Germanium detectors that search for neutrinoless double-beta decay, a hypothesized radioactive decay whose half-life would exceed $10^{25}$ years, have to recognize a signal event from the shape of a single electrical pulse. A true decay deposits its energy at one point in the crystal. Most gamma-ray backgrounds scatter several times and deposit energy at several points. The standard tests compare each pulse against an idea of what a "single-site" pulse should look like. That idea is always somewhat wrong, because charge drift through the crystal depends on where the event happened. Radar and sonar engineers know this failure mode well and call it steering-vector mismatch: the assumed signal signature differs from the true one. They have also learned how it produces self-cancellation, where the detector suppresses the very signal it is looking for, and they have tools that bound the damage.
This note proposes to bring one of those tools into pulse shape discrimination. The tool is a detector that accepts any signature lying inside a cone around a nominal signal subspace and rejects energy outside it. The two-sentence version is this. We treat the family of single-site pulses as the "steering vector," replace the single fixed template with a cone around a physics-derived subspace whose width follows from charge-drift variability, and test each whitened waveform by the angle it makes to that subspace. We predict, and will test, that fixed-template tests lose true signal events as energy rises because their shape error grows with pulse amplitude while noise does not, and that an angle-based cone test does not.
Core Claim
Our hypothesis is that the most common way pulse shape tests fail is the waveform analog of self-nulling. Take any test that scores a pulse by a residual against a fixed template, or that estimates the "background" statistics from data containing signal-like events. Such a test turns template mismatch into apparent background, and it does so more strongly the brighter the event. The analogy comes from beamforming. In beamforming, higher input signal-to-noise ratio (SNR) makes the mismatched part of the signal dominate the covariance, and the beamformer nulls the target, as described in A novel adaptive beamforming scheme for array signal data processing and Robust adaptive beamforming based on zone orthogonal constraint and steering vector estimation. In germanium, the calibration peak used to tune cuts sits at 1593 keV, and the search region sits near 2039 keV. The same mismatch in shape therefore costs more signal acceptance exactly where it matters.
Our remedy borrows from ADAPTIVE DETECTION OF MULTIPLE POINT-LIKE TARGETS UNDER CONIC CONSTRAINTS. That paper builds radar detectors that accept a target whose signature lies anywhere in a cone around a nominal vector. Its orthogonal-rejection variant also discards energy that falls outside the cone. We adapt this to temporal waveforms and change it in three ways. The cone is placed around a low-dimensional subspace of simulated single-site pulses rather than around a single vector. The cone half-angle depends on the event's measured drift time rather than being one fixed angle. The noise statistics come from pre-trigger baseline samples, which cannot contain signal, so the whitening step cannot be contaminated by target energy.
To our reading of the file, no published work applies mismatch-robust array detection to pulse morphology. The closest pulse-side work uses scalar cuts or learned features without any bound on allowed shape variation.
Method
Each event gives a time-aligned, energy-normalized waveform $x \in \mathbb{R}^d$ (we use $d = 256$ samples, following Deep learning based pulse shape discrimination for germanium detectors). The detector noise covariance $R$ is estimated from baseline traces. Because $d$ is large relative to a practical number of clean traces per detector and run, we regularize $R$. We use the closed-form, detection-optimal shrinkage of Optimal Design of the Adaptive Normalized Matched Filter Detector. That paper shows that tuning the shrinkage of a covariance estimate for detection probability gives a different, better setting than tuning it for estimation error. We reuse its method; unlike that paper, we do not assume a known signature.
A charge-drift field simulation of each detector generates single-site pulses over the crystal volume. Their principal components, whitened by $R^{-1/2}$, span a $k$-dimensional subspace with orthogonal projector $P$. We score a whitened waveform $\tilde{x} = R^{-1/2}x$ by the squared cosine of its angle to that subspace:
$$ c(x) = \frac{\tilde{x}^\top P \tilde{x}}{\tilde{x}^\top \tilde{x}}, \qquad \text{accept as single-site if } c(x) \ge \cos^2\theta(\tau), $$
Here $\tau$ is the measured drift time, used as a proxy for depth. $\theta(\tau)$ is the cone half-angle, set at the 99th percentile of simulated mismatch between real-geometry pulses and the subspace at that drift time.
Because $c(x)$ is a ratio, it does not change when the pulse is scaled. This is the property the energy-dependence prediction relies on. The orthogonal-rejection step follows the orthogonal-rejection detector of Hao et al. It adds a second statistic, the fraction of whitened energy in the best-fitting two-template multi-site model, and accepts an event only if the single-site cone wins.
Scoring an event costs $O(dk)$ after precomputation, with no iterative solve. This matters because the heavier robust beamformers in the file, such as the semidefinite programs with bilinear constraints of Robust Adaptive Beamforming via Worst-Case SINR Maximization With Nonconvex Uncertainty Sets, are too slow for streaming millions of pulses. We deliberately keep only the cone idea from that literature and drop the optimization machinery.
Research Program
The work has three parts.
The first part measures the phenomenon before proposing the fix. On thorium-228 calibration data, we will track the signal acceptance of existing tests as a function of energy. We use the single-site-rich double-escape peak and single-site-dominated Compton continuum events, cross-checked against simulation with known ground truth. If acceptance for fixed-template residual tests does not fall with energy, the self-nulling analogy is wrong for this domain, and the proposal reduces to "a new subspace classifier." We would report that honestly.
