Multirate Signal Processing

By:
Douglas L. Jones
Multirate Signal Processing

By:
Douglas L. Jones

Online:
< http://cnx.org/content/col10287/1.1/ >

CONNEXIONS

Rice University, Houston, Texas

This selection and arrangement of content as a collection is copyrighted by Douglas L. Jones. It is licensed under the
Creative Commons Attribution 2.0 license (http://creativecommons.org/licenses/by/2.0/).
Collection structure revised: June 15, 2005
PDF generated: October 25, 2012
For copyright and attribution information for the modules contained in this collection, see p. 37.
 Table of Contents
1 Overview of Multirate Signal Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
2 Interpolation, Decimation, and Rate Changing by Integer Fractions . . . . . . . . . . . . . . . . . . . . . . . . 5
3 Ecient Multirate Filter Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
4 Filter Design for Multirate Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
5 Multistage Multirate Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
6 DFT-Based Filterbanks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
7 Quadrature Mirror Filterbanks (QMF) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
8 M-Channel Filter Banks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
Attributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
iv

Available for free at Connexions <http://cnx.org/content/col10287/1.1>

Chapter 1

Overview of Multirate Signal Processing 1

Digital transformation of the sampling rate of signals, or signal processing with dierent sampling rates in
the system.

1.1 Applications
1. CD to DAT format change, for example.
2. oversampling converters; which reduce performance requirements on anti-aliasing or reconstruction
lters
3. bandwidth of individual channels is much less than the overall bandwidth
4. Eyes and ears are not as sensitive to errors in higher frequency bands, so many coding schemes split
signals into dierent frequency bands and quantize higher-frequency bands with much less precision.

1.2 Outline of Multirate DSP material

1. General Rate-changing System (Section 1.3: General Rate-Changing Procedure)
2. Integer-factor Interpolation and Decimation and Rational-factor Rate Changing (Chapter 2)
3. Ecient Multirate Filter Structures (Chapter 3)
4. Optimal Filter Design for Multirate Systems (Chapter 4)
5. Multi-stage Multirate Systems (Chapter 5)
6. Oversampling D/As (Chapter 6)
7. Perfect-Reconstruction Filter Banks and Quadrature Mirror Filters (Chapter 7)

1.3 General Rate-Changing Procedure

This procedure is motivated by an analog-based method: one conceptually simple method to change the
sampling rate is to simply convert a digital signal to an analog signal and resample it! (Figure 1.1)
1 This content is available online at <http://cnx.org/content/m12777/1.3/>.

Available for free at Connexions <http://cnx.org/content/col10287/1.1>

1
2 CHAPTER 1. OVERVIEW OF MULTIRATE SIGNAL PROCESSING

Figure 1.1

1 if || <
T1
Haa () =
0 otherwise
 
sin T1 t
haa (t) =
T1 t
Recall the ideal D/A:  
(tnT0 )
X sin T0
x0a (t) = x0 (n) (tnT0 )
(1.1)
n= T0
The problems with this scheme are:
1. A/D, D/A, lters cost money
2. imperfections in these devices introduce errors
Digital implementation of rate-changing according to this formula has three problems:
1. Innite sum: The solution is to truncate. Consider sinc (t) ' 0 for t < t1 , t > t2 : Then mT1 nT0 < t1
and mT1 nT0 > t2 which implies
mT1 t2
N1 = d e
T0
mT1 t1
N2 = b c
T0
N2
X
x1 (m) = x0 (n) sincT 0 (mT1 nT0 )
n=N1

note: This is essentially lowpass lter design using a boxcar window: other nite-length
lter design methods could be used for this.
2. Lack of causality: The solution is to delay by max {|N |} samples. The mathematics of the analog
portions of this system can be implemented digitally.

x1 (m) = haa (t) x0a (t) |t=mT1

(1.2)

(mT1 nT0 )
R P sin T0 sin
T1
= n= x0 (n) (mT1 nT0 ) d
T0 T1

P sin( T0 (mT1 nT0 ))

x1 (m) = x0 (n)
(mT1 nT0 ) |T 0=max{T0 ,T1 }
n=
P T0 (1.3)
= n= x0 (n) sincT 0 (mT1 nT0 )
So we have an all-digital formula for exact digital-to-digital rate changing!
Available for free at Connexions <http://cnx.org/content/col10287/1.1>
 3

3. Cost of computing sincT 0 (mT1 nT0 ): The solution is to precompute the table of sinc (t) values.
However, if TT01 is not a rational fraction, an innite number of samples will be needed, so some
approximation will have to be tolerated.
note: Rate transformation of any rate to any other rate can be accomplished digitally with
arbitrary precision (if some delay is acceptable). This method is used in practice in many
cases. We will examine a number of special cases and computational improvements, but in
some sense everything that follows are details; the above idea is the central idea in multirate
signal processing.

