pn uj ) :=
0 if j < −n or j > n.
Given any operator A on lZ2 , the sequence (An )n>0 defined by An := Pn APn
approximates A in the strong sense, that is, (A − An )u → 0 for any fixed u ∈ lZ2 .
Approximating the operator A by the operator An defined above is applying the
Finite Section Method to A.
Example 1.1. Let A = I − 2S. This is an invertible operator, as can easily be seen
by writing
1
A = −2S(I − S −1 ).
2
The inverse is given by the Neumann series
 −1 5+∞  k 6  
1 −1  1 1
−1 −1 −1
A = I− S (−2S) = S − S −1
2 2 2
k=0

+∞
=− 2−k−1 S −k−1 .
k=0

1 Inorder not to overburden the notation and proofs with technical details, we will consider in
general the index set to be Z+ or R+ . But most results carry on to more general index sets, as
can be seen for instance in [55, Chapter 4].
 40 Years of Algebraic Techniques in Numerical Analysis 319

Representing A and A−1 as infinite matrices using the canonical basis, one obtains
the Laurent matrices
⎡ ⎤
.. .. .. .. .. ..
. . . . . .
⎢ ⎥
⎢. . .. ⎥
⎢ . 1 0 0 0 0 .⎥
⎢ ⎥
⎢. . .. ⎥
⎢ . −2 1 0 0 0 .⎥
⎢ ⎥
⎢. . .. ⎥
A = ⎢ . 0 −2 1 0 0 . ⎥
⎢ ⎥
⎢. . ⎥
⎢ .. 0 0 −2 1 0 . .⎥
⎢ ⎥
⎢. .. ⎥
⎢ .. 0 0 −2 1 .⎥
⎣ 0 ⎦
.. .. .. .. ..
. . . . .

and
⎡ ⎤
.. .. .. .. .. ..
. . . . . .
⎢ ⎥
⎢. . .. ⎥
⎢ . 0 2 −1
2 −2
2 −3
2 −4 .⎥
⎢ ⎥
⎢. . .. ⎥
⎢ . 0 0 2−1 2−2 2−3 .⎥
⎢ ⎥
⎢ .. ⎥
A−1
= − ⎢. . . 0 0 0 2−1 2−2 .⎥ ,
⎢ ⎥
⎢. ⎥
.. ⎥
⎢ .. 0 0 0 0 2−1 .⎥
⎢
⎢. .. ⎥
⎢ .. .⎥
⎣ 0 0 0 0 0 ⎦
.. .. .. .. ..
. . . . .
where the 00-entry is underlined. Now the operator An = Pn APn is represented
as the (2n + 1) × (2n + 1) finite matrix
⎡ ⎤
1 0 0 ... 0
⎢ .. ⎥
⎢−2 1 0 . 0⎥
⎢ ⎥
⎢ . .. ⎥
An = ⎢ 0 −2 1 .. .⎥
⎢ ⎥
⎢ . . . . ⎥
⎣ .. .. .. .. 0⎦
0 0 . . . −2 1
with inverse
⎡ ⎤
1 0 0 ... 0
⎢ .. ⎥
⎢ 2 1 0 . 0⎥
⎢ ⎥
⎢ .. .. ⎥ .
(An )−1 =⎢ 4 2 1 . .⎥
⎢ ⎥
⎢ .. .. .. .. ⎥
⎣ . . . . 0⎦
22n ... 4 2 1
320 P.A. Santos

So in this case A and its finite section are both invertible (for any n > 0) but the
inverses (An )−1 do not approximate A−1 in any way.
Other types of mismatch between an infinite-dimensional operator and its
finite-dimensional approximations also occur when one considers the operator spec-
trum. In the above example, for instance, σ(I − 2S) is a circle of radius 2 centered
on the point 1, while σ(An ) = {1}.
The applicability of an approximation method Aτ x = yτ to an operator A is
closely related to the stability of the operator sequence (Aτ ), i.e., to the existence
of an index τ0 ∈ I, the invertibility of the operators Aτ for τ > τ0 and to the
uniform boundedness of the norms of their inverses. The result is due to Polski:
Proposition 1.2 (Polski, 1963). If Aτ → A strongly, then the approximation method
applies to A if and only if the sequence (Aτ ) is stable and the operator A is in-
vertible.
Proof. (⇒) If the approximation method applies to A then A−1
τ Rτ y converges
for all y ∈ Y . Then, by the Uniform Boundedness Principle we obtain that
supτ >τ0 (A−1
τ )Rτ  < ∞. Regarding the invertibility write

Pτ x − A−1 −1
τ Rτ Ax ≤ Aτ Rτ  Aτ Pτ x − Ax.

Because of the stability and the strong convergence of Aτ Pτ to A, the left-hand

side goes to zero as τ → ∞. If x is in the kernel of A, this implies in particular
that Pτ x → 0, that is, x = 0. So the kernel of A is trivial. Because the range of A
is Y by definition of applicability of an approximation method, we conclude that
A is invertible.
(⇐) Let xτ denote the solution of the approximate equation Aτ Pτ xτ = Rτ y and
x that of the original equation. If the sequence (Aτ ) is stable, then
x − xτ X = (Pτ + I − Pτ )x − (Aτ )−1 Rτ yX
= (Aτ )−1 Aτ Pτ x + (I − Pτ )x − (Aτ )−1 Rτ AxX
≤ (Aτ )−1 Rτ Aτ Pτ x − AxX + (I − Pτ )xX → 0. 
We have thus a necessary and sufficient condition for an approximation
method to apply, the stability of its approximation sequence. One can say that
the above statement can be resumed in the expression applicability = approxima-
tion + stability.
There are some instances where the stability problem is easy to solve. For
example, let H be a Hilbert space, A a positive definite operator, and (Pn )n≥1
be a sequence of self-adjoint projections which converges strongly to the identity
operator. By the positive definiteness, there is a positive constant C such that
Ax, x ≥ C x2 for all x ∈ H.
Then it is immediate that
Pn APn xn , xn  = APn xn , Pn xn  ≥ C xn 2 for all xn ∈ Im Pn ,
 40 Years of Algebraic Techniques in Numerical Analysis 321

and the operators An : Im Pn → Im Pn are invertible for each n ≥ 1 with the

norms of their inverses not greater that 1/C. So, the sequence (Pn APn ) is stable
and one even has τ0 = 1. Outside of these particular cases, the stability problem
for more general types of operator proves to be a very delicate one. The insight
by Polsky does not yet allow us to understand what characterizes operators and
approximating methods for which there is applicability. The road to obtain those
results was a long one, and started in the second part of the 20th century.
This paper presents an historical view of the major steps of this long path
with some emphasis on the author’s almost 25 years contact and contribution to
this vast area of mathematical research. The objective is to present a readable and
interesting account for the non-specialist, which allow him or her to understand
the main issues and have pointers to a more in-depth look at some of the topics.
We start in the next section by describing the grounding ideas, and then pro-
ceed dedicating some space to each generalization and use of algebraic techniques
to tackle new problems. Finally we present a couple of possible future research
directions on this already mature subject.

2. The pioneering works

During the sixties and seventies of last century, there was a lot of interest in the
finite section method (FSM) for Toeplitz operator matrices. We introduce here the
context and definition of such operators.

2.1. The context and the problem

Let T represent the unit circle on the complex plane and let Lp (T), 1 ≤ p < ∞
denote the Banach space of (the classes of) complex-valued measurable functions
a on T where the Lebesgue integral
! 2π
|a(eiθ )|p dθ =: app
0

is finite, endowed with the norm  · p . Moreover, let L∞ (T) denote the space of all
essentially bounded measurable functions, with norm a∞ := ess sup t∈T |a(t)|.
For 1 < p < q < ∞ one has
L∞ (T) ⊂ Lq (T) ⊂ Lp (T) ⊂ L1 (T).
Given a ∈ L1 (T), it is possible to define its Fourier coefficients an (n ∈ Z) by
! 2π
1
an := a(eiθ )e−inθ dθ.
2π 0
There is an isometry between the Hilbert space L2 (T) and thedouble infinite
sequence space lZ2 . A function a belongs to L2 (T) if and only if n∈Z |an |2 < ∞

and in that case a22 = 2π n∈Z |an |2 . Moreover, the set of the trigonometric
polynomials χn (t) := tn , {χn }n∈Z for t ∈ T, forms an orthogonal basis in L2 (T).
322 P.A. Santos

The sequence spaces lZp (resp lp ), 1 ≤ p < ∞, are defined as the Banach space of
all sequences ϕ = (ϕn )n∈Z (resp ϕ = (ϕn )n∈Z+ ) of complex numbers for which
5 6
 
|ϕn | < ∞
p
resp. |ϕn | < ∞
p

n∈Z n∈Z+

with Z := {n ∈ Z : n ≥ 0}. Given a function a ∈ L∞ (T), the multiplication

+

operator M (a) : L2 (T) → L2 (T), u → au canbe written in the lZ2 space as a

convolution (un ) → ((a ∗ u)n ) with (a ∗ u)n := k∈Z an−k uk . That operator can
thus be represented as a double infinite matrix with constant entries along the
diagonals and is known as the Laurent operator with symbol a. Not all a ∈ L∞ (T)
generate bounded operators in lp for p = 2. The set of all a ∈ L1 (T) that do
generate bounded Laurent operators is denoted by M p and its elements are called
multipliers. For a detailed construction of the corresponding operators for p = 2 see
[26, Chapters 1 and 2]. The Wiener algebra W is constituted by all all a ∈ L1 (T)
whose Fourier coefficients belong to lZ1 . We have that W ⊂ M p for all p.
The discrete Riesz projection P : lZp → lp is defined as (un )n∈Z → (un )n∈Z+ .
One can also think of the Riesz projection as acting on lZp and putting to zero all
elements of the sequence with negative indices. Given a ∈ M p , then the Toeplitz
operator on lp generated by a is defined by
T (a) : lp → lp , u → P (a ∗ u) ( = P M (a)P u )
and can be represented as a one-sided infinite matrix
[aij ]i,j≥0 , aij := ai−j .
In fact, infinite Toeplitz and Laurent matrices form one of the simplest class of
infinite-dimensional operators one can think of that can be non-compact, and in
general, non-self-adjoint or positive definite. And the FSM for matrices is one of
the simplest finite-dimensional strongly convergent approximation methods.
Besides applicability of the FSM, related subjects were also studied by math-
ematicians in the mid 20th century, mainly in the USA and USSR. So, for instance
Baxter [5] obtained in 1963 results on the applicability of the FSM for Toeplitz
operators with generating function in the Wiener algebra, while also working on
determinant asymptotics. Determinant asymptotics of Toeplitz matrices are an
older subject, with several seminal results obtained by Szegő [150, 151] since 1915,
and then generalized by several authors until now [2, 26, 39] (see also Section 3.4).
From 1965 to 1971 Gohberg and Feldman developed a theory on the appli-
cability of the FSM for convolution type operators (including both the discrete
and continuous forms, paired operators, multi-dimensional case, etc) with contin-
uous generating function, which was published as a book [46]. Widom [155, 156]
studied the asymptotic inversion of convolution operators, to obtain estimates for
determinants, and behavior of the spectrum.
A result by Kozak from 1973, would completely change the perspective, af-
ter being combined with an unrelated observation from Widom. Kozak, following
a suggestion from his thesis adviser Dybin [26, Section 7.11] observed that the
 40 Years of Algebraic Techniques in Numerical Analysis 323

stability problem for a sequence was equivalent to an invertibility problem in a

suitably constructed Banach algebra [81].
Let E be the set of all bounded sequences A := (Aτ )τ ∈I of operators Aτ ∈
B(X). Provided with the operations
(Aτ ) + (Bτ ) := (Aτ + Bτ ), α(Aτ ) := (αAτ ), (Aτ )(Bτ ) := (Aτ Bτ )
and with the norm (Aτ ) := supτ Aτ B(X) , the set E becomes a unital Banach
algebra. Note that the constant sequences (A) are included in E for every operator
A ∈ B(X). If X is a Hilbert space, then we can equip E with the involution
(Aτ )∗ := (A∗τ ) which makes E a C ∗ -algebra.
Let G be the set of all sequences (Aτ ) ∈ E with limτ →∞ Aτ  = 0. This set
forms a closed two-ideal in E.
Theorem 2.1 (Kozak). A sequence A ∈ E is stable if and only if A + G is invertible
in E/G.
Proof. (⇒) If A = (Aτ )τ ∈I and there exists a τ0 such that Aτ is invertible for
τ > τ0 and the norms of the inverses are uniformly bounded, then one can define
a sequence of operators Bτ by

A−1
τ if τ > τ0
Bτ := .
0 if τ ≤ τ0
Defining B := (Bτ )τ ∈I ∈ E, we have that B + G is the inverse of A + G in E/G.
(⇐) There exists B := (Bτ )τ ∈I ∈ E such that
Bτ Aτ = I + Gτ and Aτ Bτ = I + Gτ
with Gτ , Gτ  → 0. There is a τ0 such that for τ > τ0 both Gτ  and Gτ 
are less than 1/2. Then I + Gτ and I + Gτ are invertible and Aτ has both a right
and left inverse. Moreover,
A−1
τ  ≤ (I + Gτ )
−1
Bτ  ≤ 2Bτ . 

With this algebraization of the stability problem, Kozak arrived at studying

invertibility in a non-commutative Banach algebra. The other main achievement
of [81] was that Kozak generalized Simonenko’s local principle, used for the in-
vestigation of the Fredholm property of operators, to a local principle for explor-
ing invertibility in general Banach algebras. This generalization of Simonenko’s
local principle was powerful enough to treat the finite section method for multi-
dimensional convolution operators with continuous symbols. Several attempts to
extend Kozak’s method to non-continuous symbols failed because the sequences
that resulted were not of “local type”, a key feature needed to use Simonenko’s
local principle (see, e.g., [120, Section 2.5]).
Subsequently other general local principles were introduced by several au-
thors. They are all based on the happy circumstance that if a Banach algebras has
a non-trivial center, then the invertibility problem in this algebra can be reduced
324 P.A. Santos

to a collection of invertibility problems in simpler related algebras. A concise sur-

vey of this topic is in [120, Chapter 2], with proofs. Below we present the version
which is known as local principle by Allan/Douglas. A different but equivalent
(in the complex Banach algebra setting) formulation exists in the form of the
Gohberg–Krupnik local principle introduced in [49].
Theorem 2.2 (Allan/Douglas). Let A be a Banach algebra with identity e and let
C be closed central subalgebra of A containing e. Let M (C) be the maximal ideal
space of C, and for ω ∈ M (C), let Iω refer to the smallest closed two-sided ideal of
A containing the ideal ω. Then an element a ∈ A is invertible in A if and only if
a + Iω is invertible in the quotient algebra A/Iω for all ω ∈ M (C).
In the context of the above theorem, if Iω = A one considers a+ Iω invertible
for all a ∈ A.
To apply a local principle like the one described above, a specific structure
in the algebra is needed, namely it must possess a central subalgebra. In order to
describe the search to find such a subset in the sequence algebra E when considering
finite sections of Toeplitz operators, one needs to introduce additional concepts like
that of an Hankel operator. Let Q := I − P and define the flip operator J : lZp → lZp
by the rule (Ju)n = u−n−1 . Note that J 2 = I and P J = JQ. The Hankel operator
H(a) on lp is defined as the operator
H(a) := P M (a)QJ
and can be represented as a one-sided infinite matrix
[aij ]i,j>0 , aij := ai+j−1 .
This Hankel operator “uses” only the positive-indexed Fourier coefficients of a.
The negative-indexed coefficients of a are used in H(ã) where ã(t) = a(1/t), that
is, M (ã) = JM (a)J, H(ã) = JQM (a)P . The following proposition relates the
Toeplitz and Hankel operators with the products of their symbols [26, Chapter 2].
In this form, (3) and (4) appeared for the first time in Widom’s paper [157].
Proposition 2.3. Let a, b ∈ M p . Then ab ∈ M p and
M (ab) = M (a)M (b), (2)
T (ab) = T (a)T (b) + H(a)H(b̃), (3)
H(ab) = T (a)H(b) + H(a)T (b̃). (4)

Proof. We just write the proof of (3):

T (ab) = P M (ab)P = P M (a)M (b)P = P M (a)P M (b)P + P M (a)QM (b)P
= T (a)T (b) + P M (a)QJJQM (b)P = T (a)T (b) + H(a)H(b̃). 

A result by Hartman [62] from 1958 implies that if a belongs to Cp , the closure
of W in M p , then the operators H(a) and H(ã) are compact. When a and b are
 40 Years of Algebraic Techniques in Numerical Analysis 325

piecewise continuous functions2 , the respective Hankel operators are not compact
in general, but the product H(a)H(b) is compact if a and b have no common points
of discontinuity [26, Proposition 6.29].
To return now to the main narrative, beginning with Baxter’s paper referred
above, a lot of further proofs of the applicability of the finite section method to
invertible Toeplitz operators T (a), a ∈ Cp , or a belonging to some special classes
in the l2 setting were presented (see, for instance, [31, 46, 104]). One proof is based
on Widom’s formula [157]:
Tn (ab) = Tn (a)Tn (b) + Pn H(a)H(b̃)Pn + Wn H(ã)H(b)Wn , (5)
where Tn (a) = Pn T (a)Pn , Pn (u0 , u1 , . . .) := (u0 , u1 , . . . , un−1 , 0, 0, . . .) and
Wn (u0 , u1 , . . .) := (un−1 , . . . , u0 , 0, 0, . . .). Clearly, (5) is the counterpart of for-
mula (3) for finite sections of Toeplitz operators T (a). Note that (5) can be used
to give asymptotic inverses Bn for a stable sequence (Tn (a)), a ∈ Cp , that is
Bn Tn (a) − Pn  → 0, Tn (a)Bn − Pn  → 0 as n → ∞.
Indeed, in this case one can check that a is invertible, set Bn = Tn (a−1 ) +
Pn K(a)Pn + Wn K(ã)Wn , and notice that the invertibility of T (a) implies the
invertibility of T −1 (ã) and ã−1 , where K(a) = T (a)−1 − T (a−1 ) and K(ã) =
T −1 (ã) − T (ã−1 ) are compact operators (see for instance [25]).
In 1977, Verbickiı̆ and Krupnik published the paper [153] where they proved
that if a is a piecewise continuous function with only one discontinuity, having a
mild additional multiplier property, then the finite section method is applicable to
T (a) in lp if and only if T (a) and T (ã) are invertible.
Using the above two results, around 1980, Böttcher and Silbermann devel-
oped a technique of separation of discontinuities. This technique allowed them to
reduce problems for Toeplitz operators with finitely many discontinuities in the
symbols to problems for compactly perturbed Toeplitz operators whose symbols
have only one discontinuity. In this connection the following result from [46] is of
importance.
Proposition 2.4 (Gohberg, Feldman). Let A be a bounded operator and K a com-
pact operator. If the finite section method applies to A and A + K is invertible,
then the finite section method applies to A + K.
Combining their separation technique with (3), the Verbickiı̆–Krupnik result
for a single jump, and Proposition 2.4, Böttcher and Silbermann [22] were able to
extend the Verbickiı̆–Krupnik result to symbols with finitely many jumps.

