Primordial Non-Gaussianity and the Field-Level Cramer-Rao Bound

This paper employs the field-level Cramer-Rao bound to determine the ultimate limits of extracting primordial non-Gaussianity from galaxy maps, demonstrating that while multi-tracer scale-dependent bias outperforms conservative higher-point analyses, it remains suboptimal compared to utilizing all available modes, and highlighting the critical role of theoretical modeling in forecasting constraints for future surveys.

The Gist

Imagine the universe as a giant, three-dimensional puzzle. When the universe was born in the "Big Bang," it wasn't perfectly smooth; it had tiny, random ripples. For decades, scientists have believed these ripples were perfectly random, like static on an old TV screen. This is called a Gaussian distribution.

However, some theories suggest the universe had a few "glitches" or specific patterns in those early ripples. These glitches are called Primordial Non-Gaussianity (PNG). Finding them is like finding a hidden message in the static that tells us exactly what kind of physics happened during the universe's first split-second of existence (inflation).

The problem? The universe has been evolving for 13.8 billion years. Gravity has taken those tiny, smooth ripples and crumpled them into a messy, complex web of galaxies, stars, and dark matter. It's like taking a pristine sheet of paper with a faint drawing on it, crumpling it into a ball, and then trying to guess what the original drawing looked like just by looking at the crumpled ball.

This paper is a guide on how to uncrumple that ball as best as possible.

The Core Idea: The "Perfect Map" vs. The "Real Map"

The authors ask a fundamental question: What is the absolute best we can possibly do?

They use a mathematical tool called the Cramér-Rao Bound. Think of this as the "Speed Limit" of the universe. It tells you the maximum amount of information you can extract from a dataset, regardless of how clever your computer algorithms are.

• The Ideal Scenario: Imagine you have a map of the entire universe, down to the smallest grain of sand, with no noise. This is the "Field-Level" analysis. The paper calculates the speed limit for this perfect map.
• The Real Scenario: We only have maps of galaxies (which are just a few dots on the map), and our telescopes have limitations (noise).

The paper compares these two to see how much information we are losing by using real data instead of the perfect theoretical map.

The Two Types of "Glitches"

The paper focuses on two specific shapes of these primordial glitches:

1. Local Non-Gaussianity (The "Big Picture" Glitch):

   • Analogy: Imagine a calm ocean (the universe). A "local" glitch is like a giant, slow-moving swell that makes the water deeper in one specific area.
   • Effect: This affects the biggest structures in the universe. It changes how many galaxies form in huge clusters.
   • The Paper's Finding: We can actually measure this very well! The authors show that by using a technique called "Multi-Tracer" analysis, we can get almost as much information as the perfect map.
   • The Multi-Tracer Trick: Imagine you are trying to count cars in a foggy city. If you only count red cars, the fog (noise) makes it hard. But if you count red cars and blue cars, and compare how their numbers change relative to each other, the fog cancels out. By comparing different types of galaxies (different "tracers"), we can cancel out the cosmic noise and see the "Local" glitch clearly.

2. Equilateral Non-Gaussianity (The "Small Picture" Glitch):

   • Analogy: This is like a sudden, sharp ripple that happens everywhere at once, affecting the small details of the water's surface.
   • Effect: This affects the smallest structures.
   • The Paper's Finding: This is much harder. The "crumpling" of the universe (non-linear evolution) completely scrambles these small signals.
   • The Problem: To see this glitch, we need to look at very small scales. But looking at small scales is like trying to read a book through a dirty, foggy window. The "dirt" (theoretical modeling errors and galaxy formation physics) is so thick that it hides the message. The paper warns that even with the best future telescopes, if we don't understand the "dirt" (how galaxies form), we might never find this specific glitch.

The "Bias" Problem

A major theme of the paper is Bias. In physics, "bias" doesn't mean prejudice; it means that galaxies don't form exactly where the dark matter is. They are "biased" toward the densest spots.

• The Metaphor: Imagine trying to guess the shape of a mountain by looking at where the trees grow. Trees only grow on the sunny, gentle slopes, not the steep, rocky cliffs. If you don't know exactly why trees grow where they do, you will get the wrong shape of the mountain.
• The Paper's Insight: The authors show that for the "Local" glitch, we can work around the tree-growth rules (bias) using the Multi-Tracer trick. But for the "Equilateral" glitch, the tree-growth rules are so complicated and messy that they completely hide the mountain's true shape.

What Does This Mean for the Future?