Useful references for the traditional material (everything except PRFBs) are Crochiere and Rabiner,
1981 [1] and Crochiere and Rabiner, 1983 [2]. A more recent tutorial is Vaidyanathan [4]; see also Rioul and
Vetterli [3]. References to most of the original papers can be found in these tutorials.

Available for free at Connexions <http://cnx.org/content/col10287/1.1>

4 CHAPTER 1. OVERVIEW OF MULTIRATE SIGNAL PROCESSING

Available for free at Connexions <http://cnx.org/content/col10287/1.1>

Chapter 2

Interpolation, Decimation, and Rate

Changing by Integer Fractions 1

2.1 Interpolation: by an integer factor L

Interpolation means increasing the sampling rate, or lling in in-between samples. Equivalent to sampling
a bandlimited analog signal L times faster. For the ideal interpolator,

X (L) if || <
0
X1 () = L
(2.1)
0 if ||
L

We wish to accomplish this digitally. Consider (2.2) and Figure 2.1.

X m
if m = {0, (L) , (2L) , . . . }
0 L
y (m) = (2.2)
0 otherwise

Figure 2.1

1 This content is available online at <http://cnx.org/content/m12801/1.3/>.

Available for free at Connexions <http://cnx.org/content/col10287/1.1>

5
 CHAPTER 2. INTERPOLATION, DECIMATION, AND RATE CHANGING
6
BY INTEGER FRACTIONS

The DTFT of y (m) is

P (im)
Y () = m= y (m) e
P (iLn)
= n= x0 (n) e
P (2.3)
(iLn)
= n= x (n) e
= X0 (L)

Since X0 ( 0 ) is periodic with a period of 2 , X0 (L) = Y () is periodic with a period of 2

L (see Figure 2.2).

Figure 2.2

By inserting zero samples between the samples of x0 (n), we obtain a signal with a scaled frequency
response that simply replicates X0 ( 0 ) L times over a 2 interval!
Obviously, the desired x1 (m) can be obtained simply by lowpass ltering y (m) to remove the replicas.

x1 (m) = y (m) hL (m) (2.4)

Given
1 if || <
L
HL (m) =
0 if ||
L

In practice, a nite-length lowpass lter is designed using any of the methods studied so far (Figure 2.3
(Interpolator Block Diagram)).

Interpolator Block Diagram

Figure 2.3

2.2 Decimation: sampling rate reduction (by an integer factor M)

Let y (m) = x0 (Lm) (Figure 2.4)

Available for free at Connexions <http://cnx.org/content/col10287/1.1>

 7

Figure 2.4

That is, keep only every Lth sample (Figure 2.5)

Figure 2.5

In frequency (DTFT):
P
Y () = m= y (m) e(im)
P
= x0 (M m) e(im)
m=
P P (i Mn
)|
= n= x0 (n) k= (n M k) e n=M m
(2.5)
0
P P (i n)
= n= x 0 (n) k= (n M k) e |
0= M
P
= DTFT [x0 (n)] DTFT [ (n M k)]
PM 1
Now DTFT [ (n M k)] = 2 k=0 X (k) 0 2k for || < as shown in homework #1 , where
P

M
X (k) is the DFT of one period of the periodic sequence. In this case, X (k) = 1 for k {0, 1, . . . , M 1}
PM 1
and DTFT [ (n M k)] = 2 k=0 0 2k M .
P

PM 1
= X0 ( 0 ) 2 k=0 0 2k
P

DTFT [x0 (n)] DTFT [ (n M k)] M
R  PM 1

1
= 2
X0 (0 ) 2 k=0 0 0 2k
M d 0 (2.6)
PM 1 2k


= k=0 X0 0 M

Available for free at Connexions <http://cnx.org/content/col10287/1.1>

 CHAPTER 2. INTERPOLATION, DECIMATION, AND RATE CHANGING
8
BY INTEGER FRACTIONS

digital aliasing.(Figure 2.6)

PM 1
so Y () = 2k
i.e., we get


k=0 X0 M M

Figure 2.6

Usually, we prefer not to have aliasing, so the downsampler is preceded by a lowpass lter to remove all
frequency components above || < M
(Figure 2.7).

Figure 2.7

2.3 Rate-Changing by a Rational Fraction L/M

This is easily accomplished by interpolating by a factor of L, then decimating by a factor of M (Figure 2.8).

Figure 2.8

Available for free at Connexions <http://cnx.org/content/col10287/1.1>

 9

The two lowpass lters can be combined into one LP lter with the lower cuto,

1 if || <
max{L,M }
H () =
0 if
||
max{L,M }

Obviously, the computational complexity and simplicity of implementation will depend on M:

L
2/3 will be
easier to implement than 1061/1060!

