🔭 astrophysics

$\texttt{HostSub_GP}$ : Precise Galaxy Background Subtraction in Transient Long-slit Spectroscopy with Gaussian Processes

This paper introduces $\texttt{HostSub_GP}$ , a novel Gaussian process-based toolkit that leverages multi-band archival imaging to precisely model and subtract host galaxy backgrounds from long-slit transient spectra, thereby outperforming traditional methods in recovering weak spectral features for events like SN 2019eix and AT 2019qiz.

Original authors: Chang Liu, Adam A. Miller

Published 2026-02-20

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Chang Liu, Adam A. Miller

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 you are trying to listen to a single violinist playing a solo in the middle of a massive, roaring orchestra. The violinist is a transient astronomical event (like a supernova or a tidal disruption event), and the orchestra is the host galaxy it lives in.

In astronomy, when we look at these events through a telescope, the light from the galaxy often drowns out the faint, specific notes of the event. For decades, astronomers tried to "subtract" the orchestra's noise by guessing what the background sounded like based on the few quiet spots next to the violinist. They would draw a straight line or a simple curve between those quiet spots to guess the noise in the middle.

The Problem: Galaxies aren't quiet, flat backgrounds. They have swirling arms, bright cores, and clumpy star clusters. A simple straight line is a terrible guess. It's like trying to guess the volume of a symphony by looking at a single empty seat in the back row. This leads to errors: sometimes you subtract too much (making the violinist sound like they are playing in a vacuum), and sometimes too little (leaving the orchestra drowning out the solo).

The Solution: HostSub GP
The authors of this paper, Chang Liu and Adam Miller, have built a new tool called HostSub GP. Think of it as a "smart noise-canceling headphone" for telescopes.

Here is how it works, using a few creative analogies:

1. The "Blueprint" Analogy (Using Old Photos)

Instead of guessing the background noise by looking at the empty seats next to the violinist, HostSub GP looks at high-resolution photos of the entire orchestra taken years ago.

The Old Photos: Astronomers have massive archives of images of galaxies taken in different colors (filters) by surveys like SDSS or Pan-STARRS.
The Blueprint: The software takes these old photos and creates a detailed "blueprint" of what the galaxy should look like at that specific location. It knows where the bright bulges are and where the dark dust lanes are.

2. The "Gaussian Process" (The Smart Predictor)

This is the mathy part, but think of it as a super-smart artist.

Classic Method: A classic method is like a child drawing a straight line between two dots. It's rigid and often wrong.
HostSub GP: This method uses a Gaussian Process (GP). Imagine an artist who knows the rules of perspective and lighting. They look at the "blueprint" (the old photos) and say, "Okay, the galaxy is bright here and dim there, and it fades smoothly."
The GP doesn't just draw a line; it draws a smooth, flexible curve that fits the known data perfectly but allows for small, realistic wiggles. It creates a "prior" (a best guess) of what the galaxy looks like behind the transient.

3. The "Noise Cancellation" (The Final Step)

Once the software has a perfect model of the galaxy's light (the orchestra), it compares it to the actual telescope data (the recording with the violinist).

It subtracts the model from the data.
Because the model is so accurate, the "orchestra" noise disappears, leaving behind a crystal-clear recording of the violinist (the transient).

Why This Matters: Two Real-Life Examples

The paper tests this on two real cosmic events to show how powerful it is:

Case 1: The Mystery Supernova (SN 2019eix)

The Situation: A supernova exploded near the bright center of a galaxy. For years, astronomers were confused: Was it a massive star dying (Type Ic) or a white dwarf exploding (Type Ia)? The galaxy's light was so strong it distorted the data.
The Classic Result: When they used the old "straight line" method, the data looked messy, with parts of the spectrum going negative (which is physically impossible).
The HostSub GP Result: The new tool cleaned up the noise perfectly. The "violinist" was revealed clearly. The data showed specific chemical signatures (Iron and Nickel) that proved it was a white dwarf explosion. This solved a mystery that had been stuck for years.

