Multirate Polyphase Filters and Filter Banks

Multirate
Polyphase Filters and Filter Banks

fred harris
w w w . d s p o n l i n e c o n f e r e n c e . c o m
AGENDA
Alias Cancellation and

1 Humor and Perspective 4
M Roots of Unity
Motivation for Multirate DSP Cascade Analysis and Synthesis

2 5
Polyphase M-Path Filters
M-Path Polyphase Partition Synthesizing Super Channel

3 and M-to-1 Down Sampling 6 Broad Bandwidth Cascade
Analysis and Synthesis Filters
© DSPOnlineConference.com All rights reserved 2 D S P O n lin e C o n fe r e n c e .c o m

Humor and Perspective

What The Customer Wants
MORE
M ORE
RE MORE
MORE MORE MORE MO MORE
MORE
MORE MORE MORE
MORE MORE
M O RE
MORE MORE
MORE
MORE
MORE The Second
MORE
R E
M E M More machine
O
MORE R
MORE O
MORE
MORE
MORE
MORE
MORE
MORE
M
MORE
MORE
MORE O
MORE MORE
R MORE
E
MORE
MORE
MORE
MORE
MORE
MORE
MORE
MORE
Compliments of
Don Steinbrecher
I drew this copy from a Xerox he sent me
I Later Found the original in New Yorker Magazine, (circa 1990)
What the Customer Will Pay
MORE
MORE
RE MORE
MORE
M OR E M OR E M OR E
MORE M OR E
M ORE
MORE MORE
MORE
SL
E
MORE
M OR E
S L SS M E M OR E
R E
ESS M
O
MORE R
MORE O
MORE
S EVE
MORE
MORE
N MORE
MORE
LES MORE
M
S MORE
MORE
O
MORE
MORE
R M OR E M OR E
E
MORE
MORE
MORE
MORE
M OR E
M OR E
M OR E
M OR E

When the Customer wants it.
MORE
MORE
RE MORE
MORE
M OR E M OR E M OR E
MORE M OR E
M OR E
NE MORE MORE
MORE
XT
TO WE MORE
R E
M OR E
MO EK M E M OR E
M
O
MORE
RRO MORE
MORE O R
TH W MORE
AFT IS MORE
MORE
ERN MORE
MORE
M
OO
N MORE
MORE
MORE
O
MORE
R M OR E M OR E
E
MORE
MORE
MORE
MORE
M OR E
M OR E
M OR E
M OR E

The Size Customer Wants.
MORE
M ORE
RE MORE
MORE
MORE MORE MORE
MORE MORE
MORE
MORE MORE
MORE
MORE
MORE
MORE
R E
M E M
O
MORE
MORE
MORE
MORE O R
MORE
MORE
RE MORE
MORE MORE MORE MORE MO MORE
MORE MORE
MORE MORE MORE MORE
MORE
M
MORE
MORE
MORE O MORE
MORE
R
MORE
MORE MORE
MORE
E
MORE MORE
MORE
MORE
MORE
5
MORE
MORE
R E
MORE
MORE E MORE
M M
O
MORE R
MORE MORE O MORE
MORE MORE
MORE
MORE
MORE MORE
MORE
MORE MORE MORE MORE
MORE
M
MORE
MORE
MORE O MORE
MORE MORE
R MORE MORE
E MORE
MORE MORE
MORE
MORE MORE
MORE MORE
R E
MORE M E M
O
MORE O R
MORE
MORE
MORE
M
O
R MORE MORE
E
MORE
MORE
MORE
MORE

Why do we Perform
Multirate Digital Signal Processing?
• Improved Performance
• Reduced Cost
• Reduced Time to Market
• Perform Operations not Previously Possible
• Personally Rewarding
• It is fun to be creative
• If you are not having fun,
you are probably not doing it right!

An Important Tool and Resource
Accessed in the DSP Domain
• Sample Rate
 By embedding Sample Rate Changes in a filter we can form
Linear Time Varying (LTV) or Periodically Time Varying (PTV) filters.
 In an LTV or PTV filter, we can apply an input signal at one frequency
and the response can occur at another frequency.
 In the sampled data world we call this Aliasing (traveling under a different
name).
 Aliasing is your friend!
• As an Aside
 Most of us have limited experience with LTV systems.
 Actually we all share one common LTV system,
a swing!
 Every child (and a few adults)
on the playground masters
the two modes of swing excitation.

Mode I. Pumping A Swing From a Sitting Position
Rotate Body Mass in

Positive Direction Additive
when Growth Per Cycle
Angular Velocity
is Zero and
Angular
Acceleration
is Positive
Rotate Body Mass
in
Negative
Direction
when Angular
Velocity is Zero
and Angular
Acceleration
is Negative
“How to Pump a Swing”, Stephen Wirkus, Richard Rand, and Andy Ruina,
The College Mathematics Journal, Vol. 29, No. 4, September 1998

Mode II. Pumping A Swing From Standing Position
Lower Body Mass when

Angular Velocity is Zero Multiplicative
Growth Per Cycle
Raise Body Mass when

Angular Velocity is Max
Lower Body Mass when

Angular Velocity is Zero
“How to Pump a Swing”, Stephen Wirkus, Richard Rand, and Andy Ruina,
The College Mathematics Journal, Vol. 29, No. 4, September 1998
My wife commented that there is a third way to swing on a swing and
that if you are a little girl it is the best way: have grandpa push you!

Motivation and Introduction to Multirate DSP

Processing Task:
Obtain Digital Samples of Complex Envelope
Residing at Frequency f C
Analog
Digital
Rec eiver
Multi-Channel Single-Channel
FDM Base banded
Input Signal Output Signal
When I Presented this Selec ted

Input Spectrum Narrow Band
slide at BYU, the attendees
Signal
Applauded! I asked the
reason for the applause.
f
fC
See!
Uta h

First Generation DSP Receiver
1
f
2
f
3
f
4
f
All Signal Conditioning e -jω0t
Performed in Analog Domain: Low Pass
Last Operation ADC! 1
Filter
2 3 fs 4

DSP Insertion in Communication Systems
Instinctive First Response:

Copy Legacy Analog Prototype
 We should avoid this approach!!
 If we don’t, we emulate the analog design!
 That is not the reason we invoke and apply DSP!
 DSP is inserted to improve performance and reduce cost!
 Analog prototype systems incorporate design compromises
appropriate for the time they were made!
 We don’t want to perpetuate those compromises!
 We have access to tools and resources not available to past designers!

