BayeSN-TD: Time Delay and $H_0$ Estimation for Lensed SN H0pe

The paper introduces BayeSN-TD, a probabilistic model for analyzing gravitationally lensed Type Ia supernovae that robustly infers time delays and magnifications while accounting for microlensing, and applies it to SN H0pe to derive a Hubble constant estimate of $69.3^{+12.6}_{-7.8}$ km s$^{-1}$ Mpc$^{-1}$, demonstrating its potential as a key tool for future time-delay cosmography.

The Gist

The Big Picture: Measuring the Universe's Speedometer

Imagine the universe is a giant car speeding away from us. Astronomers want to know exactly how fast it's going (a value called the Hubble Constant, or $H_0$). But here's the problem: when they look at the "early universe" (like a baby photo of the cosmos), they get one speed. When they look at the "local universe" (like a photo of the car right now), they get a different, faster speed. This disagreement is called the "Hubble Tension."

To solve this mystery, we need a new, independent way to measure the speed. Enter Gravitational Lensing.

The Magic Trick: The Cosmic Funhouse Mirror

Imagine a massive galaxy cluster sitting between Earth and a distant exploding star (a Supernova). This cluster acts like a giant, warped funhouse mirror. Because gravity bends light, the single exploding star doesn't just appear once; it appears multiple times, like a reflection in a hall of mirrors.

Here is the cool part: The light takes different paths to get to us. One path is short and straight; another is long and winding. Because light travels at a finite speed, the image that took the longer path arrives later than the one that took the shorter path.

If we can measure exactly how many days it takes for the second image to appear after the first, and we know the shape of the "mirror" (the galaxy cluster), we can calculate the distance to the star and, crucially, the speed of the universe's expansion.

The Problem: The "Static" in the Signal

There's a catch. The light doesn't just travel through empty space; it passes through a field of stars in the galaxy cluster. These stars act like tiny, shifting magnifying glasses. As the supernova explodes and its surface expands, it moves over these tiny lenses.

This causes Microlensing. Think of it like looking at a streetlight through a wavy, moving window. The light flickers, brightens, and dims unpredictably. If you try to measure the time delay between the mirror images while the light is flickering, your measurement will be messy and inaccurate.

The Solution: BayeSN-TD (The "Smart Filter")

This paper introduces a new tool called BayeSN-TD. Think of it as a super-smart noise-canceling headphone for astronomy.

1. The Base Model (BayeSN): Imagine a "perfect" supernova. We know exactly how a normal Type Ia supernova (a standard candle) should look and fade over time. BayeSN is a computer model that knows this "perfect song" by heart.
2. The New Feature (TD): When we look at the real, messy data from the lensed supernova, the "song" is distorted by the microlensing static. BayeSN-TD doesn't just ignore the static; it models the static.
   • It uses a mathematical technique called a Gaussian Process (think of it as a flexible rubber sheet) to guess how the "wavy window" is distorting the light at every moment.
   • It separates the "true song" (the supernova) from the "static" (the microlensing) simultaneously.
3. The Result: By figuring out exactly how the static is messing with the signal, the tool can tell us:
   • When the images actually appeared (Time Delay).
   • How much the mirror magnified the light (Magnification).

The Test Drive: SN H0pe

The authors tested their new tool on a real-life cosmic event called SN H0pe. This was a supernova discovered by the James Webb Space Telescope (JWST) that was split into three images by a galaxy cluster.

• The Challenge: The data for SN H0pe was tricky. Some of the images were observed very late in the supernova's life (when it was fading), and the data was a bit sparse (fewer photos than ideal).
• The Validation: Before using it on the real thing, they tested BayeSN-TD on simulations. They created fake supernovae on computers, added realistic "wavy window" static, and even made the static change colors (chromatic microlensing).
  • The Result: BayeSN-TD nailed it. Even though the simulations were based on different physics than their model, the tool recovered the correct time delays and uncertainties. It proved it could handle the "noise."

The Findings: A New Speed Estimate

Using BayeSN-TD on the real SN H0pe data, the team calculated:

• Time Delays: Image B appeared about 122 days after Image A, and Image C appeared about 63 days after Image B.
• Magnification: They figured out exactly how much brighter the images looked due to the lens.

