What is Linear Phase?
Every filter changes not just the amplitude of signals but also their phase. Phase distortion occurs when different frequencies are delayed by different amounts as they pass through a filter — the output waveform is then a smeared, distorted version of the input even if all amplitudes are perfectly preserved. Linear phase means the phase response of a filter is a linear function of frequency: φ(ω) = −αω for some constant α. This single property guarantees that every frequency component is delayed by exactly the same amount of time.
The practical consequence is profound: a linear-phase filter preserves the shape of the signal waveform. Transients arrive intact. Pulse edges remain sharp. This is the defining advantage of FIR filters over IIR designs for applications where waveform fidelity matters — audio processing, data communications, and biomedical signal acquisition.
Group Delay — The Key Metric
The group delay of a filter measures how much each frequency component is delayed in time. It is defined as the negative derivative of the phase response with respect to frequency:
An IIR filter, by contrast, has a nonlinear phase response — its group delay varies with frequency. Some frequencies are delayed more than others, causing the filter to "smear" the waveform across time. For audio, this manifests as a subtle muddiness. For pulse signals (digital communications, radar), it can distort the pulse shape enough to cause detection errors.
Symmetry — The Key to Linear Phase
Linear phase in an FIR filter is guaranteed by a simple condition on its coefficients: the impulse response must be either symmetric or antisymmetric about its midpoint. These two symmetry conditions, combined with two choices of filter length (odd or even), yield four distinct types of linear-phase FIR filters.
The Four Types of Linear-Phase FIR Filters
The four types arise from two binary choices: symmetry (even vs. odd) and filter length (odd N vs. even N). Each combination has different frequency-response characteristics that make it suitable for different applications.
| Type | Symmetry | Length | Forced Zeros | Best For |
|---|---|---|---|---|
| I | h[n] = h[N−1−n] | Odd | None | General LP/HP/BP/BS |
| II | h[n] = h[N−1−n] | Even | ω = π | LP, BP |
| III | h[n] = −h[N−1−n] | Odd | ω = 0 and π | BP, Hilbert |
| IV | h[n] = −h[N−1−n] | Even | ω = 0 | HP, Differentiator |
Why Forced Zeros Matter
The forced zeros at DC or Nyquist are not design choices — they are mathematical consequences of the symmetry and length. A Type II filter cannot pass frequency π regardless of how its coefficients are chosen, because the antisymmetry forces H(e^jπ) = 0 identically. Attempting to design a high-pass filter using Type II will always yield a degraded response near Nyquist.
This is why filter type selection is an important early design decision. For a high-pass FIR filter, always use Type I or Type IV. For a Hilbert transformer (90° phase shift across a wide band), Type III or IV is required. The windowed-sinc method automatically produces Type I by default (odd N, symmetric coefficients), which is why Type I is so prevalent in practice.
Type II and Type IV filters have a group delay of (N−1)/2 samples where N is even — so the delay is a non-integer (a "half-sample"). This is mathematically valid but means the output is delayed by a fractional number of samples relative to the input. In practice, the filter output must be interpreted carefully: it sits between input samples. For most signal processing applications this is fine, but for systems requiring strict sample-aligned outputs, Type I or III (integer group delay) is preferred.
Integer group delay: Types I and III. Half-sample group delay: Types II and IV.When Linear Phase Matters
Linear phase is essential whenever the shape of the output waveform must be a faithful reproduction of the input. The key applications are:
When Linear Phase Does Not Matter
Linear phase is not always necessary. For applications where only the power spectrum of the signal matters — not its waveform — IIR filters are often a better choice. Anti-aliasing filters in audio ADCs (Sigma-Delta converters) are IIR by design and work well because the ear is relatively insensitive to broadband phase distortion. Power spectral density estimation, envelope detection, and energy measurements are all phase-insensitive. Choosing an FIR filter purely for linear phase when the application is phase-insensitive wastes computation: a 5th-order IIR Butterworth can equal the frequency-selectivity of a 50-tap FIR at one-tenth the cost.
FIR filters guarantee linear phase through coefficient symmetry. IIR filters cannot achieve exact linear phase (all-pass equalisation can compensate but adds latency and complexity). The choice often comes down to latency tolerance: offline processing can afford large FIR filters; real-time embedded systems often prefer the efficiency of IIR designs despite their phase distortion.
Need waveform fidelity → FIR with linear phase. Need efficiency → IIR (accept nonlinear phase).- Linear phase means φ(ω) = −αω — all frequencies are delayed by the same amount α, preserving waveform shape.
- Group delay τ(ω) = −dφ/dω = (N−1)/2 is constant for a linear-phase FIR filter of length N.
- Coefficient symmetry (h[n] = h[N−1−n]) or antisymmetry (h[n] = −h[N−1−n]) is the necessary and sufficient condition for linear phase in an FIR filter.
- Four types: I (symmetric, odd N), II (symmetric, even N), III (antisymmetric, odd N), IV (antisymmetric, even N).
- Type II has a forced zero at ω = π — cannot implement high-pass or band-stop filters.
- Types III and IV have a forced zero at ω = 0 — cannot implement low-pass filters.
- Linear phase is critical for audio, communications, biomedical, and image processing applications where waveform shape matters.
- When only the power spectrum matters, IIR filters are more efficient and linear phase can be sacrificed.