IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 45, NO.

4, APRIL 1997 851

Discrete Zak Transforms, Polyphase

Transforms, and Applications
Helmut Bölcskei, Student Member, IEEE, and Franz Hlawatsch, Member, IEEE

Abstract— We consider three different versions of the Zak application of discrete ZT’s to Gabor expansions (WHF’s) and
transform (ZT) for discrete-time signals, namely, the discrete- DFT filter banks.
time ZT, the polyphase transform, and a cyclic discrete ZT. In The paper is organized as follows. In Section II, we define
particular, we show that the extension of the discrete-time ZT
to the complex -plane results in the polyphase transform, an the discrete-time ZT (DTZT), which is discrete in time but
important and well-known concept in multirate signal processing continuous and periodic in frequency, and we discuss basic
and filter bank theory. properties of the DTZT. Section III considers the applica-
We discuss fundamental properties, relations, and transform tion of the DTZT to the discrete-time Gabor expansion.
pairs of the three discrete ZT versions, and we summarize We review the Zibulski–Zeevi method for calculating dual
applications of these transforms. In particular, the discrete-time
ZT and the cyclic discrete ZT are important for discrete-time WHF’s, and we show that the DTZT diagonalizes the Gabor
Gabor expansion (Weyl–Heisenberg frame) theory since they (Weyl–Heisenberg) frame operator for critical sampling and
diagonalize the Weyl–Heisenberg frame operator for critical integer oversampling, which makes the DTZT-based calcula-
sampling and integer oversampling. The polyphase representation tion of dual and tight Gabor frames and frame bounds very
plays a fundamental role in the theory of filter banks, especially efficient. We furthermore find the (generalized) eigenfunctions
DFT filter banks. Simulation results are presented to demonstrate
the application of the discrete ZT to the efficient calculation of and eigenvalues of the Gabor frame operator for critical
dual Gabor windows, tight Gabor windows, and frame bounds. sampling and integer oversampling.
In Section IV, the DTZT is extended to the complex -
plane, which is shown to result in the well-known polyphase
I. INTRODUCTION AND OUTLINE transform (PT). Important properties and transform pairs of
the PT are provided. Section V reviews the application of the
T HE ZAK TRANSFORM (ZT), which is also known
as Weil–Brezin mapping, Gel’fand mapping, and
representation [1]–[5], plays an important role in the theory
- PT to oversampled and critically sampled DFT filter banks.
Although some of the observations of Section V hold for more
of Gabor expansions and Weyl–Heisenberg frames (WHF’s) general filter banks [27], [29], we shall restrict our attention to
[6]–[24]. On the theoretical side, the ZT allows the formulation DFT filter banks for the sake of coherence with other sections.
of theorems on the existence of WHF’s. On the practical side, Section VI introduces the discrete ZT (DZT), which is
the ZT is an efficient tool for the numerical calculation of dual discrete in time and frequency and, thus, particularly appro-
WHF’s, tight WHF’s, and frame bounds. The ZT is very useful priate for numerical implementations. Important properties of
in several other fields as well, such as the theory of coherent the DZT are discussed. In Section VII, we show that the
states [4], [5], wavelet transforms [25], linear equalizers [26], DZT allows an efficient, FFT-based implementation [33], [12]
and filter bank theory [27]–[29]. of a cyclic version of the discrete Gabor expansion [8] for
A discrete-time [30]–[32], [28], [10], [19] and a cyclic, critical sampling and integer oversampling. Finally, Section
discrete-time/discrete-frequency ZT version [33], [34], [12], VIII presents simulation results demonstrating the importance
[16], [10] have recently been introduced. However, while an of the DZT for the efficient calculation of dual and tight Gabor
excellent tutorial paper on the continuous-time ZT exists [2], windows as well as frame bounds.
a similar work for the discrete ZT versions is lacking. The
present paper, which is partly of survey nature but also presents II. THE DISCRETE-TIME ZAK TRANSFORM
original results, is an attempt to fill this gap. Specifically,
In this section, we consider the discrete-time ZT (DTZT)
we summarize and extend systematic discussions of some
[30]–[32], [28], [10], [19] and discuss its properties. Most of
properties of discrete ZT’s that have been provided by various
these properties are obtained by straightforward discretization
authors [30], [32]–[35], [10], [12]. We also show that the
of results found for the continuous-time case by Janssen [2].
discrete-time ZT extended to the complex -plane [28] is the
1) Definition and Interpretation: The DTZT can be defined
same as the polyphase transform important for multirate signal
by sampling the continuous-time ZT with respect to time. The
processing and filter banks [36]–[44], and we summarize the
continuous-time ZT of a signal is defined as [2]
Manuscript received May 11, 1996; revised October 14, 1996. This work
was supported by FWF Grants P10012-ÖPH and P10531-ÖPH. The associate
editor coordinating the review of this paper and approving it for publication
was Dr. Jelena Kovacević.
The authors are with INTHFT, Vienna University of Technology,
Gusshausstrasse 25/389, A-1040 Vienna, Austria. where is time, is frequency, and is fixed time period
Publisher Item Identifier S 1053-587X(97)02572-5. parameter. Evaluating the ZT at uniformly spaced time instants
1053–587X/97$10.00 © 1997 IEEE

Authorized licensed use limited to: ETH BIBLIOTHEK ZURICH. Downloaded on January 25, 2010 at 04:09 from IEEE Xplore. Restrictions apply.
852 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 45, NO. 4, APRIL 1997

, where , setting , and dividing by 3) Periodicity: The DTZT is quasiperiodic in and peri-
, we obtain odic in :

with (2)

Hence, it suffices to know the DTZT on the “fundamental

rectangle”
4) Integral/Sum Relations: Some important relations in-
volve integrals and/or sums of the DTZT. First, the signal
Introducing the discrete-time signal as a can be recovered from its DTZT by integrating the
sampled version of with sampling period , the DTZT DTZT with respect to frequency :
is finally formulated as
(3)
(1)
We can also recover the signal’s Fourier transform from
as
The convergence of the series (1) is discussed in [30]. The
DTZT is a linear signal transformation that maps a discrete-
(4)
time signal onto a 2-D function of the discrete time
index and the (continuous) normalized frequency variable
The DTZT parameter is analogous to the parameter This is related to the polyphase decomposition in multirate
in the continuous-time ZT. We shall usually omit the signal processing [36]–[44] (see Sections IV and V).
superscript in order to simplify the notation. In what Besides being invertible, the DTZT also preserves inner
follows, we assume that , where is the space products and norms:
of square-summable (finite-energy) discrete-time signals with
and (5)
inner product and norm
where the inner product and norm in the DTZT domain are
The DTZT at a fixed time index is the discrete-time defined as
Fourier transform of the signal derived by
subsampling by a factor of

In particular, the DTZT for evaluated at is the Two further sum/integral relations are
discrete-time Fourier transform

2) Frequency-Domain Expression: The DTZT can be ex-

pressed in terms of (see above) as 5) Unitarity: The DTZT is a unitary mapping from
onto , which is the space of functions
defined for and and satisfying

In the remainder of this section, we shall discuss important

DTZT properties and relations, such as periodicity properties,
relations involving integrals and/or sums of the DTZT, and The unitarity of the DTZT has two important consequences:
shift properties. We shall show that the DTZT is a unitary • The DTZT is a one-to-one mapping. In particular, to any
mapping, i.e., a one-to-one mapping that preserves norms and , we can find a unique
inner products. We shall furthermore discuss situations where signal such that This
the symmetry of a signal implies that its DTZT has one or signal can be obtained according to the inversion formula
more zeros. (The zeros of the DTZT play a fundamental role (3) as
in WHF theory; see Section III-B.) • The DTZT preserves inner products and norms [see (5)].

