🔭 astrophysics

A Novel Lensed Point Source Modeling Pipeline using GIGA-Lens with Application to SN Zwicky and SN iPTF16geu

This paper introduces a novel, GPU-accelerated modeling pipeline using GIGA-Lens that enables precise inference of lens mass parameters and the Hubble constant from strongly lensed point sources alone, successfully validating the method on SN iPTF16geu while revealing an alternative best-fit model for SN Zwicky that highlights the importance of comprehensive parameter space exploration.

Original authors: Saul Baltasar, Nicolas Ratier-Werbin, Xiaosheng Huang, W. Sheu, C. J. Storfer, Y. -M. Hsu, Sean Xu, David J. Schlegel

Published 2026-01-27

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Saul Baltasar, Nicolas Ratier-Werbin, Xiaosheng Huang, W. Sheu, C. J. Storfer, Y. -M. Hsu, Sean Xu, David J. Schlegel

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

Imagine the universe as a giant, cosmic funhouse mirror. Sometimes, massive objects like galaxies bend the light from things behind them, creating multiple copies of the same distant object. Astronomers call this "gravitational lensing." When that distant object is a supernova (a dying star that explodes), it's like seeing four or more fireworks exploding at the exact same moment, but from different angles.

This paper introduces a new, super-fast computer program designed to figure out exactly how those funhouse mirrors are shaped, using only the "fireworks" (the supernova images) as clues.

Here is a breakdown of what the authors did, using simple analogies:

1. The Problem: The "Ghost" in the Machine

Usually, to figure out how a lens (a galaxy) is shaped, astronomers look at the background galaxy itself. But sometimes, the background galaxy is too faint to see, or the lens is just a single point of light (like a supernova or a quasar).

Trying to model these point sources with old methods is like trying to solve a puzzle where the pieces are invisible until you guess the picture perfectly. If your guess is even slightly off, the computer gets confused and stops working. The authors say, "Let's try a different approach."

2. The Solution: The "GIGA-Lens" Super-Tool

The team built a new pipeline using a tool called GIGA-Lens. Think of this as a high-speed race car for math.

The Engine: It runs on four powerful graphics cards (GPUs), which are usually used for gaming but are perfect for crunching heavy numbers.
The Method: Instead of trying to match every tiny pixel of an image, their program focuses on three specific clues from the supernova:
1. Where the images are located (The positions).
2. How bright they are (The flux).
3. When they arrived (The time delay—since light takes different paths, the "fireworks" don't all go off at the exact same second).

By combining these three clues, the program can reverse-engineer the shape of the lensing galaxy with incredible speed and accuracy.

3. The Test Drive: Simulations

Before using real data, they tested their program on five different "fake" universes they created on a computer.

The Result: The program was like a master detective. It correctly identified the shape of the lens, the brightness of the source, and the time delays in almost every scenario.
The Hubble Constant: They also tested if they could measure the expansion rate of the universe (called $H_0$ ) using just one of these systems. They found that with their method, a single supernova system could give a very precise answer (within about 3.6% uncertainty).

4. The Real-World Application: Two Famous Supernovas

The team took their new tool and applied it to two real, famous supernovas that have been studied before: SN iPTF16geu and SN Zwicky.

Case A: SN iPTF16geu

The Situation: This is a well-known system where a supernova is split into four images. Previous studies used complex data, including the light from the host galaxy, to model it.
The New Approach: The authors ignored the host galaxy entirely. They looked only at the four supernova points.
The Result: Their results matched the previous studies almost perfectly. This proves you don't always need the "whole picture" (the host galaxy) to get the right answer; the "dots" (the supernova) are enough.

Case B: SN Zwicky

The Situation: This is a newer, trickier system. Previous models suggested the lens was a fairly round galaxy with the images spread out evenly.
The New Approach: Using only the supernova points, the program found a completely different shape.
The Result: The authors discovered that the lens is actually very stretched out (like a football) and the images are sitting right on the edge of a "critical curve" (a zone where light gets super-bright).
Why it matters: Previous models couldn't explain why two of the images were much brighter than the others. The new model explains it perfectly: the two bright images are sitting right on the "edge of the mirror" where the light gets amplified, while the dim ones are further away. It's like realizing the funhouse mirror isn't flat, but curved in a specific way that makes one side of your reflection huge and the other tiny.

5. The Bottom Line

This paper shows that we can now model these cosmic "mirrors" using only the bright points of light, without needing to see the faint background galaxies.

Speed: The program is incredibly fast, taking minutes instead of hours or days.
Accuracy: It finds the true shape of the lens and can even help measure how fast the universe is expanding.
Discovery: It can find solutions that older methods missed, like the specific shape of the galaxy lensing SN Zwicky.

In short, the authors built a faster, smarter way to read the cosmic map, proving that sometimes you don't need the whole picture to understand the puzzle—you just need the right clues.

Technical Summary: A Novel Lensed Point Source Modeling Pipeline using GIGA-Lens

Problem Statement
The paper addresses the "Hubble tension," a significant discrepancy (~5σ) between early and late universe measurements of the Hubble constant ( $H_0$ ). While strongly lensed time-varying events (quasars and supernovae) offer a pathway to resolve this tension via time-delay measurements, lensed supernovae (SNe) are rare. Existing modeling approaches for lensed point sources face specific challenges:

Inapplicability of Pixel-Based Methods: Traditional lens modeling (e.g., GIGA-Lens for extended sources) relies on minimizing pixel-by-pixel differences between observed and modeled images. This fails for point sources because the forward-modeled image positions often do not overlap with observed positions at the start of optimization, resulting in zero gradients and non-informative optimization.
Bias in Loss Functions: Previous attempts using only image positions often drive samplers toward unphysical parameter regions (e.g., mass slopes $\gamma \sim 1$ ) that minimize distance metrics but violate physical constraints.
Data Scarcity: Many lensed SNe lack detectable extended host-galaxy light, preventing the use of extended source modeling techniques.
Systematic Uncertainties: Mass-sheet degeneracy (MSD) and the need for full forward modeling (simultaneous inference of all parameters) remain critical hurdles in achieving precise $H_0$ constraints from single systems.

