The Galactic Center Excess is an unexplained glow of GeV gamma rays from the inner Milky Way. It is read either as dark matter annihilating or as a crowd of faint pulsars. In both readings it is a residual: what remains after astronomers subtract a model of gamma rays made when cosmic rays hit interstellar gas. The weakest part of that subtraction is molecular hydrogen. Molecular hydrogen is invisible to radio telescopes, so it is traced through carbon monoxide (CO). The common form of CO saturates in dense clouds, much as an overexposed photograph saturates, so the densest gas is undercounted. Rarer CO isotopes do not saturate, but they have been mapped only over a thin strip of the Galactic plane.
A neural network trained on that strip can predict the unsaturated map from the saturated one (Deep Learning Models of the Discrete Component of the Galactic Interstellar Gamma-Ray Emission). The network returns one best-guess map with no error bars. The prediction is then used most heavily exactly where no ground truth exists. This proposal replaces the point prediction with a non-linear regression that returns a spatially correlated uncertainty for every predicted pixel. It turns that uncertainty into a small set of "wiggle" templates that the gamma-ray fit is allowed to absorb. Then it measures whether conclusions about the excess survive. The question is not whether a kernel model beats a convolutional network at regression. The question is whether gas-map uncertainty, honestly propagated, changes how much excess flux the fit assigns and how many spurious point sources it reports.
Core Claim
Our claim is that learned corrections for saturated CO should be judged by downstream calibration, not by pixel accuracy. Two quantities matter. The first is whether the uncertainty on the inferred excess flux covers the truth at its nominal rate in controlled injection tests. The second is whether gas-induced false point sources disappear. We build a model whose predictions come with a full covariance across sky positions and velocities. We compress that covariance into a handful of principal modes and add those modes as constrained nuisance templates in the Fermi Large Area Telescope (Fermi-LAT) likelihood.
The method is model-agnostic at the propagation step. We therefore test both a kernel regressor, which gives the covariance in closed form, and an ensemble of the published network, which approximates it empirically. Which one gives better-calibrated downstream inference is an open empirical result, not an assumption.
Research Program
The motivating evidence comes from the astrophysics baseline (Improved modeling of the discrete component of the Galactic interstellar γ-ray emission…, duplicate record; Improved modeling of the discrete component of the Galactic interstellar γ -ray emission and implications for the Fermi–LAT Galactic Center excess; Improved modeling of the discrete component of the Galactic interstellar γ -ray emission and implications for the Fermi–LAT Galactic Center excess). That work computes molecular-hydrogen column densities from the unsaturated isotopes ¹³CO and C¹⁸O in the Mopra Southern Galactic Plane CO Survey. It converts them to gamma rays with the GALPROP cosmic-ray propagation code. It finds that this discrete gas component accounts for about 15.4% of hydrogen-related gamma-ray emission in the Mopra region. In simulated 12-year Fermi-LAT data, this component alone produces 23 candidate point sources with test statistic TS ≥ 16, the threshold for listing a source. Eight of those coincide with unassociated catalog sources. Gas error is therefore not a nuisance at the percent level: it manufactures discrete structure. That is the very signature used to argue for pulsars over dark matter.
The convolutional network supplies the regression setup we reuse. It takes 7×7-pixel patches across 17 velocity bins, uses a separate spatial filter for each velocity bin, and is trained with a Poisson loss. It recovers about 83.7% of the true discrete-gas flux, and 83.6% of pixels fall within 1σ of the observation-derived counts. It was trained and validated inside the Mopra strip. It reports no uncertainty, which matters because the region where the map is needed lies outside that strip.
The first component we build is the regressor. The input is the same 12CO spectral patch, a cube of 7×7 sky positions by 17 velocity bins. We use a product kernel, meaning one similarity function for the spatial mode multiplied by another for the velocity mode. This borrows the separable multi-way kernel idea from Multi-way Multi-level Kernel Modeling for Neuroimaging Classification. That paper kernelizes tensor factorizations of 3D brain scans and reports that they outperform a 3D network in small-sample settings. We use only its kernel structure, not its classifier.
Exact kernel regression costs time cubic in the number of pixels, so we approximate the kernel with a few thousand landmark points, in the style of A Scalable Kernel-Based Algorithm for Semi-Supervised Metric Learning. That paper uses such low-rank approximations to keep non-linear separation tractable. We fit a Gaussian-process version of this regressor on log ¹³CO intensity, with noise variance that depends on the input. The resulting posterior gives a mean map $\mu$ and a covariance $\Sigma$ over output pixels and velocities. The predictive variance grows automatically as inputs move away from the training data.
The second component is propagation. For a fixed cosmic-ray density, the gamma-ray map from gas is linear in gas column. The covariance of the gamma-ray template is therefore $\Sigma_\gamma = A \Sigma A^\top$, where $A$ is the linear GALPROP emissivity-and-projection operator. We eigendecompose $\Sigma_\gamma$ and keep the top $K$ modes. The template entering the fit becomes
$$t(a) = A\mu + \sum_{k=1}^{K} a_k \sqrt{\lambda_k}\, u_k, \qquad a_k \sim \mathcal{N}(0,1),$$
where $\lambda_k$ and $u_k$ are the $k$-th eigenvalue and eigenvector of $\Sigma_\gamma$, and the $a_k$ are nuisance amplitudes. The Gaussian prior on each $a_k$ is added to the Poisson likelihood. The excess and point-source components must then compete against gas fluctuations that the regressor considers plausible, and no others. We choose $K$ to capture 90% of the trace of $\Sigma_\gamma$ and report sensitivity to that choice.
