IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 54, NO.
12, DECEMBER 2008
Asymptotically Equivalent Sequences of Matrices and
Hermitian Block Toeplitz Matrices With Continuous
Symbols: Applications to MIMO Systems
Jesús Gutiérrez-Gutiérrez and Pedro M. Crespo, Senior Member, IEEE
Abstract—For the engineering community, Gray’s tutorial
monograph on Toeplitz and circulant matrices has been, and
remains, the best elementary introduction to the Szegö theory on
large Toeplitz matrices. In this paper, the most important results
of the cited monograph are generalized to block Toeplitz (BT)
matrices by maintaining the same mathematical tools used by
Gray, that is, by using asymptotically equivalent sequences of
matrices. As applications of these results, the geometric minimum
mean square error (MMSE) for both an infinite-length multi-
variate linear predictor and an infinite-length decision feedback
equalizer (DFE) for multiple-input–multiple-output (MIMO)
channels, are obtained as a limit of the corresponding finite-length
cases. Similarly, a short derivation of the well-known capacity of
a time-invariant MIMO Gaussian channel with intersymbol inter-
ference (ISI) and fixed input covariance matrix is also presented.
Index Terms—Block circulant matrices, block Toeplitz (BT)
matrices, channel capacity, Frobenius norm, functions of ma-
trices, Hermitian matrices, multiple-input–multiple-output
(MIMO) minimum mean square error decision feedback equal-
izer (MMSE-DFE), (MMSE) linear predictor, multivariate
stationary processes, spectral norm.
I. INTRODUCTION
T HE most famous theorem on large Toeplitz matrices is
the Szegö theorem [1]. This result deals with the asymp-
totic behavior of the eigenvalues of an Hermitian Toeplitz
matrix , where the complex numbers with
are the Fourier coefficients of a bounded function called
the symbol of the sequence . Unfortunately, the level of
mathematical sophistication for understanding [1] is often be-
yond the level provided by a typical engineering background.
For this reason, Gray simplified the Szegö theory in [2] and by
assuming more restrictive conditions, namely, Toeplitz matrices
with symbols in the Wiener algebra (i.e., ).
Since his approach is purely matrix-analytic (mainly based on
his concept of asymptotically equivalent sequences of matrices),
Manuscript received February 26, 2007; revised May 13, 2008. Current
version published November 21, 2008. This work was supported in part by
the Spanish Ministry of Education and Science, by the European Social Fund,
and by the European Regional Development Fund through the Torres-Quevedo
program, the MultiMIMO Project TEC2007-68020-C04-03, and the program
CONSOLIDER-INGENIO 2010 CSD2008–00010 COMONSES.
The authors are with Electronics and Communications Department, R&D
Center CEIT, 20018 San Sebastián, Spain and University of Navarra (Tecnun),
20018 San Sebastián, Spain (e-mail: jgutierrez@ceit.es; pcrespo@ceit.es).
Communicated by A. Høst-Madsen, Associate Editor for Detection and Es-
timation.
Digital Object Identifier 10.1109/TIT.2008.2006401
0018-9448/$25.00 © 2008 IEEE
Authorized licensed use limited to: Pedro M Crespo. Downloaded on December 1, 2008 at 03:32 from IEEE Xplore. Restrictions apply.
5671
he succeeded in conveying the main ideas of the Szegö theory
to a wider audience.
In the paper of Gazzah, Regalia, and Delmas [3], the Szegö
theorem for Toeplitz matrices with symbols in the Wiener al-
gebra is generalized to block Toeplitz (BT) matrices by using
asymptotically equivalent sequences of matrices (i.e., Gray’s
tools) and the concept of vec-permutation matrix.
In this paper, we go a step further. By using again the concept
of asymptotically equivalent sequences of matrices, we extend
the most important results (the Szegö theorem among others) of
the Gray tutorial [2], to BT matrices with continuous symbols
(i.e., the with are given by the Fourier coefficients of a
continuous matrix function), which is a less restrictive
condition than being in the Wiener algebra.
In [4], Bose and Boo also derived asymptotic results of BT
matrices based on Gray’s tools. However, they look at a com-
pletely different problem to the one considered in our paper.
Here, is an BT matrix with blocks, and we look,
among other things, at the asymptotic behavior of the eigen-
values of as goes to infinity while remains fixed (specifi-
cally, we study the Szegö theorem for continuous matrix-valued
symbols). On the contrary, in [4], is an block Toeplitz
matrix with Toeplitz blocks (BTTB) of order , and Bose
and Boo look at the asymptotic behavior of the eigenvalues of
when both and go to infinity (specifically, they study
the Szegö theorem for certain two-variate symbols). It should
be pointed out that their derivation cannot be particularized to
the case where is kept fixed.
Finally, it should be mentioned that, based on matrix analysis
but without using the Gray concept of asymptotically equiva-
lent sequences of matrices, Tilli [5] and Serra-Capizzano [6] de-
rive results of Szegö type for Lebesgue integrable multivariate
matrix-valued symbols that are stronger than the results given
by Bose and Boo [4] and the ones presented here. Such gener-
ality is based on introducing another concept of approximation
of matrix sequences, namely, the notion of approximating class
of sequences (a.c.s.), which is more general than that of asymp-
totically equivalent sequences. In the Appendix, we report some
analysis based on this new approach, since it has the potential of
becoming a further tool available in the engineering community.
Therefore, the main contribution of our paper is to present
a plain generalization to BT matrices of Gray’s famous tuto-
rial [2], by maintaining the Gray tools. The theory on BT ma-
trices is, in general, useful in signal processing, communica-
tions, and information theory, since this kind of matrices arise,
for instance, as covariance matrices of multivariate stationary
5672 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 54, NO. 12, DECEMBER 2008