Authorized licensed use limited to: ETH BIBLIOTHEK ZURICH. Downloaded on January 25, 2010 at 04:09 from IEEE Xplore. Restrictions apply.
BÖLCSKEI AND HLAWATSCH: DISCRETE ZAK TRANSFORMS, POLYPHASE TRANSFORMS, AND APPLICATIONS 853

TABLE I
SIGNAL TRANSFORMATION PROPERTIES OF THE DTZT

Proof of Unitarity: Following [45, pp. 109–110], it suffices For odd and even or odd, we have
to show that the DTZT maps an orthonormal basis of to The significance of the zeros of the DTZT
an orthonormal basis of Let in the context of WHF theory will be discussed in Section
for and otherwise. The DTZT of III-B.
is on the fundamental rectangle. The set 8) Time-Limited Signals: If is zero outside the interval
of signals with and , then
is easily shown to constitute an orthonormal
for
basis for The DTZT of is given by
It is If is zero outside the interval , then
easily checked that the functions constitute an
orthonormal basis for
for
6) Product of Two DTZT’s: With (2), the DTZT product
is -periodic with respect to time
and 1-periodic with respect to frequency Therefore, This formula can be considered to be a discretized version of
can be represented by a 2-D Fourier series (3).
expansion, which can be shown to read as follows: 9) Bandlimited Signals: If is bandlimited to the fre-
quency band , i.e., for ,
then

(6)
with . We note that the Fourier
series coefficients If is bandlimited in an interval , where
, then

for
are samples of the discrete-time short-time Fourier transform
of using the analysis window [46], [47] (or, equiv-
alently, samples of the discrete cross-ambiguity function of This is a discretized version of (4).
and [33]). By inversion of (6), we obtain 10) Signal Transformation Properties, Symmetry Relations,
and Transform Pairs: Table I lists further important DTZT re-
lations that show how the DTZT behaves under various signal
transformations. Table II summarizes symmetry relations of
the DTZT, and Table III provides some DTZT transform pairs.
7) Zeros of the DTZT: When is even and is even, In these tables, we have ;
i.e., , then the DTZT has a zero in the center of furthermore, is the unit step function, is the discrete-
the fundamental rectangle, When is even time delta function (or unit sample), is the Dirac delta
and is odd, i.e., , then function, and sinc

Authorized licensed use limited to: ETH BIBLIOTHEK ZURICH. Downloaded on January 25, 2010 at 04:09 from IEEE Xplore. Restrictions apply.
854 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 45, NO. 4, APRIL 1997

TABLE II
SYMMETRY RELATIONS OF THE DTZT

TABLE III
DTZT TRANSFORM PAIRS. (THE DTZT EXPRESSIONS ARE VALID FOR

III. DISCRETE-TIME GABOR EXPANSIONS A. Discrete-Time Weyl–Heisenberg Frames

The discrete-time Gabor expansion is an important time- The theory of Weyl-Heisenberg frames (WHF’s) [45], [49],
frequency signal expansion that has been studied, e.g., in [8], [50] yields important results on the Gabor expansion. For
[10], [12], [33], [48], and [20] and extended to the multiwin- , a set of functions
dow case in [16]. This section discusses the application of the with and is said to be a WHF
DTZT to the discrete-time Gabor expansion [30], [19], [31]. for if for all
After a review of the Zibulski-Zeevi method for calculating
dual frames [12], we show that the DTZT diagonalizes the
Gabor (or Weyl–Heisenberg) frame operator in the cases of
critical sampling and integer oversampling and thereby allows
with (9)
a simple calculation of the dual Gabor window, the frame
bounds, and tight Gabor windows [12], [16], [31].
The constants and are called frame bounds. For
The discrete-time Gabor expansion of a signal
the continuous-time case, analytic expressions for the frame
is defined as
bounds of some specific window functions can be found in
[51]. For synthesis window such that is a WHF,
(7) the Gabor expansion (7), (8) exists for all , and
an analysis window (or “dual” window) can be derived from
as
with , where are the Gabor
(10)
coefficients, is a suitably chosen “synthesis window,” and
with are the grid parameters. The Gabor Here, is the inverse of the frame operator defined as
coefficients can be calculated as

(8)

The frame operator is a linear, positive definite operator

where with a suitable “analysis (corresponding to a positive definite matrix of infinite size)
window” In the cases of oversampling and mapping onto If is a WHF with frame
critical sampling , the Gabor expansion (7), (8) exists bounds and , then is a WHF as well (the “dual”
for arbitrary if the windows and are frame), with frame bounds and [45], [49],
chosen properly (see below). In the case of undersampling [50]. The numerical properties of the Gabor expansion will be
, the set is incomplete in , and hence, better for closer frame bounds and A WHF is called snug
it is not possible to expand every signal into the if and tight if For a tight WHF, , where
(It is well known that this observation also holds for is the identity operator on , and hence, there is simply
more general signal expansions; see, for example, [27] and From a synthesis window generating a
[29].) WHF, we can derive a “tight” synthesis window (i.e., a

Authorized licensed use limited to: ETH BIBLIOTHEK ZURICH. Downloaded on January 25, 2010 at 04:09 from IEEE Xplore. Restrictions apply.
BÖLCSKEI AND HLAWATSCH: DISCRETE ZAK TRANSFORMS, POLYPHASE TRANSFORMS, AND APPLICATIONS 855

window generating a tight WHF with frame bound ) as and normalized frequency (the continuous index reflects
that the spectrum of is continuous). The eigenfunctions are
(11) related to the DTZT as2
where is the inverse of the positive definite operator (15)
square root of
A WHF with critical sampling is an exact frame, where The eigenvalues are related to the DTZT
i.e., the are linearly independent [45], [49], [50]. Here, as (see the Appendix).
both the analysis window and the Gabor coefficients We see that the DTZT of the Gabor synthesis window
are uniquely determined. determines the eigenvalues of the Gabor frame operator. The
dependence of on will be emphasized by writing
B. Zibulski–Zeevi Method and Diagonalization the eigenvalues as [see (13)]. For critical sampling
of the Gabor Frame Operator , (13) simplifies to
We note that a second complete and orthogonal set of
We shall now briefly review the Zibulski–Zeevi method for
eigenfunctions of is
the calculation of dual frames [12]. This method is based on the
observation that the Gabor frame operator can be expressed with These eigenfunctions
in the DTZT domain as lead to the “dual DTZT”