Available for free at Connexions <http://cnx.org/content/col10287/1.1>

 CHAPTER 2. INTERPOLATION, DECIMATION, AND RATE CHANGING
10
BY INTEGER FRACTIONS

Available for free at Connexions <http://cnx.org/content/col10287/1.1>

Chapter 3

Ecient Multirate Filter Structures 1

Rate-changing appears expensive computationally, since for both decimation and interpolation the lowpass
lter is implemented at the higher rate. However, this is not necessary.

3.1 Interpolation
For the interpolator, most of the samples in the upsampled signal are zero, and thus require no computation.
(Figure 3.1)

Figure 3.1

1 This content is available online at <http://cnx.org/content/m12800/1.3/>.

Available for free at Connexions <http://cnx.org/content/col10287/1.1>

11
12 CHAPTER 3. EFFICIENT MULTIRATE FILTER STRUCTURES

For m = Lb m
L c + mmodL and p = mmodL,

PN2
x1 (m) = m=N1 hLp (m) y (m)
P NL2 (3.1)
gp (k) x0 b m


= N
k= L1 Lck

gp (n) = h (Ln + p)
Pictorially, this can be represented as in Figure 3.2.

Figure 3.2

These are called polyphase structures, and the gp (n) are called polyphase lters.
Computational cost
If h (m) is a length-N lter:

No simplication: = LN computations
N
T1
 T0  sec
Polyphase structure: L N
L 1
To
computations
sec = N
T0 where L is the number of lters, N
L is the taps/lter,
0

and 1
T0 is the rate.

Thus we save a factor of L by not being dumb.

note: For a given precision, N is proportional to L, (why?), so the computational cost does
increase with the interpolation rate.

Question: Can similar computational savings be obtained with IIR structures?

3.2 Ecient Decimation Structures

We only want every M th output, so we compute only the outputs of interest. (Figure 3.3 (Polyphase
Decimation Structure))
N2
X
x1 (m) = x0 (Lm k) h (k)
k=N1

Available for free at Connexions <http://cnx.org/content/col10287/1.1>

 13

Polyphase Decimation Structure

Figure 3.3

The decimation structures are ow-graph reversals of the interpolation structure. Although direct imple-
mentation of the full lter for every M th sample is obvious and straightforward, these polyphase structures
give some idea as to how one might evenly partition the computation over M cycles.

3.3 Ecient L/M rate changers

Interpolate by L and decimate by M (Figure 3.4).

Figure 3.4

Combine the lowpass lters (Figure 3.5).

Figure 3.5

We can couple the lowpass lter either to the interpolator or the decimator to implement it eciently
(Figure 3.6).

Available for free at Connexions <http://cnx.org/content/col10287/1.1>

14 CHAPTER 3. EFFICIENT MULTIRATE FILTER STRUCTURES

Figure 3.6

Of course we only compute the polyphase lter output selected by the decimator.
Computational Cost
Every T1 = L T0 seconds,
M
compute one polyphase lter of length L,
N
or
N N
L L N multiplies
= M
=
T1 L T0
M T0 second

However, note that N is proportional to max {L, M }.

Available for free at Connexions <http://cnx.org/content/col10287/1.1>

Chapter 4

Filter Design for Multirate Systems 1

The lter design techniques2 learned earlier can be applied to the design of lters in multirate systems, with
a few twists.
Example 4.1
Design a factor-of-L interpolator for use in a CD player, we might wish that the out-of-band error
be below the least signicant bit, or 96dB down, and < 0.05% error in the passband, so these
specications could be used for optimal L lter design.
In a CD player, the sampling rate is 44.1kHz, corresponding to a Nyquist frequency of 22.05kHz, but
the sampled signal is bandlimited to 20kHz. This leaves a small transition band, from 20kHz to 24.1kHz.
However, note that in any case where the signal spectrum is zero over some band, this introduces other zero
bands in the scaled, replicated spectrum (Figure 4.1).

Figure 4.1

So we need only control the lter response in the stopbands over the frequency regions with nonzero
energy. (Figure 4.2)
1 This content is available online at <http://cnx.org/content/m12773/1.3/>.
2Digital Filter Design <http://cnx.org/content/col10285/latest/>

Available for free at Connexions <http://cnx.org/content/col10287/1.1>

15
16 CHAPTER 4. FILTER DESIGN FOR MULTIRATE SYSTEMS

Figure 4.2

The extra "don't care" bands allow a given set of specications to be satised with a shorter-length lter.

