🔭 astrophysics

Calibrating spectral siren cosmology with synthetic catalogs of binary black hole mergers

This paper introduces a Normalizing Flow-based population model trained on synthetic binary black hole catalogs to calibrate spectral siren cosmology, successfully eliminating systematic biases in Hubble constant inference and revealing a degeneracy between $H_0$ and the fraction of binaries formed via dynamical versus isolated channels when applied to GWTC-4.0 data.

Original authors: Arianna Scarpa, Simone Mastrogiovanni, Filippo Santoliquido, Manuel Arca-Sedda

Published 2026-03-17

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Arianna Scarpa, Simone Mastrogiovanni, Filippo Santoliquido, Manuel Arca-Sedda

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 Big Picture: Solving the "Hubble Tension"

Imagine the universe is a giant, expanding balloon. Astronomers want to know exactly how fast this balloon is inflating (a speed known as the Hubble Constant, or $H_0$ ).

The problem? We have two different speedometers, and they disagree.

The "Baby Photos" method: Looking at the Cosmic Microwave Background (the "baby photos" of the universe) suggests the balloon is inflating at a certain speed.
The "Adult Photos" method: Looking at nearby supernovas (stars exploding in our "adult" universe) suggests it's inflating faster.

This disagreement is called the "Hubble Tension," and it's one of the biggest mysteries in physics right now.

The New Tool: Gravitational Waves as "Standard Sirens"

Usually, to measure how fast the universe is expanding, we need to know two things about a cosmic event:

How bright it is (to tell us how far away it is).
How fast it's moving away (its redshift).

For stars, we can see them, so we know their redshift. But for Binary Black Holes (two black holes smashing into each other), we can't see them with telescopes. We only hear them through Gravitational Waves (ripples in space-time).

The Good News: Gravitational waves tell us exactly how far away the black holes are (like a loud siren tells you how far away a fire truck is).
The Bad News: We don't know their redshift (how fast they are moving away) because we can't see the galaxy they live in.

The Trick: The authors use a clever workaround. They treat the black holes like a musical choir.

If you know the distribution of voices in a choir (how many tenors, baritones, and basses there are), you can guess the size of the room they are in.
Similarly, if we know the "mass spectrum" (the distribution of weights) of black holes, we can mathematically figure out their redshift and, consequently, the expansion rate of the universe.

The Problem: The "Rigid Ruler" vs. The "Living Organism"

In the past, scientists tried to guess the black hole mass distribution using simple, rigid mathematical formulas (like a straight line or a simple bell curve). They assumed this distribution looked the same whether the black holes formed 1 billion years ago or 10 billion years ago.

The paper argues this is wrong.
Think of the black hole population like a living forest.

A forest isn't static. The types of trees, their sizes, and how they grow change depending on the season and the age of the forest.
Similarly, black holes formed in the early universe are different from those formed today. Their "mass spectrum" evolves over time.

Using a rigid, unchanging formula to describe a living, evolving forest leads to a systematic bias. It's like trying to measure the growth of a child using a ruler that doesn't stretch. The result? A wrong answer for the Hubble Constant.

The Solution: "Normalizing Flows" (The AI Painter)

To fix this, the authors used a powerful machine learning technique called Normalizing Flows.

The Old Way: Trying to paint a complex, shifting cloud using only a few straight lines and circles. It never looks quite right.
The New Way (Normalizing Flows): Using an AI painter that learns directly from the data. Instead of forcing the data into a pre-made shape, the AI learns the exact shape of the cloud, including all its wiggles, bumps, and how it changes over time.

They trained this AI on synthetic catalogs (computer simulations of how black holes actually form in the universe, based on astrophysics). This allowed the AI to learn the complex, evolving relationship between a black hole's mass and its age (redshift).

The Experiment: Did it Work?

The authors ran two tests:

The Fake Test: They created a fake universe where the black hole masses did change over time.
- When they used the old "rigid ruler" method, they got the wrong expansion speed.
- When they used their new "AI Painter" (Normalizing Flows), they perfectly recovered the correct expansion speed. The bias disappeared!
The Real Test: They applied this AI model to real data from the GWTC-4.0 catalog (the latest list of detected black hole mergers).

The Results

Using this new, flexible method on real data, they calculated the Hubble Constant to be 71.6 km/s/Mpc.

This number sits right in the middle of the disagreement between the "Baby Photos" and "Adult Photos" methods.
It suggests that the universe's expansion rate is likely closer to the "Adult" measurements, but the exact answer depends on how the black holes formed.

A Fun Twist: They found that the answer depends on the "mix" of black hole origins:

If more black holes formed in dense clusters (dynamical formation), the expansion rate leans one way.
If more formed from isolated stars (isolated evolution), it leans another way.

The Takeaway

This paper is a breakthrough because it stops forcing the universe into a simple box. By using AI to learn the complex, evolving nature of black holes, they removed a major source of error.

It's like finally realizing that to measure the speed of a river, you can't just measure the water at one spot with a stick; you have to understand how the river bends, speeds up, and changes as it flows downstream. With this new tool, we are one step closer to solving the mystery of how fast our universe is expanding.