Case 2: The Nuclear Tidal Disruption (AT 2019qiz)

The Situation: A star was ripped apart by a black hole at the center of a galaxy. This is like trying to hear a whisper in the middle of a jet engine.
The Classic Result: The background galaxy light was so strong that faint, new features (like specific helium lines) were completely hidden.
The HostSub GP Result: By subtracting the galaxy light so precisely, the team could hear the "whisper." They detected faint, new emission lines that had never been seen before, giving them a much better understanding of how the black hole was eating the star.

The Bottom Line

HostSub GP is a software toolkit that uses machine learning (Gaussian Processes) and old photo archives to create a perfect map of a galaxy's background light. By subtracting this perfect map, astronomers can finally hear the "music" of cosmic explosions that were previously drowned out by the noise of their home galaxies.

It's like upgrading from a cheap, static-filled radio to a high-definition, noise-canceling system, allowing us to hear the universe's faintest secrets with crystal clarity.

1. Problem Statement

In observational time-domain astronomy, isolating the light of extragalactic transients (e.g., supernovae, tidal disruption events) from their host galaxies in long-slit spectroscopy is a critical challenge.

Limitations of Classic Methods: Traditional background subtraction relies on linear or B-spline interpolation of local galaxy flux measured adjacent to the transient's trace in the 2D spectrum. This assumes the background is smooth and slowly varying.
Failure Modes: This assumption breaks down in complex environments (e.g., galactic nuclei, bulges, spiral arms) or when the transient overlaps steep light profiles. Classic methods often lead to over-subtraction (creating artificial negative flux) or under-subtraction, introducing systematic biases that obscure weak spectral features, hinder classification, and compromise quantitative measurements (e.g., line velocities, progenitor mass constraints).
Data Scarcity: Unlike photometry, where pre-transient reference images are standard, obtaining pre-transient spectroscopic references is rarely feasible due to resource constraints and slit-loss variations.

2. Methodology: HostSub GP

The authors propose HostSub GP, a novel framework that leverages Gaussian Processes (GPs) and archival multi-band imaging to model the host galaxy background in 2D spectra.

Core Concept

The observed 2D spectrum $f_{obs}(x, \lambda)$ is modeled as the sum of the transient, the host galaxy, and the sky. The method focuses on estimating the host component $f_{host}$ by decomposing it into a 1D spectral component and a 2D spatial profile:
$f_{host}(x, \lambda) = F_{host}(\lambda) \cdot \xi_{host}(x, \lambda)$
Where:

$F_{host}(\lambda)$ is the 1D spectrum of the galaxy.
$\xi_{host}(x, \lambda)$ is the normalized spatial profile (convolved with the PSF).

Key Technical Steps

Data Preprocessing:
- Input: Calibrated 2D spectra and archival multi-band images (e.g., SDSS, Pan-STARRS, DESI, LSST).
- Adaptive Binning: The pipeline automatically identifies nebular emission lines (where the spatial profile changes rapidly) and uses finer wavelength bins there, while coarsening bins in the continuum to boost signal-to-noise ratio (SNR).
- Trace Alignment: Uses archival images to refine the transient trace position to sub-pixel accuracy, ensuring the slit geometry matches the imaging data.
Prior Construction from Imaging:
- Archival images are resampled and aligned to the slit geometry.
- A 2D Gaussian Process is trained on these images to create a smooth prior mean function ( $\xi_{sub, img}$ ) representing the galaxy's spatial profile across the slit.
- Seeing Matching: The pipeline empirically matches the spatial resolution (seeing) of the archival images to the spectroscopic data by convolving either the spectrum or the images with wavelength-dependent Gaussian kernels, minimizing the $\chi^2$ difference between them.
Gaussian Process Modeling:
- 1D GP: Models the spectral component $F_{sub}(\lambda)$ using a kernel combining Matérn-3/2 and Matérn-5/2 to capture both continuum trends and narrow line features.
- 2D GP: Models the spatial profile $\xi_{sub}(x, \lambda)$ $ξ_{s u b} (x, λ)$ .
  - Mean Function: The prior derived from archival images.
  - Covariance Kernel: A composite kernel (Matérn-3/2 + Exponential Squared) handles both small-scale fluctuations and long-term trends.
  - Emission Line Correction: A custom multiplicative kernel term ( $k_{corr}$ ) is introduced to handle discontinuities near emission lines, reducing correlation across wavelengths straddling a line while enhancing it within the line.
Inference:
- The hyperparameters of the GPs are optimized by maximizing the marginal log-likelihood (using JAX for automatic differentiation and GPU acceleration).
- The final host-subtracted spectrum is obtained by subtracting the modeled host flux from the observed spectrum.