whose fundamental rectangle is

The dual DTZT satisfies
and with Similarly to (13),
(12) we get

where with Based on (12), it can with

be shown that the inversion of the frame operator (i.e., the
calculation of the dual frame) reduces to the inversion of
a matrix of finite size. This method can be applied in the
cases of rational oversampling and integer It can furthermore be shown that This
oversampling For integer oversampling, including interesting duality relation has also been observed in [12]
critical sampling as a special case, (12) simplifies to and [54]. For critical sampling and
with
The importance of the DTZT for the Gabor expansion is
based on the above relations of the DTZT with the eigenvalues
(13)
and eigenfunctions of the Gabor frame operator. Taking the
inner product of some signal with both sides of
Hence, the frame operator becomes a simple multiplication the eigenequation , we obtain
operator in the DTZT domain, i.e., the DTZT “diagonalizes” , and further,
the Gabor frame operator in the case of integer oversampling. , where we have used (i.e., is
This diagonalization can be interpreted in terms of the self-adjoint) with denoting the adjoint of [52]. Thus, we
(generalized) eigenvalues and (generalized) eigenfunctions have rederived (13) using (15).
of the Gabor frame operator , which are defined We shall now summarize important consequences of the
by1 [53]. For critical sampling or integer diagonalization discussed above [12].
oversampling with , we show in the • The inverse frame operator can be expressed in the
Appendix that a set of generalized eigenfunctions of that DTZT domain as a pointwise division
is orthogonal and complete (in a generalized sense [53]) is
with (16)

(14) • The frame condition (9) can be reformulated, using the

DTZT of the synthesis window , as

(this result has previously been stated without proof in [31]). with (17)
Note that the eigenfunctions are parameterized by time
In addition, the frame bounds are given by the
1 The operator is not compact and has a continuous spectrum [52]. The infimum and the supremum, respectively, of In
are not in and are hence called generalized eigenfunctions. A
2 Strictly speaking, this inner product is not well defined since the
rigorous treatment of generalized eigenexpansions of such operators requires
the theory of Gel’fand triplets of spaces [53] and is beyond the scope of this are not in It must be interpreted as the linear functional assigning to
paper. the signal the value of its DTZT at

Authorized licensed use limited to: ETH BIBLIOTHEK ZURICH. Downloaded on January 25, 2010 at 04:09 from IEEE Xplore. Restrictions apply.
856 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 45, NO. 4, APRIL 1997

particular, is a tight WHF with frame bound transform to DFT filter banks with critical sampling or integer
if and only if oversampling is discussed in Section V.
• With (16), the analysis window in (10) can be 1) Definition and Interpretation: In analogy to the -trans-
calculated via the DTZT using form, which is the extension of the discrete-time Fourier
transform to the complex -plane [55], we define the polyphase
(18) transform (PT) by extending the DTZT in (1) to the -plane.
Formally replacing by , we obtain
and deriving according to (3), i.e.,
(21)
• From a synthesis window generating a WHF, a
“tight” synthesis window generating a tight WHF
with frame bound can be derived using [cf. (11)] We use a script to distinguish the PT from the DTZT. We
shall usually suppress the superscript
(19) The PT is a discrete-time version of the modified -
transform [56], [57]. At a fixed-time , the PT is
Equations analogous to (16)–(18) have first been reported the -transform of the signal derived by
by Daubechies et al. in [49] and by Zibulski and Zeevi in subsampling by a factor of :
[12]. We furthermore note that the frame operator of the
dual frame is with eigenvalues
(assuming appropri-
ate ordering of the eigenvalues). Hence, the DTZT of the
synthesis window can be obtained from the DTZT of In particular, the PT for evaluated at is the
the analysis window as -transform of :
In the case of critical sampling , the above relations
simplify to
Furthermore, with , where , the PT is seen to
be essentially the DTZT of :
(20)

with The DTZT frame condition

has an interesting implication. In the continuous-time case, DTZT
the Balian–Low theorem [45] states that only synthesis win-
dows that are either not very smooth or have poor decay In the following, we shall discuss important properties and
generate a critically sampled WHF; this, in particular, excludes relations of the PT.
the Gaussian window [49]. In the discrete-time case, i.e., for 2) -Transform Expression and Region of Convergence:
WHF’s in , Heil [30] has shown that the situation is The PT can be expressed in terms of as
different. For even, the DTZT of the discrete-time (sampled)
Gaussian has a zero; therefore, in (20)
(22)
is not satisfied, and the corresponding Weyl–Heisenberg set
is not a WHF for However, for odd,
the DTZT of the sampled Gaussian does not have a zero, This allows us to determine the region of convergence (ROC)
and therefore, is a WHF for (even though of the PT. If the -transform of converges for
must still be expected to be close to ). This shows that the with , then it follows from (22) that the
Balian–Low theorem does not carry over to the discrete-time PT’s ROC is If converges at or at 0,
case in a straightforward manner. the same need not be true for due to the factor
in (22). If the PT converges on the unit circle, then the DTZT
IV. THE POLYPHASE TRANSFORM exists and equals the PT evaluated on the unit circle, i.e.
The polyphase decomposition introduced in [36] is of fun-
damental importance in filter bank theory. In [37]–[44], the
application of the polyphase decomposition to the analysis 3) Periodicity: The PT satisfies the “quasiperiodicity” re-
and design of filter banks is discussed. In this and the next lation
section, we shall show that the extension of the DTZT to
the -plane yields the transform underlying the polyphase
decomposition. Fundamental properties of this transform will
be studied, thereby establishing known and new results on the Therefore, it suffices to consider the PT on the fundamental
polyphase decomposition. The application of the polyphase time interval

Authorized licensed use limited to: ETH BIBLIOTHEK ZURICH. Downloaded on January 25, 2010 at 04:09 from IEEE Xplore. Restrictions apply.
BÖLCSKEI AND HLAWATSCH: DISCRETE ZAK TRANSFORMS, POLYPHASE TRANSFORMS, AND APPLICATIONS 857

4) Integral/Sum Relations: We next summarize important such that This signal

relations involving integrals and/or sums of the PT. The signal can be obtained according to the inversion formula (23)
can be recovered from its PT via a contour integral in the as
-plane: • The PT preserves inner products and norms [see (24)].
Proof of Unitarity: Analogously to the DTZT case consid-
(23) ered in Section II, it can be shown that the PT maps the
orthonormal basis for given by
The contour of integration is assumed to be counterclockwise,
with and (where, again,
to enclose the origin, and to lie entirely in the ROC of
for and otherwise) onto the
If the contour of integration is the unit circle,
orthonormal basis for given by
then (23) can be shown to simplify to the DTZT inversion
formula (3). We can also recover the signal’s -transform from
: 6) Product of Two PT’s: The PT product
can be expressed as

This is equivalent to the polyphase decomposition of

[28] (see Section V). (25)
The PT also preserves inner products and norms with . By inversion of (25),
we obtain
and (24)
where the inner product and norm in the PT domain are defined
as3

7) Time-Limited Signals: If is zero outside the interval

, then

for

If is zero outside the interval , then

Two further sum/integral relations are

for

8) Initial Value Theorem: If is a causal sequence, i.e.,

for , then

for

9) Reverse PT: We define the reverse PT as the PT with

reversed time index [38]:
5) Unitarity: The PT is a unitary mapping from onto
, which is the space of functions
defined on (with the set of complex numbers) (26)
and satisfying

We note that The reverse PT

will be used in Section V.
10) Signal Transformation Properties, Symmetry Relations,
The PT’s unitarity has two important consequences: and Transform Pairs: Table IV lists PT relations involving
• The PT is a one-to-one mapping. In particular, to any various signal transformations. We note that the ROC of
, we can find a unique signal is given by the intersection of the
3 The ROC’s of and Table V summarizes symme-
paraconjugate of is defined as
[37]. Note that on the unit circle paraconjugation reduces to try relations, and Table VI provides some basic PT transform
conjugation, i.e., pairs.

