TUTORIAL ON POLYPHASE TRANSFORMS

fred harris
CUBIC Signal Processing Chair
Communication Systems and Signal Processing Institute
San Diego State University, San Diego, California 92182
e-mail: fred.harris@sdsu.edu

1. INTRODUCTION
A polyphase transform is an algorithm for simultaneously (and efficiently) performing the functions required to synthesize a multichannel receiver for digitized
data. These functions entail the spectral translation of each center frequency to
baseband, the bandwidth reduction of the translated spectrum to match the signal bandwidth, and resampling of the output series to match the reduced channel
bandwidth. This process, called channelization or de-multiplexing, is equivalent
to synthesizing a bank of narrowband filters, which span the spectral range of the
input signal. When this process is performed by the polyphase transform, the
center frequencies of the filter bank are equally spaced but otherwise arbitrary
and the spectral shapes of the filters are identical but otherwise arbitrary.
The direct approach to this task is the digital simulation of the bank of conventional homodyne (single) conversion receivers. The three-step process of frequency translation, bandwidth reduction, and resampling is most easily visualized
as separate and sequential operations preformed in the listed order. In the direct
implementation of the homodyne filter bank, the workload required for a single
channel is repeated for each channel. Suppose the workload per channel is N1
(operations/output) and the channel resampling ratio is M (inputs/output) then,
the parameter describing the filter complexity, the per channel workload is the
ratio N1/M (operations/input). To form N2 identical channels the total workload
for the channel bank is N2*N1/M (operations/input).
In the polyphase transform implementation of the channel bank the order of the
three required operations (frequency translation, bandwidth reduction, and resampling) are reversed. At first glance this seems to violate the Nyquist criterion
that in order to avoid aliasing the bandwidth reduction must occur prior to the resampling. For a time invariant linear system this is true. This is not true for the
polyphase transform. In polyphase processing, the resampling of the input is accomplished by partitioning the filter into subfilters (called polyphase segments) to
which the input is sequentially commutated. The sequential commutation performs the resampling and also converts the filter to a "periodically time varying
filter". This description reflects the fact that the impulse response of this partitioned filter depends upon the commutator position when the impulse arrives at
the filter. As strange as it might seem, the aliasing that occurs at the input to the
periodically time varying filter is exactly cancelled at the output of the filter. The

unaliased bandwidth reduction occurs when the aliased terms are cancelled
during the summation of the outputs from the polyphase segments. By performing a phase-shifted summation over the polyphase outputs different aliasing
terms can be selected to survive the summation. This phase shifted summation
performs the heterodyning or frequency selection of the filter bank and is performed by an FFT.
In the polyphase transform a single prototype filter is shared by the entire filter
bank, and the spectral positioning of the filter bank is determined by an FFT after
the resampling and filtering. Note the computational efficiency of this structure.
The workload for the filter is N1/M (operations/input) and the workload for the
transform is (N2*log (N2) ) or log (N2) (operations/output). Since for each input
data block (of length M) the filter bank computes an output block (or length N2)
we can refer the workload to the input data rate to obtain a total workload for the
entire filter bank of [log (N2 ) + N1 /M] (operations/input) as opposed to N2*N1 /M.
For instance, for a filter length N1 = 256, resampling rate M=16, and number of
output channels N2 = 32, the direct implementation would requi4re 512 (operation/input) while the polyphase transform implementation would require only 21
(operations/input).

2. BASEBAND POLYPHASE FILTERS

The tapped delay line model of a standard finite impulse response (FIR) filter with
M - to -1 resampling is shown in figure 1.
x(n)

x(n-1)

x(n-2)

h(0)

h(1)

h(2)

........

x(n-N+1)

x(n-N)

h(N-1)

h(N)

....x(n+1)

........

M:1

......
y(n),....

M:1
h(n)
x(n)

y(n),....

y(nM),....

Figure 1. TAPPED DELAY LINE MODEL OF FIR FILTER

y(nM),....

