IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—II: ANALOG AND DIGITAL SIGNAL PROCESSING, VOL. 50, NO.

9, SEPTEMBER 2003

497

On The Eigenfilter Design Method and Its
Applications: A Tutorial
Andre Tkacenko, Student Member, IEEE, P. P. Vaidyanathan, Fellow, IEEE, and
Truong Q. Nguyen, Senior Member, IEEE

Abstract—The eigenfilter method for digital filter design
involves the computation of filter coefficients as the eigenvector of
an appropriate Hermitian matrix. Because of its low complexity as
compared to other methods as well as its ability to incorporate various time and frequency-domain constraints easily, the eigenfilter
method has been found to be very useful. In this paper, we present
a review of the eigenfilter design method for a wide variety of
filters, including linear-phase finite impulse response (FIR) filters,
nonlinear-phase FIR filters, all-pass infinite impulse response
(IIR) filters, arbitrary response IIR filters, and multidimensional
filters. Also, we focus on applications of the eigenfilter method
in multistage filter design, spectral/spacial beamforming, and in
the design of channel-shortening equalizers for communications
applications.
Index Terms—Channel-shortening equalizers, constrained filter
design, eigenfilter, least-squares filter design.

I. INTRODUCTION

T

HE EIGENFILTER design method for discrete time filters involves the determination of filter coefficients as
the eigenvector of a particular Hermitian positive definite (and
often real and symmetric) matrix. As opposed to other filter design algorithms such as the least-squares approach [48], which
requires the computation of a matrix inverse, the eigenfilter
method only requires the computation of a single eigenvector,
which can be found efficiently via the iterative power method
[49]. In addition to its inherently low design complexity, the
eigenfilter method can also incorporate a variety of time and
frequency-domain constraints into the design problem with relative ease, in contrast to other well known filter design methods
such as the McClellan–Parks algorithm [16]. Furthermore, because of the myriad of design problems that can be posed
as an eigenfilter problem, the method has been shown to be
useful for a variety of applications, ranging from spectral/spacial filtering or beamforming to communications regarding the
design of channel-shortening equalizers for discrete multitone
(DMT) systems.
The notion of such a filter design technique was introduced
by Slepian in 1978 [42] in the context of the design of window
functions for the ideal low-pass filter response. Slepian considered the problem of designing a window with a minimum stopManuscript received February 5, 2003; revised April 23, 2003. This work was
supported in part by the Office of Naval Research Grant N00014–99-1–1002.
A. Tkacenko and P. P. Vaidyanathan are with the Department of Electrical
Engineering, California Institute of Technology, Pasadena, CA 91125 USA
(e-mail: andre@systems.caltech.edu).
T. Q. Nguyen is with the Department of Electrical and Computer Engineering,
University of California at San Diego, La Jolla, CA 92093 USA.
Digital Object Identifier 10.1109/TCSII.2003.816942

band energy (subject to a unit norm constraint on the window
coefficients to avoid a trivial solution). It was found that the optimal window coefficients could be found from the eigenvector
of a real, symmetric, positive definite, Toeplitz matrix corresponding to its smallest eigenvalue.
In 1987, Vaidyanathan and Nguyen [50] generalized
Slepian’s method for window design to the design of
linear-phase finite impulse response (FIR) filters and formally
introduced the eigenfilter design method. They generalized
Slepian’s method to account for both passband and stopband
conditions and showed how to design a variety of filters (mainly
low pass, high pass, and band pass) with various time- and
frequency-domain constraints, including the Nyquist constraint
and flatness constraints [49]. Numerous simulation results were
provided showing the power and usefulness of the eigenfilter
method.
Since then, various generalizations of the eigenfilter method
have been proposed. The design of linear-phase FIR Hilbert
transformers and arbitrary order digital differentiators was considered by Pei and Shyu in [24], [25], [28], [36]. More general
FIR design methods were considered by Nguyen as well as Pei
and Shyu, including the design of linear-phase filters with arbitrary amplitude response [20], the design of arbitrary complex
coefficient nonlinear-phase filters [27], [29], [21], and the design of th band nonlinear-phase filters by Wisutmethangoon
and Nguyen [52]. Multidimensional extensions to the eigenfilter
method (mostly two dimensional) were originally proposed by
Nashashibi and Charalambous [18] and later considered by Pei
and others [26], [30], [31], [35], [8]. The eigenfilter method has
even been used to design infinite impulse response (IIR) filters,
including all-pass filters [14], [41], [33], [22], [55], arbitrary
IIR filters using a time-domain [32] and a frequency-domain
approach [1], [2], and IIR filters with time and frequency-domain constraints [39]. Recently, eigenfilter methods for filters
with general linear constraints were proposed [37] along with
methods based on a total least-squares-error criterion [38], [54].
As many design problems can be posed as an eigenfilter
problem, the eigenfilter method has been found to be useful for
several applications. For example, it has been used to design
multistage interpolated FIR (IFIR) filters [9] as well as arbitrary
log magnitude and phase response filters [34]. In addition, it has
been used to develop prototype filters for pseudo-quadrature
mirror filter (pseudo-QMF) uniform and nonuniform cosine
modulated filter banks [3], [4]. The generalization of the eigenfilter approach to solve general least-squares approximation
problems was proposed in [10]. Applications to spectral/spatial
filtering were also shown in [10] and more recently in [11].

1057-7130/03$17.00 © 2003 IEEE

Authorized licensed use limited to: San Diego State University. Downloaded on February 7, 2010 at 04:56 from IEEE Xplore. Restrictions apply.

498

IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—II: ANALOG AND DIGITAL SIGNAL PROCESSING, VOL. 50, NO. 9, SEPTEMBER 2003

In addition, the eigenfilter method has been found useful for
cancelling selective signals in an acoustic environment [6],
[7]. Use of this method for image scaling or size conversion
was shown in [53]. Recently, the eigenfilter method has been
shown to be very important in communications, especially for
the design of channel-shortening equalizers in DMT systems
[43]. Melsa et al. [17] were the first to apply the eigenfilter
method to the problem of channel shortening. Since then, many
other eigenfilter based channel-shortening equalizers have
been proposed [40], [45]–[47], [5]. Several of these methods
have been found to perform nearly optimally in terms of the
observed bit rate.

namely
, is a real coefficient polynomial in
of order
given by
. The prolate spheroidal wave
sequence is the one that minimizes
(1)
subject to the unit norm constraint
(2)
column vectors:

If we define the following

A. Outline
This paper is organized as follows. In Section II, we give a
brief review of the early work on eigenfilters and focus mainly
on the work of Slepian [42] and Vaidyanathan and Nguyen [50].
Various generalizations of the eigenfilter approach for filter design are discussed in Section III. In Section IV, applications of
the eigenfilter method are considered, such as multistage filter
design, spectral/spacial beamforming, and channel shortening
for communications. Finally, concluding remarks are made in
Section V.
B. Notations
All notations are as in [49] unless mentioned otherwise.
In particular, ( ), ( ), and ( ) correspond, respectively, to
the conjugate, transpose, and conjugate transpose operations.
Vectors and matrices are in bold font with lowercase letters
corresponding to vectors and uppercase letters to matrices. The
th element of a vector will be denoted by
, whereas the
th element of a matrix will be denoted by
.
II. HISTORICAL BACKGROUND
A. Prolate Spheroidal Wave Sequences
A classical approach for designing FIR low-pass filters is the
windowing method [49], in which the impulse response of an
.
ideal low-pass filter is multiplied by a window function
Typically, the window is also low-pass and it turns out that two
parameters of the frequency response of the window
most prominently affect the quality of the overall response [49].
increases, the transition bandAs the main lobe width of
width of the windowed response likewise tends to get larger,
increases, the peak
while as the peak sidelobe level of
passband and stopband ripples of the overall response also tend
, in
to as well. To mitigate the sidelobe level effects of
[42], Slepian considered designing a real window of unit norm
is
for which the energy in the frequency region
. Such a window is
minimized, for some with
a discrete time prolate spheroidal wave function, also called a
prolate spheroidal wave sequence [49]. Slepian showed that the
coefficients of a prolate spheroidal wave sequence could be obtained from the eigenvector corresponding to the smallest eigenvalue of a real, symmetric, positive definite, Toeplitz matrix, as
we now show.
is a real causal sequence of length
Suppose that
for some nonnegative integer . Then, the transform of
,

then

and
is real). Using this in (1) yields
where

(as

(3)

is clearly Hermitian, we can decompose as
, where and are, respectively, real symmetric and antisymmetric matrices [12]. Then, as is real, we have
, and so from (3) we get

As

(4)
The

th element of

is given by

(5)
depends only on
Clearly, is a Toeplitz matrix (as
), in addition to being real, symmetric, and positive definite
) [12]. The unit norm constraint of (2) is equivalent
(as
to the statement
(6)
Combining (4) and (6), the optimization problem can be posed
as follows:
Minimize

subject to

As
is Hermitian, it follows by Rayleigh’s principle [12]
where
is the smallest
that the minimum value of is
eigenvalue of . Furthermore, this minimum value is achieved
, where
denotes a unit norm eigenvector of
if
corresponding to . (More generally,
iff is any unit
norm vector in the eigenspace corresponding to . However,
for sake of clarity, we will ignore this scenario.)
The magnitude response of a prolate spheroidal wave sequence obtained using Rayleigh’s principle is shown in Fig. 1.
and the frequency
Here, we chose the order to be
. As can be seen from Fig. 1, the window
parameter as
or
and
exhibits a large attenuation near
most of the energy is concentrated in the region
as expected. Windows designed using the above method were

Authorized licensed use limited to: San Diego State University. Downloaded on February 7, 2010 at 04:56 from IEEE Xplore. Restrictions apply.

