🔭 astrophysics

A constrained linear model for continuum normalization of stellar spectra

This paper introduces a fast, convex, and stable constrained linear model that simultaneously infers stellar parameters, telluric transmission, and continuum-instrument response by factorizing theoretical spectra into non-negative basis components, thereby eliminating the biases of independent continuum normalization and achieving state-of-the-art consistency in high-resolution stellar spectra analysis.

Original authors: Andrew R. Casey, Adam Wheeler, Megan Bedell, David W. Hogg, Andrew Sayjdari, Lily Zhao

Published 2026-01-29

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Andrew R. Casey, Adam Wheeler, Megan Bedell, David W. Hogg, Andrew Sayjdari, Lily Zhao

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

The Problem: Cleaning a Dirty Window

Imagine you are trying to read a sign through a window. The sign has letters (the stellar lines or absorption features), but the window is also covered in smudges, rain streaks, and dust (the continuum and telluric lines from Earth's atmosphere).

To read the letters clearly, you need to know exactly how dirty the window is so you can mentally "clean" it. In astronomy, scientists do this by taking the raw light from a star and dividing it by an estimate of what the "clean" background light should look like. This process is called continuum normalization.

The problem is that figuring out where the "clean" background is usually very subjective. It's like trying to guess how much dirt is on a window just by looking at the letters. If you guess wrong, your measurement of the letters is wrong. Traditional methods often require a human to manually pick "clean spots" on the window, which is slow, inconsistent, and depends on the person's experience.

The Solution: A Smart, Constrained Puzzle

The authors of this paper propose a new way to solve this puzzle. Instead of guessing the background or manually picking clean spots, they use a constrained linear model.

Think of the star's light as a complex song. The authors break this song down into three separate layers that they can adjust simultaneously:

The Star's Song: The actual absorption lines (the letters on the sign).
The Earth's Noise: The atmospheric interference (the rain on the window).
The Background: The smooth, overall brightness (the dirty window).

How It Works: The "Legos" Analogy

The core of their method is a mathematical technique called Non-Negative Matrix Factorization (NMF). Here is how they use it:

The Library of Legos: Before looking at any real stars, the scientists take a massive library of theoretical star models (computer-generated stars). They chop these models up into small "basis vectors." Imagine these as a set of Legos that represent different types of star features (like a specific hydrogen line or a molecular band).
The "No Negative" Rule: This is the most important part. The scientists impose a strict rule: You can only add Legos; you cannot remove them. In math terms, everything must be "non-negative."
- Why does this matter? In the real world, a star can only absorb light (make it dimmer), it can't spontaneously create extra light in the gaps between lines. By forcing the math to only "add" absorption, the model guarantees that the "cleaned" star light never goes above 100% brightness.
- This rule acts like a guardrail. It forces the model to say, "If the data is brighter than our star model predicts, that extra brightness must be the background (the dirty window), not the star itself."
Fitting the Puzzle: When they look at a real star, they don't need to guess the star's temperature or size first. They simply ask the computer: "How many of each Lego do I need to stack up to match this star's light?" Because the math is linear and the rules are strict, the computer can solve this puzzle incredibly fast and without getting stuck in bad guesses.

The Results: A Consistent Clean

The team tested this method on high-quality telescope data (from the HARPS instrument).

The Test: They took pictures of the same star (Alpha Centauri A) at different times and with different levels of "noise" (signal-to-noise ratios).
The Outcome: Even when the data was very noisy (like trying to read the sign through a heavy storm), their method cleaned the window consistently.
- At high quality, the results were consistent to within 0.2%.
- At lower quality, they were still consistent to within 0.5%.
The Comparison: This is better than previous "best-in-class" methods, which often struggle when the data gets noisy or when the star is very hot or very cool.

Why This Is a Big Deal

No Guessing Required: You don't need to know the star's properties beforehand. The model figures out the "cleaning" and the "star features" at the same time.
Speed: Because the math is linear (straightforward addition), it runs in seconds, even for massive datasets with millions of stars.
Reliability: It removes the human element. Two different scientists using this code will get the exact same result, whereas manual methods often vary from person to person.

Summary

The authors built a mathematical "cleaning machine" that uses a library of theoretical star parts and a strict "no-negative" rule to separate the star's true signal from the background noise and Earth's atmosphere. It works fast, works on all types of stars, and gives a very consistent "clean" view of the universe, even when the data is messy.