When they combined these numbers with models of the galaxy cluster's gravity, they calculated a new speed for the universe:

• $H_0 \approx 69$ km/s/Mpc (with some uncertainty).

Why This Matters (and Why We Need More)

This result sits right in the middle of the "Hubble Tension" debate. It agrees with both the "baby photo" and the "current photo" measurements, but the uncertainty is still too big to declare a winner. It's like saying the car is going "somewhere between 60 and 80 mph." We need to know if it's 65 or 75 to solve the mystery.

The Future:
The authors are optimistic. They say that with better "template" images (cleaner photos of the background galaxy to subtract out the noise) and more lensed supernovae being discovered by future telescopes (like the Rubin Observatory), tools like BayeSN-TD will become the standard way to measure the universe's speed.

Summary Analogy

Imagine trying to time a race between two runners (the light images) who are running through a field of people holding mirrors (the stars). The mirrors make the runners look like they are speeding up or slowing down randomly.

BayeSN-TD is the referee who knows exactly how the runners should look. It watches the race, ignores the confusing mirror reflections, and calculates the true time difference between the runners. This allows us to measure the size of the track (the universe) with much greater precision.

Technical Summary

1. Problem Statement

Gravitational lensing of Type Ia supernovae (glSNe Ia) offers a promising, independent method to measure the Hubble constant ($H_0$) and address the "Hubble tension" between early-Universe (CMB) and local-Universe (distance ladder) measurements. However, analyzing glSNe Ia presents significant challenges:

• Microlensing: Stars in the lensing galaxy create time-varying magnification maps that distort the light curves of individual images. This effect can be chromatic (wavelength-dependent) and time-dependent, potentially biasing time-delay estimates if not properly modeled.
• Late-Time Photometry: Many discovered glSNe, such as SN H0pe, are observed significantly after peak brightness (up to +60 days rest-frame). Existing Spectral Energy Distribution (SED) models for SNe Ia often lack coverage at these late phases, particularly in the Near-Infrared (NIR).
• Model Mismatch: Previous analyses often relied on fitting color curves to mitigate achromatic microlensing or used models (like SALT) that may not perfectly match the physical properties of the specific supernova being analyzed.

2. Methodology

The authors introduce BayeSN-TD, an enhanced probabilistic framework designed specifically for fitting multiply-imaged glSNe Ia.

A. The BayeSN-TD Model

BayeSN-TD is built upon the hierarchical Bayesian SED model BayeSN. It treats the light curves of multiple images of the same supernova as a joint inference problem:

• Shared Parameters: Intrinsic SN properties (light curve shape parameter $\theta_1$, host galaxy dust extinction $A_V$ and $R_V$, and residual intrinsic color surface $\epsilon_s$) are shared across all images.
• Image-Specific Parameters: Each image $i$ has its own time of maximum ($T_{max}^i$), distance modulus ($\mu^i$, which encodes magnification), and a specific microlensing curve.
• Microlensing Treatment: The model incorporates a Gaussian Process (GP) with a Gibbs kernel to model microlensing.
  • The GP assumes an achromatic treatment (the same magnification curve applies to all bands for a given image), which is a valid approximation for the first ~3 weeks post-explosion for SNe Ia.
  • The Gibbs kernel allows the length scale of the microlensing variation to change over time, accommodating the rapid evolution of the SN photosphere crossing stellar caustics.
  • The microlensing curve is marginalized over, ensuring its impact is included in the statistical error budget rather than ignored.

B. Phase-Extended SED Model

To address the lack of late-time data coverage, the authors trained a new version of the BayeSN model:

• Extended Coverage: The model now covers phases up to +85 rest-frame days after peak (compared to previous limits of +40 or +50 days).
• Training Data: Trained on a combined dataset of 278 SNe Ia (Foundation DR1, Avelino et al., and CSP-I), including U-band data to better constrain the blue end of the spectrum (relevant for JWST F090W filters).
• Implementation: The model uses cubic spline surfaces for the zeroth-order template and functional principal components, with additional knots added at +55, +70, and +85 days.