Authorized licensed use limited to: ETH BIBLIOTHEK ZURICH. Downloaded on January 25, 2010 at 04:09 from IEEE Xplore. Restrictions apply.
858 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 45, NO. 4, APRIL 1997

TABLE IV
SIGNAL TRANSFORMATION PROPERTIES OF THE PT

TABLE V
SYMMETRY RELATIONS OF THE PT

TABLE VI
PT TRANSFORM PAIRS. (THE PT EXPRESSIONS ARE VALID FOR

V. DFT FILTER BANKS The transfer functions ( -transforms) of the analysis

We now discuss the application of the PT to DFT filter and synthesis filters are frequency-translated versions of the
banks with critical sampling or integer oversampling. Our transfer functions and of baseband “prototype”
discussion essentially follows [37], [41]–[44]. Although some analysis and synthesis filters, respectively, i.e.
of the results of this section can be extended to more general
filter banks [27], [29], we shall restrict our attention to DFT
filter banks for the sake of coherence with Section III. (For with Equivalently, the filters’ impulse
a detailed discussion of the relations between filter banks and responses are
frame theory, see [58]–[62], [27]–[29], and [54].) Based on
the equivalence of DFT filter banks and WHF’s [28], [59],
[63], we shall also reformulate some of the DTZT relations of The reconstructed signal is given by
Section III in terms of the PT.
(27)
A. Perfect Reconstruction DFT Filter Banks
We consider an -channel, perfect reconstruction (PR), with the subband signals
DFT filter bank with zero delay, such that (see
Fig. 1). We assume integer oversampling or critical sampling, (28)
i.e., , where is the number of channels, is
the decimation factor, and is the oversampling factor where , and stands
[37]–[44], [58]–[60], [27], [29]. Note that in the critical case, for complex conjugation. Setting and

Authorized licensed use limited to: ETH BIBLIOTHEK ZURICH. Downloaded on January 25, 2010 at 04:09 from IEEE Xplore. Restrictions apply.
BÖLCSKEI AND HLAWATSCH: DISCRETE ZAK TRANSFORMS, POLYPHASE TRANSFORMS, AND APPLICATIONS 859

where is the PT of as defined in (21). The PT’s

of the filters and are related to the PT’s of the Gabor
windows and , respectively, as

The DTZT relations (13) and (16)–(19) formulated in Section

III in a Gabor expansion setting can now be reformulated in
terms of the PT, using the analysis and synthesis prototypes
Fig. 1. -channel DFT filter bank with subsampling factor and This amounts to extending
the DTZT relations from the unit circle to the entire -plane.
• In the PT domain, the effect of the frame operator
using the PR property, we obtain
or inverse frame operator becomes a simple
multiplication or division, respectively [28], [60]

Hence, the PR DFT FB provides an expansion of the input

signal into the synthesis functions
gener- with
ated by The expansion coefficients ( subband signals)
are obtained by projecting onto the
dual functions generated
by Comparing the Gabor synthesis and
analysis formulas (7) and (8) with the filter bank synthesis where Note that the eigenvalues
and analysis formulas (27) and (28), respectively, we see that in (13) are evaluated on the unit circle, i.e.
these formulas are fully equivalent if the Gabor analysis and
synthesis windows are related to the analysis and synthesis • In the context of DFT filter banks, the frame property is
prototype filters, respectively, according to important as well since it guarantees a certain degree of
numerical stability (as characterized by the frame bounds
and of the underlying synthesis WHF) [27], [29].
In this case, the Gabor coefficients equal the filter bank sub-
The frame condition (17) can be reformulated in terms
band signals up to a phase factor
of the PT as
Hence, the Gabor expansion (or, equivalently, WHF analy-
sis/synthesis) can be interpreted as a PR DFT filter bank4 [59], with
[60], [28], [63]. Critically sampled DFT FB’s correspond to and the frame bounds and are the infimum and
orthogonal or biorthogonal WHF’s, whereas oversampled DFT supremum, respectively, of In particular, an
FB’s correspond to redundant (overcomplete) WHF’s. integer oversampled or critically sampled DFT FB is
paraunitary [37] if and only if
B. Application of the Polyphase Transform
for (29)
The polyphase decomposition is a well established and
important tool in filter bank theory [37]. The polyphase Note that a DFT filter bank is paraunitary if and only if
decomposition of the analysis prototype reads the underlying WHF is tight [59], [28], [58], [27], [29].
• From a given synthesis prototype generating a WHF,
the analysis prototype with minimum norm is given
by [60]
where is the reverse PT of as defined in
(26). Similarly, the polyphase decomposition of the synthesis
prototype reads5

In the paraunitary (tight) case, there is simply

• For a given synthesis prototype generating a WHF,

4 There exists, however, the following basic difference: In filter bank a “paraunitary” synthesis prototype (generating a
theory, one usually specifies the analysis prototype and computes the paraunitary DFT filter bank or, equivalently, a tight WHF)
corresponding synthesis prototype for PR. In Gabor expansion (or WHF) is obtained using
theory, on the other hand, the synthesis window is first specified,
and then, the corresponding analysis window is computed.
For reasons of consistency with Section III, we shall here adopt the Gabor
expansion viewpoint for filter banks as well.
5 In the filter bank literature, the polyphase filter transfer functions where is defined by and
and are often denoted by and , respectively [37].

Authorized licensed use limited to: ETH BIBLIOTHEK ZURICH. Downloaded on January 25, 2010 at 04:09 from IEEE Xplore. Restrictions apply.
860 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 45, NO. 4, APRIL 1997

In the critical case , we have function defined on the fundamental rectangle

, and the above relations simplify The number of DZT samples
accordingly. In particular, (29) reduces to is , which equals the number of signal samples. For a signal
for , which implies of given length (or period) , the DZT parameters must
Hence, the polyphase filters are allpass filters, be chosen such that We shall usually omit the
and thus, the design of critically sampled paraunitary DFT superscript in
FB’s reduces to finding an arbitrary set of allpass filters. The DZT at a fixed time index is, up to a factor
For , a paraunitary DFT FB with frame bound can , the -point discrete Fourier transform (DFT) [55] of
be constructed by choosing the polyphase filters such that the the signal :
following power symmetry condition [37] holds:

for
DFT
A detailed treatment of the relation between WHF’s and
DFT FB’s is given in [59], [60], and [28]. We note that the
In particular, the DZT for and evaluated at
PT can be applied to the more general class of uniform filter
is the DFT of up to a factor:
banks as well [37]–[39], [58], [59], [27], [29].

