Bol. Soc. Mat. Mex.

DOI 10.1007/s40590-016-0112-z

INVITED REVIEW

The part of my path I walked together with Sergei

Grudsky

Albrecht Böttcher1

Received: 3 September 2015 / Revised: 28 October 2015 / Accepted: 29 October 2015

© Sociedad Matemática Mexicana 2016

Abstract This is an essay containing personal reminiscences and describing selected

topics of joint mathematical work of the author and Sergei Grudsky. The topics have
their focus on Toeplitz operators and large finite Toeplitz matrices.

Keywords Sergei Grudsky · Toeplitz matrix · Toeplitz operator

Mathematics Subject Classification 01A70 · 15-03 · 47-03 · 47B35 · 65F15

1 First encounter

Our first personal encounter was short and cold: Sergei handed over a few pages
to me and all he said was “perepisat i vernut”, which is Russian and means
“transcribe and return”. As I expected a little more than two mere infinitives, namely at
least two imperatives or even a couple of friendly or explanatory words, this encounter
left an unpleasant flavor on my side.
This was in 1982 or 1983. I then was a PhD student at Rostov-on-Don State Uni-
versity. My supervisor was Vladimir Borisovich Dybin. He was also the supervisor of
Sergei, who defended an outstanding thesis in 1981 and whose name was of course

For Sergei on his 60th birthday with thanks and admiration.

B Albrecht Böttcher
albrecht.boettcher@mathematik.tu-chemnitz.de

1 Fakultät für Mathematik, Technische Universität Chemnitz, 09107 Chemnitz, Germany

 A. Böttcher

known to me. To defend a thesis in Russia, one had to obey strong rules. First, it was
required that the main results of the thesis are published in generally accessible jour-
nals at least three months before the date of the defence. A PhD program lasted three
years, and due to the backlog of the journals, one had to submit the “main results”
at the end of the first year or the mid of the second year at latest. Secondly, one had
to pass three examinations: one in Russian language for foreigners, one in mathemat-
ics, and one in philosophy. The last was the pure horror for almost all PhD students,
including myself, but that is another story. The first was no problem for me. Well, and
my examiner in mathematics was Igor Borisovich Simonenko. He had a list of about
40 concrete topics for the preparation to the exam, Dybin told me that Sergei had this
list, so I turned to Sergei, and this was just the encounter I described above.

2 Our first joint paper

When sitting down and beginning to write this essay, I realized to my astonishment that
by now Sergei and I have about 50 joint papers and two joint books. Our first joint paper
is the 60-pager [10], which appeared in the proceedings of the joint German–Israeli
workshop organized by Israel Gohberg and myself in Tel Aviv and Jerusalem in 1995.
I do not remember how we came together to write that paper. In 1993, Sergei
participated in a conference in Chemnitz. This was our second encounter in 10 years,
and on that occasion we probably decided to combine our experiences in Toeplitz
operators with oscillating symbols to make some things more explicit and to push
others forward. Anyway, what resulted was the talk I gave in Israel and then the
paper [10]. And clearly, this was obviously the time when we became true friends and
started our joint walk, which has now lasted proud 20 years.

The Toeplitz operator T (a) generated by a function a ∈ L ∞ (T), its so-called

symbol, is the operator induced on 2 (N) by the infinite matrix (a j−k )∞
j,k=1 formed
by the Fourier coefficients
 2π
an = a(eiθ )e−inθ dθ/2π
0
The part of my path I walked together with Sergei Grudsky

of a. Every point t ∈ T\{1} may be represented as t = (x − i)/(x + i) with x ∈ R. Let

us denote a((x − i)/(x + i)) by a  (x). Paper [10] is on Toeplitz operators T (a) with
oscillating symbols. Writing a  (x) = ei f (x) with a real-valued continuous function f
on R, the type of the oscillation is governed by the behavior of f (x) as x → ±∞. If
f (−∞) = −∞ and f (+∞) = +∞, one speaks of an orientation preserving whirl,
the case f (−∞) = f (+∞) = +∞ produces an orientation changing whirl, and
periodic but bounded functions f lead to what we call undecided oscillations.

Sergei had a very deep understanding of the behavior of the arguments of infinite
Blaschke products. See, for example, his papers [49,50]. He knew in particular that
if f is monotone with f (−∞) = −∞ and f (+∞) = +∞, then there is an infinite
Blaschke product B such that B  (x) = eiϕ(x) and f − ϕ is bounded on R. Thus, every
monotone growth of the argument of the symbol a from −∞ to +∞ is essentially
modeled by the argument of some Blaschke product. This observation is one of the
keys to the results of [10]. I here confine myself to one highlight of that paper, which
is concerned with the finite section method. For more on Toeplitz operators with
oscillating symbols I refer to the monograph [47] by Vladimir Dybin and Sergei. This
book shows who the actual wizards of all aspects of oscillating symbols are.
One says that the finite section method is applicable to the operator T (a), and writes
T (a) ∈ {Pn } in that case, if the n × n principal truncations Tn (a) = (a j−k )nj,k=1
are invertible for all sufficiently large n and the norms of the inverses Tn−1 (a) :=
[Tn (a)]−1 remain bounded as n → ∞. The implication “T (a) is invertible ⇒ T (a) ∈
{Pn }” had been studied by many mathematicians, including Baxter, Reich, Gohberg,
Feldman, Ambartsumyan, Widom, and Silbermann, and it was shown to be true if a is
in C + H ∞ or if a is in P QC, the C ∗ -algebra of piecewise quasi-continuous functions.
See the books [42,48] for more on this topic. However, all that previous work would
have been for nothing had it turned out that the implication is valid for all a ∈ L ∞ (T).
Fortunately, in 1987, Treil [58] was able to find a function a ∈ L ∞ (T) such that
T (a) is invertible but T (a) ∈ / {Pn }. In [10], we constructed an explicit function
a  (x) = ei f (x) with this property. This function belongs to A P W , the algebra of
almost periodic functions with absolutely convergent Fourier series, and it is given by
 A. Böttcher

