Forward vs Backward: Improving BAO Constraints with Field-Level Inference

This paper demonstrates that field-level inference using the LEFTfield EFT-based forward model improves constraints on the baryon acoustic oscillation scale by a factor of 1.2–1.4 compared to standard reconstruction methods by explicitly sampling initial conditions alongside bias and noise parameters.

The Gist

Imagine the universe as a giant, expanding ocean. Long ago, when the universe was a hot, dense soup of particles, sound waves rippled through this ocean, creating a specific, repeating pattern in how matter was distributed. Think of this pattern like a giant, cosmic ruler made of ripples, with each "tick mark" spaced about 500 million light-years apart.

Astronomers call this the Baryon Acoustic Oscillation (BAO). By measuring the distance between these "ticks" in the distribution of galaxies today, we can figure out how fast the universe is expanding and how dark energy is behaving.

However, there's a problem. Over billions of years, gravity has acted like a heavy hand, smudging and stretching this cosmic ruler. The clear "ticks" have become blurry, and the ruler has been slightly warped. This makes it hard to measure the distance precisely.

The Old Way: "Erasing the Smudge"

For years, the standard method to fix this has been called BAO Reconstruction.

Imagine you have a photo of a face that has been slightly blurred and shifted. The old method tries to guess how the face moved and then tries to "undo" that movement to sharpen the image.

• The Catch: This guess relies on a simplified, low-resolution map of how gravity works. It's like trying to un-blur a photo using a basic filter. It helps, but it leaves some smudges (specifically, the complex ways galaxies cluster) and assumes we know the exact rules of the game beforehand.

The New Way: "Rewinding the Movie"

This paper introduces a new, more powerful technique called Field-Level Inference (FLI).

Instead of just guessing how to un-blur the photo, FLI tries to rewind the movie of the universe back to the very beginning.

• The Analogy: Imagine you are watching a chaotic scene in a movie where people are running around. The old method tries to guess where everyone started based on where they are now. The new method (FLI) uses a super-computer to simulate the entire physics of the scene, working backward from the current chaos to reconstruct exactly what the scene looked like before anyone started running.
• The Magic: Because this method simulates the physics of gravity and galaxy formation in extreme detail (using something called "Effective Field Theory"), it doesn't just guess the starting position; it figures out the entire history of how the galaxies moved.

Why is this better?

The authors tested this new method against the old one using massive computer simulations of the universe (like a video game world filled with billions of virtual galaxies).

1. Sharper Ruler: The new method recovered the "cosmic ruler" with 10% to 40% more precision than the old method.
2. No Guesswork: The old method had to assume certain things about how galaxies form (like assuming all galaxies are simple dots). The new method accounts for the messy, complex reality of how galaxies actually cluster together.
3. The "Broadband" Secret: The paper found a surprisingly simple reason for the improvement. The old method left behind a lot of "noise" (a broad, fuzzy background) that confused the measurement. The new method successfully stripped away this noise, leaving only the clean, rhythmic pattern of the BAO ruler.

The Bottom Line

Think of the old method as trying to read a blurry book by squinting and guessing the words. The new method is like having a magic scanner that can see through the blur, understand the context of every sentence, and reconstruct the original, crystal-clear text.

By using this "rewind" technique, astronomers can now measure the expansion of the universe with much greater accuracy, helping us solve the biggest mysteries of the cosmos: What is dark energy? How heavy are neutrinos? And how did the universe evolve?

In short: They found a way to look at the universe's "baby photos" (the initial conditions) with much higher resolution, allowing them to measure the cosmic ruler with unprecedented sharpness.

Technical Summary

Here is a detailed technical summary of the paper "Forward vs Backward: Improving BAO Constraints with Field-Level Inference" by Ivana Babić, Fabian Schmidt, and Beatriz Tucci.

1. Problem Statement

The Baryon Acoustic Oscillation (BAO) feature is a critical standard ruler for measuring cosmological distances ($d_A(z)$) and the Hubble rate ($H(z)$). However, the precision of BAO measurements is limited by the nonlinear evolution of large-scale structure (LSS).

• Damping and Shift: Gravitational displacements of matter and galaxies smooth out (dampen) the BAO feature and induce small systematic shifts in the inferred scale.
• Limitations of Standard Reconstruction: The current state-of-the-art method, "standard BAO reconstruction," attempts to reverse these displacements using the Zel'dovich approximation (1st-order Lagrangian Perturbation Theory, 1LPT). This approach has three main drawbacks:
  1. It relies on assumptions of a fiducial cosmology and a fixed linear bias.
  2. It operates at linear order only, failing to account for nonlinear galaxy formation (nonlinear bias) and higher-order gravitational evolution.
  3. It does not remove the "broad-band" power spectrum distortions caused by nonlinear mode coupling, which increases the variance of the measurement.

