🔭 astrophysics

Mapping the Information Geometry of an Unresolved Dark Matter Population using a Differentiable Strong Lensing Simulator

This paper introduces a differentiable strong-lensing simulator and spectral basis framework to quantify how macro-model and source-model degeneracies limit the ability to detect unresolved dark matter subhalos, ultimately proposing the Fisher Graph Laplacian prior as a diagnostic tool to regulate and optimize sensitivity to these substructures.

Original authors: Alexandre Adam

Published 2026-08-20

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Alexandre Adam

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

Galaxies are not solitary islands of stars floating in the void; they are the visible tips of much larger, invisible mountains made of dark matter. According to our best theories of how the universe formed, these dark matter mountains should be covered in countless smaller bumps and peaks, known as subhalos. While we cannot see this dark matter directly, we can detect its gravity. When a massive galaxy sits between us and a distant light source, its gravity acts like a lens, bending the light and creating distorted, stretched images of the background object. If the dark matter around the lensing galaxy is lumpy with subhalos, it creates tiny, subtle ripples in these distorted images. Finding these ripples would allow astronomers to weigh the dark matter and test whether it behaves as the standard "cold" theory predicts, or if it is "warm" and smoother, which would change our understanding of how the universe grew.

The problem is that these ripples are incredibly faint and easily confused with other features. The background galaxy itself has a complex, irregular shape, and the main lensing galaxy is not a perfect sphere. When astronomers try to map the dark matter, the natural bumps in the background galaxy can look exactly like the bumps caused by dark matter subhalos. It is like trying to hear a whisper in a room where someone else is speaking at the same volume; the two sounds blend together, making it impossible to tell which voice belongs to whom. For years, this confusion has made it difficult to use these cosmic lenses to measure the small-scale structure of the universe with any precision.

A researcher has now developed a new way to untangle this confusion, not by taking better pictures, but by building a smarter mathematical model of how the light behaves. They created a highly flexible computer simulator that can generate images of these cosmic lenses, but with a crucial twist: the simulator is "differentiable." In simple terms, this means the computer can not only create an image but also instantly calculate how that image would change if any single part of the model were tweaked. This allows the researcher to trace exactly how much of the signal they see comes from the dark matter and how much is just an illusion created by the background galaxy's shape.

To do this, the researcher focused on the specific region where the light is most distorted, a ring-like area around the lensing galaxy. They broke down the potential dark matter bumps in this ring into a set of basic building blocks, similar to how a complex sound can be broken down into individual musical notes. They then fed these building blocks into their simulator along with models of the background galaxy. By running thousands of simulations, they measured how much of the dark matter signal the background galaxy model could accidentally "absorb" or mimic. They found that the background galaxy model is surprisingly good at hiding the dark matter signal. When the model of the background galaxy is allowed to be very flexible and complex, it can mimic the dark matter ripples across a wide range of sizes, effectively swallowing the signal and making it invisible to the observer.

The study reveals that the ability to detect dark matter subhalos depends heavily on how strictly the researcher constrains the shape of the background galaxy. If they let the background galaxy model be too free, it steals the information, and the dark matter signal disappears. However, the researcher introduced a new mathematical tool, which they call a "Fisher Graph Laplacian prior," to act as a regulator. This tool acts as a diagnostic to regulate the sensitivity of the data to an unresolved population of dark matter subhalos by testing how the inferred sensitivity changes conditional on the strength of this regularization. When they applied this rule, the background galaxy stopped absorbing the dark matter signal as aggressively. In their simulations, this resulted in the diagnostic metric stabilizing at a value of approximately 0.6, indicating that a significant portion of the information about the dark matter remained identifiable, whereas it would have been largely lost without this regulation.

The researcher also discovered that the main lensing galaxy, the massive object doing the bending, is less of a problem than the background source. The main galaxy's shape mostly confuses the large-scale, smooth parts of the dark matter distribution, but it does not hide the smaller, finer details. The real challenge comes from the background galaxy itself. The more detail the researcher tried to include in the background galaxy model, the more it confused the results, unless they used their new regulating tool to keep the model in check.

Ultimately, this work does not claim to have found the dark matter subhalos yet. Instead, it provides a rigorous map of the obstacles standing in the way. It shows that the biggest barrier to seeing these invisible structures is not the quality of the telescope or the brightness of the light, but the difficulty of separating the signal from the noise created by our own models. The study suggests that to move forward, astronomers must be careful not to let their models of the background universe become too flexible. By using these new mathematical tools to balance flexibility with realism, future observations could finally reveal the true texture of the dark matter that holds our universe together.

Technical Summary: Mapping the Information Geometry of an Unresolved Dark Matter Population

Problem Statement
Strong gravitational lensing offers a unique probe of the matter power spectrum on small scales, where the abundance of cold dark matter (CDM) subhalos can distinguish standard $\Lambda$ CDM cosmology from alternatives like warm dark matter (WDM). While individual subhalos have been detected, the unresolved population of low-mass subhalos is expected to produce a stochastic imprint on lensed images of extended background galaxies. However, extracting this signal is computationally challenging due to severe degeneracies. Perturbations in the lensing potential caused by substructure can be mimicked by adjustments to the macro-model (the foreground mass distribution) and, more critically, by highly flexible models of the background source. Current constraints remain limited because rigorous frameworks for quantifying these high-dimensional degeneracies are lacking.

