986

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 47, NO. 4, APRIL 1999

A General Formulation of Modulated Filter Banks

Peter Niels Heller, Member, IEEE, Tanja Karp, and Truong Q. Nguyen, Senior Member, IEEE
Abstract This paper presents a general framework for maximally decimated modulated lter banks. The theory covers the known classes of cosine modulation and relates them to complexmodulated lter banks. The prototype lters have arbitrary lengths, and the overall delay of the lter bank is arbitrary, within fundamental limits. Necessary and sufcient conditions for perfect reconstruction (PR) are derived using the polyphase representation. It is shown that these PR conditions are identical for all types of modulationmodulation based on the discrete cosine transform (DCT), both DCT-III/DCT-IV and DCT-I/DCT-II, and modulation based on the modied discrete Fourier transform (MDFT). A quadratic-constrained design method for prototype lters yielding PR with arbitrary length and system delay is derived, and design examples are presented to illustrate the tradeoff between overall system delay and stopband attenuation (subchannelization). Index TermsCosine-modulated lter bank, DCT, lter bank, MDCT, modulated lter bank.

I. INTRODUCTION ULTIRATE lter banks are used in many applications, ranging from data compression (speech, audio, image, and video) to multicarrier modulation (xDSL) and feature detection. Fig. 1 shows an -channel maximally decimated lter bank with the corresponding nearly ideal lter responses. The input signal is bandpass ltered and downsampled to yield subband signals. The subband critical sampling of the signals can be processed independently to extract essential information, and they can be combined to yield a reconstructed signal, which may be similar to or very different from the original signal. A lter bank where the output is a delayed copy of the input is a perfect reconstruction (PR) system. The theory and design of such PR lter banks have been extensively studied [13], [31]. An important subclass are the modulated lter banksthose -channel lter bank is obtained by for which the full modulation of a single lowpass prototype lter. The rst examples were the pseudo-QMF banks [14], [22], [25], which offered approximate but not true PR. The rst PR lter bank where the analysis and synthesis lters were limited to a was presented in [24]. Later work extended length of 2 the idea to longer lters and other modulation schemes based
Manuscript received April 6, 1994; revised April 17, 1997. This work was supported in part by Deutsche Forschungsgemeinschaft under Grant Fli 116/10 and the National Science Foundation under Grant MIP 9501589. The associate editor coordinating the review of this paper and approving it for publication was Prof. Roberto H. Bamberger. P. N. Heller is with Aware Inc., Bedford, MA 01730-1432 USA (e-mail: heller@aware.com). T. Karp is with Mannheim University, Mannheim, Germany (e-mail: karp@rumms.uni-mannheim.de). T. Q. Nguyen is with the Department of Electrical and Computer Engineering, Boston University, Boston, MA 02215 USA (e-mail: nguyent@bu.edu). Publisher Item Identier S 1053-587X(99)02139-X.

on the Type IV discrete cosine transform or DCT-IV [8], [10], [12], [15], [24] and the modied discrete Fourier transform (MDFT) [5], [9]. These lter banks offer the advantage of high computational efciency (employing fast DCT or DFT algorithms) and high-performance lter designs [17], [21]). The rst fundamental constructions of modulated lter banks were based on a linear-phase prototype lter of length with reconstruction delay 2 1. Later work ex2 tended the PR constructions to arbitrary length prototypes [19] and DCT-II modulation [7], [11]. Systems with linearphase prototype lters face a tradeoff between latency and lter performanceshorter lters will decrease the system latency at the cost of poorer subchannelization (as measured by stopband attenuation). Several authors [17], [18], [20], 1 for the [30] have introduced more general values 2 reconstruction delay, which force the modulated lter bank to be biorthogonalthe synthesis lters will no longer be timereversed versions of the analysis lters. The authors of [29] and [34] describe the construction of pseudo-QMF (near-perfectreconstruction) and PR modulated lter banks, respectively, with fully general reconstruction delay. Such biorthogonal modulated lter banks ameliorate but do not eliminate the tradeoff between system latency and subchannelization. Each of these extensions offers advantagesDCT-II modulation yields modulated lter banks with linear-phase lters, which is useful in image compression, while reduced reconstruction delay is important for applications to speech, audio [28], and telecommunications. This paper combines and extends the cited results into a single framework that encompasses PR with arbitrary reconstruction delay and all known modulation schemes (DCT-IV, DCT-II, and MDFT). In particular, the following results are demonstrated: necessary and sufcient conditions for prototype lters with arbitrary length and reconstruction delay to yield PR biorthogonal modulated lter banks; variable overall system delay (within the fundamental limits imposed by the lter lengths); arbitrary modulation with DCT-IV, DCT-II, or MDFT schemes; equivalence of the various modulationsthe same conditions for PR apply to all the modulation schemes; a prototype that satises these conditions will yield PR in any scheme; a quadratic-constrained least-squares method for highperformance arbitrary-delay PR modulated lter bank designs; an application to subband coding of audio signals. Outline of the Paper: The paper begins with cosine-

1053587X/99$10.00 1999 IEEE

HELLER et al: GENERAL FORMULATION OF MODULATED FILTER BANKS

987

Fig. 1.