1. Problem Statement

Gravitational Wave (GW) observations of Binary Black Hole (BBH) mergers offer a "standard siren" method to measure cosmic expansion and the Hubble constant ( $H_0$ ) without relying on the cosmic distance ladder. However, a critical challenge exists:

Redshift-Mass Degeneracy: GW detectors measure the detector-frame mass and luminosity distance, but not the redshift ( $z$ ). The redshift must be inferred statistically using the source-frame mass distribution.
Systematic Bias: Current analyses often rely on "strongly modeled" phenomenological prescriptions (e.g., Power Laws, Gaussian Peaks) that assume the BBH mass distribution is static or evolves simply with redshift.
The Consequence: If the true BBH population exhibits complex redshift evolution (e.g., mass features shifting with cosmic time) that is not captured by the assumed model, it introduces a systematic bias in the inferred value of $H_0$ . Recent studies suggest mismodeling the mass spectrum can significantly skew cosmological results.

2. Methodology

The authors propose a non-parametric population model based on Normalizing Flows (NF) to replace rigid analytical prescriptions.

Normalizing Flows (NF): A generative machine learning technique that learns complex probability density functions (PDFs) directly from data using invertible neural networks. Unlike parametric models, NFs do not assume a fixed functional form (like a Gaussian) and can capture intricate correlations between variables.
Training Data: The NFs are trained on synthetic BBH catalogs generated by the B-pop pipeline. This pipeline simulates BBH populations based on astrophysical prescriptions, including:
- Isolated Binary Evolution: Binaries formed from isolated stellar evolution.
- Dynamical Formation: Binaries formed in dense stellar environments (globular clusters, nuclear clusters).
Hierarchical Bayesian Inference (HBI): The framework integrates the NF model into the standard HBI likelihood. The detector-frame rate is factorized to include a joint probability density $p_{NF}(m_{1s}, m_{2s}, z)$ that captures the correlation between source masses and merger redshift.
Mixture Model: To account for different formation channels, the authors define a mixture fraction $\phi$ :
$p_{NF} = \phi p_{NF}^{dyn} + (1-\phi) p_{NF}^{iso}$
where $\phi$ represents the fraction of binaries formed dynamically.

3. Key Contributions

Development of a Non-Parametric Framework: The paper introduces a flexible, data-driven population model using Normalizing Flows trained on state-of-the-art astrophysical simulations (B-pop), eliminating the need for fixed analytical assumptions about the mass spectrum.
Bias Elimination Demonstration: Through rigorous benchmarking on mock data, the authors demonstrate that their NF framework successfully recovers the true $H_0$ and population parameters, whereas standard phenomenological models (Power Law + Peak) fail to capture redshift evolution, leading to significant biases.
Joint Inference of Cosmology and Astrophysics: The method allows for the simultaneous inference of cosmological parameters ( $H_0$ ) and astrophysical population parameters (the mixture fraction $\phi$ of formation channels), revealing the degeneracy between the two.
Application to Real Data: The model is applied to the GWTC-4.0 catalog (153 events), providing updated constraints on $H_0$ and the formation channel fractions.

4. Results

A. Validation on Mock Data

Scenario: A mock BBH population was generated with a "Power Law + Evolving Peak" distribution, where the Gaussian peak mean shifts linearly with redshift.
Phenomenological Model Failure: When analyzed with a static Power Law + Peak model, the inferred $H_0$ was biased, falling outside the 99.7% credible interval of the injected value ( $H_0 = 67.7$ ).
NF Success: When analyzed with the Normalizing Flow model, the systematic bias was fully corrected, and the injected $H_0$ was recovered within the credible interval.

B. Application to GWTC-4.0

Using the NF model trained on B-pop synthetic data applied to real GWTC-4.0 observations:

Hubble Constant: The inferred value is $H_0 = 71.62^{+4.04}_{-4.00}$ km s $^{-1}$ Mpc $^{-1}$ (68.3% credible interval). This value is consistent with both early-universe (Planck) and late-universe (SH0ES) measurements, though the central value leans slightly toward the local universe measurements.
Formation Channels: The fraction of binaries formed in dynamical environments ( $\phi$ $ϕ$ ) was inferred as $\phi = 0.30^{+0.06}_{-0.05}$ .
- This is consistent with the B-pop fiducial prediction of $\phi \approx 0.37$ .
- Degeneracy: The study highlights a degeneracy between $H_0$ and $\phi$ . A cosmology favoring a lower $H_0$ (Planck-like) prefers a higher fraction of dynamically formed binaries ( $\sim 35\%$ ), while a cosmology favoring a higher $H_0$ (SH0ES-like) prefers a lower fraction ( $\sim 25\%$ ). This is because dynamically formed binaries tend to be more massive; to match observed detector masses, a lower $H_0$ (implying lower redshifts) is required if the population is dominated by these massive systems.

5. Significance and Conclusion

Unbiased Cosmology: This work establishes that non-parametric, astrophysically calibrated models are essential for unbiased cosmological inference with GWs. It proves that "strongly modeled" analytical approaches are insufficient for the precision era of GW astronomy.
Astrophysical Insight: The method successfully bridges the gap between astrophysical simulations and observational data, allowing researchers to constrain the relative contributions of isolated vs. dynamical formation channels directly from GW data.
Future Outlook: The approach provides a robust framework for future analyses with larger catalogs (e.g., from the Einstein Telescope or Cosmic Explorer) and synthetic populations, ensuring that cosmological constraints are not limited by population modeling assumptions.

Note on Circularity: The authors acknowledge a potential circularity: the B-pop simulations assume a cosmology to generate merger rates. However, they argue that within the range of current $H_0$ uncertainties, the dependence is weak enough to be negligible, though future work with extreme $H_0$ values may require re-simulating catalogs for each $H_0$ hypothesis.