Low-Rank Tensor Background Models for Unbinned Rare-Event Searches

Large liquid-xenon detectors hunt for rare events, such as dark-matter scatters or neutrinoless double-beta decay, that may occur only a handful of times a year. Each recorded event carries a 3D position and an energy. Where an event sits in the detector says a lot about whether it is background, because radioactivity from the walls piles up near the edges while a signal would spread evenly through the xenon. Most analyses still throw that information away. They cut to a central "fiducial" volume and fit a 1D energy histogram, as PandaX-4T does over 1.54 ton-years of exposure in Exploring New Physics with PandaX-4T Low Energy Electronic Recoil Data. The reason is not conceptual. Building a smooth, correctly normalized probability density for the background over position and energy, from a finite and expensive Monte Carlo simulation, is hard.

We propose to build that density from the factorization that made fast 3D scene reconstruction practical. TensoRF's vector-matrix (VM) decomposition writes a 3D field as a sum of products of 1D curves and 2D maps, one pair per axis (TensoRF: Tensorial Radiance Fields). If every factor is constrained to be nonnegative, the field is a valid event rate, and its integral over any box-shaped region has an exact closed form. That integral is exactly what an unbinned Poisson likelihood needs and what neural density models make awkward. The central bet concerns the regime where simulation statistics are the bottleneck. A low-rank intensity trained directly on simulated events should give lower-variance background templates than fine 3D histograms, and sharper boundary modeling than generic density estimators. Together these should translate into a measurably better expected upper limit per unit of simulation budget.

Core Claim

The claim is specific and falsifiable. Take a fixed Monte Carlo budget of background events. A nonnegative VM intensity field over cylindrical detector coordinates and energy, $(r, z, E)$, fitted by maximizing the Poisson point-process likelihood of the simulated events, will outperform two alternatives: fine binned 3D templates, and a normalizing-flow density. The comparison is on the median 90% confidence-level upper limit on a uniform signal rate in a toy xenon time projection chamber (a detector that records scintillation and ionization to reconstruct event position and energy). We expect the advantage over histograms to be largest when simulation is scarce and to shrink as simulation grows, because low rank acts as a smoothing prior that trades variance for bias. We expect the advantage over flows to come mostly from exact normalization over fiducial regions and from boundary behavior aligned with the axes. If neither advantage appears, the idea is wrong and should be dropped.

The model for background source $k$ is

$$\lambda_k(r,z,E) = \sum_{m=1}^{R} u^{(m)}_{z}(z)\,M^{(m)}_{rE}(r,E) + u^{(m)}_{r}(r)\,M^{(m)}_{zE}(z,E) + u^{(m)}_{E}(E)\,M^{(m)}_{rz}(r,z).$$

Here each $u$ is a nonnegative 1D curve and each $M$ is a nonnegative 2D map, both stored on grids and linearly interpolated. $R$ is the rank. The angle $\phi$ is treated as uniform, or given its own low-rank factor if the geometry requires it. Because each term is a product of functions of disjoint variables, its integral over a box factorizes into integrals of the individual curves and maps. For piecewise-linear interpolation those integrals are exact trapezoid sums. The physics fit then maximizes the extended likelihood

$$\log L(s, b) = -\Big(s\,\Lambda_s + \sum_k b_k \Lambda_k\Big) + \sum_{i=1}^{n} \log\Big(s\,\lambda_s(x_i) + \sum_k b_k\,\lambda_k(x_i)\Big).$$

In this expression $x_i$ are the $n$ observed events, $s$ and $b_k$ are the signal and background strengths, and $\Lambda$ denotes the exact integral of each intensity over the analysis region. This is the extended maximum likelihood structure CLEO II used in two kinematic variables to reach $\mathcal{B}(\tau\to\mu\gamma) < 1.0\times10^{-6}$ from an estimated 1.8 signal events (Update of the search for the neutrinoless decay τ→μγ). That paper's densities were simple parametric forms. Our contribution is a density family that scales to three or four continuous variables with irregular detector-driven structure.

Relation to Prior Work

The method comes from the tensorial scene-representation literature. TensoRF introduced VM and CP factorizations and showed that interpolating the factors before multiplying them gives fast, compact fields: under 75 MB against over 600 MB for dense voxels, with CP models under 4 MB. TensoSDF showed VM fields represent sharp surface boundaries better than neural networks, which matters here because backgrounds change abruptly at detector walls. D-TensoRF extended the factorization to 4D, which we would need if we add a second energy-like variable. All of these fit fields by pixel-color regression. None imposes nonnegativity, uses normalization, or trains on a point-process likelihood. That shift is the change we make.

On the physics side, several papers motivate the gap. PandaX-4T fits a 1D binned energy likelihood (PandaX-4T), and the multi-ten-tonne projections in PandaX-xT rely on conventional fiducialization. Both leave the spatial information unused. EXO-200 shows that adding event-topology information buys about 25% sensitivity. That result is evidence that extra discriminating variables pay off. EXO-200 obtained it with a classifier score used as an auxiliary variable, not with a spatial density. Our approach is complementary, and a classifier score could later become a further tensor axis.