processes or as matrix representations of linear time-invariant Lemma 1:

discrete-time multivariate filters. As applications of this theory 1) If , then .
to multiple-input–multiple-output (MIMO) systems, the geo- 2) If and , then .
metric minimum mean square error (MMSE) for both an infi- 3) If , then for every .
nite-length multivariate linear predictor and an infinite-length 4) If and , then
decision feedback equalizer (DFE) for MIMO channels are ob- and .
tained as a limit of the corresponding finite-length cases. Simi- Notice that is not reflexive.
larly, a short derivation of the well-known capacity of a time-in-
variant MIMO Gaussian channel with intersymbol interference B. Functions of Matrices
(ISI) and fixed input covariance matrix is also presented. We first review the concept of polynomial function of a ma-
The remainder of this paper is organized as follows. The trix.
next section states preliminary mathematical considerations.
Section III derives a new result on matrix sequences that will Definition 3: If is a given
be used in Section IV. Section IV is the main part of the paper polynomial with coefficients in , then for every matrix
and derives several fundamental theorems on large Hermitian , we define , where
BT matrices. We conclude with Section V where the furnished denotes the identity matrix.
theoretical results are applied to three problems in the area of We now introduce the concept of function of a matrix. For a
signal processing and communications. thorough treatment, we refer the reader to [10] or [11]. We also
recommend the text [12] for a treatment of functions of matrices,
II. PRELIMINARY MATHEMATICAL DEFINITIONS AND RESULTS
developed in a signal processing context.
A. Asymptotically Equivalent Sequences of Matrices Definition 4: We consider an diagonalizable matrix
If is an diagonalizable matrix, then we will indicate . If is a complex function
by , all the eigenvalues of counted with on the set , we define the matrix
their multiplicities. The following well-known matrix norms .
(see, e.g., [7] or [8]) will be used.
Observe that the last definition agrees with Definition
Definition 1: The Frobenius norm and the spectral 3. Furthermore, if is a complex function on the set
norm of an matrix are defined of the eigenvalues of an diag-
as onalizable matrix , there exists a polynomial of degree
at most such that . This is a direct
consequence of the fact that the interpolation problem of deter-
mining a polynomial of degree at most satisfying
, has a solution (see, e.g.,
[8]). Therefore, Definition 4 is independent of the chosen
eigenvalue decomposition of the matrix .

where denotes trace, denotes conjugate transpose, and is C. Block Circulant Matrices
a column vector. We now review a special type of BT matrices.
The Frobenius norm and the spectral norm are special in-
stances of unitarily invariant norms (see, e.g., [7] or [8]). We Definition 5: An block circulant matrix with
recall that a norm is unitarily invariant if blocks is one having the form
for every matrix and for every unitary matrices
and . .. ..
. .
The next definition is due to Gray (see [2] or [9]).
..
.
Definition 2: Let and be matrices for all . .. .. .. ..
We say that the sequences and are asymptotically . . . .
equivalent, and write , if

(1) where , and is the set of all

and complex matrices.
The next well-known lemma characterizes block circulant
(2) matrices.
Lemma 2: , with , is an
In the following lemma, the basic properties of this kind of se- block circulant matrix with blocks if and only if
quences are reviewed. Most of them are directly from Gray’s
monograph [2].

