Matched Filtering and Template Banks

A gravitational-wave detector delivers one number, the strain $s(t)$ , many thousand times a second — and buried in that stream, when two black holes collide a billion light-years away, is a signal $h(t)$ of amplitude $10^{-21}$ : a length change smaller than a proton across LIGO's four-kilometre arms, a hundred to a thousand times fainter than the detector's own noise. At any single instant it is entirely lost. What rescues it is that general relativity, through the quadrupole formula and its post-Newtonian corrections, predicts the shape of the waveform with exquisite accuracy — the chirp of the inspiral, the crescendo of merger, the ringdown — as a deterministic function of the source's masses and spins. Matched filtering turns that known shape into a detection: it cross-correlates the data against a theoretical template, weighting each frequency by how quiet the detector is there, so noise averages toward zero while the matched part of the signal adds coherently over $10^3$ – $10^4$ cycles. That coherent gain is what lifts a $10^{-21}$ strain above the floor. This article builds the framework from its foundation — a noise-weighted inner product that turns the space of signals into a geometry — proves that the matched filter is the optimal linear detector by nothing more than Cauchy–Schwarz, and follows the idea into practice: how the continuous parameter space is tiled by a discrete template bank, how a Riemannian metric fixes that tiling, and how the whole apparatus produced the announcement, in February 2016, of GW150914.

Key takeaways

Matched filtering is the provably optimal linear method for extracting a known waveform from stationary Gaussian noise: cross-correlate the data with a template, weighting each frequency by the inverse noise power $1/S_n(f)$ .
The whole theory lives on one object — the noise-weighted inner product $\langle a|b\rangle$ — which turns signals into vectors, signal-to-noise into a length, and matching a template into aligning two vectors.
Optimality is one line of Cauchy–Schwarz: no linear statistic beats $\rho\le\sqrt{\langle h|h\rangle}$ , with equality when the filter equals the template — the Wiener matched filter.
Because the true parameters are unknown, one filters against a discrete template bank whose density is set by a metric on parameter space and a chosen minimal match (typically 0.97); real banks hold $10^4$ – $10^6$ templates.
Maximising over arrival time is free (one inverse FFT gives the SNR at every time at once) and over phase and amplitude is analytic, so each template probes only the intrinsic masses and spins.
The method underlies every confirmed detection since GW150914, recovered at combined signal-to-noise ratio 24 with a false-alarm rate below one per 200 000 years.
Key idea: define $\langle a|b\rangle = 4\,\mathrm{Re}\int_0^\infty \tilde a^*(f)\tilde b(f)/S_n(f)\,\mathrm{d}f$ ; the matched filter is the statistic $\langle h|s\rangle$ , which maximises the expected SNR among all linear filters and peaks at $\rho=\sqrt{\langle h|h\rangle}$ when the template equals the true signal.

The core insight

Define the noise-weighted inner product $\langle a|b\rangle = 4\,\mathrm{Re}\int_0^\infty \tilde a^*(f)\tilde b(f)/S_n(f)\,\mathrm{d}f$ . The matched filter applies the template $h$ to the data $s$ through $\langle h|s\rangle$ ; this statistic maximises the expected signal-to-noise ratio among all linear filters, with peak value $\rho=\sqrt{\langle h|h\rangle}$ attained when the template equals the true signal.

Where this lives today

Matched filtering is the beating heart of every operating gravitational-wave observatory. The two production pipelines of the LIGO–Virgo–KAGRA collaboration, PyCBC and GstLAL, filter the incoming strain against banks of hundreds of thousands of inspiral–merger–ringdown templates in near-real time, generating an SNR time series per template per detector and combining them into ranking statistics calibrated by time-slides. This is what produces the public alerts, within seconds to minutes, that sent telescopes chasing the kilonova of the neutron-star merger GW170817. Once a candidate survives, the same inner product drives Bayesian parameter estimation, sampling the likelihood $\ln\Lambda=\langle h|s\rangle-\tfrac12\langle h|h\rangle$ ; because posterior widths scale as $1/\rho$ , the loudest events give the tightest masses, spins, and distances. The method is also a precision test of general relativity: matched filtering is so sensitive to phase that comparing the recovered inspiral against the post-Newtonian and numerical-relativity predictions bounds anomalous dispersion, the graviton mass, and the spin-2 character of the radiation. The frontier is computational: banks for the next-generation detectors (Einstein Telescope, Cosmic Explorer) and the space-based LISA mission will need $10^7$ or more templates over far longer signals, driving reduced-order and machine-learning surrogate models, hierarchical and GPU-accelerated searches, and the continuous-wave $\mathcal{F}$ -statistic searches for spinning neutron stars whose parameter-space metric is so vast that only semi-coherent stacking is affordable.