Baselines
The naive reference is the standard practice: a ¹²CO-only gas template scaled by an empirical CO-to-hydrogen conversion factor and passed through GALPROP. It is run on the same simulated sky with the same likelihood code. Its known failure is the missing 15.4% of hydrogen emission and the 23 spurious TS ≥ 16 candidates.
A second simple baseline is a per-velocity-bin monotone curve that maps ¹²CO intensity to ¹³CO intensity, fit on the training region. It encodes saturation directly with almost no capacity. If it matches the richer models, spatial context is not needed.
The strongest existing method is the convolutional network of Deep Learning Models…, reimplemented to its published architecture and loss. Its reported numbers are 83.7% flux recovery, 83.6% of pixels within 1σ, and a detection significance of √TS = 27.5 for its template on the simulated 12-year data. The observation-derived template in the companion paper is reported as "mean TS of 48.30 ± 1.02." The file states these two figures in different units, so we will recompute both in one pipeline rather than compare them as quoted.
The strongest uncertainty-aware competitor is a deep ensemble of ten copies of that network, trained with different random seeds and data orderings. Its empirical covariance across members is fed through the identical propagation step. All learned models get the same training pixels, the same spatial cross-validation folds and the same budget of 50 hyperparameter configurations, selected on a validation block.
Two ablations isolate the key ingredient. The first uses the kernel model's mean map only, with $K = 0$. The second uses only the per-pixel variance, so that spatial correlations are discarded. An oracle fit using the actual ¹³CO-derived template bounds what any method can achieve.
The proposal beats the CNN in three cases:
- it matches or exceeds 83.7% flux recovery on held-out blocks;
- its 68% intervals on injected excess flux cover the truth in 60–76% of injections where the CNN's point-estimate fit undercovers;
- it reduces spurious TS ≥ 16 sources toward the oracle count.
If the deep ensemble matches the kernel model on these criteria, we will report that the propagation step, not the kernel, is the contribution.
Why not just…?
Why not just use the published CNN, which already recovers most of the flux? Pixel-level recovery of 83.7% leaves roughly a sixth of the discrete flux unexplained. That missing structure is clumpy, and clumpy gas residuals are what the companion paper shows turning into fake point sources. A single map gives the downstream fit no way to tell modeled gas from unmodeled gas. The point-estimate ablation tests exactly this: if it gives calibrated excess-flux intervals, our propagation is unnecessary.
Why not just train an ensemble of CNNs for uncertainty? This is a real threat, and we treat it as a co-equal arm rather than a strawman. Ensembles can be overconfident far from training data, because members often agree in regions none of them has seen. A Gaussian process's variance, by construction, reverts to its prior there. That theoretical edge may not survive contact with real CO cubes. The leave-longitude-block-out calibration curves will settle it. If ensembles calibrate as well, we adopt them.
Why not just free the CO-to-hydrogen conversion factor in Galactocentric rings, as template fits already do? Ring-wise scaling adjusts gas amplitude over large annuli. It cannot create the sub-degree, velocity-resolved clumps that saturation hides, which are the clumps that seed false point sources. We run this as an additional baseline and expect it to leave the spurious-source count near the naive level. If it does not, the simpler fix wins.
Proposed Experiments
All training uses Mopra Data Release 3 ¹²CO and ¹³CO cubes. Data are split into contiguous longitude blocks, holding out one block at a time, so that every test mimics extrapolation rather than interpolation. The regression metrics are flux recovery, the fraction of pixels within 1σ, and calibration: the empirical coverage of 50%, 68% and 95% predictive intervals, plotted against distance in kernel space from the training set. The last metric tests directly whether uncertainty grows where it should.
The downstream experiments use the simulated 12-year Fermi-LAT sky from the baseline papers. For each held-out block, we build the "true" sky from the observed ¹³CO gas. We then inject one of two signals: a smooth, dark-matter-like excess with a range of fluxes, or a population of faint point sources with the same total flux. We fit each sky with every baseline's gas template. For each method we record the bias and interval coverage of the recovered excess flux, the preference for point sources over smooth emission, and the number of TS ≥ 16 detections that coincide with gas rather than injected sources.
The decisive check is the smooth-injection case. A method whose gas model is wrong but overconfident will tend to attribute gas clumps to point sources, and will therefore misread a smooth signal as pulsars. The proposal succeeds only if propagating uncertainty restores correct classification at a rate close to the oracle.
The final step applies the same pipeline outside the Mopra strip, across the inner Galaxy, and reports how much the excess flux and its uncertainty change once gas-map uncertainty is included. That applied result is only as trustworthy as the calibration established in the held-out tests, and we will say so.
References
- Deep Learning Models of the Discrete Component of the Galactic Interstellar Gamma-Ray Emission
- Improved modeling of the discrete component of the Galactic interstellar <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mrow><mml:mi>γ</mml:mi></mml:mrow></mml:math>-ray emission and implications for the Fermi–LAT Galactic Center excess
- Multi-way Multi-level Kernel Modeling for Neuroimaging Classification
- A Scalable Kernel-Based Algorithm for Semi-Supervised Metric Learning