 
8 1 1
f (x) = π − cos x + 2 cos 3x + 2 cos 5x + · · · .
π 3 5

3 Condition numbers

In the years between 1997 and 2006, Sergei visited me in Chemnitz annually. Each time
we embarked on a specific problem, “zadaqka” in Russian, sometimes we solved
it still during his visit, but frequently the solution required the whole year until his
subsequent visit.
If the finite section is not applicable to T (a), that is, if T (a) ∈ / {Pn }, then the
(spectral) norm of Tn−1 (a) does not remain bounded as n goes to infinity. Getting an
understanding of the behavior of these norms has been one of our favorite topics of
research since almost the beginning of our joint walk. In [32] we still had the words
“norms of inverses” in the title. But subsequently we made a trick. Since Tn (a)
converges to the L ∞ norm a∞ as n goes to infinity, the behavior of Tn−1 (a) is
essentially the same as that of κ(Tn (a)) := Tn (a)Tn−1 (a), and so we henceforth
went over to speaking of “condition numbers”, which resulted in an increase of the
attention from the linear algebra and numerical analysis communities.
Suppose a ∈ C(T). Then T (a) ∈ {Pn } if and only if T (a) is invertible, which
in turn is the case if and only if a has no zeros on T and winding number 0 about
the origin. In [11] we studied symbols with zeros on T, and our approach was based
on two ideas. To get lower bounds for κ(Tn (a)), we had to know vectors xn such
that Tn−1 (a)xn /xn  is close to Tn−1 (a). We managed this by choosing xn as the
coefficients of the trigonometric polynomial
 j
imθ/2 sin 2 θ
m+1
j iθ
pm (e ) = (1 + e + · · · + e ) = e
iθ imθ j
sin θ2

with cleverly adjusted m and j. In this way we proved, for example, that if a(eiθ ) =
O(|θ − θ0 |α ) as θ → θ0 for some eiθ0 ∈ T, then there exists a constant C > 0 such
that κ(Tn (a)) ≥ Cn α for all n ≥ 1.
Our argument to get upper bounds was more tricky. Let conv a(T) denote the
convex hull of the range a(T). A famous theorem by Brown and Halmos says that if
d := dist(0, conv a(T)) > 0, then Tn (a) is invertible and Tn−1 (a) ≤ 1/d for all
n ≥ 1. Now consider the n × n Toeplitz matrix
⎛ ⎞
2 −1 0 ··· 0
⎜ −1 2 −1 ··· 0⎟
⎜ ⎟
⎜ −1 ··· 0⎟
Tn (b) = ⎜ 0 2 ⎟,
⎜ .. .. .. .. .. ⎟
⎝ . . . . .⎠
0 0 0 ··· 2

which, according to Gilbert Strang, is the mother of all matrix theory. The symbol
is b(eiθ ) = −e−iθ + 2 − eiθ = 2(1 − cos θ ) and hence b(T) = [0, 4]. Thus, the
Brown–Halmos theorem is not applicable. But now replace b(eiθ ) by
The part of my path I walked together with Sergei Grudsky

a(eiθ ) = b(eiθ ) + i cos nθ = b(eiθ ) + i(einθ + e−inθ )/2.

Since only the Fourier coefficients ak with k = −(n − 1), . . . , n − 1 enter Tn (a),
we have Tn (b) = Tn (a). The point is that dn =√dist(0, conv a(T)) is strictly positive.
Some plane geometry shows that dn > (4/3 2) sin2 (π/(6n)), and so Brown and
Halmos give √
−1 −1 1 3 2 1
Tn (b) = Tn (a) ≤ < < 10n 2 .
dn 4 sin (π/(6n))
2

This bound is not bad, since Tn−1 (b) is known to be asymptotically equal to n 2 /π 2 .
This previous argument can be extended to more general situations. We did this in [11]
and obtained the following result in this way. If Re a ≥ 0 on T, if Re a is not identically
zero, and if Re a has only a finite number of zeros on T, then there is a constant D < ∞
such that κ(Tn (a)) ≤ Dn α , where α is the maximal order of the zeros.
Now let a be a Laurent polynomial. Then T (a) is banded. If a has zeros on T,
then, by what was said in the two previous paragraphs, κ(Tn (a)) grows polynomially.
But suppose a has no zeros on T but nonzero winding number. In [12] we proved
that then κ(Tn (a)) increases at least exponentially. We even showed that κ(Tn (a))
may grow arbitrarily fast. Namely, consider the function a(eiθ ) = eiθ + e−iθ /4. Then
a(T) is (the boundary of) an ellipse with the foci at −1 and 1, and given any function
ϕ : N → [0, ∞), for example, the function ϕ(n) = exp(n n ), there exists a point
λ ∈ [−1, 1] such that κ(Tn (a − λ)) < ∞ for all n ≥ 1 but κ(Tn k (a − λ)) > ϕ(n k )
for infinitely many n k .

Together with Anatoli Kozak (also a PhD student of Dybin’s) and Bernd Silbermann
we subsequently studied some more sophisticated questions. It was known from [7]
that if a ∈ C(T) and T (a) is invertible, then

lim Tn−1 (a) = T −1 (a). (1)

n→∞

In [25], we determined the speed of convergence in (1). The inverses Tn−1 (a) have a
specific structure, which in [26]
made
 us to investigate the speed with which the norms
of sum-products of the form j k Tn (a jk ) (a jk ∈ C(T)) converge to the norm of
 A. Böttcher

 
j kT (a jk ). Matrices of the form j k Tn (a jk ) are special so-called Toeplitz-
like matrices. Sergei and I continued working on condition numbers and in [13], the
proceedings of a conference in Boulder, Colorado, in 1999, we were finally able to
make precise the following interesting insights.
(a) Within the class of Toeplitz-like matrices, fast convergence of the norms, of the
norms of the inverses, and of the condition numbers is generic.
(b) Within the class of pure Toeplitz matrices, norms converge generically slowly,
norms of inverses converge generically quickly, and condition numbers converge
generically slowly.
It was Rajendra Bhatia who invited me to hold a mini-course at the Indian Statistical
Institute in New Delhi in January 2000. Therefore, Sergei and I prepared lecture notes
for this course during Sergei’s visit to Chemnitz in the late fall of 1999. Sergei was
always accompanied by his wife Olga, who in a virtuoso manner managed all the
tasks connected with his correspondence and with editing and typing his papers. It
took Sergei and me two weeks to write down the lecture notes by hand, and then, it
sounds unbelievable, Olga produced the LaTeX masters of a 100-pager within two
days! In the end and thanks Olga, in January 2000, the listeners of the mini-course
had the booklet [14] in their hands.

