How Many Background Models Make a Trustworthy Galactic Center Likelihood Ratio?

Gamma-ray telescopes see more light from the center of our galaxy than standard models predict. The central dispute is whether that extra light comes from a large population of faint, unresolved pulsars or from a smooth glow produced by dark matter particles annihilating. Researchers settle this with likelihood-ratio tests: they compare how well a "many point sources" model and a "smooth halo" model fit the photon map. Those numbers are only as trustworthy as the model of everything else in the sky, chiefly gamma rays from cosmic rays hitting interstellar gas. Only a handful of genuinely different background models exist. Forensic speaker recognition has studied this exact problem for fifteen years under another name. When the reference population used to turn a similarity score into a likelihood ratio contains fewer than about 30 independent members, the reported evidence can swing by many orders of magnitude even though the system still discriminates well.

This note proposes treating each distinct interstellar-gas background model as one "speaker" in a reference population. The Galactic Center test statistic is then treated as a raw score that must be calibrated into a log-likelihood ratio (a number that can be read directly as evidence and combined with prior odds). Calibration uses Bayesian and regularised methods taken from forensic voice comparison, and is evaluated on held-out background models rather than on held-out simulations. The core claim is that the effective sample size of Galactic Center evidence is the number of independent background models, not the number of Monte Carlo skies. With that counting, plug-in evidence for or against dark matter is strongly overconfident in a way that standard simulation-based significance estimates cannot detect. The matter–antimatter connection enters downstream. Dark-matter models that tie the relic abundance to the baryon asymmetry, or route annihilation through the neutrino-mass sector, cite the excess as support, so the proposal asks how much evidential weight the excess can actually carry for them.

Core Claim and Why It Is New

The forensic literature establishes three facts we reuse.

  • Ishihara and Kinoshita computed multivariate likelihood ratios on Japanese pitch features while varying the reference population from 10 to 120 speakers. Below about 30 speakers, likelihood ratios varied by up to 13 orders of magnitude across population draws. At 10 speakers the equal-error threshold sat at an average log LR of −4.02 instead of 0, even though the equal error rate was a usable 17.2%. Good discrimination, then, did not mean the numbers were calibrated.
  • Brümmer et al. showed why. Plug-in calibration treats parameters fitted on scarce background data as exact, and so produces extreme likelihood ratios. Averaging over the parameter posterior pulls them toward 1.
  • Wang and Hughes 2022 showed that skewed score distributions make this worse.

These papers motivate the proposal and supply its methods. All of them work with speakers and recordings, and none touches astrophysical hypothesis tests.

On the astrophysics side, Improved modeling of the discrete component of the Galactic interstellar γ-ray emission supplies the motivating evidence. It used less optically thick carbon-monoxide tracers from the Mopra survey to build a gas template that the standard maps miss. That template carried 15.4% of the molecular-gas-related emission in the surveyed strip. The same analysis found 23 new point-source candidates, of which 8 matched unassociated catalog sources (p = 0.012). The direct implication is that gas clumps the background model omits get read as point sources. Swapping one background model for another therefore moves the point-source-versus-smooth evidence. Background models are few, correlated, and expensive to produce, which makes them the "speakers" of this problem.

The research file contains no prior work that calibrates Galactic Center likelihood ratios, or counts background models as the reference population. That grouping is the specific contribution. Simulations drawn under one background model play the role of repeated recordings of one speaker. They add precision about that speaker and nothing about the population.

Research Program

The score. For each sky, the raw score is a point-source-versus-smooth test statistic from a template fit that includes a pulsar-like population component. We also run a variant where the score is the output of a classifier trained on simulated skies, so the conclusion does not depend on one scoring method.

The background models. We build a pool of K independent background models. They come from GALPROP cosmic-ray propagation runs with different gas maps, propagation parameters, and source distributions. They are deliberately made to differ in how they treat dense gas: with and without the Mopra-informed modified template, and with different carbon-monoxide-to-hydrogen conversions.