. Despite these advantages. Here.. Restrictions apply. we may be interested in filters other than low-pass filters. As such. by first focusing on designing a low-pass filter. even if the window is optimal in some sense. Several approaches have been made to overcome these problems. In [50]. is of the form . over a particular frequency region .TKACENKO et al. Early Work on Eigenfilters: Low-Pass Eigenfilters Vaidyanathan and Nguyen [50] introduced the notion of an “eigenfilter. Magnitude response of a prolate spheroidal wave sequence (N = 32. By subjecting to a unit norm constraint to avoid trivial solutions. For simplicity. 1) The Least-Squares Approach for FIR Filter Design: Often.e. is even. To say to have linear prevent phase distortion. 2010 at 04:56 from IEEE Xplore.  = 0:2 ). Downloaded on February 7. shown to yield good windowed low-pass FIR filters. This method for filter design is known as the least-squares approach [48]. proposed a recursive method based on constructing a set of orthonormal functions. i. Vaidyanathan and Nguyen suggested the optimization of a different objective function which can be expressed as a quadratic form in terms of the coefficient vector . (8) to approximate the desired response One approach to get over the region is to choose the coefficients to and . efficient FIR filter . minimize the mean-squared error between given by This solution can be obtained by using the trick of completing the square [12]. Transition From Optimal Windows to Eigenfilters Although designing FIR filters using the window method is easy. Slepian’s approach for calculating such sequences paved the way for more general filters to be designed via Rayleigh’s principle.: ON THE EIGENFILTER DESIGN METHOD AND ITS APPLICATIONS: A TUTORIAL 499 negative weighting function mean-squared-error measure in (9). C. Their method was shown to be both lower in complexity and less susceptible to numerical errors than the conventional least-squares design method in that no direct matrix inversion is involved. is of the form response of (7) where and is given by . Okuda et al. in which case it is not immediately clear how to use the window method. which adds to the complexity of the algorithm. . its computation requires the calculation of a matrix inverse which is high in complexity and subject to numerical inaccuracies. we constrain phase. They considered approximating a low-pass and stopband frequency . where is the amplitude . the resulting optimal can be obtained via Rayleigh’s principle [12] in much the same way that prolate spheroidal wave sequences are obtained [42]. this case. and it can be shown [48] that the optimal coefficient vector which minimizes the objective function in (10) is given by (11) where is an vector given by matrix and is an B. Though the filter coefficient vector from (11) is optimal in the least-squares sense. A slight generalization is to incorporate a non- as in (12). filter with passband frequency whose desired response is (12) (9) don't care. 1. we have from (7). in general the resulting filters are not optimal in any sense. two quantities were considTo approximate and a passband error ered in [50]. their method suffers from the drawback that the actual filter coefficients must be found from the coefficients of the orthonormal function expansion.” namely a filter whose coefficients are the components of an eigenvector. Furthermore. we would like the amplitude response of a real coto approximate a real desired response. In [23]. yielding a weighted (10) vectors: By defining the following Fig. although other types can be used as well. a stopband error measure Authorized licensed use limited to: San Diego State University. we will focus on Type 1 linear-phase filters here. and satisfies the symmetry condition where [49]. In is of order .

we chose for the least-squares approach and for the eigenfilter method).and Frequency-Domain Constraints: Another advantage of the eigenfilter method shown by Vaidyanathan and Nguyen is the ease with which certain time and frequency-domain constraints can be accounted for in the design. given by . In be- . 2. 2010 at 04:56 from IEEE Xplore. NO. the authors first considered the Nyquist constraint. which is formally used to calculate the largest eigenvalue of a Hermitian positive semidefinite matrix and its corresponding eigenvector. The power method. the magnitude responses using the eigenfilter method as well as the least-squares approach are plotted for and . In light of (13) and (14). As of all 1’s. show the merits of the eigenfilter method.e. When of the design emphasis is on the stopband. W (!) = 1. In [50]. the resulting expression will not be in the form of a quadratic form in terms of the vector . the authors in [50] also showed how to design bandpass and multiband filters with the introduction of different reference frequencies. 9. positive definite matrix given by To jointly minimize both and . the advanced response Authorized licensed use limited to: San Diego State University. VOL. (! = 0:3 . From the plots. the vector was chosen to minimize (15) where is a tradeoff parameter between stopband and passband performances. the resulting filter coefficients could be found using the eigenfilter technique. In particular. such cases. the complexity of the eigenfilter method is much less than that of the least-squares approach. where is an vector measured. N = 24. 2. The eigenfilter was rescaled to have unity . the mean-squared error (13) where is a real. It was shown that even with such constraints. 50. was chosen to be (14) where is a real. In particular. along with the lower complexity of the latter. For the stopband error of (9) was used. positive def. a reference frequency was introduced into the design problem and the passband error measure was taken to be the deviation of the amplitude response from its value at the reference frequency. then the optimal norm. they focused on the time-domain Nyquist constraint [49] and also a frequency-domain flatness constraint. As opposed to the least-squares approach for filter design which requires the computation of a matrix inverse as can be seen from (11). Here. to avoid trivial solutions. i. 1) Incorporating Time.. If we impose that has unit inite matrix since . most performances of the passband and stopband. Restrictions apply. the deviation of from its value at the reference frequency was . To bypass this dilemma. [49]. Magnitude responses obtained using the least-squares approach along with the eigenfilter approach. the two responses become more and more alike. As increases. the a low-pass filter with and equal weighing filter order was chosen to be was used for both the passband and stopband (i. symmetric. is known to converge quickly when the largest eigenvalue is much larger in magnitude than the next largest eigenvalue.500 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—II: ANALOG AND DIGITAL SIGNAL PROCESSING.. [50] (17) is some positive integer and where other words. In Fig. which minimizes in (16) is simply the eigenvector correof by Rayleigh’s sponding to the minimum eigenvalue principle [12]. Downloaded on February 7. symmetric. As mentioned earlier. the authors in [50] considered minimizing a convex combination [12] of the two. if the mean-squared error of (9) for the passband error is used. although the former performs slightly worse in some parts of the stopband. A . In Fig. of order we have focused on The causal FIR filter thus far is said to satisfy the Nyquist constraint [or said to be Nyquist ( )] if [49]. = 1=2). the tradeoff parameter controls the .e. the eigenfilter method only requires the computation of a single eigenvector of a Hermitian positive definite matrix. The close agreement between the least-squares approach and the eigenfilter method. In particular. it can be seen that the eigenfilter gain at is very similar to the least-squares filter. symmetric. This eigenvector can be calculated efficiently using the iterative power method [12]. similar scenario holds for the passband when In addition to designing low-pass filters using the eigenfilter approach. positive definte matrix given by Unlike the stopband error measure. in (15) can be expressed as follows: where (16) Note that the matrix is itself a real. and consequently the stopband ripple sizes are smaller than for other values of . SEPTEMBER 2003 measure . ! = 0:35 . is some constant. for .

but with the smaller matrix . Restrictions apply. given by with 501 . Here. As was suggested ative procedure to identify the appropriate and denote based on the above observation. as shown in the second equation at the bottom of the page and is the matrix obtained by removing the rows and columns of whose indices are nonzero multiples of . we have plotted the magnitude response of an equiripple filter designed using the iterative eigenfilter method along with that obtained by the Remez exchange algorithm used in the McClellan–Parks program. To compare both methods.and frequency-domain constraints into the design. the . it is apparent that the rows and columns of whose indices are nonzero multiples . From (16) and (18). From (18). Recall that from (7) is any arbitrary frequency . Downloaded on February 7. is widely used for mizes the equiripple filter design on account of its rapid convergence and good performance. if for . respectively. the is equivalent to . we designed a low-pass filter without any time or frequency-domain constraints. If the passband and stopband error curves after the th iteration. 3. the authors in [50] focused on a flatness and later showed how to accomodate for constraint at .e. . In addition to showing how the Nyquist constraint could be incorporated in the design of eigenfilters. . which mininorm of the response error. the design constraint subject to . Alteran expansion in terms of the functions can be expressed in terms of the functions natively. As such. However.: ON THE EIGENFILTER DESIGN METHOD AND ITS APPLICATIONS: A TUTORIAL comes an impulse after being decimated by must satisfy the coefficients [49]. it can easily be shown that the flatness . such as those from above. Assuming this to be the case. is often very difficult. Under the unit norm con. (18) Authorized licensed use limited to: San Diego State University. the objective function from where and . can be minimized using Rayleigh’s principle straint degree as before. Note that the vector is completely we have arbitrary and that the flatness constraint is automatically satisfied by construction. yielding the expansion (20) With this expansion. and iff for . it was also shown how to incorporate flatness constraints in the frequency domain. For uni). In other words. By is said to have a definition [49]. of do not contribute to the overall error can be rewritten as where is simply the vector of (18) with the zeros removed. The relative ease with which the eigenfilter method can accomodate such constraints was the primary motivation for Vaidyanathan and Nguyen to develop an iterative eigenfilter technique for designing equiripple filters.. Using (8). Hence. To generalize this approach to have a of flatness at any arbitrary frequency . Thus. incorporating time. we need only expand in terms of the functions . an iterfrom them. given by (19) For simplicity. where is as in (16) but where . shown at the bottom of the page. The iteration is carried out until the magnitude of the difference in the solutions is smaller than a prescribed small constant . then has a degree of flatness . 2010 at 04:56 from IEEE Xplore. the authors in [50] introfor the passband and stopband duced a weight function error measures of (14) and (13). which problem is to minimize is the traditional eigenfilter problem. constraints of (19) are satisfied for . either and or their envelopes can be used. iteration is carried out until In Fig. the coefficient vector is of the form in (18). then by defining the at following vectors: (21) . were observed to be large near the band edges and smaller away and are not known a priori. To formulate an iterative technique. th iteration was chosen as the weighting function for the with the initialization . Here. Thus. as the McClellan-Parks algorithm is incapable of incorporating them.TKACENKO et al. In this case. the errors and form weighting (i. a frequency response if we have degree of flatness of at the frequency (15) takes on the form and . 2) Iterative Eigenfilter Method for the Design of Equiripple Filters: The McClellan–Parks algorithm [16].

