Filter Ban s with Perfect Reco

T. Karp, N. J. Fliege
Hamburg University of Technology. Telecommunication Institute,
Eissendorfer Str. 40, 21071 Hamburg, Germany,
E-mail: karpCitu-harburg.d40O.de

Abstract-In [l, 21, the Modified D F T ( M D F T ) polyphaseimpulse responses of the complex modulated filters p k ( n )
filter bank has been introduced. It has been shown
that this filter bank is able to cancel adjacent chan-
nel alias spectra resulting in almost perfect reconstruc-
tion. Compared to cosine modulated filter banks, the
MDFT filter bank offers advantages such as linear phase
analysis and synthesis filters, complex valued signal
processing with a different mapping into the subbands which, in the z domain, correspond to
and low propagation delays.
In this paper, it is shown that the MDFT filter bank Pk(Z) = P(zWk)), (2)
structure provides a cancellation of all odd alias spec- nhere Z { p ( n ) }= P ( z ) .
tra. Furthermore, it is demonstrated that even perfect
In order to get causal analysis and synthesis filters, the
reconstruction can be obtained by an appropriate filter
design. impulse responses p e ( n ) are delayed by ( N - l ) / 2 samples.
Hence. the time domain representation of the analysis and
nthesis filters is
51

P Introduction: MDFT Filter Banks

N-l
The MDFT filter bank is a complex modulated 31 channel Ilk (n)= f n ( n )= p h ( n - T), n = 0, . . . N - 1. (3)
filter bank with a two step decimation of the subband sig-
nals. Upon decimating the sampling rate by the factor iU/2 Equixalently, we obtain in the z domain
( M must be even), another decimation by tn-o is realized He(z) = Fk(Z) = z
Y-1
-~

with and without a delay. The second step corresponds t o

2 Pk(Z). (4)
a polyphase decomposition. In t,hr 2-11wbbands, either To demonstrate that all analysis and synthesis filters are
the real or the imaginary part is used. respectively. Fig. 1 linear phase, we just have to take into account that the
shows a detail of this filter bank. complex modulation in equation (1) has no influence on
the zero-phase property, so that Pk (2) is still zero-phase.
X(Z)
The multiplication with z - ( ~ - ' ) / ~guaranteeing causality,
),

changes zero-phase into linear phase.

2 Alias and Distortion Functions

In order t o get the alias and distortion functions of the
SIDFT filter bank, we have t o determine the subband sig-
i nals first. Then, the reconstructed signal * ( z ) at the
output of the synthesis filter bank can be described in terms
of the input signal. The overall transfer function T ( z )can
be divided into the alias and distortion functions Talzas(z)
and Tdzst( z ), respectively.
The input signal z ( n ) is complex valued

z ( n ) = z ' ( n )+ j . z " ( n ) (5)

and its z transform can be written as
From a symmetrical, zero-phase, lowpass prototype p ( n ) .
n = -?, . .., with bandwidth 2njM (incl. positive X ( z )= X ' ( Z )4 X"(Z) (6)
and negative frequencies) and real coefficients we get the
where the real and imaginary part of z ( n )have been trans-
formed separately.
This work is supported by the "Deutsche Forschungsgemeinschaft"
(contract # F1 116-10/1). This paper as well as further publications X'(z) = 2{z'(n)}
are available via anonymous ftp to 'abnt2.et2.tu-harburg.de'.
X " ( z ) = Z { j .z"(n))

0-7803-2127-8194 $4.00 01995 IEEE 744

Let us also define

X ( z ) = X ' ( z )- X " ( z ) . 1 k odd

In a first step we ignore the sampling rate decimatio:n in

(7)
.={ 0 k even '
and the output signal of the k-th channel is
the subbands. We just regard the influence of taking the
real and imaginary part, respectively. M/2-1
1
&(z) = 4(z)[Hk(zM'")X(zWZ')
1=0

+ ( - 1 ) " M ~ k ( z W 2 " f ' ) ~ ( z ~ 2 ~ + ' )(11)

].

Inserting equations (4) and (2) into equation (11) and

adding the A4 output signals, we obtain the reconstructed
signal
1
+ Z[hk((n)
+ h M - k ( n ) ] * j .z " ( n ) (9) Z-(N-l)
M-1 M/2-1
X(z)= ~ P ( z W k )[P(ZW"+k)X(%W")
Here, we haven taken into account that the following M
k=O 1=0
relation holds true due to equations (3) and (1).