VI. THE DISCRETE ZAK TRANSFORM

For numerical implementations, a ZT version that is discrete
in both time and frequency is required. This transform, called with DFT
discrete Zak transform (DZT), has been introduced in [33] and Thus, in order to compute the DZT for , we
[34]. The DZT allows an efficient, FFT-based implementation have to compute DFT’s of length This can be done
of Gabor analysis and synthesis [12], [33], [34], [16], [10]. efficiently using an FFT algorithm. Specifically, the DZT is
1) Definition and Interpretation: If we sample the DTZT obtained by computing the DFT’s (or FFT’s) of the columns
in (1) with respect to frequency at uniformly of the matrix
spaced sampling points and include the normaliza-
tion factor , we obtain

.. .. .. ..
. . . .

2) Frequency-Domain Expression: The DZT can be ex-

pressed in terms of the -point DFT (or, equivalently, discrete
From , we can recover only a periodized version Fourier series) as
of the signal :

with . The periodized signal We next study fundamental properties of the DZT.
will contain aliasing in time unless the signal has finite We assume , where is the space of
length Specifically, let us assume that is zero -periodic discrete-time signals with inner product
outside the interval In this case, for and norm
Furthermore, for As before, we assume that the DZT
can now be expressed by the finite sum parameters are chosen such that
3) Periodicity: Let us, for the moment, extend the DZT
definition (30) to and (note, however, that is
still assumed -periodic). The extended DZT is quasiperiodic
(30) in and periodic in , i.e.

This is the final definition of the DZT of a finite-length

signal defined for Equivalently, we with
may assume that is periodic with period Thus,
the DZT is a linear signal transformation that maps a signal Therefore, it is indeed sufficient to know the DZT on the
with length (or an -periodic signal) to a 2-D fundamental rectangle

Authorized licensed use limited to: ETH BIBLIOTHEK ZURICH. Downloaded on January 25, 2010 at 04:09 from IEEE Xplore. Restrictions apply.
BÖLCSKEI AND HLAWATSCH: DISCRETE ZAK TRANSFORMS, POLYPHASE TRANSFORMS, AND APPLICATIONS 861

4) Sum Relations: Next, we summarize relations involving 6) Product of Two DZT’s: The DZT product
sums of the DZT. The signal can be recovered by allows the 2-D DFT representation
summing the DZT over the frequency index :

(31)

We can furthermore recover the -point DFT according with Conversely

to

(32) 7) Zeros of the DZT: For odd and even, the DZT will
assume characteristic zeros in the case of signals satisfying
Inserting the DZT definition (30), we obtain certain symmetry properties: If , then
, and if , then

8) Time-Limited Signals: If has finite time support

inside the fundamental period , i.e.,
Here, the -point DFT is expressed in terms of DFT’s for , then
of length This is the relation underlying the
decimation-in-time FFT algorithm [55]. for
Besides being invertible as seen above, the DZT also
preserves inner products and norms:
If is zero outside the interval with ,
and (33) then

where the inner product and norm in the DZT domain are for
defined as
This formula can be viewed as a subsampled version of (31).
9) Bandlimited Signals: If the DFT of is limited
to the interval , i.e., for ,
then

for
Two further sum relations are
If is limited to the interval with ,
then

for

5) Unitarity: The DZT is a unitary mapping from onto This is a subsampled version of (32).
, where denotes the space of square-summable 10) Signal Transformation Properties, Symmetry Relations,
functions defined for and and Transform Pairs: The effect of various signal transfor-
with inner product mations on the DZT is summarized in Table VII. We note
The unitarity of the DZT can be proved in a similar manner as that and that all signals
the unitarity of the DTZT (see Section II). It has two important are assumed -periodic. Table VIII summarizes symmetry
consequences. relations of the DZT, and Table IX provides some basic DZT
• The DZT is a one-to-one mapping. In particular, to transform pairs.
any , we can find a unique signal
such that This signal VII. CYCLIC GABOR EXPANSIONS
is obtained according to the inversion formula (31) as
We now discuss the application of the DZT to the cyclic (or
“discrete”) Gabor expansion [8], [10], [33], [20]. This version
• The DZT preserves inner products and norms [see (33)]. of the Gabor expansion assumes that all signals and windows

Authorized licensed use limited to: ETH BIBLIOTHEK ZURICH. Downloaded on January 25, 2010 at 04:09 from IEEE Xplore. Restrictions apply.
862 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 45, NO. 4, APRIL 1997

TABLE VII
SIGNAL TRANSFORMATION PROPERTIES OF THE DZT

TABLE VIII
SYMMETRY RELATIONS OF THE DZT

TABLE IX
DZT TRANSFORM PAIRS. (THE SIGNAL EXPRESSIONS ARE VALID FOR AND THE DZT EXPRESSIONS ARE VALID FOR )

have finite length (or, equivalently, they are -periodic). only for , although the DZT can also be applied to cyclic
The cyclic Gabor expansion of a signal is given by Gabor expansions with rational oversampling [12], [16], [54].

A. Cyclic Weyl–Heisenberg Frames

(34)
The theory of frames (cf. Section III-A) can also be applied
to the cyclic Gabor expansion. For a detailed discussion of
with with a suitable synthesis cyclic WHF’s, see, for example, [10]. A set of functions
window and The parameters and are
related to the signal length (period) and the oversampling is said to be a WHF for if for all
factor as , i.e., the number of Gabor
coefficients is times the number of signal samples
With , we get Using a suitable
analysis window , the Gabor coefficients can be with (36)
calculated as
where For synthesis window such
that is a WHF, the cyclic Gabor expansion (34), (35)
(35)
exists for all , and an analysis (or dual) window
can be derived from as
with (37)
In the following, we discuss the cases of critical sampling or
integer oversampling, i.e., This restriction is motivated where the frame operator is defined as
by the fact that the DZT diagonalizes the Gabor frame operator Note that the frame operator

Authorized licensed use limited to: ETH BIBLIOTHEK ZURICH. Downloaded on January 25, 2010 at 04:09 from IEEE Xplore. Restrictions apply.
BÖLCSKEI AND HLAWATSCH: DISCRETE ZAK TRANSFORMS, POLYPHASE TRANSFORMS, AND APPLICATIONS 863

corresponds to a matrix of size Thus, calculation and inverting according to (31)

of the dual window necessitates the inversion of an
matrix, which is computationally intensive for large signal • From a synthesis window generating a WHF, a tight
lengths Fast algorithms for the calculation of the dual window (with frame bound ) can be derived
Gabor window and the Gabor coefficients are discussed in using
[21], [31], and [64]–[68].
For a tight WHF (where , we have , where (41)
is the identity operator on , and hence, there is simply
From a synthesis window generating
In the case of critical sampling , the above relations
a WHF, we can derive a “tight” synthesis window
simplify to
(generating a tight WHF) as
(38)
where is the inverse of the positive definite operator
(matrix) square root of

