Proof-Budgeted Pulse Shape Discrimination for Verifiable Rare-Event Counts

When a germanium experiment searching for neutrinoless double-beta decay (a hypothetical nuclear decay that would show neutrinos are their own antiparticles) publishes "zero events in the signal window," outsiders must trust two things they cannot check. They must trust that the waveform cuts were frozen before the signal window was opened. They must also trust that those frozen cuts were actually applied to the recorded data. Blind analysis protects against the first failure only as a matter of collaboration discipline, and raw waveforms are rarely released, so the second failure cannot be checked from outside at all. Zero-knowledge proofs could close this gap. A zero-knowledge proof is a short cryptographic certificate that a stated computation was run correctly on committed inputs without revealing those inputs. The difficulty is that the discriminators physicists increasingly want, such as neural networks reading the full detector pulse, are expensive to prove. The cheap discriminator, a single ratio, is known to miss some backgrounds.

This note proposes to treat proof cost as a design constraint on the discriminator itself. We would train pulse shape classifiers whose inference cost inside a proof system is fixed in advance, and ask how much background rejection each unit of proof cost buys. We would then build a batched proof showing that the count of events passing a pre-committed cut equals the published count. Checking this proof should cost the verifier far less than a check that grows linearly with the number of events. The idea borrows its scaling question from anonymous group signatures, where the long-standing problem is that signature size grows linearly with group size. It borrows its membership machinery from those schemes to attest that each waveform came from a certified digitizer channel.

Core Claim

We claim that a discriminator built from a small number of fixed linear projections of the 256-sample current pulse, plus a low-degree polynomial output layer, can recover most of the advantage that a deep autoencoder has over the scalar amplitude-to-energy ratio. We further claim it can do so at a proof cost within a small constant factor of proving the scalar cut alone. The design question is therefore not "can we prove a neural network on a waveform." It is "what is the rejection-versus-proof-cost frontier for pulse shape discrimination, and where on it should a collaboration sit?" No paper in the file measures this frontier, and none proves anything about one-dimensional detector pulses.

The relevant physics is as follows. In p-type point-contact germanium detectors, the electric field is concentrated near a tiny electrode. A compact single-site deposit, the signature of double-beta decay, arrives there as one charge packet and produces a sharp current peak $A$ for its energy $E$. Gamma-ray backgrounds that scatter at several sites produce broader pulses with lower $A/E$. Multisite event discrimination for the majorana demonstrator turns this into the standard energy-corrected $A/E$ cut. It retains $(90\pm3.5)\%$ of double-escape-peak events and reduces single-escape-peak events to about 6%. Deep learning based pulse shape discrimination for germanium detectors compresses the waveform into a 7-dimensional autoencoder bottleneck and feeds it to a 141-parameter classifier. This reaches 90% background rejection at 90% signal acceptance, which matches $A/E$ overall while catching multi-site pulses with a sharp leading spike that fool the ratio.

Both papers motivate the proposal. The first gives a trivially provable baseline. The second gives evidence that extra waveform information exists, but only in specific subpopulations. That is exactly why the gain must be measured per unit of proof cost rather than assumed.

Research Program

The first component is a family of proof-costed discriminators. Every model has the form: take $k$ learned linear projections $W\in\mathbb{R}^{k\times 256}$ of the baseline-subtracted, fixed-point-quantized pulse $x$; append $A/E$ and energy; and apply a polynomial head of degree $d$. Linear maps and polynomials are native to arithmetic constraint systems, so their proof cost can be written in closed form:

$$ C(k,d) \approx 256\,k + P(k+2,\,d), $$

Here $C$ is the number of multiplication constraints per event, the first term counts the projection multiplications, and $P(m,d)$ is the number of monomials of degree at most $d$ in $m$ variables. This matters because training can then minimize classification loss subject to $C(k,d)\le B$ for a chosen budget $B$. We would sweep $B$ from the cost of the scalar cut up to the cost of the full convolutional autoencoder. For the upper end we would use the parametrized matrix-multiplication constraints of Architecture-private Zero-knowledge Proof of Neural Networks, which proves image-classifier inference and reports matrix checks 3 to 4 times faster than ordinary constraint systems. We reuse its matrix-check technique. We do not need its architecture privacy, since a physics cut should be public.

Labels follow the established calibration proxies: thorium-228 double-escape-peak events as the single-site class, and full-energy and single-escape-peak events as the multi-site class. The known weakness of this choice is inherited, not solved here. Calibration sources sit outside the crystal while signal decays occur uniformly through it, which is the origin of the 3.5% systematic.

The second component is provenance and batching. Each digitizer channel holds a signing key. Rather than signing every event, which would reproduce the linear-scaling problem, each channel appends event hashes to a hash chain and signs only the chain head at the end of each run. Certified channels are the leaves of a Merkle tree, the accumulator used in Zero-Knowledge Arguments for Lattice-Based Accumulators: Logarithmic-Size Ring Signatures and Group Signatures Without Trapdoors. The proof attests that each event belongs to a signed chain whose key sits in the tree. We do not use that paper's full group signature, whose 61.5 MB signature at $N=1024$ would be absurd per event. We keep only its membership-in-a-tree structure inside the proof. Channel anonymity comes for free, but we will say plainly that it is secondary. Its only honest use is to let a collaboration withhold channel-level detail, such as which detectors are being used as the blinded subset, while still proving that the data came from certified hardware.

