Burst and Excess-Power Searches

An excess-power search asks whether calibrated data contain more energy than the noise model predicts in some time–frequency region, without requiring a precise phase-evolution template. This broadens sensitivity to short signals and modeling surprises, but it replaces template mismatch by choices of transform, clustering, duration, bandwidth, network consistency, and background. This page derives the ideal statistic, shows how clustering changes the trials problem, and supplies an executed chi-square control with software and search configurations dated to 2026-08-18.

Required background. Probability and Statistical Inference supplies chi-square distributions, tail probabilities, multiple testing, and Monte Carlo error.

Helpful background. Gravitational Radiation and Waveforms supplies burst morphologies and polarization content, while Computational Relativity supplies simulated transient signals and validation principles.

Energy in a whitened subspace

Select a calibrated data segment, estimate its one-sided PSD, and whiten it so that ideal noise coefficients have unit covariance. Let $x_j$ be $k$ independent real coefficients spanning a prescribed time–frequency region $R$. The excess-power statistic is

$$E_R=\sum_{j=1}^{k}x_j^2.$$

Under zero-mean Gaussian noise,

$$E_R\sim\chi^2_k, \qquad \mathbb E[E_R]=k, \qquad \operatorname{Var}(E_R)=2k.$$

The statistic follows directly from a generalized likelihood ratio. If the signal coefficients $s_j$ inside this fixed subspace are otherwise unrestricted, then

$$\ln\Lambda(\boldsymbol s) =\sum_j\left(x_js_j-\frac12s_j^2\right).$$

Maximizing gives $\widehat s_j=x_j$ and $2\ln\Lambda_{\max}=E_R$. This explains both the appeal and the cost: every direction within the chosen subspace can be fitted, including noise, so the null mean grows with its dimension.

If a deterministic signal contributes whitened coefficients $s_j$, the statistic is noncentral chi-square with noncentrality

$$\lambda=\sum_j s_j^2,$$

which is the squared signal-to-noise ratio captured by that subspace. Anderson and collaborators derive the excess-power statistic and its time–frequency implementation in Anderson et al. 2001, §§II–IV, PDF.

The degrees of freedom must match the representation. A complex Fourier pixel contributes two real quadratures unless symmetry or overlap removes one. Windowed and overlapping transforms correlate neighboring pixels, so simply counting displayed pixels can give the wrong null distribution. The effective statistic is then calibrated from noise realizations or off-source data using the exact transform and clustering rule.

Unknown duration and bandwidth create a search

A single fixed region has one chi-square tail. A burst search examines many start times, durations, frequencies, bandwidths, resolutions, and cluster shapes. Pixels may be thresholded and joined by adjacency; coherent network projections may separate allowed signal energy from null energy. The loudest cluster is therefore an order statistic of a correlated search, not a draw from the one-region distribution.

A transparent search record specifies:

• segment and quality-mask selection;
• calibration and PSD versions;
• window, transform, overlap, and whitening filter;
• pixel threshold and cluster connectivity;
• duration, frequency, sky, and polarization domains;
• coherent, null, and detector-balance cuts;
• category definitions and background construction; and
• injection family used to measure efficiency.

Changing any of these can change both the foreground statistic and the background tail. The detector-noise guide explains why whitening, line treatment, glitches, and data-quality selections cannot be separated from transient extraction LIGO Scientific Collaboration and Virgo Collaboration 2020, §§4–6 and 8, PDF.

A fixed-region chi-square benchmark

Choose $k=8$ independent real whitened coefficients and set the threshold at the analytic 99th percentile,

$$E_*=F^{-1}_{\chi^2_8}(0.99)=20.090235\ldots .$$

The expected single-region tail probability is $0.01$. With Python 3.9.6, NumPy==2.0.2, SciPy==1.13.1, and seed 271828, the measured tail and mean converge as follows:

Background trials	Tail probability result	Mean-energy result
10000	0.008300	7.981330
100000	0.010020	7.992019
1000000	0.009999	8.001905

The million-trial estimate recovers both the 1% tail and mean eight. This validates the fixed independent-pixel calculation only. A clustered search needs its own background because overlapping regions and data conditioning violate the benchmark’s independence assumptions.