M -channel maximally-decimated lter bank.

where . Note that the modulation does not depend on the lter length but only on the delay of the system. is an odd integer, this modulation can be When the delay is even via a DCTimplemented via a DCT-IV, and when leads only to a rearrangement of the III. The phase term DCT terms and not a fundamental change in the type of cosine transform (see [31]). In this section, we derive the equations , and must satisfy for PR; that the prototype lters design examples will be presented in Section V. th Begin by writing out the prototype lters in their polyphase component form ([13], [31]) using type-1 polyphase and for both the analysis and the synthesis lters

modulation based on the DCT-III/DCT-IV. Section II derives necessary and sufcient conditions for PR for a general biorthogonal DCT-III/IV cosine-modulated lter bank, using the polyphase representation. The analysis and synthesis lters are allowed to have arbitrary length and are not necessarily linear phase. The overall delay of the system is arbitrary within fundamental limits and is incorporated in the modulation process. Section III presents a modulation scheme based on the DCT-I/DCT-II and demonstrates that the PR conditions of the previous section apply equally well to prototype lters for DCT-I/DCT-II modulation. Section IV presents two types of MDFT lter banks and demonstrates the equivalence of the MDFT scheme with DCT-I/II modulation for realvalued signals. Section V formulates the prototype lter design problem using a quadratic-constrained approach and presents design examples that exploit the freedom of the arbitrarydelay, arbitrary-modulation framework. Finally, Section VI gives an application of a low-delay lter bank to MPEG-like audio compression. Notations: Boldfaced letters indicate vectors and matrices. Superscript denotes transposition, and the tilde on a function is dened such that , , where the subscript denotes the conjugation of coefcients asterisk and (of the polynomial in each matrix entry). Moreover, denote the identity and reverse identity matrices, respectively. Unless otherwise noted, the dimension of and is . denotes the element of . II. DCT-IV COSINE-MODULATED FILTER BANKS Consider the general case of a biorthogonal lter bank with DCT-III or DCT-IV modulation. The impulse responses of are cosine-modulated versions of a the analysis lters of length , and the synthesis lters prototype lter prototype are obtained via cosine modulation of a length . The overall delay of the lter bank can be xed lter . For a arbitrarily in the range (where ), the analysis given delay and synthesis lters are given by

The lengths of the polyphase lters and are determined by the prototype lter lengths and , respectively. Dene the diagonal matrices diag and diag , and let denote the DCT modulation matrix

The polyphase form of the analysis lter bank is , where now is the analysis polyphase can matrix. The synthesis polyphase matrix , be written as diag , where diag , and denotes the DCT modulation matrix

(1)

The corresponding vector of synthesis lters

is

(2)

988

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 47, NO. 4, APRIL 1999

Using the above matrix

and is

, the overall polyphase transfer

same positions as are

in (6). The nonzero entries of

Now, it can be shown (see Appendix A) that form

has the and the nonzero entries of (3) are

(7)

Thus, the overall polyphase transfer matrix is (8)

(4) (9) However, for the lter bank to satisfy PR [31], have the form must for (5) and in the ranges and for some integers . In this case, the PR lter bank reproduces samples. the input signal with a delay of Recall that we dened the lter bank reconstruction delay to samples, where . Hence, be , , and , whereas if if , , and . Thus, for the cosine-modulation scheme to satisfy PR with delay , must satisfy and

for of both

. A nonzero entry occurs at the same position and only when

(6)

and if (10)

given in (4) with (6), we Comparing the description of obtain necessary and sufcient conditions on the polyphase and for PR. In order to express them components in (4) into the sum of two explicitly, we rst split , where matrices

and if (11)

and (12) and are integers. Note that where the tuples is an integer tuple if and have the same parity, is an integer tuple if is even. For all other and and do not have common nonzero entries. cases, In the following, we rst derive the necessary and sufcient and conditions on the polyphase components for . then give the corresponding ones for

and consider the position of the nonzero entries of , , and . As can be easily veried, the nonzero entries are on the antidiagonal, and those of are in the of

HELLER et al: GENERAL FORMULATION OF MODULATED FILTER BANKS

989

PR Constraints for with and have to solve

: For and from (10) and (12), respectively, we (13)

and are arbitrary for even (21) In both cases, the nonarbitrary polyphase components can have only one nonzero coefcient. Choosing

and (14) (15) in order to fulll (6). These are necessary and sufcient conditions for the lter bank to satisfy PR. From (7) and (13), , we obtain after substituting for

(20) and (21) t into the framework derived for polyphase components satisfy (18). When must satisfy

, and the IN,

(22) The th and th polyphase components of each prototype cannot have common zeros away from the origin, for then, the could polyphase matrix would fail to have full rank, and not have the form (5). Thus, the polyphase components and must be multiples of one another, and we obtain a necessary condition for the analysis and synthesis type-1 polyphase components

(23) By (22), this polynomial contains only even powers of ; thus, we can split the right-hand side of (23) into two separate conditions (24)

(16) and were supposed Recall that both prototype lters to be causal in (1) and (2), respectively. Thus, we either have , or we have to perform a zero padding by to choose zeros on one set of polyphase lters, depending on the sign of . For PR, we still have to solve (14) and (15). Combining them with (8) and (9) and again substituting for yields

(25) Equation (24) has the same form as (16), and (25) the same , and thus, the same conclusions form as (17) for can be drawn as for general . However, we can state from and can only have one (25) that both nonzero entry. : For , the PR Constraints for necessary conditions on the polyphase lters can be derived , also resulting in a similar way as described for in (16) and in the equations

(17)

(18) We now consider the special cases and are integers. For solve Using (7) and (9) , assuming that , we have to .