Data is delivered to the filter at the input sample rate fs . The filter reduces the
bandwidth by a factor of M hence the filter's output sample rate is reduced by the
same factor to fs/M. A filter normally produces one output for each M inputs.
Since this filter is non-recursive, the resampling is accomplished by computing a
filter output after the arrival of M new inputs. Thus each data point in the filter will
shift M positions between computations of an output point. If we follow a single
data point through the filter we observe that it intersects or engages only a subset of the filter coefficients, namely those that are separated by M points. This is
indicated in figure 2.
If the first coefficient that an input data point engages is h(r) as it shifts through
the filter, it will also engage coefficients h (r+2M, h (r+3M), , h (r+ [L - 1] M). In
general, a given data point will engage those coefficients with a fixed index r1 for
coefficient indices of the form shown in equation 1.

N = r1 + M r2

r1 = 0, 1,2, M 1
r2 = 0, 1, 2, L - 1

(1)

We now partition the filter set into sub-filters called polyphase segments with
each segment identified by the r1 index. This is indicated in equation 2.

r1

{h0 (r2 ) } = h (0 + M r2),

(2)
{h1 (r2 ) } = h (1 + M r2),
{h2 (r2 ) } = h (2 + M r2),
r2 = 0, 1, 2, , L - 1
:
{hM - 1 (r2) } = h (M -1 + M r2)

The polyphase segments are accessed by delivering the data to their inputs via
an input commutator, which starts at the r1 index M-1 and decrements to index 0.
After the commutator has executed one cycle and has delivered M input samples
to the filter, a single output is taken as the summation of the outputs from the
polyphase segments. This is shown in figure 3.

....x(n+1)

h(r)

h(r+M)

h(r+(L-1)M)

h(r+2M)

........
......
yr(nM)

h (n) = h(r+nM)
r
x(n),....

yr(nM)

FIGURE 2. M - POINT SHIFT FOR A RESAMPLING FIR FILTER

N
M
n

h(0+ nM)

y(nM)

n-1

h(1+ nM)
n-2

h(2+ nM)
x(n)
n-3

h(3+ nM)

n-(M-2)

h(M-2+nM)
n-(M-1)

h(M-1+nM)

FIGURE 3. POLYPHASE PARTITION OF M-TO-1 RESAMPLED FIR FILTER

Note that commutating the input data to the M polyphase segments in figure 3 is
identical to delivering M new samples to the filter shown in figure 1 hence the fi lter outputs must be also be identical.

3. CARRIER - CENTERED POLYPHASE FILTERS

We now address the carrier-centered version of the polyphase filter. We invoke
the modulation theorem to convert a prototype baseband filter to its equivalent
carrier centered, or spectrally shifted version. The modulation theorem is presented succinctly in equation 3. Here the finite sequence h (n) and the (periodic)
spectrum H() are a Fourier series pair. To paraphrase the theorem, multiplying
a series h(n) by a complex heterodyne of the form exp (j n) shifts the origin of
the spectrum H() by radians/sample
if
h (n)
< - - - - > H ( )
then

h (n) e j n

(3)

< - - - - > H ( - )

Thus given the impulse sresponse h (n) of the prototype low pass filter, we form
the impulse response for the carrier centered version by simply multiplying the
coefficients h (n) with the complex terms exp (j n) to obtain new coefficients g(n)
of the form shown in equation 4.
g (n) = h (n) exp (j n)

(4)

Precisely which center frequencies are desirable will be addressed shortly.

We now perform a polyphase partition of this impulse response as was done in
equation 2. We obtain the partition for the r1-th filter indicated in equation 5.
{gr1 (r2) } = h (r1 + M r2) e j ( r1 + M r2)
{gr1 (r2) } = h (r1 + M r2) e

j r 1

e j M r2

(5a)
(5b)

We now select the digital frequency to which we translate the prototype filter.
We select so that a single period of the series exp (j n) is harmonically related to M, the depth of the commutated polyphase filter partition; this is indicated
in equation 6.
= k (2/M)

(6)

Substituting the frequency defined in equation 6 into equation 5, we obtain the

polyphase partition shown in equation 7.

