What Is a Filter Bank?

A filter bank is a collection of bandpass filters applied in parallel to decompose a signal into a set of frequency sub-bands, or conversely to reconstruct a signal from multiple sub-band signals. The analysis stage passes the input through a bank of filters, each centered on a different frequency, and the synthesis stage recombines the filtered outputs to recover or modify the original signal. Filter banks are a foundational tool in multirate signal processing, audio coding, speech analysis, and wireless communications, where operating on individual frequency components is more efficient or more informative than processing the full-band signal.

The mathematical framework for filter banks draws from linear systems theory, digital signal processing, and approximation theory. Multirate processing is integral to practical filter bank design: after filtering, each sub-band signal is typically downsampled (decimated) to reduce the data rate to the minimum needed for that band, and upsampled before synthesis. The classic paper by P.P. Vaidyanathan, Multirate Digital Filters, Filter Banks, Polyphase Networks, and Applications (Caltech Authors Repository), established the theoretical foundation for perfect reconstruction filter banks and polyphase representations that continue to underpin modern designs.

Analysis and Synthesis Structure

In the analysis stage of a maximally decimated M-band filter bank, M bandpass filters divide the input spectrum into M sub-bands, each of bandwidth approximately 1/M of the original. Each filter output is decimated by a factor of M, so the total number of samples in all sub-bands equals the original sample count. In the synthesis stage, each sub-band signal is upsampled by M and passed through a corresponding synthesis filter, and the M outputs are summed. For the filter bank to achieve perfect reconstruction, the combined effect of the analysis filters, decimation, upsampling, and synthesis filters must equal a pure delay with no aliasing or distortion. The polyphase decomposition is the primary computational tool for achieving this efficiently, reducing the M-band filter bank to a polyphase matrix operation followed by a single DFT or modulation. The WPI DSP course notes on multirate filter banks provide a worked derivation of polyphase structures and their computational savings.

Quadrature Mirror Filter Banks

The two-channel quadrature mirror filter (QMF) bank is the simplest and historically most studied filter bank structure. It splits the input into a lowpass and a highpass sub-band, each decimated by two. The QMF condition requires the highpass filter to be the frequency-reversed mirror of the lowpass filter, which ensures that aliasing introduced by decimation is cancelled in the synthesis stage. Maximally flat QMF banks, such as the Johnston and Smith-Barnwell designs, were widely used in early speech coding and audio compression systems. M-band extensions of the QMF structure, including cosine-modulated filter banks and the MDCT (modified discrete cosine transform) used in MP3 and AAC audio coding, generalize the two-channel principle to arbitrary numbers of sub-bands. Wavelet filter banks are a special case in which the lowpass sub-band is recursively decomposed, producing a dyadic tree decomposition whose basis functions are the discrete wavelet transform coefficients.

Applications in Communications and Coding

Filter banks play a central role in multicarrier modulation, where an OFDM (orthogonal frequency-division multiplexing) transmitter can be interpreted as an inverse DFT followed by a cyclic prefix, equivalent to a synthesis filter bank. FBMC (filter bank multicarrier) schemes replace the rectangular prototype filters of OFDM with well-designed bandpass prototypes that achieve much lower spectral sidelobes, improving robustness to frequency offset and interference. In audio and speech coding, filter banks decompose the signal into psychoacoustically motivated sub-bands, allowing quantization noise to be shaped to stay below the masking threshold. Image compression standards such as JPEG 2000 use two-dimensional filter banks based on the Cohen-Daubechies-Feauveau 9/7 biorthogonal wavelet. The McMaster University ECE 765 course on multirate filter banks and wavelets surveys design methods for communication and coding applications.

Applications

Filter banks have applications in a wide range of signal processing and communication systems, including:

• Audio and speech coding (MP3, AAC, Opus) using MDCT-based filter banks
• Multicarrier wireless communications (OFDM, FBMC) for 4G LTE and 5G NR
• Subband image compression in JPEG 2000 and wavelet-based coding
• Hearing aids and auditory prostheses that process frequency bands independently
• Spectrum sensing and cognitive radio systems monitoring multiple frequency channels
• Vibration analysis and structural health monitoring decomposing sensor signals into frequency components