Sequential Burst Vetoes for Ionization-Only Dark Matter Searches

The faintest dark matter signals in liquid-xenon detectors are recorded in "ionization-only" mode. In this mode the detector keeps events that show only the delayed electron signal and no prompt flash of light, which lets it see recoils far too small to produce scintillation. The price is that the detector loses the event's depth coordinate, so it can no longer reject spurious electrons emitted from the cathode or produced in micro-discharges, which are small electrical breakdowns in the detector. In Search for Light Dark Matter with 259 Days of Data in PandaX-4T, the micro-discharge background model is named as a main source of systematic uncertainty on light-mediator limits. These artifacts have a known statistical signature. They arrive in bursts that switch on, persist, and switch off, while a dark matter signal would be a tiny excess rate that stays constant over the whole exposure.

Covert-communication theory has a detailed account of this distinction between bursty and stationary signals. LPD Communication: A Sequential Change-Point Detection Perspective models an eavesdropper who watches a stream in real time with sequential change-point tests. These are recursive statistics such as CUSUM (cumulative sum) and Shiryaev–Roberts, which accumulate evidence sample by sample and raise an alarm when it crosses a threshold. The paper shows that a transmitter who sends a short, loud burst can finish before the alarm fires, because these tests need time to accumulate evidence. The square-root law, as stated in Capacity Region for Covert Secret Key Generation over Multiple Access Channels, says the opposite case also holds: a weak signal spread uniformly over the whole observation is the hardest thing for any detector to notice. We propose to apply both results to detector data. The experimenter plays the eavesdropper, and micro-discharges play the burst transmitter. A sequential veto then removes bursts while leaving a stationary dark matter excess almost untouched. The detection-delay theory also predicts how many burst events leak through before the alarm fires. That prediction is the new ingredient, because it supplies a constrained, data-driven residual background in place of the template that currently dominates the systematics.

Core Claim

A CUSUM-type veto run on the time-ordered stream of ionization-only events should do two things. It should remove micro-discharge bursts with a controlled loss of exposure, and its alarm statistics should yield a calibrated prediction of the residual burst leakage. Together, these should tighten light-mediator limits relative to the static cuts plus nuisance-template approach of PandaX-4T. The improvement should show up mainly as a smaller systematic uncertainty, not only as a lower background.

To make this concrete, treat the selected events as a point process with quiescent rate $\lambda_0$ and burst rate $\lambda_1$. At the $k$-th event, with waiting time $\Delta t_k$ since the previous event, the statistic updates as

$$W_k = \max\big(0,\; W_{k-1} + \log(\lambda_1/\lambda_0) - (\lambda_1-\lambda_0)\,\Delta t_k\big),$$

and the veto opens when $W_k \ge h$. The update is a log-likelihood-ratio increment for a Poisson process, and the full statistic could also include per-event features such as electron count and position. Standard quickest-detection results give two consequences. The mean time between false alarms on quiescent data grows roughly as $e^{h}$, which fixes the exposure lost. The mean delay before a true burst is flagged is about $h$ divided by the per-unit-time divergence between the two rates. The expected number of leaked burst events per burst is therefore approximately

$$N_{\text{leak}} \approx \lambda_1 \, \frac{h}{\lambda_1\log(\lambda_1/\lambda_0) - (\lambda_1-\lambda_0)}.$$

This plays the role of "Case 3" in the covert-communication paper, a burst that finishes before detection. It is the quantity we will validate against data.

Research Program

The work has three parts. The first part, which supplies the method, adapts the sequential tests of LPD Communication from Gaussian samples to marked Poisson event streams. CUSUM targets worst-case delay and Shiryaev–Roberts targets average delay. We will compare both, because micro-discharge bursts vary in onset sharpness.

The second part characterizes the trade-off between kept exposure and residual background. This is the detector analogue of the covertness-throughput curve: the threshold $h$ trades false alarms (exposure thrown away) against missed detections (bursts that leak through). We will derive the leakage prediction and test it on sidebands and calibration periods, where bursts can be identified without unblinding the signal region.

