Tipping-Point Lipschitz Bounds for Gamma-Ray Emission Claims

When a gamma-ray observatory reports that a source accelerates particles to a peta-electron-volt, or that its emission comes from protons rather than electrons, the claim rests on a fitted physical model. That model is always an approximation of the true physics. A recent study showed how much this matters. Replacing the usual smooth approximation of how electrons lose energy with a simulation of individual, discrete collisions moved the inferred particle cutoff of one source from 1.57 to 1.17 PeV, a shift of more than 4σ (Discrete treatment of inverse Compton scattering). The error bars on such claims cover statistical noise but not the error of the approximation itself. We propose a way to put a hard, checkable bound on that approximation error and to report how robust each physical conclusion is to it.

The idea borrows one tool from safe machine learning. A function is Lipschitz-continuous with constant $L$ if it can change by at most $L$ per unit of input change. That single number lets you turn a few known values into a guaranteed envelope everywhere else. We apply this assumption to the gap between a cheap forward model and an expensive accurate one, not to the physics itself. A small number of expensive simulations then bounds the gap across the whole parameter space. Choosing $L$ is notoriously knife-edged: too small breaks the guarantee, too large makes the bound useless. So we do not pick it. We report the tipping-point constant $L^\ast$, the smallest smoothness level at which a stated conclusion (for example "cutoff above 1 PeV" or "leptonic preferred over hadronic") would flip. We compare it with an empirical lower bound on the true constant. A conclusion is called robust only when $L^\ast$ sits well above anything the simulations show.

Core Claim

Leptonic–hadronic ambiguity is usually framed as two smooth spectral models fitting equally well. In practice the ambiguity is made worse by unquantified forward-model error. The leptonic process is inverse Compton scattering, in which fast electrons boost low-energy photons. The hadronic process is pion decay, in which proton collisions with gas produce neutral pions that decay into gamma rays. Each has its own approximations: continuous energy loss for the first, parametrized collision cross sections for the second. Each approximation shifts the fitted spectrum in an energy-dependent way.

We claim three things. First, the discrepancy between an approximate and an accurate forward model can be bounded with Lipschitz cones built from a handful of expensive simulations. Second, this bound can be folded into a likelihood fit to produce intervals that stay honest under model misspecification. Third, the tipping-point constant $L^\ast$ is a better summary of evidential strength for physical claims than a single σ value.

On the fourth gap in the research file, discrete stochastic jumps versus smooth cones, our position is that the gap dissolves once the Lipschitz assumption is placed correctly. Individual electron trajectories jump. The expected photon spectrum, as a function of source parameters, is an average over many trajectories and is smooth. Monte Carlo noise in estimating that average is a bounded-noise term of exactly the kind Lipschitz-only safe optimization already handles.

Relation to Prior Work

Two papers supply the method. On Safety in Safe Bayesian Optimization introduced LoSBO, which certifies safety using only a Lipschitz constant and a deterministic noise bound, with no reliance on a correctly specified Gaussian process. It reached 90.9% of the global optimum against 82.45% for its rigorous Gaussian-process counterpart. Lipschitz Safe Bayesian Optimization for Automotive Control extended these cones to several simultaneous constraints on real hardware. We reuse their cone construction and noise-bound bookkeeping. We point it at a different object: the discrepancy between two forward models rather than a safety constraint. We also drop their assumption that $L$ is known.

Smooth Sailing supplies the inferential template and the motivation. It bounds worst-case bias under distribution shift with a Lipschitz assumption and obtains nominal 95% coverage where ordinary least squares, sandwich, and generalized least squares estimators fall toward 0%. It works on two-dimensional geographic coordinates with an expert-supplied $L$. We move the construction to bias that varies over physical source parameters and replace the supplied $L$ with tipping-point reporting.