3. Key Contributions

Novel Framework: First method to integrate archival multi-band imaging priors directly into 2D spectroscopic background modeling using GPs, avoiding expensive Spectral Energy Distribution (SED) fitting.
Software Release: HostSub GP, a scalable, open-source toolkit built on JAX and tinygp, enabling efficient computation (minutes on a laptop) and future GPU acceleration.
Robust Handling of Complex Profiles: Unlike interpolation, the GP approach naturally handles steep gradients and complex substructures (spiral arms, bulges) by learning the spatial profile from high-quality imaging data.
Adaptive Binning & Line Correction: Specific algorithms to handle the rapid spatial variation of flux near nebular emission lines, preventing the smoothing of real astrophysical features.

4. Results

The method was validated on synthetic data and real archival observations:

Synthetic Data (MUSE Cube):
- Tested on a spiral galaxy (LCRS B110709.2-121854) with 100 random slit placements.
- Performance: The GP method reduced the scatter of fractional flux residuals by a factor of ~5–8 compared to classic linear and B-spline interpolation.
- Bias: Classic methods showed systematic under-subtraction (positive bias) in curved profiles; GP effectively corrected this.
- Resolution: The method remained robust even when the galaxy was spatially unresolved (simulating lower resolution), though performance degrades if the galaxy size is comparable to the seeing.
Real Transient 1: SN 2019eix (Nebular Phase SN):
- Context: A peculiar SN near a galaxy bulge, previously ambiguous between Type Ia and Type Ic.
- Outcome: Classic methods over-subtracted the background, creating negative flux. HostSub GP revealed a clean spectrum dominated by [Fe II], [Ni II], and [Co III] lines with no [O I].
- Scientific Impact: Provided "smoking-gun" evidence that SN 2019eix is a thermonuclear explosion of a sub-Chandrasekhar-mass white dwarf, resolving previous classification ambiguities.
Real Transient 2: AT 2019qiz (Tidal Disruption Event):
- Context: A TDE in a galactic nucleus where the host light dominates.
- Outcome: The GP method recovered smooth continua and weak emission features invisible to classic methods.
- Scientific Impact: Uniquely detected persistent broad He II $\lambda$ 4686, emerging narrow He II and H $\beta$ , and [Fe VII] $\lambda$ 6087. These detections allow for tighter constraints on the TDE's ionizing SED and the gas properties surrounding the supermassive black hole.

5. Significance and Applications

Improved Science Return: Enables the study of transients in high-background environments (nuclei, bulges) that were previously unusable for detailed spectroscopic analysis.
Classification Accuracy: Reduces systematic errors in spectral features, leading to more reliable transient classification (e.g., distinguishing SNe Ia from Ic).
Scalability: The JAX-based implementation allows for rapid processing, making it suitable for integration into real-time pipelines for upcoming large-scale surveys (e.g., LSST, DESI).
General Applicability: While tested on SNe and TDEs, the method is applicable to any long-slit observation where the host background is spatially resolved and stationary.

Conclusion: HostSub GP represents a paradigm shift in spectroscopic data reduction, moving from local interpolation to global, image-informed probabilistic modeling, significantly enhancing the precision of transient astrophysics.