4.1 Direct polyphase lter design

Note that in an integer-factor interpolator, each set of output samples x1 (Ln + p), p = {0, 1, . . . , L 1},
is created by a dierent polyphase lter gp (n), which has no interaction with the other polyphase lters
except in that they each interpolate the same signal. We can thus treat the design of each polyphase lter
independently, as an NL -length lter design problem. (Figure 4.3)

Figure 4.3

Each gp (n) produces samples x1 (Ln + p) = x0 n + Lp , where n + Lp is not an integer. That is, gp (n) is

to produce an output signal (at a T0 rate) that is x0 (n) time-advanced by a non-integer advance Lp .
The desired response of this polyphase lter is thus
ip
HDp () = e L

for || , an all-pass lter with a linear, non-integer, phase. Each polyphase lter can be designed
independently to approximate this response according to any of the design criteria developed so far.
Exercise 4.1 (Solution on p. 18.)
What should the polyphase lter for p = 0 be?
Example 4.2: Least-squares Polyphase Filter Design
Deterministic x(n): Minimize

X 2
(|x (n) xd (n) |)
n=

Given x (n) = x (n) h (n) and xd (n) = x (n) hd (n). Using Parseval's theorem, this becomes
nP o n R o
2 1 2
min n= (|x (n) xd (n) |) = min 2
(|X () H () X () Hd () |) d
n R
2
o (4.1)
1
= min 2
|H () H d () |(|X () |) d

Available for free at Connexions <http://cnx.org/content/col10287/1.1>

 17

2
This is simply weighted least squares design, with (|X () |) as the weighting function.
stochastic X():
n h io h i
2 2
min E (|x (n) xd (n) |) = E (|x (n) (h (n) hd (n)) |)
n R
2
o (4.2)
1
= min 2
(|Hd () H () |) Sxx () d

Sxx () is the power spectral density of x.

Sxx () = DTFT [rxx (k)]

h i
rxx (k) = E x (k + l) x (l)

Again, a weighted least squares lter design problem.

Problem
Is it feasible to use IIR polyphase lters?
Solution
The recursive feedback of previous outputs means that portions of each IIR polyphase lter must
be computed for every input sample; this usually makes IIR lters more expensive than FIR im-
plementations.

Available for free at Connexions <http://cnx.org/content/col10287/1.1>

18 CHAPTER 4. FILTER DESIGN FOR MULTIRATE SYSTEMS

Solutions to Exercises in Chapter 4

Solution to Exercise 4.1 (p. 16)
A delta function: h0 (n) = (n0 )

Available for free at Connexions <http://cnx.org/content/col10287/1.1>

Chapter 5

Multistage Multirate Systems 1

Multistage multirate systems are often more ecient. Suppose oneQpwishes to decimate a signal by an integer
factor M , where M is a composite integer M = M1 M2 . . . Mp = i=1 Mi . A decimator can be implemented
in a multistage fashion by rst decimating by a factor M1 , then decimating this signal by a factor M2 , etc.
(Figure 5.1 (Multistage decimator))

Multistage decimator

Figure 5.1

Multistage implementations are of signicant practical interest only if they oer signicant computational
savings. In fact, they often do!
The computational cost of a single-stage interpolator is:
N taps
M T0 sec
The computational cost of a multistage interpolator is:
N1 N2 Np
+ + +
M1 T0 M1 M2 T0 M T0
1 This content is available online at <http://cnx.org/content/m12803/1.3/>.

Available for free at Connexions <http://cnx.org/content/col10287/1.1>

19
20 CHAPTER 5. MULTISTAGE MULTIRATE SYSTEMS

The rst term is the most signicant, since the rate is highest. Since Ni Mi for a lowpass lter, it is not
immediately clear that a multistage system should require less computation. However, the multistage struc-
ture relaxes the requirements on the lters, which reduces their length and makes the overall computation
less.

5.1 Filter design for Multi-stage Structures

Ostensibly, the rst-stage lter must be a lowpass lter with a cuto at M1 , to prevent aliasing after the
downsampler. However, note that aliasing outside the nal overall passband || < M
is of no concern, since
it will be removed by later stages. We only need prevent aliasing into the band || < hM
; thus we need onlyi
specify the passband over the interval || < M , and the stopband over the intervals M1 M
2k 2k
, M1 + M
,
for k {1, . . . , M 1}. (Figure 5.2)

Figure 5.2

Of course, we don't want gain in the transition bands, since this would need to be suppressed later,
but otherwise we don't care about the response in those regions. Since the transition bands are so large,
the required lter turns out to be quite short. The nal stage has no "don't care" regions; however, it is
operating at a low rate, so it is relatively unimportant if the nal lter turns out to be rather long!

