1

Part V.F: Cosine-Modulated Filter Banks This part shows how to eciently implement a multi-

Organization of This Pile of Lecture Notes This pile can be roughly divided into the the following subparts: 1. Introduction 2. Cosine-modulated FIR lter banks T. 3. Nearly perfect-reconstruction lter banks 4. Perfect-reconstruction lter banks 5. Filter banks with constraints on the amplitude distortion 6. Filter banks with constraints on both the amplitude and aliasing distortions 7. Practical examples 8. Implementation aspects

channel (M > 2) analysis synthesis bank using a single prototype lter and a proper cosine modulation scheme. It is mostly based on the following article: ter banks, in of Transforms First and Filter Saram aki, A generalized class of cosine-modulated lBanks. on Proceedings International Workshop

Transforms and Filter Banks (Tampere, Finland), edited by K. Egiazarian, T. Saram aki, and J. Astola, pp. 336-365, 1998.

1. Introduction During the past ten years, the subband coding by M -channel critically sampled FIR lter banks have received a widespread attention [1], [2], [3] (see also references in these textbooks). Such a system is shown in Fig. 1.
x(n) H0(z)
M M

In the synthesis bank, the coded symbols are con-

verted to their appropriate digital quantities, interpolated by a factor of M followed by ltering by the corresponding lters Fk (z ). Finally, the outputs are added to produce the quantized version of the input. These lter banks are used in a number of communication applications such as subband coders for speech signals, frequency-domain speech scramblers, image cod-

F0(z) y(n)

H1(z)

F1(z)

ing, and adaptive signal processing [1]. The most eective technique for constructing both the analysis bank consisting of lters Hk (z ) for k =

HM-1(z)

FM-1(z)

Figure 1: M -channel maximally decimated lter bank. In the analysis bank consisting of M parallel band-

ters Fk (z ) for k = 0, 1, , M 1 is to use a cosine single linear-phase FIR prototype lter.

0, 1, , M 1 and the synthesis bank consisting of lmodulation [1], [2], [3] to generated both banks from a Compared to the case where all the sublters are de-

HM 1(z ) are lowpass and highpass lters, respectively), subband signals.

pass lters Hk (z ) for k = 0, 1, , M 1 (H0(z ) and the input signal is ltered by these lters into separate

signed and implemented separately, the implementation of both the analysis and synthesis banks is signicantly more ecient since it requires only one prototype lter and a unit performing the desired modulation operation [1], [2], [3]. Also, the actual lter bank design becomes much faster and more straightforward since the only parame-

tized, and edcoded for transmission to the synthesis 0, 1, , M 1.

These signals are individually decimated by M , quan-

bank consisting also of M parallel lters Fk (z ) for k =

ters to be optimized are the coecients of a single prototype lter. Another alternative is to use the modied DFT [3]. However, this is a very new idea lacking a good tutorial article. Therefore, this alternative is not considered in this course. The purpose of this pile of lecture notes is to introduce ecient design techniques for generating properfect-reconstruction cases using a cosine modulation. Several examples are included illustrating that by allowing small amplitude and aliasing errors, the lter bank performance can be signicantly improved. Alternatively, the lter orders and the overall delay caused by the lter bank to the signal can be considerapplications. In many applications such small errors are tolerable and the distortion caused by these errors to the signal is smaller than that caused by coding. ably reduced. This is very important in communication totype lters for both perfect-reconstruction and nearly

2. Cosine-Modulated FIR Filter Banks This section shows how M -channel critically sampled

FIR lter banks can be generated using proper cosinemodulation techniques and a proper prototype lter.

2.1 Input-Output Relation for an M -Cannel Filter Bank For the system of Fig. 1, the input-output relation in the z -domain is expressible as Y (z ) = T0(z )X (z ) + where 1 T0(z ) = M
M 1 k =0 M 1 l=1 M 1 k =0

Tl (z )X (zej 2l/M ),

(1a)

Fk (z )Hk (z )

(1b)

and for l = 1, 2, , M 1 Tl (z ) = 1 M Fk (z )Hk (zej 2l/M ). (1c)

Here, T0(z ) is called the distortion transfer function and determines the distortion caused by the overall system for the unaliased component X (z ) of the input signal. 1, 2, , M 1 are called the alias transfer functions and of the input signal are attenuated. The remaining transfer functions Tl (z ) for l =

determine how well the aliased components X (zej 2l/M )

For the perfect reconstruction, it is reguired that T0(z ) = z N with N being an integer and Tl (z ) = 0 for l = 1, 2, , M 1. If these conditions are satised, then the output signal is a delayed version of the input signal, that is, y (n) = x(n N ). It should be noted that the perfect reconstruction is exactly achieved only in the case of lossless coding. For lossy coding, it is worth studing whether it is

approximately the value of unity at the zero frequency. The desired input-output relation is then achieved in the nal implementation by multiplying the Fk (z )s by M. This is done in order to preserve the signal energy after using the interpolation lters Fk (z ). In this case, the lters in the analysis and synthesis scaled, as is desired in many practical applications.

banks of the overall system become approximately peak

benecial to allow small amplitude and aliasing errors causing smaller distortions to the signal than the coding or errors that are not very noticable in practical applications. For nearly perfect reconstruction cases, the abovetolerances. quence of the decimation and interpolation processes. For simplicity, this term is forgotten in the sequel. In this case, the passband maxima of the amplitude responses of the Hk (z )s and Fk (z )s will become approximately equal to unity. Also the prototype lter to be considered later on can be designed such that its amplitude response has The term 1/M in Eqs. (1b) and (1c) is a consementioned conditions should be satised within given

10

2.2 Generation of Filter Banks from a Prototype Filter Using Cosine-Modulation Techniques For the cosine-modulated lter banks, both the Hk (z )s and Fk (z )s are constructed with the aid of a linear-phase FIR prototype lter of the form
N