2.2. Silbermann’s insight

The new result using the separation technique would only be published in 1982
but, already in 1980, Silbermann had started to look for a way to generalize it

2A function a ∈ L∞ is said to be piecewise continuous, in short a ∈ P C(T), if a has well-defined

one-sided limits at all points t ∈ T.
326 P.A. Santos

for a countable number of discontinuities and also for systems of Toeplitz opera-
tors, that is, for generators in the set P Cp (T) := P C ∩ M<p> , with P C the set
of piecewise continuous functions and M<p> the set of all functions from L∞ (T)
that belong to all Mp̂ for p̂ in a neighbourhood of p 3 [12]. He thought to use
the algebraization technique from Kozak’s thesis with the local principle of Go-
hberg and Krupnik. But there was a problem. The algebra E/G resulting from the
application of Kozak’s method, with E defined by
E = {(An )n∈Z+ : An : Im Pn → Im Pn },
seemed to have non-useful center for applying the local principle (it is connected
with each index of the sequence and isomorphic to l∞ – the local algebras are either
trivial or isomophic to E/G again). Something must commute were his thoughts
at the time, as he recalled many years later [14]. In fact, for operators in B(lp ), if
f ∈ C(T) and a ∈ P Cp (T), then T (a)T (f ) − T (f )T (a) is compact and invertibility
in the Calkin algebra generated by the cosets of the form T (a) + K, a ∈ P Cp (T)
can be studied using local principles because there exists a central subalgebra
C = {T (f ) + K : f ∈ C(T)}. Applying the idea to finite sections of Toeplitz
operators one obtains from Widom’s formula
Tn (a)Tn (f ) − Tn (f )Tn (a) = Pn K1 Pn + Wn K2 Wn
with K1 , K2 compact operators. Silbermann’s incredible fruitful idea was then to
check if the set
 
J := (Pn K1 Pn + Wn K2 Wn + Gn ), K1,2 compact, (Gn ) ∈ G
would form a closed ideal of E or some related subalgebra, and if it was possible to
relate invertibility in E/G with invertibility in the new algebra. It turned out that
in both cases his ingenious intuition was right. The set J is an ideal if there are
operators A and à ∈ B(lp ) such that the sequences (An ) ∈ E have strong limits
s-lim An = A, s-lim Wn An Wn = Ã,
n→∞ n→∞
(6)
s-lim A∗n ∗
=A , s-lim (Wn An Wn )∗ = Ã∗ ,
n→∞ n→∞

where A∗ and Ã∗ are bounded operators on lq , with 1p + 1q = 1. We will denote

the set of the sequences that possess such strong limits as F . The limit operators
depend on the sequence (An ) under consideration. These sequences form a closed
Banach subalgebra F of E, and one just needs now to relate invertibility in E/G
with invertibility in F /J . The proposition below, which appears in [142], is the first
of what in this context were latter called lifting theorems. Its proof makes use of
the interplay between multiplication by compact operators and weak/strong/norm
convergence. For a modern and general perspective on lifting theorems, see [120].
Proposition 2.5 (Silbermann). A sequence A ∈ F is stable if and only if the cor-
responding operators A and à are invertible, and the coset A + J is invertible
in F /J .
3 The use of M<p> over Mp facilitates some technical details in the proofs.
 40 Years of Algebraic Techniques in Numerical Analysis 327

Invertibility in F/J (in fact, in specific subalgebras like alg {(Tn (a))+J , a ∈
P Cp }) could then be studied using the local principle of Gohberg and Krupnik, in
analogy with studying invertibility in the Calkin algebra for regular operators. In
fact, in the case under study, invertibility of the coset A + J in F/J was already
implied by the invertibility of the two operators A and Ã. So, Silbermann obtained
the elegant result:
Theorem 2.6 (Silbermann). If T (a) is a bounded Toeplitz operator on [lp ]n with
symbol a ∈ [P Cp (T)]n×n , then the FSM applies to T (a) if and only if T (a) and
T (ã) are invertible.
Silbermann was not only able to prove the desired result regarding the appli-
cability of the FSM to Toeplitz operators (or even systems of Toeplitz operators)
with a countable number of discontinuities in their symbol, but the fundamen-
tal idea, finding a compact-like sequence ideal, would form the basis of a whole
new branch in theoretic numerical analysis. The approach taken by Silbermann to
study stability was later systematized by Böttcher, Silbermann and Roch into the
following steps (see Figure 1).
Let A be an algebra containing sequences of operators for which one wants
to obtain stability criteria.

Figure 1. The algebraic approach to stability problems

1. Algebraization: Find a unital Banach algebra E which contains A and a

closed ideal G ⊂ E such that the stability problem becomes equivalent to an
invertibility problem in the quotient algebra E/G.
328 P.A. Santos

2. Essentialization: Find a unital subalgebra F of E which contains A and G,

and a closed ideal J of F such that the algebra AJ := (A + J )/J has a
large center.
To be able to use the latter property, one has to guarantee that no
essential information is lost regarding the invertibility of a coset of a sequence
A ∈ F in the algebra E/G and the invertibility of the same sequence modulo
the ideal J . This is done employing inverse-closedness results and a lifting
theorem, which relates invertibility in those algebras with the invertibility of
a collection of limit operators, indexed by a set T.

3. Localization: Use a local principle to translate the invertibility problem in the

algebra AJ to a family of simpler invertibility problems in local algebras. In
this step one must, if using Allan/Douglas’ local principle, to find a suitable
central subalgebra and its maximal ideal space M . The central subalgebra is
connected with the ideal J of the previous step.
Then one uses the local principle to translate the invertibility problem
from F /J to the family of local algebras AJ w , with w ∈ M .

4. Identification: Find necessary and sufficient conditions for the invertibility of

elements in the local algebras.
In some cases, we are in the situation that the local algebras have a
relatively simple structure and can be described as being generated by two
projections, or isomorphic to known operator algebras. In the latter case the
isomorphism can be proved using the concept of limit operators, which were
already introduced by Favard [43]. Occasionally, in order to treat more in-
volved local algebras, a repeated localization can be useful. The identification
of the local algebras often is the “hardest” part of the stability problem.
Of course, all the steps described above are closely related. For instance,
finding and using a larger ideal J in the essentialization step will in principle in
the identification step wield simpler local algebras.
Two simple but concrete illustrations of the above approach can be found in
[120, Sections 6.4 and 6.5], and an alternative description of the method in [56,
Section 6.1.1]. Here we will provide a very simple illustration of the application
of the method to Toeplitz operators on l2 with continuous generating function.
We will also include some remarks on the case of piecewise continuous functions.
The interested reader can find the detailed treatment of that later case in [56,
Section 4.2].

2.3. Finite Section Method for Toeplitz operators with

continuous generating function
Let a ∈ C(T) or P C(T), and consider the FSM applied to T (a) : l2 → l2 . That is,
to solve the infinite system of equations

T (a)x = y
 40 Years of Algebraic Techniques in Numerical Analysis 329

we consider systems of n equations in n unknowns, given by

Tn (a)xn = Pn y
by selecting the first n lines and n columns of the infinite matrix T (a). The pro-
jection Pn is the one defined after equation (5), selecting the first n entries of a
vector. We would like then to study the stability of the sequences (Tn (a))n∈Z+ .
The algebras we will consider are then exactly the algebras generated by those
sequences, C := alg {(Tn (a))n∈Z+ , a ∈ C(T)}, or the larger A := alg {(Tn (a))n∈Z+ ,
a ∈ P C(T)}. In both cases, the first two steps of the method have no differences.
2.3.1. Algebraization. Let E be the algebra defined by
E = {(An )n∈Z+ : An : Im Pn → Im Pn },
with pointwise defined operations and involution, and the supremum norm. The
algebra E is a Banach (in fact C∗ ) algebra containing A and its subset G of all
sequences of operators tending in the norm to zero forms a closed two-sided ideal.
Kozak’s result then implies that a sequence in E is stable if and only if its coset is
invertible in E/G. But it is not easy in general to check invertibility in E/G, even
for the sequences belonging to C or A.
2.3.2. Essentialization. Consider now the subset F of E formed by all sequences
A = (An ) such that the following four strong limits exist:
s-lim An = W0 (A), s-lim Wn An Wn = W1 (A),
n→∞ n→∞
(7)
s-lim A∗n = W0 (A)∗ , s-lim(Wn An Wn )∗ = W1 (A)∗ .
n→∞ n→∞
The set F forms a closed subalgebra of E containing A. Note that the above
strong limits are in reality unital algebra homomorphisms from F into B(l2 ) and
that W0,1 (G) = 0, which means they induce well-defined unital homomorphisms
acting from F /G into B(l2 ). Define the subset
 
J := (Pn K0 Pn + Wn K1 Wn + Gn ), K0,1 compact, (Gn ) ∈ G
of F . One can prove that in fact J is a closed two-sided ideal of F . To illustrate
the interplay between the several concepts (strong and norm convergence, compact
operators, etc) we will show here that it is a left ideal. For that, given sequences
(An ) ∈ F and (Jn ) ∈ J one wants to check that (An )(Jn ) is in J .
An (Pn K0 Pn + Wn K1 Wn + Gn ) = An Pn K0 Pn + An Wn K1 Wn + An Gn .
The sequence (An Gn ) is clearly in G and thus in J . We just have to concern
ourselves with the other parts. Let us start with the sequence (An Pn K0 Pn ). One
can write
(An Pn K1 Pn ) = (Pn W0 (A)K0 Pn ) + (Pn (An − W0 (A))K0 Pn )
and the first part of the left-hand side is clearly in J . The second part is in G
because
Pn (An − W0 (A))K0 Pn  ≤ (An − W0 (A))K0  → 0
330 P.A. Santos

as n goes to infinity due to the strong convergence of An − W0 (A) to zero and the
compactness of K0 . Likewise, for the sequence (An Wn K1 Wn ),
(An Wn K1 Wn ) = (Wn W1 (A)K1 Wn ) + (Wn (Wn An Wn − W1 (A))K1 Pn )
because Wn2 = Pn and so the same above reasoning can be made. The result is
proved.
Now it is necessary to relate invertibility in F /G with invertibility in F /J .
This is done using the special properties of the homomorphisms. Usually one uses
a general lifting theorem. For this example, and also to give some insight on the
lifting theorem itself, I will give a sketch of a proof.
Theorem 2.7 (Lifting). An element A + G is invertible in F /G if and only if the
operators W0 (A) and W1 (A) are invertible and the coset A + J is invertible in
F /J .
Proof. (⇒) It is clear that if A + G is invertible then A + J is invertible in F /J
and, being W0,1 unital homomorphisms, W0 (A) and W1 (A) are also invertible. The
difficulty of the proof is in the other direction.
(⇐) Consider the kernels of the homomorphisms W0,1 . They are ideals of F and it
is not difficult to see that (note that the sequence of operators (Wn ) tends weakly
to zero)
J ∩ Ker W0 ∩ Ker W1 = G. (8)
Because F /Ker W0,1 is isomorphic to W0,1 (F ), and the W0,1 (A) are invertible,
there exist sequences C0 and C1 in F such that
AC0 + K0 = I and AC1 + K1 = I
where I := (Pn ) is the identity sequence in F and K0,1 ∈ Ker W0,1 . And because
A + J is invertible in F /J we also have that there exists a sequence B ∈ F such
that AB = I + J, with J ∈ J . Then the product
(AC0 − I)(AC1 − I)(AB − I) = AB − I
for some B ∈ F and it is in G by (8). Thus A + G is right-invertible. The proof for
left-invertibility is the same. 
2.3.3. Localization. In this step, one considers the smallest closed algebra formed
by the sequences in C and in J . To apply Allan/Douglas’ local principle we must
find a useful central subalgebra of
C J := (C + J )/J ⊂ F /J .
In fact, in the simple case of continuous functions only, the algebra C J is commu-
tative, as can easily be seen using Widom’s formula (5) and Hartman’s result on
compactness of Toeplitz operators with continuous generator functions:
Tn (a)Tn (b) − Tn (b)Tn (a)
= Pn H(b)H(ã)−H(a)H(b̃) Pn + Wn H(b̃)H(a)−H(ã)H(b) Wn .
 40 Years of Algebraic Techniques in Numerical Analysis 331

For commutative algebras the local principle is equivalent to the usual Gelfand
theory. The local algebras are all isomorphic to the complex field C. The maximal
ideal space of C J coincides with the maximal ideal space of C(T) that is, with T.
Things get more interesting if one considers the case of piecewise continuous
functions. In that case the formula above gives that C J is a central subalgebra of
AJ := (A + J )/J ⊂ F /J .
So, each point w ∈ T corresponds to a maximal ideal of C J which we denote by the
same symbol, and let Iw be the smallest closed two-sided ideal of AJ containing w.
Applying Allan/Douglas’s local principle, invertibility of an element (Tn (a)) + J
in AJ is equivalent to the invertibility of all the local cosets (Tn (a)) + J + Iw ,
w ∈ T.
2.3.4. Identification. In the case of continuous functions only, for each w ∈ T, the
coset (Tn (a)) + J + w ∈ AJ /w coincides with
(Tn (a(w))) + J + w = a(w)I + J + w,
which means in some sense that, locally, the continuous generating function be-
haves as a constant function. The cosets of A = (Tn (a)) are thus invertible in each
local algebra AJ /w if and only if a(w) = 0. So, the coset (Tn (a)) + J is invertible
in F /J if and only if a is invertible in C(T). The invertibility of the generator a is a
consequence of the Fredholm property of the Toeplitz operator T (a), which in our
case already appears as W0 (A). So, in the case under consideration, the condition
of invertibility in F /J gives no extra conditions for the applicability of the FSM,
besides the two strong limits. On the other hand, it is known that for (scalar)
Toeplitz operators, T (a)(= W0 (A)) is invertible if and only if T (ã)(= W1 (A))
is invertible. We can conclude that the FSM is applicable for Toeplitz operators
T (a) : l2 → l2 with a continuous if and only if T (a) is invertible.
Consider now a piecewise continuous function a. At the points w where a
is continuous, the picture is as above: we obtain a scalar local algebra, with only
multiples of the identity. This happens because the coset (Tn (bw )) + J with the
function bw defined by bw (t) := a(w) − a(t) is in Iw ( bw (w) = 0 ) and so belongs
to the 0 coset in the local algebra. If a has a discontinuity at the point w we can
use a similar reasoning. But now, the local behaviour will be piecewise constant.
Let χw be the piecewise constant function on T with the value 1 on the arc from
w to −w (with respect to the common orientation of T) and 0 otherwise. Now the
coset (Tn (aw )) + J + Iw with aw defined by
aw := a(w+ )χw + a(w− )(1 − χw )
coincides with the original (Tn (a))+J +Iw because aw −a is a function continuous
at w and taking the value 0 there, and so (Tn (aw − a)) + J ∈ Iw . The above
reasoning signifies that the local algebra at w is generated by the identity and the
coset (Tn (χw )) + J + Iw . Obtaining invertibility conditions for the local algebras,
as refered in the previous section, is usually the hardest part, and we will not enter
into details here. The idea in this case is that, being a singly generated algebra,
332 P.A. Santos

it is commutative and it can be described via Gelfand theory, and the maximal
ideal space of a singly generated algebra is homeomorphic to the spectrum of its
generator. The spectrum in this case can be shown to be the interval [0, 1] (see
[56, Proposition 4.26] for details). Thus, the coset (Tn (aw )) + J + Iw is invertible
in AJ /Iw if and only if the line segment in the complex plane joining a(w+ ) to
a(w− ) does not contain the point 0. In fact, in this case also, that condition is a
necessary condition for the invertibility of the operator T (a) = W0 (A), and so the
end result is as above.

3. Extension and application

Direct applications of the new method to powers of Toeplitz operators, sums
and products and elements of the algebra generated by those types of opera-
tors appeared almost immediately in [107, 108, 122]. As if to show the hard road
ahead, there was an overlook when applying the complex machinery of Gohberg–
Krupnik’s local principle in [107, 108] that rendered the general result obtained in
those papers false. The issue was detected and solved by Andreas Rathsfeld (see
[55, Section 4.7] for more details). A first abstract and general view of these results
was published in 1985 [123].
Incidentally, a concrete invertible Toeplitz operator for which the Finite Sec-
tion Method is not applicable was found by Treil in 1987 [152]. That example
shows the applicability problem has indeed no simple solution to be found.
The new approach to stability problems pioneered by Kozak and given wide
application possibilities by Silbermann would over the next 35 years be extended
in several dimensions and applied to a great variety of problems. It would also
reflect back into Banach and C∗ -algebra theory, by supplying that field with many
new examples and new problems.

