One Run, Every Rank: Self-Truncating Neural Operator SVD Across the Gamma-Ray Gap

Many scientific datasets are really a function of two variables that happens to be low rank: a sum of a few "profile times spectrum" pieces. Pulsar gamma-ray data are a clean example. Photon arrival rate depends on the rotational phase of the star and on photon energy, and each emission region contributes one pulse shape multiplied by one energy spectrum. The number of such pieces, which is the rank of the phase–energy intensity, is the number of physically distinct emission components. That number is exactly what astrophysicists argue about. The data that would settle it are also badly incomplete. Space telescopes such as Fermi-LAT run out of photons above about 10 GeV. Ground-based Cherenkov telescopes only begin to see showers at tens of GeV. The two instruments disagree on energy scale and flux normalization by 10–17%.

We propose one method that recovers such a low-rank structure and its rank together, from raw photon event lists, in a single training run. The model is a neural network that learns the leading singular functions of the intensity, meaning the "profile" and "spectrum" factors in order of importance. It is trained so that every prefix of its first $k$ outputs is itself a valid rank-$k$ fit. Rank selection then becomes reading off where held-out error stops improving along those prefixes, with no retraining. We also give an unbiased, unbinned training objective computed directly from Poisson photon events. Finally, we treat inter-instrument calibration as explicit nuisance parameters, so that a calibration mismatch is not mistaken for an extra emission component.

Core Claim

Neural eigenfunction solvers need the number of modes fixed in advance. NeuralEF learns kernel eigenfunctions with one network per mode, solving a sequence of variance-maximization games. It reports that modes with eigenvalues below about 1% of the leading one are lost to gradient noise, so choosing $k$ means training many models. NeuralSVD learns top-$L$ singular functions with an unconstrained "joint nesting" objective, a weighted sum of rank-1, rank-2, …, rank-$L$ approximation losses that forces outputs to sort themselves by singular value. It is evaluated only on analytic operators and an image-retrieval kernel, with $L$ fixed.

Our claim is that joint nesting already contains a rank estimator nobody has used. Because each prefix is trained as its own low-rank approximation, one overspecified run yields an entire rank-versus-error curve. Held-out error on that curve, computed without bias from event data, picks the rank. The claim has three parts. It must work when observations are Poisson counts with a hole in the middle of the energy axis. It must hold when the two sides of the hole carry multiplicative calibration offsets. And it must do better than the obvious alternative of fitting many separate smooth factorizations and comparing them.

The objective is what makes event data usable. Let $x=(\phi,E)$ be phase and true energy, $I(x)$ the intrinsic intensity, and $\epsilon_b(x)$ the known exposure of instrument $b$, which is zero inside that instrument's blind energies. Let $F_k(x)=\sum_{j\le k}\sigma_j u_j(\phi)v_j(E)$ be the rank-$k$ prefix of the network's output. The exposure-weighted squared error expands as

$$ \int \epsilon\,(I-F_k)^2\,dx \;=\; \text{const} \;-\; 2\,\mathbb{E}\Big[\sum_{i} F_k(x_i)\Big] \;+\; \int \epsilon\, F_k^2\,dx, $$

where the $x_i$ are the detected photons. The middle term is an exact sum over events, because the expected sum of any function over a Poisson process equals its integral against the rate. The last term is a Monte Carlo integral over the exposure. The loss is quadratic in $F_k$, so it plugs directly into NeuralSVD's nested objective. It needs no binning, and the same identity gives an unbiased held-out error on masked events. Energy dispersion is handled by folding $v_j$ through the published response kernel inside both terms.

Calibration enters as per-instrument energy scale $s_b$ and normalization $n_b$. Instrument $b$ sees $n_b\,I(\phi, s_b E)$. Gaussian priors are set from reported systematics, namely 15–17% on energy scale and about 11% on flux normalization in the MAGIC Crab analysis. Without these parameters, a normalization jump at the instrument boundary is itself a separable piece, a step function in energy times the mean pulse profile. It would inflate the estimated rank by one.

Research Program

We first establish the rank-readout mechanism where the truth is known. For this we use truncated, noise-corrupted versions of the 2D harmonic oscillator and hydrogen-atom operators from NeuralSVD, and finite-rank kernels with controlled spectral decay. There we measure two things. The first is whether trailing prefixes stop improving held-out error at the true rank. The second is whether the singular values of surplus modes collapse toward zero, as Ding et al. prove for overspecified factored matrix recovery. That paper shows a factored subgradient method with a median-based diminishing step size lets redundant rank dimensions collapse, with sublinear-then-linear convergence. Its proofs cover only finite matrices, Gaussian measurements, and positive semidefinite factorizations. We use it as a hypothesis about neural function space, not as a guarantee, and test its step-size rule as an ablation. The weak-mode failure reported by NeuralEF constrains the design. We expect true components near 1% of the leading singular value to be the hardest case, and we report the detectable ratio as a function of photon count rather than assuming it away.

We then move to a semi-synthetic gamma-ray regime. We simulate Geminga-like intensities with two or three separable components: two pulse peaks, optionally with an inter-peak "bridge," and sub-exponential cutoff spectra. We apply a Fermi-like exposure and an LST-like exposure with a 20 GeV threshold, then inject calibration offsets drawn from the reported systematic ranges. The hardest variant hides a weak third component whose spectrum lives mostly inside 5–20 GeV. That component can only be seen through its tails in the two instruments.

