Signals: Multirate Filter Banks

Multirate filter banks are the engine behind modern audio, image, and communication compression. By splitting a signal into multiple frequency bands, these systems allow us to process each subband independently—compressing some, discarding others, or modifying them for transmission—before seamlessly stitching the signal back together. Mastering their design is key to building efficient systems that manipulate signal bandwidth without losing critical information.

What is a Filter Bank?

A filter bank is a collection of filters used to separate an input signal into two or more frequency subbands. Think of it as an audio equalizer in reverse: instead of boosting or cutting existing bands, a filter bank first creates those isolated bands from the full signal. The primary structure you will encounter is the two-channel quadrature mirror filter (QMF) bank. It consists of an analysis bank, which downsamples the signal after filtering, and a synthesis bank, which upsamples and filters before combining the outputs. The analysis bank's job is to decompose the signal; the synthesis bank's job is to reconstruct it. The goal is often perfect reconstruction (PR), where the output is a perfect, possibly delayed, copy of the input.

The Two-Channel QMF Bank in Detail

The standard two-channel QMF bank is your foundational model. The input signal $x[n]$ is fed into two parallel analysis filters: a lowpass filter $H_0(z)$ and a highpass filter $H_1(z)$. These filters split the signal into approximate low-frequency and high-frequency components, respectively. Immediately after filtering, each branch undergoes downsampling by a factor of 2 (denoted by $\downarrow 2$), which discards every other sample to reduce the data rate. These downsampled subband signals, $v_0[n]$ and $v_1[n]$, are now available for processing (e.g., compression, coding, or transmission).

On the synthesis side, each subband signal is first upsampled by a factor of 2 (denoted by $\uparrow 2$), which inserts a zero between every sample. The upsampled signals are then passed through synthesis filters: a lowpass $G_0(z)$ and a highpass $G_1(z)$. Finally, the outputs of these synthesis filters are summed to produce the reconstructed signal $\hat{x}[n]$.

Achieving Perfect Reconstruction and Canceling Aliasing

The central challenge in filter bank design is perfect reconstruction. Downsampling causes aliasing—high frequencies fold back into low frequencies, distorting the signal. The synthesis bank must be designed specifically to cancel this aliasing. For the two-channel case, the condition for aliasing cancellation is straightforward. It requires the synthesis filters to be related to the analysis filters as: $$G_0(z)=H_1(-z)\text{and}{G}_1(z)=-H_0(-z)$$ This relationship ensures that the aliasing components generated in the analysis stage are perfectly canceled in the synthesis stage. With aliasing removed, the next condition is to achieve perfect reconstruction of the original signal's content. The perfect reconstruction condition dictates that the overall transfer function of the aliasing-cancelled system must be a pure delay: $$T(z)=\frac{1}{2}[H_0(z)G_0(z)+H_1(z)G_1(z)]=c{z}^{-k}$$ Here, $c$ is a constant and $k$ is an integer delay. For a QMF bank using real-coefficient filters, a common simplifying choice is to define the highpass analysis filter as $H_1(z)=H_0(-z)$. When combined with the aliasing cancellation formulas above, this leads to the synthesis filters being mirrors of the analysis filters. However, with this real-coefficient constraint, exact perfect reconstruction is impossible except for a trivial two-tap filter. In practice, we design filters that come very close to the PR condition, making the error negligible for applications like audio coding.

Polyphase Decomposition for Efficient Implementation

Direct implementation of a filter bank, as described, is computationally wasteful. It requires filtering at the high input rate before immediately discarding half the samples. Polyphase decomposition is a crucial structural transformation that enables highly efficient implementation. The idea is to decompose each analysis filter $H_i(z)$ into its even and odd components (its polyphase components), which correspond to the filter's impulse response at even and odd time indices.

For example, the lowpass analysis filter can be decomposed as: $$H_0(z)=E_0(z^2)+z^{-1}{E}_1(z^2)$$ where $E_0(z)$ and $E_1(z)$ are the polyphase component filters. This clever rearrangement allows us to commute the downsampling operation to the front of the structure. The input is split into even and odd sample streams, each of which is then processed by a parallel bank of smaller, simpler polyphase filters running at half the rate. This structure is called the polyphase implementation. It reduces the computational load by roughly 50% because all filtering occurs at the lower, downsampled rate. This efficiency is non-negotiable in real-time systems like audio codecs and software-defined radios.

Applications in Audio Coding and Communications

The theory of filter banks is directly applied in cornerstone technologies. In audio coding standards like MP3 and AAC, a filter bank (often a modified discrete cosine transform, which can be viewed as a uniform filter bank) splits the audio signal into multiple subbands. Psychoacoustic models then determine which subbands contain inaudible information, allowing those bands to be coarsely quantized or discarded entirely, achieving high compression with minimal perceptual loss.

In communication systems, filter banks enable technologies like orthogonal frequency-division multiplexing (OFDM). OFDM uses a bank of closely spaced, orthogonal subcarriers to transmit data. Each subcarrier is effectively a frequency subband. This approach makes the system robust against frequency-selective fading, as a deep fade affects only a small subset of the subbands, and error correction can recover the data. The efficient polyphase implementation is critical here for meeting the low-latency, high-throughput demands of modern wireless standards.

Common Pitfalls

1. Ignoring Aliasing in Non-PR Designs: Assuming that because your filters have a "nice" frequency response, the reconstructed signal will be fine. Downsampling always introduces aliasing. If you are not explicitly using a structure that guarantees aliasing cancellation (like the QMF relationships), your output will be severely distorted. Always check the aliasing term in the system's transfer function.
2. Confusing Analysis and Synthesis Filter Roles: The analysis filters $H_i(z)$ and synthesis filters $G_i(z)$ are not interchangeable. They are designed as a matched pair to satisfy the aliasing cancellation and PR conditions. Using an analysis filter in a synthesis slot, or vice versa, will lead to complete system failure.
3. Overlooking the Efficiency of Polyphase: Implementing a filter bank in its naive "filter-then-downsample" form for a class project or simulation might work, but it misses a fundamental engineering lesson. The polyphase structure isn't just an academic trick; it's the only practical way to implement such systems in real-world, resource-constrained hardware and software.
4. Assuming Perfect Reconstruction is Always the Goal: While PR is a beautiful theoretical benchmark, many practical systems are near-perfect reconstruction (NPR) banks. In perceptual audio coding, for instance, the filter bank is designed to be alias-free but may have some amplitude distortion, as this minor error is masked by the human auditory system and allows for better compression filters.

Summary

• Filter banks decompose a signal into frequency subbands for independent processing, a fundamental process in compression and communications.
• The two-channel QMF bank is the canonical structure, comprising analysis filters/downsamplers and synthesis upsamplers/filters.
• The core design challenge is aliasing cancellation, achieved by linking synthesis to analysis filters via $G_0(z)=H_1(-z)$, with the ultimate goal of perfect reconstruction $T(z)=c{z}^{-k}$.
• Polyphase decomposition is an essential implementation technique that commutes downsampling to the front, allowing all filtering to occur at the lower subband rate, drastically improving computational efficiency.
• These concepts are directly applied in audio coding standards (MP3, AAC) for perceptual compression and in communication systems (OFDM) for robust data transmission over challenging channels.