3.1. Convolution operator algebras

The new approach’s first application, from the original 1981 paper [142], was heav-
ily inspired by the study of the Calkin algebra connected to Toeplitz operators and
their commutation properties. To apply the method to other types of operators,
the structure of the algebras they generated and their commutation properties had
to be known. Continuous convolution operators were an obvious early candidate
for application, and thus a great deal of study of the algebras generated by opera-
tors of that type (singular integral and Wiener–Hopf operators) followed. Certain
operators on a contour Γ ⊂ C can be formally defined as
!
(Au)(t) := λu(t) + k(t − s)u(s) ds
Γ
where λ ∈ C and k can be understood to be either a function or a distribution,
depending on Γ. In particular, when k(x) = − πi 1 1
x we obtain the Singular Integral
Operator (SIO) on Γ. With Γ = R we have a convolution, acting on Lp (R) for
instance, and its restriction to Lp (R+ ) is called a Wiener–Hopf operator. If Γ is
 40 Years of Algebraic Techniques in Numerical Analysis 333

the unit circle T in the complex plane, one defines the Hardy spaces H p of the
functions with only positive indexed Fourier coefficients and the continuous version
of the Riesz projection is
5 +∞ 6
 +∞
P :L →H ,
p p
P ai t →
i
a i ti
i=−∞ i=0

with the Toeplitz operator defined as

Ta : H p → H p , u → P (au).
For the theory of these types of operators see, e.g., [26] and the references cited
therein.
Silbermann applied his newly discovered method to the FSM for singular
integral equations in [143]. Böttcher [6, 7] was the first to prove criteria for the
applicability of the finite section method to Wiener–Hopf integral operators on
Lp (R+ ). He used Silbermann’s localization technique to reduce the problem to
case of a symbol with a single jump. While in the discrete case, that is, for Toeplitz
operators on p whose symbol has a single jump, the analogous problem had been
solved earlier by Verbickiı̆ and Krupnik [153], Böttcher had to find a new approach
to settle the Lp case. He observed in particular for the first time that it is not
only two operators but rather a whole family of operators whose invertibility is
necessary and sufficient for the applicability of the finite section method. This was
the first appearance of the lentiform massive local spectra in the theory of singular
integral operators [13].
The corresponding result for Toeplitz operators on the unit circle followed in
1986 [20], where, for a positive integer n, the finite section of a Toeplitz operator Ta
is defined as the compression of Ta to the (n+1)-dimensional space of trigonometric
polynomials of degree at most n.
Many other results followed (e.g., [21, 125, 127], see also the following sec-
tions) using, systematizing and extending results from authors like Gohberg and
Krupnik [47, 48, 82], Simonenko and Min [149], Sarason [138], Duduchava [37],
etc, sometimes leading to fruitful collaboration with those authors.
There was also an effort to understand the relation and differences between
the three local principles (Simonenko’s [147, 148], Gohberg–Krupnik’s [49] and
Allan/Douglas’ [1, 35]) which culminated in [19]. The developments and collabo-
rations described above fed into applications of the algebraic techniques in many
cases.
3.2. Different operators, function and symbol spaces
Some operators related to Toeplitz operators like band-dominated operators were
a natural venue for the new technique [101, 102]. Banded matrices are matrices
where the non-zero entries are situated on a band around the main diagonal.
The concept can be extended to other type of operators, that are then named
banded operators. Band-dominated operators are norm limits of sequences of band
operators and a generalization of Toeplitz operators with symbols in Cp (see [113]
334 P.A. Santos

for an overview of the theory for operators in l2 ). Other suitable candidates with
some relation to Toeplitz or Wiener–Hopf operators were Hankel operators and
other operators derived from sums and products of simpler multiplication and
convolution operators [116, 118]. In fact, algebraic techniques are ideal to study
properties of operators that result from sums and products of simpler building
blocks. For operators acting in Hilbert spaces, the corresponding algebras are C∗ -
algebras which makes the problems somewhat easier to tackle, due to the nice
norm and inverse-closedness properties of such algebras (see, e.g., [120, Section
1.2] for a short discussion on this subject or [11] which condenses very well many
of the issues in the C∗ -algebra setting). So some results were first obtained for L2
spaces for instance, and only years later for Lp with p = 2.
Besides the above-mentioned pioneer works of Böttcher [6, 7] on Lp (R), Roch
and Silbermann [125] developed a symbol calculus for finite sections of singular
integral operators with flip and piecewise continuous coefficients.
Starting in 1995, as part of his PhD thesis, the author tried to find conditions
for the applicability of the FSM for algebras that included Wiener–Hopf plus
Hankel operators acting on L2 (R+ ). The main difficulties lay in the essentialization
and the identification steps. Regarding the essentialization step, simple adaptations
of the before used homomorphisms did not work because of the flip J, defined by
(Ju)(t) = u(−t), which appears in the Hankel operators, and it was not clear
what these new homomorphisms would look like. Finally it was possible to find a
new family of matrix homomorphisms that could be used. After several months
of fruitless search the inspiration came when looking at the relations expressed
on Section 7.13 of the first edition of [26], which is treating the discrete case
without explicit Hankel operators in the algebra. It crossed the author’s mind
that similar strong limits in the continuous case might be compatible with the
flip, and thus be well defined on the whole algebra generated by Wiener–Hopf
plus Hankel operators, and that those strong limits could be arranged as a single
matrix-operator-valued homomorphism.
When including the flip, the central subalgebra must be constituted by op-
erators defined by even functions. This implies a “double point” localization at
non-fixed points of the flip (which are 0 and ∞). A standard technique for dealing
with this problem is doubling the dimension (see, e.g., [120, Section 1.1.5]), ending
in operator matrices.
In the identification step, a complicated local algebra at infinity appeared,
generated by four non-commuting idempotents, which had to be analysed with
novel techniques [118].
Also, the approximation algebra building blocks were made of the constant
sequences of multiplication operators (aI) and convolution operators (W 0 (b)) for
suitable generators a and b, the flip (J), together with the FSM projection sequence
(Pτ )τ ∈I , acting on the whole space L2 (R). Before, a more restricted definition
was used considering as generators for instance the sequences (Pτ W (a)Pτ ) acting
in Im (Pτ ). This different approach has the advantage of treating the sequence
algebra as a generalization of the usual operator algebra for those operators, and
 40 Years of Algebraic Techniques in Numerical Analysis 335

the stability analysis as an extension of the operator algebra’s Fredholm theory.

The disadvantage is the appearance of complicated local algebras at infinity. In
fact, Roch had use a similar idea a few years earlier [109] in an unpublished work,
but that preprint was not known to the author at the time.
Other examples of Toeplitz-related operator sequences for which the algebraic
methods apply and were more recently focus of attention are the variable coefficient
Toeplitz matrices [92, 146] and random Jacobi operators [87]. Variable coefficient
Toeplitz matrices are of the form
[aij ]ni,j=0 , aij := ai−j (i/n, j/n),
where the ak (x, y) are the Fourier coefficients of a function t → a(x, y, t) with
a : [0, 1] × [0, 1] × T → C. They appear in discretizations of ordinary differential
equations with variable coefficients, and many other situations.

3.2.1. Beyond the C∗ -case. The generalization of the above results to Wiener–
Hopf and Hankel operators in Lp (R) for 1 < p < ∞ turned out to be much
harder than expected. As a first step in 2010 the author, together with Steffen
Roch and Bernd Silbermann, managed to study the applicability of the FSM to
operators belonging to the closed subalgebra of B(Lp (R)) generated by operators
of multiplication by piecewise continuous functions on Ṙ and operators of con-
volution by piecewise continuous Fourier multipliers [119]. This involved the use
of certain intermediate algebras to guarantee inverse-closedness properties. When
trying to extend the results of [119] to algebras with flip and to combine the usual
homomorphisms with the doubling of the dimension, the authors were not able to
overcome the occurring inverse-closedness problem.
In 2012, the author together with Steffen Roch [115] introduced two new
families of homomorphisms, which were suitable to describe Banach operator al-
gebras that include the flip and to tackle the inverse-closedness problem. These
homomorphisms possess, in some sense, a built-in doubling of dimension. Besides
studying their properties, these homomorphisms were used to give a description of
the Calkin algebra generated by convolution, multiplication and the flip operators
which is alternative to the description in [120, Section 5.7] and is interesting in
its own right. The application of these homomorphisms to sequence algebras and,
thus, to study the stability was the subject the next paper [116], published in 2014.
The type of generating function of the convolution operators can also be
extended. The original result involved Toeplitz operators with piecewise contin-
uous symbols and systems of equations with such operators. Alexei Karlovich,
Helena Mascarenhas and the author [75] considered the applicability of the FSM
to operators belonging to the algebra generated by operators of multiplication by
piecewise continuous functions and convolution operators with symbols in the alge-
bra generated by piecewise continuous and slowly oscillating functions. The space
the operators act on was Lp (R), which raised some non-trivial issues regarding
inverse-closedness of the subalgebras that the usual methods to solve those type
of problems use. This latter work built on the previous work of Bastos, Bravo, and
336 P.A. Santos

Yuri Karlovich on the compactness of commutators involving convolutions with

slowly oscillating and piecewise continuous matrix functions [4]. In both of the
above works there was a gap in the proof of the full result on the analysis of the
local algebra related to the point ∞ [76]. That gap was resolved with the help of
Markus Seidel [121, 140].
Lindner [86] considered approximations in the limit cases l1 , L1 , l∞ and L∞
for band-dominated operators. One of the difficulties is that for p = ∞ the FSM
is no longer strongly convergent. Some problems and consequences of this setting
are described in the next section.
We must mention also the book by Didenko and Silbermann [32] where they
use a real (not complex) C ∗ -algebra approach to obtain conditions for the conver-
gence of approximation methods for additive convolution-like operators.
3.2.2. Changing the domain. The multi-dimensional case (which originated the
original paper from Kozak [80]) was further tackled by Böttcher and Silbermann
in [23], extending the previous results to the FSM for Toeplitz operators on l2 × l2
with piecewise continuous symbols defined on the torus. For this they needed to
use a variant of the local principle called bilocalization. Böttcher [6] considered
two-dimensional Wiener–Hopf integral operators in Lp (R2 ), Hagen, Roch and Sil-
bermann [54] explored spline approximation methods for multi-dimensional pseu-
dodifferential operators on Lp (Rn ) (see also next section) and Mascarenhas and
Silbermann [90, 91] studied finite sections of convolution type operators on cones
on Lp (R2 ), with symbols in the Wiener algebra.
In their work on singular integral operators, Gohberg and Krupnik had con-
sidered contours in C besides the real line [47, 48, 51, 52]. The study of the corre-
sponding numerical approximations followed for instance in [33, 55] where special
care had to be taken on corner points, namely in some cases introducing using mul-
tiindiced sequences. The effect of the extra index is to cut-off part of the contour
nearby the corner, simplifying the stability conditions.
3.3. Other approximation methods
While being very important in theoretical terms, the FSM for operators acting in
spaces of non-periodic functions defined on sets like Rn (e.g., Lp (R)) is only used
in practice as a first step in the numerical solution, in order to replace unbounded
domains by compact ones, because the resulting approximations are still infinite-
dimensional operators. Thus immediate and natural applications of the algebraic
methods would be to polynomial and spline approximation. It is a strength of
the new method that it could be applied easily to these different settings. This
happens because they all share an underlying structure (see Section 3.5).
3.3.1. Polynomial approximation. Consider the unit circle T and the set of all
trigonometric polynomials of degree less or equal to n:
 3
n
Πn := un (t) = βk tk , t = eiθ , βk ∈ C . (9)
k=−n
 40 Years of Algebraic Techniques in Numerical Analysis 337

Given a bounded function f defined on the unit circle T, the Lagrange interpolation
operator Ln onto Πn is uniquely defined by the conditions
Ln f (tj ) = f (tj ), tj := e2πij /(2n + 1), j = −n, . . . , n,
and is a projection. For u ∈ L (T), the Galerkin projection of u onto Πn is de-
p

fined as
n ! 2π
1
Pn u(t) := β k tk , βk = u(eis )e−iks ds.
2π 0
k=−n
If S is the Cauchy singular integral operator defined on T,
!
1 u(s)
(Su)(t) := ds, t ∈ T,
πi T s − t
which is a bounded operator on Lp (T) for 1 < p < ∞, one can define the projec-
tions P = (I + S)/2 and Q = (I − S)/2 and consider the equation
(aP + bQ)u = f, f ∈ Lp (T) (10)

nfunctions on T. The solution u of Equation (10)

where a and b are known bounded
can be approximated by un := k=−n βk tk where the βk are the solutions of the
finite system

n −1

a(tj ) tkj βk + b(tj ) tkj βk = fj , j = −n, . . . , −1, 0, 1, . . . , n
k=0 k=−n

where fj can be either f (tj ) if f is bounded, or (Pn f )(tj ). The above is equivalent
to the operator equation
Ln (aP + bQ)Pn un = Ln f (or = Pn f ).
Already in 1982 Junghanns and Silbermann [71] applied the new algebraic
techniques to study the convergence of the above-described collocation method for
one-dimensional singular integral equations on the unit circle. The results were
published in English [72] in 1984 and included the treatment of systems of such
equations with piecewise continuous coefficients which may have a countably in-
finite number of discontinuities. Incidentally, this is one of the first appearances
of the new technique in English language. It is interesting to note that English
started becoming the lingua franca of mathematics only in the last 40 years. Be-
fore, it was common to publish research papers written in the national languages
and, in particular, it was common in the former German Democratic Republic to
publish in either German or Russian.
For singular integral equations on a interval it is possible to use polynomial
approximation based on orthogonal (Chebyshev or other) polynomials. Again the
structure of the underlying algebras is similar to the previous cases. The applica-
bility of the finite section (Galerkin) method for the Cauchy SIO with piecewise
continuous coefficients was obtained in [70] and that of the collocation method was
obtained in [68, 69, 73, 74]. Operators involving Mellin components were treated
338 P.A. Santos

in [67]. More recent work on this subject appeared in [64, 65]. In all cases there
were considered Hilbert function spaces.
The non-Hilbert case appears with delay over the Hilbert case. For instance,
in [66] Junghanns et al presented necessary and sufficient conditions for the sta-
bility of collocation methods applied to Cauchy singular integral equations on an
interval for weighted Lp spaces. They also studied the behavior of the approxima-
tion numbers and the k-splitting property of the respective sequence of discretized
operators (see Section 3.4.2). The main impact of the change to the non C∗ -case,
besides the added technical difficulty, was the appearance of additional conditions
in terms of a family of limit operators indexed on the points of the interval, in a
similar way to what had already be observed by Böttcher years earlier [7].

3.3.2. Spline approximation algebraic methods. We will illustrate the matter with
a simple example of splines for functions defined on the real line, but the principle
is the same for other domains. Consider functions on Lp (R+ ). Let (+n )n∈N denote
(n)
a sequence of meshes (or partitions) in R+ , +n = {xk : 0 ≤ k ≤ n} such that
(n) (n) (n)
0 = x0 ≤ x1 ≤ · · · ≤ xn = ∞. To the sequence of meshes (+n )n∈N we
associate a sequence of piecewise polynomial spline spaces
Sn = {u ∈ L2 (R+ ) : u| (n) (n)
∈ Pd , u | (n)
= 0},
]x ,x [ ]x ,∞[
k−1 k n−1

where k runs from 1 to n − 1 and Pd represents the set of polynomials of degree

less than or equal to d (d fixed). The above definition is for the so-called max-
imum defect splines, that is, there are no restrictions on the smoothness of the
spline functions on the mesh points. It is of course possible to define splines with
continuity or other restrictions. On the other end of the scale are the smoothest
splines, where the function u must be continuously differentiable a number of
times, depending on d (u ∈ C d−1 ).
Let P n be a projection from L2 (R+ ) onto S n , which is a closed subspace of
+
L2 (R ).
It is now possible to consider the approximation method
P n AP n un = P n v, un ∈ Im (P n ) (11)
for approximating the equation
Au = v, u, v ∈ L2 (R+ ).
In general, one can choose the projection P n to be the orthogonal projection,
which in this case corresponds to the Galerkin projection, a collocation projec-
tion (in a subspace of continuous functions), or any other projections associated
with various numerical approximation methods. As long as the projection satisfies
certain conditions on the convergence, it is possible to build the algebras of ap-
proximation sequences and apply algebraic techniques to obtain conditions for the
stability of those sequences. It is also possible to use (infinite) uniform meshes, or
substitute the maximum defect spline spaces for smoother splines.
 40 Years of Algebraic Techniques in Numerical Analysis 339

Algebraic techniques in the case of spline approximation appeared first with-

out algebras. Inspired by the results obtained by Silbermann in the algebraic set-
ting described above and its own earlier work [98] that involved implicitly a discrete
commutator property, Prössdorf [93, 94] in 1984 defined a localization principle
applicable to spline approximation sequences directly, by proving the commuta-
tor properties, without needing an algebraic structure. But indeed, the algebraic
structure is implicit. This variant of the algebraic method, including the intro-
duction of associated operators, corresponding to the identification step of the
algebraic method, continued to be used by Prössdorf and some co-workers (e.g.,
[30, 95–97, 99]).
Silbermann and his student Roland Hagen first managed to obtain results
using algebraic techniques for the finite element collocation method applied to
singular integral operators on the unit circle [57] and on the torus [58]. In [59] the
analysis was extended to the qualocation method, which is similar to the Galerkin
method, but with the exact inner products in the Galerkin method replaced by
certain quadrature rules. This method is a compromise between the Galerkin and
collocation methods, aiming to achieve the benefits of the first at a cost comparable
to the second method. In [60], Hagen and Silbermann make a bridge with the work
of Prössdorf and Rathsfeld [97] by reframing the problem using a C∗ -algebra and
showing that it falls on the template for the algebraic techniques.
Overcoming the singular behavior of spline approximation methods for one-
dimensional singular integral operators caused by discontinuities in the coefficients
or by non-smooth geometries of the underlying curves was the object of [110] where
Steffen Roch showed the singularities can be cut off by a modification of the ap-
proximation method which is closely related to the finite or infinite section method
for discrete Toeplitz operators. Also in that year Hagen, Roch and Silbermann [54]
tried to obtain a general formulation of stability problems for spline approximation
methods in Lp (Rn ). Those results were generalized and extendend to a wide set of
approximation methods with regard to singular integral operators and some types
of Mellin operators in the book “Spectral theory of approximation methods for
convolution equations” [55] by the same authors published in 1995.
Regarding Wiener–Hopf operators, Roch [111] treated the infinite Galerkin
method (i.e., the Galerkin method in which the spline space has infinite dimension)
with smoothest splines for Wiener–Hopf operators with piecewise continuous sym-
bols. In the case of a uniform mesh and maximum defect finite-dimension spline
approximations for Wiener–Hopf plus Hankel operators on the real line, the stabil-
ity problem was solved in [133] by the author. Several new ideas were introduced:
an infinite spline projection and a finite section projection were added as indepen-
dent members of an algebra that also included the operators being approximated
as constant elements. The combination of the two projections generated the finite-
dimensional spline projection. This allowed the use of both the techniques for
infinite Galerkin method from [55] and the new essentialization homomorphisms
developed ealier [118]. A main problem that was solved was related to some of
the local algebras, which were generated by four elements instead of the usual two
340 P.A. Santos

or three. The solution was to carefully analyse the relations between the gener-
ators and extend the knowledge about these types of algebras. Finally in [135],
Silbermann and the author tackled non-uniform meshes in L2 (R+ ).
On another line of research, Didenko and Silbermann studied approximation
methods for the Muskhelishvili equation on a simple closed curve Γ with a finite
number of corner points. Approximate solutions of the equation were constructed
as piecewise constant functions by the Galerkin and quadrature methods [34]. Since
the curve is not smooth, the integral operators appearing in this equation are non-
compact, which generates additional difficulties with the Fredholm property of the
corresponding operator and the stability of the approximation methods.
In [134], the author studied the applicability of maximum defect polynomial
(Galerkin) spline approximation methods with graded meshes to Wiener–Hopf op-
erators with matrix-valued piecewise continuous generating function defined on Ṙ.
The continuous symbol case for Lp was already studied in [42] and [100, Chapter
5], using factorization techniques. This work extended the L2 piecewise continuous
case in [135]. While the Banach algebra techniques were in some sense more clas-
sical, this work involved calculating norm estimates related to the approximations
used, in order to be able to assert that the commutators indeed belong to the
correct ideals.
3.3.3. Non-strongly convergent approximation sequences. When the sequence of
projections Pτ is strongly convergent, most proofs involving the characterization
of sequences as belonging to the ideal J make use of the close relationship between
strong convergence and the ideal of compact operators. In this context, the ob-
servation that the multiplication of a strongly convergent sequence by a compact
operator gives uniform convergence plays a fundamental role. For non-strongly
convergent projections, as it is the case for the finite section projections in lZ∞ or
L∞ (R), S. Roch and B. Silbermann [126] (see also [100, 4.36 et seq.]) initiated the
idea of changing the above-mentioned connection between compactness and strong
convergence to a definition. Starting with the projection sequence P := (Pn )n∈Z+ ,
they substituted the usual compact operators by compact-like operators related to
that sequence. The P-compact operators will be those operators K ∈ B(X) such
that
KPτ − K and Pτ K − K tend to zero as τ → ∞.
This new approach with the adapted ideal of P-compact operators happens
to be useful also in the cases 1 < p < ∞, and is at the heart of recent results
[88, 89, 141].
In [88] the author, together with Helena Mascarenhas and Markus Seidel
managed to use the new approach to study, together with the stability of oper-
ator sequences, a more general Fredholm theory in sequence algebras, including
the splitting property (see Section 3.4 below) and an index formula in the case
of P-strongly converging finite section methods. This setting covered, in partic-
ular, convolution type operators in Lp (R), with p ∈ [1, ∞]. A new class Q of
quasi-banded operators was introduced, extending the previously known notion of
 40 Years of Algebraic Techniques in Numerical Analysis 341