(26)

(27) and, for (19) resulting in and are arbitrary for odd (20) IN and are arbitrary for and are arbitrary for odd (29) even (28)

990

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 47, NO. 4, APRIL 1999

where, again, we have the freedom to choose

We can state these results as a theorem. Theorem 1: Given a DCT-III/IV modulated lter bank as , sufcient in (1) and (2) with an overall delay for conditions on the analysis polyphase components as PR can be formulated for the case

(30)

(31) and for as

Consequences of the PR Condition: Theorem 1 states the PR conditions for the biorthogonal case with arbitrary length and delay. As in the paraunitary case [19], some combinations of lter length , delay , and number of channels can and have impose restrictions on the lter coefcients. If the same parity, we have seen that two analysis polyphase components, as well as two synthesis polyphase components, are forced to contain only one nonzero coefcient, whereas two other analysis and synthesis polyphase components can be chosen arbitrarily. For even , another two analysis and synthesis polyphase components may only contain one nonzero coefcient. Thus, only for lter banks with an even number of subbands and an odd delay are all PR constraints given by the most general conditions (17), (18) and (26), (27). Further restrictions on the polyphase lters might arise if they are of different length, i.e., if the prototype lter length is not . These restrictions can limit the achievable a multiple of stopband attenuation of the lters. III. DCT-II COSINE MODULATED FILTER BANKS It is also possible to modulate a lowpass prototype lter using a DCT-I or DCT-II. In this section, we present such a scheme and derive conditions for PR, generalizing the approach in [11]. The necessary and sufcient conditions on the prototype lters for PR are found to be identical to those stated in Theorem 1 for DCT-III/DCT-IV modulation. However, a DCT-I/DCT-II lter bank must have channels, with each subband away from the frequencies 0 and being covered twice, as noted in [11], and the lter bank has an . In addition, if increased system delay of and have linear-phase symmetry, the prototype lters then so are all the constituent lters of the DCT-I/DCT-II lter and , we distinguish bank. Depending on the parity of two different cases. A. and Have Same Parity

(32)

(33) The corresponding synthesis polyphase components then given by are

(34) , further solutions exist, as stated in (20) and (21), For for and in (28) and (29) for . , they form Together with the previous equations for the set of necessary and sufcient conditions on the polyphase lters for the modulated lter bank to satisfy PR. , some zero padding has to Note that when choosing be introduced for the polyphase lters in order to keep both and , , prototype lters causal. For the analysis and synthesis prototype lters are the same. Derivation of the Paraunitary Case from the Biorthogonal are Case: In the paraunitary case, the synthesis lters the ipped and conjugated versions of the analysis lters : and . is This relationship is satised if the analysis prototype and has linearchosen identical to the synthesis prototype . As shown in [10], [12], phase symmetry and [19], this reduces the conditions for PR stated in Theorem 1 to the well-known single equation

and have the same parity (i.e., is even, When as in [11]), we dene the analysis lter bank by means of DCT-I modulation as

(35)

(36) The corresponding synthesis lter bank is

(37)

(38) with and restricts the lter bank to have a delay of samples. if or otherwise. (39)

HELLER et al: GENERAL FORMULATION OF MODULATED FILTER BANKS

991

We use the convention that and are nonzero for and zero otherwise. The polyphase -band lter banks can be described with matrices of the and are the aid of several xed transform matrices. matrices with entries

As shown in Appendix A, when

(42) whereas if

(43) and are matrices with entries As in the DCT III/IV modulation, two matrices splits into a sum of with

and

We will also need the matrices

and

dened by

and

as well as The nonzero entries of and can be written are again on the antidiagonal

and

with , , , , as before. Then, the and synthesis polyphase matrices are

analysis (44)

(40) Thus, the overall polyphase matrix of size is

while those of (41). For

are at the same positions as for

in

For PR to hold with a system delay of must satisfy

, (45)

(41)

(46)

992

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 47, NO. 4, APRIL 1999

and for

(47)

PR Constraints for : For , , , we obtain from (41) and (44) the same necessary conditions on the polyphase lters as already stated in (16) for the DCT-IV lter bank and from (47) and (48) in connection with (41) the same conditions as in (26) and (27) on the polyphase lters. , we obtain For

(48) B. . Common nonzero entries of when and if and (49) and and if and (50) and are integers if and have The tuples the same parity, which is the case at hand. : For , PR Constraints for , we obtain from (41) and (44) the same necessary conditions on the polyphase lters as already stated in (16) for the DCT-IV lter bank, whereas from (45) and (46) in connection with (41), we obtain the same conditions as in (17) and (18). However, note that and are different for DCT-IV and DCT-I banks. , we obtain from (41) the rst case of (44) and For the second case of (45) and arbitrary For and , adding from the second case of (44) from (46) yields, in connection with (41) and arbitrary These results for and demonstrate that , , , and are all constrained to contain just one nonzero coefcient and can also be described by (16) and (18). and are also of dimension with entries (53) and (52) (51) and occur and Have Different Parity

and have different parities ( is odd), When we dene a slightly different pair of lter banks based on the channels: DCT-II, each with

(54) with from (39). For a matrix description, we need the matrices and :

With these denitions, the polyphase matrices are still given by (40) (with substituting for

and , etc.), and

HELLER et al: GENERAL FORMULATION OF MODULATED FILTER BANKS

993

Fig. 2. Type-1 MDFT FB. The total number of analysis and synthesis lters is 2 . Depending on the subband, the real part is taken alternately from the zeroth or rst subband polyphase component. The imaginary part is taken from the remaining subband polyphase component.