0, 1, , M 1 and

From the above equations, it follows that for k = fk (n) = hk (N n) Fk (z ) = z N Hk (z 1). (5a) (5b)

Hp(z ) =
n=0

hp(n)z n,

(2a)

Another alternative is to construct the impulse responses hk (n) and fk (n) as follows [1]: fk (n) = 2hp(n) cos[ 1 M +1 (k + )(n + )] 2M 2 2

where the impulse response satises the following symmetry property: hp(N n) = hp(n). (2b) and

(6a)

One alternative is to construct them as follows [2]:

Hk (z ) = k k Hp(zej (2k+1)/(2M ) )+k k Hp(zej (2k+1)/(2M ) )

hk (n) = fk (N n) = 2hp(n) cos[

1 M +1 (k + )(N n+ )]. 2M 2 2 (6b)

(3a) and
Fk (z ) = k k Hp(zej (2k+1)/(2M ) ) + k k Hp(zej (2k+1)/(2M ) ,

In [1], instead of the constant of value 2, the constant of value

2/M has been used. The reason for

this is that the prototype lter is implemented using special butteries. The most important property of the above modulation schemes lies in the following facts. By properly designing the prototype lter transfer (3b)

where k = e

j (1)k /4

and k = e

jN (2k +1)/(4M )

The corresponding impulse responses for k = 0, 1, , N hk (n) = 2hp(n) cos[(2k + 1) (n ) + (1)k ] (4a) 2M 2 4 and N fk (n) = 2hp(n) cos[(2k + 1) 2M (n ) (1)k ]. (4b) 2 4 M 1 are then given by

function Hp(z ), the aliased components generated in the analysis bank due to the decimation can be totally or partially compensated in the synthesis bank. Secondly, T0(z ) can be made exactly or approximately equal to the pure delay z N .

11

12

Hence, these modulation techniques enable us to de-

2.3 Conditions for the Prototype Filter to Give a Nearly Perfect-Reconstruction Property The above modulation schemes guarantee that if the impulse response of
2N

sign the prototype lter in such a way that the resulting overall bank has the perfect-reconstruction or a nearly perfect-reconstruction property.

Hp(z ) = [Hp(z )]2 =

n=0

hp(n)z n,

hp(2N n) = hp(n) (7a) (7b) (7c)

satises ( x stands for the integer part of x.) and then [4] hp(N 2rM ) 0 T0(z ) = hp(N ) 1/(2M ) for r = 1, 2, , N/(2M ) ,
M 1 l=0

Fk (z )Hk (z ) z N .

(8)

comes very small.

In this case, the amplitude error |T0(ej ) ejN | be-

If the conditions of Eqs. (7b) and (7c) are exactly satised, then the the amplitude error becomes zero. It should be noted that since T0(z ) is an FIR lter of order 2N and its impulse-response coecients, 0, 1, , 2N , there exists no phase distortion. 2M th-band linear-phase FIR lter [5], [6]. denoted by t0(n), satisfy t0(2N n) = t0(n) for n =

Equation (7) implies that [Hp(z )]2 is approximately a

 13

14

Based on the properties of these lters, the stopband edge of the prototype lter Hp(z ) must be larger than /(2M ) and is specied by s = (1 + )/(2M ), where > 0. Furthermore, the amplitude response of Hp(z ) achieves approximately the values of unity and 1/ 2 at = 0 and = /(2M ), respectively.
Amplitude in dB Amplitude in dB

Prototype filter 0 20 40 60 80 0 0.1 0.2 0.3 0.4 0.5 0.6 Angular frequency / Filter bank 20 0 20 40 60 80 0 0.1 0.2 0.3 0.4 0.5 0.6 Angular frequency / 0.7 0.8 0.9 1 H0(z),F0(z) H1(z),F1(z) H2(z),F2(z) H3(z),F3(z) 0.7 0.8 0.9 1

(9)

As an example, Fig. 2 shows the prototype lter fre-

quency response for M = 4, N = 63, and = 1 as well as the responses for the lters Hk (z ) and Fk (z ) for k = 0, 1, 2, 4. 1, 2, , M 2 are bandpass lters with the center fre It is seen that the lters Hk (z ) and Fk (z ) for k =

quency at = k = (2k + 1)/(2M ) around which the amplitude response is very at having approximately the value of unity. The amplitude response of these lters achieves ap proximately the value of 1/ 2 at = k /(2M ) and the stopband edges are at = k s. H0(z ) and F0(z ) [HM 1(z ) and FM 1(z )] are lowpass

Figure 2: Example amplitude responses for the prototype lter and for the resulting lters in the analysis and synthesis banks for M = 4, N = 63, and = 1. The stopband edge is at = (2 + )/(2M ) [ = (2 + )/(2M )]. The impulse responses for the prototype lter as well as those for the lters in the banks are shown in Figs. 3 and 4, respectively. In this case, the impulse responses for the ltes in the bank have been generated according to Eq. (4).

(highpass) lters with the amplitude response being at around = 0 ( = ) and achieving approximately the value 1/ 2 at = /M ( = /M ).