it can be shown from (22) that we have the equation shown at the bottom of the page. we can create approximately equiripple eigenfilters that satisfy certain time and frequency-domain constraints. Arbitrary Desired Response Eigenfilters Nguyen originally considered the problem of using the eigenfilter method to approximate an arbitrary real desired amplitude response using linear-phase FIR filters [20]. and . In addition. is chosen by construction to be Note that the error function was chosen to be zero at . A decibel plot For the eigenfilter method. The eigenfilter is only approximately equiripple prevents since the reference frequency condition at this frequency from being an extremal frequency. it could be seen was large in magniheuristically that choosing where tude yielded better results. albeit only slightly larger. here we include an arbitrary is the reference frequency which is chosen by the designer. . 3. Restrictions apply. but (In [20]. many generalizations have been made to the eigenfilter method. Using the fact that . With the usual on . GENERALIZATIONS OF THE EIGENFILTER METHOD Since the work of Vaidyanathan and Nguyen [50]. 1) Arbitrary Amplitude Response Case: In order to approxwith a linearimate any real desired amplitude response phase FIR filter using the eigenfilter method. we will review a recent eigenfilter method based on a total least-squares-error criterion. (b) Passband details using a nondecibel scale. the error eigenfilter exhibits a larger. we will focus on designing multidimensional FIR filters using the eigenfilter method. we will focus on several such extensions. 3(b) on an nondecibel scale. denotes the boundary of the region was chosen to be . . including using the eigenfilter method to design arbitrary amplitude and frequency response filters and incorporating general linear constraints into the design. (a) Magnitude response plot on a decibel scale. 3(a). Downloaded on February 7. the optimal coefficients unit norm constraint can be obtained using the eigenfilter method as befor fore. thus justifying the choices of suggested by Nguyen in (23) and (24). Finally. Nguyen considered minimizing the objective function (22) Fig. whereas in (24) Pei and Shyu [29] showed through a design example that the had a noticable effect on the overall performance choice of of the resulting filter. 2010 at 04:56 from IEEE Xplore. which is equivalent quency condition that . A. Magnitude response of equiripple filters designed using the eigenfilter and Remez exchange algorithms. Here. In [20]. Nguyen [20] considered modifying the least-squares-error criterion (9) by incorporating a reference frequency at which the FIR filter ampliwould exactly match . More speciftude response ically. Then. 9. we see that both methods are in excellent agreement.) Here. we chose of the magnitude response is shown in Fig. 3(a). It should be noted that after the optimal is found. and we chose for both methods. Nguyen considered the special case for sake of generality. symmetric. 50. As before. From their simulations. Using the methods of the previous subsection. while the passband details are shown in Fig. Later. as well as those of [29]. to Authorized licensed use limited to: San Diego State University. NO. In this section. As such. From Fig. the matrix is a real. (23) where [21]. SEPTEMBER 2003 design IIR filters including all-pass filters. we present the methods of [20]. sired response [29]. III. the method was generalized to approximating a general complex dewith a complex coefficient FIR filter [21]. we will show how the eigenfilter method can be used to . than the filter designed using the Remez exchange algorithm.502 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—II: ANALOG AND DIGITAL SIGNAL PROCESSING. positive definite matrix. the solution must be scaled in order to satisfy the reference fre. VOL.

Under the unit . they considered minimizing the objective (27) . suppose that in addition to choosing . we wish to choose to satisfy (28) Note that this is similar to the flatness constraints of (19) considered in Section II-C-1. After obtaining the optimal coefficients.: ON THE EIGENFILTER DESIGN METHOD AND ITS APPLICATIONS: A TUTORIAL 503 2) Arbitrary Desired Response Case: Nguyen as well as Pei and Shyu independently developed different methods to design a complex coefficient FIR filter to approximate a generally complex valued desired response using the eigenfilter method [29]. Downloaded on February 7. Namely. Similarly to what was response done in the previous section. 5(a) and (b). the plex-valued Hermitian. a As an example. Pei and Shyu modified (26) to incorporate a reference frequency at which the frequency response of the filter would match the desired response. then . and uniof form weighting. positive definite matrix. a reference frequency of . 2010 at 04:56 from IEEE Xplore. In [21]. Pei and Shyu considered modifying the following sponse least-squares objective function. 4. . the resulting eigenfilter magnitude and group delay responses are shown in Fig. Nguyen’s method involved volved an real matrix. (b) Desired group delay response. (This example was considered by Nguyen [21]. the eigenfilter designed here yielded a good fit to the desired response in the frequency regions of interest. B. In particular. if we consider the complex coefficient FIR filter from to Section III-A-2. it can be shown that we have the equation shown at matrix is a comthe bottom of the page. (26) in order to express it as a quadratic form in terms of the vector . we present the method of Pei and Shyu [29]. With a filter length . Incorporating General Linear Constraints Using the Eigenfilter Approach Often times in the design of FIR filters. Whereas Pei and Shyu’s method inimaginary parts of complex matrix.TKACENKO et al. For example. be it real or complex. (a) Desired magnitude response. Nguyen solved the same problem. (In [29]. Defining the vectors (25) . we would like as well as its first approximate derivatives to match those of at some reference frequency . which applies for any Hermitian matrix. Namely.) is the negative of the derivaRecall that the group delay tive of the phase with respect to [49]. As we have . Using subject to the unit norm constraint . Here. the resulting filter must be scaled in order to satisfy the reference frequency condition . 4(a) and (b). As we can see. Restrictions apply. of the desired response respectively. to pose the design problem as an eigenfilter problem. Here. so that Let . if Fig. the authors did not include the weight function or the constant in their analysis. [21]. but decouinto real and imaginary parts and then expressed the pled objective as a quadratic form in terms of the vector of real and . (29) . We have decided to include both quantities for sake of completeness. be a causal FIR filter of length . it is desired that the filter coefficients satisfy a general linear constraint. the optimal filter coefficients can be norm constraint found using Rayleigh’s principle. the constraints in (28) can be rewritten in the form Authorized licensed use limited to: San Diego State University. suppose that the magnitude and group delay are as shown in Fig. To develop an eigenfilterwe clearly have based method for approximating a complex valued desired re.) The frequency and the desired region is now a subset of the interval is allowed to be complex.

the filter design problem becomes Minimize subject to The key to solving this constrained minimization problem is to iff lies in the null space of [12]. [37] for Recall that for the complex case..e. Note that the constraint equation in (29) is linear with respect to the coefficient vector . SEPTEMBER 2003 [12] in terms of the true coefficient vector . Pei. the eigenfilter problem is typically formulated as follows: Minimize subject to where is some Hermitian matrix. 5. . . As is clearly Hermitian. where is chosen tion. Incorporating general linear constraints as in (29) using the eigenfilter approach was first considered by and later generalized Chen [10] for the special case where . where is a vector given by matrix and is a . Downloaded on February 7. ! = 0:85 ). by Pei et al. Any such note that can be expressed as . the design problem becomes Minimize over all nonzero vectors . the minimum value of is the smallest eigenvalue of which occurs iff is in the eigenspace corresponding to . the length of is then chosen to satisfy the condition (30) If must satisfy linear constraints of the form given in (29). the true coefficient vector that we use is the optimal scaled by a factor in order to satisfy the reference frequency condi.504 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—II: ANALOG AND DIGITAL SIGNAL PROCESSING.. Tseng. In the eigenfilter design problem. By expressing as a Rayleigh quotient Hence. With this decomposition of . Pei and Tseng suggested a novel modification to the Authorized licensed use limited to: San Diego State University. 2010 at 04:56 from IEEE Xplore. and Yang [37] used the equivalent form of the reference frequency condition in a of (30) to express the linear constraint equation more simplified form. basis for the null space of and is any arbitrary vector [12]. the design problem becomes subject to Minimize To solve this optimization problem. By Rayleigh’s principle. this was done by introducing a reference frequency condition into the objective. In other words. As mentioned earlier. Using (30). Here. the length of must be chosen to satisfy the reference frequency condition of (30). It should be noted that the constraint here is really just included to prevent trivial solutions from occuring. where is a rectangular unitary ) whose columns form an orthonormal matrix (i.. . Note that here the length of is arbitrary. NO. Eigenfilter Method Based on a Total Least-Squares-Error Criterion Most eigenfilter techniques arose by altering the least-squares-error criterion of (10) or (26) in order to express it as a quadratic form in terms of the filter coefficient vector. 50. We present the results of [37]. Typically after the optimal is found. Simulation results in [37] using the derivative constraints of (28) showed the usefulness of this eigenfilter technique for designing filters with generalized linear constraints. C. Recently [38]. Restrictions apply. the linear constraint equation can be expressed as follows: where Fig. we use (which implies such that we have ). the design problem becomes Minimize where over all nonzero vectors . 9. (b) Group delay response. we can use Rayleigh’s principle to find the optimal and then use to find the optimal . VOL. (N = 50. (a) Magnitude response of the complex eigenfilter. column .

A more thorough description of this principle is provided in [44. To derive the total least-squares eigenfilter objective function. where iff . If we define and . is chosen to minimize the objective where (33) matrix represents the error in both the data The and desired response. Pei and Tseng analyzed the real least-squares-error criterion of (10). In the total least-squares method. pp. a linear model of the form (where is an vector) is formed to best approximate a desired vector in some sense. consider the complex least-squares-error criterion of (26). and is a constant chosen to (assuming that such a make the last component of equal constant exists). is chosen to minimize sumption \bfrom the second equation in (31) is that the error is associated with the desired vector .: ON THE EIGENFILTER DESIGN METHOD AND ITS APPLICATIONS: A TUTORIAL least-squares objective based on a total least-squares error criterion. this method has been shown through simulations to come closer to the desired least-squares solution than the conventional eigenfilter method. For the traditional least-squares method. the vector is chosen to minimize the “total” error in both and . From (35) and (36). an error is not only associated with the desired vector . is chosen to minimize the squared magnitude of the error (31) . whereas then represents the minimum distance between and the hyperplane . In particular. it is clear that an appropriate total least-squares-error criterion is given as follows: (37) Using (36). (38) . An implicit asNamely. 6. If is defined as the linear approximation to . that it can easily be shown that . given a set of data in the form of a matrix . From (32). we first briefly review the principle of total least-squares. First note squares objective which can be expressed in the form Authorized licensed use limited to: San Diego State University. eigenspace corresponding to . To properly introduce the total least-square eigenfilter method. where is as in (31).TKACENKO et al. 6. The resulting objective is still in the form of a Rayleigh quotient in terms of the filter coefficient vector. but also with the data matrix . Downloaded on February 7. and so the optimal vector can be found using Rayleigh’s principle as before. then the error equation \bfrom (31) becomes 505 Fig. Another advantage is that for low filter orders. then clearly we have (35) . i.) Note that analyzed above. it has been shown to be less susceptible to numerical inaccuracies than the traditional least-squares method which requires matrix inversion. the only difference being that it depends on the variable . represents the length of the path . from Rayleigh’s principle. (37) becomes (34) There is an interesting geometric interpretation between the \bfrom (34) and the leasttotal least-squares objective . Note that it is of the form (32) For the total least-squares method. Furtheris . an alternative total least-squares-error criterion can be chosen as where (36) (In [38].. we get where and It can be shown [44]. In many typical linear modeling problems. 2010 at 04:56 from IEEE Xplore. In addition. Geometric interpretation of the least-squares and total least-squares problem. This is shown \bfor each other. where is any vector in the more. for high filter orders. Because of this. One advantage of the total least-squares eigenfilter objective over the traditional ones of (22) and (27) is that the former does not require a reference frequency. where the one-dimensional case in Fig. Restrictions apply.e. 533–535]. We have opted to focus on the complex least-squares criterion for sake of generality and notational is essentially the same quantity uniformity. the path of which lies in a direction normal to the hyperplane. that the minimum value of in (33) is the smallest eigenvalue of . These paths are always separated by an angle of \bfrom . If is a matrix representing the error in the matrix .