Authorized licensed use limited to: Pedro M Crespo. Downloaded on December 1, 2008 at 03:32 from IEEE Xplore. Restrictions apply.
GUTIÉRREZ-GUTIÉRREZ AND CRESPO: ASYMPTOTICALLY EQUIVALENT SEQUENCES OF MATRICES 5673

where is the Kronecker product, is the From Lemma 1, the theorem holds for any polynomial .
Fourier unitary matrix Now, let be an arbitrary continuous function on . Since
the spectral norm is unitarily invariant, we deduce that

and with being the

Kronecker delta and satisfying
for all , where . Obviously, it can
be analogously shown that
.. .. for all .
. . On the other hand, by the Stone–Weierstrass approximation
theorem (see, e.g., [13]), given , there is a polynomial
such that . Furthermore, from the
If is an block circulant matrix, it is
fact that , there exists such that
customary to write .

D. Two Properties of Partitioned Matrices

Finally, we review two well-known results regarding the de-
terminant and the inverse of a partitioned matrix (see, e.g., [8]), Consequently, by using the triangle inequality and the fact that
which are based on block two-sided Gaussian elimination steps. the Frobenius norm is unitarily invariant, we have

Lemma 3: Let be a square matrix partitioned as

where and are also square. If and are nonsin-

gular, then

Lemma 4: Let be a square matrix partitioned as in Lemma

3. If and are nonsingular, then the corresponding
partitioned presentation of is as presented in the expression
at the bottom of the page.

III. A NEW RESULT ON ASYMPTOTICALLY EQUIVALENT for every . This ends the proof.
SEQUENCES OF MATRICES This result will be very useful as shown in the following sec-
Theorem 1: Consider two asymptotically equivalent se- tion. It should be mentioned that Theorem 1 can be stated in a
quences of Hermitian matrices and . Let be more general way (see the Appendix).
a closed interval containing the eigenvalues of the matrices of
these sequences. Then IV. LARGE HERMITIAN BLOCK TOEPLITZ MATRICES
In this section, we will derive several fundamental theorems
on Hermitian BT matrices with blocks, for large
where denotes the set of all continuous complex func- .
tions on . Fix , and let and be matrices for all
Proof: Because and are sequences of Her- . We say that the sequences and are asymp-
mitian matrices, (1) implies that there is such that totically equivalent, and write , if they satisfy the
. There- expressions (1) and (2). Note that Lemma 1 and Theorem 1 still
fore, there exists such an interval , for instance, . hold in this more general context.

Authorized licensed use limited to: Pedro M Crespo. Downloaded on December 1, 2008 at 03:32 from IEEE Xplore. Restrictions apply.
5674 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 54, NO. 12, DECEMBER 2008

In the sequel, let be an matrix tian. By reasoning as in [2] and for the scalar case, it is not dif-
function, which is continuous and -periodic. We denote by ficult to show (see, e.g., [14, p. 871]) that the eigenvalues of the
the sequence of Fourier coefficients of matrices are contained in the closed interval with finite
endpoints

and by the sequence of partial sums of the Fourier

series of
and

Therefore, from the Parseval theorem (see, e.g., [13]), we de-

duce that Observe that the eigenvalues of the matrices also lie in-
side , since they are given by
(3)

Let us consider the BT matrix ,

with blocks, that is We are now in a position to give an interesting example of
asymptotically equivalent sequences of matrices.