Signal Conditioning for DSP Receiver
1
f Input Spectrum
f Bandwidth Limiting Filter
.... .... Sampled Data Spectrum

f
-fs/2 fs/2
Low Pass
Filter
1 2 fs 3 4

Duplicate Analog Processing in DSP
3
.... ....
f Input Sampled
-fs/2 4 fs/2 data spectrum
.... ....
Move Desired
f Spectrum
-fs/2 5 fs/2 To baseband
.... ....
Limit Bandwidth
f With DSP Filter
~
~
-fs -fs/2 6 fs/2 fs
.... ....
f Reduce Sample Rate
-fs/M fs/M Proportional to
e -jθ0n
fs/M
Reduced Bandwidth
Low Pass Low Pass
Filter Filter
Ignoring 1 2 fs 3 5 6
4 M:1
Good Advice
Fundamental Operation
Select Frequency,
Limit Bandwidth,
Select Sample Rate
s(t) s(n) -jθ0n

s(n) e r(n) r(nM)
LOWPASS
ADC
FILTER
h(n) M:1
-jθ0n
e
CLK LO
The Obvious way!

Signal and Filter are at Different Frequencies
Which One to Move??
CHANNEL OF INTEREST
f
0
FILTER RESPONSE
f
0
TRANSLATED SPECTRUM
Second Option
Armstrong’s Heterodyne
f
0
TRANSLATED FILTER
First Option
f Tuned RF
0

Tuned RF VRS Super Heterodyne
Filter Output
DETECTOR
RF AMP RF AMP RF AMP AUDIO
1 2 3 AMP
THREE STAGE TRF RECEIVER
Filter Output
IF IF
DETECTOR
AMP AMP
RF AMP
AUDIO
& Narrow Narrow AMP
MIXER Band Band
Pass Pass
SUPERHETERODYNE RECEIVER
OSCILLATOR
LOCAL

Down Sample Complex Digital IF
3
.... ....
f Input Sampled
-fs/2 4 fs/2 Data Spectrum
.... ....
Move Filter to Spectrum
f of Interest and Reduce BW
~
-fs -fs/2 5 fs/2
.... .... Reduce Sample Rate and

f Alias Spectra to Lower Frequency
~
-fs/M 6 fs/M fs
.... ....
Heterodyne Aliased Spectrum
f to Baseband at Reduced
-fs/M fs/M
e jθ 0n Output Sample Rate
e -jMθ 0n
Low Pass Ba nd Pa ss fs/M
Filter Filter
1 2 fs 3 5 6
4 M:1

Fundamental Operation Modified
Equivalency Theorem
s(t) s(n) jθ0n r(n) r(nM)

r(n) e
BANDPASS
ADC
FILTER
M:1
jθ0n -jθ0n
h(n) e e
CLK LO

Equivalency Theorem
− jθ0 n
r ( n ) = s ( n )e * h( k )
= ∑ s ( n − k )e
k
− jθ0 ( n − k )
h( k )
= e − jθ0 n
∑ s ( n − k ) h ( k )e
k
jθ0 k
− jθ0 n jθ0 n
=e {s (n) * h(n)e }

Signal Flow Description
of Equivalency Theorem
Heterodyne moved in each Path from In-Line Multiply of
Signal Samples to Off-Line Multiply of Filter Coefficients
c os(nθk) c os(nθk)
x(n) h(n) x(n) h(n) c os(nθk)

y(n) y(n)
h(n) h(n) sin(n θk)
-sin(nθk) -sin(nθk)
This is No Bargain Unless

There is Another Step

Reorder Heterodyne and Resample
s(t) s(n) jθ0n jθ0Mn

r(n) e r(nM) e
r(nM)
BANDPASS
ADC
FILTER
M:1
jθ0n -jθ0Mn
h(n) e e
CLK LO
This is that next Step!

SPECTRAL DESCRIPTION
REORDERED FUNDAMENTAL OPERATION
INPUT ANALOG FILTER RESPONSE CHANNEL OF INTEREST
f
-fs/2 0 fs/2
FILTERED SPECTRUM TRANSLATED FILTER
f
-fs/2 0 fs/2
ALIASED REPLICATES AT DOWN-SAMPLED RATE
.... ....
f
-fs/M 0 fs/M
TRANSLATED REPLICATES ATDOWN-SAMPLED RATE
.... ....
f
-fs/M 0 fs/M

Successive Transformations Turn Sampled Data
Version of Edwin Armstrong’s Heterodyne Receiver
to Tuned Radio Frequency (TRF) Receiver
to Aliased TRF Receiver.
-j θ n Digital
e k Low-Pass Digital -j θ kn
Band-Pass e
H(Z) -jθ k
H(Ze )
M-to-1
M-to-1
Armstrong Nyquist Equivalency Theorem
Digital
Digital -j Mθkn Band-Pass
Band-Pass e
-jθ k
-jθ k H(Ze )
H(Ze )
M-to-1
M-to-1
M ⋅ θ k = k ⋅ 2π
Common Sense Any Multiple of
2π Output Sample Rate
or θ k =k ⋅ Aliases to Baseband
M
Polyphase Partitioned Filter
Down Sample at Input to Filter,
Alias Spectrum at Input De-Alias at Output
Don’t Down sample After BW Reduction

An Interesting Property of a Filter
Applying an operator to the output time series of

a low-pass filter is same as applying the operator
to the impulse response of that filter: the
modified filter will, in turn, apply the operator to
the input time series

h(n)
d
dn
.
y(n)
x(n) y(n)
Low Pass Derivative
Filter Operator
d
x(n) dn
h(n) .
y(t)
Low Pass
Derivative
Filter

