Estimation of the Hubble parameter from unedited compact object merger catalogues

This paper presents a novel framework for estimating the Hubble parameter using unedited compact object merger catalogues and detection-level information, enabling cosmological inference from marginal candidates without relying on per-candidate parameter estimation or additional selection cuts.

The Gist

The Big Picture: Measuring the Universe's Speedometer

Imagine the universe is a giant, expanding balloon. Scientists want to know exactly how fast it is inflating. This speed is called the Hubble Constant ($H_0$).

For decades, measuring this speed has been tricky. Usually, scientists look at "Bright Sirens"—cosmic events like colliding neutron stars that flash a light (electromagnetic signal) so we know exactly where they are. But these are rare. Most cosmic collisions are "Dark Sirens"—black holes smashing together that make a sound (gravitational waves) but no light. We hear the crash, but we don't know exactly where or how far away it happened.

The Problem: The "Noise" in the Room

Traditionally, to measure the universe's expansion using these Dark Sirens, scientists have been very picky. They only listen to the loudest, clearest crashes (high-significance candidates). They ignore the quiet, crackly ones because they assume those are just static or noise.

The authors of this paper say: "Wait a minute! By ignoring the quiet ones, we are throwing away a huge amount of information. The quiet ones are actually the ones coming from the very edge of the universe, which is exactly what we need to measure the expansion rate!"

The problem is that the quiet ones are hard to distinguish from random static (noise). If you try to analyze them, you risk confusing a real crash with a glitch in your microphone.

The Solution: The "Unedited Playlist"

The authors propose a new method. Instead of trying to figure out the details of every single crash (which takes forever and requires supercomputers), they suggest looking at the entire unedited list of candidates generated by the search software.

Think of it like this:

• Old Way: You have a playlist of 100 songs. You only listen to the top 10 hits because the rest sound like static. You try to guess the genre of music based only on those 10 hits.
• New Way: You listen to the entire playlist, including the static and the quiet tracks. You use a clever statistical trick to figure out, "Okay, 80% of this playlist is real music, and 20% is just radio static." You then use the pattern of the whole playlist to guess the genre.

How the Method Works (The Metaphors)

1. The "Volume Knob" vs. The "Song"

In the old method, scientists would take a candidate, run a massive simulation to figure out its mass, distance, and spin (the "Song"). This is computationally expensive.
In this new method, they only look at the Detection Statistic. Think of this as just the Volume Knob reading.

• "Is the volume high enough to be a song, or is it just static?"
• The authors realized they don't need to know the lyrics (the detailed physics) of every song. They just need to know the volume distribution of the whole playlist to figure out the genre (the Hubble Constant).

2. The "Ghost in the Machine" (Noise Modeling)

The biggest fear is that the "quiet tracks" are actually just the machine making noise.
The authors use a "Noise Model." Imagine you have a recording of a room with no music, just the hum of the air conditioner. You know exactly what the "static" sounds like.
When you listen to the full playlist, your computer compares every track to the "air conditioner hum."

• If a track sounds like the hum, it's noise.
• If it sounds slightly different, it might be a faint song.
• The math calculates the probability: "There is a 60% chance this is a song and a 40% chance it's the air conditioner."

3. The "Crowded Room" Analogy

Imagine you are in a crowded room trying to hear a specific conversation (the universe's expansion).

• Old Method: You only listen to the people shouting clearly. You ignore the people whispering.
• New Method: You listen to everyone. You know that some people are whispering real secrets, and others are just clearing their throats (noise).
• The authors developed a way to count the "whispers" and the "throat clearings" simultaneously. They found that even the "throat clearings" (marginal candidates) contain clues about how far away the real conversations are, helping them measure the size of the room more accurately.

The "Proof of Concept" (The Simulation)

Since they couldn't test this on real data immediately without risking errors, they built a Virtual Universe (a "Mock Catalogue").

• They created a fake universe with a known speed of expansion.
• They generated 10,000 fake "crashes" (some real, some noise).
• They ran their new method on this fake data.

The Result:
The method worked! It successfully recovered the "speed of the universe" they had programmed into the simulation.

