Assumption Maps for PeVatron Verdicts

Whether a gamma-ray source is a "PeVatron" (an object that accelerates protons to the peta-electron-volt energies needed to explain the knee of the cosmic-ray spectrum) is decided by fitting a physical emission model to data from several telescopes. That fit is run under fixed assumptions: the density of the gas the protons hit, the distance to the source, the infrared photon field that electrons scatter off, and the shape of the injected particle spectrum. The verdict can hinge on those assumptions. We propose to stop reporting one fit at one set of assumptions. Instead, we train a small neural network that maps any assumption vector to the best-fit physical parameters and fit quality at those assumptions. Reading this map shows exactly where in assumption space "hadronic" beats "leptonic", and where the inferred proton cutoff passes 1 PeV.

The optimization problem is single-objective for each assumption setting: one joint likelihood over all instruments, which gives a total ordering of candidate fits. What makes it hard is that its curvature and gradient noise change by orders of magnitude as the assumptions vary. We therefore train the map with step-size rules whose guarantees do not require knowing smoothness or noise constants. These are AdaGrad-Norm and normalized SGD. The deliverable is a tool that turns published single-point verdicts, such as the Cassiopeia A proton energy limit, into verdict regions with checked accuracy. It also gives a test of whether amortized parametric optimization beats a dense grid of independent fits once there are more than a few assumptions.

Core Claim

The research file frames this intersection as multi-objective: trace Pareto fronts between radio, X-ray and TeV channels. We think that framing misreads the physics. Independent instruments combine through a joint likelihood whose relative weights are fixed by the counting statistics, not chosen by hand. So there is no preference vector to sweep, and a scalar fit is the correct objective rather than a crude scalarization. The real hidden freedom sits in the fixed nuisance assumptions.

The Cas A analysis in Does or Did the Supernova Remnant Cassiopeia A Operate as a PeVatron? is an example. It derives a 100–1000 TeV proton energy upper limit of $3.6\times10^{47}$ erg explicitly "assuming a gas density of 10 cm$^{-3}$" and an $E^{-2}$ injection. From that it concludes Cas A supplies under 1% of the Galactic knee injection rate. The multiwavelength study of LHAASO J2108+5157 leaves a leptonic reading (magnetic field $B<1.2$–$1.9\,\mu$G) and a hadronic reading (molecular clouds at 2.0 or 3.1 kpc) side by side. The Crab (Peta–electron volt gamma-ray emission from the Crab Nebula) and 1LHAASO J0056+6346u likewise admit leptonic and lepto-hadronic fits with very different energetics. These papers motivate the proposal: they document verdicts that depend on assumptions. None of them reports how the verdict moves when the assumptions move.

Formally, let $\lambda$ be the assumption vector (gas density, distance, photon-field energy density and temperature, injection index, cloud filling factor). Let $\theta$ be the fitted physical parameters (particle cutoffs, spectral indices, $B$, energy budgets). Let $\ell(\theta;\lambda)$ be the joint negative log-likelihood of all instrument data under a hadronic or leptonic forward model. We train a network $\theta_w(\lambda)$ with weights $w$ by minimizing

$$\mathcal{J}(w)=\mathbb{E}_{\lambda\sim q}\,\big[\ell(\theta_w(\lambda);\lambda)\big],$$

where $q$ is a sampling distribution over the physically allowed assumption box. At the optimum, $\theta_w(\lambda)$ approximates the profile solution $\theta^*(\lambda)=\arg\min_\theta \ell(\theta;\lambda)$ for every $\lambda$ at once.

The verdict map is the sign of the likelihood difference between the hadronic and leptonic profiles at each $\lambda$, together with whether the fitted proton cutoff exceeds 1 PeV. The method borrows from Pareto Front Shape-Agnostic Pareto Set Learning, which trains a network to map samples from a latent distribution to optimal solutions. We keep that machinery but replace "preference over objectives" with "physical assumption", and replace hypervolume with a likelihood. This removes the reference-point and scale-imbalance failures that the GPSL paper reports on Disc Brake Design.

Research Program

Forward model. We will re-implement the standard non-thermal radiation kernels in an automatic-differentiation library, so that $\ell$ has exact gradients in both $\theta$ and $\lambda$. These kernels are synchrotron, inverse Compton with Klein–Nishina corrections, and proton–proton pion decay, i.e. what naima and jetset compute. Before any learning, we validate the kernels against naima outputs.

Data. We will use published flux points and upper limits from the cited papers: Crab (LHAASO), Cas A (LHAASO KM2A limits plus the $^{12}$CO gas column), G106.3+2.7 head and tail (MAGIC, with DRAO radio and XMM-Newton), LHAASO J2108+5157 (LST-1, XMM-Newton, 12 years of Fermi-LAT), and LHAASO J0341+5258. Raw air-shower events are collaboration-internal. We therefore state plainly that the likelihood is built on published spectral points, with upper limits treated as one-sided Gaussian or Poisson terms.

Optimization. Gradients of $\mathcal{J}$ are stochastic because each step samples a minibatch of $\lambda$ values. We expect their scale to vary strongly across the box. At low gas density the hadronic model needs enormous proton budgets. Near an X-ray upper limit, the synchrotron penalty is steep. SGD with AdaGrad Stepsizes proves high-probability convergence without knowing the smoothness constant or the noise level, under affine variance. From Gradient Clipping to Normalization for Heavy Tailed SGD shows that normalized SGD needs no knowledge of the tail index under heavy-tailed noise. Both supply our method. Both are proved for smooth objectives, and the kink at an upper limit breaks smoothness. We will therefore smooth the one-sided terms and report how the damping constant in AdaGrad's denominator affects results, since that paper shows the tuning burden shifts onto it.