Rather than Resample the output of a
filter, we can resample the impulse
response of the filter and the modified
filter will resample the time series!
h(n) y(n) y(nM)
x(n)
Low Pass
Filter
M:1
h(nM)
x(n) y(nM)
M:1
Re-sam pled
Low Pass Filter

Coefficient Assignment of Polyphase Partition
For M-to-1 resample start at Index r and Increment by M
For 3-to-1 resample start at index r and increment by 3
C0 C1 C2 C3 C4 C 5 C 6 C 7 C 8 C 9 C 10 C 11
C0 0 0 C3 0 0 C6 0 0 C9 0 0
0 C1 0 0 C4 0 0 C7 0 0 C 10 0
0 0 C2 0 0 C5 0 0 C8 0 0 C 11
C0 0 0 C3 0 0 C6 0 0 C9
0 C1 0 0 C4 0 0 C7 0 0 C 10
0 0 C2 0 0 C5 0 0 C8 0 0 C 11
This mapping from 1-D to 2-D is used by Cooley-Tukey FFT.

Extract Delays To First
Non-Zero Coefficient
Polyphase Filters and CT-FFT are kissing cousins!

Polyphase Partition of Low Pass Filter
1-Path to M-Path Transformation
x(n) y(n) M:1 y(nM)
H( Z)
N −1
H ( Z ) = ∑ h( n) Z − n x(n) M
H0( Z )
y(n) M:1 y(nM)
n =0
M −1 N −1 Z
-1 M
H1( Z )
H (Z )
= ∑∑
=
h
r 0=
(
n 0
r + nM ) Z − ( r + nM )
-2 M
Z H2( Z )
M −1 N −1
=H (Z ) ∑ ∑
Z −r
h ( r + nM ) Z − nM
....
....
....
=r 0=n 0
-(M-2) M
Z HM-2( Z )
M-Path Partition Supports M-to-1 Down Sample
Also Supports Rational Ratio M-to-Q Down Sample!
-(M-1) M
Z HM-1( Z )

Polyphase Partition of Band Pass Filter
1-Path to M-Path Transformation
Modulation Theorem of Z-Transform:
Spin Delays rather than Spin Data Samples
N −1 N −1
=G(Z )
=n 0=n 0
∑
= h( n) e Z ∑=
jθ k n
h ( n ) (e Z )
−n − jθ k −n
H (e − jθ k
Z)
M −1 N −1
G(Z )
= ∑ ∑ h(r + nM ) e
=r 0=
n 0
jθ k ( r + nM )
Z − ( r + nM )
M −1 N −1
=G(Z )
M ⋅ θ k = k ⋅ 2π = r 0= n 0
∑ e jθ k
Z −r
∑ h ( r + nM ) e jθ k nM − nM
Z
2π M −1 j 2π k N −1
or θ k =k ⋅
= M G(Z ) e M Z
=
−
r 0=
r
n 0
∑ h(r + nM ) Z − nM ∑
Select Center Frequencies to be Multiples of Output Sample Rate:
Rotators in Filter Default to Rotators Attached to Delay Lines!
Polyphase M-Path Narrowband, Resampling Filter

Noble Identity: Commute M-units of Delay followed
by M-to-1 Down Sample
M Delays M Delays
Input Cloc k, T
1 Delay 1 Delay
Output Clock, MT
M-Units of Delay at Input Rate Same as 1-Unit of Delay at Output Rate

M:1 M:1
-M -1
Z Z
M:1 M:1
-M -1
H(Z ) H(Z )

Apply Noble Identity to Polyphase Partition
M:1 j 2π
M
k0
e y(nM,k)
x(n)
H0( Z )
j 2π k1
M:1 e M
Z
-1 H1( Z )
M:1 j 2π
M
k2
e Pull Resampler through
-2 H2( Z )
Z Polyphase Filter Paths.
Filter Paths Operate at
Reduced Output sample rate.
....
....
....
M:1 j 2π
M
k(M-2)
e
-(M-2)
Z HM-2 ( Z )
M:1 j 2π
M
k(M-1)
e
-(M-1)
Z HM-1( Z )

Move Scaler Phase Spinners to
Output of Polyphase Filter Paths
j 2π k0
M:1 e M
y(nM,k)
x(n)
H0( Z )
j 2π k1
M:1 e M
Z
-1 H1( Z )
Input Delays and Synchronous
M-to-1 Down Sample Switches j 2π k2
M:1 e M Don’t make Input Data Samples
Are equivalent to Input Commutator. -2 H2( Z ) Complex on the way into the
Z
Path Filters.
No need to hold Earlier Samples in Make Output Data Samples
....
....
....
Memory Waiting till all M Samples Complex on the way out of ths
j 2π k(M-2)
are Available. We can Deliver M:1 e M
Path filters.
Successive Input Samples to -(M-2)
Z HM-2( Z )
Successive Input Ports as they Arrive. j 2π k(M-1)
M:1 e M
-(M-1)
Z HM-1( Z )
Want phase spinners as far away from resampler as possible

Resampled M-Path Narrowband Filter
Hmm... this is very

good stuff....