Methodology
The author introduces a framework to quantify these degeneracies using a fully differentiable strong-lensing simulator built on the open-source package Caustics and implemented in PyTorch. The approach relies on a locally Gaussian approximation of the Bayesian posterior, where the information geometry is encoded by the Fisher matrix.

Key methodological components include:

Differentiable Simulator & Linearization: The simulator is linearized around fiducial parameters. The Jacobian matrix of the simulator defines the tangent space of the data manifold. Degeneracies are analyzed by projecting substructure anomalies onto the tangent spaces of nuisance parameters (macro-model and source model).
Spectral Basis for Substructure: Substructure is represented as perturbations to the lensing potential within an annular domain surrounding the lensed image (specifically the Einstein ring). These perturbations are expanded in a spectral basis of eigenfunctions (azimuthal Fourier modes and radial Bessel functions) with Neumann boundary conditions. This maps the collective signal of an unresolved population of NFW subhalos into a finite vector space of coefficients.
Information Geometry & Marginalization: The Fisher matrix ( $F = J^\top C^{-1} J$ ) is used to define the metric of the parameter space. The author employs a "matrix-free" approach using automatic differentiation and conjugate-gradient methods to compute matrix-vector products without explicitly storing large matrices. This allows for the calculation of the Schur complement to determine the marginal information remaining after integrating out nuisance parameters.
Transfer Function: A transfer function ( $T_a$ ) is defined as the ratio of marginalized to nominal information for a specific substructure mode. $T_a \approx 1$ indicates the mode is identifiable, while $T_a \approx 0$ indicates it is fully degenerate with nuisance parameters.
Fisher Graph Laplacian (FGL) Prior: To address the ill-posed nature of high-resolution source models (where the Fisher matrix becomes low-rank), the author introduces the FGL prior. This prior constructs a graph from the off-diagonal entries of the source Fisher matrix, using the internal degeneracies of the source model to regulate its expressivity. It penalizes differences between source pixels proportional to their mutual degeneracy.

Key Contributions

Differentiable Framework: The paper establishes a pipeline for quantifying degeneracies in high-dimensional lensing problems using automatic differentiation, avoiding the need to materialize large Fisher matrices.
Annular Spectral Basis: The introduction of a spectral basis confined to an annulus provides a finite vector space for encoding the stochastic signal of unresolved subhalos, matching the geometry of highly magnified arcs.
FGL Prior: The proposal of the Fisher Graph Laplacian prior as a diagnostic tool to study how source regularization impacts the sensitivity to dark matter substructure.
Quantification of Degeneracy Mechanisms: The work rigorously separates the effects of macro-model degeneracies from source-model degeneracies.

Results

Macro-Model Degeneracies: The macro-model (EPL profile, multipoles, external shear) primarily absorbs low-order, large-scale perturbations of the lensing potential. Most high-frequency substructure modes remain largely unaffected by the macro-model, with the transfer function remaining close to unity outside a narrow band of low azimuthal orders.
Source Model Degeneracies: In contrast, the source model absorbs substructure responses across a broad range of scales. As the expressivity (resolution) of the source model increases, the transfer function is strongly suppressed, approaching zero for high-resolution models ( $N_s = 2048$ ) in the weak-prior regime. This indicates that inferred sensitivity to substructure is heavily conditional on the source model's flexibility.
Impact of Regularization: The application of the FGL prior significantly dampens the dependence on source resolution. With the FGL prior at full strength, the average transfer function for high-resolution models stabilizes at $T \approx 0.6$ , demonstrating that regularization can recover a significant portion of the substructure signal that would otherwise be lost to source degeneracies.
Spectral Analysis of Population Models: Using the derived transfer functions, the author propagates the sensitivity limits to dark matter signals by generating coefficients from concordance CDM and alternative WDM population models. The observations are explicitly confined to the spectral content of these different population models, serving as a diagnostic of how nuisance absorption affects the distinguishability of these spectral signatures rather than presenting final cosmological constraints.

Significance
The paper argues that the primary bottleneck in constraining the small-scale matter power spectrum via strong lensing is not the macro-model, but the expressivity of the source model. The author claims that without rigorous regularization (such as the FGL prior), increasing source resolution to reduce systematic biases paradoxically exacerbates degeneracies, suppressing sensitivity to substructure. The framework provides a blueprint for tracking how information is absorbed during marginalization, offering a diagnostic tool to regulate the sensitivity of data to unresolved dark matter populations. The work emphasizes that precise constraints on the small-scale sector require a careful balance between source model flexibility and regularization strength.

Technical Summary: Mapping the Information Geometry of an Unresolved Dark Matter Population

More like this