Recently. They showed that the optimality criterion for the phase error in the Chebyshev sense could be posed as a generalized eigenvalue problem and proposed an iterative eigenfilter method to design equiripple phase responses. and Thus. VOL. 7(b) are . is of the form which .e. Designing IIR Filters Using the Eigenfilter Method Fig. 2010 at 04:56 from IEEE Xplore. . Following an example from Pei and Tseng [38]. Hence. we focus on the method of Pei and Shyu in [33]. we have expressed as a Rayleigh quotient in terms of the vector . 7(a) are low order with parameters . 9. the . Pei and Shyu [33]. . the conventional eigenfilter method (which requires a reference frequency). The merits of the total least-squares eigenfilter method are best seen via simulations. ! = 0:4 ). The phase of . It can be observed that both eigenfilter methods performed about the same. The first to consider this problem were Laakso et al. Though difficult to see. To obtain the proper . 50. are much less susceptible to numerical inaccuracies caused by ill-conditioned matrices than filters designed using the least-squares method. Downloaded on February 7. Several methods for designing IIR filters have also been considered in the literature. By construction. D. SEPTEMBER 2003 reference frequency was chosen to be . is an matrix of the form (39) where . i. This example serves to show the merits of the total least-squares eigenfilter method for both low order filters as well as high order ones. Zhang and Iwakura [55] considered the design of phase equiripple all-pass filters based on an eigenfilter method. sponse of the only design freedom present is in the selection of the phase . ! = 0:2 . The filters shown in . The filters in Fig. Furthermore. Note the similarities between the matrix from (39) and that from (34). which is a modification of the phase error the methods proposed in [14] and [22] are only approximations to the phase error norm. (N = 148. the and 1) Design of All-Pass Filters: Thus far. norm. ! = 0:3 ) (b) Magnitude responses of several high order low-pass filters. Here. Restrictions apply. whereas the least-squares filter yielded a poor response. we must scale the optimum to satisfy (38). . . in Fig. which we denote here by of . In fact. [14] for the design of all-pass phase compensators. respectively. as they number was 1. The transfer function of a causal th order all-pass function is of the form [49] where and Here. which is optimal with respect to its error criterion. we have plotted the magnitude response of several low-pass filters designed using the traditional least-squares method. The reason for this is that the matrix which needed to be inverted for the least-squares method was extremely ill-conditioned. the frequency reis of unit magnitude. By Rayleigh’s principle. the condition . For this example. Fig. (a) Magnitude responses of several low order low-pass filters. we will focus on real coefficient all-pass functions for is real for all . . For the conventional eigenfilter method. Both eigenfilter methods. and the total least-squares eigenfilter method. ! = 0:25 . we can find the optimum and thus . (N = 32. 7. and 1 1 quantities are. [22]. is of the form Authorized licensed use limited to: San Diego State University. the total least-squares eigenfilter is noticably closer to the optimal least-squares filter than the conventional eigenfilter. An extension of the total least-squares method was recently proposed by Zhang and Chen [54] which asymptotically approaches the conventional least-squares solution. which requires matrix inversion. and high order with ( ). the design complexity of the total least-squares eigenfilter is much less than that of the least-squares filter. In this case. NO. In contrast. (40) . later suggested alternative design methods for these filters. it is indistinguishable from the least-squares filter. 7(a) and (b). Here. as well as Nguyen et al.506 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—II: ANALOG AND DIGITAL SIGNAL PROCESSING.7012 do not require matrix inversion. we only focused on designing FIR filters using the eigenfilter method..

enforcing (43) for Following an example considered by Pei and Shyu in [33]. [2] proposed different methods to design arbitrary IIR filters using the eigenfilter approach. 8(b). we have (43) is the group delay of . Downloaded on February 7. The all-pass filter coefficients were found using a method similar to that described in the previous subsection. even if 2) Design of Arbitrary IIR Filters: Pei and Shyu [32]. However. the pole farthest away from the origin had magnitude 0. Restrictions apply. which were found necessary to yield a good ( phase approximation for the rest of the frequency region of interest. we would have. The phase of a real coefficient all-pass filter sponse as in (45) introduces It should be noted that choosing ) and two transition bands into the design.: ON THE EIGENFILTER DESIGN METHOD AND ITS APPLICATIONS: A TUTORIAL where is the phase of given by 507 with poles inside the unit circle and side the unit circle satisfies [22] poles out- (41) Hence. In Fig. Howdoes not guarantee stability. namely ( ). for a stable all-pass filter of order and denote the desired phase Let and its correresponses for the all-pass filter . ever. the authors designed such an all-pass filter using the eigenfilter objective of (42). In order to have a causal. Howsired impulse response Authorized licensed use limited to: San Diego State University.TKACENKO et al. using (41) (45) otherwise.9262. In general. 8(a). we can see that the eigenfilter method yielded a good approximation to the desired phase. Pei and Shyu [33] have to match . In 70 would be best suited to approximate [33]. In [32]. One problem with designing IIR filters not present when designing FIR filters is that of stability [49]. From both plots. [39]. In particular. [41]. in (44) satisfies (43) for It can be verified that . has whereas in Fig. obtained from In obtaining a criterion to choose the coefficients of the poly. for all-pass filters. then we for all . the error metric can be the “desired response” of expressed as a quadratic form in terms of . it should be stressed that an all-pass filter designed using the eigenfilter satisfies (43). then heuristically we would have mately equal to . Since is 0. we have (42) and so the optimal denominator coefficient vector can be found using Rayleigh’s principle as before. Defining the vectors shown at the bottom of the page. as well as by Nguyen et al. To choose to would have approximately ensure this condition as best as possible. [22]. it can be inferred that an all-pass filter of order as in (44). which need not be causal. Pei and Shyu [33] considered minimizing a mean-squared-error measure of the form subject to a unit norm constraint on (namely ). For the eigenfilter designed here. as well as Argenti and Del Re [1]. often times it is possible to obtain stable solutions by imposing constraints on the desired phase re. In [41]. From this. method need not be stable. all of its poles must lie inside the unit circle. that by conto satisfy straining the desired group delay response (43). the authors proposed a time-domain method to approximate a de. . stable transfer function. From (40). the group delay of the 70th-order all-pass eigenfilter designed using the above parameters is shown. It was observed by where Pei and Shyu [33]. which can be considered choosing using the above formula. suppose that the desired phase response is (44) . Shyu and Pei proposed a method for designing multiband IIR filters using sums and differences of all-pass filters. we sponding denomiator polynomial . the phase error been plotted. and so the filter is stable. often times stable all-pass filters would be obtained. Pei and Shyu noted that if was approxinomial . 2010 at 04:56 from IEEE Xplore. Despite the fact that in this example we have stability. this constraint is very difficult to enforce. in which the weighing function was chosen as Equivalently.

. as we now show. suppose that the desired response is a and stopband low-pass filter with passband frequency that has been delayed by for some . 2010 at 04:56 from IEEE Xplore. Recently. By subjecting to the . the authors of [1] and [2] proposed choosing them to minimize the following error measure: where (46) can be argued to be a valid error crite. 0 ever. of the form Suppose that we have a causal IIR filter where and Instead of choosing and to minimize the usual weighted mean-squared-error criterion of (26). SEPTEMBER 2003 Define the following vectors: It should be noted that here we define the vector vector for any nonnegative integer . Restrictions apply. the resulting magnitude. [N = 70. since if and hence is a quadratic form in terms of the vector of numerator and denominator coefficients. Heuristically. it should be noted that the results are very sensitive to the choices of .9334 and so is stable. and . as well as very approximately linear phase in the passband region. From this. As an example. we have and . (b) Phase error  (! )  (! ). Downloaded on February 7. the pole of thest away from the origin had magnitude 0. However. frequency is Namely. Pei and Shyu considered the design of special classes of IIR eigenfilters satisfying certain time and frequency-domain constraints. computations must be performed on matrices of very large sizes (in [32]. 50. . . 9(a)–(c). the fluctuations in the group delay in the passband are noticably less than those of elliptic and Chebyshev filters designed for the above observed passband and stopband ripple characteristics. are shown in and group delay plots of the IIR eigenfilter Fig. which is difficult to do. where we desire linear phase. To obtain this kind of performance with a linear-phase FIR filter. then we should also have and conversely. The advantage can be expressed as a quadratic form in terms here is that of the polynomial coefficients. matrices of size 512 were used). 8. Clearly. and . Argenti and Del Re [1]. phase. rion. the filter order would have to be much larger than the ones used here. If we choose Suppose that . As such. suppose that don't care . from (46) is simply Let us now define the following vector nating and : obtained by concate- Then we have Using this in (46). we see that the phase becomes less linear as we approach the passband frequency . If we are not interested in the phase of and we obtain an unstable filter with an adequate magnitude response. NO.508 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—II: ANALOG AND DIGITAL SIGNAL PROCESSING.08 and stopband attenuation of about 60 dB. to get good performance. respectively. The phase and group delay are only plotted for the passband region. 9(c). Many choices yielded a poor response either in the magnitude or phase (or both) and many farchoices yielded an unstable filter. the optimal is obtained usual unit norm condition using Rayleigh’s principle. there is no guarantee that the resulting filter will be stable. Furthermore. we have Fig. and . From the plots. The deviation of the phase from linearity is best seen from the group delay plot of Fig. we see that the IIR eigenfilter has a relatively good magnitude response with a peak passband ripple size of about 0. Despite this. [2] proposed an IIR eigenfilter design method using a frequency-domain approach which we present here. Here. Authorized licensed use limited to: San Diego State University. 9. In [39]. W (! ) as in (45)]. VOL. (a) Group delay response of the all-pass eigenfilter. showing the merit of the designed IIR filter.

