Multirate Signal Processing Basics With easy examples.

Multirate signal processing is the art of changing a signal's sampling rate inside a system: lowering it when less data will do, raising it when finer timing is needed, or converting between unrelated rates. Different stages then run at the rates that suit them best instead of one fixed rate everywhere.

This matters because processing at a lower rate means proportionally fewer computations per second, saving power and cost. Oversampled converters, software-defined radios, audio format converters and image scalers all rely on multirate techniques. The celebrated trick: a filter running after downsampling does a fraction of the work of the same filter at full rate.

Key Concepts

• Downsampling by M: keep every M-th sample; the output rate is fs/M.
• Upsampling by L: insert L−1 zeros between samples; the rate becomes L·fs but no new information is added yet.
• Decimation: anti-aliasing lowpass filter followed by downsampling. The filter must cut off below fs/(2M) or high frequencies fold back as aliases.
• Interpolation: upsampling followed by an anti-imaging lowpass that fills the zero gaps with smooth values.
• Rational change L/M: upsample by L, filter once, downsample by M; used when rates share no simple integer ratio.
• Noble identities: rules letting a filter hop across an upsampler or downsampler so it runs at the lower rate.
• Polyphase decomposition: splitting a filter into parallel sub-filters so only surviving computations are performed.

Formulas

• Downsampled sequence: y[m] = x[M·m]; new rate fs' = fs/M.
• Anti-alias cutoff: the decimation filter passes 0 to fs/(2M) and stops everything above.
• Rate-change factor: converting fs_in to fs_out uses L/M ≈ fs_out/fs_in; one lowpass with cutoff at the lower Nyquist serves as both anti-alias and anti-image filter.
• Assumption: the signal is band-limited to the new Nyquist frequency before downsampling, or aliasing corrupts it permanently.

Worked Example

Decimate CD audio at 44,100 Hz by M = 4 for a voice application. The new rate is 11,025 Hz, so the anti-aliasing filter must cut off at or below 5,512.5 Hz before any sample is dropped. Skipping the filter would fold 5.5–22 kHz content into the voice band as distortion.

The savings: 60 s of stereo 16-bit audio at 44,100 Hz is 60 × 44,100 × 2 × 2 = 10,584,000 bytes (about 10.1 MB). After decimation by 4 it is 2,646,000 bytes (about 2.5 MB), and every downstream filter runs one quarter of the multiplications per second. Radios decimate as early as possible for exactly this reason.

Common Misconceptions

• Downsampling alone is decimation. Dropping samples without the anti-aliasing filter is just aliasing. The filter is what makes it decimation.
• Upsampling adds detail. Zeros add no information; the interpolation filter only reconstructs smooth values between originals.
• Higher rate is always better. It costs proportionally more computation, memory and power. Use the lowest rate that preserves the bandwidth.

Real-World Applications

• Software-defined radios decimate the wideband ADC stream to just the channel of interest.
• Sigma-delta ADCs oversample at megahertz rates then decimate to high-resolution audio-rate output.
• Audio converters change 48 kHz studio audio to 44.1 kHz CD audio with rational L/M resampling.

Multirate processing runs each stage at the rate it deserves: filter before decimating, interpolate after upsampling, and keep every filter at the cheapest possible rate.