Signal Processing

Final Project Report on Multirate Signal Processing

Dhruv Hirpara - 2020102029
Knv Karthikeya - 2020102003
August 7, 2021

Abstract
Our project report revolves around Multirate signal processing. The objective of our project is to
discuss and implement various
methods that reduce the overall complexity of Two channel multirate filter banks.

1 Introduction
Multirate Signal Processing belongs to an area of signal processing which uses more than one sampling
rate to perform operations. A Multirate system contains discrete/analog analysis and synthesis filter
banks accompanied with Up-samplers and Down-samplers. Our main goal from such systems is to obtain
two kinds of outputs one being PR(perfect reconstruction) and the other being PMR(perfect magnitude
reconstruction).

2 Digital Filter Banks

We encounter two types of filter banks in this paper, analysis and synthesis filter banks. The analysis
filter bank is made up of a bunch of filters with their system functions denoted by Hk (z) and all the
filters are arranged parallel to each other.

The synthesis filter banks are made up of a set of filters with their system functions Gk (z). They

1
Signal Processing Report Page 2

are also arranged parallel to each other. The outputs given by the filters are summed up together to form
ˆ
the synthesized signal x(n).
Here the analysis filters are equiripple (Cauer) half band infinite impulse response (IIR) filters. While,
the synthesis filters are finite impulse response filters (FIR).

2.1 Down-sampling
Down-sampling of the given signal is done by taking only the M th sample of the signal and removing all
the M − 1 samples in between where M is the factor with which we down-sample.

Decimation

∞
X
Y (z) = x[n]z −n
−∞
∞
X
Y (z) = x[M n]z −n
−∞

The main problem that arises while down-sampling is that aliasing occurs. We need to prevent aliasing
and we can do it by applying the following points during our implementation. 1. Down sample the signal
π
by a factor of M. 2. Band limit the original signal to M .
H(z) behaves as an anti-aliasing filter. It doesn’t reduce aliasing to 0 but decrease it to a tolerable limit

Original signal vs Downsampled signal

2.2 Up-sampling
Up-sampling the given signal by a factor of M is the process of adding M - 1 zeros between every two
samples.
Signal Processing Report Page 3

Upsampler
∞
X n −n
Y (z) = x[ ]z
−∞
M

Original signal vs Upsampled signal

2.3 Two-channel Digital Filter Bank

As we are working with a Two-channel digital filter bank our sampling rate M = 2.
By looking the proposed filter bank we can relate the z-transforms of y(n) and x(n) of the filter bank by

Y (z) = V0 (z)X(z) + V1 (zX (−z)

H0 (z)G0 (z) + H1 (z)G1 (z)

V0 (z) =
2

H0 (−z)G0 (z) + H1 (−z)G1 (z)

V1 (z) =
2

V0 (z) and V1 (z) functions here are the magnitude distortion and aliasing transfer functions respectively .
To obtain a PR filter bank V0 (z) = cz −d and V1 (z) = 0 for any integer d and non-zero constant c.
Following these guidelines we can see that the output signal is nothing but a scaled and delayed version
of our input signal y(n) = cx(n − d)
To obtain a PMR filter bank we need the magnitude response of the distortion function to be constant
and the aliasing function to be 0. |V0 (ejωT | = c and V1 (z) = 0 We can see that there is no magnitude
distortion but we have phase distortion.
Signal Processing Report Page 4

2.3.1 Perfect reconstruction(PR)

V1 (z) function which is responsible for aliasing should be equal to 0 for perfect reconstruction to happen.
The distortion function V0 (z) should just delay our input signal and make no other changes to it. So we
take the distortion function to be V0 (z) = cz d which ensures X(ω) has a linear phase and the output
signal y(n) = cx(n − d) which is a delayed version of the input signal x(n).

2.4 Perfect Magnitude Response(PMR)

We get a PMR if our magnitude response of the distortion function is constant |V0 (ejωT | = c and our
aliasing function is 0, V1 (z) = 0. Hence the PMR has no magnitude distortion but has phase distortion.

Proposed Filter Bank

3 Implementation
We design the analysis filter with very low-complexity We can also follow the same method for synthesis
filter by interchanging the synthesis and analysis filters as they are related to each other.
Implementing the proposed digital filter banks gives PR filter banks(with magnitude distortion but no
phase distortion). We design the filter bank by first optimizing the digital analysis filters(IIR filters) and
then optimizing the digital synthesis filters(FIR Filters).
We designed the analysis and synthesis filters for 2 cases PR and PMR. Its quite possible to obtain the
analysis filter of very low order and complexity. The overall complexity also is quite low.
For our analysis filters we use equiripple half band IIR filters which are designed using closed-form
solutions. For the synthesis filters we use FIR filters in case Perfect Reconstruction and IIR filters in case
PMR.

3.1 Analysis Filters

Realization of analysis filter bank

We use equiripple (Cauer) half band IIR low pass and high pass IIR filters. They can be written in
polyphase in this way,
A0 (z 2 ) + z −1 A1 (z 2 )
H0 (z) =
2
A0 (z 2 ) − z −1 A1 (z 2 )
H1 (z) =
2
Here, A0 (z 2 ) and A1 (z 2 ) are both taken to be real, causal all-pass filters.
For convenience we use the following expressions,

z −Ki Di (z −1 )
Ai (z) = , i = 0, 1
Di (z)

Ki
X
Di (z) = dik
k=0
Signal Processing Report Page 5

Here Ki denotes the order of Ai (z) and Di (z). Now, the analysis filters H0 (z) and H1 (z) can be written
in the form,
N (z) N (z)
H0 (z) = H1 (z) =
D(z 2 ) D(z 2 )

Here,
N (z) = 0.5[z 2K0 D0 (z −2 D1 (z 2 ) + z −1 z −2K0 D0 (z 2 D1 (z −2 )]

Nc (z) = 0.5[z 2K0 D0 (z −2 D1 (z 2 ) − z −1 z −2K0 D0 (z 2 D1 (z −2 )]

D(z 2 ) = D0 (z 2 )D1 (z 2 )
We can see that, Nc (z) = N (z). It is known that they both are linear-phase FIR filters with symmetric
and anti symmetric impulse responses.