.. Lemma 5 (BT Matrices): Let be a Her-

..
. . mitian matrix function, which is continuous and -periodic.
.. Then, .
.
.. Proof: For every , because contains
.. .. ..
. . . . the eigenvalues of the Hermitian matrices and ,
we have .
On the other hand, from (3), given , there exists such
The function is called the symbol of the BT matrices. that
Because
(6)

Thus, by the triangle inequality

by Parseval’s theorem, we have

(4)

We also associate with the function a sequence of block cir-

culant matrices given by

(7)

From (4) and (6), we obtain

Our definition of is the natural generalization of the
one given by Gray for the scalar case [2]. Notice that when the
are banded, that is, for certain , it is deduced (8)
from Lemma 2 that
On account of (5), we have

(5)
Therefore, there exists such that
for all , where denotes the zero matrix.
Suppose moreover that is Hermitian for all . Con-
(9)
sequently, the and matrices will also be Hermi-

Authorized licensed use limited to: Pedro M Crespo. Downloaded on December 1, 2008 at 03:32 from IEEE Xplore. Restrictions apply.
GUTIÉRREZ-GUTIÉRREZ AND CRESPO: ASYMPTOTICALLY EQUIVALENT SEQUENCES OF MATRICES 5675

Furthermore, because the Frobenius norm is unitarily invariant, By using the Cauchy–Schwarz inequality and the fact that
we deduce that is a unitary matrix, we see that

for every . Thus, converges uniformly

where the continuity of on guarantees the ex- to on . Because the are continuous
istence of the above limit as a Riemann integral. Consequently, on , we conclude that is also continuous on .
there is such that We are now ready to prove the assertion of the theorem. We
denote the real part and the imaginary part of by and
, respectively. Obviously, and
are Hermitian for all . Therefore, from Theorem 1,
Lemma 1, and Lemma 5, we deduce that
(10)

for all . Finally, from (7)–(10), we conclude that

The product of two BT matrices and is not, in As in the earlier case, based on advanced tools from func-
general, BT. However, the next theorem shows that this product tional analysis, a stronger result linking to
is asymptotically equivalent to the BT matrix of the product of can be obtained (see [15, Sec. 2.7]).
the two symbols, i.e., . By particularizing in Theorem 3 to
, and , we obtain the next result.
Theorem 2 (Product of BT Matrices): Let
be Hermitian matrix functions, which are continuous and Theorem 4 (Exponential, Inverse, and Square Root of BT Ma-
-periodic. If is also Hermitian, then trices): Let be as in Lemma 5.
. 1) Then .
Proof: By combining Lemmas 1 and 5, we have 2) If , then .
3) If , then .
. We now derive the Szegö theorem for BT matrix sequences
It should be mentioned that Widom, by using advanced tools with continuous symbols.
from functional analysis, proved a stronger and more general Theorem 5 (Trace of Functions of BT Matrices): Let be as
result on the product of two BT matrices that can be found, for in Lemma 5. Then
instance, in [6, Sec. 3.3.1], [15, p. 192], or [16].
From Theorem 2, with
, i.e., . Based on Theorem
1 and Lemma 5, we will show that this is also true for more
general functions. for all .
Theorem 3 (Functions of BT Matrices): If is as in Lemma Proof: Because for all matrix
5, then . , from Theorem 3, we have
Proof: Fix . We begin by showing
that is continuous on . We
will denote by
an eigenvalue decomposition of for all . By
the Stone–Weierstrass approximation theorem, there exists a
sequence of polynomials that converges uniformly to The extension of Theorem 5 to Lebesgue integrable symbols
on . Therefore, given , we can find can be found in [5] and [6]. It should be mentioned
such that that when is bounded , the assumption of bounded
support for the test function is not necessary. Moreover, for
, the assumption of bounded support

Authorized licensed use limited to: Pedro M Crespo. Downloaded on December 1, 2008 at 03:32 from IEEE Xplore. Restrictions apply.
5676 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 54, NO. 12, DECEMBER 2008

can be replaced by an assumption of mild growth at infinity. the theorem for the Frobenius norm. Using the Cauchy–Schwarz
More precisely, it is required that the test functions are inequality, we have
continuous and is bounded (see [17]).
The following theorem is a straightforward consequence of
Theorem 5, and it deals with determinants instead of traces.
Theorem 6 (Determinant of Functions of BT Matrices): Let
be as in Lemma 5. Then

for all positive function .