4 Ilya Spitkovsky and spectra

In the late 1990s, we had a very fruitful collaboration with Ilya Spitkovsky. If a is in
C(T), then the spectrum of T (a) is the union of the range a(T) and of all points in
the plane which are encircled by a(T) with nonzero winding number. Clearly, small
perturbations of a in the L ∞ norm lead to only small perturbations of the spectrum
σ (T (a)) in the Hausdorff metric. One of the open questions in those days was whether
this is also the case for arbitrary a ∈ L ∞ (T). The answer was in the affirmative for a
in C + H ∞ , in P QC, and in A P (the almost periodic functions). In our paper [33]
with Ilya, we showed that the answer is negative for symbols a in S A P, Sarason’s
C ∗ -algebra of semi-almost periodic functions.
The part of my path I walked together with Sergei Grudsky

A similar question can be (and was) asked for the limiting set

(a) := lim sup σ (Tn (a));

that is, for the set of all λ such that λ = limk→∞ λn k for a suitable subsequence of
eigenvalues λn k ∈ σ (Tn k (a)). In [16], Sergei and I showed that the answer is negative
even for continuous symbols: there exist symbols an and a in C(T) such that an → a
in the L ∞ norm but (an ) does not converge to (a) in the Hausdorff metric.
Another question in this vein is whether Fredholmness of Toeplitz operators is stable
under frequency modulation (it was known to be stable under amplitude modulation).
Together with Ilya, we showed in [36] that it is not: there exist a ∈ S A P and orientation
preserving homeomorphisms ϕ of T\{1} onto itself such T (a) is Fredholm but T (a◦ϕ)
is not Fredholm. Due to the results of papers [33,36] and even more counterexamples
one can construct using the class S A P of semi-almost periodic functions (see, e.g.,
Chapter 4 of [39]), in the years after 2000, we gave several talks at colloquia and
conferences with titles like “All evil begins with S A P”.
Finally, Ilya’s great love are matrix functions. In the papers [34,35], we fully acted
out this love with him. In the first of these papers we proved that if we are given n ≥ 2
and arbitrary vectors μ, ν ∈ Zn , then there exists a bounded n × n matrix function
that has a right Wiener–Hopf factorization in L 2 with the partial indices μ and a
left Wiener–Hopf factorization in L 2 with the partial indices ν. In the second paper
we embarked on similar questions for matrix functions with a prescribed number of
discontinuities.

5 Pseudospectra

In 2002, Sergei moved to Mexico and was appointed professor at the CINVESTAV
del I.P.N. With Sergei Grudsky and Vladimir Rabinovich from Rostov-on-Don and
Nikolai Vasilevski and Yuri Karlovich from Odessa, Mexico received a strong impetus
in operator theory and became a worldwide respected address in this field.
 A. Böttcher

Part of my and Sergei’s work in the years from 2000 to 2008 was inspired by Nick
Trefethen and Mark Embree. They taught us that in the realm of non-normal operators
pseudospectra are telling us much more than spectra. For ε > 0, the ε-pseudospectrum
of an operator or a matrix A is defined by

σε (A) = σ (A) ∪ {λ ∈ C\σ (A) : (A − λI )−1  > 1/ε}. (2)

As n goes to infinity, the spectra σ (Tn (a)) converge to the spectrum σ (T (a)) in rare
cases only. However, as shown in [7,55], the pseudospectra σε (Tn (a)) converge to
the pseudospectrum σε (T (a)) in many cases, at least if a is in PC, the C ∗ -algebra of
piecewise continuous functions on T. Looking at (2) and replacing a by a −λ in (1), we
see that convergence of pseudospecra is, in a sense, nothing but convergence of norms
of inverses. However, in 1993 or so, when working this out to precise statements,
I was led to the question whether pseudospectra may jump, or equivalently, whether
the level sets

{λ ∈ C\σ (A) : (A − λI )−1  = 1/ε}

may have interior points. I posed this question at a Banach Semester in Warsaw in
1994, and the young Polish mathematician Andrzeij Daniluk then immediately proved
that in Hilbert space the level sets cannot have interior points. In our paper [32], Sergei,
Bernd Silbermann, and I showed that the same is true in L p spaces for 1 < p < ∞. But
the case of general Banach spaces had remained open over 15 years until about 2008,
when Shargorodsky [57] solved the problem completely (by constructing a Banach
space in which the level sets may have interior points).

In control theory, it is useful to replace pseudospectra by structured pseudospectra

(also called spectral value sets). These are defined by

σεB,C (A) = σ (A) ∪ {λ ∈ C\σ (A) : C(A − λI )−1 B > 1/ε},

The part of my path I walked together with Sergei Grudsky

where C and B are given fixed operators. In [15], we addressed the question whether
level sets of structured pseudospectra can have interior points. Under the assumption
that

q
a(eiθ ) = ak eikθ , p ≥ 0, q ≥ 0, a p = 0, a−q = 0
k=− p

and that C = Pm and B = Pk are the orthogonal projections onto the first m and
k coordinates, respectively, we were led to fascinating mathematics, namely, to the
monodromy group of the Riemann surface associated with the Laurent polynomial a.
One result of [15] says that the sets σεPm ,Pk (T (a)) cannot jump in each of the following
cases: (a) C\(a) is connected, (b) m = k = 1, (c) T (a) is Hessenberg, (d) p + q
is a prime number and p or q equals 2, (e) p + q ≤ 5 or p + q = 7, (f) m = k and
p + q = 6. Looks strange, doesn’t it?
As stated above, σε (Tn (a)) → σε (T (a)) for each ε > 0 provided a is piecewise
continuous. In [9], it was shown that the convergence may be dramatically slow, which
is connected with the fact that (Tn (a) − λI )−1  only grows slowly. Half of the proof
was based on the observation that, by Cramer’s rule, the (n, 1) entry of (Tn (a)−λI )−1
is the quotient of two Toeplitz determinants with so-called Fisher–Hartwig symbols.
Sergei surmised that the slow convergence is not exclusively caused by jumps of the
symbol and may also happen for continuous symbols. This was subsequently proved
in our paper [17].
A vector xn ∈ Cn is called an ε-pseudomode for the matrix Tn (a) at a point
λ ∈ σε (Tn (a)) if (Tn (a) − λI )xn  < εxn . A sequence {xn } of vectors xn ∈ Cn is
referred to as an asymptotic pseudomode for T (a) at a point λ ∈ C if