, , and have the same form as stated in (41) and (44)(48), respectively, resulting in the same conditions and from (49) and (50) are not integers, for PR. Since and do there is no need to consider them, and not have any common nonzero entries. We can summarize the results derived in this section in the following theorem. -channel lter bank constructed Theorem 2: Given a by the DCT-I modulation (35)(38) or DCT-II modulation and with overall (51)(54) of the prototype lters , then the necdelay essary and sufcient conditions on the analysis and synthesis and , polyphase components for PR are the same as those derived for the DCT-III/IV modulation in Theorem 1. IV. MDFT FILTER BANKS In this section, we present two types of modied DFT (MDFT) lter banks, derive the PR conditions for the prototype lters, and relate them to the ones already derived for cosine-modulated lter banks. Historically, the rst modulated lter banks known to offer an efcient implementation were the complex-modulated FBs. They are also called DFT FBs since they can be realized in a highly efcient form by means of polyphase lters and an IDFT transform [1]. The -channel DFT lter analysis lters and synthesis lters of a bank are obtained by complex modulation of the prototype and . However, it is well known that the lters -channel DFT FB with FIR analysis critically subsampled and synthesis lters satises PR only in the case where the [31]. prototype lters are rectangular windows of size This case corresponds to a pure block transform. The MDFT lter bank [5], [6] is a critically sampled -channel complex-modulated lter bank derived from a DFT FB with oversampling factor 2 by introducing several modications in the subbands. a two-step decimation of the subband signals: After dec(corresponding to an overimating by the factor subbands is sampling factor of 2), each of the decomposed into two (even and odd) polyphase components that are complex valued.

Critical subsampling is obtained by taking the real part of one polyphase component and the imaginary part of the other polyphase component in each subband and alternating from one subband to the next. On the synthesis side, the subband signals are rst reconstructed from these 2-ary polyphase components and then processed by the oversampled DFT synthesis FB. Depending on the complex modulation of the analysis and synthesis lters, we get two different types of MDFT lter banks that differ in the phase of the modulation and the way one takes the real and imaginary part of the subbands polyphase components. In the following, we assume a complex-valued input signal. The case of real input signals will be discussed at the end of the section. A. Type-1 MDFT Filter Banks In this section, we generalize the results shown for the paraunitary case in [9]. The complex analysis and synthesis lters are obtained from the prototypes and , respectively, by

where . After ltering with the analysis lters , each subband signal is broken into even and odd polyphase components. For even-indexed subbands, we take the real part of the even subband polyphase component and the imaginary part of the odd polyphase component and vice versa for odd-indexed subbands, as in Fig. 2. satisfyTheorem 3: Given an analysis prototype lter ing the PR conditions (30) and (31) or (32) and (33) as well as (34) for an -channel lter bank with DCT-III/IV modulation samples, then the channel and an overall delay of MDFT lter bank will also be PR, and the overall delay of the lter bank will be samples.

994

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 47, NO. 4, APRIL 1999

Proof of Theorem 3: From [9], we know that the lter can be written as1 bank output signal

(55) denotes the transform of the conjugated complex where odd and the input signal. The + sign is valid for even. When expressing the analysis and synthesis sign for and , respectively, lters by the prototype lters the second part of the sum in (55) cancels (see Appendix B) so that the output signal is given by
Fig. 3. Type-2 MDFT FB. The real part is always taken from the zeroth subband polyphase component and the imaginary part from the rst component.

(56) Thus, we have shown that all odd aliasing terms cancel within the MDFT lter bank whenever the analysis and synthesis prototype lters satisfy (34). Note that (56) describes the -channel DFT FB with analysis transfer function of a , synthesis lters , and lters an oversampling factor of 2 in the subbands. PR solutions for DFT FBs oversampled by 2 (see, e.g., [2][4]) are therefore also valid for MDFT lter banks if the prototypes polyphase components satisfy (34). analysis polyphase matrix of the oversampled The DFT lter bank is given by

(57)

for and expressing this in terms of the Substituting , we obtain the solution already given polyphase lters , (32) and (33) for by (30) and (31) for , and (34). B. Type-2 MDFT Filter Banks Another way to obtain MDFT lter banks is by modulating the analysis and synthesis lters as

where denotes the with entries containing the delay matrix

DFT-modulation matrix . The synthesis polyphase of (56) is

(58) and always taking the real part in the for undelayed branch of the subbands and the imaginary part in the delayed branch, as in Fig. 3. The result is called a type-2 MDFT lter bank. channel type-2 Theorem 4: The PR conditions for channel typeMDFT lter banks are the same as for 1 MDFT lter banks and are thus identical to those for channel CMFBs with DCT-IV or DCT-II modulation. The . delay introduced by the lter bank is Proof of Theorem 4: It can be shown that the input signal and the reconstructed signal for type-2 MDFT lter banks are related by (56) if we use the analysis and synthesis lters (58). Thus, the overall polyphase matrix satises

with the size

so that overall polyphase matrix can be written

An argument similar to that of Appendix B shows that . Choosing the overall delay of the lter bank to be samples, must satisfy

where
For k = 0, H2M 0k (z ) = H2M (z ) must be interpreted as H2M (z ) = 0 D H0 (z).
1

( 1)

HELLER et al: GENERAL FORMULATION OF MODULATED FILTER BANKS

995

and

shown by an argument similar to that of Appendix B. Hence, the PR condition for the overall polyphase matrix of type-2 MDFT lter banks is also given by (57), with solutions given by (30)(34). Relationship Between MDFT Filter Banks and DCT-II Cosine-Modulated Filter Banks: For real-valued input signals, the type-2 MDFT lter bank is equivalent to the DCT-I/II lter bank of Section III since taking the real and imaginary part of the subband signals is equivalent to taking the real or imaginary part of the lter impulse response. If we incorporate samples previously placed in the subbands into the delay of the lter impulse responses and take the real and imaginary part of the impulse responses, we obtain Re (59) Im