5.2 L-innity Tolerances on the Pass and Stopbands

The overall response is a cascade of multiple lters, so the worst-case overall passband deviation, assuming
all the peaks just happen to line up, is
Yp
1 + pov = 1 + pi
i=1
p
Y
1 pov = 1 pi
i=1

So one could choose all lters to have equal specications and require for each-stage lter. For pov  1,

1 + p+i p 1 + pov ' 1 + p1 pov

p

1 pi 1 pov ' 1 p1 pov

p
p

Alternatively, one can design later stages (at lower computation rates) to compensate for the passband ripple
in earlier stages to achieve exceptionally accurate passband response.
s remains essentially unchanged, since the worst-case scenario is for the error to alias into the passband
and undergo no further suppression in subsequent stages.

Available for free at Connexions <http://cnx.org/content/col10287/1.1>

 21

5.3 Interpolation
Interpolation is the ow-graph reversal of the multi-stage decimator. The rst stage has a cuto at
L
(Figure 5.3):

Figure 5.3

However, all subsequent stages have large bands without signal energy, due to the earlier stages (Fig-
ure 5.4):

Figure 5.4

The order of the lters is reversed, but otherwise the lters are identical to the decimator lters.

5.4 Ecient Narrowband Lowpass Filtering

A very narrow lowpass lter requires a very long FIR lter to achieve reasonable resolution in the frequency
response. However, were the input sampled at a lower rate, the cuto frequency would be correspondingly
higher, and the lter could be shorter!
The transition band is also broader, which helps as well. Thus, Figure 5.5 can be implemented as
Figure 5.6.

Available for free at Connexions <http://cnx.org/content/col10287/1.1>

22 CHAPTER 5. MULTISTAGE MULTIRATE SYSTEMS

Figure 5.5

Figure 5.6

and in practice the inner lowpass lter can be coupled to the decimator or interpolator lters. If the dec-
imator and interpolator are implemented as multistage structures, the overall algorithm can be dramatically
more ecient than direct implementation!

Available for free at Connexions <http://cnx.org/content/col10287/1.1>

Chapter 6

DFT-Based Filterbanks 1

One common application of multirate processing arises in multirate, multi-channel lter banks (Figure 6.1).

Figure 6.1

One application is separating frequency-division-multiplexed channels. If the lters are narrowband, the
output channels can be decimated without signicant aliasing.
Such structures are especially attractive when they can be implemented eciently. For example, if
the lters are simply frequency modulated (by e(i L n) ) versions of each other, they can be eciently
2k

implemented using FFTs!

Furthermore, there are classes of lters called perfect reconstruction lters which are of nite length
but from which the signal can be reconstructed exactly (using all M channels), even though the output of
each channel experiences aliasing in the decimation step. These types of lterbanks have received a lot of
research attention, culminating in wavelet theory and techniques.

6.1 Uniform DFT Filter Banks

Suppose we wish to split a digital input signal into N frequency bands, uniformly
spaced at center frequencies
1 if || <
k = 2k
N , for 0 k N 1. Consider also a lowpass lter h (n), H () ' N
. Bandpass
0 otherwise

1 This content is available online at <http://cnx.org/content/m12771/1.3/>.

Available for free at Connexions <http://cnx.org/content/col10287/1.1>

23
24 CHAPTER 6. DFT-BASED FILTERBANKS

lters can be constructed which have the frequency response

 
2k
Hk () = H +
N

from
hk (n) = h (n) e(i )
2kn
N

The output of the k th bandpass lter is simply (assume h (n) are FIR)

x (n m) h (m) e(i )
PM 1 2km
x (n) hk (n) = N
m=0
(6.1)
= yk (n)

This looks suspiciously like a DFT, except that M 6= N , in general. However, if we x M =

N , then we can compute all yk (n) outputs simultaneously using an FFT of x (n m) h (m): The
kth FFT frequency output = yk (n)! So the cost of computing all of these lter banks outputs is O [N logN ],
rather than N 2 , per a given n. This is very useful for ecient implementation of transmultiplexors (FDM
to TDM).
Exercise 6.1 (Solution on p. 25.)
How would we implement this eciently if we wanted to decimate the individual channels yk (n)
by a factor of N , to their approximate Nyquist bandwidth?
Exercise 6.2 (Solution on p. 25.)
Do you expect signicant aliasing? If so, how do you propose to combat it? Eciently?
Exercise 6.3 (Solution on p. 25.)
How might one convert from N input channels into an FDM signal eciently? (Figure 6.2)

Figure 6.2

note: Such systems are used throughout the telephone system, satellite communication links,
etc.

Available for free at Connexions <http://cnx.org/content/col10287/1.1>

 25

Solutions to Exercises in Chapter 6

Solution to Exercise 6.1 (p. 24)
Simply step by N time samples between FFTs.
Solution to Exercise 6.2 (p. 24)
Aliasing should be expected. There are two ways to reduce it:

1. Decimate by less ("oversample" the individual channels) such as decimating by a factor of N2 . This is
eciently done by time-stepping by the appropriate factor.
2. Design better (and thus longer) lters, say of length LN . These can be eciently computed by
producing only N (every Lth) FFT outputs using simplied FFTs.