Phase Rotated Sum
M −1 2π
+j rk
y(nM , k) = ∑ yr (nM) e M
Phase Rotator
r =0 On r-th Row,
Selecting Alias
From k-th
Sum Over Nyquist Zone
nM-th Output nM-th Output
Time Sample Row Output Time Sample
From K-th Time Samples From r-th Row
Nyquist Zone
Recognize Sum as Scaled IFFT of Row Vector of Time Samples at Time nM

M-Channel Channelizer
with Rotators Replaced by M-Point IFFT
Polyp hase
Pa rtition
h0(n)
h1(n)
fs h2(n)
h3(n) M-PNT
IFFT
FDM
.....
.....
.....
.....
TDM
hM-2(n)
hM-1(n)
hr(n)= h(r+ nM)

Armstrong to Tuned RF with Alias
Down Conversion to Polyphase Receiver
Digital
-j θ n Digital
e k Low-Pass
Band-Pass
-jθ k
H(Z) H(Ze )
M-to-1
M-to-1
M-Path Digital -j 2π rk
Polyphase e M Reducing Sample Rate at Input
Causes M-fold aliasing of Input
H(Z)
r
Spectrum. Phase Coherent Sum
Extracts Selected Alias while
M-to-1 Canceling Remaining Aliases by
Destructive Cancellation.
Rather than selecting center frequency at input and reduce sample rate at
output, we reverse the order, reduce sample rate at input and select
center frequency at output. We perform arithmetic operations at low
output rate rather than at high input rate!
Single Channel Armstrong and
Multirate Aliased Polyphase Receiver
j nk 2π Standard DDC
e M
y(n,k) y(nM,k) Two Filters in Armstrong Receiver

x(n)
H(Z ) Because Signal made Complex at
1 2 2 2 Filter Input Port.
2-Polypha se M-to-1
Filters
Polyphase DDC j nk 2π One Polyphase Filter in Polyphase
x(n) x(n) y (nM) e M
Receiver Because Signal made
r
H r (Z ) y(nM,k) Complex at Filter Output Ports.
1 M-to-1 M M 2
1-Polypha se
Filter

Dual Channel Armstrong and
Multirate Aliased Polyphase Receivers
j nk1 2π Standard DDC
e M
x(n) y(n,k1) y(nM,k1)

H(Z )
1 2 2 2
2-Polypha se M-to-1
j nk2 2π
e M Filters Second Channel Requires
y(n,k2) y(nM,k2) Second Receiver.
H(Z ) Total of 4 Filters
1 2 2 2
2-Polypha se M-to-1
Filters
Polyphase DDC j nk1
2π
x(n) x(n) y (nM) e M
r
H r (Z ) y(nM,k1)
M
1 M-to-1 M 1-Polypha se 2 Second Channel Requires
Filter j nk2 2π Second set of Phase Rotators.
e M
Total of 1 Polyphase Filter
y(nM,k2)
M 2

Workload for Multiple M-Path Filters
•1-Channel M-to-1 Down Sample
• 1-Filter and M Complex Phase Rotators
•2-Channels M-to-1 Down Sample
• 1-Filter and 2M Complex Phase Rotators
•K-Channels M-to-1 Down Sample
• 1-Filter and kM Complex Phase Rotators
•M-channels M-to-1 Down Sample (use FFT)
• 1-Filter and [log2(M)/2]M Complex Phase Rotators
When k > Log2(M)/2 Build all channels and discard the channels you don’t need!
IF M=16, Log2(16)/2 = 2: thus if you want 2 or more, Build them all!
IF M=128, Log2(128)/2 = 3.5: thus if you want 4 or more, Build them all!
IF M=1024, Log2(1024)/2 = 5: thus if you want 5 or more, Build them all!
Suppose you want 10 Channels: Build all 1024 and discard 1014:
This is half the cost of building 10

M-Channel Channelizer
with Rotators Replaced by M-Point IFFT
Polyp hase
Pa rtition
h0(n)
h1(n)
fs h2(n)
h3(n) M-PNT
IFFT
FDM
.....
.....
.....
.....
TDM
hM-2(n)
hM-1(n)
hr(n)= h(r+ nM)

Phase Profiles of M-Path Filters
Alias Separation by
Destructive Cancellation
Of Roots of Unity

Impulse Response and Frequency Response of
Prototype Low Pass FIR Filter
Impulse Response, Prototype Lowpass Filter
1
0.8
Bandwidth is 1/6th
Amplitude 1
0.6
of Sample Rate
0.4 fs 6-Samples Between
0.2 1
BW Successive Zero
0
Crossings in Time Domain
-0.2
0 2 4 6 8 10 12 14 16 18
Normalized Time (1/BW)
Spectrum
0
0.2
Zoom to Passband
Log Mag (dB)
-20
0.1
-40 0 Filter Bandwidth

-0.1 1/6th of Sample Rate
-60
-0.2
-0.4 -0.2 0 0.2 0.4 0.6
-80
-100
-3 -2 -1 0 1 2 3
Normalized Frequency

Impulse Response of 6-Path Polyphase Partition
Prior to 6-to-1 Resampling (by Discarding Zeros)
By Not Discarding
The Interstitial Zeros,
We have not Changed
the Sample Rate, there
Is no Aliasing, and
the Sampled Data
Sequence Still has a
Transfer Function

Frequency Response of Zero Packed 6-Path Polyphase
Partition Prior to 6-to-1 Resampling: Can See 6-Nyquist Zones
The Zero Packed

Samples In M-Path
Polyphase Filter
have Spectral Copies in
M-Nyquist Zones.
We can Extract
the Spectra Residing
in any of the
M-Nyquist Zones

Phase Response of 6-Path Polyphase Partition
Prior to 6-to-1 Resampling
There are Dog-Leg

Phase Transitions
at the Boundaries
Between Adjacent
Nyquist Zones.
Can be Seen
Better at Larger
Scale Image

Overlay Phase Response of 6-Path Polyphase Partition
-3 Prior to 6-to-1 Resampling
Nyquist
Zone
-2
Nyquist
Zone
-1
Nyquist
-2π/3 Zone
Phase 0
Shifts -2π/6 Nyquist
Phase Zone +1
Shifts Nyquist
Phase Zone
Aligned +2
Nyquist +3
2π/6 Zone Nyquist
Phase Zone
Shifts 2π/3
Phase
Shifts

De-Trend Causality Tilt of Overlay Phase Responses:
6-Path Partition Prior to 6-to-1 Resampling

3-D Paddle-Wheel Phase Profiles,

Overlay 3-D Paddle-Wheel Phase Profiles,

Showing Phase Shifts in +1 Nyquist Zone

Phase Shifted to Align Phases in +1 Nyquist Zone

Tone in Nyquist Zone 0, Phase Shifts Identical, Phasors Coherently Sum to Maximum
Channel Spectral Centers and Shifting Tone Frequency Response
Log Mag (dB)