2. Methodology: Field-Level Inference (FLI)

The authors propose Field-Level Inference (FLI) as a superior alternative. Instead of reconstructing displacements and then analyzing a power spectrum, FLI performs a joint Bayesian inference of the BAO scale ($r_s$), large-scale displacements, and bias parameters directly from the observed galaxy/halo density field.

• Forward Model (LEFTfield): The core of the method is the LEFTfield model, based on the Effective Field Theory of Large Scale Structure (EFTofLSS).
  • It uses a Lagrangian formulation that evolves initial conditions (linear density field $\delta^{(1)}$) to the observed field.
  • It includes bias operators up to third order in perturbation theory, capturing nonlinear bias and evolution.
  • It enforces Einstein's equivalence principle, ensuring matter and galaxies co-move on large scales.
• Inference Pipeline:
  • Initial Conditions: The linear density field is sampled from a Gaussian random field, modulated by a transfer function that explicitly varies the BAO scale parameter $r_s$.
  • Likelihood: A Gaussian likelihood is constructed comparing the observed halo density field to the forward-modeled prediction. The model marginalizes over stochastic noise and bias coefficients (except the linear bias $b_\delta$).
  • Sampling: High-dimensional posterior sampling is performed using Hamiltonian Monte Carlo (HMC) for the initial conditions and slice sampling for cosmological parameters ($r_s$) and noise/bias terms.
• Data: The study utilizes two N-body simulation halo catalogs:
  1. SNG Sample: $z=0.5$, mass range $10^{12.5}-10^{14.0} h^{-1}M_\odot$.
  2. Uchuu Sample: $z=1.03$, mass range $10^{12.0}-10^{13.5} h^{-1}M_\odot$ (higher resolution).

3. Key Contributions

• First Application to Nonlinear Tracers: This is the first demonstration of field-level BAO inference on fully nonlinear tracers (dark matter halos in N-body simulations), moving beyond previous tests on linear or mock data.
• Joint Inference Framework: The method consistently infers the BAO scale alongside the initial conditions and nonlinear bias parameters, avoiding the "fixed cosmology/bias" assumptions of standard reconstruction.
• Identification of Information Gain: The paper analytically derives the source of improved precision, identifying that the gain comes primarily from the reduction of variance in the power spectrum denominator (by recovering the linear power spectrum shape) rather than just improved displacement reconstruction.

4. Results

The authors compare FLI against a standard BAO reconstruction pipeline applied to the same data and scales ($k_{max}$).

• Precision Improvement: FLI improves the constraint on the BAO scale ($r_s$) by a factor of $\sim 1.2 - 1.4$ (a 10–40% reduction in uncertainty) compared to standard reconstruction.
  • SNG Sample: Improvement up to 40%.
  • Uchuu Sample: Improvement up to 40% (with higher signal-to-noise).
• Bias and Systematics:
  • Standard reconstruction yields unbiased results.
  • FLI showed slight bias at higher cutoffs when the EFT cutoff ($\Lambda$) was set equal to the analysis scale ($k_{max}$). However, increasing $\Lambda$ to $1.5 k_{max}$ removed this systematic shift, confirming the robustness of the EFT expansion.
• Power Spectrum Analysis:
  • Standard reconstruction leaves a significant "broad-band" enhancement in the power spectrum due to unresolved nonlinearities.
  • FLI successfully reconstructs the initial linear density field, resulting in a power spectrum that closely matches the linear theory ($P_L$) without the broad-band excess. This directly reduces the variance in the Fisher information calculation.

5. Significance and Conclusion

• Mechanism of Improvement: The paper clarifies that the primary source of the information gain is the incorporation of nonlinear bias and higher-order gravity. By jointly inferring these parameters, FLI effectively removes the nonlinear broad-band power that contaminates standard reconstruction, allowing the BAO signal to be measured with lower variance.
• Future Outlook: While currently tested on halos, the authors argue that because the model relies on the equivalence principle rather than specific tracer details, it generalizes to real galaxies.
• Implications: This approach represents a paradigm shift from "reconstruct then measure" to "infer jointly." It suggests that future surveys (like DESI, Euclid, and SKA) can achieve significantly tighter cosmological constraints by adopting field-level inference, provided computational costs and model validation (against real data systematics) are managed.

In summary, the paper demonstrates that Field-Level Inference using EFT-based forward models is a robust and superior method for BAO analysis, offering up to 40% better precision by fully exploiting the information content of the density field and mitigating the variance introduced by nonlinear structure formation.