Proof: We have

for all . Notice that by virtue of Theorem 3, (11) now

follows. The proof for (12) is similar.
Observe that all the results of this section still hold when
The result now follows from Theorem 5 and the continuity of the sequence of matrices is replaced by another se-
the exponential function. quence of Hermitian matrices with ,
By particularizing in Theorem 6 to , we obtain the and is replaced by a bigger closed interval con-
next result on limits of BT determinants. taining the eigenvalues of the matrices.
Finally, notice that most of the main results of this section
Theorem 7 (BT Determinants): Let be as in Lemma 5. If (Lemma 5, Theorem 2, assertion 2 of Theorem 4, Theorem 5,
, then and Theorem 7) were proved for the scalar case by Gray [2],
assuming more restrictive conditions, namely, for functions
in the Wiener algebra rather than with continuous symbols. It
should also be noted that, as in [2], the asymptotic results pre-
sented in our paper do not concern individual entries of the
As we have seen, Theorem 5 deals with the asymptotic behavior matrices considered but rather, they describe an “average” be-
of the arithmetic average (scalar) of the entries belonging to the havior. For recent results on Toeplitz matrices regarding ele-
main diagonal of , i.e., . The last mentwise convergence with engineering applications, we refer
result of the paper is stronger than Theorem 5, because it deals the reader to [18]–[20].
with the arithmetic average ( matrix) of the block entries
belonging to an arbitrary diagonal of , for large . V. APPLICATIONS
Theorem 8: Let be as in Lemma 5. If is a nonnegative We now present three examples of problems in the area of
integer, then for every , signal processing and communications that can be solved by
using the theoretical results of the previous sections.

A. Infinite-Length Multivariate MMSE Linear Predictor

(11) In the first example, we evaluate the geometric MMSE for an

infinite-length multivariate linear predictor, as a limit of the cor-
responding finite-length cases. The observed multivariate sta-
tionary process is given by . The process
(12) is a moving average (MA) process obtained at the output of a
finite-impulse response (FIR) MIMO filter (i.e., a multivariate
where is the block entry of FIR filter)
.
Proof: Because in the finite-dimensional vector space
all norms are equivalent, we will prove the assertion of

Authorized licensed use limited to: Pedro M Crespo. Downloaded on December 1, 2008 at 03:32 from IEEE Xplore. Restrictions apply.
GUTIÉRREZ-GUTIÉRREZ AND CRESPO: ASYMPTOTICALLY EQUIVALENT SEQUENCES OF MATRICES 5677

0 1 taps.
Fig. 2. MIMO communication system: discrete time channel and equalizer.
Fig. 1. Multivariate forward linear predictor with n

The matrix can be partitioned as

where denotes the input process and
the filter taps are matrices. The vari-
ables , with and , are uncorrelated random
variables with zero mean and variance . The observation noise
is uncorrelated with the input, i.e., and, consequently
for all ; here is the
zero matrix. Furthermore, the noise variables , with
and , are uncorrelated random variables with
zero mean and variance . Combining Lemmas 3 and 4, we deduce that
Let be the BT matrix with
blocks given by (13) shown at the bottom of the page. We
denote by the banded Hermitian BT matrix with
blocks defined by . This matrix can be
written as , with where
and, therefore, . Be-
is the transfer function of the MA
cause , using a well-known property regarding
filter.
limits of sequences of positive numbers yields
We linearly estimate from the previous vectors
. Let denote
such an estimate. Fig. 1 shows this forward linear predictor.
The weight matrices are chosen to minimize the
mean square error (MSE) defined by the trace1 of the error au-
tocorrelation matrix with .
By reasoning as in [21] for the univariate case of the forward Finally, by Theorem 6, the geometric MMSE for an infinite-
linear predictor, it is not difficult to prove that the resulting length multivariate forward linear predictor can be written as
error autocorrelation matrix can be expressed as

where and is the

block entry of . Thus, the corresponding geometric
MMSE is now computed as

B. Infinite-Length MIMO MMSE-DFE

In the second example, we evaluate the geometric MMSE for
1Because we are dealing with multidimensional error random processes, ei- an infinite-length DFE in MIMO channels, as a limit of the cor-
0
ther the trace or the determinant of R (n 1) can be used as a mean square responding finite-length cases. We consider the communication
error measure. It turns out that by using either criterion, the obtained optimal
system parameters and the associated error autocorrelation matrix R (n 0 system shown in Fig. 2 where the MIMO channel has inputs
1) are the same. and outputs. Its discrete channel impulse response is given

.. .. .. .. .. .. ..
. . . . . . . (13)
.. .. .. .. ..
. . . . .