Executable rerun

import numpy as np
from scipy.stats import chi2

degrees = 8
threshold = chi2.ppf(0.99, degrees)
print(f"threshold={threshold:.9f}")
for trials in (10_000, 100_000, 1_000_000):
    rng = np.random.default_rng(271828)
    samples = rng.normal(size=(trials, degrees))
    energy = np.sum(samples**2, axis=1)
    tail = np.mean(energy >= threshold)
    print(trials, f"{tail:.6f}", f"{np.mean(energy):.6f}")

Save this as excess_power_null_check.py. The clean command sequence is:

python3 -m venv .venv
. .venv/bin/activate
python3 -m pip install numpy==2.0.2 scipy==1.13.1
python3 excess_power_null_check.py

An end-to-end public-data rerun must additionally pin the GWOSC file hashes, detector and GPS interval, quality category, calibration statement, sample rate, PSD interval, transform code, clustering configuration, software commit, random seed, and off-source livetime. The analytic test should be rerun after each transform or normalization change.

Coherent network bursts

For a sky position, whitened network data can be projected onto the two-polarization response subspace. Coherent energy tests the part consistent with a gravitational wave; null energy tests the orthogonal part. Searching over sky position changes the response matrix and adds trials. A linearly polarized or elliptically constrained model reduces degrees of freedom but can lose signals outside that polarization family.

Coincident excess power is less dependent on calibration phase but uses network information incompletely. A coherent statistic gains discrimination when timing and responses are accurate, yet a correlated environmental transient or a calibration error can mimic or spoil coherence. Detector-balance and null tests therefore supplement rather than replace empirical background.

Sensitivity, evidence, and failure modes

At the evidence cutoff, the GWTC-5.0 methods paper documents the versioned O4b burst-search configurations, sensitivity studies, candidate checks, and waveform sets used for injections LVK Collaboration 2026, §§2–4, PDF. Those choices should not be converted into timeless duration bands, thresholds, or performance claims.

A burst search fails its stated interpretation if whitening leaves coherent lines, gating erases part of a signal, cluster tuning uses the on-source event, time shifts retain terrestrial correlations, the simulated morphologies do not cover the claimed source class, or a high-dimensional trials factor is estimated from too little background. Sensitivity must be reported conditional on waveform morphology, amplitude definition, sky and orientation distribution, detector state, and threshold.

The durable conclusion is methodological: excess power is the generalized likelihood-ratio statistic for an unrestricted deterministic signal in a fixed subspace with known Gaussian covariance. That does not establish a universally most powerful test for every burst population. The representation, polarization constraints, and cluster family define which signals are easy to collect.

Check your understanding

For $k=8$, a preselected region has $E_R=25$. The ideal single-region tail is $P(\chi^2_8\ge25)\simeq1.55\times10^{-3}$. Why is that not the false-alarm probability of the loudest cluster in a full burst search?

Solution

The quoted tail applies to one region chosen independently of the data, with eight independent unit-variance Gaussian coefficients. A search selects the loudest among many correlated regions, resolutions, sky points, and categories. Windowing and clustering can also change the effective covariance and degrees of freedom. The search false-alarm probability must therefore be measured for the maximum of the complete pipeline statistic using the same conditioning, clustering, vetoes, and trials—not obtained from the single-region chi-square tail.

References

• Anderson, Warren G., Patrick R. Brady, Jolien D. E. Creighton, and Éanna É. Flanagan. “Excess Power Statistic for Detection of Burst Sources of Gravitational Radiation.” Physical Review D 63 (2001): 042003. DOI. Open PDF.
• LIGO Scientific Collaboration and Virgo Collaboration. “A Guide to LIGO–Virgo Detector Noise and Extraction of Transient Gravitational-Wave Signals.” Classical and Quantum Gravity 37 (2020): 055002. DOI. Open PDF.
• LIGO Scientific Collaboration, Virgo Collaboration, and KAGRA Collaboration. “GWTC-5.0: Methods for Identifying and Characterizing Gravitational-Wave Transients.” arXiv:2605.27224 [gr-qc] (2026). Open PDF.