The physical picture

Imagine straining to hear one precisely known melody hidden inside a roar of static. If you know the tune exactly, you can multiply the incoming sound, instant by instant, against a stored copy and keep a running sum. The random static is as often out of step as in step, so its contributions cancel and the sum wanders near zero; the matched portion always multiplies positive-with-positive and negative-with-negative, so it accumulates. Over $N$ notes the coherent sum grows like $N$ while the noise grows only like $\sqrt N$ , so the signal-to-noise ratio climbs like $\sqrt N$ . A gravitational-wave chirp sweeps through the band for thousands of cycles, and matched filtering banks every one — which is how a $10^{-21}$ strain, hopeless in any single sample, becomes a confident detection. The noise weighting is what separates the real matched filter from naive correlation. A detector is not equally trustworthy at all frequencies: seismic noise dominates below tens of hertz, quantum shot noise above a kilohertz, with a quiet valley between. The optimal filter trusts the detector where it is quiet, dividing each frequency by the noise power $S_n(f)$ there. This whitens the data, flattening coloured noise into featureless white noise, after which a plain correlation is optimal. The first figure shows the pay-off, using the discovery event itself.

Figure

The matched filter in action for GW150914: the chirp buried in the top strain trace is invisible to the eye, yet correlating it against the general-relativity template makes the output ρ(τ) spike to a combined signal-to-noise ratio of 24 at the coalescence time, far above the ρ = 8 detection threshold shown in red

Because the source parameters — the two masses, the two spins, the coalescence time and phase, the distance and orientation — are unknown in advance, one cannot use a single template. The waveform's shape changes with the masses and spins, and a template even slightly off accumulates a phase error over $10^4$ cycles that destroys the coherent buildup. The answer is a template bank: a discrete grid laid down so densely that any real signal, wherever it falls in the continuum, lies close — in the inner-product geometry — to at least one grid point. How close is "close enough" is a design choice, the minimal match, and it fixes both the spacing and, through a Riemannian metric on parameter space, the number of templates. The second figure shows that tiling.

Figure

The intrinsic parameter space of a binary tiled by a template bank: each green ellipse is the patch a single template covers to better than the minimal match of 0.97, and neighbouring cells deliberately overlap so no true signal falls through a gap. A production Advanced-LIGO bank holds roughly 2.5×10⁵ such templates spanning masses and aligned spins

The core equations

Everything begins with the data model: the detector returns the sum of whatever signal is present and its own noise, and the whole task is to decide whether the signal term is there and what its parameters are.

s(t)=h(t)+n(t) \qquad (1)

where: $s(t)$ — the detector strain, the data; $h(t)$ — the astrophysical signal, modelled by a template; $n(t)$ — noise, assumed stationary and Gaussian. Reading: the signal is additive, so "is $h$ present?" is a question about the statistics of $s-h$ versus $s$ alone. The single object carrying the whole theory is a noise-weighted inner product on the space of signals. It divides each frequency by the noise power there and folds the negative frequencies of a real signal onto the positive axis — the origin of the factor $4\,\mathrm{Re}$ . Once it exists, signals are vectors, signal-to-noise is a length, and the matched filter is a projection.