(Tn (a) − λI )xn /xn  → 0

as n → ∞. In [19,37] we showed that these notions hold more surprises than

one could expect. Here is a result from our paper [37] with Jérémie Unterberger:
if a ∈ L ∞ (0, 2π ) is continuous in an open neighborhood of π , then the sequence {xn }
given by
  n
2 jπ
xn = (−1) j+1
f
n+1 n+1
j=1

is an asymptotic pseudomode for T (a) at λ = a(π ) for every Riemann integrable

function f on (0, π ).

6 Olga Grudskaya

As already said, Sergei was always accompanied by his wife Olga, also a PhD student
of Dybin’s. She was an extraordinarily pleasant person, a wonderful friend, and an
irreplaceable colleague. When visiting Chemnitz, Olga and Sergei usually had a small
apartment in the guest house of the university. A highlight of every visit was the
invitation to the Russian dinner Olga prepared once during each visit. She cooked
 A. Böttcher

truly masterly, and the dinners we had with Olga and Sergei in their apartment are
unforgettable in the minds of my wife Sylvia and myself.
Olga also was the private secretary of Sergei. She took care of all tasks that had to
be done with the computer, from writing emails up to creating Matlab codes and per-
forming numerical experiments. A young colleague and friend of Sergei once said that
Sergei is a prominent exception from the rule according to which good mathematicians
always have great gift for languages. Sergei is a brilliant mathematician, but as soon
as things came to writing something in English, he was depending on Olga (or me).

Tragically, Olga died in a car accident in 2004. Her early death left great emptiness
in all of us. In late 2003, she began working on the illustrations for our book [21]
with great enthusiasm. She could not accomplish her visions and we were left with
including some of her drafts. We invite the reader to browse in the book to get an idea
of the beauty that would have emerged if Olga had been able to complete her work.

7 Probability arguments
Another topic of our joint research arose from a talk by Siegfried Rump at a conference
in Marrakesh in 2001. The condition number κ(Tn (a), x) associated with Tn (a) and
a vector x with n components is of importance in connection with pertubations of
Tn (a) by arbitrary matrices of small norms. However, since Tn (a) is a structured
matrix, it makes sense to restrict its perturbations to perturbations that respect the
structure, that is, to perturbations that are themselves Toeplitz matrices. This leads
to the notion of the structured condition number κ Toep (Tn (a), x). In many cases of
practical relevance, the symbol a is a rational function and hence, as outlined in Sect.
3, the numbers κ(Tn (a), x) may grow exponentially. Rump showed that the structured
condition numbers κ Toep (Tn (a), x) may grow essentially slower. As this could call in
question the relevance of our studies of condition numbers, Sergei and I embarked on
the topic and this led to our papers [18,20]. Incidentally, the results of [18] have found
nice applications in the papers [40,41,43] by Daniel Potts, David Wenzel, and myself.
The part of my path I walked together with Sergei Grudsky

In [18], we first considered arbitrary real or complex n×n matrices An and computed
the expected value and the variance of the random variable An x2 /An 2 where x is
drawn from the uniform distribution on the unit sphere of Rn or Cn . We then applied
the result to Toeplitz matrices and showed that, for large n, the values of the quotient
Tn (a)x2 /Tn (a)2 cluster fairly sharply around a2 /a∞ if a ∈ L ∞ (T) and
around zero in case a ∈ / L ∞ (T); here a2 and a∞ stand for the norms of a in
L 2 (T) and L ∞ (T), respectively. In [20], we proved that if the components of x are
independent standard normal or independent Rademacher variables, then there are
universal positive constants ε and n 0 such that
 
κ Toep (Tn (a), x) ε 99
Probability ≥ 3/2 >
κ(Tn (a), x) n 100

for all rational symbols a without poles on T and all n ≥ n 0 . This reveals that struc-
tured condition numbers do with high probability also increase exponentially, which
supports the claim that one is not likely to win much on the average Toeplitz input
by passing from the usual condition numbers of Toeplitz matrices to their structured
counterparts.
 A. Böttcher

8 The Fox–Li operator

In 2010, Arieh Iserles of Cambridge acquainted us with another fascinating research

topic: the spectrum of the Fox–Li operator. This operator acts on L 2 (0, τ ) by the rule
 τ
1 2
(Fτ u)(x) = √ ei(x−y) u(y) dy, x ∈ (0, τ ),
πi 0

and hence it is a truncated Wiener–Hopf operator, the continuous analog of a finite

Toeplitz matrix. The physics background says that τ is a large number. The symbol
of such an operator is the Fourier transform of its kernel. For the Fox–Li operator, the
symbol is
 ∞ 2
ei x
√ eiξ x dx = e−iξ /4 , ξ ∈ R,
2
f (ξ ) =
−∞ πi

and we write Wτ ( f ) instead of Fτ . Notice that f is an orientation changing whirl in

the sense of Sect. 2.
The operator Wτ ( f ) is in the trace class and its eigenvalues are conjectured to lie
along a spiral commencing at 1 and rotating clockwise to the origin. Vainshtein [60]
even raised a conjecture on the shape of the spiral: its parametric representation is
z(t) = exp(−α(τ )t ν − iβ(τ )t ν ), t ∈ (0, ∞), with