0.5
0
-15 -10 -5 0 5 10 15
Frequency
In-Band Tone Time Response, Path (0) In-Band Tone Time Response, Path (1) In-Band Tone Time Response, Path (2)
1 1 1
0 0 0
Amplitude
Amplitude
Amplitude
-1 -1 -1
0 20 40 60 80 100 0 20 40 60 80 100 0 20 40 60 80 100
In-Band Tone Time Response, Path (3)

Time Index In-Band Tone Time Response, Path (4)
Time Index
1 1 1
0 0 0
Amplitude
Amplitude
Amplitude
-1 -1 -1
0 20 40 60 80 100 0 20 40 60 80 100 0 20 40 60 80 100
In-Band Tone Response Sum

Time of 6-Paths
Index Frequency Response, Sum
Time of 6-Paths
Index Nyquist Plot, Each ofIndex
Time 6-Paths
1 0
1
-50
Michelle_6_path_3
Log Mag (dB)
0 0
Amplitude
Real
-100 -1
-1
0 20 40 60 80 100 -3 -2 -1 0 1 2 3 -1 0 1
Time Index Frequency (kHz) Imaginary

Tone in Nyquist Zone +1, Phase Shifts Differ by 60o, 6-Roots of Unity, Phasors Destructively Sum to Zero
Channel Spectral Centers and Shifting Tone Frequency Response
Log Mag (dB)

0.5
0
-15 -10 -5 0 5 10 15
Frequency
In-Band Tone Time Response, Path (0) In-Band Tone Time Response, Path (1) In-Band Tone Time Response, Path (2)
1 1 1
0 0 0
Amplitude
Amplitude
Amplitude
-1 -1 -1
0 20 40 60 80 100 0 20 40 60 80 100 0 20 40 60 80 100
In-Band Tone Time Response, Path (3)

Time Index
1 1 1
0 0 0
Amplitude
Amplitude
Amplitude
-1 -1 -1
0 20 40 60 80 100 0 20 40 60 80 100 0 20 40 60 80 100
In-Band Tone Response Sum

Time of 6-Paths
Index Frequency Response, Sum
Time of 6-Paths
Index Nyquist Plot, Each ofIndex
Time 6-Paths
1 0
1
-50
Log Mag (dB)
0 0
Amplitude
Real
-100 -1
-1
0 20 40 60 80 100 -3 -2 -1 0 1 2 3 -1 0 1
Time Index Frequency (kHz) Imaginary

Cascade Analysis and Synthesis
Polyphase M-Path Filters

Down Sample M-to-1 Up Sample M-to-1
H0(Z) H0(Z)
H1(Z) H1(Z)
x(n,k1)
y(n,0)
H2(Z) H2(Z)
y(nM,0)
.
.
...
...
H(M-1)(Z) H(M-1)(Z)
f f f
0 0 0

Down Sample M-to-1 Up Sample M-to-1
Alias From Nyquist Zone k1 Alias to Nyquist Zone k1
j0k12π j 2π
e M e 0k1 M
H0(Z) H0(Z)
j 2π j 2π
e 1k1 M e 1k1 M
H1(Z) H1(Z)
x(n,k 1) 2π 2π
j j
e 2k1 M e 2k1 M y(n,k1)
H2(Z) H2(Z)
y(nM,k 1)
j(M-1)2k1 2π j (M-1)2k1 2π
.
.. ..
...
e M e M
H(M-1)(Z) H(M-1)(Z)
f f f
k1 k1

Up Sample M-to-1
Alias to Nyquist Zone k2
j0k12π j 2π
e M e 0k2 M
H0(Z) H0(Z)
j 2π j 2π
e 1k1 M e 1k2 M
H1(Z) H1(Z)
x(n,k 1) 2π 2π
j j
e 2k1 M e 2k2 M y(n,k2)
H2(Z) H2(Z)
y(nM,k 1)
2π j(M-1)k2 2π
.. ..
.
j
e (M-1)k1 M
...
e M
H(M-1)(Z) H(M-1)(Z)
f f f
k1 k2

j0k22π j0k2
2π
e M
j0k12π j0k12π
e M
e M e M
y(nM,k 2)
H0(Z) H0(Z)
H1(Z) H1(Z)
x(n,k 1)
H2(Z) H2(Z)
y(n,k 1)+y(n,k 2)
.. ..
.. ..
H(M-1)(Z) y(nM,k 1) H(M-1) (Z)
f
f f f
k2 k1 k2 k1

Alias Multiple Narrow Band Signals Alias Multiple Narrow Band
from Different Center Frequencies Baseband Signals to Different Center
Down to Baseband By Down Sampling Frequencies By Up Sampling
y(nM,0) y(n,0) x(nM+ 0)

H0 ( Z )
H0 ( Z )
y(nM,1) y(n,1) x(nM+ 1)

H1 ( Z ) H1 ( Z )
x(n) y(nM,2) y(n,2) x(nM+ 2)

H2 ( Z ) M-Point H2 ( Z ) x(n)
M-Point
IFFT
....
IFFT
.... y(n,k)
....
....
y(nM,k)
....
....
y(n,M-2) x(nM+ M-2)
y(nM,M-2)
HM-2( Z ) HM-2( Z )
y(n,M-1) x(nM+ M-1)

y(nM,M-1)
HM-1( Z ) HM-1( Z )

Circular Buffer
fs
flg=0
M
Path 2 fs/M
Poly M-PNT
Phase
.. ...
IFFT
Filter
flg=1
flg=
in Z 2
1
flg= 0
0.5T
mT T
0
(m+1)T T
0.5T 1.5T
0.5T T 1.5T 2T
Phase Continuity

0 0
1 1
M-Path Input Data Buffer
M-Path Polyphase Filter
Circular Output Buffer

2 2
FDM
3
....
M-Point IFFT
M/2-1 M M M
........
M/2
M/2+ 1
TDM
….
M-2
M-1 M-1
State Eng ine

0 0
1 1
FDM
M-Path Input Data Buffer

M-Path Polyphase Filter
Circular Output Buffer
2 2
3
....
TDM
M-Point IFFT
M M/2-1
M M
........ M/2
M/2+ 1
….
M-2
M-1 M-1
State Eng ine

nM/2 x nM/2
.. .. .. ..
.. .. .. ..
x
Point Point
M Shift M M Shift M
Path Point Point Path
..
..
..
..
..
..
Filter M-Point IFFT M-Point Filter
IFFT
Circ ula r Circ ula r
Buffer x Buffer
x
Looks Like Fast Convolution. It Isn’t!