Authorized licensed use limited to: Pedro M Crespo. Downloaded on December 1, 2008 at 03:32 from IEEE Xplore. Restrictions apply.
5678 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 54, NO. 12, DECEMBER 2008

by , where the are matrices. The perhaps for the upper left corner, and therefore,
channel output can be written as . The matrix can be partitioned as

Thus
where and
denote the input and noise, respectively. The input components
symbols are discrete uncorrelated random variables with
and . The observation noise is un-
Combining Lemmas 3 and 4, we deduce that
correlated with the input, and its samples are uncorrelated
random variables with zero mean and variance .
The equalizer block is a standard DFE that comprises an
tapped delay line feedforward filter (FFF), and an infinite
tapped delay line feedback filter (FBF). We linearly estimate the
input vector , where is the decision (14)
delay, from the channel outputs and
from the inputs (assuming Without any loss of generality, suppose that
correct previous decisions). Thus, such an estimate has where is a function such that
the form , and (for example,
). Consequently, for sufficiently large , the
matrix is equal to and by using (14), we have

with the and being matrices and ma-

trices, respectively. These matrix taps and are chosen
to minimize the MSE defined by the trace2 of the error au-
tocorrelation matrix with
. From [22], it can be easily shown that the re-
sulting error autocorrelation matrix can be expressed as
Because , from Theorem 6, the geometric
MMSE for an infinite-length MIMO DFE can be written as
where is the matrix obtained by deleting the last
rows and columns of the matrix , with
being the channel matrix (13); the function is now

and is the block

entry of . Thus, the corresponding geometric MMSE
can be computed as C. Capacity of a Time-Invariant MIMO Gaussian Channel
With ISI and Fixed Input Covariance Matrix
Finally, we derive the capacity of a time-invariant MIMO
Let now be the matrix function Gaussian channel with impulse response subject to an
where is the transfer function of the input constraint, namely, a flat input power spectral density. The
channel. Although the Hermitian matrix input–output relation is given by
is in general not BT, it is equal to the banded Hermitian
BT matrix except perhaps for the upper
left and lower right corners. Furthermore, the entries of these
corners do not depend on the order. Consequently, if we define where and are the input
as the matrix obtained by deleting the last and the output of the channel, respectively. The channel noise
rows and columns of , then is equal to except is independent of the input, and the noise random vectors
are assumed to be independent identically dis-
2As in the first application, the trace or the determinant of R (n) can be used
tributed (i.i.d.) and circularly symmetric Gaussian with zero
as a mean square error measure. In [22], it was shown that the obtained optimal
system parameters and the associated error autocorrelation matrix R (n) mean and covariance . For every , the output of
are the same by using either criterion. the channel up to time can be written in matrix notation as

Authorized licensed use limited to: Pedro M Crespo. Downloaded on December 1, 2008 at 03:32 from IEEE Xplore. Restrictions apply.
GUTIÉRREZ-GUTIÉRREZ AND CRESPO: ASYMPTOTICALLY EQUIVALENT SEQUENCES OF MATRICES 5679

, where is given by expression All the Schatten norms are special instances of unitarily in-
(13), variant norms (see, e.g., [7] or [8]).
and . Using these norms, the definition of asymptotically -equiv-
Assuming we restrict the covariance of the input random alent sequences of matrices is as follows.
vector to be , the maximum
Definition 7: Consider a strictly increasing sequence of nat-
of the normalized mutual information is ural numbers . Let and be matrices for all
bits per channel use, and it . We say that the sequences and are asymp-
is obtained when the input process is Gaussian, with zero mean totically -equivalent, , and write , if
[23]. By denoting , the signal-to-noise ratio at the
input of the MIMO channel, the channel capacity is now given (15)
by the limiting expression and
(16)