band-dominated operators. It includes all convolutions with generating function

continuous at all finite points of R, but does not include the general piecewise con-
tinuous case. Results on its inverse and Fredholm closedness, and results on the
stability of the FSM for such operators were obtained. This new context allowed
for the treatment of the FSM for convolution type operators with symbols in a
wide variety of function spaces. It was also possible to combine the convolution
operators with multiplication operators by functions of again a variety of spaces
and to obtain relatively simple conditions for stability. The results cover the cases
p ∈ {1, ∞}, in which all Fourier convolution operators are contained in Q. By
using a special transformation, the flip J could also be included. An interesting
observation is that the general results obtained by this method do not need the use
of local principles. Without local principles, it was possible to cover convolution
operators whose symbol has a jump at infinity.
The techniques pioneered in [88] were added to and reworked into a new
approach in [89], which the authors named “Rich Approach”. It completed the
picture by describing also non-stable sequences in terms of the asymptotic behav-
ior of their approximation numbers and Fredholm indices. It also covered much
larger classes of data, including almost periodic, even bounded uniformly continu-
ous multipliers b, piecewise continuous as well as operators of multiplication with
almost arbitrary functions a having restrictions only on the behavior at infinity.
The new approach involved the introduction of the concept of rich sequences,
where the existence of strong (or other) limits for the whole sequence is not neces-
sary, as long as there are subsequences with certain properties. Using rich sequences
with the concept of P-convergence made constructions and proofs much simpler.
In the previous works, when considering certain function spaces as generators for
the convolution and multiplication operators, the localization had to be done over
involved commutative subalgebras of sequences arising from functions whose defi-
nition was technical and the characterization of their maximal ideal spaces required
hard proofs. This was now replaced by the much simpler algebra of the continuous
functions. The approach used allowed to build upon much simpler (and probably
even more general) definitions of the multiplier algebras.
An interesting point is that the role of the homomorphisms changed. While in
the classic method the W homomorphisms served for the construction a sufficiently
large ideal J such that the respective quotient has a large center, and other family
H of homomorphisms provided the desired representatives for the identification
of the local cosets, in this work only the W homomorphisms were used in the
identification of the local cosets, whereas the H homomorphisms were used to
build the ideal corresponding to J .

3.4. Asymptotics, behavior of eigenvalues, approximation numbers

and numerical ranges
Asymptotics of Toeplitz matrices were one of the main drivers for the development
of the algebraic techniques described in this work. For instance, already in 1958,
Widom was interested in the distribution of the eigenvalues for Hermitian Toeplitz
342 P.A. Santos

matrices [154] and in 1974, his aim to calculate asymptotic determinants was the
approximation of the inverses of (large) finite matrices by the known inverse of their
infinite counterparts [156] following on the pioneer works of Szegő, Hartwig and
Fisher, and others [63, 150, 151]. The above-mentioned paper [157] that presented
for the first time formula (5) had the aim of extending some of the earlier results
to Block Toeplitz matrices. The topic of Toeplitz determinants, which inspired
the discovery of the algebraic techniques in numerical analysis, was touched also
by those techniques. Already in 1985, Böttcher and Silbermann [24] proved the
Fisher–Hartwig conjecture on Toeplitz determinants for a large class of symbols
and after the fall of the Berlin wall, normal academic contacts between Widom in
the USA and Böttcher and Silbermann in (now) Germany developed [27] further.
See also Böttcher’s nice expository paper [9].
Torsten Ehrhardt, a student from Silbermann, continued the work on the
subject of Toeplitz Determinants with Fisher–Hartwig Singularities, publishing a
series of works starting in 1997 ([3, 41]). In 2003 he gave a proof of the Szegő–
Widom limit theorem using an algebraic approach [38].
The author also made a modest contribution to this topic by proving, together
with A. Karlovich, a generalization of the Szegő strong limit theorem for Toeplitz
determinants with the symbol having a nonstandard smoothness [78]. There it
was assumed that the symbol belonged to the Wiener algebra and, moreover,
the sequences of Fourier coefficients of the symbol with negative and nonnegative
indices belonged to certain weighted Orlicz classes.
A book that contains a good overview of the issues and research done on
Asymptotics and spectral properties of banded Toeplitz matrices up to the early
2000s is [17]. As was mentioned earlier, banded matrices are matrices where the
non-zero entries are situated on a band around the main diagonal.
Work has continued on the subject, as is shown by [130], where the authors
present a unified C∗ -algebra approach to classical Szegő limit theorems that applies
to general (including non-selfadjoint) Toeplitz operators and in particular obtain
results for operators with almost periodic diagonals. Other example is a recent
generalization of the Szegő’s strong limit theorem for certain classes of band-
dominated operators [39] or perturbed operators [15].

3.4.1. Spectral approximation. The basic question in Spectral Approximation is

whether the spectra of the operators in a convergent approximating sequence
(Aτ )τ ∈I do approximate the spectrum of the limit operator A, σ(A), in some
sense. For that, one can define the partial limiting set of a sequence of subsets
(Uτ ) ⊂ C as the set of all partial limits of all sequences (λτ ) of λτ ∈ Uτ , τ ∈ I and
denote it by p-limτ →∞ Uτ . Then the question above can be more precisely stated
as whether
p-lim σ(Aτ ) = σ(A).
τ →∞

In the case of Hermitian Toeplitz operators acting in l2 or L2 the answer is

affirmative. In general it is not. A simple counter-example is given for instance by
 40 Years of Algebraic Techniques in Numerical Analysis 343

the unilateral forward shift V on l2 , defined by (V u)0 = 0, (V u)j = uj−1 , j ∈ N.

The spectrum of V is the closed unit disk, but approximations An = Pn APn are
represented as the n + 1 × n + 1 finite matrices
⎡ ⎤
0 0 0 ... 0
⎢ .. ⎥
⎢1 0 0 . 0⎥
⎢ ⎥
⎢ .. .. ⎥
An = ⎢0 1 0 . .⎥
⎢ ⎥
⎢. . ⎥
⎣ .. . . . . . . . . 0⎦
0 0 ... 1 0

with spectrum {0}. In fact, since the 1960 work of Schmidt and Spitzer [139] it is
known that in the case of a trigonometric polynomial a, while the spectrum of the
associated Laurent operators corresponds with the range of a, the limiting set of
the spectra of the finite sections of that operator is a union of analytic arcs that
are “inside” the closed curve a(T) ⊂ C. But while we have a negative answer for
non-Hermitian operators to the spectral approximation question when considering
the spectrum, the picture changes when one considers the -pseudospectrum of
the operators. This notion was first introduced by Landau [85] in 1975. The -
pseudospectrum is defined as the set

σ (A) := {λ ∈ C : (A − λI)−1  ≥ 1/}.

In 1992 Reichel and Trefethen [105] showed that for some classes of non-real sym-
bols, the pseudospectra of the corresponding discrete convolution operators would
mimic exactly the pseudospectra of the limit operator. Inspired by those results
Böttcher in 1994 used algebraic techniques to extend this result to convolution
operators on L2 (R) for a much wider class of symbols [8]. His results also carried
over to the discrete case. Later, Böttcher, Grudsky and Silbermann considered the
case 1 < p < ∞ for continuous (including matrix) symbols [18]. Roch [112] gener-
alised the results for almost periodic symbols and included Numerical Ranges. In
[56, Chapter 3] Hagen, Roch and Silbermann went on to prove that for the C∗ -case,
fractality of the approximating algebra (see Section 3.5) guarantees convergence
of the pseudospectra and also of the numerical ranges.

3.4.2. Approximation numbers. Given an operator A acting on the Banach space

X, the lower approximation numbers from the right and from the left,

srk (A) := inf{A − T  : T ∈ B(X), dim Ker T ≥ k},

slk (A) := inf{A − T  : T ∈ B(X), dim Coker T ≥ k}.

measure the distance to (or degree of) injectivity and surjectivity, respectively. It
is clear that
r/l r/l r/l
0 = s0 (A) ≤ s1 (A) ≤ s2 (A) ≤ · · ·
344 P.A. Santos

r/l
for both right and left approximation numbers and if dim X = n then sn (A) =
A. It is also not difficult to show (see, for instance, [10]), that

l r 1/A−1  if A is invertible
s1 (A) = s1 (A) = (12)
0 if A is not invertible.
Also notice that in case of X being a Hilbert space, these approximation numbers
coincide with the lower singular values of A or A∗ , respectively, that is, the square
roots of the eigenvalues of A∗ A and AA∗ . In the early 1990s, Silbermann and
Roch started [128, 144] to use algebraic techniques to extend Widom’s earlier
results [158] on the limiting set of singular values. A little later, by using the
C∗ -algebra concept of Moore–Penrose invertibility in the approximation sequence
algebra, they were able to derive a certain behavior of the singular values of the
approximating matrices. They defined (An ) to be a Moore–Penrose sequence if the
coset (An )+G would be Moore–Penrose invertible as an element of the algebra E/G
and showed that in this case the set of the singular values of An would split into two
parts. One of these parts remains bounded away from zero by a positive constant
l/r
for all n, and the other part tends to zero as n → ∞. Otherwise, sk (Tn (a)) → 0
as n → 0 for each fixed k. The results were quite general, showing the power of
C∗ -algebra techniques. When applied to the finite sections of a Toeplitz operator,
the number k of singular values going to zero was given by
k = dim Ker T (a) + dim Ker T (ã). (13)
The results were written in preprint form in 1996, but were only published in 1998
[129].
Böttcher tried to extend these results to the Banach algebra case and in 1997
managed to prove the scalar version of the splitting property for some Toeplitz
matrices with piecewise continuous symbol in lp [10]. He settled the cases where
either the limit operator T (a) would be not normally solvable in lp or lq , with
1/p + 1/q = 1, or the operator would be Fredholm in the whole scale of spaces lr
for r ∈ [p, q]. In those cases, the result is the same as for p = 2. The third case, when
T (a) is normally solvable in lp and lq but not on some lr with r ∈ [p, q], was only
solved in 2006 by Rogozhin and Silbermann [131, 132]. They were able to consider
the splitting property for all elements in the algebra of approximating sequences.
In the third case above, the k-splitting property still occurs, but the number k
must consider also the kernels of the non-invertible limit operators associated the
local algebras at each point t ∈ T.
Already in [129] the concept of a Fredholm sequence was introduced as a
analogy to the notion of Fredholm operator. Broadly, Fredholm sequences are
those sequences in F that are invertible modulo the ideal J (see Section 2.2). For
Fredholm sequences the third case above cannot occur as all local representatives
are invertible. When studying concrete approximation methods, it is possible to
rely on this framework to extract the splitting property for Fredholm sequences.
Silbermann and Seidel [141] revisited this whole framework in the context
of non-strongly convergent approximating sequences. In [88] the author, together
 40 Years of Algebraic Techniques in Numerical Analysis 345

with H. Mascarenhas and M. Seidel, building on those results, formulated and

proved results on the splitting property for the large class of rich sequences. There,
to a sequence A we associate numbers α(A) and β(A) as the splitting number k
corresponding to the right (resp. left) approximation numbers of the sequence A
and were able to relate those numbers to the kernels and cokernels of the limit
operators of the sequence.

3.5. Structure of the approximation algebras

Algebraic methods being applicable in such a wide range of spaces, operators and
numerical methods, as we have seen in the sections above, hints at the existence
of an underlying common structure in the occurring algebras. It is clear that the
approximating algebras are related to the original Toeplitz operator algebra, which
can be thought of as the algebra generated by the unilateral shift operator V .
In the case of C∗ -algebras, Coburn showed in the sixties [29] that the C∗ -
algebra generated by any non-unitary isometry is ∗ -isomorphic to the C∗ -algebra
alg {V }, where V : l2 → l2 is the unilateral shift of multiplicity one. That algebra
is exactly the algebra T (C) := alg {T (c) : c ∈ C(T)} with T (c) : l2 → l2 . This
algebra includes the ideal K of all compact operators and it was also known that
the Calkin algebra T (C)/K is isometrically isomorphic to the C∗ -algebra of the
continuous functions C(T).
With the above framework in mind, Böttcher and Silbermann in [23], when
considering the convergence of the finite sections method for quarter plane Toeplitz
operators, described for the first time the structure of the algebras of approximat-
ing sequences. They considered the algebra
A(T (C)) := alg {(Pn T (c)Pn )n∈Z+ , c ∈ C(T)}
containing all discretizations of Toeplitz operators with continuous generating
function and obtained the following result:
Theorem 3.1. The algebra A(T (C)) consists of all sequences of the form
(An ) = (Pn T (c)Pn + Pn K0 Pn + Wn K1 Wn + Gn ) (14)
with K0,1 compact operators and (Gn ) ∈ G and this representation is unique for
each element. Moreover, the strong limits (6) exist for the elements of A(T (C))
and for each sequence A ∈ A(T (C)) the coset A + G is invertible in A/G if and
only if the strong limit operators A = T (c) + K0 and à = T (c̃) + K1 are invertible.
The above result implies that the quotient algebra A(T (C)) is ∗-isomorphic
to the C∗ -algebra of all pairs
(T (c) + K0 , T (c̃) + K1 ) ∈ B(l2 ) × B(l2 ).
with c continuous, and K0,1 compact operators.
As in the case of Silbermann’s original paper [142], the set
 
J = Pn K0 Pn + Wn K1 Wn + Gn , K0,1 compact, Gn ∈ G
346 P.A. Santos

is a closed two-sided ideal of A(T (C)). Write W0 and W1 for the ∗-homomorphisms
corresponding to the strong limits in (6), such that
W0 (An ) := s-lim An and W1 (An ) := s-lim Wn An Wn .
n→∞ n→∞

It can be easily checked that they are in fact ∗-homomorphisms. Then F ⊂ E is

the sequence algebra where those homomorphisms are well defined. Introduce then
the sets in F
K0 = {(Pn KPn ) + G, K ∈ K}
K1 = {(Wn KWn ) + G, K ∈ K}.

They can be checked to be ideals of F /G, and the following structure emerges:
1. G ∈ Ker W0,1 ;
2. the quotient homomorphisms F /G → B(l2 ), A + G → W0,1 (A) are ∗-iso-
morphisms between K0,1 and K;
3. K0 ∈ Ker W1 and K1 ∈ Ker W0 ;
4. an element A ∈ A is invertible in F if and only if the operators W0,1 (A) are
invertible.
A most interesting feature is that a similar structure (with a finite or infinite
number of homomorphisms) was found in many sequence algebras, even if the
operators are far from Toeplitz operators and the methods are different from the
Finite Section Method. That it must be so was only recently proved by Steffen
Roch [106] for separable C∗ -subalgebras of F .
A further important observation derived from these earlier works is the notion
of fractality applied to sequence algebras. Because the structure of the C∗ -algebra
A/G is given by the strong limit homomorphisms (that are in fact irreducible
representations), all essential information is stored in each subsequence of the
original sequence. Roch et al [103] attempted to describe rigorously the method of
limit operators and to collect and unify known information about their applications
in the study of Fredholm properties of band and band-dominated operators.
In [117], Roch and the author proved that all compact operators acting on
Lp (R) belong to the algebra generated by the operator of multiplication by the
characteristic function of the positive half-axis and by the convolution operators
with continuous generating function. This result, together with classical results,
were used to prove that certain ideals of compact-like operator sequences (ideals
of the type of K0 and K1 above), which are building blocks of the ideal J , are
included in the algebra generated by convolution and multiplication operators and
the finite section projection sequence, in the case of Banach algebras.
In [36] the author, together with some students, showed in fact that the
Banach algebra case also possessed a structure similar to the one referred above in
this section, but considered a concrete example (see Figure 2). A key observation
in that work was that it is possible to build the ideals Kj as the inverse of algebra
automorphisms applied to the set of constant sequences K ⊂ F containing the
 40 Years of Algebraic Techniques in Numerical Analysis 347

compact operators. The homomorphisms are built by applying the strong limit to
the sequences resulting from the automorphisms.

Figure 2. The algebraic approach to stability problems, as proposed in [36]

Here the usual role of the strong limit homomorphisms changes. While before
the W homomorphisms served for the construction a sufficiently large ideal J such
that the respective quotient F /J has a large center, and the H homomorphisms
provided the desired representatives for the identification of the local cosets, now
the H homomorphisms are also used to enlarge the ideal J , in such a way that the
local algebras are now almost trivial, and invertibiity of the local cosets is implied
by the invertibility of the essentialization homomorphisms. . .
There were many other attempts to understand and explore the structure of
approximation algebras. Bernd Silbermann and the author [136, 137] generalized
Coburn’s theorem to approximating algebras where the operators being approxi-
mated are functions of a shift and the approximation method possess certain simple
properties. It was possible in this abstract setting to recover many concrete results
on applicability of concrete numerical methods. Roch [114] recently observed that
(a minor modification of) the FSM-algebra A(T (C)) is also distinguished by a
universal property: it is ∗ -isomorphic to the universal C∗ -algebra generated by a
power partial isometry, that is, an element v of a C∗ -algebra with the property
that every power v n is a partial isometry.
The theory for algebras generated by two idempotents, that originated inde-
pendently with many authors (see [28, Remark 1.3]) was object of intense attention
because, as was referred above, those structures appear when trying to identify
local algebras related to singular integral operators. Important contributions in-
cluded the generalization of Halmo’s two projections theorem [61] to the case of
two idempotents in Banach algebras, due to Roch and Silbermann [124], later com-
pleted by Gohberg and Krupnik [50, 53]. Generalization of the results for elements
with massive spectra followed in [44] and the case of two idempotents and one
348 P.A. Santos

flip in [45, 84] (there are several types of shifts, the original paper from Roch and
Silbermann already considered one type). Algebras possessing more idempotents
were also studied [16, 83]. Two gentle introductions to the subject can be found
on [28] and [120, Chapter 3].