B. Diagonalization of the Gabor Frame Matrix VIII. SIMULATION RESULTS

In the cases of critical sampling or integer oversampling, We now present simulation results demonstrating the appli-
i.e., , an orthogonal and complete set of eigenfunctions cation of the DZT to the calculation of dual Gabor windows,
(or eigenvectors) of the Gabor frame operator (or matrix) frame bounds, and tight Gabor windows within the framework
is given by of the cyclic Gabor expansion (see Section VII). Fig. 2(a)
shows a Gaussian synthesis window with period
The corresponding dual Gabor window for over-
with , where sampling factor and grid parameters
The DZT is related with these eigenfunctions as is depicted in Fig. 2(b). Straightforward calculation of
according to (37) would require the inversion of the
frame matrix, which is on the order of (recall
that , followed by a matrix-vector multiplication
and the eigenvalues of can be written in terms of the DZT
(on the order of ). In contrast, the DZT-based method
of as6
according to (40) requires the calculation of the DZT
FFT’s of length , which is on the order of
(39) divisions, and DZT inversion
according to (31) FFT’s of length , which is on the order
(the proof of these relations is similar to that in the Appen- of
dix). It again follows that the DZT diagonalizes the Gabor Fig. 2(c) shows the eigenvalues of the frame matrix that
frame matrix (cf. Section III-B), which has the following were calculated using the DZT according to (39). The frame
consequences [12], [16]: bounds and (which characterize important numerical
• In the DZT domain, the effect of the Gabor frame matrix properties of the Gabor expansion and equal the minimum and
or the inverse Gabor frame matrix becomes a maximum, respectively, of the eigenvalues) were obtained as
simple multiplication or division, respectively: Finally, Fig. 2(d) shows the tight Gabor
synthesis window calculated via the DZT according to
(41). Note that direct calculation of the tight Gabor window
according to (38) would require the calculation of the inverse
square root of the frame matrix.
• The frame condition (36) can be reformulated in terms of
the DZT of as IX. CONCLUSION
We considered three different versions of the Zak trans-
with
form (ZT) for discrete-time signals, emphasizing fundamental
The frame bounds and are the minimum and maxi- properties and relations. We also discussed the important role
mum, respectively, of In particular, is played by these transforms in the theory of Gabor expansions
a tight WHF for if and only if (Weyl–Heisenberg frames) and in the related field of DFT filter
• The Gabor analysis window can be computed using banks. This role is due to the fact that the ZT diagonalizes
fundamental operators and matrices in the cases of critical
(40) sampling and integer oversampling. We note that the ZT can
also be applied to rationally oversampled Weyl–Heisenberg
6 Note that and are ordinary (i.e., not generalized, cf. frames [12], [16], [69] and DFT filter banks [60], as well as
Section III-B) eigenvalues and eigenfunctions, respectively. to a more general class of frame expansions and filter banks

Authorized licensed use limited to: ETH BIBLIOTHEK ZURICH. Downloaded on January 25, 2010 at 04:09 from IEEE Xplore. Restrictions apply.
864 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 45, NO. 4, APRIL 1997

(a similar expression has been found for the continuous-

time Gabor expansion in [12] and [70]). We now restrict
our attention to the cases of integer oversampling or critical
sampling, i.e., with Inserting
[see (14)]
in (42), we obtain

(a) (b)

(c) (d)
Fig. 2. DZT-based calculation of the dual (analysis) window, the eigenvalues
and frame bounds, and the tight synthesis window. (a) Gaussian synthesis win-
dow with period . (b) Corresponding dual Gabor window
for oversampling factor and grid parameters
(c) Eigenvalues (sorted in ascending order) of the frame matrix (d) Tight
synthesis window

[27], [29], even though it does not perform a diagonalization

in these cases.
We have furthermore shown that the extension of the
discrete-time ZT to the complex -plane is equivalent to the
polyphase representation traditionally used in multirate signal
processing and filter bank theory. It is interesting that the
essentially equivalent concepts of polyphase representation
and ZT have been developed independently in filter bank
theory and Gabor expansion theory, respectively.
Finally, simulation results have demonstrated the application
of the discrete ZT to the efficient calculation of dual windows,
tight windows, and frame bounds for critically sampled and
integer oversampled, cyclic Gabor expansions. This shows that is an eigenfunction of with eigen-
value
APPENDIX We shall next prove that the are orthogonal and
EIGENVALUES AND EIGENFUNCTIONS complete in a generalized sense [53]. According to [53, p.
OF THE GABOR FRAME OPERATOR
186], one has to show that for all

We recall from Section III-B that the Gabor frame operator

is defined as and

Interchanging the order of summations and using the Poisson where The first equation follows from
sum formula , it (5), and the second equation is easily verified to be equivalent
is possible to derive the following alternative expression: to (3).

ACKNOWLEDGMENT
The authors would like to thank Dr. A. J. E. M. Janssen for
(42) valuable suggestions that led to an improvement of this paper.

Authorized licensed use limited to: ETH BIBLIOTHEK ZURICH. Downloaded on January 25, 2010 at 04:09 from IEEE Xplore. Restrictions apply.
BÖLCSKEI AND HLAWATSCH: DISCRETE ZAK TRANSFORMS, POLYPHASE TRANSFORMS, AND APPLICATIONS 865

