MULTIRATE SIGNAL

PROCESSING
 Multirate Signal Processing
• Definition: Signal processing which uses more than
one sampling rate to perform operations
• Upsampling increases the sampling rate
• Downsampling reduces the sampling rate
• Reference: Digital Signal Processing, DeFatta, Lucas,
and Hodgkiss

B. Baas, EEC 281 431

 Multirate Signal Processing
• Advantages of lower sample rates
– May require less processing
– Likely to reduce power dissipation, P = C V2 f,
where f is frequently directly proportional to the sample rate
– Likely to require less storage
• Advantages of higher sample rates
– May simplify computation
– May simplify surrounding analog and RF circuitry
• Remember that information up to a frequency f
requires a sampling rate of at least 2f. This is the
Nyquist sampling rate.
– Or we can equivalently say the Nyquist sampling rate is ½
the sampling frequency, fs
B. Baas, EEC 281 432
Upsampling
 Upsampling or Interpolation

• For an upsampling by a factor of I, add I‐1 zeros between samples in

the original sequence
• An upsampling by a factor I is commonly written I
For example, upsampling by two: 2
• Obviously the number of samples will approximately double after 2
• Note that if the sampling
frequency doubles after an
upsampling by two, that t
the original sample
sequence will occur at the
same points in time

B. Baas, EEC 281 434

 Original Signal Spectrum

• Example
signal with
most energy
near DC
• Notice 5
spectral
“bumps”
between
large signal
“bumps”

B. Baas, EEC 281 π 2π435

 Upsampled Signal (Time)

• One zero is
inserted
between the
original
samples for 2x
upsampling

B. Baas, EEC 281 436

 Upsampled Signal Spectrum
(Frequency)
• Spectrum of
2x
upsampled
signal
• Notice the
location of
the (now
somewhat
compressed)
five “bumps”
on each side
of π
B. Baas, EEC 281 π 2π437
 Post Upsampling Processing
• We likely want to attenuate images centered at π
from π/2 to 3π/2 because these are artificial artifacts
caused by the upsampling process
• A low‐pass filter will keep the original desired signal
and remove the artifacts (“image”) from π/2 to 3π/2
• This low‐pass filter is so common, that it has its own
name: an anti‐image filter

B. Baas, EEC 281 438

Downsampling
 Downsampling or Decimation

• To decimate by a factor D, keep one of every D

samples—on a periodic basis
• Downsampling by a factor I is commonly written I
For example, downsampling by two: 2
• Obviously the
number of samples
t
will be approximately
cut in half after 2

B. Baas, EEC 281 440

Decimation Example 1—
Time Domain

• 32 samples in original
waveform
• Downsampled by 2
• 16 samples in
downsampled waveform

B. Baas, EEC 281 441

Decimation Example 1—
Frequency Domain

• Downsampled by 2
• Frequencies in
downsampled waveform
are 2x higher in the
“digital frequency”
domain
• They interfere when input
frequencies reach π/2
(downsampled frequencies
reach π)

B. Baas, EEC 281 442

π 2π
 Decimation Example 2
Spectrum, Linear Scale
• Notice
signal
frequency
slope from
0–0.4π
• Plotted with
abs(fft(•))

B. Baas, EEC 281 443

0.4π 2π
downsamp2.m
 Decimation Example 2
Spectrum, dB Scale

• Spectral
slope from
0–0.4π not
visible with
dB vertical
scale
• Notice
signal goes
to “zero” at
0.4π
• Plotted
with psd(•)

B. Baas, EEC 281 444

0.4π π
downsamp2.m
 Decimation Example 2

• Decimated
by 2
• Spectral
slope now
from 0–0.8π
• Plotted with
abs(fft(•))

B. Baas, EEC 281 445

0.8π 2π
downsamp2.m
 Decimation Example 2

• Signal now
from 0–0.8π
though
sloping
shape is still
not visible
with dB
vertical scale
• Plotted with
psd(•)

B. Baas, EEC 281 446

0.8π π
downsamp2.m
 Decimation

• Signals above ½ the Nyquist frequency alias to lower

frequencies when downsampled
– Normally this is bad
– In rare cases, this can be exploited to some benefit

B. Baas, EEC 281 447

Original

0.4π

Down‐
sampled
by 2

B. Baas, EEC 281 0.8π 2π

448
Original

0.4π 0.6π

Down‐
sampled
by 2

B. Baas, EEC 281 0.8π 1.2π 2π

449
 Anti-alias filtering
• Normally we will want to filter the original signal
with a filter before decimation to attenuate signal
components which would alias into the desired
signal band(s)
• This filter also has its own name: an anti‐alias filter

B. Baas, EEC 281 450

 Decimation Example
• Example
– 1 MSample/sec signal, interested in 0–100 KHz portion only,
decimate as much as possible
– Need to watch “head room” above signal
– Decimate by three  new fs at 333 KHz
• Leaves 67 KHz between highest freq components and fs/2
– Decimate by four  new fs at 250 KHz
• Leaves 25 KHz between highest freq components and fs/2
– Decimate by five  new fs at 200 KHz
• Leaves 0 KHz between highest freq components and fs/2
• Some aliasing will occur

B. Baas, EEC 281 451