We need to follow certain conditions for efficient implementation. For the Low-pass filter H0 (z),

1 − δs2 ≤ |H0 (ejωT )|2 ≤ 1, ωT ∈ [0, π − ωs T ]

|H0 (ejωT )|2 ≤ δs2 , ωT ∈ [ωs T, π]

For High-pass filter H1 (z)

1 − δs2 ≤ |H0 (ejωT )|2 ≤ 1, ωT ∈ [ωs T, π]

|H0 (ejωT )|2 ≤ δs2 , ωT ∈ [0, π − ωs T ]

3.2 Synthesis Filter Banks

To construct a PMR we can make the synthesis filters using the analysis filters mentioned above. The
synthesis filters are taken to be,

G0 (z) = 2H0 (z) G1 (z) = −2H1 (z)

So now the distortion function and aliasing function would be,

V0 (z) = z −1 A0 (z 2 )A1 (z 2 )

V1 (z) = 0
This solution is very attractive when we can tolerate phase distortion. The complexity would also be
very low as we are only using IIR filters.

Synthesis filter bank

Signal Processing Report Page 6

To obtain a distortion function with exact linear phase, we make the synthesis filter with FIR filters.
Now the transfer functions are,
G0 (z) = 2F0 (z 2 )F1 (z)F2 (z 2 )

G1 (z) = 2F0 (z 2 )F1c (z)F2 (z 2 )

Here, F0 (z) is a non-linear phase FIR filter, while F1 (z), F1c (z), F2 (z) are linear-phase FIR filters.
Role of the following FIR filters, 1. F0 eliminates phase distortion. 2. F1 (z) and F1c (z) eliminates or
suppresses aliasing. 3. F2 (z) tries to eliminate the magnitude distortion. N (z) and Nc (z) are linear
phase FIR filters hence the phase distortion is due to D(z 2 ). So keeping this in mind the filter F0 (z) is
chosen as shown below in order to obtain a distortion function with exact linear phase.

F0 (z) = D(z)

To obtain the distortion and aliasing function

V0 (z) = z −1 D(z 2 )D(z 2 )F2 (z 2 )

V1 (z) = 0

we select F1 (z) and F1c (z) as to follow the criteria of the distortion and aliasing functions.

F1 (z) = N (z) F1c (z) = −Nc (z)

By using these we get a exactly zero aliasing function and distortion function with linear phase, provided
that F2 (z 2 ) is a linear-phase filter. F2 (z 2 ) must be an even-order linear-phase filter with symmetric
impulse response. F2 (z) should satisfy the following conditions.
1.
X M
−cM +1 + cp cos(pωk T )DR (ωk T ) ≤ 1
p=0

M
X
−cM +1 − cp cos(pωk T )DR (ωk T ) ≤ −1
p=0

2.
f2 (n) = f2 (KF 2 − n)

(
c0 n=M
f2 (n) =
0.5cM −n 0≤n≤M −1
3.
V0ˆ(z) = D(z)D(z −1 )F2 (z)
F2 (z) is required as it maintains magnitude distortion . This problem can be solved using passing the
above data through an simplex algorithm
Signal Processing Report Page 7

4 Results
4.1 Analysis Filter bank for both cases PR and PMR filters

Analysis filter bank(elliptical 5 th order half band filter pair)

4.2 Synthesis Filter bank for PMR

PMR Synthesis filter bank

Signal Processing Report Page 8

4.3 Synthesis Filter bank for PR without F2 (z) blocks

PR Synthesis filter bank without F2 (z) block

4.4 Example PMR with input sin(2πt) + sin(0.5πt) + cos(0.1) + rect(0, 0.5)

PMR Example with input sin(2πt) + sin(0.5πt) + cos(0.1) + rect(0, 0.5)

Signal Processing Report Page 9

4.5 Example PR with input sin(2πt) + sin(0.5πt)

PR example with input sin(2πt) + sin(0.5πt) using matlab inbuilt function

5 Conclusions and Observations

Filter design techniques and optimisation methods to obtain a lower complexity for application. The
synthesis are digital FIR filters for linear phase(PR) and they are digital IIR filters for PMR. Analysis
filters are IIR in both cases. They have very low complexity. The proposed digital filter banks are
approximately PR filter banks.
We had implemented both PMR and PR filters and their results are shown above.
First we improve the analysis filters and then move onto the synthesis filters. By reducing the individual
complexity of each filter bank, we finally obtain a Digit Two-channel multirate filter bank with very low
complexity analysis and synthesis filters.

6 References
1. Digital Signal Processing 4th Edition by John Proakis, Dimitris Manolakis (Multirate Signal Process-
ing(Chapter 11)).
2. Multirate Digital Filters, Filter Banks, Polyphase Networks, and Applications: A Tutorial by P.
P. Vaidyanathan.
3. Upsampling(wikipedia) and Downsampling(wikipedia)
4. Multirate Filter Banks(wikipedia)
5. Multirate Digital Signal Processing: Part 4 by Dr. Deepa Kundur.