like . [30]. Multidimensional Eigenfilters The eigenfilter method can easily be extended to the multidimensional case. is nonzero. [31]. a subset (of finite cardinality) of the set of where is a real sequence. including T. (a) Magnitude response of the IIR low-pass eigenfilter. if not all. E. Here. Here. and for appropriate arranged of according to some order. (b) Phase response in the passband region. Argenti and Del Re proposed an iterative equiripple eigenfilter technique to improve the magnitude characteristic of the filter in this case. However. The frequency response of is a function of variables given by the following [49]: where is a real constant vector and sional function of the form -dimen(48) is. (c) Group delay response in the passband region. we mean is only nonzero for a finite number of integer vectors that . Note that and are simply -dimenif sional extensions of the objectives (22) and (27) previously considered for the one-dimensional (1-D) case. as well as H. In [1]. Suppose that we have an FIR -dimensional signal where is an -dimensional integer vector. . then one approach is to choose the filter response coefficients to minimize either (50) if the desired response is a real function. M = 12. Hence. in (51). By FIR. contributions regarding multidimensional eigenfilters have been for the 2-D case and have come from Pei and Shyu [26]. Downloaded on February 7. Here.: ON THE EIGENFILTER DESIGN METHOD AND ITS APPLICATIONS: A TUTORIAL 509 Fig. Most. K = 12). although others have also considered this problem. . (! = 0:3 . by imposing where where Authorized licensed use limited to: San Diego State University. ! = 0:4 . and consist of the respective elements where . If the goal is to design a filter to approximate a desired . and is all -D integer vectors. Chen [10]. N = 5. then we can replace each pole outside the unit circle with its reciprocal conjugate pair. the region in (50) is a subset of the -dimensional interval . [2]. Chen and Ford [8]. Note that if we impose some for sort of lexicographical ordering on the coefficients of as in (47) or for as in (48). . then we have is a real is complex. Nashashibi and Charalambous [18] were the first to design multidimensional eigenfilters by considering the two-dimensional (2-D) case. . it can be shown that we get the equations shown at the bottom of the page. Using the decompositions of whereas and from (49). we focus on the general -dimensional case by generalizing the objective functions (22) and (27). IIR eigenfilters have not been as widely used as their FIR counterparts. If has linear integer vectors such that phase [49]. 2010 at 04:56 from IEEE Xplore. which was shown to work well through simulations. since we often desire exact or approximate linear phase in practice. then we have (49) . [35]. . or alternatively (51) (47) is a subset of the set of all -dimensional integer where has finite cardinality and consists of those vectors. . 9.TKACENKO et al. Restrictions apply. a real function consisting of trigonometric functions (products of sines and cosines in particular).

it has been used to design multistage IFIR [19] filters [9]. where . In this section. in image processing. 10. One simple way to overcome this problem is to implement the single FIR filter as a cascade of two filters as shown in Fig. Recently. A. the eigenfilter method chosen to be produced a good approximation to the desired response. quency as is that The advantage to implementing and . 10. most notably 2-D. that the filter order approximately varies inversely with the length of the is a linear-phase low-pass transition band [49]. Restrictions apply. for fixed passband and stopband ripples sizes. In addition to designing circularly symmetric filters. then designing a single filter to accomodate the given specifications often becomes cumbersome due to the large filter order required. called the IFIR filter implementation [19]. [40]. These images are suppressed via the . Several such channel-shortening eigenfilters have been shown to be nearly optimal in terms of observed bit rate. The eigenfilter technique has been shown useful for spectral/spatial selectivity. Here. First. Namely. As stopband frequency of consists of a desired low-pass such. may be much larger than . the expanded version and . Several of these include fan filters. we will focus on applications in communications regarding the shortening of channels for DMT systems. then If the transition band length is relatively small. 2010 at 04:56 from IEEE Xplore. With the advent of DMT systems such as the digital subscriber loop (DSL) [43] in recent years. we show how the eigenfilter technique can be used for spectral/spacial beamforming. eigenfilters which satisfy the Nyquist constraint have been shown to be suitable prototype filters for interpolation filters for image size conversion [53]. 11. For example. denoted here by the transition bands of and . it can be shown. the one in which it has had the most impact is in communications. 2 12 circular low-pass a unit norm constraint on or . VOL. Authorized licensed use limited to: San Diego State University. IFIR filter implementation. In particular. we have only focused on the use of the eigenfilter method to design single filters of various desired responses. However. [5]. Among the various applications where the eigenfilter method has been shown to be beneficial. Furthermore. 50. For this problem. elliptical filters. arbitrary log magnitude and phase response filters [34]. we can minimize or via Rayleigh’s principle as before. Here. Then. is a posiis detive integer called the interpolation factor. The filter and a signed to be low pass with a passband frequency of (assuming that ). much attention has been given to the shortening of channels encountered in practice. The frequency response of a zero-phase 12 12 eigenfilter designed for this desired response is shown in Fig. the eigenfilter approach has been found to be useful in other applications as well. [11]. is a 2-D circularly symAs an example. 11.510 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—II: ANALOG AND DIGITAL SIGNAL PROCESSING. Multistage IFIR Eigenfilters When designing linear-phase low-pass FIR filters. as well as prototype filters for uniform and nonuniform cosine-modulated filter banks [3]. IV. . thus transforming the problem of designing one filter with a high order into that of two filters with much smaller order. equal uniform weighing was used and the reference frequency was . Eigenfilters have also been successfully employed for selective signal cancellation in acoustic environments [6]. Finally. and quadrantally symmetric filters. Downloaded on February 7. As we can see. design methods such as the McClellan–Parks algorithm and the least-squares approach become subject to numerical inaccuracies. which is the primary criterion of interest in DMT systems. NO. filter of order . which is designed to be low-pass with passband frefilter and stopband frequency . or beamforming. suppose that metric low-pass filter of the form don't care otherwise. An alternative method to design 2-D filters using a 1-D prototype with the McClellan transformation [15] has been considered by Pei and Shyu [31] as well as Chen and Ford [8]. [7]. SEPTEMBER 2003 Fig. [45]–[47]. we will focus on three applications of the eigenfilter method. APPLICATIONS OF THE EIGENFILTER METHOD Thus far. as the filter order increases. as portion with passband and stopband frequencies ) undesired bandpass images [49] which occur well as ( by virtue of the expansion. if . the eigenfilter method has been found to be very useful [17]. the eigenfilter method has been used to design a plethora of other multidimensional filters. for antenna and microphone arrays [10]. [4]. 9. Fig. we will show how the eigenfilter approach can be applied to the design of multistage IFIR filters. Frequency response of a zero-phase 12 eigenfilter.

both assuming .TKACENKO et al. . respectively. the identity and where reversal matrices. We should note that once the optimal for the model filter is found. Authorized licensed use limited to: San Diego State University. . Let also a Type 1 linear-phase filter of order and denote the vectors of coefficients of and . . 2010 at 04:56 from IEEE Xplore. If is the vector of filter coefficients . ... . As such. . . and denote the folded versions of and by and . 12. In [9]. the folded version (the same vector considered in Section II) was chosen miniwith is as in (16).. Here. . . Restrictions apply. the folded version of . subject to mize the objective . suppose we wish to design a and using the IFIR low-pass filter with . we can express the and of version of and we have ’s in terms of the ’s. .. This shows that the eigenfilter problem constraint can be posed as an eigenfilter problem for the original . . To properly describe their method. the objective function from (16) becomes To avoid trivial solutions.. . . the matrices . For example Using (8). as we now show. where . . the authors considered choosing eigenfilter approach of Section II... is we must introduce a few new quantities.. It was shown that this the usual unit-norm condition filter design problem could be posed as an eigenfilter problem in terms of the vector . With these definitions. consisting of the ( ) identity matrix with ( zeros inserted between each row.. (53) . the resulting solution should be scaled so that . then is the folded . where is the matrix have . . . namely the vector .. Magnitude response of a low-pass IFIR eigenfilter. . and denote.. . is the . Then clearly is ters of orders . As an example (taken from [9]). Now note that the coefficients of can be expressed as Toeplitz convolution matrix .: ON THE EIGENFILTER DESIGN METHOD AND ITS APPLICATIONS: A TUTORIAL and of which are larger than 511 . If we choose (so that ). (52) where and and ) rows of are. . Downloaded on February 7. 12. eigenfilter approach. Chen and Vaidyanathan showed how to design the filter for the specifications of the original filter using the eigenfilter method. . Suppose that for some even . Hence. consisting of the coefficients of the predesigned filter . . and are real Type 1 linear-phase filSuppose that and . if secutive zeros between each coefficient of denotes the vector of coefficients of . respectively. Then note that First note that are obtained by inserting ( ) conthe coefficients of . In [9].. some Type I linear-phase filter of order Then Fig. . . is given by the following: With this. assuming that the image suppressor has already been designed.. the folded version of .. using (52) and (53). we obtain the magnitude response and was chosen to be a low-pass filter shown in Fig. Recall that a good low-pass narrowband the original goal is to make using the filter. we subject to the usual unit norm . . . Namely. . we can now proceed to for the specifications of using show how to design the eigenfilter approach.. . then we .