The paper is a mix of optimism and a reality check:

1. Good News: For the "Local" glitch, we are on the right track. Future surveys (like the ones planned for the next decade) can likely beat the current records set by the Cosmic Microwave Background (CMB) if they use the Multi-Tracer technique.
2. Bad News: For the "Equilateral" glitch, simply building bigger telescopes isn't enough. We need better theories and simulations. We need to understand the "dirt" on the window (how galaxies form) better than we do now. Without better physics models, we might never reach the "Speed Limit" of what we can learn.

Summary in One Sentence

This paper calculates the ultimate limit of what we can learn about the universe's birth from galaxy maps, showing that while we can easily find "big picture" glitches using clever comparison tricks, finding "small picture" glitches will require us to first master the messy physics of how galaxies actually form.

Technical Summary

1. Problem Statement

The primary goal of modern cosmology is to constrain the physics of the inflationary epoch, specifically through the detection of Primordial Non-Gaussianity (PNG). While the Cosmic Microwave Background (CMB) provides tight constraints, it is effectively a 2D linear system. Future Large Scale Structure (LSS) surveys (e.g., DESI, Euclid, MegaMapper) offer 3D volume but introduce significant challenges:

• Nonlinear Evolution: The late-time universe is nonlinear, mixing primordial signals with late-time non-Gaussianity generated by gravitational clustering.
• Degeneracies: Distinguishing primordial signals from late-time effects requires modeling galaxy bias, which introduces "nuisance parameters."
• Uncertainty in Forecasts: Forecasts for PNG constraints (especially for the equilateral shape) vary by orders of magnitude in the literature. This variance stems from differing assumptions about nuisance parameters (bias evolution), the maximum usable wavenumber ($k_{max}$), and the treatment of cosmic variance.
• Optimality: It is unclear whether standard summary statistics (power spectrum + bispectrum) saturate the information limit, or if "field-level" inference (using the full density map) offers superior constraints.

2. Methodology

The authors employ a Field-Level Cramér-Rao (CR) Bound analysis to determine the ultimate theoretical limit of information extraction from galaxy maps.

• Field-Level Likelihood: Instead of analyzing summary statistics (like the power spectrum $P(k)$ or bispectrum $B(k)$) directly, the authors construct a likelihood function for the observed galaxy density field $\delta_g^{obs}$ given the initial conditions $\delta_{lin}$.
  $$P[\delta_g^{obs} | b_i, \theta_i] = \int \mathcal{D}\delta_{lin} \exp\left[ -\frac{1}{2} \int_k \frac{|\delta_{lin}|^2}{P_{lin}} - \frac{1}{2} \int_k \frac{|\delta_g(\delta_{lin}) - \delta_g^{obs}|^2}{N^2} \right]$$
  Here, $\delta_g$ is modeled via a forward model including bias and nonlinear evolution.
• Saddle-Point Approximation: The functional integral is evaluated using a saddle-point (maximum likelihood) expansion. The initial conditions are reconstructed perturbatively from the observed field.
• Bias Modeling: The galaxy bias is expanded in terms of the matter density and primordial potential. Crucially, the authors include:
  • Scale-Dependent Bias: Arising from local PNG, modifying the linear bias as $b(k) \propto f_{NL}/k^2$.
  • Quadratic Bias: Including terms like $b_2 \delta^2$ and the primordial bispectrum contribution to the quadratic term.
  • Redshift-Space Distortions (RSD): Modeled via the Kaiser effect and Finger-of-God (FoG) damping.
• Fisher Matrix Calculation: The Fisher information matrix is derived for the cosmological parameters ($\theta$) and bias parameters ($b$) by marginalizing over the uncertainty in the initial conditions $\delta_{lin}$. This yields the CR bound: $\text{Var}(\theta) \geq (F^{-1})_{\theta\theta}$.
• Halo Model: To relate bias parameters to physical observables, the authors use the Press-Schechter formalism to link halo mass, number density, and bias coefficients ($b_1, b_\Phi$).

3. Key Contributions

A. Theoretical Clarification of Information Content

The paper establishes that for Local PNG ($f_{NL}^{loc}$), the field-level Fisher information is saturated by the combination of the power spectrum and the bispectrum.

• Single-Tracer vs. Multi-Tracer:
  • Single-Tracer: The signal-to-noise ratio (SNR) scales as $k_{min}^{-1}$ (dominated by large scales where cosmic variance is low). However, it is far from optimal because it cannot cancel cosmic variance.
  • Multi-Tracer: By cross-correlating different galaxy populations (tracers), cosmic variance is canceled. The SNR scales as $\log(k_{max}/k_{min})$, matching the scaling of the underlying matter bispectrum.
  • Result: A multi-tracer power spectrum analysis effectively saturates the field-level CR bound for local PNG, provided the analysis includes modes up to the scale of the halo size ($k_{max} \sim 2\pi/R_{halo}$).