Finally, we apply the method to public Fermi-LAT Geminga events. To create a known gap, we artificially mask energy bands that Fermi actually observes. If LST-1 event data become accessible through the collaboration, we then run the full two-instrument fit. The observational setting motivates the work because standard practice fits hand-chosen templates. Abe et al. had to move from exponential to sub-exponential cutoffs to describe Geminga across 20–65 GeV. The Crab analysis found an inverse-Compton peak broader than simple models predict. PKS 1424+240 modeling ruled out one-zone emission. In each case the number of components was decided by the modeler, not the data.

We do not claim the learned singular functions are individually physical, since SVD factors are mixed and signed. We claim only that their count estimates the number of separable components, and that the fitted intensity inside the gap is a template-free prediction.

Baselines

The naive reference is the published parametric analysis of Abe et al.. It reports P2 at 12.2σ with power-law index $\Gamma = 4.5 \pm 0.4_{\text{stat}}{}^{+0.2}_{-0.6}{}_{\text{sys}}$ over 20–65 GeV, and P1 at only 2.6σ. It implicitly assumes one spectral component per peak. On semi-synthetic data we fit the same template family by forward-folded Poisson likelihood with identical exposures and calibration priors. Our method beats it if it recovers the injected rank, including the hidden gap component, while the template fit either misses that component or shows a significant misfit. We also require lower held-out Poisson deviance inside the masked band.

The strongest competitor is a binned Poisson non-negative matrix factorization with spline-parameterized factors. It is fit separately for each candidate rank from 1 to $K$, with rank chosen by BIC and by held-out deviance, using the same calibration nuisances. For a two-variable problem this is a genuinely strong method. It gets the same data, the same calibration priors, and a tuning budget in spline knots and bins equal to our architecture search. We report wall-clock time for its full rank sweep against our single run. Beating it means matching or exceeding its rank-recovery accuracy and in-gap deviance at a fraction of the compute, or clearly outperforming it at low photon counts where binning discards information. No reported result found on Lacuna for this baseline on these data.

Two more baselines isolate our ingredients. The first is NeuralSVD with fixed $L$, trained once per candidate $L$ and selected by the same held-out estimator. This isolates the "one run" claim: prefixes of a jointly nested model should select the same rank as separately trained models. The file gives no scalar for NeuralSVD itself, only that it beats SpIN and NeuralEF by roughly an order of magnitude on hydrogen and oscillator eigen-accuracy, so we reproduce it. The second is Riemannian Pursuit, which grows the rank one manifold step at a time while observed-entry residuals stay high, run on the binned exposure-corrected matrix. It represents greedy rank growth for incomplete matrices.

Ablations remove the calibration nuisances, which should inflate rank by one when offsets are injected. They also replace nesting with an unordered overspecified model that relies only on Ding-style implicit collapse, which tests whether ordering or step size is doing the work.

Why not just…?

Why not just fit spline-based Poisson NMF at every rank and pick by BIC? This is the real threat. For Geminga's two-variable phase–energy problem it may tie with our method on accuracy. What such a tie would leave our method is the lower cost of a single run, an unbinned estimator that should matter at the lowest photon counts, and a formulation that extends to further inputs such as observation time for variable sources, where tensor splines grow awkward. The semi-synthetic sweep over photon count settles this directly. If NMF matches us at every count level and cost is not decisive, we will report that the neural parameterization is unnecessary for pulsar phase–energy data. The contribution would then shrink to the nested rank readout and the calibration-aware objective, both of which transfer to spline factors.

Why not just use the physical templates, since the emission physics is known? The file's own evidence shows those templates keep being revised at exactly the gap boundaries: sub-exponential cutoffs for Geminga, a broader-than-expected Crab peak, failed one-zone blazar fits. Our method does not replace physical modeling. It tells the modeler how many components the data support before templates are chosen. The hidden-component simulations test whether that information is real.

Why not just bin the data, build the matrix, and read the rank from a scree plot, via Nyström or randomized SVD, or from stochastic spectral density estimation? These methods need the full matrix or matrix–vector products with it. Here an entire energy block is missing, and the two observed blocks are miscalibrated relative to each other. Imputing the gap first presupposes the rank. Running these tools on the naively corrected matrix is included as a check, and we expect the calibration step to show up as a spurious singular value.

Proposed Experiments

The decisive checks are as follows. The first is rank-recovery accuracy on truncated oscillator and hydrogen operators and on semi-synthetic Geminga data, as a function of photon count and of the weakest component's singular-value ratio, compared against the per-rank NeuralSVD sweep and spline NMF. The second is whether the injected gap-dwelling component is detected by our method and missed by per-peak templates. The third is whether removing calibration nuisances inflates the estimated rank by exactly one. The fourth is agreement between nested-prefix rank selection and separately trained models at a fraction of the compute.

On real Fermi data with artificially masked bands, the fit must predict the withheld photons better than the template fit. If LST-1 data are obtained, we will test whether the estimated rank across 20–65 GeV exceeds the two-peak assumption underlying the 2.6σ and 12.2σ figures.