15

16

Prototype filter 0.15

Impulse response 0.3 0.2 0.1 0 0.1 0

H0(z) 0.2

H1(z)

Impulse response

0.1 0.05 0 0.05 0

Impulse response 20 40 n in samples H2(z) 60

0.2 0 20 40 n in samples H3(z) 60

10

20

30 n in samples

40

50

60

Impulse response

Impulse response 20 40 n in samples F0(z) 60

Figure 3: Impulse response for the prototype lter in the case of Fig. 2. If Hp(z ) satises Eq. (7), then both of the abovementioned modulation schemes have the very important property that the maximum amplitude values of the aliased transfer functions Tl (z ) for l = 1, 2, , M 1 are guaranteed to be approximately equal to the max-

0.2

0.2

0.2 0

0.2 0 20 40 n in samples F1(z) 0.2 60

0.3 Impulse response 0.2 0.1 0 0.1 0 Impulse response 20 40 n in samples F2(z) 60

imum stopband amplitude values of the lters in the bank [2], as will be seen in connection with the examples of Section 5. If smaller aliasing error levels are desired to be achieved, then additional constraints must be imposed on the prototype lter. This will be considered in the next section. In the case of the perfect reconstruction, the additional constraints are so strict that they actually retotype lter, as will be seen in Section 4. duce the number of adjustable parameters of the pro-

0.2 0 20 40 n in samples F3(z) 60

Impulse response

Impulse response 20 40 n in samples 60

0.2

0.2

0.2 0

0.2 0 20 40 n in samples 60

Figure 4: Impulse responses for the lters in the bank in the case of Fig. 2.

 17

18

3 General Optimization Problems for the Prototype Filter This section states two optimization problems for deoverall lter bank possesses a nearly perfect-reconstruction property. Ecient algorithms are then briey described for solving these problems. 3.1 Statement of the Problems We consider the following two general optimization problems: Problem I. Given , M , and N , nd the coecients of Hp(z ) to minimize

Problem II. Given , M , and N , nd the coecients of Hp(z ) to minimize E = max |Hp(ej )|
[s , ]

(11)

signing the prototype lter in such a way that the

subject to the conditions of Eqs. (10c) and (10d).

E2 =
s

|Hp(ej )|2d,

(10a)

where s = (1 + )/(2M ) subject to 1 1 |T0(ej )| 1 + 1 , and |Tl (ej )| 2, [0, ] for l = 1, 2, , M 1. [0, ] (10c) (10b)

(10d)

19

20

3.2 Algorithm for Solving Problem I This contribution concentrates on the case where N , N + 1 is even). is restricted to be odd, as will be seen in the following section. lter is expressible as Hp(, e ) =
(N +1)/2 j

where (, ) = 4 2 2s 2 sin[(2 1)s] , = 2 1 = 2 sin[( + 1)s] 2 sin[( )s] , = . + 1 (13b) |Tl (, ej )| for l = 0, 1, , M 1, in turn, can be written as shown in the article mentioned in the rst page of this part of lecture notes. into the discrete points j 1 , 2 , , J0 . at a very accurate solution. To solve Problem I, we discretize the region [0, /M ] [0, /M ] for j = s cos( 1/2) cos( 1/2)d

the order of the prototype lter, is odd (the length This is because for the perfect-reconstruction case N

For N odd, the frequency response of the prototype

2ej (N 1)/2
n=1

hp[(N + 1)/2 n] cos[(n 1/2) ], (12a)

where = [hp(0), hp(1), , h[(N 1)/2] type lter. After some manipulations, Eq. (10a) is expressible as E2() E2 =
(N +1)/2 (N +1)/2 =1 =1

In many cases, J0 = N is a good selection to arrive

(12b)

The resulting discrete problem is to nd to minimize = E2(), where E2() is given by Eq. (13) subject to gj () 0, where j = 1, 2, , J, (14b) (14a)

denotes the adjustable parameter vector of the proto-

hp[(N + 1)/2 ]hp[(N + 1)/2 ](, ), (13a)

J = (M + 2)/2 J0, gj () = ||T0(, ejj )| 1| 1 ,

(14c)

j = 1, 2, , J0, (14d)

 21

22

and glJ0+j () = |Tl (, ejj )| 2 , l = 1, 2, , M/2 , j = 1, 2 , , J0 . (14f )

and hl () = 0, where J = M/2 J0, L = J0 , g(l1)J0 +j () = |Tl (, e

jj

l = 1, 2, , L,

(15b) (15c)

In the above, the region [0, /M ], instead of [0, ], periodicity equal to 2/M . has been used since the |Tl (, ej )|s are periodic with

)| 2,

Furthermore, only the rst (M + 2)/2 |Tl (, ej )|s

j j

j = 1, 2 , , J0 ,

l = 1, 2, , M/2 , (15d)

1, 2, , (M 1)/2 .

have been used since |Tl (, e )| = |TM l (, e )| for l =

and hl () = ||T0(, ejj )| 1|, the Dutta-Vidyasagar algorithm. l = 1, 2, , L. (15e)

The above problem can be solved conveniently by us-

ing the second algorithm of Dutta and Vidyasagar [7]. The details can be found in the article mentioned in the rst page of this part of lecture notes. Since the optimization problem is nonlinear in nature, a good initial starting point for the vector is needed. This can be obtained by nding a corresponding vecin the next section. If it is desired that |T0(, ej )| 1 [4], then the resulting discrete problem is to nd to minimize given by Eq. (14a) subject to gj () 0, j = 1, 2, , J (15a) as tor in the perfect-reconstruction case to be considered

This problem can also be solved conveniently using

23

24

3.3 Algorithm for Solving Problem II To solve Problem II, we discretize the region [s, ] In many cases, I = 20N is a good selection. The

4. Perfect-Reconstruction Filter Banks This section introduces algorithms for designing eec-

into the discrete points i [s, ] for i = 1, 2, , I .

tively the prototype lter for perfect-reconstruction lter banks either in the least-mean-square sense or in the minimax sense. Both of these algorithms are characterized by the desired property that they enable us to design banks of lters with much higher lengths than the other existing synthesis techniques. 4.1 Conditions for the Perfect Reconstruction Under certain conditions, the output of the lter bank is a delayed version of the output. 1, 2, , M 1 [8][14]. ditions are T0(z ) z N These conand Tl (z ) 0 for l =

resulting discrete minimax problem is to nd to minimize = E() = max fi() subject to gj () 0, where fi() = |Hp(, eji )|, and (14e). Again, the Dutta-Vidyasagar algorithm can be used to solve the above problem. Also, the optimization of the prototype lter for the lem I. i = 1, 2, , I (16c) and J and the gj ()s are given by Eqs. (14c), (14d),
iI

(16a) (16b)

j = 1, 2, , J,

In this case, N + 1 = 2KM with K being an integer and only K M/2 parameters are adjustable. In the sequel, the conditions for the prototype lter

case where |T0(, ej )| 1 can be solved like for Prob-

are given in the form suggested in [13]. This form can be used in a straightforward manner in optimizing the prototype lter in the minimax or in the least meansquare sense. The suggested form for the perfect reconstruction has been derived by slightly modifying the conditions given in [10].