Methodology
The authors introduce a novel modeling pipeline built on the GIGA-Lens framework, utilizing the JAX platform to run on four A100 GPUs. The core innovation is a differentiable loss function designed specifically for point sources, utilizing only centroid positions, observed fluxes (or magnifications), and time delays, without requiring extended host-galaxy data.

The likelihood function $\log \mathcal{L}$ combines three terms with empirically determined relative weights ( $\omega_D, \omega_F, \omega_{TD}$ ):

Source Plane Compactness ( $\mathcal{L}_D$ ): Instead of pixel comparison, the pipeline "delenses" each observed image position back to the source plane using the lens model parameters. The loss minimizes the squared distance between these delensed positions and their mean, enforcing that all images originate from a single point source.
Flux/Magnification ( $\mathcal{L}_F$ ): To constrain the mass slope and break degeneracies, the pipeline models the inverse squared observed magnifications. A key technical contribution is the decomposition of the non-differentiable magnification function (which has singularities near critical curves) into a smooth, differentiable component. This allows the use of Automatic Differentiation (AD) for computing magnifications and their derivatives, avoiding finite-difference approximations and improving computational speed by 20–40%.
Time Delay ( $\mathcal{L}_{TD}$ ): The pipeline incorporates time delays by minimizing the difference between observed and predicted Fermat potentials (or time delays directly, treating $H_0$ as a free parameter).

Technical Implementation Details:

Differentiability & Bijectors: To handle parameter boundaries (e.g., restricting the mass slope $\gamma > 1.5$ ) without introducing numerical instabilities, the authors use bijectors to map unconstrained parameter spaces to constrained physical spaces.
Weighting Scheme: The authors demonstrate that relative weights between loss terms cannot be set solely by observational uncertainties. They employ an empirical weighting scheme where weights are increased (e.g., for flux) to ensure the posterior components contribute comparably during Hamiltonian Monte Carlo (HMC) sampling, preventing bias toward unphysical solutions.
Sampling: The pipeline uses a three-stage optimization: Multi-start gradient descent for the Maximum a Posteriori (MAP), Stochastic Variational Inference (SVI) for a Gaussian surrogate, and HMC for full posterior sampling.

Key Results

Simulated Systems: The pipeline was tested on five archetypal lens configurations (cross, long-cusp, short-cusp, fold, and double).
- It successfully recovered image positions, fluxes, and time delays.
- It constrained the mass slope $\gamma$ using point source data alone.
- For a simulated system with $H_0$ as a free parameter, the pipeline achieved a statistical uncertainty of $\sim 3.6\%$ ( $\sim 2.5$ km s $^{-1}$ Mpc $^{-1}$ ) on $H_0$ from a single system using full forward modeling.
- Convergence metrics ( $\hat{R} < 1.01$ and ESS $\gtrsim 10^3$ ) were achieved for most systems, with modeling times ranging from minutes to ~23 minutes on four A100 GPUs.
Application to SN iPTF16geu:
- The pipeline modeled the system using only the four lensed SN images (positions and magnifications), ignoring the extended host galaxy.
- The inferred mass slope ( $\gamma \approx 1.92$ ) and other parameters were consistent within $1\sigma$ with previous studies (e.g., Mörtsell et al. 2020) that utilized velocity dispersion and host galaxy data.
- The results confirmed that a smooth lens model alone cannot explain flux anomalies, supporting the microlensing hypothesis.
Application to SN Zwicky:
- The pipeline produced an alternative best-fit model compared to previous studies (Goobar et al. 2023; Pierel et al. 2023).
- Previous models suggested a low ellipticity and images symmetric relative to the critical curve, failing to explain the observed brightness asymmetry (images A and C being significantly brighter than B and D).
- This model inferred a higher ellipticity ( $q \approx 0.38$ ) and a mass slope $\gamma \approx 2.11$ . This geometry places images A and C directly on the critical curve (explaining their brightness) while B and D are further away.
- The model successfully predicted the observed positions, fluxes, and the revised time-delay ordering (Image A arriving first), which is consistent with recent refined measurements (Larison et al. 2024).

Significance and Claims
The paper claims that:

Point Source Data Sufficiency: It is possible to infer complex lens mass distributions, including the mass density slope $\gamma$ , using only lensed point source data (positions, fluxes, time delays) without extended host-galaxy information.
Full Forward Modeling: Simultaneous inference of all lens parameters and $H_0$ is feasible and yields precise constraints (3.6% uncertainty) from a single system, provided the system has a sufficiently large Einstein radius ( $\theta_E \gtrsim 1''$ ).
Robustness to Singularities: The decomposition-based approach to magnification allows for robust modeling of images near critical curves, a regime where previous methods often failed or required manual intervention.
Exploration of Parameter Space: The SN Zwicky case study underscores the importance of fully exploring the parameter space; the authors' model reveals a physically plausible configuration that resolves the tension between image symmetry and brightness asymmetry, which previous models missed.
Efficiency: The use of JAX and GPU acceleration enables rapid, statistically rigorous modeling (achieving strict convergence criteria $\hat{R} < 1.01$ ) in minutes, making it a practical tool for the increasing number of discovered lensed SNe.

The authors conclude that while the current focus is on SNe Ia, the framework is readily extensible to quasars and non-Ia SNe via flux ratios, and can be combined with host-galaxy modeling for even tighter cosmological constraints.