The notion of asymptotically -equivalent sequences is a con-

cept of approximation in norm of matrix sequences. Observe
that for and is the same as
(the standard notion introduced by Gray).
The second and last concept is the notion of a.c.s. due to
where is the matrix function , with Serra-Capizzano [26]. It deals with approximation in norm and
, and . Now, in rank.
by applying Theorem 5, we obtain the well-known expression
for the desired capacity [24] Definition 8: Consider a strictly increasing sequence of nat-
ural numbers . Let be a matrix for all .
We say that is an a.c.s. for if, for all suffi-
ciently large , the following splitting holds:

with

APPENDIX
OTHER CONCEPTS OF APPROXIMATION OF MATRIX SEQUENCES where and depend only on and, moreover
In this Appendix, we present the notions of asymptotically
-equivalent sequences of matrices and that of a.c.s. Both are
concepts of approximation of matrix sequences that generalize
Gray’s concept of asymptotically equivalent sequences. We The a.c.s. notion is more general than that of asymptotically
present these two notions to show that Theorem 1 can be -equivalent sequences. In fact, we show next that if
stated in a more general way, by proving that this theorem can , then is an a.c.s. for . Moreover, the
be considered as a special instance of certain known results, requirement of uniform boundedness in (15) is not necessary.
namely, [25, Th. 2.1 and 2.2]. Lemma 6: Consider and a strictly increasing se-
We begin with the definition of asymptotically -equivalent quence of natural numbers . Let and be ma-
sequences of matrices that can be obtained from the definition trices for all . If expression (16) holds, then
of asymptotically equivalent sequences of matrices by substi- is an a.c.s. for .
tuting the Frobenius norm for any other Schatten -norm. We Proof: Let be a sin-
first review these well-known norms (see, e.g., [7] or [8]). gular value decomposition of the matrix with
Definition 6: The Schatten -norm of an matrix . We define
, is the norm on of the vector of the singular
values of , i.e.,
if
if
where , with , are the singular values of
counted with their multiplicities. Obviously, , because
Notice that (the spectral radius if is Her- . On the one hand
mitian) and (the norm on of the vector
containing the eigenvalues of when is Hermitian).

Authorized licensed use limited to: Pedro M Crespo. Downloaded on December 1, 2008 at 03:32 from IEEE Xplore. Restrictions apply.
5680 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 54, NO. 12, DECEMBER 2008

