Experiments searching for extremely rare decays read out each event as a short electrical waveform from a germanium crystal. They then have to answer two questions about it. Is this a real particle interaction at all, or an electronic glitch? If it is real, did the energy land in one spot, as the sought-after neutrinoless double-beta decay would, or in several spots, as a background gamma ray that scatters around the crystal would? Today these questions get separate answers. A one-number cut on peak current over energy handles the second. Learned classifiers trained on hand-labeled clusters handle the first. This note proposes one geometric object that answers both. The set of waveforms a germanium detector can physically produce is a convex cone: every real waveform is a non-negative mixture of single-deposit pulse shapes, and the mixture weights must add up to the measured energy. A glitch lies outside this cone. A multi-site event lies inside it but needs several pulse shapes to build. A single-site event lies on one of the cone's edges.
The idea comes straight from quantum state tomography. There, noisy estimates fall outside the cone of valid density matrices (the matrices that describe physically possible quantum states), and enforcing that constraint during reconstruction cuts error sharply when data are scarce. The tomography literature also warns that projecting onto the edge of the cone distorts statistics. We use both lessons. Positivity-constrained decomposition of each waveform supplies the discriminant. The known shape of the noise distribution at the cone's boundary supplies a principled threshold. Neither has been used in rare-event pulse-shape analysis.
Core Claim
Write a waveform as $w \in \mathbb{R}^T$, a vector of $T$ current samples, already whitened using the noise covariance measured on pre-trigger baselines. Let $D$ be a dictionary whose columns $d_j$ are unit-charge single-site current pulses after the detector's electronic response. The columns cover a grid of interaction positions and start times. Two residuals summarize the event:
$$r_{\mathrm{phys}}(w) = \min_{c \ge 0,\ \mathbf{1}^\top c = E} \lVert w - Dc \rVert^2, \qquad r_{\mathrm{SS}}(w) = \min_{j} \lVert w - E\, d_j \rVert^2 .$$
Here $c$ is the vector of non-negative charge weights and $E$ is the reconstructed energy. The first residual, $r_{\mathrm{phys}}$, is the squared distance from the waveform to the physical cone; it serves as a glitch score. The gap $\Delta = r_{\mathrm{SS}} - r_{\mathrm{phys}}$ is a constrained likelihood-ratio statistic for "several sites" against "one site."
Because the multi-site fit is constrained to lie on a cone, under the single-site hypothesis $\Delta$ does not follow the textbook chi-squared distribution. It follows a mixture of chi-squared distributions with different degrees of freedom, known as a chi-bar-squared distribution. Ignoring this is the waveform version of the projection bias documented in tomography. Noise alone will fit small spurious positive weights, and a naive threshold will throw away real single-site events.
We make three predictions. First, $\Delta$ matches learned classifiers at high energy and beats them at low signal-to-noise. Second, $\Delta$ needs no multi-site labels. Third, because the dictionary covers the whole crystal volume rather than inheriting the calibration source's spatial bias, its single-site acceptance carries less calibration-to-bulk systematic error than the current cut does.
Where This Sits
The method draws on three tomography results.
Quantum State Tomography via Linear Regression Estimation and A two-stage solution to quantum process tomography both estimate without constraints first and then project onto the physical set. That split gives us our main ablation: unconstrained least squares with a projection afterwards, versus fitting with positivity enforced inside the fit.
Mitigating shot noise in local overlapping quantum tomography with semidefinite programming reports that enforcing physicality reduced the measurement samples needed by a factor of 10 to 100. That result motivates our prediction that the constraint helps most where the data are noisiest, which for us means low-energy events.
On how neural networks enhance quantum state tomography with limited resources builds physicality into a network's output by having it produce a Cholesky factor, a triangular matrix whose product with its own transpose is always a valid state. That paper also finds the benefit concentrated in the low-sample regime.
We do not port semidefinite programs to waveforms literally. A current pulse is not a matrix. The constraint that carries over is non-negativity of deposited charge together with charge conservation. That constraint is convex and cheap to enforce, and it is exactly the structure of the physics.
On the detector side, four papers set the problem.
Multisite event discrimination for the majorana demonstrator defines the operational baseline. Its cut on peak current over energy (called A/E, or AvsE in that paper) is a one-number proxy for the multi-site statistic. Its ±3.5% signal-efficiency uncertainty comes from the calibration source's spatial distribution, which is the gap our volume-covering dictionary targets.
Deep learning based pulse shape discrimination for germanium detectors shows that a learned model catches multi-site events with late sub-structure that A/E misses. Our decomposition should catch these explicitly as additional non-negative charge weights. Unlike that model, it needs no training labels and gives an interpretable output.
Machine learning-powered data cleaning for LEGEND cleans glitches with wavelet features, clustering and a support vector machine. Its 15% leakage at high rate is attributed mainly to pileup, meaning two real pulses overlapping in one readout window. In our geometry pileup sits inside the cone, as two weights well separated in start time, while a glitch sits outside it. The decomposition therefore separates the two cases that this paper's classifier confuses.
Ad-hoc Pulse Shape Simulation using Cyclic Positional U-Net supplies the method for building the dictionary. It translates ideal simulated pulses into the detector's real shape, and it reports that the raw simulation matched the real tail-slope distribution 0% of the time. That negative result rules out a purely first-principles dictionary. The same paper notes that its translator produces occasional non-physical artifacts, so translated atoms must themselves be screened.
Research Program
The work has four parts.