ζ (1/2)π 3/2 π2
ν = 2, α(τ ) ≈ √ , β(τ ) ≈ 2 ,
8 2τ 3 4τ

and the nth eigenvalue is approximately given by t = n. So far these conjectures are
still open. It is certainly interesting to mention that it was the search for the spectrum
of the Fox–Li operator which led Henry Landau to the invention of the notion of the
pseudospectrum [53]. Considering the operator Wτ ( f )Wτ∗ ( f ), one gets an operator
whose eigenvalues were studied by Landau and Widom in [54]. Their result yields
information about the singular values of the Fox–Li operators [8], but in the case at
hand, the singular values are nothing but poor shadows of the eigenvalues.
The joint attempts of Arieh Iserles, Sergei, and myself try to prove the conjecture
on the spiral failed, but we could nevertheless push things an epsilon forward. With
the help of Daan Huybrechs, we were able to find the first-order asymptotics of the
traces of the powers of the Fox–Li operator. In [22] we proved that

τ
tr [Wτ ( f )]k = √ + o(τ ) as τ → ∞
πik

for each fixed natural number k. Such trace formulas can be used to test conjectures
on asymptotic eigenvalue distribution. In the concrete case of the Fox–Li operator, we
could so prove the following. Assume the eigenvalues of Wτ ( f ) are indeed samples
of the function exp(−α(τ )t ν − iβ(τ )t ν ) at t = n, where ν is a natural number and
α(τ ), β(τ ) are positive real numbers. Then necessarily
The part of my path I walked together with Sergei Grudsky

   
1 π2 1
ν = 2, α(τ ) = o , β(τ ) = 2 + o .
τ 4τ τ2

For more on the topic see the papers [8,22,24].

If physicists cannot solve a problem, they are allowed to abstain from mathematical
rigor and to follow solely their intuition, but they have to stay with the same problem. If
mathematicians cannot solve a problem, they must not leave aside rigor, but they have
the right to replace it by another problem. The Fox–Li operator is a truncated Wiener–
Hopf operator with a complex-symmetric kernel, but its symbol is strongly oscillating.
In [23], we considered truncated Wiener–Hopf operators Wτ (a) with complex-
symmetric kernels under the assumption that the symbol a is a rational function.
We there constructed a certain function b : (0, ∞) → C with the following property:
there is a numbering {λn }∞ n=1 of the eigenvalues of Wτ (a) such that, with ξn := nπ/τ ,
 
1  i  1
λn = a(ξn ) + a (ξn ) arg b(ξn ) − a (ξn ) log |b(ξn )| + O .
2τ 2τ τ2

The approach of [23] is based on the explicit formulas for Fredholm determinants
of Wiener–Hopf operators with rational symbols that were established in my early
paper [6] on the one hand, and on an ingenious fixed-point argument introduced by
Sergei on the other. Incidentally, the same problem in the discrete case, that is, for
complex-symmetric Toeplitz matrices with rational symbols, is more complicated than
its continuous counterpart. It was solved by Sergei in collaboration with Alexander
Batalshchikov and Vladimir Stukopin only recently in [1].

9 With Manuel and Egor in the crowd of the eigenvalues

On the most recent part of our joint walk, Sergei and I are accompanied by two highly
talented young men: Egor Maximenko of Mexico City and Manuel Bogoya of Bogotá.
Egor is a PhD student of Simonenko’s and I know him personally since he, still being a
student in Rostov-on-Don, visited Chemnitz in 2000. Manuel received his PhD under
the supervision of Sergei, and I am acquainted with him since his 3-month visit to
Chemnitz in 2010.
Theorems of the Szegő type say that, under certain conditions on a and F, including
that a be real-valued,
 2π
1
n
1
lim F(λ j (Tn (a))) = F(a(eiθ )) dθ, (3)
n→∞ n 2π 0
j=1

where λ1 (Tn (a)) ≤ · · · ≤ λn (Tn (a)) are the eigenvalues of the Hermitian Toeplitz
matrix Tn (a), while theorems of the Avram–Parter type state that, again under appro-
priate assumptions on a and F,
 2π
1
n
1
lim F(s j (Tn (a))) = F(|a(eiθ )|) dθ, (4)
n→∞ n 2π 0
j=1
 A. Böttcher

where s1 (Tn (a)) ≤ · · · ≤ sn (Tn (a)) are the singular values of Tn (a)). The functions F
are called test functions. It is assumed that F is real-valued and continuous on R when
considering (3) and real-valued and continuous on [0, ∞) when dealing with (4). In
1971, B. L. Golinskii and I. A. Ibragimov proved that the strong Szegő limit theorem,
which tells us that
 ∞
det Tn (a)
lim = exp k(log a)k (log a)−k
n→∞ exp(n(log a)0 )
k=1

is true for positive a whenever it makes sense. Serra Capizzano [56] described large
classes of test functions F for which (3) and (4) are valid, and in our paper [31] with
Moshe Schwartz, we asked whether (3) and (4) are also true whenever they make sense.
To our big surprise, the answer was negative. In our paper [31] with Moshe, we con-
structed a counterexample with a highly oscillating test function F and then together
with Egor a counterexample with a monotone test function in [27,28]. In the other
direction, we showed that (3) and (4) are true for all essentially convex test functions.
Formula (3) concerns the collective asymptotic behavior of the eigenvalues. Much
attention has also been paid to the asymptotics of the extreme eigenvalues, that is,
to λ j (Tn (a)) as n → ∞ but j or n − j remains fixed. Significantly less was known
about individual inner eigenvalues. We started exploring this terrain together with Egor.
Tackling the case of general real-valued symbols is still far away from what we are able
to do. To get a feeling for what happens and to find the right language for the problem,
we considered smooth so-called simple-loop symbols. These are smooth real-valued
functions a which increase strictly from their minimum to their maximum, and then
decrease strictly back from the maximum to the minimum, having nonzero second
derivatives at the minimum and the maximum. Under this assumption, we established
asymptotic formulas for λ j (Tn (a)) as n → ∞ and j/n → x ∈ [0, 1]. In [29,30], we
still required that a be a Laurent polynomial. Deift, Its, and Krasovsky [46] proved
analogous results for C ∞ symbols. The restriction to C ∞ symbols in [46] was for
simplicity only, and at the price of increasing the technical effort, the method employed
there would also work under significantly weaker smoothness assumptions. Together
with Manuel and Egor, we developed another approach to the problem and could so
derive the results for symbols with reasonably weak smoothness in [4].
In [2,3], we considered individual eigenvalues and eigenvectors of Hessenberg
Toeplitz matrices whose symbol has a power singularity. To be more precise, the
symbol is of the form

a(eiθ ) = e−iθ (1 − eiθ )α h(eiθ ),

where α ∈ [0, ∞)\Z and h is a C ∞ function in the Hardy class H ∞ . This class
of symbols and the class introduced in paper [45] by Dai, Geary, and Kadanoff are
overlapping, and our results gave in particular a rigorous justification for an asymptotic
formula that was conjectured in [45] on the basis of numerical experiments.
I want to finish this essay with one of the most recent adventures with Sergei,
Manuel, and Egor. We consider (4) with compactly supported test functions F. If |a|
is Riemann integrable, then so also is F ◦ |a|, and hence (4) implies that
The part of my path I walked together with Sergei Grudsky