Downloaded on February 7. Chen was the first to consider using the eigenfilter method for the design of beamformers in [10]. To obtain the same ripple of specifications using a single eigenfilter would require an order of 36 [9]. Spectral/Spacial Beamforming A common problem in array signal processing is that of beamforming [13]. and In contrast. B. As a result. by using the single eigenfilter. of the arc array.512 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—II: ANALOG AND DIGITAL SIGNAL PROCESSING. then the gain pattern can be expressed as Analogous to the multidimensional objective of (51). 9. the problem of steering the gain of the array to accept certain temporal frequencies and spacial directions while rejecting others is analogous to designing a filter to match a desired response. One advantage from using the IFIR approach comes in the cascade form in the savings from implementing . and is the length of the unit delay. For example. . Fig. 13. we clearly have Here. 2010 at 04:56 from IEEE Xplore. Waves from a particular source impinging upon the array are assumed to be in the far field and arrive at the array at an angle of as shown in Fig. sensors that are placed along a circle The array consists of of radius . Restrictions apply. 13. SEPTEMBER 2003 with equal and maximal flatness at [49]. The resulting and a stopband response has a passband ripple of (or ). Magnitude of the gain pattern of an arc array with delay lines. in which a group of waves (such as electromagnetic or audio) impinge upon an array of sensors (such as antennas or microphones) and the goal is to tune the output of these sensors to focus only on a specific set of waves corresponding to a certain temporal frequency arriving from a certain spacial direction. VOL. where is some reference point. 50. we would require multipliers and 36 adders. the th sensor is at an with respect to a certain reference point and has angle delay lines attached to it. As can be seen in Fig. Namely. Each delay path is weighed by the which is chosen at the receiver and used to steer quantity the beam or gain of the array to a desired response. The signals from the various sensors and delay paths are given by the following combined to obtain a gain pattern expression: where is the delay of the th tap of the to wave propagation and taps given by [10] Fig. Arc antenna array with delay lines. In particular. if a radio station is transmitting a signal in an environment with reflective paths between the transmitter and antenna array at the receiver. he considered the design of an arc array with delay lines shown in Fig. the eigenfilter approach can be applied to the problem of beamforming. 13. 14. was chosen to minimize th sensor due As . such as the arc array with delay lines that we shall soon consider. NO. Authorized licensed use limited to: San Diego State University. It turns out that for delay-and-sum beamformers [13]. we only require adders using the IFIR implementation [49]. the receiver may wish to focus only on waves near the carrier frequency of the radio signal coming directly from the transmitter (following a path known as the line-of-sight path). is the speed of the wave in the given medium (such is the radius as the speed of light or the speed of sound). Chen to make best approximate a desired gain chose scaled by a factor . As and are linear phase of orders 18 multipliers and 6. If we define the weight vector where the ’s are simply the ’s arranged according to some order and similarly define the vector where the ’s are just the ’s arranged according to the same order. 13.

As we can see. each attached with ten arc array with ten sensors ( for all ). 17. Suppose that we 2 . 4 . if the channel is of length construction. we want the cyclic prefix length as small as possible. As the channel may have zeros near is usually not or outside the unit circle. the gain in the regions of interest match closely with the desired response. i. The beauty of this redundancy is that it can be shown that in the absence of noise. Downloaded on February 7. the DMT system is able less than or equal to to equalize an FIR channel using only FIR components (namely and the blocking components along with the DFT matrices as well as the frequency-domain equalizer). Instead. Decomposition of the effective channel c (n). we have perfect re. the desired response is a delay parameter which satisfies . . C. The goal is to shorten the channel length . and .e. This is only possible because of the inherent redundancy. From a practical point of view. Let the be chosen to be 0 . FIR filter chosen to concentrate the energy of the effective in a window of length . 17. Then. A typical DMT system is shown in Fig. though the gain pattern takes on a wide range of values in the “don’t care” regions. 6 . 14. whereas we desire zero gain for for all . and . and a residual response which consists of whatever remains of after removing . don't care otherwise.: ON THE EIGENFILTER DESIGN METHOD AND ITS APPLICATIONS: A TUTORIAL 513 Fig. However. Channel-Shortening Eigenfilter Equalizers for DMT Systems Fig. 2010 at 04:56 from IEEE Xplore. . and desire the response to be unity in the range . the optimal can be found using Rayleigh’s principle as before. the effective channel is of . Such an equalizer is typically called a time-domain equalizer or TEQ [43]. is given by Fig. As an example (adapted from [10]). This is illustrated in Fig. delay taps ( . although the channels are typically hundreds of samples long. as this may result chosen to exactly shorten the channel is typically an in spurious noise amplification. suppose that we have an ).. and [43]. as is the case in practical DMT systems such as ADSL in which the channel is a telephone wire line [43]. In other words. Perhaps the most important feature of the DMT system of Fig. Typical DMT system. have become very popular and have revolutionized telephone wireline communications. to concentrate most of the energy of to a length to a window of length . We can decompose as a sum of two responses. since it represents a redundancy which hinders the overall rate of the system by a factor .e. Practical DMT systems such as the asynchronous digital subscriber line (ADSL). Essentially. For example. in ADSL. (n) into a desired channel . 8 . namely a desired response which is exactly of length . as shown in Fig. 16. Recently this communications system has received much attention on account of its low complexity and excellent performance compared to traditional modem systems.TKACENKO et al. the equalizer . Restrictions apply. For the fictitious example in which the optimal gain pattern we obtain is shown in Fig. 15 [43]. i. 15. 15 is the inclusion of redundancy in the form of a cyclic prefix of length [43]. Here. This suggests the need for an equalizer at the receiver which shortens the channel. channel 1) The Channel-Shortening Problem: Suppose that the is and that of the equalizer length of the channel is . With the usual unit norm constraint on . c (n) and residual channel c Authorized licensed use limited to: San Diego State University. the channel may be very long. 10 . Channel/equalizer model. .. . . 16.

. which can be solved via Rayleigh’s principle as before. then as varies over all nonzero vectors. the optimal equalizer was calculated for all in and the actual used was the one which the range yielded the largest SSNR. we have whereas and . 2010 at 04:56 from IEEE Xplore.. As we now show. since it maximizes the shortening signal-to-noise ratio defined by In Fig. 9. Downloaded on February 7. VOL.. at the bottom of the page. and are . which we denote here.. 50. . are simply the following: tively. the channel is carrier-service area (CSA) loop # 1. however. . In this example. If we define the vector .. Hence. chose the equalizer coefficients to minimize subject to keeping the the energy of the residual response at unity. will as well. . [17] is tantamount to the following problem: (55) This problem (in terms of the vector ) is in fact the eigenfilter problem. .3312 dB.. the equalizer is of length 16. Here. respecand . Hence. . along with the equalized effective channel designed using the MSSNR method. . and the deis of length (corresponding to a cyclic prefix length sired length is of 32). Assuming that the matrix is positive definite. and are. . NO. Define the following vectors and matrices: where and are Hermitian positive semidefinite matrices. . Authorized licensed use limited to: San Diego State University. SEPTEMBER 2003 Melsa et al.. the original optimization problem of (55) can still be posed as an eigenfilter problem [17]. The channel . . If the matrix is strictly positive semidefinite. . . we perform a Cholesky decomposition of the matrix [12]. To show that the problem of (55) can be posed as a traditional eigenfilter problem. .514 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—II: ANALOG AND DIGITAL SIGNAL PROCESSING. . . the optimal is found using . By the convolution is a square nonsingular matrix. . . the effective channel appears to be “shortened” to the desired length.. Here. Clearly. energy of the desired response this problem can be posed as an eigenfilter problem.. respectively. a common subscriber loop encountered in an ADSL system [43]. 18. As can be seen. [17]. Once the optimal is found. . as shown in the equation windowing matrices and are . the problem in (55) is equivalent to . Here. The method for channel shortening given in (55) is known as the maximum shortening signal-to-noise ratio (MSSNR) method [5]. the best equalizer occured when was 25 corresponding to a SSNR of 33. the impulse response of a typical channel encountered in a DMT system is shown. . as well as the Now define the following vectors and . then it admits a Cholesky-like decomposition of the form [12] where . by Here we have used the fact that subject to and . and (54) and .. . where we have used the convolution equation of (54). Here. Minimize where . . the method by which this is done is much more complicated. . the design problem considered by Melsa et al. The energy of the desired and residual responses. . so that . Restrictions apply. .