Solution to Exercise 6.3 (p. 24)

Use an FFT and an inverse FFT for the modulation (TDM to FDM) and demodulation (FDM to TDM),
respectively.

Available for free at Connexions <http://cnx.org/content/col10287/1.1>

26 CHAPTER 6. DFT-BASED FILTERBANKS

Available for free at Connexions <http://cnx.org/content/col10287/1.1>

Chapter 7

Quadrature Mirror Filterbanks (QMF) 1

Although the DFT lterbanks are widely used, there is a problem with aliasing in the decimated channels. At
rst glance, one might think that this is an insurmountable problem and must simply be accepted. Clearly,
with FIR lters and maximal decimation, aliasing will occur. However, a simple example will show that it
is possible to exactly cancel out aliasing under certain conditions!!!
Consider the following trivial lterbank system, with two channels. (Figure 7.1)

Figure 7.1

^
Note x (n) = x (n) with no error whatsoever, although clearly aliasing occurs in both channels! Note
that the overall data rate is still the Nyquist rate, so there are clearly enough degrees of freedom available
to reconstruct the data, if the lterbank is designed carefully. However, this isn't splitting the data into
separate frequency bands, so one questions whether something other than this trivial example could work.
Let's consider a general two-channel lterbank, and try to determine conditions under which aliasing can
be cancelled, and the signal can be reconstructed perfectly (Figure 7.2).

Figure 7.2

1 This content is available online at <http://cnx.org/content/m12770/1.3/>.

Available for free at Connexions <http://cnx.org/content/col10287/1.1>

27
28 CHAPTER 7. QUADRATURE MIRROR FILTERBANKS (QMF)

^
Let's derive x (n), using z-transforms, in terms of the components of this system. Recall (Figure 7.3) is
equivalent to
Y (z) = H (z) X (z)

Y () = H () X ()

Figure 7.3

and note that (Figure 7.4) is equivalent to

X
x (m) z (Lm) = x z L


Y (z) =
m=

Y () = X (L)

Figure 7.4

and (Figure 7.5) is equivalent to

M 1
1 X  1 k
Y (z) = X z M WM
M
k=0

M 1  
1 X 2k
Y () = X +
M M M
k=0

Figure 7.5

Available for free at Connexions <http://cnx.org/content/col10287/1.1>

 29

Y (z) is derived in the downsampler as follows:

X
Y (z) = x (M m) z m
m=

Let n = M m and m = M,
n
then

n
X X
Y (z) = x (n) (n M p) z M
n= p=

Now h
i
P PM 1
x (n) p= (n M p) = IDFT x () 2
M k=0 M
2k
h P
i
M 1
= IDFT 2M k=0 X 2k
M
(7.1)
1
PM 1 nk
= M k=0 X (n) WM | i2
W M=e M

so  
P PM 1 nk n
Y (z) = n=
1
M x (n) WM
k=0 z M
 n
(7.2)
1
PM 1 +k M 1
= M k=0 x (n) WM z
1
PM 1  1 k 
= M k=0 X z WM
M

^ ^
Armed with these results, let's determine X (z) x (n). (Figure 7.6)

Figure 7.6

Note
U1 (z) = X (z) H0 (z)

1
1 X  1 i2k   1  1  1  1 1  1  1
U2 (z) = X z 2 e 2 H0 z 2 e(ik) = X z 2 H0 z 2 + X z 2 H0 z 2
2 2 2
k=0

1 1
U3 (z) = X (z) H0 (z) + X (z) H0 (z)
2 2
1 1
U4 (z) = F0 (z) H0 (z) X (z) + F0 (z) H0 (z) X (z)
2 2
and
1 1 1 1
L4 (z) = F1 (z) H1 (z) X (z) + F1 (z) H1 (z) X (z) = F1 (z) H1 (z) X (z) + F1 (z) H1 (z) X (z)
2 2 2 2

Available for free at Connexions <http://cnx.org/content/col10287/1.1>

30 CHAPTER 7. QUADRATURE MIRROR FILTERBANKS (QMF)

Finally then,

^
X (z) = U4 (z)+L4 (z) = 12 (H0 (z) F0 (z) X (z) + H0 (z) F0 (z) X (z) + H1 (z) F1 (z) X(7.3)
(z) + H1 (z) F1 (
1 1
2
(H0 (z) F0 (z) + H1 (z) F1 (z)) X (z) + 2 (H0 (z) F0 (z) + H1 (z) F1 (z)) X (z)

Note that the X (z) X ( + ) corresponds to the aliasing terms!