Problem Statement
Continuum normalization is a critical pre-processing step for inferring stellar parameters and chemical abundances from spectral line strengths. However, current practices are often subjective, inconsistent across signal-to-noise (S/N) ratios, or biased because they treat continuum normalization as an independent step separate from stellar parameter inference. Traditional methods, such as iterative sigma-clipping or manual selection of "continuum pixels," struggle with metal-rich stars, low S/N data, and spectral regions devoid of continuum points (e.g., blue wavelengths of cool stars). While joint modeling of continuum and stellar parameters is theoretically ideal, it is often computationally expensive or involves non-convex optimization. Existing data-driven models often require large training sets and can be unstable.

Methodology
The authors propose a constrained linear model that simultaneously fits stellar absorption, telluric transmission, and the joint continuum-instrument response. The core of the method relies on Non-Negative Matrix Factorization (NMF) applied to a grid of theoretical, continuum-rectified spectra.

Model Formulation: The observed flux is modeled as the product of three components: stellar lines ( $f$ ), telluric transmission ( $g$ ), and continuum-instrument response ( $h$ ). The model operates in log-transformed space to convert this multiplicative relationship into a linear summation: $Y = \log f + \log g + \log h$ .
Stellar and Telluric Absorption: Both stellar and telluric components are modeled by factorizing a grid of theoretical spectra into two non-negative matrices: a set of basis vectors and a set of weights.
- The stellar grid (BOSZ) is transformed using a negative logarithm to ensure sparsity and non-negativity, then factorized into $K$ basis vectors.
- The telluric grid (from the PypeIt project) is similarly factorized into $L$ basis vectors.
- Crucially, the non-negativity constraint ensures that the reconstructed rectified flux is strictly less than or equal to unity ( $f \le 1$ ). This physically separates the absorption features from the continuum, as any flux above the stellar model prediction must be attributed to the continuum component.
Continuum Modeling: The continuum-instrument response is modeled using a Fourier basis (sine and cosine functions), providing a flexible, linear representation of smooth variations.
Inference: The inference problem is reduced to a weighted least-squares optimization with non-negativity constraints on the stellar and telluric weights ( $\alpha \ge 0, \beta \ge 0$ ). Because the model is linear and the constraints are convex, the solution is stable, fast, and requires no initial guess for stellar parameters.

Key Contributions

Constrained Linear Framework: The paper introduces a method that unifies continuum normalization with nuisance parameter fitting (tellurics) within a single, convex optimization problem.
Physical Constraints via NMF: By enforcing non-negativity on the basis vectors and weights, the model guarantees that the stellar component represents only absorption. This prevents the continuum model from "absorbing" spectral features and ensures the rectified flux remains physically bounded ( $\le 1$ ).
Efficiency and Scalability: The method avoids expensive spectral synthesis codes or non-linear emulators at inference time. Training is performed once on a theoretical grid, while inference on a high-resolution spectrum (e.g., HARPS) takes only a few core-seconds.
Extensibility: The linear nature of the model allows for the potential inclusion of other linear nuisances, such as radial velocity shifts and rotational broadening, without breaking the convexity of the optimization.

Results
The authors validated the method using ESO/HARPS high-resolution echelle spectra:

Experiment 1 (Spectral Types): The model successfully fitted stars ranging from B-type to M-type, including a B4 star (HD 45418) that exceeded the temperature bounds of the training grid. The continuum estimates were consistent across evolutionary stages, including red giants.
Experiment 2 (S/N Consistency): Using repeat observations of $\alpha$ $α$ -Centauri A across a wide range of S/N ratios (10 to 200), the authors demonstrated high consistency in continuum normalization.
- For S/N > 100, the standard deviation of the 95th percentile of normalized flux was 0.22%.
- For S/N > 30, the scatter was 0.46%.
- These results indicate that the method produces consistent pseudo-continua even at low S/N ratios where traditional methods often fail or require manual tuning.

Significance and Claims
The paper claims that this constrained linear model offers a robust, reproducible, and physically motivated alternative to subjective continuum normalization. By distinguishing the continuum from absorption features through non-negativity constraints, the method provides a "true" continuum estimate rather than a pseudo-continuum, particularly in spectral regions lacking continuum pixels. The authors emphasize that while the model is designed for continuum normalization and not necessarily for precise stellar parameter inference (as it can mask deficiencies in the underlying physics), its ability to fit nuisances simultaneously and consistently across varying S/N ratios makes it a superior tool for large-scale spectroscopic surveys. The method achieves sub-percent consistency at S/N ratios an order of magnitude lower than previous state-of-the-art methods (e.g., RASSINE) applied to stacked spectra.