⎛ ⎞
1 ⎝ 
n n
(n) ⎠
lim F(s j (Tn (a))) − F(|a(t j )|) = 0, (5)
n→∞ n
j=1 j=1

(n)
where, for example, t j = exp(2πi j/(n + 1)). A result stronger than (5) would be
that
1   
n
(n) 
lim F(s j (Tn (a)) − F(|a(t j )|) = 0. (6)
n→∞ n
j=1

This is in general not true as stated, because now the ordering of the singular values
and the sample values comes into play. We assume that

s1 (Tn (a)) ≤ · · · ≤ sn (Tn (a))

and we let σ denote any permutation of the set {1, . . . , n} such that

|a(tσ(n) (n)
(1) )| ≤ · · · ≤ |a(tσ (n) )|.

It turns out that

1   
n
(n) 
lim  F(s j (Tn (a)) − F(|a(tσ ( j) )|) = 0 (7)
n→∞ n
j=1

is indeed true. This was shown by Bill Trench in a cycle of papers culminating with [59].
Recall that a∞ is the L ∞ norm of a. The components of the vectors

Sn := (s1 (Tn (a)), . . . , sn (Tn (a))), An := (|a(tσ(n) (n)

(1) )|, . . . , |a(tσ (n) )|)

lie in [0, a∞ ], and taking F so that F(s) = s on this segment, we infer from (7) that
Sn − An 1 /n → 0 and hence Sn − An  p /n 1/ p → 0 whenever 1 ≤ p < ∞, where
 ·  p stands for the  p norm. The question is whether this is also true for p = ∞,
that is, whether Sn − An ∞ → 0. The answer was given in our paper [5]: if |a| is
Riemann integrable and the essential range R(|a|) of |a| is connected and contains
the origin, then Sn − An ∞ → 0, but if R(|a|) is disconnected or if 0 ∈ / R(|a|), then
Sn − An ∞ need not converge to zero.
To ensure (7) we had to order the values of |a| in the right way. The tool which
performs this task automatically for Riemann integrable functions and which perfectly
works even beyond Riemann integrability is the quantile function of |a|. In probabilistic
language, |a| is a real-valued random variable on the unit circle T with normalized
Lebesgue measure μ being the probability measure. Let F|a| : R → [0, 1] be the
cumulative distribution function of |a|, that is, F|a| (v) := μ{t ∈ T : |a(t)| ≤ v}.
Then the corresponding quantile function Q |a| : (0, 1] → R is defined by Q |a| (u) :=
inf{v ∈ R : u ≤ F|a| (v)}. Herewith another main result of [5]: if a ∈ L ∞ (T) and
R(|a|) is connected and contains the origin, then
 A. Böttcher

  
 j 
lim max s j (Tn (a)) − Q |a|  = 0.
n→∞ 1≤ j≤n n+1 

This result shows that the quantile function is more than a tool for reordering the
values. The quantile function is a concept from abstract measure theory, working with
pre-images of sets of values, whereas reordering of values is based on partitions of
the domain of the function. In a sense, the quantile function replaces reordering in the
same way as the Lebesgue integral relieves the Riemann integral.

10 Zum Schluss

I want to give Egor and Manuel the final words. Egor once characterized Sergei as
follows. “Sergei is one of my best teachers and the most important of my benefactors.
The year of my post-doc in Mexico was the most productive year in my life. Sergei
likes to work on problems that don’t have nice and easy solutions and therefore are
avoided by other mathematicians in the area. During my first years of collaboration
with Sergei, I imagined him as a berserker who attacks these terrible problems in
the fighting madness, rejecting fear and common sense. Now I understand that from
Sergei’s height above ground, from his knowledge and capacities, these problems
don’t look to him as terrible as to me.” Yes, Egor loves strong words.
Manuel is a passionate swimmer. He wrote me the following. “My favorite swimmer
is Ian Thorpe, he put 21 world records, one of them being the 400 mt freestyle 3:40.08
result. That’s amazing to me, and I have seen the video hundreds of times. I cannot
believe how he can swim so nicely, so efficiently, so smoothly, and so fast. And exactly
that is the feeling I have for the work of Sergei.” Egor, Manuel, Sergei, and I are four
very different characters with very different skills. I think it is the lucky combination
of our different peculiarities which makes us a really good team.

There it is. I hope I succeeded in conveying at least an idea of the adventures Sergei
and I have experienced on our joint walk and an idea of the fascinating topics we
The part of my path I walked together with Sergei Grudsky

struggled with and of the many people we encountered and who accompanied us along
several stages of our way. I did not have the space to discuss the lots of interesting
results Sergei established with other authors. See his recent papers [44,51,52], for
example.
Well, the list of references might make me look like a self-citation maniac, but it
should be viewed a list of publications by Sergei contracted to items co-authored and
commented upon by myself. Note that I even completely omitted about 15 further
papers with him, in particular our paper On the composition of Muckenhoupt weights
and inner functions, J. London Math. Soc. (2) 58, 172–184 (1998), where we showed
that if w ∈ A p and u is an inner function, then w ◦ u is again in A p for p = 2 but w ◦ u
does not necessarily belong to A p for p = 2. This answered a question that then had
been open for a long time. I also left off the paper European double-barrier options with
a compound Poisson component, in: Progress in Economics Research, Vol. 18, 315–
331, Nova Science Publishers, Huntington, NY, 2011 by Sergei, R. Michael Porter,
and myself, which explores a Black–Scholes system that is perturbed by a Toeplitz
matrix.
Anyway, I want to draw the reader’s attention to the fact that the list of references
contains exactly 60 items, which perfectly fits with Sergei’s 60th birthday. I consider
it as a great luck that I made the acquaintance of Sergei, I am grateful to him for all
the common years, and I wish him all the best for the future.