There are four things we would like to have:
1. No aliasing distortion
2. No phase distortion (overall linear phase simple time delay)
3. No amplitude distortion
4. FIR lters
No aliasing distortion
^
By insisting that H0 (z) F0 (z) + H1 (z) F1 (z) = 0, the X (z) component of X (z) can be removed, and
all aliasing will be eliminated! There may be many choices for H0 , H1 , F0 , F1 that eliminate aliasing, but
most research has focused on the choice
F0 (z) = H1 (z) : F1 (z) = H0 (z)
We will consider only this choice in the following discussion.
Phase distortion
The transfer function of the lter bank, with aliasing cancelled, becomes T (z) =
1
2 (H0 (z) F0 (z) + H1 (z) F1 (z)), which with the above choice becomes T (z) =
1
2 (H0 (z) H1 (z) H1 (z) H0 (z)). We would like T (z) to correspond to a linear-phase lter to
eliminate phase distortion: Call
P (z) = H0 (z) H1 (z)
Note that
1
T (z) = (P (z) P (z))
2
n n
Note that P (z) (1) p (n), and that if p (n) is a linear-phase lter, (1) p (n) is also (perhaps of the
opposite symmetry). Also note that the sum of two linear-phase lters of the same symmetry (i.e., length of
p (n) must be odd) is also linear phase, so if p (n) is an odd-length linear-phase lter, there will be no phase
distortion. Also note that

2p (n) if n is odd
n
Z 1 (p (z) p (z)) = p (n) (1) p (n) =
0 if n is even

means p (n) = 0, when n is even. If we choose h0 (n) and h1 (n) to be linear phase, p (n) will also be linear
phase. Thus by choosing h0 (n) and h1 (n) to be FIR linear phase, we eliminate phase distortion and get
FIR lters as well (condition 4).
Amplitude distortion
Assuming aliasing cancellation and elimination of phase distortion, we might also desire no amplitude
distortion ( |T () | = 1). All of these conditions require
1
T (z) = (H0 (z) H1 (z) H1 (z) H0 (z)) = cz D
2
where c is some constant and D is a linear phase delay. c = 1 for |T () | = 1. It can be shown by considering
that the following can be satised!

2p (z) = 2c (n D) if n is odd
T (z) = P (z) P (z) = 2cz D
p (n) = anything if n is even

Available for free at Connexions <http://cnx.org/content/col10287/1.1>

 31

Thus we require
0
N
X
P (z) = p (2n) z (2n) + z D
n=0

Any factorization of a P (z) of this form, P (z) = A (z) B (z) can lead to a Perfect Reconstruction lter bank
of the form
H0 (z) = A (z)

H1 (z) = B (z)
[This result is attributed to Vetterli.] A well-known special case (Smith and Barnwell)


H1 (z) = z (2D)+1 H0 z 1

Design techniques exist for optimally choosing the coecients of these lters, under all of these constraints.
Quadrature Mirror Filters
H1 (z) = H0 (z) H1 () = H0 ( + ) = H0 ( ) (7.4)

for real-valued lters. The frequency response is "mirrored" around = 2 . This choice leads to T (z) =
H0 2 (z) H0 2 (z): it can be shown that this can be a perfect reconstruction system only if

H0 (z) = c0 z (2n0 ) + c1 z (2n1 )

which isn't a very exible choice of lters, and not a very good lowpass! The Smith and Barnwell approach
is more commonly used today.

Available for free at Connexions <http://cnx.org/content/col10287/1.1>

32 CHAPTER 7. QUADRATURE MIRROR FILTERBANKS (QMF)

Available for free at Connexions <http://cnx.org/content/col10287/1.1>

Chapter 8

M-Channel Filter Banks 1

The theory of M-band QMFBs and PRFBs has been investigated recently. Some results are available.

8.1 Tree-structured lter banks

Once we have a two-band PRFB, we can continue to split the subbands with similar systems! (Figure 8.1)

Figure 8.1

Thus we can recursively decompose a signal into 2p bands, each sampled at 2p th the rate of the original
signal, and reconstruct exactly! Due to the tree structure, this can be quite ecient, and in fact close to the
eciency of an FFT lter bank, which does not have perfect reconstruction.

8.2 Wavelet decomposition

We need not split both the upper-frequency and lower-frequency bands identically. (Figure 8.2)
1 This content is available online at <http://cnx.org/content/m12775/1.3/>.

Available for free at Connexions <http://cnx.org/content/col10287/1.1>

33
34 CHAPTER 8. M-CHANNEL FILTER BANKS

Figure 8.2

This is good for image coding, because the energy tends to be distributed such that after a wavelet
decomposition, each band has roughly equal energy.