B. The Equilateral PNG Challenge

For Equilateral PNG ($f_{NL}^{eq}$), the signal peaks at small scales (high $k$), where nonlinear evolution and bias degeneracies are severe.

• The field-level CR bound shows that $f_{NL}^{eq}$ is degenerate primarily with linear and quadratic bias parameters.
• Unlike local PNG, where scale-dependent bias moves information to the power spectrum, equilateral PNG relies heavily on the bispectrum and higher-order correlations.
• The analysis reveals that the vast variation in literature forecasts for $f_{NL}^{eq}$ is driven by how nuisance parameters (bias evolution across redshift bins) are marginalized.

C. Forecasting Framework

The authors develop a "greedy marginalization" framework to quantify how different nuisance parameters degrade constraints. They test three surveys: BOSS, DESI, and MegaMapper.

4. Key Results

Local Non-Gaussianity ($f_{NL}^{loc}$)

• Optimality of Multi-Tracer: Multi-tracer analyses of the power spectrum alone can achieve constraints very close to the theoretical limit set by the full matter bispectrum.
• Halo Mass Dependence: The information content depends on the halo mass. Lighter halos have higher number densities (lower shot noise) but smaller scale-dependent bias coefficients. Heavier halos have larger bias but are rarer. The optimal balance is found in the multi-tracer approach.
• Conclusion: For local PNG, current and future surveys are limited more by the ability to model bias than by the raw data volume. Multi-tracer methods are sufficient to reach the CMB limit and potentially surpass it.

Equilateral Non-Gaussianity ($f_{NL}^{eq}$)

• Sensitivity to Assumptions: Forecasts for $f_{NL}^{eq}$ vary by orders of magnitude depending on:
  1. The choice of $k_{max}$ (conservative vs. aggressive).
  2. Whether bias parameters are shared across redshift bins or treated as independent nuisance parameters in each bin.
• Degeneracy: Marginalizing over bias parameters independently in each redshift bin introduces weakly constrained directions that severely degrade $f_{NL}^{eq}$ constraints. Sharing bias parameters (assuming a known redshift evolution) significantly improves constraints.
• Survey Potential:
  • DESI: Can reach Planck-level sensitivity ($\sigma \sim 47$) only with strong priors on bias evolution and aggressive $k_{max}$.
  • MegaMapper: Has the statistical power to reach $\sigma(f_{NL}^{eq}) \approx 1$ (a transformative measurement), but this is highly contingent on theoretical inputs regarding bias evolution and modeling of the quasi-linear regime.
• Bispectrum Utility: While the bispectrum contributes little in conservative $k_{max}$ regimes, it is crucial for breaking degeneracies when nuisance parameters are marginalized.

5. Significance and Implications

1. Resolution of Forecast Discrepancies: The paper explains why previous forecasts for $f_{NL}^{eq}$ varied wildly. The difference is not in the data but in the marginalization strategy for nuisance parameters. Treating bias evolution as a known physical process (shared across bins) rather than a free parameter in every bin is essential for competitive constraints.
2. Theoretical Bottleneck: The paper argues that the limiting factor for future PNG measurements is not the size of the survey (volume) but the theoretical modeling of galaxy bias.
   • For $f_{NL}^{loc}$, the multi-tracer power spectrum is nearly optimal; the bottleneck is understanding the bias evolution.
   • For $f_{NL}^{eq}$, the bottleneck is the degeneracy with quadratic bias and the need to model nonlinear evolution accurately up to $k \sim 0.5-1.0 \, h \text{ Mpc}^{-1}$.
3. Path Forward: To achieve the ambitious goal of $\sigma(f_{NL}^{eq}) \approx 1$, the community must invest in:
   • Simulation-based priors: Using N-body simulations to constrain the redshift evolution of bias parameters.
   • Improved Modeling: Developing robust effective field theory (EFT) descriptions for the quasi-linear regime.
   • Field-Level Inference: While standard summary statistics can approach the CR bound, field-level methods provide the rigorous framework to define these limits and identify exactly which parameters (e.g., $b_2$, $b_{s2}$) are the dominant sources of uncertainty.

In summary, the paper demonstrates that while future surveys have the statistical power to revolutionize our understanding of inflation, realizing this potential requires a parallel revolution in the theoretical modeling of galaxy formation and bias evolution.