Fast Convolution Occurs in Frequency Domain,
Alters Spectrum and Returns to Time Domain.
Channelizer Never Leaves Time Domain!

Change Bandwidth with Binary Mask,
Change Sample Rate with Transform Size
x
.. .. .. ..
.. .. .. ..
x
M M-Point M M M-Point M
Path Point Point Circ ular Path
..
..
Circ ular
..
..
Filter Buffer IFFT x FFT Buffer Filter
x
x
x
.. .. .. ..
M M-Point M P P-Point P
Path Circ ular Point Point Path
..
.. Circ ular
Filter Buffer IFFT Buffer
FFT Filter
..
Spectrum, Composite Complex Baseband Signal Sampled 480 MHz
0
Log Mag (dB)

-20
-40
-60 Channelizer Bandwidth

-80
-200 -150 -100 -50 0 50 100 150 200
Normalized Frequency
0
-20
-24-200 -23-200 -22-200 -21-200 -20-200 -19-200 -18-200 -17
-40 -40 -40 -40 -40 -40 -40 -40
-60 -60 -60 -60 -60 -60 -60 -60
-80 -80 -80 -80 -80 -80 -80 -80
-10 0 10 -10 0 10 -10 0 10 -10 0 10 -10 0 10 -10 0 10 -10 0 10 -10 0 10
0
-20
-16-200 -15-200 -14-200 -13-200 -12-200 -11-200 -10-200 -9
-40 -40 -40 -40 -40 -40 -40 -40
-60 -60 -60 -60 -60 -60 -60 -60
-80 -80 -80 -80 -80 -80 -80 -80
-10 0 10 -10 0 10 -10 0 10 -10 0 10 -10 0 10 -10 0 10 -10 0 10 -10 0 10
0
-20
-8 -200 -7 -200 -6 -200 -5 -200 -4 -200 -3 -200 -2 -200 -1
-40 -40 -40 -40 -40 -40 -40 -40
-60 -60 -60 -60 -60 -60 -60 -60
-80 -80 -80 -80 -80 -80 -80 -80
-10 0 10 -10 0 10 -10 0 10 -10 0 10 -10 0 10 -10 0 10 -10 0 10 -10 0 10
0
-20
0 -200 1 -200 2 -200 3 -200 4 -200 5 -200 6 -200 7
-40 -40 -40 -40 -40 -40 -40 -40
-60 -60 -60 -60 -60 -60 -60 -60
-80 -80 -80 -80 -80 -80 -80 -80
-10 0 10 -10 0 10 -10 0 10 -10 0 10 -10 0 10 -10 0 10 -10 0 10 -10 0 10
0
-20
8 -200 9 -200 10 -200 11 -200 12 -200 13 -200 14 -200 15
-40 -40 -40 -40 -40 -40 -40 -40
-60 -60 -60 -60 -60 -60 -60 -60
-80 -80 -80 -80 -80 -80 -80 -80
-10 0 10 -10 0 10 -10 0 10 -10 0 10 -10 0 10 -10 0 10 -10 0 10 -10 0 10
0
-20
16 -200 17 -200 18 -200 19 -200 20 -200 21 -200 22 -200 23
-40 -40 -40 -40 -40 -40 -40 -40
-60 -60 -60 -60 -60 -60 -60 -60
-80 -80 -80 -80 -80 -80 -80 -80
-10 0 10 -10 0 10 -10 0 10 -10 0 10 -10 0 10 -10 0 10 -10 0 10 -10 0 10

Synthesis
Arbitrary Matc hed
Channelizer
Interpolator Filter
fSYM
10-Path
1-to-5 44.7 MHz
7
100 89.4 89.4
MHz MHz MHz
48-Path
Ana lysis fSYM
4-Path
Channelizer 1-to-5
3 10.7 MHz
40 21.4 21.4
48-Path Input
MHz MHz MHz
Multiple M-Path
Analysis Channelizer Sa m ple Rate
16-Path
fSYM Synthesis Channelizers
480 MHz
1-to-5 74.0 MHz
11
160 148.0 148.0
24-to-1 Down Sam ple MHz MHz MHz
24-to-1
fSYM
Output 8-Path
1-to-5 22.1 MHz
48 Baseba nd 5
80 44.2 44.2
10 MHz BW MHz MHz MHz
Channels
Sam pled at fSYM
20 MHz 32-Path
1-to-5 147.5 MHz
21
320 295.0 295.0
MHz MHz MHz

7-Aliased to Baseband Analysis Channels
0 bin 2 0 bin 3 0 bin 4 0 bin 5 0 bin 6 0 bin 7 0 bin 8
Log Mag (dB)

-20 -20 -20 -20 -20 -20 -20
-40 -40 -40 -40 -40 -40 -40

-60 -60 -60 -60 -60 -60
-60
-80 -80 -80 -80 -80 -80
-80
-10 0 10 -10 0 10 -10 0 10 -10 10 -10 0 10 -10 0 10
0 10 -10 0
Synthesized Spectrum of 44.7 MHz Signal at f C = 100 MHz
0
Log Mag (dB)
-20
-40
Synthesized Baseband Super Channel
-60
-80
from 7 Baseband Analysis Channels
-50 -40 -30 -20 -10 0 10 20 30 40 50
Frequency
Synthesized Spectrum of 44.7.1 MHz Signal at f C = 89.4 MHz
0
Log Mag (dB)
-20
-40
Interpolated Baseband Super Channel
-60
-80
from Synthesized Baseband Super Channel
-40 -30 -20 -10 0 10 20 30 40
Frequency

