Certain-Answer Topology for Projected Particle Tracks

Many rare-event detectors do not record a particle's path in three dimensions. They record two flat shadows of it. The 2D strip readout studied for the PandaX-III neutrinoless double-beta decay search works this way: one set of strips measures charge along x and time, the other along y and time. Every downstream decision is made on a single 3D track reconstructed from those two views, for example "this event has two dense charge blobs at its ends, so it looks like signal." That reconstruction is a guess. Many 3D tracks can produce the same pair of shadows. Some of them look like signal and others look like a gamma-ray background. Signal claims in these experiments rest on fewer than one count per tonne-year, so an event classified on one arbitrary reconstruction is a systematic risk that nobody currently quantifies.

We propose to treat this ambiguity the way database theory treats incomplete data. An event's classification should be its certain answer: a topological property counts only if it holds in every 3D track consistent with the recorded projections. Knowledge-representation research has mapped when certain answers are cheap or intractable to compute. It has also found that the shape of the data decides which side of the line a problem falls on. We will carry that analysis over to projected tracks, prove where the problem is tractable, build an exact solver for the tractable fragment, and test whether certain-answer topology recovers the background rejection of a full 3D readout from a cheaper 2D one.

Core Claim

The decisive topological quantities in a projected-track event can be defined as minima and maxima over all consistent 3D explanations. Examples are the number of high-charge endpoints ("blobs") and the number of disconnected clusters. These bounds can be computed exactly and quickly when the projected event is path-like, which covers most signal and single-Compton background events. The hard cases are a small, identifiable set of branched or dead-strip-affected events. Classifying on the bounds, rather than on one best-guess reconstruction, should lower worst-case background leakage at a fixed signal efficiency. It should also flag exactly which events are ambiguous.

Written out, an event $e$ consists of x-strip hits and y-strip hits with charges and drift times. $\mathrm{Mod}(e)$ is the set of 3D charge assignments whose projections reproduce $e$ within detector resolution. For a topological count $c$, such as blobs, we define

$$ c_{\min}(e)=\min_{M\in \mathrm{Mod}(e)} c(M), \qquad c_{\max}(e)=\max_{M\in \mathrm{Mod}(e)} c(M). $$

An event is certified signal-like if $c_{\min}(e)\ge 2$ and certified background-like if $c_{\max}(e)\le 1$. Everything in between is reported as ambiguous instead of being resolved silently.

Where the Idea Comes From

The method comes from open-world counting. Cardinality Queries over DL-Lite Ontologies asks for the minimum number of objects across all models of an incomplete knowledge base. It computes this with "merging strategies" that collapse unnamed individuals, and finds that complexity ranges from highly parallel ($\mathrm{TC}^0$), through a class tied to the complement of perfect matching, up to coNP-complete. The structural parallel to strip readouts is direct. In each time slice, deciding which x-strip hit belongs with which y-strip hit is a bipartite pairing problem. Deciding whether two hits are one charge deposit or two is a merging decision.

We reuse this way of thinking, but not their logic. Our "data" are geometric and our "axioms" are continuity and charge-conservation constraints, so their theorems do not transfer verbatim. The proofs have to be redone.

A second line of work motivates the bet that tractability depends on event shape. Query Answering in Description Logics with Transitive Roles shows that conjunctive query answering in the logic $\mathcal{S}$ falls to EXPTIME when the data are tree-shaped, while the general case sits between co-NEXPTIME-hardness and a 2EXPTIME upper bound. Particle tracks are close to paths, with occasional branches from delta electrons or bremsstrahlung. Our conjecture mirrors that result: certain-answer blob counting is polynomial for path-like projections and becomes hard when branching or dead strips leave many pairings open.

The Price of Selfishness constrains the design. It shows that one innocuous-looking construct can push query answering up an exponential level. We will therefore add physical constraints one at a time, each with its own complexity result, rather than adopting a rich formal language wholesale.

The application regime comes from Topological background discrimination in the PandaX-III neutrinoless double beta decay experiment. That paper defines four track observables: energy ratio, blob charge, end-track twist, and length. It reports roughly seven orders of magnitude background suppression at above 20% signal efficiency, and compares a 3 mm 2D strip readout with 1–10 mm 3D pixel readouts. Its reconstruction uses heuristic spatial clustering in the REST software and returns one track per event. This proposal replaces that single track with bounds over all consistent tracks.

Research Program

The theoretical part has three steps.

First, we formalize $\mathrm{Mod}(e)$ for a strip readout. Space is discretized into voxels at strip pitch and drift-time resolution. A model assigns non-negative charge to voxels so that the x-projection and y-projection match the recorded strip charges within a resolution tolerance. A blob is a connected high-charge region at a track extremity. Dead strips are treated as unconstrained: they are the "open world" of the problem.

Second, we prove complexity results along two axes. One axis is event shape: single path, path with bounded branching, or arbitrary. The other is constraint set: pairing only, then adding charge conservation, continuity, and dead strips. The target is a classification of the same kind as the DL-Lite trichotomy. Our working hypotheses are polynomial time for paths via dynamic programming along drift time, a matching-flavored intermediate case when slices contain several hits with distinct charges, and coNP-hardness for $c_{\min}$ in general, shown by reduction from a covering-style problem.

