Filter Banks & Polyphase Structures
Split a wideband signal into M subbands, process each independently, then reassemble — the architecture behind MP3, OFDM, and wavelet compression.
What is a Filter Bank?
An analysis bank decomposes the input into M subbands. A synthesis bank reconstructs the original. When designed correctly, x̂[n] = x[n − k] — perfect reconstruction.
Filtering & Decimating
Apply M bandpass filters then downsample each by M. Total data rate stays the same: M subbands × fs/M = fs.
Upsampling & Summing
Upsample each subband by M, filter with synthesis filters Gk, then sum all M branches. Joint design of Hk & Gk achieves perfect reconstruction.
Polyphase Decomposition
Split H(z) into M shorter branches Ek(zM). All arithmetic runs at the subband rate fs/M. Combine branches with an FFT — computation drops by a factor of M.
Quadrature Mirror Filters
- M = 2: one lowpass H0, one highpass H1
- H1(z) = H0(−z) → h1[n] = (−1)n h0[n]
- Aliasing from the two decimators cancels exactly
- Johnston & paraunitary (CQF) designs achieve exact PR
M-Channel Bank
Adjust M to see how the subbands tile the spectrum. Each colored band is one subband of width π/M.
MP3 & AAC
- MP3: 32-channel polyphase bank + MDCT subbands
- AAC: pure MDCT filter bank, better resolution
- Allocate bits by perceptual importance per subband
- Masked subbands receive fewer (or zero) bits
- Wavelet (JPEG 2000): recursive QMF tree → multi-resolution
OFDM as a Filter Bank
The OFDM transmitter is a polyphase synthesis bank — each subcarrier is one channel. The IFFT combines M data streams into one wideband signal. The receiver's FFT is the analysis bank.
Multirate DSP Done
- Filter banks split & recombine signals via analysis/synthesis pairs
- Polyphase decomposition cuts cost by factor M — all work at fs/M
- QMF banks (M=2) cancel aliasing; PR conditions ensure exact recovery
- Powers MP3, AAC, JPEG 2000, OFDM, hearing aids, wavelets