Available for free at Connexions <http://cnx.org/content/col10287/1.1>

Bibliography
[1] R.E. Crochiere and L.R. Rabiner. Interpolation and decimation of digital signals: A tutorial review.
Proc. IEEE, 69(3):300331, March 1981.

[2] R.E. Crochiere and L.R. Rabiner. Multirate Digital Signal Processing . Prentice-Hall, Englewood Clis,
NJ, 1983.

[3] O. Rioul and M. Vetterli. Wavelets and signal processing. IEEE Signal Processing Magazine , 8(4):1438,
October 1991.

[4] P.P Vaidyanathan. Multirate digital lters, lter banks, polyphase networks, and applications: A tutorial.
Proc. IEEE, 78(1):5693, January 1990.

Available for free at Connexions <http://cnx.org/content/col10287/1.1>

35
36 INDEX

Index of Keywords and Terms

Keywords are listed by the section with that keyword (page numbers are in parentheses). Keywords
do not necessarily appear in the text of the page. They are merely associated with that section. Ex.
apples, 1.1 (1) Terms are referenced by the page they appear on. Ex. apples, 1

D digital aliasing, 8 P perfect reconstruction lters, 23

I Interpolation, 5 polyphase lters, 12

polyphase structures, 12
M multirate signal processing, 1(1), 2(5), T transmultiplexors, 24
3(11), 4(15), 5(19), 6(23), 7(27),
8(33)

Available for free at Connexions <http://cnx.org/content/col10287/1.1>

ATTRIBUTIONS 37

Attributions
Collection: Multirate Signal Processing
Edited by: Douglas L. Jones
URL: http://cnx.org/content/col10287/1.1/
License: http://creativecommons.org/licenses/by/2.0/
Module: "Overview of Multirate Signal Processing"
By: Douglas L. Jones
URL: http://cnx.org/content/m12777/1.3/
Pages: 1-3
Copyright: Douglas L. Jones
License: http://creativecommons.org/licenses/by/2.0/
Module: "Interpolation, Decimation, and Rate Changing by Integer Fractions"
By: Douglas L. Jones
URL: http://cnx.org/content/m12801/1.3/
Pages: 5-9
Copyright: Douglas L. Jones
License: http://creativecommons.org/licenses/by/2.0/
Module: "Ecient Multirate Filter Structures"
By: Douglas L. Jones
URL: http://cnx.org/content/m12800/1.3/
Pages: 11-14
Copyright: Douglas L. Jones
License: http://creativecommons.org/licenses/by/2.0/
Module: "Filter Design for Multirate Systems"
By: Douglas L. Jones
URL: http://cnx.org/content/m12773/1.3/
Pages: 15-18
Copyright: Douglas L. Jones
License: http://creativecommons.org/licenses/by/2.0/
Module: "Multistage Multirate Systems"
By: Douglas L. Jones
URL: http://cnx.org/content/m12803/1.3/
Pages: 19-22
Copyright: Douglas L. Jones
License: http://creativecommons.org/licenses/by/2.0/
Module: "DFT-Based Filterbanks"
By: Douglas L. Jones
URL: http://cnx.org/content/m12771/1.3/
Pages: 23-25
Copyright: Douglas L. Jones
License: http://creativecommons.org/licenses/by/2.0/

Available for free at Connexions <http://cnx.org/content/col10287/1.1>

38 ATTRIBUTIONS

Module: "Quadrature Mirror Filterbanks (QMF)"

By: Douglas L. Jones
URL: http://cnx.org/content/m12770/1.3/
Pages: 27-31
Copyright: Douglas L. Jones
License: http://creativecommons.org/licenses/by/2.0/
Module: "M-Channel Filter Banks"
By: Douglas L. Jones
URL: http://cnx.org/content/m12775/1.3/
Pages: 33-34
Copyright: Douglas L. Jones
License: http://creativecommons.org/licenses/by/2.0/

Available for free at Connexions <http://cnx.org/content/col10287/1.1>

Multirate Signal Processing
An electrical engineering course on multirate signal processing.

About Connexions
Since 1999, Connexions has been pioneering a global system where anyone can create course materials and
make them fully accessible and easily reusable free of charge. We are a Web-based authoring, teaching and
learning environment open to anyone interested in education, including students, teachers, professors and
lifelong learners. We connect ideas and facilitate educational communities.
Connexions's modular, interactive courses are in use worldwide by universities, community colleges, K-12
schools, distance learners, and lifelong learners. Connexions materials are in many languages, including
English, Spanish, Chinese, Japanese, Italian, Vietnamese, French, Portuguese, and Thai. Connexions is part
of an exciting new information distribution system that allows for Print on Demand Books. Connexions
has partnered with innovative on-demand publisher QOOP to accelerate the delivery of printed course
materials and textbooks into classrooms worldwide at lower prices than traditional academic publishers.