• The Catch: They found that if the "fake noise" in their computer simulation wasn't perfectly smooth (a bit of "jitter"), it could slightly skew the results, especially when there were lots of real signals mixed in. It's like if your air conditioner hum had a weird, random rhythm; it might confuse the computer about what is a real song.

Why This Matters

1. Speed: This method is much faster. It skips the heavy lifting of analyzing every single candidate in detail.
2. Inclusivity: It allows scientists to use the "marginal" candidates—the faint, distant signals that were previously ignored. These are crucial for seeing deep into the history of the universe.
3. Future-Proofing: As our detectors get better (like the next generation of gravitational wave observatories), we will hear thousands of collisions. We won't be able to analyze them all individually. This "playlist" method is the only way to handle that much data.

The Bottom Line

The authors have built a new statistical toolkit that lets us listen to the entire cosmic symphony, not just the loudest instruments. By treating the "static" and the "faint music" as a single, mixed dataset, we can measure the expansion of the universe more efficiently and potentially more accurately, even if we aren't sure exactly what every single sound is.

It's a shift from asking "What is this specific sound?" to "What does the pattern of all the sounds tell us about the room?"

Technical Summary

1. Problem Statement

Current methods for estimating cosmological parameters (specifically the Hubble constant, $H_0$) using gravitational wave (GW) "dark sirens" (compact binary coalescences without electromagnetic counterparts) face two primary limitations:

• Reliance on High-Significance Candidates: Traditional hierarchical Bayesian inference requires full parameter estimation (PE) for individual candidates to generate posterior samples in the source parameter space. Due to the high computational cost of PE, analyses are typically restricted to candidates with high statistical significance (e.g., false-alarm rate > 4 years), discarding "marginal" candidates that could probe the distant universe.
• Selection Bias and Noise Modeling: Including low-significance candidates introduces a mixture of real signals and noise (false alarms). Rigorously modeling the noise population is difficult because the distribution of noise-originating candidates in the parameter space is not well-understood. Furthermore, conventional methods often rely on large-scale injection campaigns to estimate selection effects, adding complexity and computational overhead.

The authors propose a framework to bypass per-candidate PE, utilizing only detection-level information (ranking statistics) from an unedited candidate list produced directly by a search pipeline (GstLAL).

2. Methodology

The paper introduces a Bayesian hierarchical framework that operates directly on the detection statistic ($x$) assigned to each candidate, rather than raw strain data or PE posteriors.

A. Core Framework

• Dataset: The dataset $D$ consists of a catalogue of $N_{tot}$ detection statistics ($x_i$) from a search pipeline (e.g., GstLAL).
• Detection Criteria ($\Delta$): The framework explicitly accounts for the search threshold ($x \geq x_{th}$). This inherently incorporates selection effects without needing external injection campaigns.
• Likelihood Construction: The hierarchical likelihood is constructed by modeling the probability of observing a detection statistic $x$ as a mixture of signal ($H_s$) and noise ($H_n$) hypotheses:
  $$p(D|\lambda, \bar{\eta}, \Delta) \propto \prod_{i=1}^{N_{tot}} \left[ \bar{\eta} p(x_i | H_s, \lambda, \Delta) + (1-\bar{\eta}) p(x_i | H_n, \Delta) \right]$$
  Where:
  • $\lambda$: Population parameters (including cosmological parameters like $H_0$ and mass distribution parameters).
  • $\bar{\eta}$: The expected fraction of signal-originating candidates.
  • $p(x|H_n, \Delta)$: The PDF of the ranking statistic for noise, derived from the search pipeline's noise model.
  • $p(x|H_s, \lambda, \Delta)$: The PDF of the ranking statistic for signals, which depends on the population model $\lambda$.

B. Modeling the Signal Likelihood

To compute $p(x|H_s, \lambda, \Delta)$ without full PE, the authors re-parameterize the GstLAL signal model:

1. Marginalization: They marginalize over the combined Signal-to-Noise Ratio (SNR, $\rho_{comb}$) and template parameters.
2. SNR Distribution: They model the distribution of observed SNR ($\rho_{obs}$) based on the optimal SNR ($\rho_{opt}$), which is a function of the source's intrinsic mass, redshift ($z$), and cosmological model ($H_0$).
3. Importance Sampling: Since the analytical form of the ranking statistic PDF is unknown, they use importance sampling to map the assumed astrophysical population (mass distribution and redshift distribution) to the distribution of the detection statistic ($x$).