The two subbands equal to zero can be omitted without , changing the overall transfer function. For if we multiply the amplitude we can omit the subband of the subband by two, which can be done by multiplying . The the corresponding analysis and synthesis lters by . number of relevant, i.e., nonredundant subbands is now to the synthesis Furthermore, if we move the factor in lters, we nally end up with the same analysis and synthesis lters as in the DCT-I/II lter bank of Section III given by (35)(38). V. QUADRATIC-CONSTRAINED PROTOTYPE FILTER DESIGN We now turn to the problem of designing lowpass prototype lters for modulated lter banks with arbitrary length and delay. The conditions for such a lter to produce a PR system are encapsulated in (30)(33). Equations such as these are known as Bezout identities [27], [33]. Given two polynomial and , the elements from such an equation equation may be solved for the other two terms if and only if and are coprime (have no common factors). is Thus, the space of admissible solutions is quite large; , and is an arbitrary polynomial of degree or factors of simply restricted to not share any of the or . Given the large solution space and the requirement for high-performance lters, a time-domain design scheme was chosen for constructing biorthogonal modulated lter banks with arbitrary length and delay. The quadratic-constrained least-squares (QCLS) method [21] for lter bank design is suitable for creating prototype lters for the variable-delay biorthogonal modulated lter banks described in the previous sections with high stopband attenuation and satisfying the PR conditions to arbitrary accuracy. The QCLS algorithm works with the vector containing the coefcients of the time-domain impulse response of the unknown lowpass prototype lter. The PR conditions can be expressed as a quadratic form in , i.e., equations (30)(33) leads to a set of each of the constraints that can be expressed using matrices (63) such quadratic Thus, there are a total of constraints on , expressing the PR conditions (30)(33). Similarly, the objective function for the optimization [taken ] can also to be the stopband energy of the prototype lter be expressed as a quadratic form via the eigenlter formulation [32]

(60) Re

(61) Im (62) and and from (58). with Exploiting the symmetries of the sine and cosine function even yields, for

and for

odd

Here, is a real, symmetric and positive-denite matrix as and described in [32]. Since both the objective function the constraints (63) are quadratic forms in , they admit closed-form expressions for the gradient and Hessian, enabling efcient solution of the constrained minimization problem subject to (64)

996

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 47, NO. 4, APRIL 1999

(a) Fig. 4.

M = 8 length 96. (a) Impulse responses h(n). (b) Magnitude responses H (ej! ) .
j j

(b)

Consider the case where the lter length and the for DCT-IV modulation. For simplicity, delay . We must express the PR also assume that conditions (32) and (33) on the polyphase components as a set of quadratic forms in . The polyphase components can be written

where , and with entries otherwise. Equation (32) can be written in terms of as is the

, matrix
Fig. 5. Frequency responses of = 16 prototype lters with lengths 64 and 96 and reconstruction delay 63 samples.

Substituting otherwise we obtain the rst set of QCLS constraints

(65) Similarly, (33) yields a second set of QCLS constraints

(66) and are taken to be the zero matrix. where and for lter length In summary, given and and for overall system delay (for a DCT-IV and . Using lter bank), we can compute the matrices to denote the constant matrix dening the quadratic to denote the right-hand side form in (65) and (66) and , the PR constraints (32) and (33) become . Given the stopband edge of ,

we can compute the objective function . A nonlinear constrained optimization algorithm such as NLPQL [26] is then used to solve the problem (64). The solution is the time-domain impulse response of a lowpass prototype lter with minimal stopband energy, satisfying the PR conditions within a specied tolerance. Given , we can then construct and for a PR the analysis and synthesis lters modulated lter bank with system delay in the case of DCT-IV modulation or system delay with DCT-II or MDFT modulation. Design Examples: We present a number of design examples, illustrating the possibilities and tradeoffs of the arbitrarydelay formulation. , length ( ), different values of . a) b) , delay 63, different lengths. c) and xed, different values of . d) same prototype ( , , ), different modulations. and , yielding lters with length 96. We a) Let choose , which provides a large overlap between and . Prototype lters for and were designed using the QCLS formulation. The impulse responses for these six cases are plotted in Fig. 4(a). Note that is minimum phase for and linear phase for , as expected. The frequency responses

HELLER et al: GENERAL FORMULATION OF MODULATED FILTER BANKS

997

Fig. 6.

M = 4, N = 32 prototype lters with delay D = 3, 5, and 7. (a) Impulse responses. (b) Frequency responses.

(a)

(b)

(a)

(b)

Fig. 7. Magnitude frequency responses of lter banks obtained by DCT-IV and DCT-II modulations of the same

M = 8, N = 96, D = 95 prototype lter.

(c)

are plotted in Fig. 4(b). We see that the system with has the highest stopband attenuation, whereas the system with has the poorest stopband attenuation. b) A second set of design examples demonstrate the effect constant while varying the of keeping the delay parameter lter length. Fig. 5 displays the magnitude responses of two , reconstruction delay of 63 samples ( lters with and ), and lengths 64 and 96. The longer prototype has superior stopband attenuation (43 dB rst sidelobe height versus 34 dB for the shorter prototype). c) A third set of design examples shows the effect of and xed. Here, and changing while keeping , whereas and takes the values 3, 5, and 7. The impulse responses of the three lters are shown in Fig. 6(a) and the magnitude frequency responses in Fig. 6(b). As before, decreasing delay leads to lower stopband attenuation (lowerhas two zero performance lters). The prototype for prototype has coefcients near its tail, whereas the has none. These one zero coefcient, and the zeros of the impulse response, forced by the PR conditions (30)(33), affect the lter performance, as described in [19]. d) The nal example explores the use of a given prototype for different types of modulation. First, consider a prototype , , and . Since and lter with have different parity, we can modulate the prototype using