Third, for the tractable fragment we build an exact solver. The design is a dynamic program over time slices whose state is the current pairing frontier. For hard events we fall back to an integer program with a time budget and report the outcome as solved or ambiguous.

The empirical part attaches $c_{\min}$ and $c_{\max}$, along with min/max versions of track length and energy ratio, to the PandaX-III cut sequence. We then measure what changes.

Baselines

The naive reference point is the existing PandaX-III pipeline run on the 2D 3 mm strip readout. It uses heuristic REST reconstruction followed by the four topological cuts in Topological background discrimination in the PandaX-III neutrinoless double beta decay experiment. We will rerun it on our own simulated samples so that every method sees identical events. The paper reports about $10^7$ background suppression at above 20% signal efficiency. The file gives no separate figure for the strip readout versus the pixel readouts, so there is no reported per-readout result for the comparison we need, and we will regenerate both.

The strongest competitor on the actual question, "how much does projection ambiguity cost?", is the same topological analysis run on 3D pixel readout at matched pitch. It sees the true track and is the ceiling our method tries to approach in software.

The strongest learned competitor is a convolutional network trained directly on the two projection images. This follows the evidence in Learning to identify electrons, where a CNN on raw calorimeter images reached an AUC of 0.972 against 0.945 for standard physics features. That shows low-level learners can absorb structure that hand-built features miss. The CNN gets the same training events and a hyperparameter search budget matched in GPU-hours to our solver's development budget. It also gets the same signal efficiency operating point.

The key ablation keeps everything in our pipeline but replaces the bounds with a single most-likely 3D reconstruction from the same voxel model. This isolates the all-models ingredient from the effect of simply having a better reconstruction.

Beating the baselines means the following at matched signal efficiency. Background rejection should be higher than the 2D heuristic pipeline. The gap to 3D pixels should be closed by a measurable fraction. Worst-case leakage under dead-strip and charge-noise perturbations should be lower than for the CNN and the ablation.

Why not just…?

Why not just train a neural network on the two projections and let it learn the ambiguity? This is the most serious threat. A CNN may match or exceed our nominal rejection, as the electron-identification results suggest. The case against relying on it alone is robustness and accountability. A network trained on one dead-strip map gives no statement about events it is uncertain about. Rare-event searches need to justify each surviving count, which is why Learning to identify electrons itself distilled its network back into physical observables. Our experiment settles the question directly. If the CNN matches us both nominally and under unseen dead-strip patterns, the contribution narrows to the ambiguity flag and the complexity map, and we will say so. We will also test the combination, feeding $c_{\min}$ and $c_{\max}$ to the CNN as extra inputs.

Why not just hand the problem to an integer-programming solver and skip the complexity theory? An integer program computes the same bounds, and we use one as the fallback and as a correctness check. The theory is what tells us which events the solver will finish on, and it supplies a dedicated polynomial algorithm for the path-like majority. We will report wall-clock time per event for the integer program alone versus our dynamic program. If the generic solver is already fast enough on every event, the algorithmic contribution shrinks and the complexity map becomes purely explanatory.

Why not just build the 3D pixel readout? PandaX-III evaluated pixel readouts, and they remove the ambiguity in hardware. They also multiply channel count and cost, which is why strip readouts are studied at all. Our result reports how much of the strip-to-pixel gap software can recover. If the answer is "almost none," that is a useful negative result for detector design.

Proposed Experiments

We will simulate $^{136}$Xe neutrinoless double-beta signal events and $^{214}$Bi and $^{208}$Tl gamma backgrounds in high-pressure xenon with REST, matching the PandaX-III configuration. Readouts will be 3 mm strips plus 3D pixels at 3 mm and coarser pitches.

We will check the theory by measuring empirical solver time against event branching and the number of dead strips. The tractable fragment should show polynomial scaling. Any event shape the theory calls easy but that blows up in practice refutes the analysis.

The main empirical test is background rejection at fixed signal efficiency, compared across five settings:

  • the 2D heuristic pipeline;
  • the single-reconstruction ablation;
  • our bounds;
  • the projection CNN;
  • 3D pixels.

The decisive check is a stress test. Random dead-strip maps and charge noise that none of the methods saw during development are injected at analysis time, and we measure how much each method's background leakage grows.

We will also report the fraction of events labeled ambiguous, and how often the pipeline's single reconstruction contradicts a certain answer. The second number directly measures how often current analyses classify an event on a coin flip among consistent tracks.

As a secondary test of generality, we will apply the same min/max formulation to the multi-plane wire readouts in Vertex-finding and reconstruction of contained two-track neutrino events in the MicroBooNE detector. That paper reports about 69% full-track reconstruction efficiency and failures where tracks overlap at small opening angles. These overlap failures are precisely where certain-answer counts of tracks should be informative.