{gr1 (r2) } = h (r1 + M r2 ) e j k (2/M) r1 e j k (2/M) M r2

(7a)

{gr1 (r2 )} = h (r1 + M r2 ) e j k (2/M) r1

(7b)

Note that the complex exponential multiplying the resampled prototype coefficient
only has one index, the r1 index which identifies the polyphase segment and that
the r2 index which identifies positions within the segment have been eliminated
by the appropriate selection of the sinusoid's period. The remarkable results
indicated in equation 7 is that the resampling of the complex exponential has
resulted in a complex exponential term which is a constant for each polyphase
segment. This constant can be applied after the polyphase segment summation
is performed. This is shown in figure 4 where the first index associated with the
putput indicates time and the second index indicates frequency.

N
M

j0k

h0(n)
e

y(nM,k)

j1 k

h1(n)
e
h2(n)
x(n)

k= 2M

j2k

j3k

h3(n)

e
hM-2(n)
e

j(M-2) k

j(M-1) k

hM-1(n)
FIGURE 4. M-TO-1 RESAMPLED, CARRIER CENTERED POLYPHASE FILTER

4. POLYPHASE TRANSFORM FILTER BANK

The conplex weights which follow the polyphase partial summations shown in
equation 7 and indicated in fiigure 4 are samples of a complex sinusoid with an
interger number of cycles in the span of the commutator. This summation must
be performed over the values of the index k (which indicates cycles per interval
of length M). In fact, this weighted summation is easily seen to be a DFT of the
polyphase outputs. The DFT can of course be performed with an FFT provided
the length M is hightly composite (ie; is the product of many small integers ). The
implementation of the full filter bank as a polyphase transform is shown in figure
5.
N
M

h0 (n)
h1 (n)
h2 (n)
x(n)
h3 (n)

y0 (nM)

y(nM,0)

y1 (nM)

y(nM,1)

y2 (nM)

y(nM,2)

y3 (nM)

M
Point
DFT

y(nM,3)

(FFT)

hM-2 (n)
hM-1 (n)

yM-2 (nM)

y(nM,M-2)

yM-1 (nM)

y(nM,M-1)

FIGURE 5. M-TO-1 RESAMPLED, M-OUTPUT POLYPHASE TRANSFORM

In the sequence of modifications through which we have observed the FIR filter
evolve into the polyphase transform we have unnecessarily coupled the spectral
spacing of the filters with the time domain resampling ratio. This was intentionally
done to simplify the sequence of transformations. The parameter relationships
which determine the characteristics of the polyphase transform are the following.
The spacing between center frequencies of the filter bank is determined by the
length N2 of the FFT. The spectral spacing is fs/N2. The spectral shape of the
filters is determined by the original prototype low pass filter of length N1. In the
earlier presentation, the filter length was N and the transform length was M. But

M was also the resampling ratio. It is not necessary to have the resampling ratio
match the lenghtof the transform. The resampling ratio can be larger than, equal
to, or smaller than the transform length. When they were equal, the commutator
delivered data to all N2 (formally M) polyphase sub filters before the transform
was exercised. For any other resampling ratio say M, which is less than N2, the
cummutation operation becomes slightly more complicated. We can best
examine the additional complexity by way of illustration. For instance, we assume
a filter of lneegth 32 a factor of 3-to-1. The indices of the input data in the
polyphase partition of the linerar fitler are indicated in Table 1. For three
successive output condition. The bold entries in the table indicate the latest three
data points.
h0:
h1:
h2:
h3:
h4:
h5.
h6:
h7:

