The Universe left an acoustic imprint

Just after the Big Bang, the Universe was a scalding soup of matter and energy. In this soup, giant sound waves propagated — like ripples on a pond, but at the scale of the entire Universe. When the Universe cooled enough, these waves froze, leaving an imprint in the distribution of galaxies. This imprint is called baryon acoustic oscillations, or BAO.

It's a kind of cosmic ruler: physicists know exactly at what distance this imprint should appear. By measuring it in the sky, they can calculate distances across the Universe — and thus understand its nature, its expansion, and the dark energy that accelerates it.

The problem: the signal has blurred

For 14 billion years, galaxies have moved, drawn toward each other. The original imprint has been "spread out" by this evolution. To measure it precisely, you have to rewind this motion — find where the galaxies started. This is called BAO reconstruction.

Classical methods do this approximately, assuming galaxies move in a simple, regular way. But that's a simplification: reality is far more complex.

The solution: learning to rewind

Researchers trained a neural network — an artificial intelligence program — on 100 complete Universe simulations. The network learned to look at a "today" galaxy map and figure out how they were distributed "at the start".

And the results were verified on 1,000 different simulations. The method works even when starting assumptions are slightly wrong — crucial for applying it to real DESI telescope observations.

BAO: a cosmic ruler

Baryon acoustic oscillations form a characteristic scale of about 150 comoving Mpc — the distance at which galaxies have a slight tendency to cluster, a fossil of primordial-plasma sound waves. By measuring this scale at different redshifts, we constrain H(z)·r_d and D_A(z)/r_d, and thus cosmological parameters including dark energy.

The BAO signal has attenuated since decoupling under non-linear gravitational evolution. BAO reconstruction aims to partially reverse this evolution to sharpen the acoustic peak — and thus improve the precision of cosmological measurements.

Three competing methods

The study compares three approaches on simulated (DESI-like) galaxy catalogs:

• Traditional reconstruction: Zel'dovich approximation — estimating Lagrangian displacements by inverting the smoothed density field. Simple, fast, but limited to large scales.
• Explicit field-level inference: a differentiable model (HEFT — Hybrid Effective Field Theory) simulates the galaxy field. Initial conditions are optimized by gradient descent (L-BFGS-B with annealing) to match observations.
• Implicit inference: a convolutional neural network (CNN) directly learns the mapping between observed galaxy field and initial linear field, over 100 training simulations.

Quantitative results

Uniform priors on nuisance parameters. For BGS with kmax=0.4 h/Mpc, the CNN reaches −46% and FoM ×3.2.

What it changes for cosmology

The method is designed to integrate directly into DESI's standard pipeline — no break with existing analyses. The precision gain applies equally to the power spectrum P(k) and to the correlation function ξ(s), confirming that the method saturates the available BAO information.

Protocol and mocks

1,000 FastPM realizations (L=1 Gpc/h, Np=1024³ LRG / 2048³ BGS, Planck cosmology: Ωm=0.3175, h=0.6711, σ8=0.834), HOD calibrated on Ding et al. 2025. Snapshots at z=0.7 (LRG) and z=0.2 (BGS). 100 simulations for CNN training, 900 for evaluation.

CNN architecture and training

9-layer doubly-convolutional network with ReLU activations (widths: 32/64/128 channels per 3-layer block). Input: 2 channels (δg and δtrad), output: δl on sub-grid (Nsub−18)³. Spectral smoothing via weighted Fourier loss: M(k)=10 for k∈[0.08; 0.5] h/Mpc, 1 otherwise. Inference via overlapping patches Nsub=50 (~195 Mpc/h), stride=10.

Explicit inference: annealing and HEFT

MAP optimization via L-BFGS-B with annealing on kiter (1 Mpc/h final). Second-order HEFT model via pmwd (Nc=256³, force grid 256³). P₁ prior de-wiggled (Vlah et al. 2016) for robustness to cosmological misspecification. Error σ²g(k)=A+Bk²+Cμ²+Dk²μ² parameterized on 4 degrees of freedom.

BAO fitting and scale parameters

Chen et al. 2024 model: P(k,μ)=B(k,μ)P_nw(k)+C(k,μ)P_w(k)+D(k), fitted via emcee (desilike) over kmin=0.02 to kmax=0.3–0.4 h/Mpc by dk=0.005. Scale parameters αiso=α⊥^(2/3)α∥^(1/3) and αap=α∥/α⊥. Leave-one-out covariance with Ledoit–Wolf regularization.

Robustness and coverage tests

• Ωm misspecification (0.29 vs 0.3175): unbiased constraints for all FLI methods. Traditional reconstruction with fixed priors shows slight overconfidence.
• AP distortions (z extension ×3.9%): αiso and αap recovered without bias for implicit and traditional methods (explicit not tested — cuboidal geometry unsupported in pmwd).
• Nominal coverage: maintained at all confidence levels across the 900 test realizations. Implicit inference produces a diagonal covariance matrix matching the Gaussian prediction at the percent level.