( n )= h M - k ( n ) (10)
+ (-,)k,(,,2l+l+M-k)x(Zw21+1
.
The 2nd expression of this sum is equal to zero, which
)I
Consequently, the M subband signals decimated with'out will be demonstrated by the following modification:
a delay are: M-1 M/2-1

(-1)"(ZW")P(zW2'+1-")X(zM'2'+1 1
k=O I=O

=
M/2-1
X(ZW21+1) [ M/2-1
p(zW2")p(zW"+l-2k 1
k 0 k=O
M/2-1
- P(zW2"+l)P(zW21-2k
k=O 11
Substituting I-k for k in the last sum leads to the desired
+[Hk(zW') + HM-k(21t")]X"(zW1)k odd
For the other M subband signals, decimated with a delay
1 result. Hence, k ( z ) can be rewritten as

2(z)=
M-1 M/2--1

P ~ ( Z ) H & M ' 2 ' ) X ( Z W ~ ~ )(12)

of M/2 samples, the factor W - 1 M /occurs
2 in the spectrum k=O 2=O

of the I-th modulation component. and is therefore half of the output signal of an A4 channel
DFT filter bank with a subsampling factor of M / 2 for the
subband signals (see Fig. 2).

+ [H k ( Z W 1 ) + H M - k ( z W " ] X I f (.W")] w-lMI2 k even

With the above relations and equations (6) and (7), the
subband signals are 2iz(z)

Fig. 2: Complex modulated M channel filter bank with a

YL")(z") = 1 M-l [ H k ( Z W ' ) X ( Z W " )
subsampling factor of M / 2
1=0
The distortion function of the MDFT filter bank is

745
 The influence of Tdist(z) can be ignored with practical
applications if the prototype is maximally flat for 0 < w <
rr/M and has a sufficiently high attenuation for w > 27r/?,l.
A Pseudo-QMF design with almost perfect reconstruction
based on a square root raised cosine prototype is given
in [3]. z-'
If the prototype is a spectral factor of an M-th band
filter, the distortion function simplifies to a delay and a A
2X(z)
scaling factor.
The alias function is given by Fig. 4: DFT polyphase realization of the filter bank in
Fig. 2

The multiplications with Wk((N-')/2between the IDFT

It consists of the same alias spectra as the non-critically m d DFT operations are due to the causality of the analysis
subsampled filter bank in Fig. 2 . Additional aliasing caused and synthesis filters. Remember that the symmetrical,
by critical subsampling is exactly compensated i n the. zero-phase prototype was first modulated and then time
MDFT filter bank, see Fig. 3. shifted.
If the order of the prototype is an integer multiple of
.U. is equal to one. The subband signals after
the DFT operation are identical to those before the IDFT
ktaled b) the factor M .
In the general case N = T . M + 5 , the output signal of
the DFT differs from the scaled input signal of the IDFT
because of the multiplication with Wk((N-l) = W k ( " - ' )in
the k-th subband This multiplication can be transferred
after the DFT by exploiting the effect of a frequency shift
on the DFT.

1-0 input frequency D F T { z ( ~ )= )} -s +I ) ) ~

} D F T { W - ~ ( ~ - ~ ) X (=~ x((n
output frequency
(13)
Here ((n- s + l ) ) n 4 means ''(n - s + 1) niodulo M " .
Fig. 3: Aliasing of an 8 channel MDFT filter bank using a In order to apply equation (13) to the filter bank in
windowed square root raised cosine filter (T = 0.5)for the Fig 4. n e have to interpret Z ( k ) as the k-th output signal
prototype of the IDFT, z ( k ) as the k-th input signal of the DFT, and
S ( r i - s + l ) as the ( n - s + l ) - t h output signal of the DFT.
This way. we obtain the filter bank depicted in Fig. 5.
3 Perfect Reconstruction (PR)
The results of the last section show that the output signal
of the MDFT filter bank is half of that of the non-critically
subsampled filter bank in Fig.2. Hence, it is sufficient t o
regard this filter bank with respect to PR. This simplifies
the following considerations as we are not required to take
the real and imaginary parts, respectively, in the subbands
any longer.
For PR, the following conditions have to be satisfied by
the filter bank:
e exact compensation of aliasing
e no distortion
In order t o impose conditions for PR on the analysis
and synthesis filters in Fig.2 , it is useful to regard the
DFT polyphase realization of the MDFT filter bank. In W Z )
the analysis filter bank, type 1 polyphase filters are used
Fig. 3: Simplified polyphase structure for N = r . M +s
gk(m) = ho(m. M +k), IC = 0 , . . . M - 1
The polyphase filters Gk(z) can be expressed by type 1
In the synthesis filter bank we use
;?k(7n) = h l J ( m . M - k ) , k=O, ...M - 1

