Fundamentals of Multirate Systems

Shahzad Saleem
ssaleem@student.tugraz.at
Signal Processing and Speech Communication (SPSC) Lab
Graz University of Technology, Austria
Abstract
There has been significant advancement in the field of digital signal
processing over the past several decades. Classical digital signal processing structures are the so-called single-rate systems because the sampling
rates are the same at all points of the system. There are many applications where the signal of a given sampling rate needs to be converted
into an equivalent signal with a different sampling rate. Such systems are
called multirate systems. This paper presents the fundamentals of multirate building blocks and filter banks and describes some applications of
multirate systems.

Introduction

During the last several years, the multirate processing of digital signals has
attracted many researchers. The multirate approach increases the computation speed, decreases the overall filter order, reduces word-length effects, and
decreases power consumption. Consequently, one of the main characteristics of
multirate systems is their high computational efficiency. Multirate digital signal
processing has different applications, such as efficient filtering, subband coding
of speech, audio and video signals, analog/digital conversion, communications
etc.
The two basic operations in multirate digital signal processing are decimation
and interpolation. These operations can be performed by building blocks known
as decimators and expanders.
An M -fold decimator (Figure 1) that takes an input x(n) and produces the
output sequence[1]
yD (n) = x(M n)
(1)
where M is an integer. Decimation results in aliasing unless x(n) is bandlimited
in a certain way. In general, however, it may not be possible to recover x(n)
from yD (n) because of loss of information.
An L-fold expander (Figure 2) takes an input x(n) and produces an output

Figure 1: M-fold decimator

sequence [1]
1

yE (n) =

x(n/L) if n is integer multiple of L

0
otherwise

(2)

Figure 2: L-fold expander

1.1
1.1.1

Transform Domain Analysis of Decimators and Expanders

Expander

We have [1]
YE (z) =
=

n=

yE (n)z n =

yE (n)z n

n=mul. of L

yE (kL)z kL =

k=

x(k)z kL

(3)

k=

= X(z ).
So Y (ej ) = X(ejL ). This means that Y (ej ) is an L-fold compressed version
of X(ejL ). Hence expander creats an imaging effect.
1.1.2

Decimator

For the M -fold decimator (1), we can write output YD (ej ) in terms of X(ej )
as [1]
M1
1 X
j
YD (e ) =
X(ej(2k)/M )
(4)
M
k=0

This can be graphically interpreted as follows: (a) stretch X(ej ) by a factor

M to obtain X(ej/M ), (b) create M 1 copies of this stretched version by
shifting it uniformly in successive amount of 2, and (c) add all these shifted
stretched versions to the unshifted stretched version X(ej/M ) and divide by
M . The aliasing created by decimation can be avoided if x(n) is a lowpass signal
bandlimited to the region || < /M .

1.2

Decimation and Interpolation Filters

In most applications, the decimator is preceded by a lowpass digital filter called

the decimation filter. This filter ensures that the signal being decimated is
bandlimited. The exact bandedges of the filter depend on how much aliasing is
permitted. For example, in QMF banks a certain degree of aliasing is usually
permitted because this can eventually be canceled off.
An interpolation filter is a digital filter that follows an expander.The typical
purpose is to suppress all the images. Typically interpolation filter is lowpass
2

with cutoff frequency /L.

Both types of filters with their typical responses are shown in Figure 3 and 4.

Figure 3: (a) The complete decimation circuit, and (b)typical response of the
decimation filter

Figure 4: (a) The complete interpolation circuit, and (b)typical response of the
interpolation filter

1.3

Fractional Sampling Rate Alteration

Uptill now we have seen that sampling rate of a signal can be altered by an
integer factor. In certain applications, it is required to change the sampling
rate by a rational fraction. An example of such a system is shown in Figure. 5.
Here L = 2, M = 3, so M/L = 1.5. Here X(ej ) is bandlimited to || 2/3.
If signal is decimated by two, then that would create aliasing error. But it is
possible to decimate by factor 1.5 (as shown in the Figure. 5 by broken lines)

1.4

Digital Filter banks

A digital filter bank is a collection of digital filters having common input or

common output. There are two types of filter banks known as analysis bank
and sysntesis bank. An analysis bank together with analysis filters Hk (z) splits
a signal x(n) into M subband signals xk (n), while the task of a synthesis bank is

to combine the M subband signals into a single signal x (n) using the synthesis
filters Fk (z). Both types of filter banks have been shown in Figure. 6.

Figure 5: Fractional decimation

Figure 6: (a)Anaysis Bank (b) Synthesis Bank

Interconnection of Building Blocks

Figure. 7 shows the interconnection of commonly used building blocks in multirate systems.

Figure 7: Simple Identities for interconnected systems

2.1

Noble Identities

Noble Identities are very useful in the theory and implementation of multirate
systems(c.f Figure. 8).

