🔭 astrophysics

Summary statistic for pulsar timing arrays

This paper demonstrates that the information content of gravitational-wave signals in pulsar timing arrays can be effectively captured by a low-dimensional summary statistic, specifically requiring only the lowest-order spherical harmonic coefficients of the Earth term map and the monopole (or dipole for point sources) of the pulsar term variance to retain approximately 95% of the signal information.

Original authors: Gabriela Sato-Polito, Matias Zaldarriaga, Barak Zackay

Published 2026-08-07

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Gabriela Sato-Polito, Matias Zaldarriaga, Barak Zackay

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Cosmic Symphony and the Tiny Orchestra

Imagine the universe is a giant, echoing concert hall. For decades, scientists have been trying to hear the music of the cosmos, specifically the deep, rumbling vibrations caused by massive objects colliding, like black holes dancing together. These vibrations are called gravitational waves. To catch them, astronomers don't use microphones; they use pulsars. Pulsars are the universe's most reliable metronomes—dead stars that spin incredibly fast and beam radio waves toward Earth with the precision of a laser. By timing the arrival of these "ticks" from dozens of pulsars scattered across the sky, scientists can detect if a gravitational wave has stretched or squeezed space itself, causing the ticks to arrive slightly early or late.

This field of study is called a Pulsar Timing Array (PTA). The big mystery right now is what exactly is making the noise. Is it a chaotic, static hiss from billions of black holes colliding everywhere at once (an isotropic background)? Or is it a specific, loud solo from a single, massive pair of black holes (a point source)? The challenge is that the data is messy. It's like trying to figure out who is singing in a choir when you can only hear the singers standing in a few specific spots in a large room, and you don't know exactly how far away each singer is. The signal is buried in noise, and the math to separate the "Earth's echo" from the "singer's echo" is incredibly complex.

The Paper's Big Idea: Summarizing the Symphony

This paper, written by Gabriela Sato-Polito, Matias Zaldarriaga, and Barak Zackay, proposes a clever shortcut to solve this mess. They suggest that instead of trying to analyze every single piece of timing data from every pulsar—which is like trying to memorize every word of a thousand-page book to understand the plot—you can just read the "summary statistics."

The authors realized that the gravitational wave signal has a very specific shape. No matter where the source is or what kind of source it is, the way the signal correlates across the sky follows a pattern that drops off very quickly. They compare this to a musical chord: while a chord has many notes, the lower notes (the bass) carry almost all the weight and character of the sound. In the language of math, they use "spherical harmonics," which are like the notes of a spherical instrument. The paper shows that you only need to listen to the first few "notes" (specifically, the lowest levels of these harmonics) to understand 95% of what's happening.

The Earth Term vs. The Pulsar Term

To understand their trick, you have to know about two parts of the signal. When a gravitational wave passes, it affects two places:

The Earth Term: The wave hits Earth, shifting the timing of the pulsar ticks we hear right now. This part is coherent; it creates a specific pattern across the whole sky.
The Pulsar Term: The wave hit the pulsar itself thousands or millions of years ago, changing the timing of the signal before it even started its journey to us. Since we don't know the exact distance to these pulsars, this part looks like random, incoherent noise.

The authors developed a new way to write the math (the "likelihood") that treats these two parts separately. They found that the Earth term looks like a smooth map that can be described by just a few numbers (specifically, the "quadrupole" or $\ell=2$ mode). The pulsar term, which looks like noise, can be summarized by its average strength (the "monopole").

What They Found: Less is More

Using data from the NANOGrav 15-year dataset (a collection of 67 pulsars), the team ran simulations to see how much information they would lose if they threw away the complex details and only kept the simple summary.

Their results are surprisingly efficient:

For a general background noise: If you keep the Earth term map up to the second level of detail ( $\ell_{max} = 2$ ) and just the average strength of the pulsar noise (the monopole), you retain about 95% of the information needed to detect the signal. It's like realizing you only need the first three chapters of a mystery novel to guess the ending.
For a single, specific source: If there is one loud black hole pair causing the signal, the story is slightly more complex. The "noise" from the pulsar term actually contains a clue called a "dipole" (a one-sided bump in the noise). If you include this dipole and go up to the third level of detail ( $\ell_{max} = 3$ ) for the Earth map, you can capture almost all the information about the source, no matter where it is in the sky.

Why This Matters

The paper suggests that the universe is "low-dimensional," meaning the complex, high-resolution data we collect can be compressed into a tiny list of numbers without losing the story. Instead of wrestling with terabytes of raw timing data for every new analysis, scientists could first compress the data into these few spherical harmonic coefficients.

This doesn't mean the detailed data is useless; it means the detailed data is often redundant. Once you have this compressed summary, you can run different tests—looking for a background hum or a specific soloist—very quickly and cheaply. The authors emphasize that while this works best for the current generation of pulsar arrays, as we add more pulsars and map the sky more evenly, we might need even fewer numbers to get the same result. It's a way of turning a chaotic cosmic roar into a clear, concise message.