The second part builds the cone detector and its drift-time-dependent width.
The third part addresses the geometry problem flagged by Holl et al. Point-contact assumptions break in large inverted-coaxial crystals, where drift varies more. Our claim is that, in those crystals, a width tied to physics degrades more gracefully than a single learned decision boundary. The surface-alpha channel from $\alpha$-event characterization and rejection in point-contact HPGe detectors serves as a stress test. There, delayed charge release adds a slow tail, so these events should fall outside the cone without a dedicated tail parameter.
Baselines
The naive reference point is the amplitude-over-energy test, the maximum current amplitude divided by total energy. It is the standard single-site test described in Holl et al. We run it with its usual cut tuned to keep 90% of double-escape-peak events, on the same events we use for our own method.
The strongest existing method is the two-stage network of Deep learning based pulse shape discrimination for germanium detectors. It compresses waveforms with a convolutional autoencoder to a 7-dimensional code and then classifies them with a small network. It reports 90% background rejection at 90% signal acceptance, separating double-escape-peak from full-energy-peak events. We will retrain it on our exact training split, with an equal hyperparameter search budget, and evaluate it on identical held-out runs. Beating it means one of two things. Either we match its rejection at the double-escape peak and show a statistically significant smaller loss of single-site acceptance when moving to energies near 2039 keV and to inverted-coaxial detectors. Or we exceed 90% rejection at 90% acceptance outright.
For surface alphas, the reference is the amplitude-over-energy test combined with the delayed-charge-recovery tail slope from Arnquist et al. That combination rejected all surface alphas beyond 3 mm, within statistical uncertainty, while sacrificing 0.2% of bulk events. Matching it without a hand-designed tail parameter would count as success.
Four ablations isolate the key ingredient:
- a fixed cone angle instead of the drift-time-dependent one;
- whitening from calibration-event covariance instead of baselines, which tests the contamination mechanism;
- a single mean template instead of a subspace;
- removing the orthogonal-rejection term.
Why not just…?
Why not just correct amplitude-over-energy for drift time? This is the cheap fix a practitioner would try first, and it may recover much of the energy-dependent loss. It is included as an additional baseline, tuned on the same data. If it matches the cone detector on the energy-dependence measurement, the scalar correction wins on simplicity. The remaining argument for our method would then be multi-site events that the scalar misclassifies, which Holl et al. show exist.
Why not just fit each pulse against a large library of simulated templates and cut on the best chi-square? This is the real threat, because a dense library also covers position dependence. The difference is that a chi-square residual grows with amplitude, which is exactly the self-nulling mechanism we hypothesize. Our angle statistic does not grow with amplitude. The first experiment decides between them directly: if library chi-square acceptance is flat in energy, our scale-invariance argument fails.
Why not just train a larger network with simulated augmentations? This is plausible for the double-escape-peak benchmark. However, the network's labels come from calibration peaks that are known to be contaminated: double-escape-peak events contain multi-site events and vice versa. It also has no explicit bound on acceptable shape variation, so its acceptance near 2039 keV is unknown until measured. We will train this augmented network with the same simulation budget. It beats us if its acceptance also stays flat in energy and geometry.
Proposed Experiments and Decisive Checks
We will use thorium-228 calibration data in the style of Holl et al. The file does not name a public release, so this requires access through an experimental collaboration. We will also use the MAJORANA TUBE-scanner alpha data of Arnquist et al., and simulated waveforms for point-contact and inverted-coaxial geometries where ground truth is exact.
Metrics are:
- background rejection at 90% single-site acceptance;
- the slope of single-site acceptance versus energy between 1593 keV and the region near 2039 keV;
- the same acceptance comparison between point-contact and inverted-coaxial crystals;
- surface-alpha rejection and bulk sacrifice;
- per-event latency.
Two checks are decisive. The first is whether fixed-template and chi-square tests actually lose acceptance with energy while the cone test does not. The second is whether the baseline-whitening ablation shows that estimating noise from contaminated events costs acceptance, as the self-cancellation account predicts.
A further stress test targets the high-rate regime reported in Machine learning-powered data cleaning for LEGEND, where anomalous-event leakage reached 15.034±0.224%. Pileup waveforms should fall outside the single-site cone. We will measure whether the cone test reduces that leakage on comparable data, while noting it was not designed as a data-cleaning tool.
References
- A novel adaptive beamforming scheme for array signal data processing
- Robust adaptive beamforming based on zone orthogonal constraint and steering vector estimation
- ADAPTIVE DETECTION OF MULTIPLE POINT-LIKE TARGETS UNDER CONIC CONSTRAINTS
- Deep learning based pulse shape discrimination for germanium detectors
- Optimal Design of the Adaptive Normalized Matched Filter Detector
- Robust Adaptive Beamforming via Worst-Case SINR Maximization With Nonconvex Uncertainty Sets
- $$\alpha $$-event characterization and rejection in point-contact HPGe detectors
- Machine learning-powered data cleaning for LEGEND: a semi-supervised approach using affinity propagation and support vector machines