\langle a\,|\,b\rangle = 4\,\mathrm{Re}\int_0^\infty \frac{\tilde a^*(f)\,\tilde b(f)}{S_n(f)}\,\mathrm{d}f \qquad (2)

where: $\tilde a(f)$ — Fourier transform of $a(t)$ ; $S_n(f)$ — one-sided noise power spectral density, defined by $\langle\tilde n^*(f)\tilde n(f')\rangle=\tfrac12 S_n(f)\,\delta(f-f')$ . Reading: the weighting $1/S_n(f)$ whitens the noise, giving it unit variance per mode so a plain Euclidean geometry applies. The detection statistic is the matched-filter signal-to-noise ratio: the inner product of a template with the data, normalised by the template's norm so it has unit variance in pure noise. Its square is what is thresholded.

\rho^2(\theta)=\frac{\langle h(\theta)\,|\,s\rangle^2}{\langle h(\theta)\,|\,h(\theta)\rangle} \qquad (3)

where: $\rho$ — SNR recovered by the template at parameters $\theta$ ; $\langle h|h\rangle$ — template norm-squared, normalising the noise variance to one. Reading: in noise alone $\langle\rho^2\rangle=1$ , so $\rho=24$ is a $24\sigma$ excursion. When the template equals the true signal, the expected SNR takes its largest possible value — simply the norm of the signal in the noise-weighted geometry.

\rho_{\rm opt}=\sqrt{\langle h\,|\,h\rangle}=\left(4\int_0^\infty \frac{|\tilde h(f)|^2}{S_n(f)}\,\mathrm{d}f\right)^{1/2} \qquad (4)

where: $\rho_{\rm opt}$ — expected SNR when the template matches the signal exactly. Reading: the integrand $|\tilde h|^2/S_n$ shows only frequencies where the signal is strong and the detector quiet contribute; the loud bands dominate. Maximising over the unknown arrival time looks expensive, but a time shift multiplies $\tilde h$ by a pure phase $e^{2\pi i f\tau}$ , so a single inverse Fourier transform returns the correlation at all offsets at once; using the complex template makes $|z(\tau)|$ also maximise over the unknown coalescence phase.

z(\tau)=4\int_0^\infty \frac{\tilde h^*(f)\,\tilde s(f)}{S_n(f)}\,e^{2\pi i f\tau}\,\mathrm{d}f \qquad (5)

where: $z(\tau)$ — complex matched-filter output versus trial coalescence time; $|z(\tau)|/\sqrt{\langle h|h\rangle}$ — the phase-maximised SNR at each time. Reading: this is why the output is an SNR time series with a sharp peak, exactly the lower trace of the first figure. To quantify how badly a slightly wrong template performs one uses the match: the normalised overlap of two waveforms, maximised over time and phase, equal to one only when they coincide.

\mathcal{M}(\theta,\theta')=\max_{t_c,\phi_c}\frac{\langle h(\theta)\,|\,h(\theta')\rangle}{\sqrt{\langle h(\theta)|h(\theta)\rangle\,\langle h(\theta')|h(\theta')\rangle}} \qquad (6)

where: $\mathcal{M}$ — the match in $[0,1]$ ; $\theta,\theta'$ — intrinsic parameter vectors of two templates. Reading: $1-\mathcal{M}$ is the fractional SNR lost when the signal sits at $\theta'$ but is recovered with a template at $\theta$ . Near coincidence the match falls off quadratically, and the coefficient of that fall-off is a positive-definite Riemannian metric on parameter space. Divide the metric volume of the search region by the volume one template covers and you get the bank size.

\mathcal{M}\approx 1-g_{ij}\,\Delta\theta^i\,\Delta\theta^j,\qquad N_{\rm templates}\approx \frac{1}{(2\sqrt{(1-\mathrm{MM})/D}\,)^{D}}\int \sqrt{\det g}\;\mathrm{d}^D\theta \qquad (7)

where: $g_{ij}=-\tfrac12\,\partial_i\partial_j\mathcal{M}|_0$ — the metric on intrinsic parameter space; $\mathrm{MM}$ — the minimal match (e.g. 0.97); $D$ — the intrinsic dimension. Reading: denser metric volume means more templates; a higher minimal match shrinks each cell and multiplies the count. Finally, the same inner product organises inference. In stationary Gaussian noise the log-likelihood ratio for "signal $h$ present" versus "noise only" is exactly the matched filter minus half the template energy.