0 bin 23 0 bin 24 0 bin 25 0 bin 26 0 bin 27
Log Mag (dB)

-20 -20 -20 -20 -20
-40 -40 -40 -40 -40
-60 -60 -60 -60 -60
-80 -80 -80 -80 -80
-10 -5 0 5 10 -10 -5 0 5 10 -10 -5 0 5 10 -10 -5 0 5 10 -10 -5 0 5 10
Synthesized Spectrum of 22.1 MZ Signal at f C =80 MHz

0
Log Mag (dB)
-20
-40
-60
-80
-40 -30 -20 -10 0 10 20 30 40
Frequency
Synthesized Spectrum of 22.1 MZ Signal at f C =88.4 MHz
0
Log Mag (dB)
-20
-40
-60
-80
-40 -30 -20 -10 0 10 20 30 40
Frequency

Modulator and Demodulator, No Spectral Partition and Reassembly
Constellation Modulator Output Constellation Channelizer Output Constellation Matched Filter Output
1.5 3 1.5
1 2 1
0.5 1 0.5
0 0 0
-0.5 -1 -0.5
-1 -2 -1
-1.5 -3 -1.5
-1.5 -1 -0.5 0 0.5 1 1.5 -3 -2 -1 0 1 2 3 -1.5 -1 -0.5 0 0.5 1 1.5
Constellation Modulator Output Constellation Channelizer Output Constellation Matched Filter Output
1.5 1.5 1.5
1 1 1
0.5 0.5 0.5
0 0 0
-0.5 -0.5 -0.5
-1 -1 -1
-1.5 -1.5 -1.5

-1.5 -1 -0.5 0 0.5 1 1.5 -1.5 -1 -0.5 0 0.5 1 1.5
-1.5 -1 -0.5 0 0.5 1 1.5
Modulator and Demodulator, With Spectral Partition and Reassembly

Synthesizing Super Channel Broad Bandwidth
Cascade Analysis and Synthesis Filters

Change of Emphasis!
Rather than Extract Multiple Narrow Bandwidth Channels,
We now Extract One Very Wide Bandwidth Super Channel!
But Wait, If the Channel is Wide Bandwidth, We Don’t Have a

Large Ratio of Sample Rate to Bandwidth and we Can’t Reduce
the Sample Rate to Access the Efficiency of the Cascade
Down-Sampling Up-Sampling Process

Compare Traditional Implementation to
Polyphase Filter Implementation
Input Ba ndwidth
f
Traditional 600 Tap s

Input Filter 600 Multip lies
f
Channelizer Filter
600 Tap s
f
f
Channelized Ba ndwidth
f
Disa bled Channels Enab led Channels Disa bled Channels
Outp ut Filter 8.5% of Workloa d

51 Multiplies 91.5% Reduc tion
f

Frequency Domain Filtering With Cascade
M/2-to-1 Analysis and 1-to-M/2 Synthesis Channelizers
nM/2 x nM/2
.. .. .. ..
.. .. .. ..
x
Point Point
M Shift M M Shift M
Path Point Point Path
..
..
..
..
..
..
Filter M-Point IFFT M-Point Filter
IFFT
Circ ula r Circ ula r
Buffer x Buffer
x
600 Tap Prototype Filter 60 Point Good-Thomas Nested Winograd FFT

60 Path Polyphase Partition 200 Real Multiples
10 Coefficients per Path Amortized over 30 Input Samples:
Input Filter: 6.7 Operations per Input Sample
20 Operations Per Input Sample
For Input and Output Polyphase Filter and IFFT:
53.3 Operations per Input-Output Sample Pair
Workload: 8.9% of 600 Coefficient Tapped Delay Line Filter With same Frequency Response!

Spectrum: 600-Tap, 60 Channel Analysis and Synthesis Filter, 29-Enabled Bins, 51 Real Multiplies per Input Sample and 600 Tap Filter With Same Spectrum
0
Log Magnitude (dB)
-20
Implement 600 Tap Filter
-40
with 60 Path, 600 Tap Cascade Red 600 Tap filter Response
Blue Cascade 600 Tap 60-Path Filter Response
-60 Analysis and Synthesis Filter
-80
-100
0 5 10 15 20 25 30
Frequency (in Channel Number)
Spectrum: Detail of Transition Bandwidth x 10
-3 Spectrum: Detail of In Band Ripple and Band Edge
0 2
Log Magnitude (dB)
Log Magnitude (dB)

-20 1
Pass Band
-40
0-14.35 MHz 0
-60
Stop Band -1
-80
14.77-30.0 MHz
-100 -2
14 14.1 14.2 14.3 14.4 14.5 14.6 14.7 14.8 14.9 15 13 13.5 14 14.5 15
Frequency (in Channel Number) Frequency (in Channel Number)

Is Open For Questions

UCSD Fa cult y a nd Co nsult a nt in DSP
THE SPEAKER Po lyp ha se Filt e r Ba nks, Physica l La ye r
Mo d e m De sig n, Synchro niza t io n, a nd
fre d ha rris, fjha [email protected] du Sp e ct ra l Est im a t io n.
Le cture s a nd De sig ns of DSP Ba se d Ra d ios,
SDR, Ra da r, Sona r, Instrum e nts, Spe ctra l
Ana lysis.
I a m a t UCSD whe re I te a ch a nd cond uct re se a rch on DSP a nd
com m unica tion syste m s. I hold som e 40 p a te nts on d ig ita l
re ce ive r a nd DSP te chnolog y. I le cture world wid e on DSP
a p p lica tions a nd consult for org a niza tions on hig h p e rform a nce ,
cost e ffe ctive DSP solutions. I ha ve writte n ove r 260 journa l a nd
confe re nce p a p e rs. My m ost we ll known p a p e r is “On the Use
of Wind ows for Ha rm onic Ana lysis with Discre te Fourie r
Tra nsform ”. I a m a uthor of the te xtb ook “Multira te Sig na l
Proce ssing for Com m unica tion Syste m s”. The se cond e d ition will
b e a va ila b le e a rly 2021. I hold a BEE, MSEE, a nd PhD from
Brooklyn Polyte ch, SDSU, a nd Aa lb org unive rsity re sp e ctive ly. A
DSP cha ir wa s re ce ntly fund e d in m y na m e a t SDSU b y Pe g g y
a nd Eric J ohnson. The sp e lling of m y na m e with a ll lowe r ca se
le tte rs is a source of d istre ss for typ ists a nd sp e ll che cke rs. A
child a t he a rt, I colle ct toy tra ins, old slid e -rule s, a nd g yroscop e s.