either a DCT-IV or a DCT-II. Fig. 7(a) shows the magnitude frequency response of the lter bank generated using DCT-IV modulation. Notice that each subband is covered once, and all the subband lters have the same amplitude. The lter bank has eight bands and reconstruction delay 95. Fig. 7(b) and (c) show the magnitude frequency responses of the DCT-II lter bank generated from the same prototype; plot (b) shows the lters generated using (51), and plot (c) shows the lters generated by (52). Notice that the subbands centered at DC and are covered once, whereas the remaining subbands are covered twice by lters with amplitudes that have been . This DCT-II lter bank has 16 bands and scaled down by reconstruction delay 103. , Now, consider the example of a prototype lter with , and . Here, and have the same parity so that we will use a DCT-III or DCT-I to modulate the prototype. Fig. 8(a) shows the magnitude frequency response of the lter bank generated from the prototype using DCT-III modulation (1). Again, each subband is covered once, and all the subband lters have the same amplitude. The lter bank has channels and reconstruction delay . Fig. 8(b) and (c) show the magnitude frequency responses of the DCT-I lter bank generated from the same prototypeplot (b) shows the lters generated using (35), and plot (c) shows the lters generated using (36). This 16-band lter bank

998

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 47, NO. 4, APRIL 1999

(a)

(b)

Fig. 8. Magnitude frequency responses of lter banks obtained by DCT-III and DCT-I modulations of the same

M = 8, N = 49, D = 48 prototype lter.

(c)

Fig. 9. Linear-phase impulse responses of the rst three lters in the DCT-II modulated lter bank of example = 95 prototype lter.

d, based on the M

= 8,

= 96,

is similar to the DCT-II modulation, except that when the full -band lter bank is broken into two pieces (corresponding to the two different formulas for modulation), they contain different numbers of subbands. Again, the reconstruction delay , which equals 56 in this example. is All the lter banks in these two gures are paraunitary. Furof the DCT-I and thermore, each of the constituent lters DCT-II lter banks have linear-phase symmetry, generalizing the examples of [11]. Fig. 9 shows the time-domain impulse responses of the rst three lters in the DCT-II modulated banknotice that the rst (lowpass) lter is simply a scaled copy of the prototype. VI. APPLICATION
TO

AUDIO COMPRESSION

Modulated lter banks are widely used as the transform element for subband audio coders [16]. These coders perform lossy compression, usually using a psychoacoustic model for the coding algorithm that incorporates the frequency-masking effects of human hearing. However, these algorithms are adapted to stationary signals and may introduce disturbing effects such as pre-echoes when dealing with highly nonstationary signals such as percussive sounds. Therefore, the lter lengths cannot be chosen to be arbitrarily long, thus limiting the number of subbands. Another objective in the design of an audio coder is the overall delay of the system, which should be as short as possible for real-time applications. The audio coder standardized in ISO/IEC 11 172-3 by the moving pictures expert group (MPEG) [16] uses a 32-channel

cosine-modulated lter bank with a linear-phase prototype of length . However, the linear-phase property of the prototype lter is not important for audio coding performance, and the analysis and synthesis lters are not linear phase. On the contrary, it has been shown [35] that the masking of the human ear in the time domain is greater after a percussive sound than before it. Both pre-echoes and the overall system delay can be reduced when using a minimalphase prototype lter. However, we must also ensure that the low-delay prototype lter has a high stopband attenuation in order to suppress disturbing aliasing, which is an issue not taken into consideration by the psychoacoustic model. Fig. 10 shows the frequency responses of the linear-phase prototype lter tabulated in [16] as well as a prototype lter designed using the QCLS algorithm of length yielding a delay of 447 samples instead of 511. Note that the original MPEG prototype is an approximate spectral factor th-band lter, providing near-perfect reconstruction of an (NPR), whereas the low-delay prototype lter satises PR. The stopband attenuation of the latter can be improved by sacricing the PR property. Fig. 11(a) shows an extract of a bells sound over time. Fig. 11(b) and (c) depict the reconstructed signal coded at 128 kb/s per channel using the MPEG-1 layer 1 audio coder and prototype lters from Fig. 10. The pre-echo is reduced when using the prototype lter with shorter delay. The results can be expanded to other layers of the MPEG coder and other subband audio coders as well.

HELLER et al: GENERAL FORMULATION OF MODULATED FILTER BANKS

999

Fig. 10. Frequency response of length 512 prototype lters with delays 447 and 511.

system delay , we have developed necessary and sufcient conditions on the polyphase components of the prototype lter for the modulated lter bank to satisfy perfect reconstruction. Furthermore, the relationships among modulated lter banks with different modulation schemes (DCT-III/IV, DCT-I/II, and MDFT) have been explored. A prototype lter satisfying perfect reconstruction for one of these modulation schemes will yield PR with any of the other types of modulation, and in fact, MDFT modulation has been shown to be equivalent to DCT-II modulation. A quadratic-constrained least-squares technique has been presented for the design of prototype lters for PR modulated lter banks, with full generality in setting the parameters , , and . Implementation of the technique shows that the best subchannelization (greatest stopband attenuation) is obtained for . Systems with have lower stopband ), leading to a attenuation (as do systems with tradeoff between subchannelization and system delay for a given lter length. The application of biorthogonal modulated lter banks to audio compression has been investigated. A comparison of pre-echo noise in MPEG-type subband audio coders with and was presented, and the system with shorter delay showed a decrease in pre-echo noise power. APPENDIX A

(a)

We prove here several identities involving the matrix prodand . ucts Proof 1: In the case of DCT-III/IV modulation, we will has the form given in (3). Consider the prove that of the matrix element

(b)

(c) Fig. 11. Original audio signal and reconstructed signals coded at 128 kb/s using the MPEG-1 prototype lter and the low-delay prototype lter.