\ln\Lambda=\langle h\,|\,s\rangle-\tfrac12\langle h\,|\,h\rangle \qquad (8)

where: $\ln\Lambda$ — log-likelihood ratio for signal versus noise. Reading: the data-dependent term is precisely $\langle h|s\rangle$ , so the matched filter is the sufficient statistic that Bayesian parameter estimation samples over.

Derivation AWhy the matched filter is the optimal linear detector

The goal is to prove — not assert — that among all linear ways of combining the data into a single detection number, the matched filter gives the largest expected signal-to-noise ratio, and that the proof is nothing deeper than the Cauchy–Schwarz inequality once the noise-weighted inner product is in hand. We take a completely general linear statistic, compute its mean and variance in stationary Gaussian noise, form the SNR, and bound it; the bound is saturated by exactly one filter.

Step 1. Write the most general real linear statistic as $X=\langle q\,|\,s\rangle$ for some filter waveform $q$ ; every linear filter can be cast this way. Why: fixing the competitor class to "all linear statistics" makes the optimality claim precise and provable.

Step 2. Insert $s=h+n$ and take expectations: the signal part gives a mean $\langle X\rangle=\langle q\,|\,h\rangle$ , and because the inner product was built so the noise has unit variance per mode, the noise part gives variance $\sigma_X^2=\langle q\,|\,q\rangle$ . Why: the earlier whitening makes the noise a unit-variance Gaussian per mode, so mean and variance are just projections in the geometry.

Step 3. Form $\rho=\langle X\rangle/\sigma_X=\langle q\,|\,h\rangle/\sqrt{\langle q\,|\,q\rangle}$ and apply Cauchy–Schwarz, $\langle q\,|\,h\rangle\le\sqrt{\langle q\,|\,q\rangle}\,\sqrt{\langle h\,|\,h\rangle}$ , valid because $\langle\cdot|\cdot\rangle$ is positive-definite whenever $S_n>0$ ; dividing gives $\rho\le\sqrt{\langle h\,|\,h\rangle}$ . Why: Cauchy–Schwarz is the geometric statement that a projection is largest when the two vectors are parallel.

Step 4. Equality holds iff $q\propto h$ , so the optimal filter is the template itself — in the frequency domain, correlation against $\tilde h(f)/S_n(f)$ , the Wiener matched filter. Why: this shows the $1/S_n(f)$ weighting is not a modelling choice but a theorem, forced by demanding the best linear SNR.

Go deeper · Derivation B · 1 minSignal-to-noise of a chirp and the detection horizon

Derivation BSignal-to-noise of a chirp and the detection horizon

The aim is to combine the quadrupole-formula waveform of an inspiral with the optimal-SNR integral and read off how loud a source is as a function of chirp mass, distance, and the noise curve — then convert a fixed detection threshold into a horizon distance, the reach of the instrument. This is where the abstract inner product meets astrophysics: it says how far out a given binary can be seen.

Step 1. From the quadrupole formula and the chirp's frequency sweep, the stationary-phase inspiral amplitude is $|\tilde h(f)|\propto (\mathcal{M}^{5/6}/D_L)\,f^{-7/6}$ , with chirp mass $\mathcal{M}=(m_1 m_2)^{3/5}/(m_1+m_2)^{1/5}$ and luminosity distance $D_L$ . Why: the chirp mass is the single mass combination controlling the leading inspiral, which is why it is the best-measured parameter of a low-mass event.

Step 2. Insert into equation (4); the phase drops out of $|\tilde h|^2$ , leaving $\rho_{\rm opt}^2\propto (\mathcal{M}^{5/3}/D_L^2)\int_{f_{\rm low}}^{f_{\rm isco}} f^{-7/3}/S_n(f)\,\mathrm{d}f$ . Why: loudness factorises cleanly into an intrinsic piece (chirp mass), a geometric piece (distance), and a detector piece (the band integral).

Step 3. Read off $\rho_{\rm opt}\propto \mathcal{M}^{5/6}/D_L$ : heavier binaries are louder, and loudness falls as one over distance — the geometric dilution of amplitude. Why: because amplitude scales as $1/D_L$ , the detectable volume scales as $\rho^3$ , which governs event rates.