The first part builds the dictionary. We generate single-site pulses with the siggen charge-drift simulator on a position grid covering the crystal. We then map them toward real detector shapes using two routes, compared directly. The first route is a translator in the style of the Cyclic Positional U-Net above. The second is a low-parameter electronics model, an RC tail and a smoothing kernel, fitted to single-site calibration events. These are double-escape-peak (DEP) events from a thorium-228 source, where two 511 keV photons escape and the remaining energy is deposited at essentially one point. Translated atoms that have negative current lobes or that violate charge conservation are rejected or repaired. This is the dictionary-level analogue of tomographic projection.
The second part is the solver. Start-time shifts make the dictionary convolutional, so an active-set non-negative least-squares solver with fast convolutions runs per event in milliseconds. We will measure throughput directly against data-acquisition rates.
The third part is the statistics. We derive the null distribution of $\Delta$. Concretely, we compute the chi-bar-squared mixture weights from the local geometry of the cone, check them by Monte Carlo against simulated single-site events, and set thresholds at fixed single-site acceptance without multi-site labels.
The fourth part is cleaning. Thresholding $r_{\mathrm{phys}}$ gives a glitch flag. The number and time spread of the active weights gives a pileup flag.
Baselines
The naive reference is the A/E cut of Multisite event discrimination for the majorana demonstrator. On thorium-228 calibration it reports $(90 \pm 3.5)\%$ single-site retention. It suppresses single-escape-peak events (SEP, a mostly multi-site population) to about 6% of their original count, giving a factor-of-3 background reduction in the region of interest. We will run it with its published energy-dependent normalization on identical events. Beating it means lower SEP survival at equal DEP acceptance, or equal survival with a smaller acceptance uncertainty.
The strongest learned competitor is the autoencoder plus small classifier of Deep learning based pulse shape discrimination for germanium detectors. It reports 90% rejection of full-energy-peak (FEP) events at 90% DEP acceptance. We will retrain it on the same training split, with the same hyperparameter-search budget we allow our own threshold and dictionary choices. Matching it at high energy without multi-site labels, and beating it below roughly 500 keV or on held-out detectors, counts as a win.
For cleaning, the baseline is the clustering-plus-classifier pipeline of Machine learning-powered data cleaning for LEGEND. It reports $0.024 \pm 0.004\%$ physics sacrifice and $15.034 \pm 0.224\%$ anomalous leakage on high-rate test-stand data. Reducing that leakage at equal sacrifice is the target.
The ablation that isolates the key ingredient keeps the dictionary fixed. It replaces the positivity-constrained fit with ordinary least squares followed by clipping negative weights to zero, mirroring two-stage tomography. A second ablation computes $\Delta$ but uses a naive chi-squared threshold, to measure what the cone-aware null distribution buys.
Why not just…?
Why not just use A/E, which already works? For point-contact detectors near 2 MeV, A/E may genuinely capture most of the separation. Holl et al. document multi-site events with late sub-structure that A/E misses. A/E also does nothing for glitches or pileup. The energy-binned comparison will show whether the decomposition adds anything over A/E. If it does not at the region of interest, the remaining claim narrows to cleaning and low energy.
Why not just train a supervised network on DEP versus SEP? This is the real threat to raw discrimination power. Predicting the single-site and multi-site event discrimination power of dual-phase time projection chambers finds that an unconstrained ResNet approaches the Cramér–Rao bound in time projection chambers, which suggests learned models may leave little headroom. We do not claim to beat such a network on in-distribution calibration data. Our claims concern label-free operation, transfer from calibration to bulk events and across detectors, and behavior at low signal-to-noise. A network trained on the same data will be scored on exactly those shifted test sets.
Why not just fit siggen pulses directly? Li et al. show that idealized simulations miss tail kinetics entirely. A first-principles dictionary would put many real single-site pulses outside the cone. We test this by measuring $r_{\mathrm{phys}}$ on DEP events with raw versus corrected atoms.
Proposed Experiments
We start with the 65,536-event siggen set used by Li et al. We add measured baseline noise and planted glitches, which gives ground truth for site multiplicity and for non-physical events. On this set we validate the chi-bar-squared null and the effect of positivity on accuracy as energy falls.
We then move to thorium-228 calibration of point-contact or broad-energy germanium detectors. DEP serves as the single-site sample. FEP and SEP serve as the mostly multi-site samples. We will report results binned by energy and by detector. Access to MAJORANA- or LEGEND-type data and to the LEGEND test-stand data has to be arranged through those collaborations. No public release is listed in our sources.
The decisive checks are three.
First, the advantage of the constrained fit over unconstrained least squares with clipping should grow as signal-to-noise decreases. If it does not, the tomography analogy fails.
Second, single-site acceptance should agree between DEP events and simulated uniformly distributed bulk events to better than the 3.5% spread reported for A/E.
Third, pileup should account for most of the clustering pipeline's high-rate leakage, and it should be separable from glitches using $r_{\mathrm{phys}}$ together with the timing of the active weights.
References
- Quantum State Tomography via Linear Regression Estimation
- A Two-Stage Solution to Quantum Process Tomography: Error Analysis and Optimal Design
- Mitigating shot noise in local overlapping quantum tomography with semidefinite programming
- On how neural networks enhance quantum state tomography with limited resources
- Multisite event discrimination for the majorana demonstrator
- Deep learning based pulse shape discrimination for germanium detectors
- Machine learning-powered data cleaning for LEGEND: a semi-supervised approach using affinity propagation and support vector machines
- Ad-hoc Pulse Shape Simulation using Cyclic Positional U-Net
- Predicting the single-site and multi-site event discrimination power of dual-phase time projection chambers