They are furthermore grateful to Dr. S. J. L. van Eijndhoven [27] H. Bölcskei, F. Hlawatsch, and H. G. Feichtinger, “Frame-theoretic
and Prof. H. G. Feichtinger for illuminating discussions and to analysis of filter banks,” submitted to IEEE Trans. Signal Processing,
Feb. 1996.
Prof. W. Mecklenbräuker for drawing their attention to [56] [28] , “Equivalence of DFT filter banks and Gabor expansions,” SPIE
and [57]. Proc., vol. 2569, part I, Wavelet Applications in Signal and Image
Processing III. San Diego, CA: SPIE, July 1995, pp. 128–139.
[29] , “Frame-theoretic analysis and design of oversampled filter
banks,” Proc. IEEE ISCAS-96, Atlanta, GA, May 1996, pp. 409–412,
REFERENCES vol. 2.
[30] C. Heil, “A discrete Zak transform,” Tech. Rep. MTR-89W000128,
[1] J. Zak, “Finite translations in solid state physics,” Phys. Rev. Lett., vol. MITRE Corp., Danvers, MA, 1989.
19, pp. 1385–1397, 1967. [31] H. Bölcskei, H. G. Feichtinger, and F. Hlawatsch, “Diagonalizing
[2] A. J. E. M. Janssen, “The Zak transform: A signal transform for sampled the Gabor frame operator,” Proc. IEEE UK Symp. Applications Time-
time-continuous signals,” Philips J. Res., vol. 43, pp. 23–69, 1988. Frequency Time-Scale Methods, Univ. Warwick, Coventry, U.K., Aug.
[3] , “The Zak transform and some counterexamples in time- 1995, pp. 249–255.
frequency analysis,” IEEE Trans. Inform. Theory, vol. 38, pp. 168–171, [32] J. J. Benedetto, C. Heil, and D. F. Walnut, “Differentiation and the
Jan. 1992. Balian–Low theorem,” J. Fourier Anal. Applications, vol. 1, no. 4, pp.
[4] H. Bacry, A. Grossmann, and J. Zak, “Proof of completeness of lattice 355–403, 1995.
states in the kq representation,” Phys. Rev. B, vol. 12, no. 4, pp. [33] L. Auslander, I. C. Gertner, and R. Tolimieri, “The discrete Zak
1118–1120, Aug. 1975. transform application to time-frequency analysis and synthesis of nonsta-
[5] M. Boon, J. Zak, and J. Zucker, “Rational von Neumann lattices,” J. tionary signals,” IEEE Trans. Signal Processing, vol. 39, pp. 825–835,
Math. Phys. vol. 24, no. 2, pp. 316–323, Feb. 1983. Apr. 1991.
[6] D. Gabor, “Theory of communication,” J. Inst. Elec. Eng., vol. 93, no. [34] Y. Y. Zeevi and I. C. Gertner, “The finite Zak transform: An efficient
3, pp. 429–457, 1946. tool for image representation and analysis,” J. Visual Commun. Image
[7] M. J. Bastiaans, “Gabor’s expansion of a signal into Gaussian elemen- Represent., vol. 3, pp. 13–23, Mar. 1992.
tary signals,” Proc. IEEE, vol. 68, pp. 538–539, Apr. 1980. [35] R. S. Orr, “Derivation of the finite discrete Gabor transform by peri-
[8] J. Wexler and S. Raz, “Discrete Gabor expansions,” Signal Processing, odization and sampling,” Signal Processing, vol. 34, pp. 85–97, 1993.
vol. 21, pp. 207–220, 1990. [36] M. G. Bellanger, G. Bonnerot, and M. Coudreuse, “Digital filtering
[9] A. J. E. M. Janssen, “Duality and biorthogonality for Weyl-Heisenberg by polyphase network: Application to sample-rate alteration and filter
frames,” J. Fourier Anal. Applications, vol. 1, no. 4, pp. 403–437, 1995. banks,” IEEE Trans. Acoust., Speech, Signal Processing, vol. ASSP-24,
[10] , “Duality and biorthogonality for discrete-time Weyl-Heisenberg pp. 109–114, Apr. 1976.
frames,” Nat. Lab. Rep. 2/94, RWR-518-RE-94001-ak, Philips Nat. Lab. [37] P. P. Vaidyanathan, Multirate Systems and Filter Banks. Englewood
Eindhoven, The Netherlands, 1994. Cliffs, NJ: Prentice-Hall, 1993.
[11] I. Daubechies, H. Landau, and Z. Landau, “Gabor time-frequency [38] M. Vetterli and J. Kovačević, Wavelets and Subband Coding. Engle-
lattices and the Wexler-Raz identity,” J. Fourier Anal. Applications, vol. wood Cliffs, NJ: Prentice-Hall, 1995.
1, no. 4, pp. 437–478, 1995. [39] A. N. Akansu and R. A. Haddad, Multiresolution Signal Decomposition.
[12] M. Zibulski and Y. Y. Zeevi, “Oversampling in the Gabor scheme,” New York: Academic, 1992.
IEEE Trans. Signal Processing, vol. 41, pp. 2679–2687, Aug. 1993. [40] M. Vetterli, “A theory of multirate filter banks,” IEEE Trans. Acoust.,
[13] , “Frame analysis of the discrete Gabor-scheme,” IEEE Trans. Speech, Signal Processing, vol. ASSP-35, pp. 356–372, Mar. 1987.
Signal Processing, vol. 42, pp. 942–945, Apr. 1994. [41] R. E. Crochiere and L. R. Rabiner, Multirate Digital Signal Processing.
[14] , “Analysis of multi-window Gabor-type schemes by frame meth- Englewood Cliffs, NJ: Prentice-Hall, 1983.
ods,” Tech. Rep. CC Pub. 101, Technion, Haifa, Israel, Apr. 1995. [42] K. Swaminathan and P. P. Vaidyanathan, “Theory and design of uniform
[15] , “Multi-window Gabor-type transform for signal representation DFT, parallel, quadrature mirror filter banks,” IEEE Trans. Circuits Syst.,
and analysis,” SPIE Proc., vol. 2569, Wavelet Applications in Signal and vol. CAS-33, pp. 1170–1191, Dec. 1986.
Image Processing III. San Diego, CA: SPIE, July 1995, pp. 116–127. [43] M. J. T. Smith and T. P. Barnwell, III, “A new filter bank theory
[16] , “Discrete multi-window Gabor-type transforms,” Tech. Rep., CC for time-frequency representation,” IEEE Trans. Acoust., Speech, Signal
Pub. 124, Technion, Haifa, Israel, Nov. 1995. Processing, vol. ASSP-35, pp. 314–327, Mar. 1987.
[17] H. G. Feichtinger and K. Gröchenig, “Gabor wavelets and the Heisen- [44] P. Vary, “On the design of digital filter banks based on a modified
berg group: Gabor expansions and short time Fourier transform from principle of polyphase,” AEÜ, vol. 33, no. 7/8, pp. 293–300, 1979.
the group theoretical point of view,” in Wavelets: A Tutorial in Theory [45] I. Daubechies, Ten Lectures on Wavelets. Philadelphia, PA: SIAM,
and Applications, C. K. Chui, Ed. New York: Academic, 1992, pp. 1992.
297–359. [46] M. R. Portnoff, “Time-frequency representations of digital signals and
[18] J. J. Benedetto and D. F. Walnut, “Gabor frames for and related systems based on short-time Fourier analysis,” IEEE Trans. Acoust.,
spaces,” in Wavelets: Mathematics and Applications, J. J. Benedetto and Speech, Signal Processing, vol. ASSP-28, pp. 55–69, Feb. 1980.
M. W. Frazier, Eds. Boca Raton, FL: CRC, 1993, pp. 97–162. [47] S. H. Nawab and T. F. Quatieri, “Short-time Fourier transform,” in
[19] M. J. Bastiaans, “Gabor’s expansion and the Zak transform for Advanced Topics in Signal Processing, J. S. Lim and A. V. Oppenheim,
continuous-time and discrete-time signals: Critical sampling and rational Eds. Englewood Cliffs, NJ: Prentice-Hall, 1988.
oversampling,” EUT Rep. 95-E-295, Tech. Univ. Eindhoven, Eindhoven, [48] D. F. Stewart, L. C. Potter, and S. C. Ahalt, “Computationally attractive
The Netherlands, Dec. 1995. real Gabor transforms,” IEEE Trans. Signal Processing, vol. 43, pp.
[20] S. Qian and D. Chen, “Discrete Gabor transform,” IEEE Trans. Signal 77–84, Jan. 1995.
Processing, vol. 41, pp. 2429–2438, July 1993. [49] I. Daubechies, A. Grossmann, and Y. Meyer, “Painless nonorthogonal
[21] R. S. Orr, “The order of computation for finite discrete Gabor trans- expansions,” J. Math. Phys. vol. 27, no. 5, pp. 1271–1283, May 1986.
forms,” IEEE Trans. Signal Processing, vol. 41, pp. 122–130, Jan. [50] C. E. Heil and D. F. Walnut, “Continuous and discrete wavelet trans-
1993. forms,” SIAM Rev., vol. 31, no. 4, pp. 628–666, Dec. 1989.
[22] R. Tolimieri and R. S. Orr, “Poisson summation, the ambiguity func- [51] A. J. E. M. Janssen, “Some frame bound calculations for Weyl-
tion, and the theory of Weyl-Heisenberg frames,” J. Fourier Anal. Heisenberg frames,” Indag. Mathem. N. S., vol. 7, no. 2, pp. 165–183,
Applications, vol. 1, no. 3, pp. 233–247, 1995. June 17, 1996.
[23] J. M. Morris and Y. Lu, “Discrete Gabor expansion of discrete-time [52] A. W. Naylor and G. R. Sell, Linear Operator Theory in Engineering
signals in via frame theory,” Signal Processing, vol. 40, pp. and Science, 2nd ed. New York: Springer, 1982.
155–181, 1994. [53] I. M. Gel’fand and G. E. Shilov, Generalized Functions, vol. 3. New
[24] S. Qian, K. Chen, and S. Li, “Optimal biorthogonal functions for York: Academic, 1967.
finite discrete-time Gabor expansion,” Signal Processing, vol. 27, pp. [54] A. J. E. M. Janssen, “The duality condition for Weyl-Heisenberg
177–185, 1992. frames,” to appear in Gabor Analysis: Theory, Algorithms, and Ap-
[25] A. J. E. M. Janssen, “The Zak transform and sampling theorems plications, H. G. Feichtinger and T. Strohmer, Eds. Boston, MA:
for wavelet subspaces,” IEEE Trans. Signal Processing, vol. 41, pp. Birkhäuser, June 1997.
3360–3364, Dec. 1993. [55] A. V. Oppenheim and R. W. Schafer, Discrete-time Signal Processing.
[26] J. W. M. Bergmans and A. J. E. M. Janssen, “Robust data equalization, Englewood Cliffs, NJ: Prentice-Hall, 1989.
fractional tap spacing and the Zak transform,” Philips J. Res., vol. 42, [56] R. H. Barker, “The pulse transfer function and its application to sampling
pp. 351–398, 1987. servo systems,” Proc. Inst. Elec. Eng., pt. IV, monograph 43, July 1952.