We will measure the empirical tail index of the gradient noise. If it is not heavy-tailed, the normalized-SGD argument does not apply, and we will say so.

Certification. Every point on the map is only as good as the network's gap to the true profile at that point. We therefore run direct per-point fits at held-out $\lambda$ values. We report the distribution of the likelihood gap, and treat a gap below 0.5 in $\ell$ (one standard deviation for one parameter) as "accurate". We considered the anytime-valid stopping rules of Beyond Optimal Rates in Stochastic Optimization and do not use them. They require the strong-convexity constant $\mu$, which is exactly what we cannot know here.

Baselines

Published single-assumption fit (naive reference). This is each paper's own fit at its stated assumptions. It is the reference point, not a competitor. Our map must reproduce each published result at the published $\lambda$ before any other result counts. The Cas A result gives a numerical target: the $3.6\times10^{47}$ erg limit at 10 cm$^{-3}$. For Crab, the target is $B\approx110\,\mu$G. For J2108 under leptonic assumptions, it is $B<1.2$–$1.9\,\mu$G. For G106.3+2.7 tail, it is a proton cutoff near 1 PeV with index $1.83\pm0.10$. Failing to reproduce these would indicate a forward-model bug.

Dense grid of warm-started independent fits (strongest competitor). This is the obvious practitioner's approach and the real threat. We cover the same assumption box with a Sobol grid, run a standard minimizer at each node warm-started from its neighbor, and interpolate the verdict. Both methods get the same forward-model and gradient budget, counted in kernel evaluations, and the same held-out test points. For $\lambda$ of two or three dimensions we expect the grid to match or win. The proposal succeeds only if, at five to eight assumption dimensions, the amortized map meets two conditions under equal budget. It must reach a lower 90th-percentile likelihood gap. It must also misclassify fewer held-out points' hadronic-versus-leptonic verdict. We found no reported result on Lacuna for amortized profile fitting of gamma-ray SEDs, so there is no external number to beat.

MCMC marginalization with broad priors. The second competitor puts priors on $\lambda$ and samples everything jointly, as naima's sampler-based fitting does. It will be run with the same forward model and budget. It answers a different question, the averaged posterior, so we compare it on whether its marginal verdict agrees with the fraction of prior mass our map assigns to each verdict.

Optimizer ablation. This ablation isolates the adaptive ingredient. The same network is trained with Adam and with SGD, each given a learning-rate grid of eight values. AdaGrad-Norm and normalized SGD get a single untuned run each. For normalized SGD versus clipping, the only prior evidence is qualitative: on Penn Treebank and WikiText-2, NSGD matched Clip-SGD loss with no threshold search, and no numerical loss is reported. Adaptive methods "win" if their single run lands within the spread of tuned Adam's best run at lower total cost including the tuning. They also win if tuned Adam fails on some sources, for example diverging on the Cas A hadronic model at low density.

Why not just…?

Why not just trace Pareto fronts between instruments, as the multi-objective literature suggests? Pareto Multi-Task Learning is right that weighted sums miss concave parts of a trade-off front. That matters when the weights are an arbitrary user preference. In statistical fitting, the weights on radio, X-ray and TeV terms are fixed by the likelihood, so a front across instruments mostly shows fits that no single dataset would accept. Where cross-instrument tension genuinely signals a wrong model, a per-channel goodness-of-fit at the profile solution detects it more directly. We will report that diagnostic on each map rather than a front.

Why not just run a grid over the assumptions? In low dimensions this is probably as good and simpler, and we concede that case up front. The grid's cost grows exponentially with the number of assumptions. The amortized map's cost is set by the smoothness of $\theta^*(\lambda)$. The equal-budget comparison at two, four, six and eight assumption dimensions will show where, if anywhere, the crossover sits. If the grid wins everywhere up to eight dimensions, the contribution reduces to the verdict maps themselves, computed by grid. That is still scientifically useful, and we would report it that way.

Why not just marginalize the assumptions in an MCMC? Marginalizing requires priors on gas density and photon fields that the cited papers do not have. The Cas A limit is quoted at a single density precisely for this reason. A marginal verdict mixes the prior with the data, and a reader cannot see which assumption drove it. The map keeps that dependence visible. The MCMC comparison above checks that the two agree when the priors are agreed on.

Why not just use Adam with a tuned learning rate? It may well work. The optimizer ablation charges tuning cost honestly, and a win for tuned Adam would simply remove the adaptive-step component from the recommendation.

Proposed Experiments

Synthetic injection–recovery. We simulate spectral points at realistic instrument sensitivities from known hadronic and leptonic truths across a six-dimensional assumption box, then compare recovered verdict boundaries with the true boundary. The measures are misclassification rate and likelihood gap, measured against both baselines at matched budget.

Real sources. We build hadronic and leptonic maps for Cas A, G106.3+2.7 tail, LHAASO J2108+5157, LHAASO J0341+5258 and the Crab. Every map must first pass the reproduction checks above. We then report scientific quantities read off the maps. For Cas A, the gas density below which the 1% conclusion fails. For J2108, the photon-field and distance region in which the leptonic reading needs a physically implausible $B$. For the Crab, whether any hadronic region survives the LHAASO PeV points.

Decisive check. Everything rests on one experiment: the equal-budget, held-out comparison of the amortized map against the warm-started grid as the number of assumptions grows. If the map does not beat the grid at six or more dimensions on both likelihood gap and verdict error, we drop the amortization claim. We keep the maps and the adaptive-optimizer ablation as the result.