Many physical searches fail because each event is too faint to see on its own, not because the instrument is poorly calibrated. A standard fix is stacking: add up many copies of the same weak signal so the signal grows faster than the noise. The best-known gravitational-wave example lines up detector data on the arrival times of gamma-ray flares from a magnetar and adds the data together. Over 18 bursts, this gains about an order of magnitude in energy sensitivity over a single search covering the whole storm (Stacking gravitational wave signals from soft gamma repeater bursts). That method depends entirely on bright external flares to say where each copy sits in time. If the flares are missing, the method has nothing to align on, and the faint population stays invisible.
We propose to stack without the external trigger. The alignment, the shape of the shared signal, and which candidate windows contain any signal at all are all estimated from the data. The main risk is well known and easy to underrate. When you align pure noise to a template learned from that same noise, you get a convincing-looking template out, and the resulting "detection" is a product of the optimization. Our central contribution is therefore a way of assigning significance that accounts for this self-alignment. The null distribution is built by running the entire pipeline, including trigger selection and alignment, on data known to contain no signal. We then measure how much sensitivity survives once significance is computed honestly. If the approach works, sources of repeating signals with unknown or poorly known shape become searchable even when every individual event sits below threshold and no outside timing marker exists.
Core Claim
A latent-alignment stack is a mixture model. In it, each candidate window either contains a time-shifted, rescaled copy of an unknown shared waveform or contains only noise. Such a stack can recover most of the sensitivity of trigger-aligned incoherent stacking, provided its detection statistic is calibrated against background realizations that pass through the same alignment optimization. Without this calibration, the stack reports false detections on pure noise. With it, the sensitivity lies strictly between single-event search and oracle-triggered stacking, and we want to measure exactly where.
For each candidate window $i = 1,\dots,N$ (a short segment of detector data flagged by a cheap first-pass detector), we model the data as
$$y_i = z_i \, a_i \, S_{\tau_i} x + \epsilon_i,$$
where $x$ is the unknown shared waveform, $S_{\tau_i}$ shifts it by an unknown offset $\tau_i$, $a_i$ is an unknown per-burst amplitude, $z_i \in \{0,1\}$ says whether window $i$ holds signal at all, and $\epsilon_i$ is detector noise. The model assigns $z_i = 1$ with probability $\pi$. The stacked statistic $\Lambda$ is the log likelihood ratio, maximized over $x$, $\pi$, and the per-window latent variables, of this model against $\pi = 0$.
Because $\Lambda$ is maximized over a flexible waveform, the textbook chi-square asymptotics do not hold. Its null distribution must be estimated empirically.
Positioning Against Prior Work
The method draws on two sources. The alignment machinery comes from the Continuous Profile Model (Multiple Alignment of Continuous Time Series). That model aligns replicate speech and chromatography series to a latent trace using a hidden Markov model trained by expectation-maximization. It assumes every replicate contains the signal, and it never asks whether the recovered trace is real. We add the presence variable $z_i$ and a detection test on top of it.
The motivating result is the magnetar stacking paper cited above. It compares a coherent time-series stack against an incoherent spectrogram stack. The coherent stack has a strain scaling of $N^{-0.55}$ but needs timing errors of at most 100 μs. The incoherent stack scales as $N^{-0.27}$ and tolerates millisecond jitter. Both rely on flare times. Our model interpolates between the two: shifts are estimated rather than given, and phase can be treated either as aligned or as marginalized.
Two further results constrain the design. First, neural trigger generators flag candidates down to a signal-to-noise ratio of 7.79, but at about one false alarm per 19.5 minutes, so their scores are not significances (Convolutional neural networks: A magic bullet for gravitational-wave detection?). Any candidate catalog we stack is therefore mostly noise, and $\pi$ must be allowed to be small. Second, when the reference pool used to calibrate a threshold is contaminated with real signals, the threshold becomes conservative and detection power drops (Robust Conformal Outlier Detection under Contaminated Reference Data). Off-source background data may contain real sub-threshold bursts, so we expect exactly this power loss. That paper fixes contamination by having experts label a few points. Physically, nobody can label a sub-threshold fluctuation, so our remedy has to be different: we build the null from time-shifted data between the two detectors.
Two further papers show that aggregating before inference beats per-object estimation at low signal-to-noise ratio. These are the lensing challenge, where stacking Fourier power spectra reached $Q \approx 211$ against $Q \approx 10$–$100$ for per-galaxy fitting (Results of the GREAT08 Challenge★), and the proof that sample-mean stacking restores convergent regularization when the noise level is unknown (Beyond the Bakushinkii veto). Both papers assume that the objects being stacked are real. Ours does not.
Research Program
We will first build the estimator. It uses expectation-maximization over the waveform $x$, the signal fraction $\pi$, and per-window posteriors over $(z_i, \tau_i, a_i)$. The shift posterior is computed on a grid, which is cheap because shifts are one-dimensional. We add a smoothness penalty on $x$ and whiten the data per detector. For two-detector data we tie the offsets across detectors to the physically allowed light-travel window.