Similarly, a good initial vector is obtained by ndcase.

ing a corresponding vector for the perfect-reconstruction

 25

26

In the following subsection, the connections of this form to those described in [10], [11] will be given. Eq. (2) can be expressed as The prototype lter transfer function as given by Hp(z ) = Gp(z )/(M 2), where
KM

maining M/2 sets of 2K g (n)s. These sets are given by

Second, there are K constraints for each of the re-

(17a)

r ={g (r ), g (M + 1 r ), g (M + r ), g (2M + 1 r ), g (KM + 1 r )} for r = 1, 2, , M/2 .

g (2M + r ), g (3M + 1 r ), , g ((K 1)M + r ), (20)

Gp(z ) = z

(KM 1/2) n=1

g (n)[z

(n1/2)

+z

(n1/2)

].

(17b)

For the r th set r , the conditions can be expressed in a convenient manner in terms of crk = cos(rk ) and srk = sin(rk ), justable parameters. in z 1 are rst determined: Pr(K )(z ) = To do this, the following (K 1)th-order polynomials
K 1 n=0

Here, the hp(n)s are related to the g (n)s through hp(n) = hp(2KM 1 n) =g (KM n)/(M 2)

(21a)

for n = 0, 1, , KM 1. (18)

(21b)

The basic reason for expressing the prototype lter

where the rk s for k = 1, 2, . . . , K are the basic ad-

transfer function in the above form lies in the fact that the conditions for the perfect reconstruction can be stated conveniently with the aid of the g (n)s for n = 1, 2, , KM . First, these conditions imply that for M odd, g ((M + 1)/2) = and g ((M + 1)/2 + kM ) = 0 for k = 1, 2, , K 1. For M even, this condition is absent. (19b) 1/2 (19a)

pr (n)z n

(22a)

and
K) Q( r (z ) =

K 1 n=0

qr (n)z n

(22b)

recursively as
(k 1) Pr(k)(z ) = srk Pr(k1)(z ) + crk z 1Qr (z )

(23a)

27

28

and
k) (k 1) (k 1) Q( (z ) srk z 1Qr (z ) r (z ) = crk Pr

and (23b) g ((2k + 1)K + 1 r ) =pr ( K 1 k) 2

with the initializations Pr(1)(z ) = sr1 and Q(1) r (z ) lowing form: = cr1 . (23d) For K even, the desired conditions take then the folK K + k) for k = 0, 1, , 1, (24a) 2 2 K K g ((2k + 1)M + r ) = qr ( + k) for k = 0, 1, , 1, 2 2 (24b) K K g (2kM + 1 r ) = pr ( k) for k = 1, 2, , , (24c) 2 2 and K K g ((2k 1)M + 1 r ) = qr ( k) for k = 1, 2, , . 2 2 (24d) g (2kM + r ) = pr ( For K odd, they become K 1 K1 g (2kM + r ) = qr ( + k) for k = 0, 1, , , 2 2 (25a) K 1 K 1 g ((2k1)M +r ) = pr ( +k) for k = 1, 2, , , 2 2 (25b) K 1 K 1 g (2kM + 1 r ) = qr ( k) for k = 1, 2, , , 2 2 (25c) (23c)

for k = 0, 1, ,

K1 . 2 (25d)

The key point for using the rk s as the basic adjustable parameters lies in the fact that for any selecerty can be achieved. The zero-phase frequency response of the lter with the g (n)s as
KM

tion of these parameters the perfect-recontruction prop-

the transfer function Gp(z ) can be written in terms of

Gp(, ) = 2
n=1

g (n, ) cos[(n 1/2) ],

(26a)

where is an K M/2 k = 1, 2, , K , that is,

length vector containing the and

adjustable parameters rk for r = 1, 2, , M/2 = [11, 12, , 1K , 21, 22, , 2K , ,

M/2 ,1 , M/2 ,2 , , M/2 ,K ].

(26b)

The above notations are used to emphasize the dejustable parameters.

pendences of g (n, ) and Gp(, ) on the basic ad-

 29

30

4.2 Relations of the Proposed Adjustable Angles to Those of Existing Implementation Forms There are very simple relations of the angles rk

simple form:

0, 1, , K 1, the desired relations have the following r,k = /2 r+1,k+1 . (29)

to the corresponding angles used by Malvar in [14] and Koilpillai and Vaidyanthan in [10] for generating perfect-recontruction cosine-modulated lter banks. r = 0, 1 , M/2 1 and k = 0, 1, , K 1, the relar,2k = 3/2 r+1,K 2k for k = 0, 1, , K/2 1 (27a) and r,2k+1 = /2 + r+1,K 2k1 for k = 0, 1, , K/2 1 for K even. lowing form: For K odd, the corresponding relations take the folr,2k = /2 + r+1,K 2k for k = 0, 1, , K/2 and r,2k+1 = 3/2 r+1,K 2k1 for k = 0, 1, , K/2 1. (28a) For the angles considered in [14], denoted by rk for

tions are given by

(27b)

(28b)

For the corresponding angles considered in [10], again denoted by rk for r = 0, 1 , M/2 1 and k =