CaloChallenge 2022 constrains the design. It reports that exact-likelihood flows ran into memory problems at 40,500 voxels, which is a warning about flow scalability. It also shows that the best generative models reach classifier AUCs near 0.5–0.6 against Geant4, which is a warning that modern neural density models are strong. We do not claim to beat them at shower generation. We use the challenge only as a secondary test of density fidelity.

Research Program

We will build three pieces. The first is an open toy cylindrical xenon detector simulation. Its scale follows the PandaX-4T fiducial mass of about 2.4 tons and its low-energy region of interest up to 30 keV. It includes gamma backgrounds that fall off exponentially from the walls, a uniform intrinsic beta background, and a uniform signal peak. Energy smearing and position resolution are applied, and the simulation is fast enough to generate any number of events. The second piece is the nonnegative VM intensity, with softplus-parameterized factors, grid resolutions of 64 to 256 per axis, and rank chosen on held-out simulated events. The third is an extended-likelihood fitting and toy-experiment harness that computes median upper limits, discovery significance, and coverage.

The analysis has two focal questions. The first is how template variance and bias trade off against rank and simulation budget. The second is whether misfit from the low-rank model shows up as spurious signal.

Baselines

The naive reference is what the field actually does: a fiducial cut followed by a binned Poisson fit in energy alone, following PandaX-4T. It is run on identical toy data, with its fiducial radius and height optimized for expected limit on the same simulation. The file reports no sensitivity number for this setup on our toy detector, so all numbers will be regenerated. For scale, PandaX-4T reports an observed background p-value of 0.16 with that method.

The second baseline is a binned 3D Poisson fit with fine histograms in $(r,z,E)$ built from the same simulated events, with bin widths tuned per simulation budget on held-out data. This is the most natural competitor and the one most likely to tie us when simulation is plentiful.

The third baseline is a smoothed histogram, a 3D kernel density estimate with bandwidth tuned on held-out simulated events.

The strongest existing method is a normalizing flow over $(r,z,E)$. Its normalization over the fiducial region is estimated by Monte Carlo integration. It receives the same simulated events and a hyperparameter search matched in GPU-hours to the VM search. No reported result was found on Lacuna for flow-based background densities in rare-event likelihoods. The closest reported figure is CaloChallenge's classifier AUC of 0.5–0.6 for the best shower generators, which is a different task.

Two ablations isolate the proposal's key ingredients. The first is a rank-1 CP model, $p(r)p(z)p(E)$, which tests whether correlation modeling matters at all. The second is a VM model fitted by regression to a histogram rather than by point-process likelihood. Beating the baselines means a median 90% confidence-level limit at least 10% lower than the best baseline at matched simulation budget, with coverage within two percentage points of nominal.

Why not just…?

Why not just use fine 3D histograms with a binned likelihood? When simulation is abundant, this should essentially match any unbinned method, and we expect it to. The threat is real. The argument for VM rests on the regime where simulation is limited: fine bins contain few simulated events each, and that noise inflates limits or biases fits. The simulation-budget sweep below settles the question. If histograms match VM at the smallest budgets, the claim fails.

Why not just use a normalizing flow or another neural density? CaloChallenge shows flows are accurate but costly at high granularity (CaloChallenge 2022). Flows also integrate over fiducial boxes only approximately, and their smooth maps tend to round off sharp wall edges, which TensoSDF-style grids handle well (TensoSDF). A flow might nonetheless match VM on limits. If it does, the remaining VM advantages are exact normalization, speed, and interpretable factors. We would then report the result as a tie on sensitivity and not claim a sensitivity gain.

Why not just assume separable densities in each variable? That is the rank-1 CP ablation. Wall gammas couple radius, height, and energy, because attenuation hardens the spectrum with depth, so separability should fail. Measuring how much it fails tells us whether higher rank is needed at all.

Why not a coordinate MLP? Beyond its cost, an MLP shares the flow's normalization problem, which is worse here because the MLP outputs an unnormalized function. The flow baseline already covers the neural option more strongly.

Proposed Experiments

The main experiment sweeps the simulation budget from $10^4$ to $10^8$ background events and, at each budget, fits every template method. We then run 2,000 background-only toy experiments at PandaX-4T-like exposure and report the median 90% confidence-level signal limit and its coverage. The prediction is that VM's lead over histograms narrows as the budget grows, since its benefit is variance reduction. A flat or reversed curve would falsify the mechanism.

The second decisive check is spurious signal. We fit signal-plus-background to large background-only samples and measure the fitted signal bias as a function of rank. A low-rank model that systematically absorbs background mismodeling into the signal term would disqualify the method, whatever its limits look like.

The third check tests misalignment with the axes, which the file names as VM's structural weakness. We add a tilted source, for example a hot spot on a pipe that is not aligned with $r$ or $z$, and measure how much rank is needed to recover the true held-out log-likelihood.

The secondary fidelity check treats energy deposits in CaloChallenge Dataset 2 as weighted point sets on its cylindrical 6,480-voxel grid. It compares held-out log-likelihood and memory for VM, CP, and flow densities at matched parameter counts. This tests whether low-rank tensors capture real shower-scale structure and not only our toy backgrounds.