746
polyphase filters:

As we use a linear phase lowpass filter H o ( z ) , exactly

two polyphase filters depend on each other [4]. I z.1

Taking- equations (14) and (15) into account, w e can

express G ~ - ~ - l + k ( as
z ) follows:
G M - ~ + ~ +=~ ~( -z~) G , - l - k-
- z-r-l 4 Conclusion
G:k(z), IC 5 s -- 2
~ ( ZGs-l-k
G M - ~ + ~ += ) = z - ' G ~ ( z ) , IC = s -- 1 In this paper, the P R conditions for the MDFT filter bank
~ Z-'GL(Z), s <: IC << M
G M - ~ + ~ +=~Z( -z~) G , - I - = have been derived for prototypes of arbitary length. We
have demonstrated that it is sufficient to require this prop-
Inserting these modifications in Fig. 5 and leaving out the erty for the non-critically decimated DFT filter bank. The
IDFT and DFT (the effect of the IDFT is compensated by analysis and synthesis filters are linear phase and identical.
the DFT), we attain the structure shown in Fig. 6. Prototypes developed for the cosine modulated filter bank
with PR can also be used for the MDFT filter bank, and
,z-(s-2) do reveal this property if they are scaled correctly.
TMQ
_I
References
N. J. Fliege. Multirate Digital Signal Processing. John
Wiley and Sons, Chichester, UK, 1994.
N. J. Fliege. Modified DFT Polyphase SBC Filter
Banks with Almost Perfect Reconstruction. In Proc.
ICASSP, volume 3 , pages 149-152, Adelaide, Australia,
April 1994.
N. J. Fliege. Closed Form Design of Prototype Filters
for Linear Phase DFT Polyphase Filter Banks. In Proc.
ISCAS, Chicago, USA, May 1993.
T. Nguyen and R. D. Koilpillai. The Design of
Arbitary-Length Cosine-Modulated Filter Banks and
Wavelets, Satisfying Perfect Reconstruction. In Prvc.
Fig. 6: Polyphase structure from Fig. 5 after modifications ISCAS, San Diego, USA, May 1992.
R. D. Koilpillai and P. P. Vaidyanathan. Cosine-
Using the fact that subband signals delayed by A412 Modulated FIR Filter Banks Satisfying Perfect Re-
samples with respect to each other can be subsarnpled construction. IEEE Trans. on Signal Processing, SP-
together, we get the simplified structure in Fig. 7. 40:770-783, April 1992.
Hence, the condition for P R on the polyphase filters is Y. Lin and P. P. Vaidyanathan. Linear Phase Cosine
2 Modulated Maximally Decimated Filter Banks with
GI, = M / -z
( 2 ) G k (2)+ G ~ + M / ~ ( ~ ) G( z~) + (16) Perfect Reconstruction. In Prvc. ISCAS, volume 2,
pages 17-19, London, UK, June 1994.
This form of the P R condition is equivalent t o those T. Nguyen. A Quadratic-Constrained Least-Squares
for cosine modulated filter banks given in [5, 4, 61. Con- Approach to the Design of Digital Filter Banks. In
sequently, prototypes designed for an M channel cosine Proc. ISCAS, San Diego, USA, May 1992.
modulated filter bank can be applied to a 2M channel T. Nguyen. Digital Filter Banks Design - Quadratic-
MDFT filter bank if their impulse reponses are scalled by Constrained Formulation. submitted to IEEE Trans.
a. Furthermore, the lattice structure for the ]polyphase on Signal Processing, 1994.
filters proposed in [5] can be used, too.
The design of a prototype P ( z ) leading t o perfelct re-
construction has t o be done via non-linear optimization
routines. Design methodes can be found in [7, 81.

747