Skip to content

Matched Filtering

A template-correlation detector (tool) — instantiates Model-Guided Signal Separation

Detects and times a known signal shape buried in noise by correlating the observation against a template of that shape — the optimal linear detector once the noise is characterized.

Matched Filtering requires the one thing several of its siblings deliberately do without: an exact template of the target's waveform. It slides that template against the data and correlates, so the template itself is the target specification, and the correlation peak marks where — and when — the observation most resembles it. Among the mechanisms it is the detection specialist: it concentrates the scattered energy of a known shape into a single sharp peak and, for characterized noise, does so optimally, maximizing output signal-to-noise. Its question is "is the known thing present, and when?" — not "what is this unknown signal?" — which is precisely what separates it from the reconstruction and unmixing methods.

Example

A radar transmits a known chirp and listens for its echo. Each return is correlated against a stored copy of the transmitted pulse; a reflector at range implied by delay τ produces a correlation peak at that delay, and the correlation gathers the pulse's energy — spread thin and buried in receiver noise — into one clean spike that a threshold can catch. Because the receiver's noise is not white, the template is whitened to the noise spectrum first, which is what preserves the optimality. A target that would be invisible in the raw trace stands out sharply after matched filtering, timed to the resolution of the pulse.

How it works

The filter is the time-reversed, noise-whitened template; convolving it with the data yields a statistic that peaks wherever the data matches the template's shape, and a threshold on that peak declares a detection. The signal's shape must be known a priori — that is the price of admission. When the target varies over parameters (a chirp's rate, a waveform's mass), a whole bank of templates is swept and the best-matching one reported, which is powerful and also the seed of its main failure mode.

Tuning parameters

  • Template fidelity and bank density — how exactly the template matches the true signal, and how finely a parameter bank is sampled; mismatch silently loses signal-to-noise, a denser bank costs more compute.
  • Noise whitening — matching the filter to the noise spectrum; skip it on colored noise and false peaks appear.
  • Detection threshold — the peak height that declares a hit; lower it to catch weak signals at the cost of more false alarms.
  • Integration window — how much data is coherently combined; a longer window boosts signal-to-noise for a stable signal but smears a changing one.

When it helps, and when it misleads

It is the tool for known, transient signals in noise — radar and sonar returns, communication sync words, gravitational-wave templates — where it is the provably optimal linear detector.[1] It misleads when the template is wrong: mismatch quietly costs signal-to-noise and biases the estimated timing, un-whitened colored noise throws false peaks, and a large enough template bank will eventually match some template to pure noise. The classic misuse is lowering the threshold until the signal one was hoping for finally "appears." The discipline is to set the threshold from an explicit false-alarm model rather than by eye, and to confirm any candidate detection with an injected known signal before claiming it.

How it implements the components

  • target_signal_specification — the template is the specification of the target waveform; there is no separating this mechanism from a known signal shape.
  • separation_operator_specification — the whitened, time-reversed template is the correlation operator applied to the data.
  • nuisance_and_noise_model — the noise spectrum used to whiten the filter, which is what makes the detector optimal.
  • evidence_threshold — the detection threshold on the correlation peak that declares a hit versus a non-detection.

It does not model or invert how the signal was mixed or blurred (that is Deconvolution and Inverse Filtering) and carries no running uncertainty on a state (that is Kalman or Particle Filter).

References

[1] The matched filter maximizes the output signal-to-noise ratio for a known signal in additive white noise (North's theorem), and in that setting it is the Neyman–Pearson-optimal detector — no linear operation extracts a known waveform from such noise better. Whitening extends the result to colored noise by transforming it back to the white case before correlation. ↩