Step 4. Fix a threshold, conventionally $\rho_{\rm opt}=8$ per detector; the distance at which a fiducial source reaches it is the horizon distance — hundreds of megaparsecs for neutron-star binaries, several gigaparsecs for stellar-mass black holes. Why: this turns a signal-processing threshold into a survey volume, and is why lowering $S_n$ pays off as the cube in event rate.

Go deeper · Derivation C · 1 minTemplate spacing and the size of a bank

Derivation CTemplate spacing and the size of a bank

This shows how the fractional SNR loss from a slightly-wrong template controls both the spacing of a bank and, in the end, the number of templates a search must run — turning a design tolerance into a compute cost. The key is that the match, as a function of the offset from the nearest template, is smooth and peaks at unity, so near its peak it is a quadratic form, and a quadratic form is a metric.

Step 1. The match satisfies $\mathcal{M}(\theta,\theta)=1$ and $\partial_i\mathcal{M}|_0=0$ , since a normalised overlap peaks at coincidence and a smooth maximum has vanishing gradient. Why: with no linear term, the leading mismatch is second order — the regime where a metric description is exact.

Step 2. Expand to second order, $\mathcal{M}\approx 1-g_{ij}\,\Delta\theta^i\Delta\theta^j$ with $g_{ij}=-\tfrac12\partial_i\partial_j\mathcal{M}|_0$ ; this $g_{ij}$ is positive-definite once time and phase are analytically maximised out, so it is a Riemannian metric. Why: promoting the Hessian of the match to a metric turns template placement into a lattice-covering problem in a curved geometry — the second figure.

Step 3. Demand every point lie within the minimal match of some template: $g_{ij}\Delta\theta^i\Delta\theta^j\le 1-\mathrm{MM}$ , fixing the cell size and lattice spacing. Why: the minimal match is the single knob a designer turns — raising it tightens the lattice and buys sensitivity at the price of more templates.

Step 4. Count: the number of templates is the metric volume $\int\sqrt{\det g}\,\mathrm{d}^D\theta$ divided by one cell's covering volume, equation (7). For $\mathrm{MM}=0.97$ over compact-binary mass–spin space this yields $10^4$ – $10^6$ templates. Why: this closes the loop from an abstract inner product to the concrete size of the bank a pipeline filters the data against.

Limiting cases

Case	Result	Meaning
White noise	$S_n(f)=S_0=\text{const}$	The inner product becomes a plain time correlation $\langle a\|b\rangle\propto\int a(t)b(t)\,\mathrm{d}t$ ; the optimal filter is a time-reversed template — the classical radar matched filter.
Exact match	$\rho_{\rm opt}=\sqrt{\langle h\|h\rangle}$	Template equals the true signal; SNR attains its maximum expectation and no linear statistic can do better.
Small mismatch	$\mathcal{M}=1-g_{ij}\Delta\theta^i\Delta\theta^j$	The match falls off quadratically near a maximum; the metric $g_{ij}$ sets the template spacing and bank size.
Unmodelled burst	$\rho\sim h_{\rm rms}/\sqrt{S_n(f_0)\Delta f}$	With no reliable template, matched filtering gives way to excess-power and wavelet methods; the coherent many-cycle gain is lost.
Non-Gaussian noise	glitches inflate $\rho$	The optimality proof assumes stationary Gaussian noise; transient glitches break it, so a high $\rho$ must be vetoed before it counts.

Common trapA high matched-filter SNR by itself is a detection

It is not. The optimality theorem assumes stationary, Gaussian noise, and real interferometers violate both. They ring with non-Gaussian transients called glitches — from scattered light, the suspensions, anthropogenic disturbances — that can correlate strongly with a chirp template and produce a large $\rho$ with no astrophysical cause. A raw SNR of 10 in one detector, unaccompanied, is not evidence of a wave. Real searches add three layers on top of the bare filter: a signal-consistency test (the $\chi^2$ time–frequency test, checking that the SNR accumulated cycle by cycle as a real chirp would, not in one glitchy burst); multi-detector coincidence with consistent arrival times and parameters; and an empirical false-alarm rate estimated by time-sliding the detectors to build the background. GW150914's significance came not from $\rho=24$ alone but from that $\rho$ surviving all three with a false-alarm rate below one per 200 000 years. The opposite trap is a slightly wrong template: a phase error of a fraction of a cycle, accumulated over $10^4$ cycles, dephases the sum and can cost most of the SNR — which is why post-Newtonian and numerical-relativity waveforms accurate to a small fraction of a radian are indispensable.

