High-dimensional inference for the $γ$-ray sky with differentiable programming
GPU-accelerated probabilistic programming may help solve a decade-old gamma-ray mystery at the Milky Way's center.
The Thesis
For over a decade, astronomers have detected an unexplained glow of gamma rays at the center of our galaxy — called the Galactic Center Excess, or GCE — that could be dark matter annihilation, a dense population of rapidly spinning neutron stars called millisecond pulsars, or something else entirely. The core problem is that the galactic center is messy: gas, dust, magnetic fields, and competing radiation sources all overlap, and picking the right spatial model for any one of them changes your answer. This paper proposes using 'differentiable probabilistic programming' — a technique that fuses machine-learning-style gradient optimization with rigorous Bayesian statistics — to simultaneously consider a continuous range of possible spatial emission shapes rather than picking one and hoping it's right. The method runs efficiently on GPUs and uses variational inference (an approximation technique that finds the best probability distribution over all plausible models at once) to make previously intractable high-dimensional searches feasible. If the approach holds up, it represents a methodological step forward for astrophysical data analysis more broadly, not just for the GCE puzzle.
Catalyst
GPU hardware has become cheap and accessible enough that astrophysicists can vectorize large forward models — simulations that predict what a telescope would observe given a physical scenario — at scales that were impractical five years ago. At the same time, mature differentiable programming frameworks like JAX and PyTorch have made it straightforward to compute gradients through complex physical simulations, enabling variational inference methods that previously required hand-derived math. The GCE puzzle itself has sharpened: recent analyses have converged on a handful of competing hypotheses that differ mainly in their spatial morphology predictions, making a method that can compare many morphologies simultaneously particularly timely.
What's New
Earlier analyses of the GCE — including influential work using Non-Poissonian Template Fitting (NPTF), a technique that statistically distinguishes smooth emission from a sea of point sources — committed to a single spatial template for each emission component before fitting began. If the assumed template was wrong, the inferred source composition was biased in ways that were hard to quantify. This paper replaces those fixed templates with a parameterized, differentiable model that spans a continuum of spatial shapes, and uses variational inference to marginalize over that uncertainty — meaning the model averages over all plausible spatial configurations rather than picking the best single one.
The Counter
The GCE has resisted definitive explanation for over a decade despite increasingly sophisticated methods, and this paper is unlikely to change that. Variational inference — the approximation technique at the heart of the approach — trades exactness for speed; it finds the best approximation to the true posterior distribution, not the true posterior itself, and the gap between the two can be large in highly multimodal problems like the GCE. The paper showcases a framework but does not present a new physical conclusion about the GCE's origin: it doesn't, for example, definitively rule in or rule out dark matter. The comparison to prior methods is limited, and without a rigorous benchmark against NPTF or other established pipelines on identical data, it is hard to know whether the additional model flexibility produces meaningfully better-calibrated inferences or just more complex ones. Finally, the astrophysical community has seen many methodological proposals for the GCE that haven't resolved the underlying degeneracies — there is no guarantee that adding morphological flexibility breaks those degeneracies rather than just redistributing uncertainty across a larger parameter space.
Longs
- NVDA — GPU acceleration is the backbone of the entire inference pipeline
- GOOGL — JAX, the differentiable programming framework most likely underlying this work, is a Google project
- BOTZ (robotics/AI ETF) — indirect exposure to GPU-driven scientific computing infrastructure
Shorts
- Teams invested in NPTF pipelines — if flexible morphology inference becomes standard, fixed-template methods lose interpretive authority
- Frequentist profile-likelihood analysis software packages — variational Bayesian methods could supplant them for high-dimensional astrophysical searches
Enablers (Picks & Shovels)
- JAX (Google's differentiable array computing library) — enables GPU-accelerated gradient computation through physical simulations
- NumPyro / PyMC — probabilistic programming frameworks that support variational inference at scale
- Fermi-LAT public data archive — the gamma-ray telescope dataset the method is applied to
- CUDA ecosystem — underlies the GPU vectorization the paper depends on
Private Watchlist
- Kendall (private) — probabilistic programming tooling
- Alphabet/DeepMind (subsidiary) — differentiable physics simulation research
Resources
The Paper
We motivate the use of differentiable probabilistic programming techniques in order to account for the large model-space inherent to astrophysical $γ$-ray analyses. Targeting the longstanding Galactic Center $γ$-ray Excess (GCE) puzzle, we construct differentiable forward model and likelihood that make liberal use of GPU acceleration and vectorization in order to simultaneously account for a continuum of possible spatial morphologies consistent with the GCE emission in a fully probabilistic manner. Our setup allows for efficient inference over the large model space using variational methods. Beyond application to $γ$-ray data, a goal of this work is to showcase how differentiable probabilistic programming can be used as a tool to enable flexible analyses of astrophysical datasets.