Polyphase Filtering

The core principle enabling polyphase efficiency is the Noble Identity, which mathematically proves that filtering followed by downsampling is equivalent to downsampling followed by filtering with a decomposed filter. This identity allows the filter to operate at the lower output sample rate rather than the high input rate.

• Reduces multiply-accumulate operations by a factor equal to the decimation ratio
• Eliminates the computation of samples that would be immediately discarded
• Applies identically to interpolation through the transposed Noble Identity

A prototype FIR filter of length N is decomposed into M parallel sub-filters, where M is the resampling factor. Each sub-filter contains coefficients selected from the original filter at intervals of M, starting at a different phase offset.

• Sub-filter 0: coefficients h[0], h[M], h[2M], ...
• Sub-filter 1: coefficients h[1], h[M+1], h[2M+1], ...
• Each sub-filter operates at 1/M the original sample rate
• The structure is inherently parallel and suitable for hardware implementation

In a polyphase decimator, the input signal is commutated across M sub-filters in a counter-clockwise rotation. The commutator delivers each input sample to the next sub-filter in sequence, and the outputs of all sub-filters are summed to produce the decimated output.

• Input commutator rotates at the high sample rate
• Each sub-filter processes only every M-th sample
• Output is produced at the decimated rate fs/M
• Eliminates the need for a separate downsampling stage

The polyphase interpolator reverses the decimator structure. The input signal is broadcast to all M sub-filters simultaneously, and an output commutator rotates to select the output of each sub-filter in sequence, inserting M-1 computed samples between each original input.

• Input is distributed to all sub-filters at the low sample rate
• Output commutator rotates at the high sample rate M*fs
• Each sub-filter computes a different fractional delay
• Zero-stuffing and subsequent filtering are combined into one efficient operation

Polyphase filters enable arbitrary rational resampling ratios L/M by cascading an interpolator (factor L) with a decimator (factor M). The intermediate filter operates at L times the input rate, but polyphase decomposition ensures only non-zero samples are processed.

• Supports ratios like 3/2, 5/4, or 147/160
• The interpolation and decimation filters can share a single prototype
• Critical for symbol timing recovery in software-defined radios
• Enables sample rate alignment between asynchronous digital domains

For a decimation-by-M filter with N taps, a direct implementation requires N multiplications per input sample. The polyphase implementation requires only N/M multiplications per input sample, a reduction factor of M.

• Example: N=128, M=8 → 128 vs 16 multiplications per input sample
• 87.5% reduction in multiply-accumulate operations
• Power savings directly proportional for hardware implementations
• Enables real-time wideband processing on resource-constrained FPGAs and DSPs

The process of digitally translating a sampled bandpass signal to baseband using a numerically controlled oscillator (NCO) and decimating filters. Polyphase filtering is the standard computational engine within a DDC, allowing the final channel filter to operate at the lower output sample rate rather than the high input rate. This architecture is fundamental to modern software-defined radio receivers, enabling channel selection with minimal hardware complexity.

The two fundamental operations of multirate signal processing. Decimation reduces the sampling rate by a factor of M, requiring a low-pass filter to prevent aliasing. Interpolation increases the sampling rate by a factor of L, requiring a filter to suppress spectral images. Polyphase decomposition restructures these filters so that computations are only performed on non-zero samples or for retained output samples, drastically reducing multiply-accumulate operations per second.

A set of mathematical equivalences that allow the order of filtering and downsampling/upsampling operations to be swapped under specific conditions. The First Noble Identity states that filtering a signal and then downsampling by M is equivalent to downsampling first and then filtering with the M-fold polyphase decomposition. These identities are the theoretical proof that validates the computational savings achieved by polyphase filter structures.

Cascaded Integrator-Comb (CIC) filters are an efficient class of multiplierless filters often used as a first-stage decimator or interpolator in high-speed digital links. While computationally simple, they suffer from passband droop. A polyphase FIR filter is typically cascaded after a CIC stage to compensate for this droop and provide sharp transition bands, combining the efficiency of CIC structures with the precision of polyphase FIR designs.

A filter bank splits a wideband signal into multiple subbands using parallel bandpass filters, a process efficiently implemented using polyphase decomposition and the Discrete Fourier Transform (DFT). This is the foundation of OFDM transceivers and audio compression algorithms. The polyphase DFT filter bank allows for uniform channelization where the computational load is shared across all channels, making it vastly more efficient than implementing independent filters for each subband.

An optimal linear filter that maximizes the signal-to-noise ratio (SNR) by correlating the received signal with a known pulse shape. In high-sample-rate receivers, the matched filter is often implemented as a polyphase decimating filter to simultaneously perform pulse shaping and sample rate reduction. This combined operation is critical in timing recovery loops where the signal must be filtered and downsampled to the symbol rate with precise phase alignment.