Bairagi & Wandelt (2026)

The big idea¶

How much cosmological information lives in the non-linear density field, and how do you extract it without an analytic likelihood? Bairagi & Wandelt take the simulation-based inference (SBI) route: learn the posterior directly from forward-modelled simulations with a neural network, sidestepping the need for a tractable likelihood. Their PatchNet is a hierarchical architecture that fuses small-scale information from field sub-volumes (“patches”) with large-scale summaries (power spectrum and bispectrum), which is efficient in both compute and training data and avoids the memory cost of full-field training. On the Quijote dark-matter suite it enhances the Fisher information over analytic summaries and matches a very different method (wavelet statistics), suggesting it captures “most of the information content of the dark matter density field at the resolution of $\sim 7.8\,\mathrm{Mpc}/h$.”

Why it is the contrast for progenax¶

SBI is the default modern answer to a non-differentiable stochastic simulator — train a network on many runs. The Differentiable inference — natal cloud parameters from cluster substructure layer makes the opposite bet: keep the physics direct and differentiable by predicting the summary statistic analytically (Gaussianization two-point + counts-in-cells + the peaks-over-threshold tail) and differentiating that, rather than learning a black-box mapping. The two approaches are complementary:

• SBI (this paper) captures non-Gaussian, phase-coherent information a Gaussian-mapped model cannot, at the cost of interpretability and a large simulation budget.

• Physics-direct (progenax) keeps every parameter physically meaningful and the gradients exact, at the cost of restricting to the 1pt + 2pt statistics the model represents faithfully.

The shared theme is the information-vs-resolution question: Bairagi & Wandelt quantify the information content at a given grid resolution; the FDF forecast quantifies how many independent tail resolution elements a gas map needs to measure the density-PDF slope $\alpha$.

Use in progenax¶

• Cited in Differentiable inference — natal cloud parameters from cluster substructure as the SBI alternative the physics-direct layer deliberately avoids, and as a pointer for where a neural cross-check (or a future hybrid) would add the genuinely phase-coherent, higher-order information the 1pt+2pt model omits.

Notes¶

• Recent (2026) and cosmology-focused (Euclid/DESI/Rubin); the relevance to star clusters is by analogy — the FDF reframes cluster substructure as a galaxy-clustering-style inference problem.