4. Some open problems and future research

In the early 1990s, a young researcher was refused a position, with a remark that
research into linear and one-dimensional singular integral operators was no longer
current [13]. That remark proved completely wrong, as the next 25 years saw a
multitude of interesting and unexpected results for one-dimensional convolution
and related operators, whose intersection with this paper’s topic is not empty.
Nevertheless, the author would risk affirming that for one-dimensional convolution
operators in Lp spaces a lot is already known. A natural future line of research
would be to move beyond those types of spaces. Interesting candidates can be vari-
able exponent Lp type spaces, which are not rearrangement invariant. In the last
few years advances in the study of algebras of operators in such spaces, commuta-
tor properties, etc (see, e.g., [77, 79]) opened the way for treating approximation
methods in those spaces.
Another alternative are the super-Lebesgue spaces [40]. If we let 1 ≤ p < q <
∞, X be a suitable domain, one can define the super-Lebesgue space Lpq (X) by
Lpq (X) := Lp (X) + Lq (X), (15)
that is, the elements of Lpq (X) are those functions4 u that can be written as a
sum u = up + uq , with up ∈ Lp (X) and uq ∈ Lq (X).
Super-Lebesgue spaces include the usual Lp spaces and have interesting prop-
erties. Torsten Ehrhardt and the author are working on a characterization of the
Singular Integral Operator on those spaces, in such a way as to allow the further
study of Numerical Methods for convolution operators in those types of spaces.
Numerical methods applications to real world problems need multi-dimen-
sional domains with complex geometries. While tackling those complex geometries
directly with Banach algebra techniques can be a difficult challenge, the study of
multi-dimensional canonical problems, that “locally” can approximate those real
world problem geometries will constitute a first step.
Not much is known yet about sequences of variable-coefficient Toeplitz ma-
trices when the generating function is outside the Wiener algebra, or one considers
lp norms. As these matrices appear in many situations, this is also a promising
area of future research.
Finally, the study of Banach algebras directly related to or just inspired by
the applications to Numerical Analysis continues, with many recent papers on the
subject, as was seen above. It would be interesting to be able to understand if the

4 As is normal, functions which differ only on a set of measure 0 are identified.

 40 Years of Algebraic Techniques in Numerical Analysis 349

different characterizations obtained so far can lead to a deeper understanding of

these algebras’s structure.

5. Conclusions
In this paper we tried to present an brief overview of the main lines of research
connected with Banach algebra techniques in numerical analysis, which have been
developed for more than 40 years. We tried to present, more than a dry account
and enumeration of results, a lively account of the development of some of the
key ideas behind those results, emphasising the cross-pollination of ideas between
mathematicians across the globe.
The author’s idea was to preserve some of the small details that are known
by those who lived through what we described, but would be lost forever when
only the published papers and results remained, of all those lifes dedicated to
mathematics.
“Brief” here is the keyword. We set up with the objective of doing a 20 page
paper with around 120 references, and have already overstepped those objectives.
Nevertheless, much remains to be said, and the referred works are only a fraction
of all that could be cited. We hope to have included all of the most relevant ones.
If that is not the case, it is solely the author’s fault.
Acknowledgement
The author would like to thank Albrecht Böttcher, Steffen Roch and Bernd Sil-
bermann for many details and clarifications on the history of the subject during
the 1980s, and for reading and commenting earlier versions of this document. This
work has been partially supported by national funds through Fundação para a
Ciêcia e a Tecnologia (FCT) with reference UID/CEC /50021/2013.

References
[1] G. Allan. Ideals of vector-valued functions. Proc. London Math. Soc., 18(2):193–216,
1968.
[2] E. Basor. A brief history of the strong Szegő limit theorem. In Mathematical methods
in systems, optimization, and control, volume 222 of Oper. Theory Adv. Appl., pages
73–83. Birkhäuser/Springer Basel AG, Basel, 2012.
[3] E. Basor and T. Ehrhardt. Asymptotic formulas for determinants of a sum of finite
Toeplitz and Hankel matrices. Math. Nachr., 228:5–45, 2001.
[4] M. Bastos, P. Bravo, and Y.I. Karlovich. Convolution type operators with symbols
generated by slowly oscillating and piecewise continuous matrix functions. volume
147 of Oper. Theory Adv. Appl., pages 151–174. Birkhäuser, 2004.
[5] G. Baxter. A norm inequality for a “finite-section” Wiener–Hopf equation. Illinois
J. Math., 7:97–103, 1963.
[6] A. Böttcher. The finite section method for two-dimensional Wiener–Hopf integral
operators in Lp with piecewise continuous symbols. Math. Nachr., 116:61–73, 1984.
350 P.A. Santos

[7] A. Böttcher. The reduction method for Wiener–Hopf integral operators with
piecewise-continuous symbol in Lp spaces. Funktsional. Anal. i Prilozhen., 18(2):55–
56, 1984.
[8] A. Böttcher. Pseudospectra and singular values of large convolution operators. J.
Integral Equations Applications, 6(3):267–301, 1994.
[9] A. Böttcher. The Onsager formula, the Fisher–Hartwig conjecture, and their in-
fluence on research into Toeplitz operators. J. Statist. Physics (Lars Onsager
Festschrift), 78:575–585, 1995.
[10] A. Böttcher. On the approximation numbers of large Toeplitz matrices. Documenta
Mathematica, 2:1–29, 1997.
[11] A. Böttcher. C∗ -Algebras in numerical analysis. Irish Math. Soc. Bulletin, 45:57–
133, 2000.
[12] A. Böttcher. Essay on Bernd Silbermann. In Toeplitz matrices and singular integral
equations (Pobershau, 2001), volume 135 of Oper. Theory Adv. Appl., pages 1–12.
Birkhäuser, Basel, 2002.
[13] A. Böttcher. Linear and one-dimensional. In Israel Gohberg and Friends, pages
291–293. Birkhäuser, Basel, 2008.
[14] A. Böttcher, 2016. Private Communication.
[15] A. Böttcher, L. Fukshansky, S.R. Garcia, and H. Maharaj. Toeplitz determinants
with perturbations in the corners. J. Funct. Anal., 268(1):171–193, 2015.
[16] A. Böttcher, I. Gohberg, Y. Karlovich, N. Krupnik, S. Roch, B. Silbermann, and
I. Spitkovsky. Banach algebras generated by N idempotents and applications. In
Singular integral operators and related topics, volume 90 of Oper. Theory Adv. Appl.,
pages 19–54. Birkhäuser, Basel, 1996.
[17] A. Böttcher and S.M. Grudsky. Spectral Properties of Banded Toeplitz Matrices.
SIAM, Philadelphia, 2005.
[18] A. Böttcher, S.M. Grudsky, and B. Silbermann. Norms of inverses, spectra, and
pseudospectra of large truncated Wiener–Hopf operators and Toeplitz matrices.
The New York Journal of Mathematics, 3:1–31, 1997.
[19] A. Böttcher, N. Krupnik, and B. Silbermann. A general look at local principles
with special emphasis on the norm computation aspect. Integral Equations Operator
Theory, 11:455–479, 1988.
[20] A. Böttcher, S. Roch, and B. Silbermann. Local constructions and Banach algebras
associated with Toeplitz operators on H p . In Seminar Analysis, pages 23–30. Akad.
Wiss. DDR, Berlin, 1986.
[21] A. Böttcher, S. Roch, B. Silbermann, and I. Spitkovsky. A Gohberg–Krupnik–
Sarason symbol calculus for algebras of Toeplitz, Hankel, Cauchy, and Carleman
operators. In Topics in operator theory: Ernst D. Hellinger memorial volume, vol-
ume 48 of Oper. Theory Adv. Appl., pages 189–234. Birkhäuser, Basel, 1990.
[22] A. Böttcher and B. Silbermann. Über das Reduktionsverfahren für diskrete Wiener–
Hopf-Gleichungen mit unstetigem Symbol. Z. Analysis Anwendungen, 1:1–5, 1982.
[23] A. Böttcher and B. Silbermann. The finite section method for Toeplitz operators on
the quarter-plane with piecewise continuous symbols. Math. Nachr., 110:279–291,
1983.
 40 Years of Algebraic Techniques in Numerical Analysis 351

[24] A. Böttcher and B. Silbermann. Toeplitz matrices and determinants with Fisher–
Hartwig symbols. J. Funct. Anal., 63(2):178 – 214, 1985.
[25] A. Böttcher and B. Silbermann. Introduction to Large Truncated Toeplitz Matrices.
Universitext. Springer-Verlag, New York, 1999.
[26] A. Böttcher and B. Silbermann. Analysis of Toeplitz Operators. Springer-Verlag,
Berlin, Heidelberg, second edition, 2006.
[27] A. Böttcher, B. Silbermann, and H. Widom. A continuous analogue of the Fisher–
Hartwig formula for piecewise continuous symbols. J. Funct. Anal., 122(1):222–246,
1994.
[28] A. Böttcher and I. Spitkovsky. A gentle guide to the basics of two projections theory.
Lin. Alg. Appl., 432(6):1412–1459, 2010.
[29] L.A. Coburn. The C ∗ -algebra generated by an isometry. Bull. Amer. Math. Soc.,
73:722–726, 1967.
[30] W. Dahmen, S. Prössdorf, and R. Schneider. Wavelet approximation methods for
pseudodifferential equations i: Stability and convergence. Math. Zeit., 215(1):583–
620, 1994.
[31] A. Devinatz and M. Shinbrot. General Wiener–Hopf operators. Trans. Amer. Math.
Soc., 145:467–494, 1969.
[32] V. Didenko and B. Silbermann. Approximation of additive convolution-like opera-
tors: real C ∗ -algebra approach. Birkhäuser, Basel, Boston, Berlin, 2008.
[33] V.D. Didenko, S. Roch, and B. Silbermann. Approximation methods for singular
integral equations with conjugation on curves with corners. SIAM J. Numer. Anal.,
32(6):1910–1939, 1995.
[34] V.D. Didenko and B. Silbermann. On stability of approximation methods for the
Muskhelishvili equation. J. Comput. Appl. Math., 146(2):419–441, 2002.
[35] R.G. Douglas. Local Toeplitz operators. Proc. London Math. Soc. (3), 36(2):243–
272, 1978.
[36] D. Dragičević, G. Preto, P. Santos, and M. Szamotulski. On a question by Markus
Seidel. In Operator Theory, Operator Algebras and Applications, volume 242 of
Oper. Theory: Adv. Appl., pages 159–172. Birkhäuser, Basel, 2014.
[37] R. Duduchava. Integral equations with fixed singularities. B.G. Teubner Verlags-
gesellschaft, Leipzig, 1979.
[38] T. Ehrhardt. A new algebraic approach to the Szegő–Widom limit theorem. Acta
Math. Hungar., 99(3):233–261, 2003.
[39] T. Ehrhardt, S. Roch, and B. Silbermann. A strong Szegő–Widom limit theorem
for operators with almost periodic diagonal. J. Funct. Anal., 260(1):30–75, 2011.
[40] T. Ehrhardt and P.A. Santos. On the singular integral operator in super-Lebesgue
spaces. In preparation.
[41] T. Ehrhardt and B. Silbermann. Toeplitz determinants with one Fisher–Hartwig
singularity. J. Funct. Anal., 148(1):229 – 256, 1997.
[42] J. Elschner. On spline approximation for a class of non-compact integral equations.
Math. Nachr., 146:271–321, 1990.
[43] J. Favard. Sur les équations différentielles a coefficients presque périodiques. Acta
Math., 51:31–81, 1927.
352 P.A. Santos

[44] T. Finck, S. Roch, and B. Silbermann. Two projection theorems and symbol cal-
culus for operators with massive local spectra. Math. Nachr., 162:167–185, 1993.
[45] T. Finck, S. Roch, and B. Silbermann. Banach algebras generated by two idempo-
tents and one flip. Math. Nachr., 216:73–94, 2000.
[46] I. Gohberg and I. Feldman. Convolution Equations and Projection Methods for
their solution. Amer. Math. Soc., Providence, RI., 1974. First published in Russian,
Nauka, Moscow, 1971.
[47] I. Gohberg and N. Krupnik. The algebra generated by one-dimensional singular in-
tegral operators with piecewise continuous coefficients. Funkcional. Anal. i Priložen,
4(3):26–36, 1970 (in Russian).
[48] I. Gohberg and N. Krupnik. On singular integral operators on a composed curve.
Soobshch. Akad. Nauk. Gruz. SSR, 64(3):21–24, 1971 (in Russian).
[49] I. Gohberg and N. Krupnik. On a local principle and algebras generated by Toeplitz
matrices. An. Şti. Univ. “Al. I. Cuza” Iaşi Secţ. I a Mat. (N.S.), 19:43–71, 1973.
(in Russian; English translation in “Convolution equations and singular integral
operators”, Oper. Theory Adv. Appl., vol. 206 (2010), 157–184).
[50] I. Gohberg and N. Krupnik. Extension theorems for invertibility symbols in Banach
algebras. Integral Equations Operator Theory, 15:991–1010, 1992.
[51] I. Gohberg and N. Krupnik. One-dimensional Linear Singular Integral Equations
(I). Birkhäuser, Basel, 1992.
[52] I. Gohberg and N. Krupnik. One-dimensional Linear Singular Integral Equations
(II). Birkhäuser, Basel, 1992.
[53] I. Gohberg and N. Krupnik. Extension theorems for Fredholm and invertibility
symbols. Integral Equations Operator Theory, 16:514–529, 1993.
[54] R. Hagen, S. Roch, and B. Silbermann. Stability of spline approximation meth-
ods for multidimensional pseudodifferential operators. Integral Equations Operator
Theory, 19(1):25–64, 1994.
[55] R. Hagen, S. Roch, and B. Silbermann. Spectral Theory of Approximation Methods
for Convolution Equations. Birkhäuser, Basel, 1995.
[56] R. Hagen, S. Roch, and B. Silbermann. C ∗ -algebras and Numerical Analysis. Marcel
Dekker, Inc., New York, 2001.
[57] R. Hagen and B. Silbermann. A finite element collocation method for bisingular
integral equations. Applicable Anal., 19(2-3):117–135, 1985.
[58] R. Hagen and B. Silbermann. Local theory of the collocation method for the ap-
proximate solution of singular integral equations. II. In Seminar Analysis (Berlin,
1986/87), pages 41–56. Akad. Wiss. DDR, Berlin, 1987.
[59] R. Hagen and B. Silbermann. On the stability of the qualocation method. In Sem-
inar Analysis (Berlin, 1987/1988), pages 43–52. Akademie-Verlag, Berlin, 1988.
[60] R. Hagen and B. Silbermann. A Banach algebra approach to the stability projection
methods for singular integral equations. Math. Nachr., 140(1):285–297, 1989.
[61] P. Halmos. Two subspaces. Trans. Amer. Math. Soc., 144:381–389, 1969.
[62] P. Hartman. On completely continuous Hankel matrices. Proc. Amer. Math. Soc.,
9:862–866, 1958.
 40 Years of Algebraic Techniques in Numerical Analysis 353

[63] R.E. Hartwig and M.E. Fisher. Asymptotic behavior of Toeplitz matrices and de-
terminants. Arch. Rational Mech. Anal., 32:190–225, 1969.
[64] P. Junghanns, R. Kaiser, and G. Mastroianni. Collocation for singular integral
equations with fixed singularities of particular Mellin type. Electron. Trans. Numer.
Anal., 41:190–248, 2014.
[65] P. Junghanns, R. Kaiser, and D. Potts. Collocation-quadrature methods and fast
summation for Cauchy singular integral equations with fixed singularities. Linear
Algebra Appl., 491:187–238, 2016.
[66] P. Junghanns, G. Mastroianni, and M. Seidel. On the stability of collocation meth-
ods for Cauchy singular integral equations in weighted Lp spaces. Math. Nachr.,
283(1):58–84, 2010.
[67] P. Junghanns and A. Rathsfeld. On polynomial collocation for Cauchy singular
integral equations with fixed singularities. Integral Equations Operator Theory,
43(2):155–176, 2002.
[68] P. Junghanns and A. Rathsfeld. A polynomial collocation method for Cauchy singu-
lar integral equations over the interval. Electron. Trans. Numer. Anal., 14:79–126,
2002. Orthogonal polynomials, approximation theory, and harmonic analysis (Inzel,
2000).
[69] P. Junghanns, S. Roch, and B. Silbermann. Collocation methods for systems of
Cauchy singular integral equations on an interval. Vychisl. Tekhnol., 6(1):88–124,
2001.
[70] P. Junghanns, S. Roch, and U. Weber. Finite sections for singular integral op-
erators by weighted Chebyshev polynomials. Integral Equations Operator Theory,
21(3):319–333, 1995.
[71] P. Junghanns and B. Silbermann. Lokale Theorie des Kollokationsverfahrens zur
näherungsweisen Lösung singulärer Integralgleichungen. Technical report, Institut
für Mathematik, 1982. Preprint.
[72] P. Junghanns and B. Silbermann. Local theory of the collocation method for the
approximate solution of singular integral equations, I. Integral Equations Operator
Theory, 7(6):791–807, 1984.
[73] P. Junghanns and B. Silbermann. Numerical analysis for one-dimensional Cauchy
singular integral equations. Journal of Computational and Applied Mathematics,
125(1–2):395–421, 2000. Numerical Analysis 2000. Vol. VI: Ordinary Differential
Equations and Integral Equations.
[74] P. Junghanns and U. Weber. Local theory of projection methods for Cauchy singular
equations on an interval. In Boundary integral methods: numerical and mathemati-
cal aspects, volume 1 of Comput. Eng., pages 217–256. WIT Press/Comput. Mech.
Publ., Boston, MA, 1999.
[75] A. Karlovich, H. Mascarenhas, and P.A. Santos. Finite section method for a Banach
algebra of convolution type operators on Lp (R) with symbols generated by PC and
SO. Integral Equations Operator Theory, 67(4):559–600, 2010.
[76] A. Karlovich, H. Mascarenhas, and P.A. Santos. Erratum to: Finite section method
for a Banach algebra of convolution type operators on Lp (R) with symbols generated
by P C and SO. Integral Equations Operator Theory, 69(3):447–449, 2011.
354 P.A. Santos

