Polyphase Filter Bank

The core innovation of a PFB is performing filtering at the decimated output rate rather than the high input rate. By decomposing the prototype filter into M polyphase components and placing the downsamplers before the filters, the arithmetic workload is reduced by a factor of M compared to a naive channelizer. This enables wideband processing on resource-constrained edge hardware.

Unlike a simple FFT-based channelizer with rectangular windowing, a PFB uses a carefully designed prototype low-pass filter to control spectral leakage. The stopband attenuation of this filter directly determines the adjacent channel rejection.

• High stopband attenuation (>80 dB) prevents strong signals from masking weak adjacent channels.
• This spectral isolation is critical for accurate modulation classification in dense signal environments.

PFBs can be configured with different oversampling ratios to trade off channel spacing for alias protection:

• Critically Sampled (M = D): The number of channels equals the decimation factor. Maximizes spectral efficiency but suffers from aliasing at channel edges.
• Oversampled (M < D): Fewer channels than the decimation factor. Provides a flat passband and high alias rejection, simplifying downstream tasks like modulation classification and parameter estimation at the cost of increased output data rate.

A PFB designed as an analysis filter bank can be paired with a corresponding synthesis filter bank to perfectly reconstruct the original wideband signal from its sub-channels. This is achieved when the polyphase components satisfy specific orthogonality conditions. This property is essential for applications like spectrum stitching and distributed signal processing, where sub-band signals must be recombined without distortion.

The PFB's regular structure maps efficiently to FPGA fabric. The polyphase decomposition allows each branch filter to run at the lower output sample rate, enabling time-division multiplexing of a single Multiply-Accumulate (MAC) unit across multiple channels. Combined with the FFT IP core, a complete wideband channelizer can be synthesized with minimal logic resources, making it the standard architecture for direct RF sampling front-ends.

Modern PFB implementations support runtime reconfiguration of channel spacing and bandwidth by reloading prototype filter coefficients and adjusting the FFT size. This allows a single SDR platform to adapt to different communication standards (e.g., LTE, 5G NR, WiFi) without hardware changes. The ability to dynamically isolate specific frequency bands is a key enabler for cognitive radio and automatic modulation classification systems that must operate in unknown spectral environments.

The prototype filter is the single low-pass FIR filter from which all sub-band filters are derived through modulation. Its design critically determines channel isolation and reconstruction fidelity.

• Stopband attenuation: Dictates adjacent channel rejection (typically >80 dB for spectral analysis)
• Passband ripple: Controls in-band amplitude distortion
• Transition bandwidth: Sets the guard band between channels
• Common designs: Kaiser window, Parks-McClellan equiripple, and Nyquist-M filters for perfect reconstruction

The core sampling rate change operations that enable the computational efficiency of polyphase structures. Decimation by M discards M-1 out of every M samples, while interpolation by L inserts L-1 zeros between samples.

• Noble identities: Allow filter and rate-change operations to be swapped, enabling the polyphase decomposition
• Critical sampling: When decimation factor equals the number of channels, maximizing spectral efficiency
• Oversampled banks: Decimation factor less than channel count, providing redundancy for better alias rejection

A computationally efficient implementation where all sub-band filters are generated by complex-modulating a single prototype filter using a Discrete Fourier Transform. This exploits the FFT for dramatic complexity reduction.

• Uniform channelization: All channels have identical bandwidth and spacing
• Complex modulation: Shifts the prototype to center frequencies at multiples of 2πk/M
• Polyphase + FFT: The combination reduces M filters of length N to one polyphase decomposition plus an M-point FFT
• Odd-stacking vs even-stacking: Determines whether channels are centered at k/M or (k+0.5)/M of the sample rate

The polyphase filter bank acts as a front-end channelizer that separates a wideband input into parallel narrowband streams, each potentially containing a distinct modulated signal for downstream classification.

• Simultaneous multi-signal classification: Each sub-band feeds an independent classifier instance
• Dynamic channel allocation: Sub-bands can be routed to classifiers only when energy is detected
• Variable bandwidth support: Hierarchical filter banks can accommodate signals of different bandwidths
• Hardware implementation: Efficiently maps to FPGA systolic arrays and polyphase DSP blocks

The mathematical constraints ensuring that a signal can be decomposed into sub-bands and reconstructed without amplitude, phase, or aliasing distortion. Critical for analysis-synthesis applications.

• Nyquist(M) criterion: The prototype filter's autocorrelation must satisfy specific zero-crossing constraints
• Power complementarity: The squared magnitude responses of all analysis filters must sum to a constant
• Paraunitary filter banks: Orthogonal filter banks that preserve energy and are numerically well-conditioned
• Cosine-modulated vs DFT-modulated: Cosine banks achieve perfect reconstruction with real signals using only half the channels

The Weighted Overlap-Add structure is an enhanced polyphase architecture that applies an additional analysis window before the FFT, providing superior spectral containment and flexibility.

• Arbitrary overlap: Decoupling the window length from the FFT size allows independent control of time-frequency resolution
• Superior alias rejection: The additional windowing provides >100 dB stopband attenuation
• Dynamic reconfigurability: Channel spacing and bandwidth can be changed by adjusting the overlap factor
• Common in SDRs: Used in advanced software-defined radio platforms for spectrum monitoring and signal intelligence