The simulated skies. For each background model we simulate skies containing either a millisecond-pulsar population or a smooth dark-matter profile of the Navarro–Frenk–White type, with matched total excess flux. The pulsar luminosity functions and dark-matter profile slopes are varied.

The calibrators. Calibration maps score to log-likelihood ratio. We compare four methods:

  • affine logistic regression;
  • regularised logistic regression;
  • Bayesian calibration in the style of Brümmer et al., extended to skew-tolerant score models because the Gaussian assumption is unlikely to hold;
  • Empirical Lower and Upper Bounds, the bounding method tested by Wang and Hughes 2021.

The protocol. For each K, drawn many times from the pool, we fit each calibrator on skies from the K drawn models. We then score it on skies from models outside the draw. The metrics are those of Wang and Hughes 2021:

  • mean log-likelihood-ratio cost, a proper-scoring measure that penalises both poor discrimination and miscalibration and equals 1 for a system that always reports LR = 1;
  • the range of that cost across draws, which measures stability.

The prediction. Plug-in methods should show a sharp rise in cost range below some critical K, analogous to the N ≈ 30 boundary in speech. If the forensic results transfer, Bayesian calibration should flatten that rise at a modest cost in mean accuracy.

The accuracy–stability tension identified in the forensic work will reappear here. Logistic regression had mean cost 0.37–0.38 against 0.54–0.56 for the Bayesian model in Wang and Hughes 2021. We address it with a criterion the forensic papers lack. The trade-off is judged by cost on held-out background models only, because that is the quantity relevant when the true sky is not among the models we built.

The downstream link to baryogenesis models. Leptophilic cascades are needed to avoid the antiproton bounds from Probing dark matter annihilation in the Galaxy with antiprotons and gamma rays and Cosmic signals from the hidden sector. Two families of models in that space draw support from Galactic and cosmic-ray anomalies:

  • the type II seesaw scalar dark-matter models of Okada et al. and Okada et al. 2018b, whose triplet mediator also carries the lepton-number violation behind leptogenesis;
  • the assimilation mechanism of D'Eramo, Lin and Thaler, which ties the relic density to the baryon asymmetry and which its authors note predicts weak indirect signals.

We do not refit these models. We report the calibrated smooth-versus-point-source log-likelihood ratio as a quantity model builders can multiply into their own prior odds. We also report how much the "excess as dark-matter evidence" input shrinks once background-model scarcity is priced in. This is honestly an application regime, not a contribution to baryogenesis theory.

Baselines

Naive reference: Wilks' theorem. The test statistic is converted to significance via Wilks' theorem, with no calibration. It is the current default in gamma-ray analyses and the point any improvement must clear. We run it on exactly the same simulated skies. It beats nothing by construction, so it serves to show how far uncalibrated evidence sits from zero cost on held-out backgrounds.

Strongest existing practice: Monte Carlo calibration under a fiducial background. The null and alternative test-statistic distributions are estimated from many simulated skies under the analysis's chosen background model, and evidence is read off those distributions. This is the strongest competitor because it is what careful practitioners actually do, and it is often well calibrated in-model. We give it the same total simulation budget as our method, all concentrated on one model. That is the plug-in limit with K = 1 and unlimited recordings.

Strongest calibration method in the forensic literature: Bayesian calibration. The best reported result on a comparable small-population question is in Wang and Hughes 2022. With 20 speakers and skewed i-vector scores, Bayesian calibration reduced the cost range from 1.3 under logistic regression to about 0.31 (about 0.60 for GMM-UBM). Wang and Hughes 2021 reports regularised and Bayesian ranges around 0.1 or lower from 20 to 100 speakers. For the Galactic Center task itself: no reported result found on Lacuna for calibrated log-likelihood-ratio cost on pulsar-versus-dark-matter discrimination.