Two papers constrain the design. Safe Exploration for Optimization with Gaussian Processes and the safe learning survey document the two failure modes of a misset $L$: an $L$ that is too small invalidates the guarantee, and one that is too large stalls progress. This is precisely why we refuse to fix $L$. Lipschitz Bandits without the Lipschitz Constant shows that adapting to an unknown constant is possible in bandits. It motivates our empirical estimate of $L$ but does not provide certified bounds, which is why we pair the estimate with $L^\ast$. Benefits of monotonicity in safe exploration shows that monotone structure tightens cones. We use it as an optional refinement, because the discrepancy from discrete energy loss grows with cutoff energy.

The discrete inverse Compton paper is both motivation and ground truth. It established the bias but treated it with unconstrained Bayesian sampling under a single alternative model, with no bound that transfers to new sources or new approximations. Combining Bayesian optimization and Lipschitz optimization mixes the two ideas only for generic benchmark functions. We found no work applying Lipschitz bounds to gamma-ray forward models.

Research Program

Let $\theta$ be the source parameters on a log scale: injection index, cutoff energy, magnetic field, photon-field density, and for hadronic models gas density and proton cutoff. Let $\delta(\theta)$ be the vector, over energy bins, of the difference in log flux between the accurate model (discrete-collision Monte Carlo, or a high-fidelity pion production code) and the approximate model used in standard fits. Suppose we have run the accurate model at design points $\theta_1,\dots,\theta_n$, each with Monte Carlo error at most $\epsilon_i$ (from a concentration bound on the sample mean). Then a Lipschitz assumption gives, in every energy bin,

$$|\delta(\theta)| \le \min_i \big( |\hat\delta(\theta_i)| + \epsilon_i + L\,\|\theta-\theta_i\| \big),$$

where $\hat\delta(\theta_i)$ is the simulated discrepancy at design point $i$.

The robust fit treats the true spectrum at any $\theta$ as the approximate spectrum plus some unknown discrepancy inside this band. It profiles the likelihood over the worst-case discrepancy. The resulting confidence set for $\theta$, and the likelihood ratio between the leptonic and hadronic hypotheses, then depend on $L$. For a binary claim $C$ we define

$$L^\ast(C) = \inf\{L : \text{claim } C \text{ is no longer supported at the chosen confidence level}\}.$$

We set this beside $\hat L$, the largest finite-difference ratio observed among simulation pairs. $\hat L$ is a valid lower bound on the true constant. It is not an upper bound, which is why robustness needs a margin.

Three components get built. The first is the robust profiled likelihood. Because the discrepancy enters additively in log flux, profiling it is a box-constrained inner problem; for Gaussian-approximated flux points it is a quadratic program, in the spirit of the noise-variance program in Smooth Sailing. The second is an active design rule borrowed from LoS-GP-UCB. The next expensive simulation is placed where the cone band is widest within the current confidence set, because width elsewhere cannot change any conclusion. The third is a pair of discrepancy libraries: continuous-versus-discrete inverse Compton scattering, and two pion-production parametrizations. Both are applied symmetrically, so neither emission hypothesis is privileged.

Baselines

The naive reference is current practice: a standard Markov chain Monte Carlo fit with continuous inverse Compton scattering, as done with Naima and threeML in Study of the IC 443 region with the HAWC observatory. It uses the same flux points and priors as our method. On the two sources with known discrete answers it gives 133.7 TeV for Geminga and 1.57 PeV for 1LHAASO J1954+2836u, against discrete-model values of 115.0 TeV and 1.17 PeV (Discrete treatment of inverse Compton scattering). We beat this baseline if our intervals cover the discrete values while its intervals do not.

The strongest existing method is the full discrete-model fit from that same paper, which is the gold standard but expensive. Fairness means matching its answer under a fixed budget of accurate simulations. We count how many our method needs to produce an interval that contains the full-fit estimate and is no more than, say, 1.5 times its width. The full fit's own simulation count is the yardstick. We have no reported compute figure for it on Lacuna, so we will measure it ourselves.

The strongest generic competitor is a Gaussian-process model of the discrepancy with a fixed ±2σ band. This is the heuristic that On Safety in Safe Bayesian Optimization found violated its guarantee in up to 28% of runs. It gets the same design points and a hyperparameter tuning budget matched to ours. We beat it by holding nominal coverage on synthetic injections where its coverage drops.

