From Stars to Waves: Non-deterministic Inference of Microlensed Gravitational Waves

Zheng Su, Xikai Shan, Zhenwei Lyu, J. W. Zhang, Yebin Liu, Shude Mao, Huan YangarXiv 2025 · DOI · arXiv

TL;DR

Researchers have developed a new machine-learning approach using Normalizing Flows to detect when gravitational waves are "microlensed" by thousands of individual stars in a distant galaxy. By treating this as a "non-deterministic" inference problem, they show that next-generation detectors could identify these subtle stellar signatures in about 8% of lensed events, providing a new way to "weigh" the stellar populations of distant galaxies.

What Problem Does This Paper Address?

When gravitational waves (GWs) from merging black holes pass through a massive foreground galaxy, they are "strongly lensed"—split into multiple images. However, as these waves traverse the galaxy, they also encounter a "minefield" of individual stars and remnants. Each star acts as a tiny lens (a microlens), adding complex ripples to the wave's amplitude and phase.

The challenge is complexity. A typical lensing galaxy contains billions of stars. Modeling the exact position and mass of every star near the wave's path is computationally impossible—there are simply too many variables. Furthermore, different arrangements of stars can produce nearly identical signals, making the problem "non-deterministic." Traditional "deterministic" templates used in GW astronomy fail here because we can't build a unique template for every possible stellar configuration in the universe.

Methods

The authors propose a shift toward Simulation-Based Inference (SBI). Instead of trying to map every individual star, they use an artificial intelligence framework called Normalizing Flows (specifically Neural Spline Flows).

Key technical steps include:

  1. Surrogate Parameters: They defined four "surrogate" variables ($l_p, l_{d10}, l_{d60}, l_{d120}$) that capture the statistical essence of the microlensing effect (like peak magnification and phase distortion) without needing to know the location of every star.
  2. Normalizing Flows: This neural network architecture learns to map complex, high-dimensional GW data into a simple probability distribution. This allows the researchers to estimate the "posterior distribution"—the probability of certain lensing parameters given the observed data—extremely quickly.
  3. Modified Mahalanobis Distance ($D_M$): To quantify if a signal is actually microlensed or just noisy, they introduced a modified distance metric to measure how far the recovered parameters "drift" from a null (unlensed) result.

Datasets

The study relied on high-fidelity simulations:

  • Training Set: 500,000 microlensed waveforms generated by statistically sampling stellar-field realizations. They used the LALSimulation library and the IMRPhenomXPHM model for the underlying black hole signals.
  • Test Set: 1,000 events drawn from a realistic astrophysical population model, simulating what "Third Generation" (3G) detectors like the Cosmic Explorer would see.
  • Stellar Populations: They used a Salpeter initial mass function (IMF) to distribute the masses of the microlenses (stars and remnants) between 0.1 and 28 solar masses.

Key Insights and Ideas

  • Embracing Non-Determinism: The "aha moment" is realizing that while we can't reconstruct the stellar field perfectly, we can measure its collective statistical signature. The surrogate parameters act as a bridge between the messy wave physics and the underlying galactic properties.
  • The Redshift Compensation: Interestingly, the paper finds that detectability isn't just about being close or loud. While low-redshift events have higher signal-to-noise ratios, high-redshift lenses behave like more massive objects, which actually enhances their microlensing signature. This "compensation" makes microlensing detectable across a wide range of the astrophysical population.
  • Physical Correlation: The AI-recovered surrogate parameters weren't just arbitrary numbers; they correlated strongly with physical properties like the stellar convergence ($\kappa_s$), which represents the density of stars in the lensing galaxy.

Key Results

  • Detection Statistics: In the 3G detector era, approximately 8% of microlensed events can be detected with high significance ($\ge 3\sigma$). If we lower the threshold to $2\sigma$, that number jumps to 14%.
  • Annual Yield: Based on predicted 3G detection rates, this translates to roughly 3 to 77 "golden events" per year where we can clearly see the influence of individual stars on gravitational waves.
  • Efficiency: Once trained, the Normalizing Flow can perform parameter estimation in seconds to minutes, compared to the hours or days required by traditional Monte Carlo methods.

Limitations and Future Work

The authors acknowledge several simplifications that provide a roadmap for future research:

  • Simplified Lens Models: The study used a Singular Isothermal Elliptical (SIE) model for the galaxy, which ignores more complex structures like external shear or substructure.
  • Image Types: The analysis focused on "Type-I" images (minima of the time-delay surface). "Type-II" images (saddle points) often show even stronger microlensing effects and will be a crucial next step.
  • Broader Applications: The "non-deterministic" framework isn't just for microlensing. The authors suggest it could be used to study the messy, unpredictable signals from core-collapse supernovae or the chaotic "post-merger" phase of binary neutron star collisions.

This work marks a significant step toward using gravitational waves not just to study black holes, but as a "microscope" to probe the dark and luminous matter within distant galaxies.

Summary updated

Abstract

Strongly lensed gravitational waves may pass through the stellar field of a lensing galaxy with additional modulations (on both phase and amplitude) due to gravitational microlensing effect of stars/remnants near the line of sight. These microlensed waveforms depend on the mass and location of thousands or more most relevant stars, so that their deterministic reconstruction from the data is computationally prohibitive. We classify the detection and parameter estimation of such events as non-deterministic inference problem and propose a solution with the implementation of normalizing flows. As a first step, we show that $8\%$ of microlensed events can be detected with significance $\ge 3 σ$ in the third generation era, with the chosen microlensing parameters correlated with the density of the underlying stellar field. This approach opens the door to probing microlensing effects and the properties of the underlying stellar fields. A similar construction may also be applied to other non-deterministic inference problems, such as detecting post-merger gravitational waves from binary neutron star coalescence and signals from core-collapse supernovae.
BibTeX
@misc{su2025from,
  author = {Zheng Su and Xikai Shan and Zhenwei Lyu and J. W. Zhang and Yebin Liu and Shude Mao and Huan Yang},
  title = {{From Stars to Waves: Non-deterministic Inference of Microlensed Gravitational Waves}},
  year = {2025},
  eprint = {2510.17125},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2510.17125},
  howpublished = {arXiv},
}
Figure