Ablation of the key ingredient. The ablation runs the identical Bayesian calibrator but treats every simulated sky as independent, ignoring which background model generated it. This isolates the claim that grouping by background model is what matters. Beating it means a lower cost range on held-out backgrounds at equal simulation budget.

What counts as success. The proposal succeeds if grouped Bayesian or regularised calibration shows a held-out cost range at small K that is materially below plug-in logistic regression and below the fiducial Monte Carlo method. It must do so without a mean-cost penalty larger than the roughly 0.17 gap seen in speech.

Why not just…?

Why not just calibrate the test statistic with Monte Carlo under the chosen background, as everyone does? This is the real threat, and the proposal is designed to settle it directly. Such calibration is correct conditional on the background model being true. The Mopra result shows that a single omitted gas component carrying 15.4% of molecular emission manufactures point sources (Improved modeling). That is precisely the condition under which the fiducial background is not true. If fiducial Monte Carlo calibration attains held-out-background cost comparable to grouped Bayesian calibration, the core claim fails, and we will report that.

Why not just fit a few alternative background models and report the spread of results? Envelope reporting does acknowledge the problem. But a spread is not an evidence value, and it cannot be combined with prior odds or downstream model likelihoods. The forensic analogue, hard bounding with Empirical Lower and Upper Bounds, performed worst in Wang and Hughes 2021. It had mean cost 0.65–0.66, and its range at 100 speakers was as bad as logistic regression's at 20. We include both an envelope-based conservative LLR and that bounding method as calibrators, so the comparison is empirical.

Why not just train a neural classifier on skies pooled across many backgrounds? Its output already approximates a likelihood ratio. That holds only for the backgrounds it saw. With few backgrounds, the classifier inherits the small-population problem: a model can discriminate well and still report miscalibrated evidence, as with the 17.2% equal error rate sitting at log LR −4.02 in Ishihara and Kinoshita. Our classifier-score variant tests this directly. It compares raw classifier output against the same output after grouped Bayesian calibration, both on held-out backgrounds.

Why not assume Gaussian scores and use closed-form calibration? Constrained maximum-likelihood Gaussian calibration matched logistic regression on telephone speech, with mean cost 0.375 versus 0.376 (The distribution of calibrated likelihood-ratios). However, it requires Gaussian scores. Test statistics near a physical boundary are skewed, and Wang and Hughes 2022 shows that skew is exactly when small samples hurt most. We measure score skew per background model and run the Gaussian method as an additional calibrator.

Proposed Experiments and Decisive Checks

Stage 1: reproduce the forensic numbers. We first reimplement the calibrators on simulated scores matching the protocol of Wang and Hughes 2022. The implementation is accepted only if it recovers a logistic-regression range near 1.3 and a Bayesian range near 0.31 for 20 skewed i-vector-like speakers. This guards against attributing implementation bugs to physics.

Stage 2: the core experiment on simulated Galactic Center skies. We sweep K from 3 upward across a pool of roughly 20 or more GALPROP-derived background models. For each K we use many random draws and score on held-out models.

Stage 2 decisive check. The decisive check is whether plug-in cost range rises sharply at small K while grouped Bayesian range stays flat, and whether the ungrouped ablation behaves like plug-in. A flat plug-in curve would falsify the transfer of the forensic population-size effect.

Stage 3: real-data check in the Mopra strip. The Mopra strip is where a better gas truth exists. We apply calibrators fitted without the modified template to the 23 candidate sources from Improved modeling. We then check whether calibrated point-source LLRs for candidates later explained by gas fall toward zero, while plug-in values stay large.

Stage 4: application to the Galactic Center. Only after the earlier stages succeed do we apply the method to Fermi-LAT Galactic Center data and report a calibrated smooth-versus-point-source LLR. Its uncertainty is reported as a function of K. We state explicitly what weight that leaves for leptophilic and asymmetry-linked dark-matter models.