31

32

4.3 Optimization Problems We consider the following two optimization problems: Problem A. Given M , K , and , nd the unknowns mize rk for r = 1, 2, , M/2 and k = 1, 2, , K to mini E2() = where s = (1 + )/(2M ). (30b) s [Gp(, )]2d, (30a)

4.4 Design Algorithm for Problem A For Problem A, the objective function can be expressed, after some manipulations, in the form
KM KM

E2() =
=1 =1

g (, )g (, )(, ),

(32)

where (, ) is given by Eq. (13b). unknowns rn r = 1, 2, , M/2

KM KM

The partial derivatives of E2() with respect to the and n = 1, 2, , K

are given by Problem B. Given M , K , and , nd the unknowns mize rk for r = 1, 2, , M/2 and k = 1, 2, , K to miniE = max |Gp(, )|
[s , ]

E2() =2 grn(, )g (, )(, ), rn =1 =1 where grn(, ) = g (, ) . rn

(33a)

(31)

(33b)

In the rst case, the stopband energy of the prototype lter is minimized, whereas in the second case zero in the stopband region is minimized. the maximum deviation of the amplitude response from

Only the elements gnk ((k 1)M + r, ) and g (kM + r, ) for k = 1, 2, , K are nonzero. The values of these elements can be determined with

of the elements g ((k 1)M + r, ) and g (kM + r, ) for k = 1, 2, , K according to Eqs. (21)(25). The only exeption is that for k = n crn = sin(rn )

the aid of the angles rk for k = 1, 2, , K like those

and srk = cos(rn ).

Knowing the partial derivates with respect to the

unknows, the optimum solution can be found in a

 33

34

straightforward manner using, for instance, the FletcherPowell algorithm [15]. The main advantage of the above formulation of the optimization problem is the fact that when changing the unknown angles during optimization, the overall lter bank is guaranteed to have all the time the perfectrecontruction property without having additional constraints.

4.5 Practical Filter Design for Problem A This subsection illustrates, by means of an example, for large values of K and M . It is desired to design a bank with M = 32 lters of length N + 1 = 512 (K = 8) for = 1. optimum is the following (see Fig. 5): A very fast procedure to arrive at least at a local Step 1: Optimize the prototype lter for M = 2, K = 8, and = 1 using the initial values 1k = 0 for k = 1, 2, , 8. Step 2: Optimize the lter for M = 4, K = 8, and = 1 using the initial values rk = 1k for r = 1, 2 and k = 1, 2, , 8, where the 1k s are the optimized parameters of the previous step. Step 3: Optimize the lter for M = 8, K = 8, and = 1 using the initial values rk = 1k for r = 1, 2 and rk = 2k for r = 3, 4, where the 1k s are the optimized parameters of the previous step. Step 4: Optimize the lter for M = 16, K = 8, and
(4) rk (4) rk (3) (2) (2) (3) (2) (1) (2) (1) (0)

how to design eectively a cosine-modulated lter bank

= 1 using the initial values rk = 1k for r = 1, 2, = =

(3) 2k (3) 4k

(4)

(3)

for r = 3, 4,

(4) rk

(3) 3k

for r = 5, 6, and are the opti-

for r = 7, 8, where the

(3) 1k s

35
Amplitude in dB

36
0 Step 1 50 100 0 0.1 0.2 0.3 0.4 0.5 0.6 Angular frequency / 0.7 0.8 0.9 1

mized parameters of the previous step. Step 5: Optimize the lter for M = 32, K = 8, and = 1 using the initial values rk = 1k for r = 1, 2, rk = 2k for r = 3, 4, rk = 3k for r = 5, 6, rk =
(5) (4) (4) (5) (4) (5) (5) (4)

Amplitude in dB

4k for r = 7, 8, rk = 5k for r = 9, 10, rk = 6k for r = 11, 12, rk = 7k for r = 13, 14, and rk = 8k
(4) (5) (4) (5)

(5)

(4)

(5)

(4) (4)

0 Step 2 50 100 0 0.1 0.2 0.3 0.4 0.5 0.6 Angular frequency / 0.7 0.8 0.9 1

for r = 15, 16. Here, where the 1k s are the optimized parameters of the previous step. For the initial lter at Step 1, g (1) = 1 and g (n) = 0 for n = 2, 3, , KM = 8. Correspondingly, hp(KM 1) = hp(KM ) = 1/(2 2)

Amplitude in dB

0 Step 3 50 100 0 0.1 0.2 0.3 0.4 0.5 0.6 Angular frequency / Step 4 50 100 0 0.1 0.2 0.3 0.4 0.5 0.6 Angular frequency / 0.7 0.8 0.9 1 0.7 0.8 0.9 1

Amplitude in dB Amplitude in dB

and the remaining impulse response values are zero. ter at Step 1.

This has turned out to be a good starting point l The basic idea of performing Steps 2, 3, 4, and 5 as described above lies in the fact that the transfer function obtained using the given initial values is expressible as Hp(z ) = Hp(z 2 )(1 + z 1)/2, previous step. (34) where Hp(z ) is the transfer function optimized at the

0 Step 5 50 100 0 0.1 0.2 0.3 0.4 0.5 0.6 Angular frequency / 0.7 0.8 0.9 1

Figure 5: Design a perfect-reconstruction prototype lter using a ve-step procedure for M = 32, N +1 = 512, and = 1. The dashed and solid lines show the initial and optimized responses, respectively. As seen from Fig. 5, this means that the amplitude response of Hp(z 2 ) is a frequency-axis compressed version of that of H (z ) in such a way that the interval