[77] A.Y. Karlovich. Fredholmness of singular integral operators with piecewise contin-
uous coefficients on weighted Banach function spaces. J. Integral Equations Appli-
cations, 15(3):263–320, 09 2003.
[78] A.Y. Karlovich and P.A. Santos. On asymptotics of Toeplitz determinants with
symbols of nonstandard smoothness. J. Fourier Anal. Appl., 11(1):43–72, 2005.
[79] V. Kokilashvili, A. Meskhi, H. Rafeiro, and S. Samko. Integral Operators in Non-
Standard Function Spaces, volume 248 of Oper. Theory Adv. Appl. Birkhäuser,
Basel, 2016.
[80] A. Kozak. A certain projection method of solving operator equations in a Banach
space. Dokl. Akad. Nauk SSSR, 211:1042–1045, 1973 (in Russian; English transla-
tion in Soviet Math. Dokl. 14:1159–1162, 1973).
[81] A. Kozak. A local principle in the theory of projection methods. Dokl. Akad. Nauk
SSSR, 212:1287–1289, 1973 (in Russian; English translation in Soviet Math. Dokl.
14:1580–1583, 1974).
[82] N. Krupnik. Banach algebras with symbol and singular integral operators. Birk-
häuser, Basel, 1987.
[83] N. Krupnik, S. Roch, and B. Silbermann. On C∗ -algebras generated by idempotents.
J. Funct. Anal., 137(2):303–319, 1996.
[84] N. Krupnik and Y. Spigel. Invertibility symbols for a Banach algebra generated
by two idempotents and a shift. Integral Equations Operator Theory, 17:567–578,
1993.
[85] H.J. Landau. On Szegő’s eingenvalue distribution theorem and non-hermitian ker-
nels. Journal d’Analyse Mathématique, 28(1):335–357, 1975.
[86] M. Lindner. Infinite Matrices and their Finite Sections. Birkhäuser, Basel, Boston,
Berlin, 2006.
[87] M. Lindner and S. Roch. Finite sections of random Jacobi operators. SIAM J.
Numer. Anal., 50(1):287–306, 2012.
[88] H. Mascarenhas, P.A. Santos, and M. Seidel. Quasi-banded operators, convolutions
with almost periodic or quasi-continuous data, and their approximations. J. Math.
Anal. Appl., 418(2):938–963, 2014.
[89] H. Mascarenhas, P.A. Santos, and M. Seidel. Approximation sequences to operators
on banach spaces: a rich approach. Journal of the London Mathematical Society,
96(1):86–110, 2017.
[90] H. Mascarenhas and B. Silbermann. Convolution type operators on cones and their
finite sections. Math. Nachr., 278(3):290–311, 2005.
[91] H. Mascarenhas and B. Silbermann. Spectral approximation and index for con-
volution type operators on cones on Lp (R2 ). Integral Equations Operator Theory,
65(3):415–448, 2009.
[92] H. Mascarenhas and B. Silbermann. Sequences of variable-coefficient Toeplitz ma-
trices and their singular values. Journal of Functional Analysis, 270(4):1479 – 1500,
2016.
[93] S. Prössdorf. Ein Lokalisierungsprinzip in der Theorie der Spline-Approximationen
und einige Anwendungen. Math. Nachr., 119:239–255, 1984.
 40 Years of Algebraic Techniques in Numerical Analysis 355

[94] S. Prössdorf. A localization principle in the theory of finite element methods. In

Problems and methods of mathematical physics (Karl-Marx-Stadt, 1983), volume 63
of Teubner-Texte Math., pages 169–177. Teubner, Leipzig, 1984.
[95] S. Prössdorf and A. Rathsfeld. A spline collocation method for singular integral
equations with piecewise continuous coefficients. Integral Equations Operator The-
ory, 7(4):536–560, 1984.
[96] S. Prößdorf and A. Rathsfeld. On spline galerkin methods for singular integral
equations with piecewise continuous coefficients. Numer. Math., 48(1):99–118, 1986.
[97] S. Prössdorf and A. Rathsfeld. Stabilitätskriterien für Näherungsverfahren bei sin-
gulären Integralgleichungen in Lp . Z. Anal. Anwendungen, 6(6):539–558, 1987.
[98] S. Prössdorf and G. Schmidt. A finite element collocation method for singular in-
tegral equations. Math. Nachr., 100:33–60, 1981.
[99] S. Prössdorf and R. Schneider. Spline approximation methods for multidimen-
sional periodic pseudodifferential equations. Integral Equations Operator Theory,
15(4):626–672, 1992.
[100] S. Prössdorf and B. Silbermann. Numerical Analysis for Integral and Related Op-
erator Equations. Birkhäuser, Basel, 1991.
[101] V. Rabinovich, S. Roch, and B. Silbermann. Fredholm theory and finite sec-
tion method for band-dominated operators. Integral Equations Operator Theory,
30(4):452–495, 1998.
[102] V. Rabinovich, S. Roch, and B. Silbermann. Band-dominated operators with
operator-valued coefficients, their Fredholm properties and finite sections. Integral
Equations Operator Theory, 40(3):342–381, 2001.
[103] V. Rabinovich, S. Roch, and B. Silbermann. Limit operators and their applications
in operator theory, volume 150 of Oper. Theory Adv. Appl. Birkhäuser Verlag, Basel,
2004.
[104] E. Reich. On non-hermitian Toeplitz matrices. Mathematica Scandinavica, 10:145–
152, 1962.
[105] L. Reichel and L.N. Trefethen. Eigenvalues and pseudo-eigenvalues of Toeplitz ma-
trices. Linear Algebra and its Applications, 162:153–185, 1992.
[106] S. Roch. Extension-restriction theorems for algebras of approximation sequences.
to appear.
[107] S. Roch. Das Reduktionsverfahren für Produktsummen von Toeplitzoperatoren mit
stückweise stetigen Symbolen. Wiss. Z. Tech. Hochsch. Karl-Marx-Stadt, 26(2):265–
273, 1984.
[108] S. Roch. Das Reduktionsverfahren für Operatoren aus einer Toeplitzalgebra. Wiss.
Z. Tech. Hochsch. Karl-Marx-Stadt, 27(1):121–126, 1985.
[109] S. Roch. Finite sections of operators generated by singular integrals with Carleman
shift. Preprint 52. Technische Universität Karl-Marx-Stadt, 1987.
[110] S. Roch. Spline approximation methods cutting off singularities. Z. Anal. Anwen-
dungen, 13(2):329–345, 1994.
[111] S. Roch. Spline approximation methods for Wiener–Hopf operators. In Recent de-
velopments in operator theory and its applications (Winnipeg, MB, 1994), pages
282–308. Birkhäuser, Basel, 1996.
356 P.A. Santos

[112] S. Roch. Spectral approximation of Wiener–Hopf operators with almost periodic

generating function. Numer. Funct. Anal. Optim., 21(1-2):241–253, 2000.
[113] S. Roch. Finite Sections of Band-Dominated Operators, volume 191/895 of Memoirs
of the American Mathematical Society. AMS, 2008.
[114] S. Roch. The universal algebra generated by a power partial isometry. In Large
Truncated Toeplitz Matrices, Toeplitz Operators, and Related Topics, volume 259 of
Oper. Theory Adv. Appl., pages 649–662. Birkhäuser, Basel, 2017.
[115] S. Roch and P.A. Santos. Two points, one limit: Homogenization techniques for
two-point local algebras. J. Math. Anal. Appl., 391(2):552–566, 2012.
[116] S. Roch and P.A. Santos. Finite section approximations in an algebra of convolu-
tion, multiplication and flip operators on Lp (R). Journal of Approximation Theory,
186:64–97, 2014.
[117] S. Roch and P.A. Santos. A tour to compact type operators and sequences related to
the finite sections projection. In Operator theory, operator algebras and applications,
volume 242 of Oper. Theory Adv. Appl., pages 311–323. Birkhäuser/Springer Basel
AG, Basel, 2014.
[118] S. Roch, P.A. Santos, and B. Silbermann. Finite section method in some algebras
of multiplication and convolution operators and a flip. Z. Anal. Anwendungen,
16(3):575–606, 1997.
[119] S. Roch, P.A. Santos, and B. Silbermann. A sequence algebra of finite sections,
convolution and multiplication operators on Lp (R). Numer. Funct. Anal. Optim.,
31(1):45–77, 2010.
[120] S. Roch, P.A. Santos, and B. Silbermann. Non-commutative Gelfand Theories. Uni-
versitext. Springer-Verlag, London, 2011.
[121] S. Roch, P.A. Santos, and B. Silbermann. Corrigendum: “A sequence algebra of
finite sections, convolution, and multiplication operators on Lp (R)”. Numer. Funct.
Anal. Optim., 34(1):113–116, 2013.
[122] S. Roch and B. Silbermann. Das Reduktionsverfahren für Potenzen von Toeplitz-
operatoren mit unstetigem Symbol. Wiss. Z. Tech. Hochsch. Karl-Marx-Stadt,
24(3):289–294, 1982.
[123] S. Roch and B. Silbermann. Toeplitz-like operators, quasicommutator ideals, nu-
merical analysis. I. Math. Nachr., 120:141–173, 1985.
[124] S. Roch and B. Silbermann. Algebras generated by idempotents and the symbol
calculus for singular integral operators. Integral Equations Operator Theory, 11:385–
419, 1988.
[125] S. Roch and B. Silbermann. A symbol calculus for finite sections of singular integral
operators with flip and piecewise continuous coefficients. J. Funct. Anal., 78(2):365–
389, 1988.
[126] S. Roch and B. Silbermann. Non-strongly converging approximation methods.
Demonstr. Math., 22(3):651–676, 1989.
[127] S. Roch and B. Silbermann. Algebras of convolution operators and their image in the
Calkin algebra. Report, volume 157. R-MATH-05-90 des Karl-Weierstrass-Instituts
für Mathematik, Berlin, 1990.
[128] S. Roch and B. Silbermann. Limiting sets of eigenvalues and singular values of
Toeplitz matrices. Asymptotic Analysis, 8(3):293–309, 1994.
 40 Years of Algebraic Techniques in Numerical Analysis 357

[129] S. Roch and B. Silbermann. Index calculus for approximation methods and singular
value decomposition. J. Math. Anal. Appl., 225:401–426, 1998.
[130] S. Roch and B. Silbermann. Szegő limit theorems for operators with almost periodic
diagonals. Oper. Mat., 1(1):1–29, 2007.
[131] A. Rogozhin and B. Silbermann. Approximation numbers for the finite sections of
Toeplitz operators with piecewise continuous symbols. J. Funct. Anal., 237(1):135–
149, 2006.
[132] A. Rogozhin and B. Silbermann. Banach algebras of operator sequences: Approxi-
mation numbers. Journal of Operator Theory, 57(2):325–346, 2007.
[133] P.A. Santos. Spline approximation methods with uniform meshes in algebras of
multiplication and convolution operators. Math. Nachr., 232:95–127, 2001.
[134] P.A. Santos. Galerkin method with graded meshes for Wiener–Hopf operators with
PC symbols in Lp spaces. In Spectral Theory, Mathematical System Theory, Evolu-
tion Equations, Differential and Difference Equations, volume 221 of Oper. Theory:
Adv. Appl., pages 587–605. Birkhäuser, Basel, 2012.
[135] P.A. Santos and B. Silbermann. Galerkin method for Wiener–Hopf operators with
piecewise continuous symbol. Integral Equations Operator Theory, 38(1):66–80,
2000.
[136] P.A. Santos and B. Silbermann. An axiomatic approach to stability of approxima-
tion sequences. In Singular integral operators, factorization and applications, pages
369–381. Birkhäuser, Basel, 2003.
[137] P.A. Santos and B. Silbermann. An approximation theory for operators generated
by shifts. Numer. Funct. Anal. Optim., 27(3-4):451–484, 2006.
[138] D. Sarason. On products of Toeplitz operators. Acta Sci. Math. (Szeged), 35:7–12,
1973.
[139] P. Schmidt and F. Spitzer. The Toeplitz matrices of an arbitrary Laurent polyno-
mial. Mathematica Scandinavica, 8(1):15–38, 1960.
[140] M. Seidel. On an open question about the stability of the finite section method for a
class of convolution type operators. Integral Equations Operator Theory, 75(2):249–
255, 2013.
[141] M. Seidel and B. Silbermann. Banach algebras of operator sequences. Oper. Matri-
ces, 6(3):385–432, 2012.
[142] B. Silbermann. Lokale Theorie des Reduktionsverfahrens für Toeplitzoperatoren.
Math. Nachr., 104:137–146, 1981.
[143] B. Silbermann. Lokale Theorie des Reduktionsverfahrens für Singuläre Integral-
gleichungen. Z. Anal. Anwendungen, 1(6):45–56, 1982.
[144] B. Silbermann. On the limiting set of singular values of toeplitz operators. Lin. Alg.
Appl., 182:35–43, 1993.
[145] B. Silbermann. Algebraic techniques in stability analysis for spline approximation
methods. Numer. Funct. Anal. Optim., 17(7-8):833–841, 1996.
[146] B. Silbermann and O. Zabroda. Asymptotic behavior of generalized convolutions:
An algebraic approach. J. Integral Equations Applications, 18(2):169–196, 06 2006.
358 P.A. Santos

[147] I. Simonenko. A new general method of investigating linear operator equations of

singular integral equation type. I. Izv. Akad. Nauk SSSR Ser. Mat., 29:567–586,
1965 (in Russian).
[148] I. Simonenko. A new general method of investigating linear operator equations of
singular integral equation type. II. Izv. Akad. Nauk SSSR Ser. Mat., 29:757–782,
1965 (in Russian).
[149] I. Simonenko and C.N. Min . A local method in the theory of one-dimensional sin-
gular integral equations with piecewise continuous coefficients. Noethericity. Rostov.
Gos. Univ., Rostov-na-Donu, 1986 (in Russian).
[150] G. Szegő. Ein Grenzwertsatz über die Toeplitzschen Determinanten einer reellen
positiven Funktion. Math. Ann., 76(4):490–503, 1915.
[151] G. Szegő. On certain Hermitian forms associated with the Fourier series of a positive
function. Comm. Sém. Math. Univ. Lund [Medd. Lunds Univ. Mat. Sem.], (Tome
Supplémentaire):228–238, 1952.
[152] S. Treil. Invertibility of a Toeplitz operator does not imply its invertibility by the
projection method. Dokl. Akad. Nauk SSSR, 292(3):563–567, 1987.
[153] I.È. Verbickiı̆ and N.J. Krupnik. The applicability of the projection method to
discrete Wiener–Hopf equations with a piecewise continuous symbol. Mat. Issled.,
(Vyp. 45):17–28, 179, 1977 (in Russian).
[154] H. Widom. On the eigenvalues of certain Hermitian operators. Trans. Amer. Math.
Soc., 88:491–522, 1958.
[155] H. Widom. Asymptotic behavior of block Toeplitz matrices and determinants. Ad-
vances in Math., 13:284–322, 1974.
[156] H. Widom. Asymptotic inversion of convolution operators. Publications Mathéma-
tiques de l’Institut des Hautes Études Scientifiques, 44(1):191–240, 1974.
[157] H. Widom. Asymptotic behavior of block Toeplitz matrices and determinants. II.
Advances in Math., 21:1–29, 1976.
[158] H. Widom. On the singular values of Toeplitz matrices. Z. Anal. Anwendungen,
8(3):221–229, 1989.

Pedro A. Santos
Departamento de Matemática / INESC-ID
Instituto Superior Técnico
Universidade Técnica de Lisboa
Av. Rovisco Pais
1049-001 Lisboa, Portugal
e-mail: [email protected]
Operator Theory:
Advances and Applications, Vol. 267, 359–370

c Springer International Publishing AG, part of Springer Nature 2018

Linearizability of Multi-Control Systems of

the Class C 1 by Additive Change of Controls
K.V. Sklyar, G.M. Sklyar and S.Yu. Ignatovich

Abstract. We give necessary and sufficient conditions for nonlinear control

systems of the class C 1 with multi-dimensional control to be linearizable by
means of changes of variables of the class C 2 and additive change of the control
of class C 1 .
Mathematics Subject Classification (2010). Primary 93B18; Secondary 93C10.
Keywords. Systems with multi-dimensional control, linearizability problem,
C 1 -smooth nonlinear control systems, Lie brackets technique.

1. Introduction and statement of the problem

We consider the linearizability problem for systems of the form
ẋ = f (x, u), x ∈ Q ⊂ Rn , u ∈ Rr , (1.1)
where the vector function f (x, u) is continuously differentiable, i.e., f (x, u) ∈
C 1 (Q × Rr ). More specifically, we say that the system (1.1) is locally linearizable
in the domain Q if there exists a change of variables
z = F (x) ∈ C 2 (Q) such that det Fx (x) = 0, x ∈ Q, (1.2)
which reduces the system (1.1) to a linear form
ż = Az + Bu, z ∈ Rn , u ∈ Rr , (1.3)
satisfying the controllability condition
rank(B, AB, . . . , An−1 B) = n and rank B = r. (1.4)
We say that (1.1) is locally feedback linearizable in the domain Q if there exists a
change of variables (1.2) and a change of controls
v = g(x, u) ∈ C 1 (Q × Rr )
(1.5)
such that g(x, Rr ) = Rr , det gu (x, u) = 0, x ∈ Q, u ∈ Rr ,

This work was financially supported by Polish National Science Centre grant no. 2017/25/
B/ST1/01892.
360 K.V. Sklyar, G.M. Sklyar and S.Yu. Ignatovich

which reduce (1.1) to a linear form

ż = Az + Bv, z ∈ Rn , v ∈ Rr , (1.6)
satisfying condition (1.4).
The linearizability problem was deeply and completely studied for the class
of smooth systems, that is, for systems with f (x, u) ∈ C ∞ (Q×Rr ), during 70–80th
[1, 2, 4, 6, 7, 13]. The main approaches were connected with differential-geometric
tools, in particular, with Lie brackets technique.
Another ideas in the linearizability problem were proposed in [5] where a spe-
cial subclass of linearizable systems (triangular systems) was described. Within this
approach, smoothness requirements could be weaken: in fact, nonlinear systems of
the class C 1 were treated.
In [11] the problem of linearizability in the class C 1 was considered and condi-
tions of linearizability were formulated for systems with one-dimensional control.
It turned out that in both cases – linearizability and feedback linearizability –
well-known conditions for C ∞ -smooth systems should be modified. For the latter
case this adjustment was rather significant. In particular, it turned out that Lie
brackets which were the basis of the whole analysis in C ∞ -smooth case may not
exist. Instead of Lie brackets, in [11] it was proposed to introduce new vector fields
which were built recursively. In [10] and [9] these results were extended to the case
of systems with multi-dimensional control.
Along with a change of controls of the general form, more specific cases are
studied separately. In this paper we consider additive change of controls of the
form
v = g(x) + u such that g(x) ∈ C 1 (Q), x ∈ Q. (1.7)
Definition 1.1. We say that a control system of the form (1.1), where f (x, u) ∈
C 1 (Q × Rr ), is locally linearizable in the domain Q by additive change of controls
or, for brevity, A-linearizable, if there exists a change of variables of the form (1.2)
and a change of controls of the form (1.7) which reduce the system (1.1) to a linear
form (1.6), (1.4).
Such change of controls allows usage of results from linear control theory. For
example, in [12] for A-linearizable systems the method of constructing a bang-bang
control was proposed where the known methods for linear systems were applied.
Below we use the following common notation. By [·, ·] we denote the Lie
bracket of vector fields, i.e., [a(x), b(x)] = bx (x)a(x) − ax (x)b(x). Also, we use the
notation ad0a b(x) = b(x), adk+1a b(x) = [a(x), adka b(x)], k ≥ 0. For a vector field
a(x) and a function ϕ(x), a Lie derivative equals La ϕ(x) = ϕx (x)a(x), and Lka
denotes the kth power of the operator La , k ≥ 0.
Conditions of A-linearizability for systems with one-dimensional control were
obtained in [8], and also in [11] as a corollary of the general result. In this paper
we consider multi-control systems. Instead of a direct generalization, we propose
conditions of A-linearizability which are close to the conditions of linearizability
[10]. In particular, for A-linearizability, a system should be affine, i.e., have the
 Linearizability by Additive Change of Controls 361

form (2.4). As in the general feedback linearizable case, Lie brackets adka bi (x)
may not exist. In fact, if a(x) and bi (x) are of class C 1 (Q) then ada bi (x) =
(bi )x (x)a(x) − ax (x)bi (x) are necessarily of class C(Q), however, may not belong
to the class C 1 (Q). If this is the case, ad2a bi (x) cannot be found. This situation is
shown below in Example 2.1. So, as in the general feedback linearizable case [9],
vector fields χki (x) are introduced. However, unlike the general case, a recursive
formula for χki (x) includes only χ01 (x), . . . , χ0r (x) (see (2.2)). Moreover, it turns out
that, like in the linearizable case [10], condition (B4) of [11, Theorem 2] can be
omitted.