31
30
29
28
27
26
25
24

23
22
21
20
19
18
17
16

15
14
13
12
11
10
9
8

FIRST OUTPUT

7
6
5
4
3
2
1
0

34
33
32
31
30
29
28
27

26
25
24
23
22
21
20
19

18
17
16
15
14
13
12
11

10
9
8
7
6
5
4
3

SECOND OUTPUT

37
36
35
34
33
32
31
30

29
28
27
26
25
24
23
22

21
20
19
18
17
16
15
11

13
12
11
10
9
8
7
6

THIRD OUTPUT

TABLE 1. INPUT INDICES OF POLYPHASE FILTER OF LENGTH 32 WITH 8

OUTPUT SEGMENTS AND RESAMPLED 3-TO-1.
Note that the indices in segments 0 through 4 for the first output condition have
been pushed down three rows and are located in segments 3 through 7 for the
second output condition. Also note that the indices for segments 5 through 7 for
the first output condition have rolled around to segments 0 through 3 and have
shifted one position to the right to make room for the next three data points. This
is equivalent to a circular shift of the columns of three places followed by a one
position linear shift of the top three rows to accept new data. This is shown explicitly in table 2. For a firm understanding of the modified commutation it is worth
following the steps indicated in table 2.

h0:
h1:
h2:
h3:
h4:
h5:
h6:
h7:

31
30
29
28
27
26
25
24

23
22
21
20
19
18
17
16

15
14
13
12
11
10
9
8

7
6
5
4
3
2
1
0

INITIAL CONDITION

34 26
33 25
32 24
31
30
29
28
27

18
17
16
23
22
21
20
19

10
9
8
15
14
13
12
11

2
1
0
7
6
5
4
3

34
33
32
31
30
29
28
27

BARREL ROLL

26
25
24
23
22
21
20
19

18
17
16
15
14
13
12
11

10
9
8
7
6
5
4
3

INSERT NEW DATA

TABLE 2. SUCCESSIVE POSITION OF INPUT INDICES BETWEEN OUTPUTS

The sequence of operations consisting of a circular shift of columns followed by
linear shift of rows for data insertion can be reversed. The circular shift is actually accomplished via memory addressing. This reversed form is the one demonstrated in the polyphase transform presented in figure 6. Here the number of
filters is 8 and the resampling ratio is 3-to-1.

Initial
Data

Circular
Shift

Polyphase
Filter

d 0 (n-1)

d 0 (n)

h 0 (n)

y0 (3n)

y1 (3n)

d 1 (n-1)

d 1 (n)

h 1 (n)

d 2 (n-1)

d 2 (n)

h 2 (n)

d 3 (n-1)

d 3 (n)

h 3 (n)

y0 (3n)

y0 (3n)

FFT
d 4 (n-1)

d 4 (n)

h 4 (n)

d 5 (n-1)

d 5 (n)

h 5 (n)

d 6 (n-1)

d 6 (n)

h 6 (n)

d 7 (n-1)

d 7 (n)

h 7 (n)

y0 (3n)

x(n)
y0 (3n)

y0 (3n)

y0 (3n)

FIGURE 6. POLYPHASE TRANSFORM WTH 8 FILTERS OF EQUIV ALENT

LENGTH 32, RESAMPLED 3 - TO - 1

One final note is that there is no need for the filter length to be an integer multiple
of the transform length. If they were so related, each polyphase segment would
be of the same length. If they are not so related then there are two lengths for
the polyphase segments and the two differ by one point.

5. BIBLIOGRAPHY
Comparing Traditional FFT Based Frequency Domain Excision with Polyphase
Transform Excision
(with Hana Abusalem)
Institute of Navigation Conference (ION-99), 28-June, 1-July 1999, Boston, MA
On the Relationship between Multirate Polyphase FIR Filters and Windowed, Overlapped, FFT Processing,
23rd Annual Asilomar Conference on Signals, Systems, and Computers,
Pacific Grove, CA., 30 October to 1 November 1989.
Convolution, Correlation, and Narrowband Filtering with the DFT,
IEEE, ESIME Conference Semena de la Ingenieria en Communicaciones Electricas; Mexico City, July 1980.