 37

38

[0, 2 ] is shrunk onto [0, ]. Therefore, it has an extra passband at = . This is partially attenuated by the term (1 + z )/2.
1

two. After the optimization, this impulse response becomes smooth as seen from Fig. 6. Figure 7 shows in the perfect-reconstruction case the prototype lter designed in the least-mean-square sense and = 1. Some characteristics of the resulting lter bank can be found in Table I. In this table, the characteristics of the lter banks to be considered later on will be collected. as well as the lters in the bank for M = 32, N = 512,

After the optimization, the extra peak around = is attenuated. It is interesting to observe from Fig. 5 that as M is
1

increased, the term (1 + z )/2 provides a better attenuation and the initial lter becomes close to the optimized one. The above procedure enables us to design very fast at least suboptimum lters for very high values of K designed using only a one-step procedure. In the time domain, Eq. (34) means that the impulse response of Hp(z ) is related to that of Hp(z ) through hp(2n) = hp(2n + 1) = hp(n)/2 where Np is the order of Hp(z ). As seen from Fig. 6, the impulse response of the initial lter is obtained from the optimized impulse response of the previous step in two steps. First, the impulse-response values are divided by two. Then, the resulting values are repeated two times, increasing the length of the resulting lter by a factor of for n = 0, 1, , Np, and M compared to the case where the overall lter is

(35)

39

40

Optimized filter at Step 2 0.15 Impulse response 0.1 0.05 0 0 10 20 30 n in samples Initial filter at Step 3 0.08 Impulse response 0.06 0.04 0.02 0 0 20 40 60 n in samples 80 100 120 40 50 60

Table I: Comparison Between Various Banks of M = 32 Filters for = 1 The lter length is N + 1 = 2KM Boldface numbers show the xed values. Design K L2 L L2 L L2 L2 L2 N 1 2 E E2
3

8 511 0

Optimized filter at Step 3 0.08 Impulse response 0.06 0.04 0.02 0 0 20 40 60 n in samples 80 100 120

L2

0 dB 8 511 0 0 dB 8 511 0.0001 2.3 106 113 dB 8 511 0.0001 1.1 105 99 dB 8 511 0 9.1 105 81 dB 8 511 0.01 5.3 107 126 dB 6 383 0.001 0.00001 -100 dB 5 319 0.01 0.0001 -80 dB

Figure 6: Impulse responses for lters involved when performing Step 3 in the proposed ve-step procedure.

1.2 10 58 dB 2.3 104 73 dB 1.0 105 100 dB 5.1 106 106 dB 4.5 104 67 dB 2.4 106 112 dB 1.7 104 75 dB 8.4 104 62 dB

7.4 109 7.5 108 5.6 1013 3.8 1011 5.4 1010 4.5 1014 2.8 1010 2.7 109

 41

42

Figure 7: for = 1.
0 Amplitude in dB

Least-squared perfect-reconstruction

4.6 Design algorithm for problem B For Problem B, an ecient synthesis scheme based been described in [13]. Alternatively, the problem can be solved by using the Dutta-Vidyasagar algorithm. To do this, we we discretize the region [s, ] into In many cases, I = 20N is a good selection. The resulting discrete minimax problem is to nd to minimize = E() = max fi(),
iI

bank of M = 32 lters of length N +1 = 8 2M = 512

Prototype Filter

on the use of the Jing-Fam algorithm [16], [17] has

50

100

150 0

0.1

0.2

0.3

0.4 0.5 0.6 Angular frequency / Filter Bank

0.7

0.8

0.9

the discrete points i [s, ] for i = 1, 2, , I .

0 Amplitude in dB

50

100

(36a)

150 0

0.1

0.2

0.3

0.4 0.5 0.6 Angular frequency / Amplitude Error T0(z)zN

0.7

0.8

0.9

where fi() = |Gp(, i)|,

x 10

10

i = 1, 2, , I.

(36b)

0.5 Amplitude 0 0.5

Figure 8 shows in the perfect-reconstruction case the prototype lter designed in the minimax sense as well
0.1
15

as the lters in the bank for M = 32, N = 512, and

0 x 10 0.2 0.3 0.4 0.5 0.6 Angular frequency / 0.7 0.8 0.9 1

= 1. Some characteristics of the resulting lter bank can be found in Table I. As can be expected, the minimum stopband attenu-

WorstCase Aliased Term Tl(z)

2.5 2 Amplitude 1.5 1 0.5

0 0

ation of the prototype lter is signicantly higher for

0.1 0.2 0.3 0.4 0.5 0.6 Angular frequency / 0.7 0.8 0.9 1

the minimax design compared to the least-squared de-

43

44

sign (73 dB compared to 58 dB). However, for the least-squared design, the stopband 10 ). energy is considerably lower (7.4 109 compared to 7.5
8

Figure 8: for = 1.
0 Amplitude in dB

Minimax perfect-reconstruction bank

of M = 32 lters of length N + 1 = 8 2M = 512

Prototype Filter

50

100

150 0

0.1

0.2

0.3

0.4 0.5 0.6 Angular frequency / Filter Bank

0.7

0.8

0.9

0 Amplitude in dB

50

100

150 0

0.1

0.2

0.3

0.4 0.5 0.6 Angular frequency / Amplitude Error T0(z)zN

0.7

0.8

0.9

x 10

10

0.5 Amplitude 0 0.5

0 x 10
15

0.1

0.2

0.3

0.4 0.5 0.6 Angular frequency /

0.7

0.8

0.9

WorstCase Aliased Term Tl(z)

2.5 2 Amplitude 1.5 1 0.5

0 0

0.1

0.2

0.3

0.4 0.5 0.6 Angular frequency /

0.7

0.8

0.9

 45

46

5 Filters with Small Amplitude Distortion |T0(ej )| Filters with very small aliasing errors can be synthefor high-order lters. sized by allowing a small amplitude distortion |T0(ej )|

bank performance is improved and the aliasing errors become smaller. This is illustrated in the least-mean-square case in Figs. 11 and 12 for 1 = 0 and 1 = 0.01, respectively. It is seen from Table I that for the perfect-reconstruction lter as well as for the lters with 1 = 0, 1 = 0.0001, and 1 = 0.1 the mininimum stopband attenuations (the stopband energies) of the prototype lters are 58 dB, 67 dB, 106 dB, and 112 dB, respectively tively). (7.4 109, 5.4 1010, 5.6 1013, and 4.5 1014, respec-

N + 1 = 512, and = 1 case, where 1 = 0.0001.