2. A-linearizability conditions
Theorem 2.1. A nonlinear system of the form (1.1), where f (x, u) ∈ C 1 (Q×Rr ), is
A-linearizable in the domain Q if and only if there exist integers n1 ≥ · · · ≥ nr ≥ 1,
n1 + · · · + nr = n, and a permutation {m1 , . . . , mr } of the set {1, . . . , r} such that
the following conditions hold:
r
(A) f (x, u) = a(x) + bi (x)ui , where a(x), b1 (x), . . . , br (x) ∈ C 1 (Q);
i=1
(B1) there exist functions μpks (x) ∈ C(Q) such that the vector fields χks (x), s =
1, . . . , r, k = 0, . . . , ns , which are defined recursively by
χ0s (x) = bms (x), (2.1)

r
χk+1
s (x) = [a(x), χks (x)] + μpks (x)χ0p (x), k = 0, . . . , ns − 1, (2.2)
p=1

exist and belongs to C 1 (Q);

(B2) rank M (x) = n for x ∈ Q, where
M (x) = χ01 (x), . . . , χn1 1 −1 (x), . . . , χ0r (x), . . . , χnr r −1 (x) ,
and, additionally, for any s = 1, . . . , r and any x ∈ Q
 
χns s (x) ∈ Lin χip (x) : 1 ≤ p ≤ r, 0 ≤ i ≤ min{ns , np − 1} ; (2.3)
(B3) [χis (x), χjq (x)] = 0, x ∈ Q, for all s, q = 1, . . . , r, i = 0, . . . , ns , j = 0, . . . , nq .
In this case, the numbers n1 , . . . , nr are controllability indices of the resulting linear
system.
Proof. Necessity. Suppose the system (1.1) is A-linearizable in the domain Q.
Namely, suppose that a change of variables (1.2) and a change of controls (1.7)
reduce the system (1.1) to a linear form (1.6), (1.4). Then for new variables z =
F (x) we have
ż = Fx (x)f (x, u) = Az + Bv = AF (x) + Bg(x) + Bu.
Therefore,
Fx (x)f (x, u) = AF (x) + Bg(x) + Bu.
362 K.V. Sklyar, G.M. Sklyar and S.Yu. Ignatovich

Differentiating this equality w.r.t. u we get Fx (x)fu (x, u) = B. Hence,

fu (x, u) = (Fx (x))−1 B,

what means that f (x, u) is linear on u, i.e., it satisfies condition (A). So, the initial
system has the form
 r
ẋ = a(x) + bi (x)ui , (2.4)
i=1

where a(x), b1 (x), . . . , br (x) ∈ C (Q).

1

Now let us implement the change of controls (1.7) to the system (2.4), i.e.,
substitute u = v − g(x). We get

r 
r
ẋ = a(x) − bi (x)gi (x) + bi (x)vi ,
i=1 i=1

and, by our supposition, this system is linearizable by a change of variables only,

since the change of variables (1.2) reduces the system to a linear form (1.6), (1.4).
So, we can apply known conditions of linearizability in the class C 1 [10, 11].
Namely, let us denote
 r

a(x) = a(x) − bi (x)gi (x).
i=1

Taking into account [10, Theorem 2.1 and Remark 2.4], we get that all vector fields
adka bs (x) exist, belong to the class C 1 (Q), and satisfy equalities
7 8
adka bs (x), adja bq (x) = 0.

Let us denote by n1 ≥ · · · ≥ nr ≥ 1 the controllability indices of the corresponding

linear system. Then there exists a permutation {m1 , . . . , mr } of the set {1, . . . , r}
such that

rank{ad0a bm1 (x), . . . , adna 1 −1 bm1 (x), . . . , ad0a bmr (x), . . . , adna r −1 bmr (x)} = n

and
 
adans bms (x) ∈ Lin adia bmp (x) : 1 ≤ p ≤ r, 0 ≤ i ≤ min{ns , np − 1} .

Thus, let us denote

χks (x) = adka bms (x), s = 1, . . . , r, k ≥ 0, (2.5)

then all these vector fields exist, belong to the class C 1 (Q), and satisfy conditions
(B2) and (B3).
It remains to check that χks (x) defined by (2.5) satisfy the recursive formula
(2.1), (2.2). Equality (2.1) follows directly from the definition of χks (x). For k ≥ 0
 Linearizability by Additive Change of Controls 363

using condition (B3) we get


r
χk+1
s (x) = adk+1
a bms (x) = a(x), adka bms (x)]
[ = [a(x) − bi (x)gi (x), χks (x)]
i=1

r
= [a(x), χks (x)] − [χ0p (x)gmp (x), χks (x)] (2.6)
p=1
r r
= [a(x), χks (x)] − [χ0p (x), χks (x)] gmp (x) + (gmp (x))x χks (x) χ0p (x).
p=1
A BC D p=1
=0

Let us denote μpks (x)

= (gmp (x))x χks (x). Since gi (x) ∈ C 1 (Q), we get μpks (x) ∈
C(Q); therefore, (2.6) leads to (2.2).
Sufficiency. Suppose conditions of the theorem are satisfied. For convenience,
let us introduce the notation
Ω = {(p, i) : 1 ≤ p ≤ r, 0 ≤ i ≤ np − 1},
then condition (B2) implies that the set of columns of the matrix M (x)
P (x) = {χip (x) : (p, i) ∈ Ω}
is a basis of Rn for any x ∈ Q. Let us expand χns s (x) w.r.t. the basis P (x) and
take into account condition (2.3). We get

χns s (x) = s
νp,i (x)χip (x), x ∈ Q,
(p,i)∈Ω, i≤ns
s
where νp,i (x) are some functions. Since all functions χip (x) and χns s (x) belong to
1
the class C (Q) and P (x) is a basis, we get νp,i s
(x) ∈ C 1 (Q). Now we prove that
νp,i (x) = const. Let us consider any (m, k) ∈ Ω. Using (B3), we get
s

[χkm (x), χns s (x)] = [χkm (x), νp,i
s
(x)χip (x)]
(p,i)∈Ω, i≤ns
⎛ ⎞
 ⎜ k ⎟
= ⎝[χm (x), χip (x)] νp,i
s s
(x) + (νp,i (x))x χkm (x) χip (x)⎠ = 0.
(p,i)∈Ω, i≤ns
A BC D
=0
s
Since P (x) is a basis, we get (νp,i (x))x χkm (x) = 0 for any (p, i) ∈ Ω, i ≤ ns . Since
this equality holds for any (m, k) ∈ Ω and P (x) is a basis, we get (νp,i s
(x))x = 0.
s s s
Hence, functions νp,i (x) do not depend on x, i.e., νp,i (x) = νp,i , and

χns s (x) = s
νp,i χip (x), x ∈ Q, s = 1, . . . , r. (2.7)
(p,i)∈Ω, i≤ns

Now, for any fixed q = 1, . . . , r, we consider the system of partial differential

equations with coefficients in C 1 (Q) of the form
(ϕ(x))x χjs (x) = 0, (s, j) ∈ Ω, (s, j) = (q, nq − 1),
(2.8)
(ϕ(x))x χnq q −1 (x) = (−1)nq −1 .
364 K.V. Sklyar, G.M. Sklyar and S.Yu. Ignatovich

In a matrix form, the system can be written as (ϕ(x))x M (x) = (−1)nq −1 ep , where
ep is the unit row vector with 1 on the pth place, p = n1 + · · · + nq . Then (2.8) is
equivalent to the system

(ϕ(x))x = (−1)nq −1 ep M −1 (x),

where the right-hand side belongs to the class C 1 (Q). It can be shown analogously
to [10] that the conditions of the Frobenius Theorem [3, Chapter VI] are satisfied
and there exists a solution ϕq (x) of this system such that ϕq (x) ∈ C 2 (Q). Our
goal is to prove that for k = 0, . . . , nq − 1

Lka ϕq (x) ∈ C 2 (Q). (2.9)

Moreover, we also prove that

(Lka ϕq (x))x χjs (x) = (−1)k yq,s

k+j
= const,
(2.10)
s = 1, . . . , r, j = 0, . . . , ns ,

for k = 0, . . . , nq − 1, where, in particular, for s = 1, . . . , r

m
yq,s = 0 for 0 ≤ m ≤ nq − 2, s = q, (2.11)
m
yq,q = 0 for 0 ≤ m ≤ nq − 2, (2.12)
nq −1 nq −1
yq,q = (−1) , (2.13)
m
yq,s = 0 for 0 ≤ m ≤ ns − 1, s = q. (2.14)

For k = 0, (2.9) was justified above and (2.10) follows immediately from (2.8)
and (2.7), as well as (2.12)–(2.14). Let us prove that, additionally,
ns
yq,s = 0 if ns ≤ nq − 2, s = q. (2.15)

In fact, in this case nq − 1 ≥ ns + 1, therefore, (2.7), (2.8), and (2.14) imply


ns
yq,s = (ϕq (x))x χns s (x) = s
νp,i (ϕq (x))x χip (x) = 0.
(p,i)∈Ω, i≤ns

Thus, (2.14) and (2.15) imply

m
yq,s = 0 for 0 ≤ m ≤ min{ns , nq − 2}, s = q.

Now suppose that for some d such that 0 ≤ d ≤ nq − 2, relations (2.9), (2.10)
and
m
yq,s = 0 for 0 ≤ m ≤ min{ns + k, nq − 2}, s = q, (2.16)

hold for k = 0, . . . , d. Let us prove that they are true also for k = d + 1.
 Linearizability by Additive Change of Controls 365

First, consider any (s, j) ∈ Ω. Notice that (2.16) for k = d gives yq,p
d
= 0 for
any p = 1, . . . , r. Using (2.2) and the induction supposition we get
(Ld+1 j d j
a ϕq (x))x χs (x) = (La ϕq (x))x a(x) x χs (x)

= (Lda ϕq (x))x χjs (x) x a(x) − (Lda ϕq (x))x [a(x), χjs (x)]
A BC D
=0

r
= −(Lda ϕq (x))x χj+1 (x) + μpjs (x) (Lda ϕq (x))x χ0p (x)
A BC s
D p=1
A BC D
d+j+1
=(−1)d+1 yq,s =(−1)d yq,p
d =0

= (−1)d+1 yq,s
d+j+1
.
Thus, for any (s, j) ∈ Ω we have
(Ld+1 j
a ϕq (x))x χs (x) = (−1)
d+1 d+j+1
yq,s . (2.17)
−1
Using the matrix form we get (Ld+1
a ϕq (x))x = yM (x), where y is a constant
d+1 d+j+1
row vector with components (−1) yq,s . Since M −1 (x) ∈ C 1 (Q), we get

a ϕq (x) ∈ C (Q).
Ld+1 2
(2.18)
Now let us consider any s = 1, . . . , r and j = ns . Applying (2.7) and (2.17),
we get

(Ld+1 ns
a ϕq (x))x χs (x) =
s
νp,i (Ld+1 i
a ϕq (x))x χp (x) = (−1)
d+1 d+ns +1
yq,s
(p,i)∈Ω, i≤n
A BC D
s
d+i+1
=(−1)d+1 yq,p
(2.19)
where 
d+ns +1 s d+i+1
yq,s = νp,i yq,p = const . (2.20)
(p,i)∈Ω, i≤ns

Thus, (2.17)–(2.20) imply that (2.9) and (2.10) hold for k = d + 1.

Now let us notice that if ns + d ≥ nq − 2 then (2.16) for k = d implies (2.16)
for k = d + 1 and, moreover, in this case (2.11) holds.
Suppose d is such that ns + d + 1 ≤ nq − 2. Let us consider yq,p d+i+1
from
(2.20): we have d + i + 1 ≤ d + ns + 1 ≤ nq − 2 since i ≤ ns , and d + i + 1 ≤ d + np
since (p, i) ∈ Ω. Hence, d + i + 1 ≤ min{np + d, nq − 2}, therefore, each term in the
right-hand side of (2.20) vanishes due to the induction supposition. Hence, (2.16)
holds for k = d + 1.
The induction arguments complete the proof of (2.11) and (2.9), (2.10) for
k = 0, . . . , nq − 1.
Let us denote σ1 = 0 and σq = n1 + · · · + nq−1 for q = 2, . . . , r. We consider
the change of variables z = F (x) where
Fσq +k (x) = Lk−1
a ϕq (x), q = 1, . . . , r, k = 1, . . . , nq .
As was shown above, F (x) ∈ C 2 (Q). One can prove analogously to [10] that the
map z = F (x) is locally invertible, i.e., det Fx (x) = 0 for x ∈ Q.
366 K.V. Sklyar, G.M. Sklyar and S.Yu. Ignatovich

In the new variables, for k = 1, . . . , nq we have


r
żσq +k = (Fσq +k (x))x ẋ = Lk−1
a ϕq (x) x
a(x) + bmi (x)umi
i=1

r 
r
k−1
= Lka ϕq (x) + Lk−1
a ϕq (x) χ0 (x)umi = Lka ϕq (x) +
x i
(−1)k−1 yq,i u mi .
i=1 i=1

Taking into account (2.11)–(2.14) we get k−1

yq,i = 0 for k = 1, . . . , nq − 1, i =
n −1 n −1
1, . . . , r. For k = nq , we get yq,iq = 0 if ni = nq , i = q, and yq,qq = (−1)nq −1 .
n −1
Also, (2.14) implies that yq,iq = 0 if ni ≥ nq , i = q. Hence,
żσq +k = zσq +k+1 , k = 1, . . . , nq − 1,
 n −1
żσq +nq = Lna q ϕq (x) + umq + (−1)nq −1 yq,iq umi .
ni ≤nq −1

Now we define a change of controls. Let us introduce the constant matrix

n −1
S = {Sqi }rq,i=1 with elements Sqi = (−1)nq −1 yq,iq . Since n1 ≥ · · · ≥ nr , this
matrix is upper triangular with unit diagonal elements. Denote by h(x) the vector
n
with elements hq (x) = La q ϕq (x) ∈ C 1 (Q), then

r
żσq +nq = hq (x) + Sqi umi .
i=1
Let us define
gmi (x) = (S −1 h(x))i ∈ C 1 (Q), i = 1, . . . , r,
and introduce the additive change of controls of the form v = g(x) + u. Then

r 
r  n −1
żσq +nq = hq (x) + Sqi umi = Sqi vmi = vmq + (−1)nq −1 yq,iq vmi .
i=1 i=1 ni ≤nq −1

Thus, the change of variables z = F (x) and the additive change of controls
v = g(x) + u reduce the initial system to the linear form
żσq +k = zσq +k+1 , k = 1, . . . , nq − 1,
 n −1
żσq +nq = vmq + (−1)nq −1 yq,iq vmi , q = 1, . . . , r.
ni ≤nq −1

Finally, we note that the numbers n1 , . . . , nr are controllability indices of this

linear system. 
Example 2.1. Consider the system of the class C 1
ẋ1 = x2 + x22 |x2 |,
ẋ2 = x4 |x4 | + 1+3x12 |x2 | u1 ,
x4 (2.21)
ẋ3 = 1−3|x 3 |x3
,
ẋ4 = u1 + u2 ,
 Linearizability by Additive Change of Controls 367

in the domain Q = {x ∈ R4 : x2 > − √13 , x3 < √13 }. For brevity, denote α(x) =
x + x2 |x| ∈ C 2 , β(x) = x − x2 |x| ∈ C 2 , γ(x) = x|x| ∈ C 1 , then the system can be
rewritten as
1
ẋ1 = α(x2 ), ẋ2 = γ(x4 ) + α (x 2)
u1 ,
x4
ẋ3 = β  (x3 ) , ẋ4 = u1 + u2 .
We have
⎛ ⎞ ⎛ ⎞
α(x2 ) 0
⎜ γ(x4 ) ⎟ ⎜ 1 ⎟
a(x) = ⎜ ⎟ 0 ⎜
⎝ β x(x4 ) ⎠ , χ1 (x) = b1 (x) = ⎝
α (x2 ) ⎟ , χ0 (x) = b2 (x) = e4 .
⎠ 2
3
0
0 1
Then
⎛ ⎞ ⎛ ⎞
−1
' ( 0
⎜  ⎟ ⎜ −γ  ⎟
⎜ γ(x4 ) 1
− γ (x4 ) ⎟ (x4 )
[a(x), χ01 (x)] = ⎜ α (x2 ) ⎟, [a(x), χ02 (x)] = ⎜
⎝ − 1
⎟.
⎠
⎝ − β  (x3 )
1 ⎠ β  (x3 )
0 0

Obviously, [a(x), χ01 (x)] and [a(x), χ02 (x)] do not belong to the class C 1 (Q). Since
χ0i (x) = bi (x), we conclude that ada bi (x) = [a(x), bi (x)] do not belong to C 1 (Q),
therefore, ad2a bi (x) do not exist.
' ' (
However, let us introduce μ101 (x) = −μ201 (x) = −α (x2 ) γ(x4 ) α (x 1
−
( 2)

γ  (x4 ) ∈ C(Q) and μ102 (x) = −μ202 (x) = α (x2 )γ  (x4 ) ∈ C(Q), then

χ11 (x) = [a(x), χ01 (x)] + μ101 (x)χ01 (x) + μ201 (x)χ02 (x) = −e1 − 1
β  (x3 ) e3 ,

χ12 (x) = [a(x), χ02 (x)] + μ102 (x)χ01 (x) + μ202 (x)χ02 (x) = − β  (x
1
3)
e3 ,
so, χ11 (x), χ12 (x) ∈ C 1 (Q). One easily finds χ21 (x) = [a(x), χ11 (x)] = 0 and χ22 (x) =
[a(x), χ12 (x)] = 0. Obviously, vector fields χ01 (x), χ02 (x), χ11 (x), χ12 (x) are linearly
independent and all their Lie brackets vanish. Therefore, all conditions of Theorem
2.1 are satisfied (with n1 = n2 = 2), hence, the system is A-linearizable.
Let us find linearizing changes of variables and controls. We have two systems
of first-order partial differential equations
(ϕ1 (x))x χ01 (x) = (ϕ1 (x))x χ02 (x) = (ϕ1 (x))x χ12 (x) = 0, (ϕ1 (x))x χ11 (x) = −1,
(ϕ2 (x))x χ01 (x) = (ϕ2 (x))x χ11 (x) = (ϕ2 (x))x χ02 (x) = 0, (ϕ2 (x))x χ12 (x) = −1,
and we may choose
ϕ1 (x) = x1 , ϕ2 (x) = β(x3 ) − x1 = x3 − x23 |x3 | − x1
(the solutions are unique up to a constant). Let us define z = F (x) as
z1 = F1 (x) = ϕ1 (x) = x1 , z2 = F2 (x) = La ϕ1 (x) = α(x2 ),
z3 = F3 (x) = ϕ2 (x) = β(x3 ) − x1 , z4 = F4 (x) = La ϕ2 (x) = −α(x2 ) + x4 .
368 K.V. Sklyar, G.M. Sklyar and S.Yu. Ignatovich

Then det Fx (x) = α (x2 )β  (x3 ) = (1 + 3|x2 |x2 )(1 − 3|x3 |x3 ) = 0 for x ∈ Q. In the
new variables, we have
ż1 = z2 , ż2 = α (x2 )γ(x4 ) + u1 ,
A BC D
=v1
ż3 = z4 , ż4 = −α (x2 )γ(x4 ) + u2 .
A BC D
=v2

So, we choose v = g(x) + u, where

g1 (x) = α (x2 )γ(x4 ) = (1 + 3x2 |x2 |)x4 |x4 |, g2 (x) = −g1 (x),

then g(x) ∈ C 1 (Q). Thus, the change of variables z = F (x) and the additive
change of controls v = g(x) + u reduce the system (2.21) to the linear form

ż1 = z2 , ż2 = v1 , ż3 = z4 , ż4 = v2 .