On the other hand, because (16) holds, there exists REFERENCES

such that [1] U. Grenander and G. Szegö, Toeplitz Forms and Their Applications.
Los Angeles, CA: Univ. California Press, 1958.
[2] R. M. Gray, “Toeplitz and circulant matrices: A review,” Found. Trends
Commun. Inf. Theory, vol. 2, no. 3, pp. 155–239, 2006.
and, therefore [3] H. Gazzah, P. A. Regalia, and J. P. Delmas, “Asymptotic eigenvalue
distribution of block Toeplitz matrices and application to blind SIMO
channel identification,” IEEE Trans. Inf. Theory, vol. 47, no. 3, pp.
1243–1251, Mar. 2001.
[4] N. K. Bose and K. J. Boo, “Asymptotic eigenvalue distribution of
block-Toeplitz matrices,” IEEE Trans. Inf. Theory, vol. 44, no. 2, pp.
858–861, Mar. 1998.
[5] P. Tilli, “A note on the spectral distribution of Toeplitz matrices,”
Linear Multilinear Algebra, vol. 45, pp. 147–159, 1998.
[6] S. Serra-Capizzano, “The GLT class as a generalized Fourier analysis
and applications,” Linear Algebra Appl., vol. 419, pp. 180–233, 2006.
[7] R. Bhatia, Matrix Analysis. New York: Springer-Verlag, 1997.
[8] R. A. Horn and C. R. Johnson, Matrix Analysis. New York: Cam-
bridge Univ. Press, 1990.
[9] R. M. Gray, “On the asymptotic eigenvalue distribution of Toeplitz ma-
trices,” IEEE Trans. Inf. Theory, vol. IT-18, no. 6, pp. 725–730, Nov.
1972.
[10] F. R. Gantmacher, The Theory of Matrices. New York: Chelsea Pub-
lishing, 1977, vol. 1.
[11] R. A. Horn and C. R. Johnson, Topics in Matrix Analysis. New York:
Cambridge Univ. Press, 1994.
[12] R. A. Roberts and C. T. Mullis, Digital Signal Processing. Reading,
MA: Addison-Wesley, 1987.
[13] W. Rudin, Principles of Mathematical Analysis. New York: Mc-
Graw-Hill, 1976.
[14] M. Miranda and P. Tilli, “Asymptotic spectra of Hermitian block
Toeplitz matrices and preconditioning results,” SIAM J. Matrix Anal.
Appl., vol. 21, no. 3, pp. 867–881, 2000.
for all . [15] A. Böttcher and B. Silbermann, Introduction to Large Truncated
Toeplitz Matrices. New York: Springer-Verlag, 1999.
Finally, it should be mentioned that by using the previous [16] H. Widom, “Asymptotic behavior of block Toeplitz matrices and deter-
minants. II,” Adv. Math., vol. 21, pp. 1–29, 1976.
lemma (with and ), Theorem 1 can be obtained [17] S. Serra-Capizzano, “Test functions, growth conditions and Toeplitz
from [25, Th. 2.1 and 2.2]. In fact, Theorem 1 can be stated in a matrices,” Rend. Circolo Mat. Palermo, vol. 68, pp. 791–795, 2002,
more general way: Serie II.
[18] A. Böttcher, J. Gutiérrez-Gutiérrez, and P. M. Crespo, “Mass concen-
i) if , with some , then tration in quasicommutators of Toeplitz matrices,” J. Comput. Appl.
; Math., vol. 205, pp. 129–148, 2007.
ii) if a.c.s. for , then a.c.s. [19] P. M. Crespo and J. Gutiérrez-Gutiérrez, “On the elementwise conver-
gence of continuous functions of Hermitian banded Toeplitz matrices,”
for ; IEEE Trans. Inf. Theory, vol. 53, no. 3, pp. 1168–1176, Mar. 2007.
for all continuous function with bounded support. Theorem 2 [20] J. Gutiérrez-Gutiérrez, P. M. Crespo, and A. Böttcher, “Functions of
also holds if is replaced by for every (see [26] banded Hermitian block Toeplitz matrices in signal processing,” Linear
Algebra Appl., vol. 422, pp. 788–807, 2007.
and [27]), even for bounded symbols. If the sequence [21] S. Haykin, Adaptive Filter Theory. Englewood Cliffs, NJ: Prentice-
is substituted by , where the matrix is obtained Hall, 1986.
by minimizing over all the block circulant ma- [22] N. Al-Dhahir and A. H. Sayed, “The finite-length multi-input multi-
output MMSE-DFE,” IEEE Trans. Signal Process., vol. 48, no. 10, pp.
trices (see [28, Definition 4.1]), then Lemma 5 also holds 2921–2936, Oct. 2000.
when is replaced by for all , even for bounded [23] T. M. Cover and J. A. Thomas, Elements of Information Theory. New
symbols. In fact, it holds that (see [17], [27], and [28, Prop. 4.3]): York: Wiley, 2005.
[24] L. H. Brandenburg and A. D. Wyner, “Capacity of the Gaussian channel
i) ; with memory: The multivariate case,” Bell Syst. Tech. J., vol. 53, no. 5,
ii) for every and for every pp. 745–778, May–Jun. 1974.
[25] S. Serra-Capizzano and D. Sesana, “Approximating classes of se-
unitarily invariant norm ; quences: The Hermitian case,” in Heinig’s Memorial Volume, D.
iii) is uniformly bounded for all . Bini, V. Olshevsky, and E. Tyrtyshnikov, Eds. Cambridge, MA:
Moreover, the matrix , unlike , has a simple ex- Birkhäuser, 2008.
[26] S. Serra-Capizzano, “Distribution results on the algebra generated by
plicit expression (see [28, eq. (18)]) in terms of the Fourier co- Toeplitz sequences: A finite dimensional approach,” Linear Algebra
efficients of . Appl., vol. 328, pp. 121–130, 2001.
[27] S. Serra-Capizzano and P. Tilli, “On unitarily invariant norms of matrix
valued linear positive operators,” J. Inequalities Appl., vol. 7, no. 3, pp.
ACKNOWLEDGMENT 309–330, 2002.
[28] S. Serra-Capizzano, “The spectral approximation of multiplication
The authors would like to thank the anonymous referees for operators via asymptotic (structured) linear algebra,” Linear Algebra
their thorough review. Appl., vol. 424, pp. 154–176, 2007.

Authorized licensed use limited to: Pedro M Crespo. Downloaded on December 1, 2008 at 03:32 from IEEE Xplore. Restrictions apply.