This is illustrated in Figs. 9 and 10 for the M = 32,

Figure 9 shows the amplitude responses of the least-

squared-error prototype lter and the corresponding lter bank as well as the amplitude response of T0(z ) z N and the worst-case aliased term Tl (z ). shown in Fig. 10. The corresponding results for the minimax design are These lters have been optimized without considering the aliasing errors. When comparing these responses with those of Figs. 7 and 8, it is seen that the attenuations of both the prototype lter and the lters in the bank have been improved approximately by 40 dB at the expense of very small aliasing errors and a very small amplitude distortion (see also Table I). For the least-squared design, 2 = 2.3 106 (113 dB) and for the minimax design, 2 = 1.1 105 (99 dB) By allowing a larger value for 1, both the lter

The corresponding minimum stopband attenuations of the worst-case aliased transfer functions are dB, 81 dB, 113 dB, and 126 dB, respectively.

47

48

Figure 9: Least-squared bank of M = 32 lters of length N + 1 = 8 2M = 512 for = 1 and 1 = 0.0001.

Prototype Filter 0 Amplitude in dB

Figure 10:

Minimax bank of M = 32 lters of

length N + 1 = 8 2M = 512 for = 1 and 1 = 0.0001.

Prototype Filter 0 Amplitude in dB

50

50

100

100

150 0

0.1

0.2

0.3

0.4 0.5 0.6 Angular frequency / Filter Bank

0.7

0.8

0.9

150 0

0.1

0.2

0.3

0.4 0.5 0.6 Angular frequency / Filter Bank

0.7

0.8

0.9

0 Amplitude in dB Amplitude in dB 0.1 0.2 0.3 0.4 0.5 0.6 Angular frequency / Amplitude Error T0(z)zN 0.7 0.8 0.9 1

50

50

100

100

150 0

150 0

0.1

0.2

0.3

0.4 0.5 0.6 Angular frequency / Amplitude Error T0(z)zN

0.7

0.8

0.9

x 10

x 10

Amplitude

Amplitude 0.1
6

1 0 x 10

0.2

0.3

0.4 0.5 0.6 Angular frequency /

0.7

0.8

0.9

1 0 x 10 1.2 1 Amplitude 0.8 0.6 0.4 0.2

5

0.1

0.2

0.3

0.4 0.5 0.6 Angular frequency /

0.7

0.8

0.9

WorstCase Aliased Term Tl(z)

WorstCase Aliased Term Tl(z)

2.5 2 Amplitude 1.5 1 0.5

0 0

0.1

0.2

0.3

0.4 0.5 0.6 Angular frequency /

0.7

0.8

0.9

0 0

0.1

0.2

0.3

0.4 0.5 0.6 Angular frequency /

0.7

0.8

0.9

 49

50

Figure 11: Least-squared bank of M = 32 lters of length N + 1 = 8 2M = 512 for = 1 and 1 = 0.

Prototype Filter 0 Amplitude in dB

Figure 12: Least-squared bank of M = 32 lters of length N +1 = 82M = 512 for = 1 and 1 = 0.01.
Prototype Filter 0 Amplitude in dB

50

50

100

100

150 0

0.1

0.2

0.3

0.4 0.5 0.6 Angular frequency / Filter Bank

0.7

0.8

0.9

150 0

0.1

0.2

0.3

0.4 0.5 0.6 Angular frequency / Filter Bank

0.7

0.8

0.9

0 Amplitude in dB Amplitude in dB 0.1 0.2 0.3 0.4 0.5 0.6 Angular frequency / Amplitude Error T0(z)zN 0.7 0.8 0.9 1

50

50

100

100

150 0

150 0

0.1

0.2

0.3

0.4 0.5 0.6 Angular frequency / Amplitude Error T0(z)zN

0.7

0.8

0.9

x 10

10

0.01 0.005 Amplitude 0 0.005 0.01 0 x 10

7

0.5 Amplitude 0 0.5

0 x 10
4

0.1

0.2

0.3

0.4 0.5 0.6 Angular frequency /

0.7

0.8

0.9

0.1

0.2

0.3

0.4 0.5 0.6 Angular frequency /

0.7

0.8

0.9

WorstCase Aliased Term Tl(z) 6 5 Amplitude 4 3 2 1 0.1 0.2 0.3 0.4 0.5 0.6 Angular frequency / 0.7 0.8 0.9 1

WorstCase Aliased Term Tl(z)

1 0.8 Amplitude 0.6 0.4 0.2

0 0

0 0

0.1

0.2

0.3

0.4 0.5 0.6 Angular frequency /

0.7

0.8

0.9

51

52

6 General Examples This section illustrates, by means of two examples,

lent. These gures are directly the orders of the prototype lters.

the exibility of the general design technique for reducing the lengths of the lters in the bank to achieve practically the same lter bank performance as with the perfect-reconstruction lter banks. For both examples, M = 32, = 1, and the leasttype lter. 0.001, and 2 = 105 (a 100-dB attenuation). 0.01, and 2 = 104 (a 80-dB attenuation). Figs. 13 and 14, respectively. For the rst example, N + 1 = 384 (K = 6), 1 = mean-square criterion is used for optimizing the proto-