Remark 2.1. The true source of the results like Theorem 2.1 is the class of trian-
gular systems first proposed in the remarkable paper [5]. Namely, a control system
is called triangular if it is of the form
ẋi = fi (x1 , . . . , xi+1 ), i = 1, . . . , n − 1,
ẋn = fn (x1 , . . . , xn , u),
where x ∈ Rn , u ∈ R. Such systems arise in a plenty of applications; for example,
many studies were initiated by control problems for satellites. It was proved in
[5] that a triangular system is (globally) feedback linearizable under rather weak
requirements on smoothness of the functions fi ; namely, fi ∈ C n+1−i (Ri+1 ), i =
1, . . . , n, i.e., in fact, the right-hand side of the system is of class C 1 (Rn+1 ). It
follows from a constructive proof [5] that the change of a control becomes additive,
i.e., takes the form v = g(x) + u, iff ∂x ∂f1
2
· · · ∂f
∂u = c1 (x1 ), where c1 ∈ C (R),
n n

c1 (x1 ) = 0.

Example 2.2. In [5] a generalization for multi-control systems was mentioned where
xi and u are m-dimensional vectors rather than numbers. As a very simple exam-
ple, we consider a partial case of a triangular system with two controls
ẋ1 = f1 (x1 , x2 ) + x3 ,
ẋ2 = f2 (x1 , x2 ) + x4 ,
(2.22)
ẋ3 = f3 (x1 , x2 , x3 , x4 ) + u1 ,
ẋ4 = f4 (x1 , x2 , x3 , x4 ) + u2 ,

where f1 , f2 ∈ C 2 (R2 ), f3 , f4 ∈ C 1 (R4 ). A special form of the system allows us

to see immediately that it is A-linearizable by the change of variables y1 = x1 ,
y2 = x2 , x3 = f1 (x1 , x2 ) + x3 , x4 = f2 (x1 , x2 ) + x4 , and by the change of controls
v1 = (f1 )x1 (f1 + x3 ) + (f1 )x2 (f2 + x4 ) + f3 + u1 , v2 = (f2 )x1 (f1 + x3 ) + (f2 )x2 (f2 +
x4 ) + f4 + u2 .
 Linearizability by Additive Change of Controls 369

Concerning conditions of Theorem 2.1, we have χ01 = b1 = e3 , χ02 = b2 = e4 ,

and ⎛ ⎞ ⎛ ⎞
−1 0
⎜ 0 ⎟ ⎜ −1 ⎟
[a(x), χ01 (x)] = ⎜ ⎟
⎝ −(f3 )x3 ⎠ , [a(x), χ02 (x)] = ⎜ ⎟
⎝ −(f3 )x4 ⎠ .
−(f4 )x3 −(f4 )x4
Obviously, [a(x), χ0i (x)] are continuous, however, they may be non-differentiable.
We choose χ11 (x) = [a(x), χ01 (x)] + (f3 )x3 χ01 (x) + (f4 )x3 χ02 (x) = −e1 , χ12 (x) =
[a(x), χ02 (x)] + (f3 )x4 χ01 (x) + (f4 )x4 χ02 (x) = −e2 , which satisfy conditions of The-
orem 2.1.
Acknowledgement
We are grateful to the anonymous referee for his/her careful review of our manu-
script and helpful comments.

References
[1] R.W. Brockett, Feedback invariance for nonlinear systems. Proc. 7 World Cngress
IFAC, Helsinki, 1978, 1115–1120.
[2] W. Dayawansa, W.M. Boothby, and D.L. Elliot, Global state and feedback equivalence
of nonlinear systems. Systems and Control Lett. 6 (1985), 229–234.
[3] P. Hartman, Ordinary Differential Equations. John Wiley and Sons, New York, Lon-
don, Sydney, 1964.
[4] B. Jakubczyk and W. Respondek, On linearization of control systems. Bull. Acad.
Sci. Polonaise Ser. Sci. Math. 28 (1980), 517–522.
[5] V.I. Korobov, Controllability, stability of some nonlinear systems. Differential Equa-
tions 9 (1973), 466–469 (1975).
[6] A. Krener, On the equivalence of control systems and the linearization of non-linear
systems. SIAM J. Control 11 (1973), 670–676.
[7] W. Respondek, Global aspects of linearization, equivalence to polynomial forms and
decomposition of nonlinear control systems. In: Algebraic and Geometric Methods
in Nonlinear Control Theory, D. Reidel P.C., 1986, 257–284.
[8] E.V. Sklyar, On the class of non-linear control systems mapping onto linear systems.
Mat. Fiz. Anal. Geom. 8 (2001), 205–214.
[9] K.V. Sklyar and S.Yu. Ignatovich, Linearizability of systems of the class C 1 with
multi-dimensional control. Systems and Control Lett. 94 (2016), 92–96.
[10] K.V. Sklyar, S.Yu. Ignatovich, and V.O. Skoryk, Conditions of linearizability for
multi-control systems of the class C 1 . Commun. Math. Anal. 17 (2014), 359–365.
[11] G.M. Sklyar, K.V. Sklyar, and S.Yu. Ignatovich, On the extension of the Korobov’s
class of linearizable triangular systems by nonlinear control systems of the class C 1 .
Systems and Control Lett. 54 (2005), 1097–1108.
[12] K.V. Sklyar, G.M. Sklyar, and Yu.I. Karlovich, On bang-bang controls for some
nonlinear systems. Commun. Math. Anal. 14 (2013), 163–178.
370 K.V. Sklyar, G.M. Sklyar and S.Yu. Ignatovich

[13] R. Su, On the linear equivalents of nonlinear systems. Systems and Control Lett. 2
(1982), 48–52.

K.V. Sklyar and G.M. Sklyar

University of Szczecin
Wielkopolska 15
Szczecin 70-451, Poland
e-mail: [email protected]
[email protected]
S.Yu. Ignatovich
Department of Applied Mathematics
V.N. Karazin Kharkiv National University
Svobody sqr. 4,
Kharkiv 61022, Ukraine
e-mail: [email protected]
Operator Theory:
Advances and Applications, Vol. 267, 371–376

c Springer International Publishing AG, part of Springer Nature 2018

A Distance Formula Related to a Family of

Projections Orthogonal to Their Symmetries
Ilya M. Spitkovsky

Abstract. Let u be a Hermitian involution, and e an orthogonal projection,

acting on the same Hilbert space H. We establish the exact formula, in terms
of eue, for the distance from e to the set of all orthogonal projections q
from the algebra generated by e, u, and such that quq = 0.

Mathematics Subject Classification (2010). Primary 47A05; Secondary 47A30.

Keywords. Orthogonal projection, involution, C ∗ -algebra, W ∗ -algebra.

1. Introduction
Let H be a Hilbert space and let B(H) stand for the C ∗ -algebra of all bounded
linear operators acting on H. Given a Hermitian involution u ∈ B(H), denote by
Qu the set of all orthogonal projections q ∈ B(H) for which quq = 0.
Theorem 1.2 of [6] can be stated as follows:
Theorem 1. Let e ∈ B(H) be an orthogonal projection such that
eue < ξ,
where ξ(≈ 0.455) is the positive root of
 
4
x2 2 + √ = 1.
1 + 1 − 4x2
Then there exists q ∈ Qu for which
1
e − q ≤ eue + 4 eue2 . (1)
2
Further, q is in the C ∗ -subalgebra of B(H) generated by e, ueu∗ .

The author was supported in part by Faculty Research funding from the Division of Science and
Mathematics, New York University Abu Dhabi.
372 I.M. Spitkovsky

Note that the distance between any two orthogonal projections does not
exceed one. So, estimate (1) is useful only when eue is smaller than the positive
root of 8x2 + x − 2, that is, approximately 0.441 (< ξ).
We will provide an explicit formula for the distance from e to the intersection
of Qu with the W ∗ -algebra W(e, u) generated by e, u, as well as for the element on
which this distance is attained. No a priori restriction on eue is needed, and the
respective q indeed lies in the C ∗ -algebra C(e, ueu∗) generated by e, ueu∗ whenever
eue < 1.
Theorem 2. Let e, u ∈ B(H) be, respectively, an orthogonal projection and a Her-
mitian involution. Denote by H± the eigenspace of u corresponding to its eigen-
value ±1. Then the distance d from e to Qu ∩ W(e, u) is one if the range of e has
a non-trivial intersection with H+ or H− , and is given by the formula
4  , 
1 2
d= 1 − 1 − eue (2)
2
otherwise.
For small values of eue, it is instructive to compare (1) with the Taylor
expansion of (2):
d = 12 eue + 161
eue3 + · · ·

Figure 1. Estimate (1) versus formula (2) as functions of eue

 A Distance Formula Related to a Family of Projections 373

2. Proof of the main result

Using the canonical representation [3] (see also [4] or a more recent survey [1]) of
the pair e, (u + I)/2 of orthogonal projections, we can find an orthogonal decom-
position of H into six summands,
H = M00 ⊕ M01 ⊕ M10 ⊕ M11 ⊕ (M ⊕ M) , (3)
with respect to which
u = I ⊕ I ⊕ (−I) ⊕ (−I) ⊕ diag[I, −I],
 + 
+ H H(I − H) (4)
e=I ⊕0⊕0⊕I ⊕ .
H(I − H) I −H
(Here and in what follows we use the notation diag[X1 , . . . , Xk ] for block diagonal
matrices with X1 , . . . Xk as their diagonal blocks.) Note that in (3) the subspaces
M00 and M11 (resp, M01 and M10 ) are the intersections of the range (resp, the
kernel) of e with H+ and H− . The (Hermitian) operator H is the compression of
e onto M := H+ , (M00 ⊕ M01 ). By construction, H has its spectrum Δ lying
in [0, 1] and 0, 1 are not its eigenvalues.
Elements of W(e, u) with respect to the same decomposition (3) look as
q = a00 I ⊕ a01 I ⊕ a10 I ⊕ a11 I ⊕ Φ(H), (5)
 
φ φ01
where Φ = 00 , aij ∈ C, and the functions φij are Borel-measurable and
φ10 φ11
essentially bounded on Δ, in the sense of the spectral measure of H ([2], see
also [1, 4]). Consequently, q ∈ W(e, u) is an orthogonal projection if and only
if aij ∈ {0, 1}, the functions φ00 , φ11 are real-valued, while φ01 , φ10 are complex
conjugate, and
2
φ00 − φ200 = φ11 − φ211 = |φ01 | , (φ00 + φ11 − 1)φ01 = 0. (6)
On the other hand, direct computations immediately reveal that condition quq = 0
is equivalent to
φ200 = φ211 = φ01 φ10 , (φ00 − φ11 )φ01 = (φ00 − φ11 )φ10 = 0. (7)
Solving the system of equations (6)–(7) yields
1 1 1
φ00 = φ01 = χ, φ01 = χω, φ10 = χω
2 2 2
with χ being a characteristic function of some subset of Δ and unimodular ω.
So, elements of Qu ∩ W(e, u) have the form
 
1 χ χω
q =0⊕0⊕0⊕0⊕ (H). (8)
2 χω χ
The rest of the reasoning depends on whether or not the subspaces M00 ,
M11 are actually present in the decomposition (3).
Case 1. At least one of the subspaces M00 , M11 is different from zero, that is, the
range of e contains some eigenvectors of u.
374 I.M. Spitkovsky

Since for any q of the form (8) the restriction of e − q on M00 ⊕ M11 is the
identity, we then have e − q = 1. Consequently, d = 1. Note that in this case
also eue = 1.
Case 2. M00 = M11 = {0}. Since both e given by (4) and q given by (8) have zero
restrictions onto M01 ⊕ M10 , we may without loss of generality suppose that in
place of (3) simply H = M ⊕ M, and respectively
 +   
H H(I − H) 1 χ χω
e= + , q= (H). (9)
H(I − H) I −H 2 χω χ
So, e − q = Φχ,ω (H), where
 + 
t − 12 χ(t) t(1 − t) − 12 χ(t)ω(t)
Φχ,ω (t) = + .
t(1 − t) − 12 χ(t)ω(t) 1 − t − 12 χ(t)
Consequently,
e − q = ess supt∈Δ λχ,ω (t),
where λχ,ω (t) is the positive eigenvalue of Φχ,ω (t), and ess is understood in the
sense of the spectral measure of H.
If χ(t) = 0 for some t ∈ Δ, then the respective λχ,ω (t) equals one, guaran-
teeing e − q = 1. We should concentrate therefore on elements q with χ(t) ≡ 1.
Then we have
 + 
+ t − 12 t(1 − t) − 12 ω(t)
Φ1,ω (t) = ,
t(1 − t) − 12 ω(t) 2 −t
1

and J
1 +
λ1,ω (t) = − t(1 − t) Re ω(t).
2
Since ω is unimodular, to minimize λ1,ω (t) for any given t we should take ω(t) = 1.
The respective element q is simply
 
1 I I
q0 = , (10)
2 I I
, +
λ1,1 (t) = 12 − t(1 − t), and
J 4  J 
1 + 1 2
e − q0  = − min t(1 − t) = 1 − 1 − max |2t − 1| .
2 t∈Δ 2 t∈Δ

In order to justify (2), it remains only to observe that

max |2t − 1| = eue . (11)
t∈Δ

But this is indeed the case, since eue = Φ(H) with the matrix
 + 
t t(1 − 1)
Φ(t) = (2t − 1) + ,
t(1 − 1) 1−t
the eigenvalues of which are zero and 2t − 1.
 A Distance Formula Related to a Family of Projections 375

3. Additional remarks
1. As in the case of W ∗ -algebras, the C ∗ -algebra C(e, u) generated by e and u is
the same as the C ∗ -algebra generated by the orthogonal projections e, (u + I)/2.
Invoking the description from [5] we observe that elements of C(e, u) are those
of the form (5) for which the functions φij are continuous on Δ and such that
φ01 (j) = φ10 (j) = 0, aij = φii (1 − j) if j ∈ Δ (i, j = 0, 1). From (10) we therefore
conclude that the element q0 ∈ W(e, u) on which the distance from e to Qu is
attained does not lie in C(e, u) if the spectrum of H contains 0 or 1.
On the other hand, due to (11) condition eue < 1 guarantees that 0, 1 ∈ / Δ,
and thus the invertibility of the operator H(I − H). Moreover, M00 = M11 = {0},
as was observed in Section 2. So, without loss of generality e is given by the first
formula in (9), while u = diag[I, −I]. From here:
 + 
1 0 H(I − H)
z := (e − ueu∗ ) = + ∈ C(e, ueu∗),
2 H(I − H) 0
z 2 = diag[H(I − H), H(I − H)] is positive definite and also lies in C(e, ueu∗), and
−1/2 −1/2
therefore so does (z 2 )−1/2 = diag[((H(I − H)) , (H(I − H)) ]. Along with
2 −1/2 ∗
z and (z ) , the algebra C(e, ueu ) contains their product
 
0 I
w := z(z 2 )−1/2 = .
I 0
Since w2 = I, the algebra C(e, ueu∗) is unital, and we conclude from (10) that
q0 = 12 (I + w) ∈ C(e, ueu∗).
2. The distances from e to the sets Qu and Qu ∩ W(e, u) may not coincide. To
illustrate, consider H = C3 , u = diag[1, 1, −1] and e = diag[1, 0, 0]. Then Qu
consists of zero and all matrices of the form
⎡ 2 ⎤
|x| xy x
1
qx,y = ⎣ xy |y|2 y ⎦ ,
2
x y 1
2 2
with the parameters x, y ∈ C satisfying |x| +|y| = 1. An easy computation shows
that ,
2
e − qx,y  = (1 + |y| )/2.
√
So, the distance from e to Qu equals 1/ 2 and is attained on all the matrices qω,0
with |ω| = 1, that is, having the form
⎡ ⎤
1 0 ω
1⎣
0 0 0⎦ .
2
ω 0 1
On the other hand, the algebra generated by e and u consists simply of all 3-by-3
diagonal matrices. The only diagonal matrix lying in Qu is 0, and d = 1 in full
agreement with Theorem 2.
376 I.M. Spitkovsky

Note that in this example eue = 1, and so the right-hand side of (2)
coincides with the distance d0 from e to Qu . It would be interesting to figure out
if in fact d0 always equals the right-hand side of (2), thus implying that d = d0
when the intersections of the range of e with H± are trivial.
Acknowledgement
The author is thankful to the anonymous referees for their valuable comments.

References
[1] A. Böttcher and I.M. Spitkovsky, A gentle guide to the basics of two projections theory.
Linear Algebra Appl. 432 (2010), no. 6, 1412–1459.
[2] R. Giles and H. Kummer, A matrix representation of a pair of projections in a Hilbert
space. Canad. Math. Bull. 14 (1971), no. 1, 35–44.
[3] P.L. Halmos, Two subspaces. Trans. Amer. Math. Soc. 144 (1969), 381–389.
[4] I.M. Spitkovsky, Once more on algebras generated by two projections. Linear Algebra
Appl. 208/209 (1994), 377–395.
[5] N. Vasilevsky and I. Spitkovsky, On the algebra generated by two projections. Doklady
Akad. Nauk Ukrain. SSR, Ser. A 8 (1981), 10–13 (in Russian).
[6] S. Walters, Projection operators nearly orthogonal to their symmetries. J. Math. Anal.
Appl. 446 (2017), no. 2, 1356–1361.

Ilya M. Spitkovsky
Division of Science
New York University Abu Dhabi (NYUAD)
Saadiyat Island
P.O. Box 129188
Abu Dhabi, UAE