References
1. Batalshchikov, A.A., Grudsky, S., Stukopin, V.A.: Asymptotics of eigenvalues of symmetric Toeplitz
band matrices. Linear Algebra Appl. 469, 464–486 (2015)
2. Bogoya, J.M., Böttcher, A., Grudsky, S.: Asymptotics of individual eigenvalues of a class of large
Hessenberg Toeplitz matrices. Oper. Theory Adv. Appl. 220, 77–95 (2012)
3. Bogoya, J.M., Böttcher, A., Grudsky, S., Maksimenko, E.A.: Eigenvectors of Hessenberg Toeplitz
matrices and a problem by Dai, Geary, and Kadanoff. Linear Algebra Appl. 436, 3480–3492 (2012)
4. Bogoya, J.M., Böttcher, A., Grudsky, S., Maximenko, E.A.: Eigenvalues of Hermitian Toeplitz matrices
with smooth simple-loop symbols. J. Math. Anal. Appl. 422, 1308–1334 (2015)
5. Bogoya, J.M., Böttcher, A., Grudsky, S., Maximenko, E.A.: Maximum norm versions of the Szegő and
Avram–Parter theorems for Toeplitz matrices. J. Approx. Theory 196, 79–100 (2015)
6. Böttcher, A.: Wiener–Hopf determinants with rational symbols. Math. Nachr. 144, 39–64 (1989)
7. Böttcher, A.: Pseudospectra and singular values of large convolution operators. J. Integral Equ. Appl.
6, 267–301 (1994)
8. Böttcher, A., Brunner, H., Iserles, A., Nørsett, S.P.: On the singular values and eigenvalues of the
Fox–Li and related operators. N. Y. J. Math. 16, 539–561 (2010)
9. Böttcher, A., Embree, M., Trefethen, L.N.: Piecewise continuous Toeplitz matrices and operators: slow
approach to infinity. SIAM J. Matrix Anal. Appl. 24, 484–489 (2002)
10. Böttcher, A., Grudsky, S.: Toeplitz operators with discontinuous symbols: phenomena beyond piece-
wise discontinuity. Oper. Theory Adv. Appl. 90, 55–118 (1996)
11. Böttcher, A., Grudsky, S.: On the condition numbers of large semi-definite Toeplitz matrices. Linear
Algebra Appl. 279, 285–301 (1998)
12. Böttcher, A., Grudsky, S.: Toeplitz band matrices with exponentially growing condition numbers.
Electron. J. Linear Algebra 5, 104–125 (1999)
13. Böttcher, A., Grudsky, S.: Condition numbers of large Toeplitz-like matrices. Contemp. Math. 280,
273–299 (2001)
14. Böttcher, A., Grudsky, S.: Toeplitz Matrices, Asymptotic Linear Algebra, and Functional Analysis.
Hindustan Book Agency, New Delhi (2000); reprinted by Birkhäuser Verlag, Basel (2000)
 A. Böttcher

15. Böttcher, A., Grudsky, S.: Can spectral value sets of Toeplitz band matrices jump? Linear Algebra
Appl. 351/352, 99–116 (2002)
16. Böttcher, A., Grudsky, S.: Asymptotic spectra of dense Toeplitz matrices are unstable. Numer. Algo-
rithms 33, 105–112 (2003)
17. Böttcher, A., Grudsky, S.: Toeplitz matrices with slowly growing pseudospectra. Factorization, Singular
Operators and Related Problems (Funchal, 2002), pp. 43–54. Kluwer Acad. Publ, Dordrecht (2003)
18. Böttcher, A., Grudsky, S.: The norm of the product of a large matrix and a random vector. Electron. J.
Probab. 8 paper no. 7 (2003)
19. Böttcher, A., Grudsky, S.: Asymptotically good pseudomodes for Toeplitz matrices and Wiener–Hopf
operators. Oper. Theory Adv. Appl. 147, 175–188 (2004)
20. Böttcher, A., Grudsky, S.: Structured condition numbers of large Toeplitz matrices are rarely better
than usual condition numbers. Numer. Linear Algebra Appl. 12, 95–102 (2005)
21. Böttcher, A., Grudsky, S.: Spectral Properties of Banded Toeplitz Matrices. SIAM, Philadelphia (2005)
22. Böttcher, A., Grudsky, S., Huybrechs, D., Iserles, A.: First-order trace formulae for the iterates of the
Fox–Li operator. Oper. Theory Adv. Appl. 218, 207–224 (2012)
23. Böttcher, A., Grudsky, S., Iserles, A.: Spectral theory of large Wiener–Hopf operators with complex-
symmetric kernels and rational symbols. Math. Proc. Camb. Philos. Soc. 151, 161–191 (2011)
24. Böttcher, A., Grudsky, S., Iserles, A.: The Fox-Li operator as a test and a spur for Wiener-Hopf theory.
In: Pardalos, P.M., Rassias, T.M. (eds.) Essays in Mathematics and Its Applications in Honor of Stephen
Smale’s 80th Birthday, pp. 37–48. Springer-Verlag, Heidelberg (2012)
25. Böttcher, A., Grudsky, S., Kozak, A., Silbermann, B.: Convergence speed estimates for the norms of
the inverses of large truncated Toeplitz matrices. Calcolo 36, 103–122 (1999)
26. Böttcher, A., Grudsky, S., Kozak, A., Silbermann, B.: Norms of large Toeplitz band matrices. SIAM
J. Matrix Anal. Appl. 21, 547–561 (1999)
27. Böttcher, A., Grudsky, S., Maksimenko, E.A.: Pushing the envelope of the test functions in the Szegő
and Avram–Parter theorems. Linear Algebra Appl. 429, 346–366 (2008)
28. Böttcher, A., Grudsky, S., Maksimenko, E.A.: The Szegő and Avram–Parter theorems for general test
functions. C. R. Math. Acad. Sci. Paris 346, 749–752 (2008)
29. Böttcher, A., Grudsky, S., Maksimenko, E.A.: On the asymptotics of all the eigenvalues of Hermitian
Toeplitz band matrices. Dokl. Akad. Nauk 428, 153–156 (2009) (Russian); Engl. translation in Dokl.
Math. 80, 662–664 (2009)
30. Böttcher, A., Grudsky, S., Maksimenko, E.A.: Inside the eigenvalues of certain Hermitian Toeplitz
band matrices. J. Comput. Appl. Math. 233, 2245–2264 (2010)
31. Böttcher, A., Grudsky, S., Schwartz, M.: Some problems concerning the test functions in the Szegő
and Avram–Parter theorems. Oper. Theory Adv. Appl. 187, 81–93 (2009)
32. Böttcher, A., Grudsky, S., Silbermann, B.: Norms of inverses, spectra, and pseudospectra of large
truncated Wiener–Hopf operators and Toeplitz matrices. N. Y. J. Math. 3, 1–31 (1997)
33. Böttcher, A., Grudsky, S., Spitkovsky, I.: The spectrum is discontinuous on the manifold of Toeplitz
operators. Arch. Math. (Basel) 75, 46–52 (2000)
34. Böttcher, A., Grudsky, S., Spitkovsky, I.: Matrix functions with arbitrarily prescribed left and right
partial indices. Integr. Equ. Oper. Theory 36, 71–91 (2000)
35. Böttcher, A., Grudsky, S., Spitkovsky, I.: On the Fredholm indices of associated systems of Wiener–
Hopf equations. J. Integral Equ. Appl. 12, 1–29 (2000)
36. Böttcher, A., Grudsky, S., Spitkovsky, I.: Toeplitz operators with frequency modulated semi-almost
periodic symbols. J. Fourier Anal. Appl. 7, 523–535 (2001)
37. Böttcher, A., Grudsky, S., Unterberger, J.: Asymptotic pseudomodes of Toeplitz matrices. Oper. Matri-
ces 2, 525–541 (2008)
38. Böttcher, A., Hurák, Z., Šebek, M.: Minimum distance to the range of a banded lower triangular
Toeplitz operator in 1 and application in 1 -optimal control. SIAM J. Control Optim. 45, 107–122
(2006)
39. Böttcher, A., Karlovich, Yu.I., Spitkovsky, I.M.: Convolution Operators and Factorization of Almost
Periodic Matrix Functions. Birkhäuser Verlag, Basel (2002)
40. Böttcher, A., Potts, D.: Probability against condition number and sampling of multivariate trigonometric
random polynomials. Electron. Trans. Numer. Anal. 26, 178–189 (2007)
41. Böttcher, A., Potts, D., Wenzel, D.: A probability argument in favor of ignoring small singular values.
Oper. Matrices 1, 31–43 (2007)
The part of my path I walked together with Sergei Grudsky