For the second example, N + 1 = 320 (K = 5), 1 = Responses for the optimized lters are shown in As seen from Table I (or by comparing Figs. 13

and 14 with Fig. 7), these two optimized prototype lters provide a better lter bank performance than the perfect-reconstruction lter band with N + 1 = 512 (K = 8). The main advantage of the above two lter banks is that the delays caused by them are 383 samples and 319 samples, respectively, compared to 511 samples caused by the perfect-reconstruction lter bank equiva-

 53

54

Figure 13: Least-squared bank of M = 32 lters of length N + 1 = 6 2M = 384 for for = 1, 1 = 0.001, and 2 = 0.00001.
0 Amplitude in dB Prototype Filter

Figure 14: Least-squared bank of M = 32 lters of length N +1 = 5 2M = 320 for for = 1, 1 = 0.01, and 2 = 0.0001.
0 Amplitude in dB Prototype Filter

50

50

100

100

150 0

0.1

0.2

0.3

0.4 0.5 0.6 Angular frequency / Filter Bank

0.7

0.8

0.9

150 0

0.1

0.2

0.3

0.4 0.5 0.6 Angular frequency / Filter Bank

0.7

0.8

0.9

0 Amplitude in dB Amplitude in dB

50

50

100

100

150 0

0.1

0.2

0.3

0.4 0.5 0.6 Angular frequency / Amplitude Error T0(z)zN

0.7

0.8

0.9

150 0

0.1

0.2

0.3

0.4 0.5 0.6 Angular frequency / Amplitude Error T0(z)zN

0.7

0.8

0.9

1 0.5 Amplitude

x 10

0.01 0.005 Amplitude 0 0.005 0.01 0 x 10

4

0 0.5 1 0 x 10
5

0.1

0.2

0.3

0.4 0.5 0.6 Angular frequency /

0.7

0.8

0.9

0.1

0.2

0.3

0.4 0.5 0.6 Angular frequency /

0.7

0.8

0.9

WorstCase Aliased Term Tl(z)

WorstCase Aliased Term Tl(z)

1 0.8 Amplitude

1 0.8 Amplitude 0.6 0.4 0.2

0.6 0.4 0.2 0 0 0.1 0.2 0.3 0.4 0.5 0.6 Angular frequency / 0.7 0.8 0.9 1

0 0

0.1

0.2

0.3

0.4 0.5 0.6 Angular frequency /

0.7

0.8

0.9

55

56

7. Implementation Aspects The most eective way of implementing cosine[1] and [11]. In this implementation, the protype lter in both the analysis and synthesis part is implemented at the lower sampling rate using special butteries. The cosine modulation is implemented using a DCT IV block. In the next version of this pile of lecture notes, more details are included.

References
[1] H. S. Malvar, Signal Processing with Lapped Transforms. Artech House, 1992. [2] P. P. Vaidyanathan, Multirate Systems and Filter Banks. Clis: Prentice-Hall , 1993. Norwood: Englewood

modulated lter banks has been proposed by Malvar in

[3] N. J. Fliege, Multirate Digital Signal Processing. Chichester: John Wiley and Sons, 1994. [4] T. Q. Nguyen, Near-perfect-reconstruction pseudo-QMF banks, IEEE Trans. Signal Proc., vol. 42, pp. 6376, Jan. 1994. [5] F. Mintzer, On half-band, third-band, and N th-band FIR lters and their design, IEEE Trans. Acousts, Speech, Signal Process., vol. ASSP-30, pp. 734738, Oct. 1982. [6] T. Saram aki and Y. Neuvo, A class of FIR Nyquist (N th-Band) lters with zero intersymbol interference, IEEE Trans. Circuits Syst., vol. CAS-34, pp. 11821190, Oct. 1987. [7] S. R. K. Dutta and M. Vidyasagar, New algorithms for constrained minimax optimization, Mathematical Programming, vol. 13, pp. 140 155, 1977. [8] H. S. Malvar, Modulated QMF lter banks with perfect reconstruction, Electron Lett., vol. 26, pp. 906907, June 21, 1990. [9] T. A. Ramstad and J. P. Tanem, Cosine-modulated analysis-synthesis lter bank with critical sampling and perfect reconstruction, in Proc. IEEE Int. Conf. Acoust., Speech, Signal Process. (Toronto, Canada), pp. 17891792, May 1991. [10] R. D. Koilpillai and P. P. Vaidyanathan, New results on cosinemodulated FIR lter banks satisfying perfect reconstruction, in Proc. IEEE Int. Conf. Acoust., Speech, Signal Process. (Toronto, Canada), pp. 17931796, May 1991. [11] H. S. Malvar, Extended lapped transforms: Fast algorithms and applications, in Proc. IEEE Int. Conf. Acoust., Speech, Signal Process. (Toronto, Canada), pp. 17971800, May 1991. [12] R. D. Koilpillai and P. P. Vaidyanathan, Cosine-modulated FIR lter banks satisfying perfect reconstruction, IEEE Trans. Signal Proc., vol. 40, pp. 770783, Apr. 1992.

 57

[13 T. Saram aki, Designing prototype lters for perfect-reconstruction cosine-modulated lter banks, in Proc. 1992 IEEE Int. Symp. Circuits and Syst. (San Diego, CA), pp. 16051608, May 1992. [14] H. S. Malvar, Extended lapped transforms: Properties, applications, and fast algorithms, IEEE Trans. Signal Proc., vol. 40, pp. 2703 2714, Nov. 1992. [15] R. Fletcher and M. J. D. Powell, A rapidly convergent descent method for minimization, Comput. J., vol. 6, pp. 163168, 1963. [16] Z. Q. Jing, Chebyshev design of digital lters by a new minimax algorithm, Ph. D. dissertation, State University of New York at Buffalo, 1983. [17] Z. Q. Jing and A. T. Fam, An algorithm for computing continuous Chebyshev approximations, Mathematics of Computation, vol. 48, pp. 691710, Apr. 1987