VII. CONCLUSIONS In this paper, the theory, structure, and design of general biorthogonal modulated lter banks (cosine-modulated and modied DFT) have been developed, using the polyphase representation. Working with the design parameters of the number of channels , the prototype lter length , and the

(67)

1000

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 47, NO. 4, APRIL 1999

First, examine the upper sum in (67) and replace by

Thus, the

th matrix entry is

Substituting

, we obtain

and the matrix as a whole has the form of (3). Proof 2: In the case of DCT-I/II modulation, we prove and have the form given in (42) and (43). that th element of First, consider the

otherwise. The value must be chosen so that , yielding are in the range

with solutions in (67) becomes

and

. Thus, the upper sum

For the second sum in (67), we proceed similarly. The can be expressed as phase term , yielding (68)

The lower sum in (68) reduces to matrix to contributing a in (68) reduces to

, . The upper sum

where becomes

, and

. When

, this

HELLER et al: GENERAL FORMULATION OF MODULATED FILTER BANKS

1001

whereas when

, the upper sum equals

Thus, the overall matrix is similar. proof for

satises (42) and (43). The

APPENDIX B We show here that the second part of (55) is equal to zero. This is equivalent to the cancellation of all odd alias spectra.

Substituting

yields

Note that we do not have to change the summation index since we sum over a whole period when substituting for and . Expressing the prototype lters by of their polyphase components and taking (34) into consideration yields that the whole expression is equal to zero. REFERENCES
[1] M. G. Bellanger and J. L. Daguet, TDM-FDM transmultiplexer: Digital polyphase and FFT, IEEE Trans. Commun., vol. COMM-22, pp. 11991204, Sept. 1974. [2] H. B olcskei, F. Hlawatsch, and H. G. Feichtinger, Oversampled FIR and IIR DFT lter banks and Weyl-Heisenberg-frames, in Proc. IEEE ICASSP, Atlanta, GA, 1996. [3] R. E. Crochiere and L. R. Rabiner, Multirate Digital Signal Processing. Englewwod Cliffs, NJ: Prentice-Hall, 1983. [4] Z. Cvetkovi c and M. Vetterli, Oversampled modulated lter banks and tight Gabor frames in `2 (Z), in Proc. IEEE ICASSP, Detroit, MI, 1995.

[5] N. J. Fliege, Computational efciency of MDFT lter banks, in Proc. 27th Asilomar Conf. Signals, Syst., Comput., Asilomar, CA, 1993. , Modied DFT polyphase SBC lter banks with almost perfect [6] reconstruction, in Proc. IEEE ICASSP, Adelaide, Australia, 1994, pp. III149III152. [7] R. A. Gopinath, Modulated lter banks and waveletsA general unied theory, in Proc. IEEE ICASSP, Atlanta, GA, 1996. [8] R. Gopinath and C. Burrus, Theory of modulated lter banks and modulated wavelet tight frames, in Proc. IEEE ICASSP, Minneapolis, MN, 1993, pp. III169III172. [9] T. Karp and N. Fliege, MDFT Filter banks with perfect reconstruction, in Proc. IEEE ISCAS, Seattle, WA, 1995. [10] R. D. Koilpillai and P. P. Vaidyanathan, Cosine modulated FIR lter banks satisfying perfect reconstruction, IEEE Trans. Signal Processing, vol. 40, pp. 770783, Apr. 1992. [11] Y.-P. Lin and P. P. Vaidyanathan, Linear-phase cosine modulated maximally decimated lter banks with perfect reconstruction, IEEE Trans. Signal Processing, vol. 43, pp. 25252539, Nov. 1995. [12] H. Malvar, Extended lapped transforms: Fast algorithms and applications, IEEE Trans. Signal Processing, Nov. 1992. [13] H. Malvar, Signal Processing with Lapped Transforms. Norwood, MA: Artech House, 1992. [14] J. Masson and Z. Picel, Flexible design of computationally efcient nearly perfect QMF lter banks, in Proc. IEEE ICASSP, Tampa, FL, 1985. [15] J. Mau, Perfect reconstruction modulated lter banks: Fast algorithms and attractive new properties, in Proc. IEEE ICASSP, Minneapolis, MN, 1993, pp. III-225III-228. [16] Coding of moving pictures and associated audio for digital storage media up to about 1.5 MBit/s, Int. norm MPEG1, ISO 11172-3, 1992. [17] K. Nayebi, T. P. Barnwell, III, and M. J. T. Smith, Time-domain lter bank analysis: A new design theory, IEEE Trans. Signal Processing, vol. 40, pp. 14121429, 1992. [18] , Low delay FIR lter banks: Design and evaluation, IEEE Trans. Signal Processing, vol. 42, pp. 2431, Jan. 1994. [19] T. Q. Nguyen and R. D. Koilpillai, The theory and design of arbitrarylength cosine-modulated lter banks and wavelets, satisfying perfect reconstruction, IEEE Trans. Signal Processing, vol. 44, pp. 473483, Mar. 1996. [20] T. Q. Nguyen, A class of generalized cosine modulated lter bank, in Proc. IEEE ISCAS, 1992, pp. 943946. [21] , Digital lter bank design quadratic constrained formulation, IEEE Trans. Signal Processing, vol. 43, pp. 21032108, Sept. 1995. [22] J. Nussbaumer and M. Vetterli. Computationally efcient QMF lter banks, in Proc. IEEE ICASSP, San Diego, CA, 1984. [23] J. Princen and A. Bradley, Analysis/synthesis lter bank design based on time domain alias cancellation, IEEE Trans. Acoust., Speech, Signal Processing, vol. ASSP-34, pp. 476492, 1986. [24] T. A. Ramstad and J. P. Tanem, Cosine modulated analysis synthesis lter bank with critical sampling and perfect reconstruction, in Proc. IEEE ICASSP, Toronto, Ont., Canada, 1991, pp. 17891792. [25] J. H. Rothweiler, Polyphase quadrature ltersA new subband coding technique, in Proc. IEEE ICASSP, 1983, pp. 12801283. [26] K. Schittkowski, NLPQL: A FORTRAN subroutine solving constrained nonlinear programming problems, Ann. Oper. Res., vol. 5, pp. 485500, 1986. [27] M. R. Schroeder, Number Theory in Science and Communication. New York: Springer-Verlag, 1986. [28] G. Schuller, A low delay lter bank for audio coding with reduced pre-echoes, in Proc. 99th AES Conv., New York, NY, Oct. 1995. [29] , A new factorization and structure for cosine modulated lter banks with variable system delay, in Proc. 30th Asilomar Conf. Signals, Syst., Comput., Pacic Grove, CA, Nov. 1996. [30] G. Schuller and M. J. T. Smith, New framework for modulated perfect reconstruction lter banks, IEEE Trans. Signal Processing, vol. 44, pp. 19411954, 1996. [31] P. P. Vaidyanathan, Multirate Systems and Filter Banks. Englewood Cliffs, NJ: Prentice-Hall, 1992. [32] P. P. Vaidyanathan and T. Q. Nguyen, Eigenlters: A new approach to least squares FIR lter design and applications including Nyquist lters, IEEE Trans. Circuits Syst., vol. CAS-34, pp. 1123, Jan. 1987. [33] M. Vetterli and C. Herley, Wavelets and lter banks, IEEE Trans. Signal Processing, vol. 40, pp. 22072232, 1992. [34] H. Xu, W.-S. Lu, and A. Antoniou, Efcient iterative design method for cosine-modulated QMF banks, IEEE Trans. Signal Processing, vol. 44, pp. 16571668, July 1996. [35] E. Zwicker and H. Fastl, Psychoacoustics. New York: Springer-Verlag, 1990.