Worked examples

1 · GW150914 — chirp mass and how far it could have been seen (fully worked)

GW150914 was a binary black hole with source-frame masses $m_1=36\,M_\odot$ , $m_2=29\,M_\odot$ , recovered at combined SNR $\rho=24$ at luminosity distance $D_L\approx410$ Mpc. Find its chirp mass and estimate the horizon distance at which such a system would just cross the single-detector threshold $\rho=8$ .

Step 1. Chirp mass from $\mathcal{M}=(m_1 m_2)^{3/5}/(m_1+m_2)^{1/5}$ : the product is $36\times29=1044\,M_\odot^2$ and the sum $65\,M_\odot$ , so $\mathcal{M}=1044^{0.6}/65^{0.2}=64.7/2.30=\mathbf{28.1\ M_\odot}$ . This single number, not the individual masses, is what the leading inspiral measures most tightly.

Step 2. Since $\rho_{\rm opt}\propto 1/D_L$ at fixed intrinsic parameters, the horizon is $D_{\rm hor}=D_L\,(\rho/\rho_{\rm thr})=410\times(24/8)=\mathbf{1230\ Mpc\approx1.2\ Gpc}$ .

Step 3. Sanity check: at twice the true distance, 820 Mpc, the SNR would halve to $24\times(410/820)=\mathbf{12}$ — still a confident detection. Only near 1.2 Gpc does $\rho$ fall to 8, so GW150914 sat at one-third of its horizon, hence it was so loud.

Step 4. Because detectable volume scales as $D_{\rm hor}^3\propto\rho^3$ , tripling the reach corresponds to $3^3=27$ times the accessible volume. Chirp mass $28.1\,M_\odot$ ; horizon $\approx1.2$ Gpc; source at one-third horizon gives $\rho=24$ .

2 · The price of a discrete bank

A bank is built to minimal match $\mathrm{MM}=0.97$ . Find (a) the worst-case fractional loss in recovered SNR, and (b) the resulting loss in detection rate. For context, estimate how many cycles a $1.4+1.4\,M_\odot$ binary-neutron-star inspiral spends above 10 Hz.

Step 1. A signal at the worst point of a cell is recovered with $\mathcal{M}=0.97$ , so the SNR is $0.97\,\rho_{\rm opt}$ : a fractional loss of $1-0.97=\mathbf{0.03}$ , i.e. 3 %. That barely-noticeable loss is exactly why 0.97 is the standard choice.

Step 2. Rate follows the cube law: the surviving fraction is $(0.97)^3=0.913$ , a worst-case rate loss of $1-0.913=\mathbf{0.087}$ , i.e. about 8.7 % (less on average, since most signals land nearer a template than the cell corner). This is why banks are not pushed to 0.99 — the gain is marginal while the template count and compute cost climb steeply.

Step 3. For a canonical BNS, the Newtonian estimate $N_{\rm cyc}\approx\frac{1}{32\pi}(\pi G\mathcal{M}f_{\rm low}/c^3)^{-5/3}$ with $\mathcal{M}=1.22\,M_\odot$ and $f_{\rm low}=10$ Hz gives $N_{\rm cyc}\approx\mathbf{1.6\times10^4}$ cycles. The matched filter adds them in phase, a coherent gain of order $\sqrt{N_{\rm cyc}}\approx125$ over a single cycle.

Step 4. By contrast GW150914's large total mass merges at low frequency (point-particle ISCO $\approx68$ Hz), so it spent only tens of cycles in band — yet was loud because its amplitude $\propto\mathcal{M}^{5/6}$ is enormous. Worst-case 3 % SNR loss, about 8.7 % rate loss; a BNS banks $\sim1.6\times10^4$ cycles, a heavy BBH only tens.