C. Inference Framework

• Likelihood: The likelihood is evaluated in flux space, comparing observed fluxes to the model SED integrated through specific filters, corrected for redshift and dust.
• Time Delays: Time delays ($\Delta T_{ij}$) are derived from the posterior distributions of the time-of-maximum for each image, converted to the observer frame.
• Magnifications: Absolute magnifications ($\beta$) are inferred by comparing the measured distance modulus of the lensed images against a cosmology-independent prediction derived from high-redshift SNe Ia in the Pantheon+ sample.

3. Key Contributions

1. Development of BayeSN-TD: A robust, joint inference framework that simultaneously fits light curves, time delays, magnifications, and microlensing deviations for glSNe Ia.
2. Validation on Simulations: The model was rigorously tested on two distinct simulation sets:
   • Roman Simulations: Based on the SALT2 model with achromatic microlensing. BayeSN-TD successfully recovered time delays with negligible bias and well-calibrated uncertainties, despite the underlying model mismatch (SALT2 vs. BayeSN).
   • LSST Simulations: Based on SALT3 with chromatic microlensing. The model remained robust, showing that the achromatic GP assumption does not introduce significant bias in time-delay estimation even when chromatic effects are present.
3. Application to SN H0pe: The first application of this framework to real JWST data for SN H0pe, providing updated constraints on time delays and magnifications.
4. $H_0$ Estimation: Integration of BayeSN-TD constraints with existing lens models (Pascale et al. 2025) to derive new constraints on the Hubble constant.

4. Results

A. SN H0pe Constraints

Using photometry from Pierel et al. (2024a), BayeSN-TD inferred the following parameters:

• Time Delays:
  • $\Delta T_{BA} = 121.9^{+9.5}_{-7.5}$ days
  • $\Delta T_{BC} = 63.2^{+3.2}_{-3.3}$ days
  • Note: The $\Delta T_{BC}$ value is $\sim15$ days larger than previous estimates (Pierel et al. 2024a), attributed to methodological differences (fitting photometry vs. colors, marginalizing over microlensing vs. ignoring it, and the extended phase model).
• Magnifications (Flux space):
  • $\beta_A = 2.38^{+0.72}_{-0.54}$
  • $\beta_B = 5.27^{+1.25}_{-1.02}$
  • $\beta_C = 3.93^{+1.00}_{-0.75}$
• Microlensing: The data for SN H0pe was too sparse to tightly constrain the time-varying microlensing curves, though the model successfully marginalized over their potential impact.

B. Hubble Constant ($H_0$)

Combining the inferred time delays and magnifications with seven independent lens models from Pascale et al. (2025):

• Photometry Only (Weighted): $H_0 = 69.3^{+12.6}_{-7.8}$ km s$^{-1}$ Mpc$^{-1}$.
• Photometry + Spectroscopy: $H_0 = 66.8^{+13.4}_{-5.4}$ km s$^{-1}$ Mpc$^{-1}$.
• Time-Delay Only (No Magnification): $H_0 = 60.9^{+5.1}_{-4.6}$ km s$^{-1}$ Mpc$^{-1}$ (Noted as less reliable due to the mass-sheet degeneracy not being broken).

The results are consistent with both CMB and local distance ladder measurements but currently lack the precision to resolve the Hubble tension.

5. Significance and Future Outlook

• Robustness: The study demonstrates that probabilistic SED models combined with GP-based microlensing treatments can yield reliable time-delay estimates even when the underlying simulation models differ from the inference model or when chromatic microlensing is present.
• Scalability: BayeSN-TD is implemented in Numpyro/JAX, enabling GPU-accelerated Bayesian inference, which is crucial for analyzing the large sample of glSNe expected from the Vera C. Rubin Observatory (LSST) and Nancy Grace Roman Space Telescope.
• Future Prospects: While the current SN H0pe analysis is limited by sparse data, the authors highlight that upcoming template images will significantly improve photometric precision. BayeSN-TD is positioned as a critical tool for the next generation of time-delay cosmography, potentially providing the precision needed to definitively address the Hubble tension.

Code Availability: The BayeSN-TD code is publicly available at https://github.com/bayesn/bayesn-td.