Melsa. Vaidyanathan. The eigenfilter method continues to find applications in several areas of research. SSNR = Fig. Such advantages. J. C. J. Shyu. 1993 [35] S. using IFIR eigenfilters as compaction filters. and R. the eigenfilter method has a much lower design complexity and remains robust even when ill-conditioned matrices are present in the design problem. Recently. C. C. the eigenfilter method was shown to be applicable to the problem of spectral/spacial filtering or beamforming for sensor arrays. 2002 Authorized licensed use limited to: San Diego State University. a similar method was proposed requiring only one Cholesky decomposition (i. and S. T. D. J. Restrictions apply. 1994 [1] F. : () 33 3312 dB = 512 = 16 515 for channel shortening was extended for the design of fractionally spaced equalizers [45]. Q. Pei and J. which is difficult to modify for certain design criteria. Argenti and E.L . Downloaded on February 7. E. E. J. Charalambous. C. Pei. Vaidyanathan and T. Shyu and S. 1978 [50] P. 1993 All-Pass Eigenfilters: [14] T. Shyu. the decomposition did not depend on the delay parameter ). J. [47] as well as for designing channel-shortening equalizers for multiple-input multiple-output (MIMO) channels [46]. 2002 [47] A. together with the myriad of design problems to which it can be applied. Q. Laakso. make it an attractive method to use for filter design. Pei and J. 2001 [11] S. Schur and J. Tkacenko and P. Vaidyanathan. whereas in [45]. the authors proposed a method to jointly shorten the channel and suppress the noise power observed after equalization. Kyriakakis. D. 1987 [18] A. and P. when coupled together with the good performance of eigenfilter method. H. since many design problems can be posed as an eigenfilter problem. Rohrs. 1998 [2] F. C. 1994 [22] T. and C. the eigenfilter method was shown to be useful in a variety of applications. L. CONCLUDING REMARKS The eigenfilter method has been shown to possess several advantages over other traditional filter design methods. C. Nguyen. Kyriakakis. W. Chen. Pei. and using the eigenfilter method to design optimal filter banks. Nguyen. which are useful when narrow transition bands are required. . on account of its numerous merits. 1988 [24] S. J. Pei and J. I. the eigenfilter method SELECTED REFERENCES BY TOPIC Early Works on Eigenfilters: [42] D. 2002 Applications in Channel-Shortening Equalizer Design: [17] P. Current open problems in filter bank theory include using the eigenfilter method to design FIR compaction filters. Tkacenko and P. 1993 [33] S. T. Evans. Here. In contrast to the McClellan–Parks algorithm. Doclo and M. Shyu. 1999 IIR Eigenfilters: [41] J. Q. 2002 Multidimensional Eigenfilters: [18] A. In addition to its numerous strengths. B. Vaidyanathan. R. Shyu. Nguyen. P. Nguyen. Bharitkar and C. Kiaei.: ON THE EIGENFILTER DESIGN METHOD AND ITS APPLICATIONS: A TUTORIAL ( ) and the equalized = 33 1 = 25. Nguyen. Chen. Shyu. C. Simulation results have shown that these methods perform nearly optimally in terms of observed bit rate (the primary figure of merit for DMT systems).TKACENKO et al. ). C. I. C. 1996 Applications in Spectral/spacial Filtering or Beamforming: [10] T. P. 2010 at 04:56 from IEEE Xplore. Nashashibi and C. the eigenfilter method has been shown to appear in the design of certain optimum compaction filters [51]. Tkacenko and P. Pei and J. Del Re. Also. Zhang and H. Moonen. Koilpillai. 1993 [10] T. 2001 [45] A. 1990 [30] S.. J. P. Pei and J. V. 1994 [8] H. Charalambous. 1993 [21] T. C. Iwakura. Q. 1993 [6] S. the eigenfilter method can easily be modified to satisfy a plethora of design constraints. Del Re. the method was found to be useful for the design of channel-shortening equalizers. In addition. Recently. Speidel. In [47]. As can be seen. Pei and J. which requires the computation of a matrix inverse which may be susceptible to numerical inaccuracies. As opposed to the least-squares approach. 1988 Arbitrary Desired Response Eigenfilters: [20] T. 1988 [26] S. 1996 [40] R. C. Pei and J. J. Laakso. and R. Bharitkar and C. C. Shyu. Finally. 2001 [5] G. 1991 [29] S. Younce. Shyu. Chen and G. (L . Hsu. a novel eigenfilter method called the minimum intersymbol interference (min-ISI) method was proposed by Arslan et al. Slepian. n Several generalizations of the MSSNR method for channel shortening have been proposed recently. Arslan. Ford. C. 2002 [46] A. Argenti and E. Wang. 18. Impulse responses of the original channel c channel c n . 1992 [32] S. which exploited the DMT subchannels which cannot be used to improve the DMT channels used for joint channel shortening and noise suppression. in which we wish to focus on temporal frequencies in a particular band arriving from a particular spacial direction. especially the min-ISI method.e. the eigenfilter method is and will continue to be a versatile design algorithm. J. 2000 [7] S. Koilpillai. Q. 1994 [55] X. which are commonly needed in popular communications systems such as ADSL. the method was shown to be useful for designing model filters for multistage IFIR filters.L . P. 1993 [31] S. 1998 [39] S. Nashashibi and C. Pei and J. [5]. J. Shyu.

Ikehara.” Proc. vol. “Interpolated th-band filters for image size conversion. 247–251. Englewood Cliffs. “Eigen-approach for designing FIR filters and all-pass phase equalizers with prescribed magnitude and phase response. . pp. R. 11–23. M Authorized licensed use limited to: San Diego State University. Signal Processing. and S. Cambridge. J. [10] T. Ford.” IEEE Trans. Commun. 40. II. 487–494. Vaidyanathan. Englewood Cliffs. IEEE ICASSP ’02. 698–701. no. . 2001. I. 2. R. IEEE ICASSP ’99. June 1984. “The design of 2-D digital filters by transformations. 35. 46. “Analytic closed forms for designing higher order digital differentiators by eigen-approach. Acoust. M. Circuits Syst. 1993. 1987. “Designing allpass filters using the eigenfilter method.” IEEE Trans. May 2002. 1985. IEEE ICASSP ’91.” IEEE Trans.. 1. Signal Processing.” IEEE Trans. P. 3028–3035.. P. . Y. T. IEEE ICC ’01. June 1998. 2886–2890. Fourier analysis. 50.” Proc. vol. Audio Electroacoust. W. U. Feb. pp. vol. 73–76. C. Kiaei. 2002. Circuit Theory. Circuits and Systems. Iwakura. 3rd IEEE SPS-2002. W. C. pp.” IEEE Trans. 1973. 9. Younce. P. pp. Shyu and S. “2-D FIR eigenfilters: A least-squares approach. P. IEEE ISCAS ’91. 4171–4174. and W. vol. 1477–1480. 48. Pei and C. Monterey. Wisutmethangoon and T. 7th Annu. 966–968. pp. Yang. “Eigenfilters for signal cancellation. K. Pei. and R. Jan. Mar. vol. Johnson and D. “Fast and stable least-squares approach for the design of linear phase FIR filters. pp. 1996. 158–162. 1993. Vaidyanathan and T. C. 3. Bharitkar and C. pp. 41. 1973. 1993. 37. 1994. Apr. Q. Pei. II. pp. Chen. 554–559. 42. . H. 505–511. pp. Pei. Princeton Conf. . vol. H. Signal Processing. 3. E. 3123–3135. IEEE ICASSP ’98.” Proc. pp. Understanding Digital Subscriber Line Technology. vol. P. Jan.” IEEE Trans. 44. Pei and J. Mar.” IEEE Trans.” Proc. IEEE ICASSP ’02. 2001. “Noise optimized eigenfilter design of time-domain equalizers for DMT systems. “Eigenfilter design of mimo equalizers for channel shortening. II. vol.” Proc. vol. Englewood Cliffs. vol. NJ: Prentice-Hall. “A unified approach to the design of optimum FIR linear-phase digital filters. 1999. Downloaded on February 7. “Eigenfilters: A new approach to least-squares FIR filter design and applications including Nyquist filters. C. “Design of IFIR eigenfilters. Asilomar Conf. pp. J. Mitra. pp. 2.” IEEE Trans. [22] T. 1994. June 2001. D. vol. 1993. [19] Y. Chen and P.” IEEE Trans. Shyu. 1992.. Wang. Q. C. “Correlation and spatial sensitivity of eigenfilters for selective [7] signal cancellation in multiple-listener environments. [28] [29] [30] [31] [32] [33] [34] [35] [36] [37] [38] [39] [40] [41] [42] [43] [44] [45] [46] [47] [48] [49] [50] [51] [52] [53] [54] [55] . Sept. Englewood Cliffs. 1–5. Apr. Nashashibi and C. Signal Processing. Q. J. 47. vol. IEEE ICASSP ’93. Signal Processing. NJ: Prentice-Hall. and C. Johnson. Mar. pp. Signal Processing. 1528–1543. II. 1981–1984. Restrictions apply. pp. Circuits Syst. C. 1998. Signal Processing. 2010 at 04:56 from IEEE Xplore. 3. Hsu. Starr. II. 1994. Q. Doclo and M. Sept. T. C. Dudgeon. pp. Circuits Syst.” IEEE Trans. no. 34. S. Signal Processing. [16] J. pp.. P. Cioffi. vol.” IEEE Trans.” in Proc. 1485–1493. Circuits Syst. “Eigenfilter design of real and complex coefficient prototypes for [3] uniform and nonuniform filter banks. vol. Circuits Syst. 1128–1139. vol. X.516 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—II: ANALOG AND DIGITAL SIGNAL PROCESSING. vol. Array Signal Processing. Video Technol. .” IEEE Trans. IEEE ISCAS ’88. May 1994. [15] J. [5] G. . vol. D. 1457–1461. 2001. A. [20] T. pp. M .” IEEE Trans. Circuits Syst. Dec. vol. E. Press. 35th Midwest Symp. vol.” IEEE Trans. NO. . vol. “A new eigenfilter based method for optimal design of channel shortening equalizers. Tufts and J. Parks. [12] R. 1992. Dec. May 2002. C.” Proc. Apr. “Unified eigenfilter approach: With applications to spectral/spatial filtering.” IEEE Trans. 1988. Rohrs. [14] T. and Computers. 1990. 1996. pp. vol. 1993. Circuits Syst. pp. pp.” Proc. . and R. pp. vol. Speidel. . 1991. vol. Signal Processing. Y. pp. 1371–1430.” IEEE Trans. “The eigenfilter for the design of linear-phase filters with arbitrary magnitude response. pp. Nguyen. vol. 44. [25] . pp. 45. Argenti and E. pp. 1037–1040.” IEEE Trans. J. 1993. S. T. vol. Apr. Oct.” IEEE Trans. 2002. Charalambous. II. Q. and S. “Eigenfilter approach for the design of allpass filters approximating a given phase response. Nguyen. NJ: Prentice-Hall. 2000.” Proc. 2. McClellan and T. Nov. SEPTEMBER 2003 REFERENCES [1] F. 962–966.K: Cambridge Univ. Okuda. pp.” in Proc. [18] A. Dong. Koilpillai.” Proc. D. Aug. 622–626. IEEE ISCAS ’92. A. pp. Vaidyanathan. J. . S. “Eigenfilter design of 1-D and 2-D IIR digital all-pass filters. pp. II. and P. submitted for publication. Signals.” Proc. II. E.” IEEE Trans.” Proc. “2-D FIR eigenfilters. Tkacenko and P. Therrien. P. “A method for design of th-band filters. 563–570. 361–364. vol.” in Proc. Horn and C. vol. 46. pp. IEEE ISCAS ’02. 4. Circuits Syst. [9] T. “Design of IIR multi-band filters using IIR all-pass eigenfilters. Matrix Analysis. Feb. 46. 43. Arslan. “Design of IIR eigenfilters with arbitrary magnitude frequency response. pp. 1992.. “Design of a class of IIR eigenfilters with time. Silverman. Nguyen. NJ: Prentice-Hall. “FIR filter design with linear constraints using the eigenfilter approach. pp. “Impulse response shortening for discrete multitone transceivers. Apr. [8] H. 264–267. Aug. 24–34. [6] S. 1996. “Design of 2-D FIR digital filters by Mcclellan transformation and least squares eigencontour mapping. L. Nguyen. 699–709. Systems. vol. D. “Toward optimal least square filters using the eigenfilter approach. June 1998. I. J.” IEEE Trans. Nov. 57. [17] P. pp. 573–576. “The design of arbitrary FIR digital filters using the eigenfilter [21] method. 1991.. and S.” Proc. S. Signal Processing. Circuits Syst. Circuits Syst. 40.. Feb. P. “Eigenfilter design of higher-order digital differentiators. 1993. “Design of two-dimensional FIR eigenfilters for sampling-structure conversion.” IEEE Trans. C. vol. pp. pp. 1998. Francis. C.. W. “Comparison of least-squares and eigenfilter techniques for broadband beamforming. Multirate Systems and Filter Banks. vol. 137–145. pp. pp. pp. Moonen. CT-20. vol. Discrete Random Signals and Statistical Signal Processing. 1. C. IEEE ISCAS ’93.” IEEE Trans. Signal Processing.” Bell Sys. 2002. 1988. 3.” IEEE Trans. vol. B. AU-18. C. “Prolate spheroidal wave functions.and frequency-domain constraints. 47. 1989. vol. 1999. McClellan. vol. [24] S. Jan. Zhang and H. 145–151. 1662–1672. and uncertainty – V: The discrete case. 1994. vol. Re.” Proc. 601–604.. 1993. 1694–1698. 1265–1268. . J. Takahashi. 232–237. I. T. 2002. Q. Schur and J. 42. “Interpolated finite impulse response filters. R. vol. vol. pp. vol. 331–334. pp. 1970. “Design of FIR Hilbert transformers and differentiators by eigenfilter. [2] Signal Processing. 3. Circuits Syste. .” IEEE Trans. Tseng. CA. “A new eigenfilter based on total least squares error criterion. June 1999. 697–701. Tech. Information Science and Systems. Nov. D. H. VOL. 546–555. Neuvo. 41. Kyriakakis. 42. Nguyen. 47. Signal Processing. Laakso. Signal Processing. 4. pp. Inc. [27] . Vaidyanathan. “A unified approach to the design of quadrantally symmetric linear-phase two-dimensional FIR digital filters by eigenfilter approach. Melsa. “An efficient equalization method to minimize delay spread in OFDM/DMT systems. vol. 2002. pp. “Designing digital lowpass filters: Comparison of some methods and criteria. 32–40. 49. Chen. H. Zhang and T. “Design of 1-D and 2-D IIR eigenfilters. [23] M. Evans. S.” Proc. “Design of IIR eigenfilters in the frequency domain. “A unified eigenfilter approach to the design of two-dimensional zero-phase FIR filters with the Mcclellan transform. 2000. and P. [11] S.. “Analytic closed-form matrix for designing higher order digital differentiators using eigen-approach.” IEEE Trans. Slepian. 787–792. 1. Dec. Tseng. vol. [13] D. Speech. 3. 39. .” IEEE Trans.” IEEE Trans. Koilpillai. “Design of IIR digital allpass filters based on eigenvalue problem. May-June 1978. 1992. Circuits Syst. June 1998. Speech. “Theory of optimal orthonormal subband coders. pp. 50.” IEEE Trans. M. “Complex eigenfilter design of arbitrary complex coefficient FIR digital filters. Acoust. Vaidyanathan. 77–80. Chen and G. 1999. 2257–2263. Yang and T. pp. “Eigenfilter design of real and complex coefficient QMF [4] prototypes. vol. pp. C. 49. “Design of arbitrary FIR log eigenfilters. IEEE ICC ’02. Jun. “Equalization for discrete multitone transceivers to maximize bit rate. 42. 1. Dec. vol.” IEEE Trans. IEEE ISPACS ’00.” IEEE Trans. . C. Signal Processing. 1578–1582. 37.” Proc. [26] . Circuits Syst.” IEEE Trans. vol. Nguyen. Laakso. pp. Aug. pp. 1669–1678.