42. Böttcher, A., Silbermann, B.: Analysis of Toeplitz Operators, 2nd edn 2006. Springer-Verlag, Berlin
(1990)
43. Böttcher, A., Wenzel, D.: How big can the commutator of two matrices be and how big is it typically?
Linear Algebra Appl. 403, 216–228 (2005)
44. Carrada-Herrera, R., Grudsky, S., Palomino-Jiménez, C., Porter, R.M.: Asymptotics of European
double-barrier option with compound Poisson component. Commun. Math. Anal. 14, 40–66 (2013)
45. Dai, H., Geary, Z., Kadanoff, L.P.: Asymptotics of eigenvalues and eigenvectors of Toeplitz matrices.
J. Stat. Mech.-Theory Exp. 2009(5), 1–25, Art ID P05012 (2009)
46. Deift, P., Its, A., Krasovsky, I.: Eigenvalues of Toeplitz matrices in the bulk of the spectrum. Bull. Inst.
Math. Acad. Sin. (N.S.) 7, 437–461 (2012)
47. Dybin, V.B., Grudsky, S.: Introduction to the Theory of Toeplitz Operators with Infinite Index.
Birkhäuser Verlag, Basel (2002)
48. Gohberg, I., Feldman, I.A.: Convolution Equations and Projection Methods for Their Solution. Amer.
Math. Soc. Transl. Math. Monographs, vol. 41. Providence (1974)
49. Grudsky, S.: Singular integral equations and the Riemann boundary value problem with infinite index
in the space L p (, ). Izv. Akad. Nauk SSSR 49, 55–80 (1985) (Russian)
50. Grudsky, S.: Singular integral operators with infinite index and Blaschke products. Math. Nachr. 129,
313–331 (1986) (Russian)
51. Grudsky, S., Rybkin, A.: Soliton theory and Hankel operators. SIAM J. Math. Anal. 47, 2283–2323
(2015)
52. Grudsky, S., Shargorodsky, E.: Applications of Blaschke products to the spectral theory of Toeplitz
operators. Fields Inst. Commun. 65, 1–30 (2013)
53. Landau, H.J.: The notion of approximate eigenvalues applied to an integral equation of laser theory.
Q. Appl. Math. 35, 165–172 (1977/78)
54. Landau, H.J., Widom, H.: Eigenvalue distribution of time and frequency limiting. J. Math. Anal. Appl.
77, 469–481 (1980)
55. Reichel, L., Trefethen, L.N.: Eigenvalues and pseudo-eigenvalues of Toeplitz matrices. Linear Algebra
Appl. 162/164, 153–185 (1992)
56. Serra Capizzano, S.: Test functions, growth conditions and Toeplitz matrices. Rend. Circ. Mat. Palermo,
Ser. II 68(Suppl.), 791–795 (2002)
57. Shargorodsky, E.: On the level sets of the resolvent norm of a linear operator. Bull. Lond. Math. Soc.
40, 493–504 (2008)
58. Treil, S.: Invertibility of Toelitz operators does not imply applicability of the finite section method.
Dokl. Akad. Nauk SSSR 292, 563–567 (1987) (Russian)
59. Trench, W.F.: An elementary view of Weyl’s theory of equal distribution. Am. Math. Mon. 119, 852–
861 (2012)
60. Vainshtein, L.A.: Open resonance for lasers. Sov. Phys. JETP 40, 709–719 (1963)