Realization Variance of Gravitational Wave Background Anisotropies from Shot Noise for Pulsar Timing Arrays

Shot-noise anisotropies in the nHz gravitational wave background (GWB) are a promising target for pulsar timing arrays (PTAs). If the nHz GWB is sourced by merging supermassive black hole binaries (SMBHBs), as current evidence suggests, the shot-noise signal is expected to be large, potentially of order unity at observing frequencies of f ∼ 1 yr−1. In this regime, the signal is dominated by rare bright binaries, and Poisson fluctuations in the discrete SMBHB population produce significant spatial anisotropies. Here, we use Monte Carlo simulations to model the realization-to-realization scatter in the shot-noise, sampling from empirically calibrated models of the SMBHB source populations. We find that the probability distribution of shot-noise amplitudes is broad, spanning a factor of ∼ 50 (95% interval) at fixed frequency, with a long tail towards high amplitudes. The most probable and median amplitudes lie significantly below the ensemble means by factors of ∼ 2 − 3, implying that the shot-noise in typical realizations is smaller than the mean. The ensemble-averaged shot-noise also differs from simple estimates based on moments of the strain, ⟨h4 ⟩/⟨h2 ⟩2, because the average of a ratio is not equal to the ratio of the averages (i.e., ⟨X/Y ⟩ ̸= ⟨X⟩/⟨Y ⟩). This difference is a factor of ∼ 3 at f = 0.1 yr−1, growing to larger than two orders of magnitude by f ∼ 1 yr−1, where the GWB is dominated by low abundance, high-strain sources. Shot-noise nevertheless provides a powerful diagnostic for understanding the GWB and SMBHB populations; interpreting PTA measurements, however, requires modeling its full probability distribution.

Improvements to AI Systems:

1. Hierarchical Bayesian Inference with Ratio-Aware Likelihoods

• Improvement: Replace point-estimate or Gaussian-noise assumptions in pulsar timing array (PTA) analysis pipelines with a full hierarchical model that explicitly samples the joint posterior of shot-noise amplitudes and source population parameters, using the simulated probability distribution (broad, skewed, heavy-tailed) as a prior.

• Capability: The AI can correctly infer SMBHB population properties (e.g., merger rate, mass function) from PTA data without bias from the ensemble-mean shot-noise, which is shown to be systematically higher than typical realizations.

2. Simulation-Based Inference (SBI) for Non-Analytic Noise Distributions

• Improvement: Train a neural density estimator (e.g., normalizing flows) on the Monte Carlo outputs to learn the full conditional distribution of shot-noise amplitudes given frequency and population parameters, bypassing the need for closed-form analytic approximations (which fail due to ⟨X/Y⟩ ≠ ⟨X⟩/⟨Y⟩).

• Capability: The AI can rapidly generate realistic shot-noise realizations for arbitrary PTA configurations and frequencies, enabling fast likelihood-free model comparison and parameter estimation for future data releases.

3. Uncertainty-Aware Anisotropy Detection

• Improvement: Incorporate the factor-of-50 scatter and long tail into anomaly detection algorithms for GWB anisotropies. Instead of thresholding on mean amplitude, use quantile-based detection (e.g., 95th percentile) to flag significant spatial fluctuations.

• Capability: The AI can distinguish true astrophysical anisotropies from Poisson noise with controlled false-alarm rates, even when the most probable shot-noise is 2–3× below the mean.

4. Multi-Frequency Extrapolation with Tail Risk Modeling

• Improvement: Use the frequency-dependent scaling (factor of 3 at 0.1 yr−1 to >100 at 1 yr−1) to train a meta-model that predicts the full shot-noise distribution at unobserved frequencies, using extreme value theory to model the high-amplitude tail.

• Capability: The AI can forecast the probability of rare, bright SMBHB events that dominate the high-frequency GWB, aiding in target selection for follow-up electromagnetic observations.

5. Realization-Aware Signal Separation

• Improvement: Modify source-separation algorithms (e.g., independent component analysis or dictionary learning) to treat the shot-noise as a stochastic, realization-dependent component with a known prior, rather than a fixed background.

• Capability: The AI can decompose PTA data into individual bright binary signals and diffuse background more accurately, improving localization of nearby SMBHBs.

6. Active Learning for Simulation Efficiency

• Improvement: Use the broad distribution to design an active learning loop that selectively runs additional Monte Carlo simulations only in regions of parameter space where the shot-noise distribution is most uncertain (e.g., high-strain tail).

• Capability: The AI can reduce computational cost by 10–100× while maintaining accurate coverage of rare, high-amplitude events, enabling real-time updates as new PTA data arrives.