Tech. he was a Visiting Lecturer at MIT and an Adjunct Professor at Northeastern University. His main research interests are in digital signal processing. 1954. on Oct. as an Assistant Professor. He received the IEEE TRANSACTIONS ON SIGNAL PROCESSING Paper Award (in the image and multidimensional processing area) for the paper he co-wrote with Prof. SPCOM’95.D degrees in electrical engineering from the California Institute of Technology. In 1999. He has published over 100 publications. 1986 and 1989. and M. Vaidyanathan on linear-phase perfect-reconstruction filter banks in 1992 and the National Science Foundation Career Award in 1995. wavelet transforms. He has written several chapters for various signal processing handbooks and has authored a number of papers in IEEE journals. K. in 1974.) degree in physics and the B. and since 1993 has been Professor of Electrical Engineering there. 1993). Dr. During the academic year 1993-1994. from August 1996 to June 2001. Vaidyanathan was elected Fellow of the IEEE in 1991. NJ: Prentice Hall. Boston. and Asilomar’88 conferences on signal processing. . His research interests include multirate systems. P. Dr. Boston. 2010 at 04:56 from IEEE Xplore. respectively. and 1979. He also received the National Science Foundation’s Presidential Young Investigator Award in 1986. Pasadena. M. Terman Award of the American Society for Engineering Education.S. In 1998. the IEEE Acoustics. In March 1983 he joined the Electrical Engineering Department.S. 1997) and the author of several matlab-based toolboxes on image compression. Santa Barbara. In 1999 he was chosen to receive the IEEE CAS Society’s Golden Jubilee Medal. wavelet transforms.TKACENKO et al. He was a post-doctoral fellow at the University of California. multirate systems. He served as Vice Chairman of the Technical Program Committee for the 1983 IEEE International symposium on Circuits and Systems. He is currently an Associate Editor for the IEEE TRANSACTIONS ON IMAGE PROCESSING. he was awarded the Graduate Division Fellowship from Caltech. all from the University of Calcutta. He is the author of the book Multirate Systems and Filter Banks (Englewood Cliffs. P. MA: Wellesley-Cambridge Press. Dr. and the S. He is currently a Professor at the Electrical and Computer Engineering Department. Restrictions apply. he was with the Electrical and Computer Engineering Department. He received the B.. Lexington. Pasadena. and M. as a Member of Technical Staff. and Ph. He was with Boston University.: ON THE EIGENFILTER DESIGN METHOD AND ITS APPLICATIONS: A TUTORIAL Andre Tkacenko (S’00) was born in Santa Clara. Nguyen served as an Associate Editor for the IEEE TRANSACTIONS ON SIGNAL PROCESSING from 1994-1996 and for the IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS from 1996-1997. and a Consulting Editor for Applied and Computational Harmonic Analysis. sponsored by Hewlett Packard Company. Authorized licensed use limited to: San Diego State University.S. from September 1982 to March 1983. University of Wisconsin. and IEEE Signal Processing Letters. 1977. degree in the field of digital signal processing at Caltech. He is coauthor (with Prof.S. from June 1989 to July 1994. and signal processing for digital communications. 1977. and 1993–1994. From August 1994 to July 1998. wavelets. 16. He is currently working toward the Ph. IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—II: ANALOG AND DIGITAL SIGNAL PROCESSING. 1992–1993. (Hons. on the topics of filter banks. CA. and as the Technical Program Chairman for the 1992 IEEE International Symposium on Circuits and Systems.. Speech.Tech. He received the B. Pasadena. He is currently the Series Editor (Digital Signal Processing) for Academic Press. Downloaded on February 7. respectively. India. He was a recepient of the Award for Excellence in Teaching at the California Institute of Technology in 1983–1984. and is currently an Associate Editor for IEEE Signal Processing Letters. and subband coders.D. Calcutta. degrees in electrical engineering from the California Institute of Technology (Caltech). he was a Guest Editor for special issues of the IEEE TRANSACTIONS ON SIGNAL PROCESSING and the IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—II: ANALOG AND DIGITAL SIGNAL PROCESSING. He is a recepient of the IEEE Signal Processing Society’s Technical Achievement Award for the year 2002. University of California at San Diego. in 1985. He was also the coauthor of a paper on linear-phase perfect reconstruction filter banks in the IEEE SP Transactions. E. MA. electrocardiogram compression. His research interests include video processing algorithms and their efficient implementation. Nguyen (S’85–M’85–SM’95) received the B.D degree in electrical and computer engineering from the University of California at Santa Barbara in 1982. respectively. 517 Truong Q. Mitra Memorial Award from the Institute of Electronics and Telecommuncations Engineers. Eusipco’98. He received the 1995 F. Vaidyanathan (S’80-M’83-SM’88-F’91) was born in Calcutta.Sc. MA. Calfornia Institute of Technology. degrees in radiophysics and electronics. and Signal Processing Senior Award for his paper on multirate perfect-reconstruction filter banks in 1989. and filter bank design. Vaidyanathan was an Associate Editor for the IEEE TRANSACTIONS ON CIRCUITS AND Systems during 1985–1987. for his joint paper in the IETE Journal in 1990. He has given several plenary talks including at the Sampta’01. G. He was with MIT Lincoln Laboratory. Strang) of the textbook Wavelets and Filter Banks (Wellesley. MA. He has been chosen a distinguished lecturer for the IEEE Signal Processing Society for the year 1996-97. for which the first author (Truong Nguyen) received the Young outstanding author award in 1993. on February 24. India. for his contributions to engineering education. Madison. especially the book Multirate systems and filter banks published by Prentice Hall in 1993. in 1999 and 2001. India. and the Ph. and their applications in digital communications and data compression.