1002

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 47, NO. 4, APRIL 1999

Peter Niels Heller (M91) received the B.S. degree in mathematics with highest honors from the University of North Carolina, Chapel Hill, in 1981 and the M.A. and Ph.D. degrees in mathematics from Princeton University, Princeton, NJ, in 1983 and 1986, respectively. He is presently a Scientist at Aware, Inc., Bedford, MA. From 1985 to 1986, he was a C. L. E. Moore Instructor of Mathematics at the Massachusetts Institute of Technology (MIT), Cambridge, and he returned to MIT as a Visiting Lecturer in Mathematics in 1989. During 1990, he developed geometric modeling software at Parametric Technology Corp., Waltham, MA. in December 1990, he joined Aware. His research interests include wavelets and multirate systems and their application to data compression and telecommunications. He is currently directing the development of a wavelet-based broadband wireline modem chip. In 1994, he copresented the Nordic postgraduate course on wavelets, lter banks, and applications at the Helsinki University of Technology, Helsinki, Finland. Dr. Heller held National Science Foundation and Sloan Foundation fellowships while a graduate student at Princeton.

Truong Q. Nguyen (S85M90SM95) received the B.S., M.S., and Ph.D. degrees in electrical engineering from the California Institute of Technology, Pasadena, in 1985, 1986, and 1989, respectively. He was with the Lincoln Laboratory, Massachusetts Institute of Technology (MIT), Cambridge, from June 1989 to July 1994, as a Member of the Technical Staff. From 1993 to 1994, he was a Visiting Lecturer at MIT and an Adjunct Professor at Northeastern University, Boston, MA. From August 1994 to July 1996, he was an Assistant Professor at the University of Wisconsin, Madison. He is now with the Boston University. His research interests are in digital and image signal processing, multirate systems, wavelets and applications, and biomedical signal processing. Prof. Nguyen was a recipient of a fellowship from Aerojet Dynamics for advanced studies. He received the IEEE TRANSACTIONS ON SIGNAL PROCESSING Award (Image and Multidimensional Processing area) for the paper he cowrote, with Prof. P. P. Vaidyanathan, on linear-phase perfect-reconstruction lter banks in 1992. He received the NSF Career Award in 1995 and is the coauthor (with Prof. G. Strang) of a textbook on Wavelets and Filter Banks (Wellesley, MA: Wellesley-Cambridge). He served as Associate Editor for the IEEE TRANSACTIONS ON SIGNAL PROCESSING and for the IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS II. He is a member of Tau Beta Pi and Eta Kappa Nu.

Tanja Karp was born in Germany in 1969. She received the Dipl.-Ing. degree in electrical engineering and the Dr.-Ing. degree from Hamburg University of Technology, Hamburg, Germany, in 1993 and 1997, respectively. Since 1997, she has been with Mannheim University, Mannheim, Germany, as a Research and Teaching Assistant. Her research interests include multirate signal processing, lter banks, source and channel coding, and signal processing for communications.