Authorized licensed use limited to: ETH BIBLIOTHEK ZURICH. Downloaded on January 25, 2010 at 04:09 from IEEE Xplore. Restrictions apply.
866 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 45, NO. 4, APRIL 1997

[57] E. I. Jury, Theory and Application of the Z-Transform Method. New [68] P. Prinz, “Calculating the dual Gabor window for general sampling sets,”
York: Wiley, 1964. IEEE Trans. Signal Processing, vol. 44, pp. 2078–2082, Aug. 1996.
[58] Z. Cvetković and M. Vetterli, “Oversampled filter banks,” to appear in [69] A. J. E. M. Janssen, “On rationally oversampled Weyl-Heisenberg
IEEE Trans. Signal Processing. frames,” Signal Processing, vol. 47, pp. 239–245, 1995.
[59] Z. Cvetković, “Oversampled modulated filter banks and tight Gabor [70] D. F. Walnut, “Continuity properties of the Gabor frame operator,” J.
frames in ” Proc. IEEE ICASSP-95, Detroit, MI, May 1995, pp. Math. Anal. Appl., vol. 165, pp. 479–504, 1992.
1456–1459.
[60] H. Bölcskei, F. Hlawatsch, and H. G. Feichtinger, “Oversampled FIR
and IIR DFT filter banks and Weyl-Heisenberg frames,” Proc. IEEE
ICASSP-96, Atlanta, GA, May 1996, vol. 3, pp. 1391–1394.
[61] J. J. Benedetto and S. Li, “The theory of multiresolution analysis frames Helmut Bölcskei (S’94) was born in Austria on
and applications to filter banks,” submitted to ACHA, 1996. May 29, 1970. He received the Dipl.-Ing. degree
[62] , “Multiresolution analysis frames with applications,” Proc. IEEE (summa cum laude) from Vienna University of
ICASSP-93, Minneapolis, MN, 1993, vol. 3, pp. 304–307. Technology, Austria, in October 1994.
[63] M. Poize, M. Renaudin, and P. Venier, “The Gabor transform as a From October to November 1994, he was a
modulated filter bank system,” in Quatorzième Colloque GRETSI, Juan- Researcher at the Department of Mathematics, Uni-
Les-Pins, France, Sept. 1993, pp. 351–354. versity of Vienna, Austria. Since December 1994, he
[64] S. Qiu and H. G. Feichtinger, “The structure of the Gabor matrix and
has been a Researcher at the Department of Commu-
efficient numerical algorithms for discrete Gabor expansions,” in Proc.
nications and Radio-Frequency Engineering, Vienna
SPIE, Visual Commun. Image Processing 94, K. Katsaggelos, Ed., 1994,
University of Technology. From February 1996 to
vol. 2308, pp. 1146–1157.
May 1996, he was a Visiting Researcher in the
[65] S. Qiu, H. G. Feichtinger, and T. Strohmer, “Inexpensive Gabor de-
Applied Mathematics Group at Philips Research Laboratories, Eindhoven, The
compositions,” in Proc. SPIE, Wavelet Applications in Signal and Image
Netherlands, where he worked on video coding. He is currently pursuing the
Processing II, A. Laine and M. Unser, Eds., 1994, vol. 2303, pp.
Ph.D. degree in electrical engineering. His research interests include filter
286–294.
[66] S. Qiu and H. G. Feichtinger, “Discrete Gabor structures and op- banks, frame theory, image coding, and time-frequency analysis.
timal representations,” IEEE Trans. Signal Processing, vol. 43, pp.
2258–2268, Oct. 1995.
[67] N. Polyak, W. A. Pearlman, and Y. Y. Zeevi, “Orthogonalization of
circular stationary vector sequences and its application to the Gabor
decomposition,” IEEE Trans. Signal Processing, vol. 43, pp. 1778–1789, Franz Hlawatsch (M’88), for photograph and biography, please see the
Aug. 1995. February 1997 issue of this TRANSACTIONS.

Authorized licensed use limited to: ETH BIBLIOTHEK ZURICH. Downloaded on January 25, 2010 at 04:09 from IEEE Xplore. Restrictions apply.