The third part sets limits. Sequentially vetoed data, with the leakage-constrained residual term, will be fed into the same profile-likelihood machinery PandaX-4T uses. We will then measure how the limit moves at fixed exposure.

The other paper in this cluster bounds what to expect. Unfolding work such as High-fidelity unfolding of 252Cf neutron spectra shows that calibrated response modeling is effective offline. We do not replace that kind of modeling. We add a time-domain layer, which offline spectral methods cannot see.

Baselines

The naive reference point is the ionization-only analysis with no time-domain treatment beyond standard quality cuts, with the micro-discharge rate floated as a nuisance parameter. Every veto's gain is measured relative to this.

The strongest existing method is the PandaX-4T analysis itself (Search for Light Dark Matter with 259 Days of Data in PandaX-4T). It applies cut-based filtering and an empirical micro-discharge background model across 1.04 tonne·year of ionization-only data. Its reported heavy-mediator electron-scattering limit at 200 MeV/c² is 1.5 × 10⁻⁴¹ cm², and its spin-independent limit at 3 GeV/c² is 1.1 × 10⁻⁴³ cm². No reported result was found on Lacuna for the light-mediator numerical limit; the value is truncated in the file. No reported result was found on Lacuna for micro-discharge veto efficiency or exposure loss.

We will rerun the baseline pipeline on identical data and exposure, with identical likelihood and signal models, and we will tune its cut thresholds on the same sidebands with the same tuning budget used to choose $h$. Beating the baseline means a lower median expected limit for the light mediator, together with a narrower uncertainty band attributable to the micro-discharge term, at no more than a few percent of exposure lost.

The time-window veto described under "Why not just…?" below is a required baseline.

The key ablation keeps the sequential veto but drops the leakage prediction, so the residual background is modeled with the original free template. If the gains disappear under this ablation, the delay theory is doing the work. If they persist, plain burst removal is.

Why not just…?

Why not just veto a fixed time window after any event cluster? This is the most serious competitor, and it may capture most of the background rejection. A fixed window responds to rate exceedances in a single bin, which makes it a Shewhart test in the language of LPD Communication. That paper shows such tests are slow on gradual onsets compared with CUSUM. The fixed window also gives no principled leakage estimate. We will tune the window length and multiplicity to the same exposure loss as the sequential veto. If it matches the sequential veto on both residual background and systematic width, we will report that the simpler veto suffices and that the contribution is the leakage calibration alone.

Why not just train a per-event classifier such as a boosted decision tree? Per-event classifiers are already in use. They score each event in isolation, so a micro-discharge electron that looks like signal on its own cannot be separated from signal. Its only distinguishing feature is its temporal neighbors. We will include the classifier score as a mark inside the sequential statistic and report the classifier alone as a separate baseline. This separates what the time correlation contributes.

Why not just improve the background template? This is complementary rather than competing. A better template still leaves the burst-rate normalization free. The leakage prediction instead ties the residual background to an observable alarm-delay distribution, which is exactly the uncertainty that light-mediator limits are sensitive to.

Proposed Experiments

The first test is closed-world validation. We will generate event streams with the PandaX-4T energy ranges (about 0.04–0.39 keV for electronic recoils) and with burst parameters fitted to sideband data. Into these we inject a stationary signal at a known rate. We then check two things: that signal acceptance equals one minus the dead fraction, and that the predicted $N_{\text{leak}}$ matches the leakage measured in the simulation.

The decisive test runs on PandaX-4T Run0/Run1 sidebands, which requires collaboration access because the data are not public. It asks whether predicted and observed leakage agree within statistical error across a scan of $h$. If they do, we set the light-mediator limit and compare it with the baseline.

As a secondary, low-rate check, we will apply the same veto to the QROCODILE stream of 15 non-periodic pulses in 415.15 hours (First Sub-MeV Dark Matter Search with the QROCODILE Experiment). There we ask whether those pulses look stationary or burst-clustered. With so few counts this can rule out strong clustering, but it cannot demonstrate a gain.

The idea fails if the leakage prediction is biased on sidebands by more than its statistical error, or if a tuned fixed-window veto matches the sequential veto on every metric.