13 12-06-20MultirateSigProcPolyphase

Lecture 13, Multirate Signal Processing
Polyphase Representation
Last time we saw how to obtain the polyphase
representation for the filtering and
downsampling operation of 1 filter. We now
extend this formulation for a bank of N filters.
Here we not only have 1 filter, but N filters, and
hence can assemble the N filter outputs or N
subbands into a vector of N elements,
We also now have N filters instead of 1 filter,
and we assemble the polyphase vectors of our
filters into a matrix, in which each column
corresponds to an anlalysis filter,
This matrix now contains all our N impulse

responses of our analysis filter bank, just like in
our block transform case. But unlike our
transform case, here we now have z-tranforms
(polynomials in z) as our matrix entries. This
now enables us to write the filtering and
downsampling operations of our entire analysis

filter bank as a simple multiplication using the
above polyphase matrix!
Y z= X z H z
Mathematically this looks similar to the block
transform case, but with z-transforms!
Observe that this equation now contains all the
samples of our input signal and also of our
subband signals and our impulse responses,
because we use the z-transformed signals. The
z-transform converts an infinite sequence into
just a scalar or an element (its z-transform),
which can be seen as a 1x1 matrix (the ztransform of that sequence is one big
polynomial). This is important because it allows
us to use longer filters than just 1 block, longer
than with the block transforms.
Synthesis Filter Bank
Just as for the block transform case, we can
also get a corresponding formulation for the
synthesis filter bank. For the synthesis filter
bank, we can now also write the reconstructed
sequence x n in terms of blocks m and
phases n, and obtain the synthesis upsampling
and convolution as (see also eq. (8) of lecture
(11))
N 1 L / N 1
x mN n=
k =0 m' =0
y k mm ' g k m ' N n
We can now also use vectors for our sequences
of blocks to simplify this equation, using a

vector for our reconstructed signal and for our
k'th synthesis filter (we start again with looking
at just 1 filter),
Now we can re-write our synthesis equation as
where we now no longer have our phase index

n, because we now have output blocks instead
of samples.
The inner sum is again a convolution, which
turns into a multiplication using the ztransform,
N 1
z = Y k z G k z
X
k =0
Now we can extend this notation to our bank of

N synthesis filters using our subband vector
Y z , and the synthesis polyphase matrix.
Since the output is the sum of all subbands, we
obtain our polyphase matrix by collecting all
our polyphase (row) vectors of our synthesis
filters into a matrix, such that the outer sum of
the above equation turns into a matrix
multiplication,
X z=[Y 0 z ,Y 1 z ,... ,Y N 1 z]G z
where each row of G(z) now contains 1

synthesis filter,
or
with
This is the synthesis polyphase matrix.

Observe that for this polyphase matrix, the
indices for the subbands k and for the phase n
are in reversed order compared to the analysis
polyphase matrix.
Again we turned the mathematically very
complex operation of upsampling and synthesis
filtering into a mathematically very simple
operation with the multiplication of the
subband vector with the polyphase matrix!
Perfect Reconstruction
Perfect reconstruction (PR) is defined as a
reconstructed signal which is identical to the
original signal except for a delay,
x n= x nnd
with some delay n d at the original sampling
rate. This delay usually results from our
filtering and the downsampling and upsampling
operations. To obtain PR we can simply take a
look at the output of our synthesis filter bank,
The structure of the polyphase analysis and
synthesis filter bank can be seen also in the
following structure,
The structure with the delays and

downsamplers on the left of the analysis
polyphase matrix converts a sequence of

samples at the high sampling rate into a
sequence of blocks at the low sampling rate
(blocking). Conversly, the structure to the
right of the synthesis polyphase matrix
converts a sequence of blocks at the low
sampling rate into a sequence of samples at
the high sampling rate (de-blocking).
Here we can see that we obtain perfect
reconstruction if we have
G z= zd H 1 z
(where d is the delay at the downsampled
rate). This is basically again like in our block
transform case. This is now the constraint for
obtaining PR. The question is, how do we obtain
filters for PR? How do we invert a polyphase
matrix, containing the polynomials? How do we
get "good" synthesis filters?
A simple approach is analog to orthogonal
block transform matrices, where the inverse is
simply the transpose matrix. There the analysis
and synthesis filters are identical, except for
the time reversal for the analysis filters.
The corresponding property of polyphase
matrices is called "para-unitary" (see also:
Vaidyanathan: Multirate Systems and Filter
Banks). It is defined as,
H 1 z = H T z1
The advantage (or rather: one of its
advantages) of a polyphase matrix with this
property is, that we don't need to explicitly
compute its inverse, we just need to transpose

it and replace all z's by z1 .
Observe: This is the correspondence to
orthogonal or orthonormal real or complex
valued matrices.
Also observe: if our polynomials in the
polyphase matrix only have zero'th order (only
a constant, no z), then the polyphase matrix is
identical to a transform matrix.

13 12-06-20MultirateSigProcPolyphase

Uploaded by

Copyright:

Available Formats

13 12-06-20MultirateSigProcPolyphase

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

13 12-06-20MultirateSigProcPolyphase

Uploaded by

Copyright:

Available Formats

Lecture 13, Multirate Signal Processing

This matrix now contains all our N impulse

downsampling operations of our entire analysis

We can now also use vectors for our sequences

of blocks to simplify this equation, using a

Now we can re-write our synthesis equation as

where we now no longer have our phase index

Now we can extend this notation to our bank of

X z=[Y 0 z ,Y 1 z ,... ,Y N 1 z]G z

where each row of G(z) now contains 1

This is the synthesis polyphase matrix.

The structure with the delays and

polyphase matrix converts a sequence of

compute its inverse, we just need to transpose

You might also like