Does QMF always mean two bands?

The classical construction is two-band, but the term is also used more broadly in some filter-bank literature.

Is each downsampled band free from aliasing?

Not necessarily. Properly designed banks can cancel alias contributions when the bands are recombined.

Does perfect reconstruction survive arbitrary band editing?

No. Quantization, deletion or other changes can alter the output even when the unmodified bank reconstructs exactly.

AI Glossary / Computer vision and multimodal AI / Intermediate

Quadrature mirror filter

Definition

A quadrature mirror filter, or QMF, is a filter used in a filter-bank arrangement that separates a signal into complementary frequency bands. In the classical two-band construction, a high-pass response mirrors a low-pass response around a specified mid-band frequency.

What it means

Filter banks divide a signal into parts that can be processed, compressed or analyzed separately. A two-band analysis stage produces a lower-frequency component and a higher-frequency component, often followed by downsampling. A synthesis stage later combines the bands. The mirror relationship is a design property, but it does not alone guarantee perfect reconstruction. The analysis and synthesis filters, sampling operations and normalization must work together. Terminology has broadened over time, so some references use QMF filter bank for more general multiband or reconstruction-oriented designs.

How it works

In a classical discrete-time two-channel setting, a high-pass analysis filter can be related to the low-pass filter by alternating signs of coefficients, corresponding to H₁(z) = H₀(−z) under the stated convention. The filtered outputs are commonly downsampled by two. Each decimated branch may contain aliasing, but a properly designed synthesis bank can cancel those alias terms when the branches are recombined.
Cancellation of aliasing is only part of reconstruction: the combined amplitude and phase response must also meet the desired conditions. Filters and conventions differ across QMF, conjugate-quadrature and broader perfect-reconstruction constructions, so equations should be taken from one consistent design rather than mixed from different sources.

Three practical examples

An audio subband system separates lower and higher frequency content before coding or analysis.
A multiresolution image transform applies related low-pass and high-pass filtering along image dimensions.
A signal-processing experiment tests whether a chosen analysis and synthesis bank reconstructs its input within numerical tolerance.

Think of it this way

Think of separating a mixture into two coordinated streams and later recombining them. The streams must be designed as a pair if their combined result is to recover the original. The analogy is limited because downsampling introduces mathematical alias terms, and reconstruction depends on their cancellation rather than merely pouring two unchanged portions back together.

Worked example

Use a simple orthonormal Haar pair on two samples, 6 and 2. The low-band value is (6 + 2)/√2 = 4√2, and the high-band value is (6 − 2)/√2 = 2√2. Reconstruction gives the first sample as (4√2 + 2√2)/√2 = 6 and the second as (4√2 − 2√2)/√2 = 2.
This pair demonstrates coordinated averaging and differencing with exact reconstruction in ideal arithmetic. If the high-band value is discarded, both reconstructed samples become 4, so detail is lost. Longer filters, boundary handling and multilevel decompositions require additional conventions; this tiny example does not define every QMF implementation.

Illustrative case study

In a fictional compression prototype, a team splits a signal into low and high bands and confirms exact reconstruction before adding quantization. After quantization, the reconstructed signal differs from the source even though the filter bank itself remains correctly designed. The team separates transform error from coding error and checks the first and last samples for boundary artifacts. This case illustrates why perfect-reconstruction filter properties do not imply a lossless end-to-end codec after the bands are altered.

Common misconceptions and mistakes

A mirror frequency response is not sufficient evidence that a complete bank reconstructs perfectly. Downsampling each band can introduce aliasing even when the full bank is designed to cancel it. Processing the branches independently can disturb that cancellation or add distortion. Another mistake is treating every use of QMF as the exact same two-filter formula. Historical and application-specific terminology should be checked, especially when implementing a design from equations.

Business use

For a product using filter-bank processing, verify the complete analysis - synthesis path before evaluating compression or enhancement. Record filter coefficients, scaling, sampling phases and boundary handling. Test impulse signals, constant inputs and representative content, then measure what changes after subband processing. The practical benefit may be efficient multiscale analysis or coding, but it should be assessed alongside latency, computation and reconstruction artifacts rather than inferred from the filter-family name.

Check your understanding

A Haar analysis produces one average-like value and one difference-like value. What happens if the difference value is discarded before reconstruction?

Answer: The pair's original difference is lost, so reconstruction retains a common average-like value rather than recovering both original samples.

Related terms

Low-pass filter
High-pass filter
Filter bank
subband coding
downsampling
alias cancellation
perfect reconstruction
Haar transform
wavelet transform

Learn next

Learn two-channel analysis and synthesis next. Reconstruct a short signal using a consistent Haar convention, then alter one subband and identify which changes arise from processing rather than from the unmodified transform.

Sources

Julius O. Smith III, Spectral Audio Signal Processing, Quadrature Mirror Filters: https://www.dsprelated.com/freebooks/sasp/Quadrature_Mirror_Filters_QMF.html Primary author-textbook page consulted; full implementation-specific and editorial review remain pending.

Previous: Quadratic unconstrained binary optimization Next: Qualification problem