The pre-registration step is a public, timestamped commitment to the quantized parameters $(W, \text{head}, \text{threshold})$ before unblinding. After unblinding, a single aggregated proof covers every event in the window. For each event it attests that the event is in a committed chain, that it lies in the committed energy window, and that it was classified by the committed model, and it attests that the passing count equals the published number. Prover cost grows as $N\cdot C(k,d)$ for $N$ events. The target is verifier cost and proof size that are essentially flat in $N$, which is the property the signature-scaling literature pursues for group size. The file gives no measured numbers for succinct aggregation on this workload, so we treat the flat-verifier property as something to measure, not assume.

Baselines

The naive reference point is a public hash commitment to the cut code and parameters before unblinding, followed by an internal re-run. It costs nothing and establishes pre-registration. It cannot show outsiders that the committed cut was applied to the committed data. Every proof-based variant must justify its cost against this point.

The strongest physics baseline is the energy-corrected $A/E$ cut of Multisite event discrimination for the majorana demonstrator. It reports $(90\pm3.5)\%$ double-escape-peak retention, single-escape-peak survival of about 6%, and a factor-of-3 background index reduction. We will re-derive it on our exact calibration split with its standard energy correction, and prove it inside the same system as our models. That places it at the lowest point of the cost axis.

The strongest learned baseline is the autoencoder-plus-classifier of Deep learning based pulse shape discrimination for germanium detectors, reported at 90% full-energy-peak rejection at 90% double-escape-peak acceptance. We will retrain it on the same split with the same hyperparameter-search budget as our family, quantize it to the same fixed-point precision, and prove it with parametrized matrix constraints. On proof cost for one-dimensional pulses: no reported result found on Lacuna for proving any pulse shape classifier.

The key ablation removes the learned projections, fixing $k=0$ and keeping only $A/E$ and energy with a polynomial head. This isolates whether any budgeted waveform information helps. A second ablation replaces learned projections with fixed wavelet coefficients, as in the transform used by Machine learning-powered data cleaning for LEGEND: a semi-supervised approach using affinity propagation and support vector machines. This tests whether training the projections matters.

Beating the baselines means the following. At matched 90% double-escape-peak acceptance, and at a proof cost no more than a stated multiple of the $A/E$ proof cost, the budgeted model must reject single-escape-peak and Compton-continuum events significantly better than $A/E$, beyond the 3.5% systematic. At a proof cost well below the autoencoder's, it must match the autoencoder's rejection within statistical error.

Why not just…?

Why not just prove the $A/E$ cut and stop? This is the real threat. Overall, the learned classifier only matches $A/E$ at 90/90 in Deep learning based pulse shape discrimination for germanium detectors. If the frontier is flat, the right recommendation is to prove $A/E$, and this proposal would say so. The pipeline would still be the first verifiable pulse shape cut. The ablation with $k=0$, together with rejection measured separately on leading-spike multi-site pulses, settles the question directly.

Why not just release a hash of the frozen code and let a trusted auditor re-run it? This handles pre-registration but relocates trust to the auditor and requires sharing raw waveforms with them. The data-auditing schemes in the file, such as Blockchain-Based Privacy-Preserving Public Auditing for Group Shared Data, audit storage integrity, not the correctness of a computation on the data. For collaborations willing to appoint one mutually trusted auditor, this alternative may be enough. Our claim applies only when verification must be public.

Why not just run off-the-shelf zero-knowledge machine learning on the full convolutional network? Nothing prevents it, and it is our upper cost point. Ordinary constraint systems pay 3 to 4 times more for matrix checks than parametrized ones, per Architecture-private Zero-knowledge Proof of Neural Networks, and convolutions over 256 samples dominate cost. If the full network's proof is affordable for the tens of events in a signal window, the budget matters less there. It matters much more for the thousands of calibration events that must also be proven to certify the acceptance number. We will report both regimes.

Why not just attach a group signature to each event? Measured costs rule this out. Pairing-based verification reaches 946 ms on automotive ARM hardware in Comparative evaluation of pseudonym-based anonymisation techniques for connected autonomous vehicles (CAVs), and lattice group signatures are 61.5 MB at 1024 members. Per-run signed hash chains avoid both costs.

Proposed Experiments

The primary experiment traces the rejection-versus-proof-cost frontier on thorium-228 calibration waveforms from point-contact detectors. We would obtain these through the collaborations behind the cited papers, with simulated point-contact pulses as a fallback if access is refused. Metrics are single-escape-peak survival and Compton-continuum survival at 90% double-escape-peak acceptance, as a function of per-event constraints, prover time, and proof size. The decisive check is whether any budgeted model at under ten times the $A/E$ proof cost beats $A/E$ beyond the 3.5% systematic, specifically on the leading-spike multi-site subpopulation.

A second experiment tests robustness. We would train on one detector and evaluate on others, since crystal non-uniformity is a known failure mode, to see whether cheap linear projections transfer worse than the ratio does.

A third experiment measures scaling of the batched proof: prover time, verifier time, and proof size against the number of events, from signal-window sizes up to full calibration runs. The expected pattern is linear prover growth and near-flat verification. A departure from that pattern would locate the practical limit.

Finally, we would repeat the frontier analysis for the anomaly-rejection stage of Machine learning-powered data cleaning for LEGEND: a semi-supervised approach using affinity propagation and support vector machines, holding its reported 0.024% physics sacrifice fixed. This shows whether the budget approach carries over from signal selection to data cleaning, which is the other cut a skeptic of a published count would question.