Physics of the portrait

The matched filter is where information theory, radar engineering, and general relativity meet on one equation: born in wartime electronics, matured into the language of optimal estimation, and finding its most spectacular application waiting in the physics of colliding black holes.

Origins · 1 minmatched filtering and how this was discovered

matched filtering

the optimal linear detector for a known signal buried in noise · gravitational-wave data analysis

Matched filtering was not discovered by an astronomer or a relativist — its origin is radar. During the Second World War, engineers needed to decide whether a faint, known pulse was hidden in a noisy receiver trace, and in 1943 Dwight O. North, at RCA, wrote down the linear filter that maximises the peak signal-to-noise ratio for a known waveform in white noise — one whose impulse response is the time-reversed signal, later universally called the "matched filter" (North's report circulated for two decades before its 1963 reprint in the Proceedings of the IEEE). In parallel, Norbert Wiener's theory of optimal linear filtering, published in 1949, gave the idea its rigorous statistical home in the geometry of stationary random processes; the noise-weighted inner product here is Wiener's whitening made concrete, which is why the optimal filter is called the Wiener filter. The transplant into gravitational-wave physics came in the early 1990s: Sathyaprakash and Dhurandhar (1991) recognised that the coalescence waveforms predicted by the quadrupole formula were precisely the "known signals" matched filtering was built for, and worked out how to choose filters for a chirp of unknown parameters. Benjamin Owen (1996) then supplied the geometric piece — the metric on parameter space and the minimal match — that turned "filter against the template" into the rigorous covering problem of a template bank. The synthesis came of age on 14 September 2015, when the LIGO detectors recorded GW150914 and the pipelines returned a combined SNR of 24: a wartime radar idea, dressed in the differential geometry of signal space, reading off the collision of two black holes 1.3 billion years ago.

History

1943

— North formulates the matched filter for radar detection (the North filter)

1949

— Wiener publishes the theory of optimal linear filtering of stationary processes

1991

— Sathyaprakash and Dhurandhar cast the inspiral search as matched filtering against chirp templates

1996

— Owen introduces the parameter-space metric and the minimal-match template bank

2015

— GW150914 is recorded and recovered by matched filtering at combined SNR 24

Going further

Texts. Misner, Thorne and Wheeler, Gravitation (1973), §36–37 · Wald, General Relativity (1984), Ch. 4 (linearised gravity) · Maggiore, Gravitational Waves, Vol. 1 (2008), Ch. 7 — the primary reference here · Creighton and Anderson, Gravitational-Wave Physics and Astronomy (2011), Ch. 7 · Jaranowski and Królak, Analysis of Gravitational-Wave Data (2009). Key relations used throughout: the optimality bound $\rho\le\sqrt{\langle h|h\rangle}$ follows from Cauchy–Schwarz; the PSD is normalised by $\langle\tilde n^*(f)\tilde n(f')\rangle=\tfrac12 S_n(f)\delta(f-f')$ , which is the origin of the $4\,\mathrm{Re}$ factor in the inner product; the stationary-phase inspiral amplitude is $|\tilde h|\propto\mathcal{M}^{5/6}f^{-7/6}$ , giving $\rho_{\rm opt}^2\propto\mathcal{M}^{5/3}/D_L^2$ ; and the match metric and bank-size formula follow from the quadratic expansion of the overlap. The GW150914 numbers — masses $36+29\,M_\odot$ , chirp mass $28.1\,M_\odot$ , combined SNR 24, $D_L\approx410$ Mpc, false-alarm rate below one per 200 000 years — are from Abbott et al. (LIGO/Virgo) PRL 116, 061102 (2016). The minimal match 0.97 and $10^4$ – $10^6$ bank range are standard practice. The BNS cycle count $\sim1.6\times10^4$ above 10 Hz and the GW150914 point-particle ISCO $\approx68$ Hz are computed from the Newtonian expressions.

Prerequisites. Quadrupole Formula · Multipole Expansion of Gravitational Radiation · Plus and Cross Polarizations · Gravitational Wave Equation. Leads to. Isaacson Stress-Energy of Gravitational Waves · Helicity and Spin-2 Nature of Gravitons.