Two ablations isolate the key ingredients. The first replaces tipping-point reporting with a single expert-fixed $L$, which shows whether $L^\ast$ adds information beyond one cone. The second replaces active design with uniform random design at the same simulation budget. A third check imposes the Lipschitz assumption on single-realization spectra instead of ensemble means, to confirm that the placement argument about stochastic jumps holds in practice.

Why not just…?

Why not just run the accurate discrete model inside every fit? This is the real threat. If accurate simulation were cheap at every likelihood call, our bound would be unnecessary for that one approximation. Our method only earns its place if it reaches comparable intervals with far fewer accurate runs. A second justification is that it covers approximations for which no full accurate fit exists, such as the pion cross-section choice. The simulation-budget experiment above settles the first point. If our method needs nearly as many simulations as the full fit, we will say so and narrow the claim to cheap screening plus tipping-point reporting.

Why not just inflate the error bars by a fractional systematic? Discrete-loss effects matter mainly in the Klein–Nishina regime above roughly 100 TeV, where each collision can strip much of an electron's energy at once. They are negligible at lower energies. A uniform inflation large enough to cover the cutoff region over-widens the well-modeled part of the spectrum. One small enough to keep the low-energy fit sharp under-covers the cutoff. We test this directly with coverage and median width per energy decade on synthetic injections.

Why not just fit a Gaussian-process discrepancy model and use its posterior band? Its band is only as honest as its kernel and multiplier. The 28% violation rate for fixed multipliers (On Safety in Safe Bayesian Optimization) and the near-zero coverage of misspecified estimators in Smooth Sailing suggest it will under-cover exactly where data are thin. The synthetic coverage experiment measures this. If the Gaussian process matches our coverage with narrower bands, we will adopt it as the inner model and keep tipping-point reporting on top.

Why not just constrain the sampler with physical priors, as in "safe exploration" of parameter space? Evaluating an unphysical parameter value in a fit harms nothing. It is a computation, not a crash. Energetic limits such as pulsar spin-down power belong in the prior, and we include them. Safe-exploration machinery adds nothing to that question, which is why we do not build it.

Proposed Experiments

The first experiment is synthetic coverage. We generate spectra from the discrete inverse Compton model and from a high-fidelity pion model over a grid of true parameters. We fold them through published response approximations for HAWC and for LHAASO's two arrays: WCDA, a field of water ponds that detect particle showers in the low-TeV range, and KM2A, a square-kilometre surface array used above 10 TeV. We then fit with the approximate models. For each method we measure empirical coverage of the cutoff and index, interval width, and the rate at which a binary claim (PeV cutoff; leptonic versus hadronic) is called wrongly. We are not aware of a reported coverage rate for standard gamma-ray fits under this misspecification; we found no reported result on Lacuna for it, so it is measured here.

The second experiment is ground-truth recovery. On Geminga and 1LHAASO J1954+2836u we check whether robust intervals built from approximate fits plus at most a few dozen accurate simulations contain 115.0 TeV and 1.17 PeV. We also check whether $L^\ast$ for the claim "cutoff above 1.5 PeV" falls below $\hat L$, which would correctly flag the continuous-model claim as fragile.

The third experiment applies the method to open questions. We compute $L^\ast$ for the hadronic reading of the IC 443 compact core, which has a power law of index about 3.0 continuing past 30 TeV and a 95% proton lower limit near 300 TeV (LHAASO IC 443). We do the same for the leptonic preference in CTA1, with a parent-electron cutoff near 373 TeV (CTA1 with LHAASO), and for the hint of a hadronic component in the 60–500 TeV deviations of the Crab Nebula (Crab PeV emission). All three use published flux points.

The decisive check is simple. The proposal succeeds if robust intervals hold nominal coverage on synthetic data where standard fits do not, recover the two known discrete answers at a small fraction of the full-fit cost, and produce $L^\ast$ values that separate claims already known to be fragile from claims that survive. If $L^\ast$ is uninformative, meaning every claim flips at implausibly small $L$ or at none, the reporting idea fails and we will say so.