Thanks to Stephane Boucher
and
Thank you for Attending and for your Participation
w w w . d s p o n l i n e c o n f e r e n c e . c o m

Multirate Polyphase Filters and Filter Banks

Uploaded by

Copyright:

Available Formats

Multirate Polyphase Filters and Filter Banks

Uploaded by

Document Information

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

Multirate Polyphase Filters and Filter Banks

Uploaded by

Copyright:

Available Formats

Multirate

Polyphase Filters and Filter Banks

Alias Cancellation and

Motivation for Multirate DSP Cascade Analysis and Synthesis

M-Path Polyphase Partition Synthesizing Super Channel

© DSPOnlineConference.com All rights reserved 2 D S P O n lin e C o n fe r e n c e .c o m

© DSPOnlineConference.com All rights reserved 3 D S P O n lin e C o n fe r e n c e .c o m

© DSPOnlineConference.com All rights reserved 5 D S P O n lin e C o n fe r e n c e .c o m

© DSPOnlineConference.com All rights reserved 6 D S P O n lin e C o n fe r e n c e .c o m

© DSPOnlineConference.com All rights reserved 7 D S P O n lin e C o n fe r e n c e .c o m

© DSPOnlineConference.com All rights reserved 8 D S P O n lin e C o n fe r e n c e .c o m

© DSPOnlineConference.com All rights reserved 9 D S P O n lin e C o n fe r e n c e .c o m

Rotate Body Mass in

© DSPOnlineConference.com All rights reserved 10 D S P O n lin e C o n fe r e n c e .c o m

Lower Body Mass when

Raise Body Mass when

Lower Body Mass when

© DSPOnlineConference.com All rights reserved 12 D S P O n lin e C o n fe r e n c e .c o m

© DSPOnlineConference.com All rights reserved 13 D S P O n lin e C o n fe r e n c e .c o m

When I Presented this Selec ted

© DSPOnlineConference.com All rights reserved 15 D S P O n lin e C o n fe r e n c e .c o m

© DSPOnlineConference.com All rights reserved 16 D S P O n lin e C o n fe r e n c e .c o m

Instinctive First Response:

© DSPOnlineConference.com All rights reserved 17 D S P O n lin e C o n fe r e n c e .c o m

f Bandwidth Limiting Filter

.... .... Sampled Data Spectrum

© DSPOnlineConference.com All rights reserved 18 D S P O n lin e C o n fe r e n c e .c o m

-fs -fs/2 6 fs/2 fs

s(t) s(n) -jθ0n

© DSPOnlineConference.com All rights reserved 20 D S P O n lin e C o n fe r e n c e .c o m

© DSPOnlineConference.com All rights reserved 21 D S P O n lin e C o n fe r e n c e .c o m

THREE STAGE TRF RECEIVER

© DSPOnlineConference.com All rights reserved 22 D S P O n lin e C o n fe r e n c e .c o m

-fs -fs/2 5 fs/2

.... .... Reduce Sample Rate and

© DSPOnlineConference.com All rights reserved 23 D S P O n lin e C o n fe r e n c e .c o m

s(t) s(n) jθ0n r(n) r(nM)

© DSPOnlineConference.com All rights reserved 24 D S P O n lin e C o n fe r e n c e .c o m

© DSPOnlineConference.com All rights reserved 25 D S P O n lin e C o n fe r e n c e .c o m

x(n) h(n) x(n) h(n) c os(nθk)

This is No Bargain Unless

© DSPOnlineConference.com All rights reserved 26 D S P O n lin e C o n fe r e n c e .c o m

s(t) s(n) jθ0n jθ0Mn

This is that next Step!

© DSPOnlineConference.com All rights reserved 27 D S P O n lin e C o n fe r e n c e .c o m

© DSPOnlineConference.com All rights reserved 28 D S P O n lin e C o n fe r e n c e .c o m

© DSPOnlineConference.com All rights reserved 30 D S P O n lin e C o n fe r e n c e .c o m

Applying an operator to the output time series of

© DSPOnlineConference.com All rights reserved 31 D S P O n lin e C o n fe r e n c e .c o m

© DSPOnlineConference.com All rights reserved 32 D S P O n lin e C o n fe r e n c e .c o m

© DSPOnlineConference.com All rights reserved 33 D S P O n lin e C o n fe r e n c e .c o m

This mapping from 1-D to 2-D is used by Cooley-Tukey FFT.

© DSPOnlineConference.com All rights reserved 34 D S P O n lin e C o n fe r e n c e .c o m

© DSPOnlineConference.com All rights reserved 35 D S P O n lin e C o n fe r e n c e .c o m

© DSPOnlineConference.com All rights reserved 37 D S P O n lin e C o n fe r e n c e .c o m

M-Units of Delay at Input Rate Same as 1-Unit of Delay at Output Rate

© DSPOnlineConference.com All rights reserved 38 D S P O n lin e C o n fe r e n c e .c o m

© DSPOnlineConference.com All rights reserved 39 D S P O n lin e C o n fe r e n c e .c o m

Want phase spinners as far away from resampler as possible

© DSPOnlineConference.com All rights reserved 40 D S P O n lin e C o n fe r e n c e .c o m