Technical Summary: Summary statistic for pulsar timing arrays

Problem Statement
Pulsar Timing Arrays (PTAs) have recently reported evidence for low-frequency gravitational waves (GWs). A primary goal following detection is to resolve the GW sky, distinguishing between an isotropic stochastic background and anisotropic signals or individual point sources (e.g., supermassive black hole binaries). Current data analysis tools often rely on template-based or anisotropy searches that operate on the full timing residual dataset. However, the timing residuals produced by GW signals consist of two components: a coherent "Earth term" (correlated across pulsars) and an incoherent "pulsar term" (uncorrelated, dependent on unknown pulsar distances). The authors posit that the full dataset contains redundant information and that the signal is inherently low-dimensional, as the angular correlation function (the Hellings and Downs curve) decays steeply with multipole order $\ell$ . The challenge is to identify a compressed set of parameters that retains nearly all information about the GW signal without requiring the full, high-dimensional dataset.

Methodology
The authors develop a new framework for PTA data analysis based on spherical harmonic expansions of the Earth term and the pulsar term variance.

Likelihood Derivation: They derive a modified PTA likelihood function where the unknown pulsar distances are marginalized over. This is achieved by approximating the pulsar term as a Gaussian random variable with a variance determined by the Earth term amplitude. The likelihood is expressed explicitly in terms of the Earth term map ( $z_e$ ) and the pulsar term variance map ( $\sigma_p^2$ ).
Spherical Harmonic Expansion: Both maps are expanded into spherical harmonic coefficients ( $a_{\ell m}$ for the Earth term and $b_{\ell m}$ for the pulsar term variance). The authors argue that these coefficients serve as a "nearly lossless summary statistic."
Theoretical Characterization: They calculate the expected power spectra for these coefficients. For the Earth term, the power spectrum follows the Hellings and Downs curve, $C_\ell \propto 1/[(\ell+2)(\ell+1)\ell(\ell-1)]$ , which decays rapidly, suggesting that low-order multipoles dominate. For the pulsar term, they analyze the coefficients for both a single deterministic source and an isotropic stochastic background, noting that a single source introduces a significant dipole component in the pulsar term variance.
Fisher Matrix Analysis: To quantify the information retention of this compression, the authors compute the Fisher matrix for the amplitude of a stochastic background and a deterministic point source. They compare the Fisher information of the full dataset against a truncated dataset containing only spherical harmonic coefficients up to a maximum multipole $\ell_{max}$ .
Realistic Simulation: The analysis is grounded in realistic noise properties and sky locations derived from the NANOGrav 15-year (NG15) dataset. They account for non-uniform sky coverage, varying noise levels across pulsars, and the marginalization over common uncorrelated red noise (CURN).

Key Contributions

Explicit Likelihood Formulation: The paper presents a likelihood expression where the Earth term map and pulsar term variance are explicit parameters, with the pulsar term marginalized as a Gaussian prior.
Summary Statistic Proposal: The authors propose that the spherical harmonic coefficients of the Earth term map and the pulsar term variance map constitute a sufficient statistic for PTA data.
Information Retention Quantification: They provide quantitative estimates of how many multipoles are required to capture the signal information. Specifically, they demonstrate that for a stochastic background, measuring the Earth term up to $\ell_{max}=2$ and the monopole of the pulsar term variance retains $\sim95\%$ of the information. For a point source, including the dipole of the pulsar term variance is crucial, and an $\ell_{max} \approx 3$ for the Earth term combined with the pulsar term dipole retains $\ge90\%$ of the information across the entire sky.
Validation of Gaussian Approximation: The authors investigate the validity of approximating the pulsar term as a Gaussian variable, even in the case of a single dominant source, concluding that the approximation is excellent for current PTA configurations due to the effective number of sources and noise contributions.

Results

Stochastic Background: Truncating the Earth term map at $\ell_{max}=2$ and including the monopole of the pulsar term variance recovers approximately 95% of the Fisher information regarding the amplitude of an isotropic stochastic background. The inclusion of the pulsar term is essential to break degeneracies between auto-correlation and cross-correlation patterns introduced by low-order truncation.
Deterministic Point Sources: For a single point source, the information retention is more sensitive to sky location due to the non-uniform distribution of pulsars. While $\ell_{max}=2$ retains a large fraction of information, it varies significantly across the sky. Including the dipole of the pulsar term variance and extending the Earth term to $\ell_{max} \approx 3$ ensures that at least 90% of the information is retained for sources at any sky location.
Noise and Sky Distribution: The study confirms that while realistic noise and sky distributions couple the spherical harmonic modes (preventing perfect diagonalization of the Fisher matrix), the steep decay of the signal power spectrum ensures that low-order modes still dominate the signal-to-noise ratio.

Significance and Claims
The paper claims that the GW sky as observed by PTAs is inherently low-dimensional. The primary significance of this work is the demonstration that the complex, high-dimensional PTA dataset can be compressed into a small number of spherical harmonic coefficients (roughly ten per frequency bin for the lowest bins) without significant loss of information.

The authors emphasize that this compressed summary is "signal-agnostic." Once the data is compressed into these coefficients, any downstream analysis—whether searching for an isotropic background, anisotropic signals, or individual continuous wave sources—can be performed efficiently and consistently on this reduced dataset. This approach avoids the computational cost of re-analyzing the full timing residuals for every new hypothesis. The work suggests that future PTA analyses could focus on extracting these specific coefficients first, streamlining the path toward characterizing the nature of the GW signal and resolving individual sources.