C. Signal Fraction Estimation

A critical component is estimating $\bar{\eta}$. The authors propose a point estimate method:

• They calculate the average of the astrophysical probability ($p(\text{astro})$) provided by the pipeline for all candidates.
• They derive a linear correction relation between the raw average and the true signal fraction to account for biases, allowing $\bar{\eta}$ to be fixed (via a Dirac delta prior) during the inference of $H_0$. This avoids the computational cost of jointly inferring $\bar{\eta}$ and $H_0$ while improving constraints.

3. Key Contributions

• Detection-Level Inference: A novel framework that performs population inference using only detection statistics, eliminating the need for computationally expensive per-candidate parameter estimation.
• Inclusion of Marginal Candidates: The method is designed to extract information from the full candidate list, including low-significance events, thereby increasing the effective volume for cosmological inference.
• Implicit Selection Correction: By using detection statistics that already reflect the search threshold, the framework naturally accounts for selection biases without requiring explicit injection campaigns.
• Linear Correction for Signal Fraction: The discovery and validation of a linear relationship between the raw $p(\text{astro})$ average and the true signal fraction, enabling more accurate point estimates of $\bar{\eta}$.

4. Results (Proof-of-Concept Mock Data Analysis)

The authors validated the framework using Mock Data Analyses (MDAs) based on LIGO S5 sensitivity data products.

• Setup: They generated 1,386 mock universes with varying injected values of $H_0$ (25–150 km/s/Mpc) and signal fractions ($\bar{\eta}$).
• Joint Inference: When jointly inferring $H_0$ and $\bar{\eta}$, the marginalized posteriors for $H_0$ showed low sensitivity to the injected $H_0$ values. This was attributed to the limited sensitivity of the S5-era detectors to distant sources (where $H_0$ dependence in the SNR distribution is strongest).
• Point Estimate Performance:
  • The corrected point estimate for $\bar{\eta}$ (using the derived linear relation) achieved high accuracy (error < 5% for $\bar{\eta} \gtrsim 0.03$).
  • Fixing $\bar{\eta}$ to this point estimate (Dirac delta prior) yielded significantly tighter constraints on $H_0$ compared to joint inference.
• Consistency Tests: Probability-Probability (p-p) plots showed that the recovered posteriors were statistically consistent with injected values (KS test p-value $\sim 0.1$), indicating the framework is largely unbiased.
• Limitations Identified:
  • Convergence Issues: The importance sampling used to construct the signal model PDF exhibited residual statistical fluctuations, particularly at high detection statistics (high SNR).
  • Systematic Bias: These fluctuations caused a systematic underestimation of the hierarchical likelihood at specific $H_0$ values (particularly large $H_0$), leading to a slight downward bias in the p-p plots for high signal fractions. This bias scales with the number of candidates and the signal fraction.

5. Significance and Future Outlook

• Scalability: This method offers a pathway to utilize the full output of future GW observing runs (e.g., LVK O4/O5), which will produce thousands of candidates, many of which are currently discarded.
• Efficiency: By removing the bottleneck of per-candidate PE, the computational cost is drastically reduced, making it feasible to analyze massive datasets for cosmological constraints.
• Future Improvements: The authors identify that improving the convergence of the importance sampling algorithm (specifically for the signal model) is critical for future precision. They also note that extending the framework to jointly infer mass distribution parameters alongside $H_0$ will require more sophisticated modeling but is theoretically possible within this detection-statistic framework.

In conclusion, the paper presents a robust, computationally efficient framework for cosmological inference from unedited GW catalogues. While current limitations in model convergence prevent perfect recovery of $H_0$ in the specific low-sensitivity mock scenario, the methodology is proven to be unbiased and highly promising for future high-sensitivity detectors where the $H_0$ dependence of the signal distribution is more pronounced.