We will then build the null. Background realizations come from time slides, in which one detector's data is shifted by far more than the light-travel time before candidate selection. Every background realization goes through the full pipeline: first-pass triggering, window extraction, expectation-maximization with the same restarts, and computation of $\Lambda$. The p-value is the rank of the on-source $\Lambda$ among these realizations. We will also quantify contamination bias by injecting real-looking sub-threshold bursts into the background pool.
Third, we will map the sensitivity trade-off by varying three factors: $N$, the true fraction of windows with signal, and cycle-to-cycle variation in the waveform. This last factor tests how the method behaves when the signal's shape changes from burst to burst, which the magnetar paper assumes away. For reference, known-alignment coherent stacking gives strain sensitivity $\propto N^{-1/2}$, because the signal adds linearly while the noise adds in quadrature. Incoherent power stacking gives $\propto N^{-1/4}$, because power sensitivity scales as $N^{-1/2}$. We expect the latent-alignment stack to approach the incoherent rate when the shift uncertainty is large relative to a period, and to exceed it when the shifts are well localized.
Baselines
The do-nothing reference is the single-window search: excess-power detection on each candidate window, with the loudest window taken and a trials correction applied for the number of windows. This is the regime the magnetar paper escaped.
The strongest existing method is the incoherent spectrogram stack (P-Stack) from Stacking gravitational wave signals from soft gamma repeater bursts, run with the true flare times. It is an oracle, because it is given the alignment information our method must infer. On the mock SGR 1900+14 storm (18 bursts, a 1090 Hz ringdown decaying over 100 ms), that paper reports a 90% confidence isotropic energy sensitivity of about $10^{48}$ erg at 10 kpc. We will re-run it on identical injections so that both methods face the same noise. We will run the coherent T-Stack the same way, as the upper oracle.
The realistic competitor is P-Stack using first-pass detector times in place of the flare times. The first-pass detector is a fully convolutional network in the style of Convolutional neural networks: A magic bullet for gravitational-wave detection?, trained on its injection set in real O1 noise.
All methods will see the same windows, the same background time slides, and the same false-alarm target, and will be tuned on the same held-out injection set. Two ablations isolate our ingredients. The first removes the pipeline-level null and uses the asymptotic chi-square instead, to show the size of the self-alignment inflation. The second fixes $\pi = 1$, as in the original Continuous Profile Model, to show the cost of assuming every window holds signal.
Beating the realistic competitor means a lower 90% energy sensitivity at equal, empirically verified false-alarm probability. Reaching within a factor of about two of the oracle's $10^{48}$ erg would count as success.
Why not just…?
Why not just use P-Stack at the neural trigger times? Most of those triggers are false alarms, and the true ones are timed only to the trigger network's resolution. Summing all of them dilutes the stack with noise windows and blurs the alignment, whereas our model down-weights windows that look like noise and refines the shifts. This alternative is our realistic-competitor baseline, so the experiment settles the question directly. If it matches us once trigger timing is good, the lesson is that the pipeline-level null, which it also needs, is the real contribution.
Why not just matched-filter each candidate and sum the squared signal-to-noise ratios? When the waveform is known, this is close to optimal, and we concede that case. Our target is repeating signals with unknown or burst-varying shape, where no template bank exists. We will run a template-bank sum on the ringdown injections, where it has an unfair advantage because the template is correct. We will also run it on injections whose shape differs from the bank, to show where it breaks.
Why not just train a network that takes a whole set of windows and outputs a detection score? This is a genuine threat, because such a network could learn alignment implicitly. It faces the same calibration problem, though: its score has no known null distribution, and it can be fooled by noise glitches. We will train a permutation-invariant set network on the same windows and give it the same time-slide null. If it matches our method, the paper's claim reduces to the calibration procedure, and we would report that honestly.
Decisive Checks
The first test runs on pure time-slid O1 noise with no injections. Here the pipeline-level p-values must be uniform, while the asymptotic ablation should show inflated false-detection rates. Showing this inflation is the core empirical finding. If it does not appear, the calibration machinery is unnecessary.
The second test is the SGR 1900+14 mock storm with the flare times removed. It reports 90% energy sensitivity for four methods: our stack, P-Stack at trigger times, the P-Stack oracle, and the single-window search.
The third test is a sweep over timing jitter from 100 μs to tens of milliseconds, the signal fraction $\pi$, and waveform variability. It traces where the latent-alignment stack sits between the $N^{-1/2}$ and $N^{-1/4}$ scalings, and measures the contamination-induced loss of power when real bursts leak into the background pool.
References
- Stacking gravitational wave signals from soft gamma repeater bursts
- Multiple Alignment of Continuous Time Series
- Convolutional neural networks: a magic bullet for gravitational-wave detection?
- Robust Conformal Outlier Detection under Contaminated Reference Data
- Results of the GREAT08 Challenge★: an image analysis competition for cosmological lensing
- Beyond the Bakushinkii veto: regularising linear inverse problems without knowing the noise distribution