Ask the tutor: Want to work through how the parameter-space metric $g_{ij}$ is computed from a post-Newtonian phase, or how a $\chi^2$ veto tells a real chirp from a glitch? Ask and we will build it step by step.

Is the matched filter really the "best possible" detector, or just the best simple one?

For stationary Gaussian noise it is genuinely the best, in two senses. Among all linear statistics it maximises the expected signal-to-noise ratio — the Cauchy–Schwarz theorem of this article. More strongly, in Gaussian noise the matched-filter output is a sufficient statistic, so by the Neyman–Pearson lemma the most powerful test at any false-alarm rate depends on the data only through $\langle h|s\rangle$ ; no other statistic, linear or not, can do better given the assumptions. The catch is entirely in those assumptions: real noise is neither perfectly stationary nor Gaussian, which is why practical pipelines bolt consistency and coincidence tests onto the optimal core rather than trusting it blindly.

Why does the noise appear as a power spectral density rather than a single number?

Because detector noise is coloured — its power depends strongly on frequency. Seismic and suspension noise dominate at low frequencies, quantum shot noise at high, with a sensitive valley of tens to a few hundred hertz between, spanning several orders of magnitude in power. A single noise number would throw away exactly the information that makes matched filtering optimal: which frequencies to trust. The one-sided power spectral density $S_n(f)$ encodes the noise variance per unit frequency, and dividing by it in the inner product weights each frequency by how quiet the detector is there. Where $S_n$ is small the detector is trustworthy and the signal counts heavily; where it is large it is discounted.

What is the difference between intrinsic and extrinsic parameters?

Intrinsic parameters change the shape of the waveform as a function of time — the two masses and two spin vectors. Extrinsic parameters change only its overall scaling, timing, or orientation — coalescence time and phase, amplitude (distance and inclination), and sky position. The distinction is computational: the extrinsic ones are handled almost for free, since maximising over arrival time is a single inverse FFT and over phase and amplitude is analytic through the complex matched filter. Only the intrinsic parameters need an actual grid of templates. This is why a template bank is described as covering "masses and spins" even though the full parameter space is much larger — the rest is maximised away analytically.

How can a bank of a hundred thousand templates run in real time?

Through a stack of tricks on the basic filter. First, the matched filter is computed in the frequency domain, where a correlation at every trial time is one multiplication and one inverse FFT — an $N\log N$ operation, not $N^2$ . Second, the extrinsic maximisations are free or analytic, so the bank spans only the intrinsic dimensions. Third, templates are ordered so neighbours share computation, and modern pipelines exploit GPUs, multi-banding (shorter segments at higher frequencies), and singular-value or reduced-order decompositions that compress thousands of templates into a handful of basis waveforms. The result is that hundreds of thousands of templates can be filtered against several detectors with a latency of seconds.

Why does a small phase error ruin the SNR when a small amplitude error does not?

Because the coherent gain comes from adding many cycles in phase. The SNR accumulates as a sum over $10^3$ – $10^4$ cycles that is large only if the template stays in step the whole way. If the phase drifts — say the chirp mass is slightly wrong — the accumulated error grows cycle by cycle, and once it reaches about a radian the later contributions subtract instead of add, collapsing the sum. An amplitude error merely rescales each term and costs SNR only mildly and linearly. This asymmetry is why waveform models match the true phase to a small fraction of a radian over the whole inspiral, and why the chirp mass — which controls the leading phase — is the best-measured parameter of a low-mass event.

Does matched filtering assume the waveform is exactly right? What if general relativity is wrong?

It assumes an accurate template for the model you are testing. If the true signal departs from the template — a modified theory, or unmodelled tidal or environmental physics — the recovered SNR drops and the fit worsens. Far from a weakness, this is the basis of the field's tests of gravity: one searches with GR templates, and any systematic mismatch between the recovered and predicted phase evolution is a bound on deviations (extra dispersion, a graviton mass, dipole radiation). Searches use banks broad enough to catch what GR predicts, and dedicated tests then probe for residuals. So the method neither assumes GR is right nor fails silently if it is wrong — it turns the degree of match into a measurement.