Figure 8: Noble Identities for multirate systems

The Polyphase Representation

The polyphase representation simplifies the theoratical results and leads to

more efficient implementations of decimation/interpolation filters as well as filter banks. Consider a filter

X
H(z) =
h(n)z n
n=

By separating the even numbered coefficients of h(n) from odd numbered ones,
we get [1]

X
X
H(z) =
h(2n)z 2n +
h(2n + 1)z 2n
(5)
n=

n=

Defining
E0 (z) =

h(2n)z n , E1 (z) =

n=

h(2n + 1)z n

(6)

n=

So H(z) can be written as

H(z) = Eo (z 2 ) + z 1 E1 (z 2 )

(7)

These representations are valid for whether H(z) is an FIR or IIR; causal or
noncausal.
Now suppose that we are given any integer M , so H(z) can be decomposed as
[1]

H(z) =

h(nM )z nM

n=

X
1

+z

h(nM + 1)z nM
(8)

n=

..
.
+z

M1

h(nM + M 1)z nM

n=

this can be written as

H(z) =

M1
X

z l El (z M ) (Type 1 Polyphase)

l=0

(9)

where
El (z) =

el (n)z n

n=

(10)

with
el (n) , h(M n + l), 0 l M 1
A variation of (8) is given by [1]
H(z) =

M1
X

z (M1l) Rl (z M ) (Type 2 Polyphase)

(11)

l=0

Type 2 polyphase components Rl (z) are permutations of El (z), i.e. Rl (z) =

EM1l (z)

3.1
3.1.1

Efficient Structures for Decimation and Interpolation

Filters
Decimation Filters

Consider the decimation filter with M = 2. If H(z) is represented as in (7), then

systems can be redrawn as shown in Figure 9(a). By invoking noble identity
1, this can be redrawn as shown in Figure 9(b). This implementation is more
efficient than a direct implementation of H(z) (Figure 9(c)). As shown, the
direct implementation computes only even numbered output samples, which
requires N + 1 multiplications and N additions. As time changes from 2n
to 2n + 1, the stored signals in the delays change. This means that above
computations must be completed in one unit of time. The speed of operation
should therefore correspond to N + 1 multiplications and N additions per unit
time. During the odd instants of time, the structure is merely resting. This is
inefficient resource utilization.
Now if we consider the polyphase representation given in Figure 9(b), let n0 and
n1 be the orders of E0 (z) and E1 (z). So El (z) requires nl +1 multiplications and
nl additions. The total cost is again N + 1 multipliers and N adders. However,
since El (z) operates at lower rate, only a total of (N + 1)/2 multiplications
per unit time (MPUs) and N /2 additions per unit time (APUs) are required.
The multipliers and adders in each of the filters E0 (z) and E1 (z) now have two
units of time available for their work, and they are continually operative (i.e.,
no resting time).
3.1.2

Inerpolation Filters

Consider an interpolation filter with L = 2. A direct form implementation of

H(z) is again inefficient because, at most 50% of the input samples to H(z)
are nonzero, which means that at any point in time, only 50% of multipliers
of h(n) have nonzero input. So the remaining multipliers are resting. And
those multipliers which are not resting are expected to complete their job in
half unit of time because the outputs of the delay elements will change by that
time. A more efficient structure can be obtained by using Type 2 polyphase
decomposition.
H(z) = R1 (z 2 ) + z 1 R0 (z 2 )
(12)
6

Figure 9: (a) Polyphase representation of decimation filter (b)Using noble identity 1 (c) direct form implementation
This is shown in Figure 10. It can be seen now that Rl (z) are opearting at the
input rate, and none of the multipliers are resting. Each multiplier gets one
unit of time to finish its task. The complexity of the system is N + 1 MPUs and
N 1 APUs. The extra adder following the expander is not counted because,
the signal y(n) is obtained merely by interlacing y0 (n) and y1 (n).

Figure 10: Polyphase representation of an interpolation filter

4
4.1

Applications of Multirate Systems

Subband Coding of Speech and Image Signals

Often, we cencounter signals that are not bandlimited, but still have dominant frequency bands. An example is shown in Figure 11(a).The information in
|| /2 is not small enough to be discarded. And x(n) cannot be decimated
without causing aliasing either. This is because of the small fraction of energy
in the high frequency region that prevent to compress the signal.
But if the signal is splitted into two frequency bands by using an analysis bank
with responses as shown in Figure : 11(b), then the subband signal x1 (n) has
less energy than x0 (n) and so can be encoded with less number of bits.
The reconstruction of the full band signal is done using the expanders and
synthesis filters as shown in Figure 12.
So a generalization of this idea can be as: split the signal into two or more
subbabnds, decimate each subband signal, and allocate bits for samples in each
subband depending on th energy content. In speech coding practice the number
of subbands, filter bandwidths and bit allocations are chosen to further exploit
7

Figure 11: Splitting a signal into subbabnd signals x0 (n) and x1 (n)

Figure 12: The analysis/synthesis system for subband coding.

the perceptual properties of human hearing. the complete analysis synthesis system (Figure 12) is called the Quadrature Mirror Filter (QMF) bank.
For image processing, multirate filters have different applications such as multiresolution systems and the Laplace pyramid.

4.2

Transmultiplexers

A complete transmultiplexer is shown in Figure 13. The components xk (n)

of the TDM version can be recovered by separating the consecutive regions of
Y (ej ) with the help of analysis bank and then decimating the outputs. If
the synthesis filters Fk (z) and analysis filters Hk (z) are non ideal, then the
reconstructed signals x
has contributions from the desired signal and xk (n) as
well as the cross talk terms xl (n), l 6= k. To reduce the extent of cross talk,
Hk (z) and Fk (z) can be designed to have non overlapping frequency responses.
To reduce the cross talk to an acceptable value requires filters of very high order.

Figure 13: The complete transmultiplexer structure

Summary

In this paper, a very brief introduction of Multirate Systems and Filter Banks
has been presented. An introduction to the technique Polyphase Representation and how it can be used to make multirate systems computationally
efficient has also been shown with examples. At the end, some applications of
the multirate systems have been discussed.

References
[1] Vaidyanathan, P.P.: Multirate Systems and Filter Banks, 2nd Ed, Prentice Hall , 1993.
[2] Vaidyanathan, P.P.: Multirate Digital Filters, Filter Banks, Polyphase
Networks, and Applications: A Tutorial, Proceedings of the IEEE, Vol. 78,
No. 1, January 1990.