What Are Filter Banks?

Filter banks are collections of parallel filters applied to a common input signal, producing multiple output signals that together represent the frequency content of the original. Each filter in the bank passes a specific frequency band while attenuating others, enabling the decomposition of a wideband signal into subband components. This structure is central to multirate signal processing, where different frequency bands can be processed, compressed, or transmitted at different sample rates.

Filter bank theory draws on classical digital filter design, Fourier analysis, and linear algebra. The concept emerged prominently in telecommunications during the 1970s and 1980s, where efficient spectral decomposition was needed for speech coding and transmission. Since then, filter banks have become a foundational tool across audio engineering, image compression, and wireless communications.

Analysis and Synthesis Banks

A filter bank system typically operates in two complementary stages. The analysis bank decomposes the input signal into subbands by passing it through a set of bandpass filters, each followed by a downsampler that reduces the data rate in proportion to the bandwidth of the subband. The synthesis bank performs the inverse operation: it upsamples each subband signal and passes it through a second set of filters, recombining the results into a reconstructed output. When the analysis and synthesis filters are designed to cancel aliasing and amplitude distortion, the system achieves perfect reconstruction, where the output matches the input exactly up to a fixed delay.

The IEEE Transactions tutorial on multirate digital filters, filter banks, and polyphase networks by P.P. Vaidyanathan is the standard reference for this topic, covering perfect reconstruction conditions, the M-band QMF bank, and the polyphase representation in detail.

Polyphase Structures

A polyphase decomposition rewrites a filter's transfer function as a sum of components, each operating on a downsampled version of the input. This representation reveals that many filter bank computations can be shared across channels, dramatically reducing the total number of multiplications and additions compared to a naive parallel implementation. Polyphase structures are the basis for efficient fast filter bank implementations and underpin practical applications in audio codecs and software-defined radio receivers.

Wavelets and Multiresolution Analysis

There is a direct mathematical correspondence between two-channel filter banks and the discrete wavelet transform. When the analysis filters satisfy the biorthogonality or orthonormality conditions, the resulting system produces a multiresolution decomposition of the input signal. This connection, formalized in the late 1980s by researchers including Stephane Mallat, linked classical subband coding to wavelet theory and opened the way for filter bank design methods grounded in functional analysis. The JPEG 2000 image compression standard uses a biorthogonal wavelet filter bank as its core transform, illustrating the direct path from filter bank theory to deployed compression systems.

Oversampled filter banks, which produce more output samples than the input, provide additional design freedom and noise robustness at the cost of redundancy. Frame theory gives the mathematical tools to characterize these systems, and oversampled designs appear in communications systems that benefit from robustness to channel impairments.

Applications

Filter banks have applications in a wide range of fields, including:

• Audio and speech coding, including the MP3 codec's modified discrete cosine transform bank
• Image and video compression in standards such as JPEG 2000 and HEVC
• Multicarrier modulation systems such as OFDM-based wireless standards
• Hearing aid signal processing and cochlear implant sound coding
• Subband adaptive filtering for acoustic echo cancellation
• Biomedical signal analysis, including EEG and ECG decomposition