SpectralUnmix: A Torch-Based Regularized Non-negative Matrix Factorization

Ana L. Chies-Santos; Niranjana P; Paula Coelho; Rafael S. de Souza; Rog\'erio Riffel

arxiv: 2603.11111 · v1 · submitted 2026-03-11 · 🌌 astro-ph.IM

SpectralUnmix: A Torch-Based Regularized Non-negative Matrix Factorization

Rafael S. de Souza , Paula Coelho , Niranjana P , Ana L. Chies-Santos , Rog\'erio Riffel This is my paper

Pith reviewed 2026-05-15 12:52 UTC · model grok-4.3

classification 🌌 astro-ph.IM

keywords non-negative matrix factorizationstellar spectraspectral unmixingregularizationR packagetorchproximal gradientastronomical data analysis

0 comments

The pith

SpectralUnmix is an R package that performs regularized non-negative matrix factorization on stellar spectra using proximal-gradient updates in torch.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents SpectralUnmix, a new R package for regularized non-negative matrix factorization implemented with torch and supporting GPU acceleration. It estimates low-rank non-negative representations of spectra using proximal-gradient updates while allowing smoothness regularization along the spectral axis. The method is demonstrated on a subset of stellar spectra by comparing the recovered components to principal component directions and representative stellar spectra. This provides a tool for decomposing complex astronomical spectra into simpler, non-negative basis elements that may be easier to interpret physically.

Core claim

SpectralUnmix estimates low-rank non-negative representations through proximal-gradient updates and allows smoothness regularization along the spectral axis to recover interpretable components from stellar spectra, as shown in comparisons with PCA and typical spectra.

What carries the argument

proximal-gradient updates with smoothness regularization along the spectral axis for non-negative matrix factorization

Load-bearing premise

Proximal-gradient updates combined with smoothness regularization along the spectral axis produce interpretable low-rank components for stellar spectra.

What would settle it

A direct comparison showing that the NMF components are no more interpretable or physically meaningful than those from standard PCA on the same stellar spectra dataset would falsify the utility of the regularization.

read the original abstract

We present SpectralUnmix, an R package for regularized non-negative matrix factorization (NMF), implemented in torch with optional GPU acceleration. The package estimates low-rank non-negative representations through proximal-gradient updates and allows smoothness regularization along the spectral axis. As a compact demonstration, we apply the method to a subset of stellar spectra and compare the recovered NMF components with principal-component directions and representative stellar spectra. The package is released under the MIT license at \href{https://rafaelsdesouza.github.io/SpectralUnmix/}{this repository}.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

SpectralUnmix is a straightforward R package release for regularized NMF on spectra via torch, useful for practitioners but not a new methodological advance.

read the letter

This paper announces SpectralUnmix, an R package that runs regularized non-negative matrix factorization on spectra using proximal-gradient updates in torch, with an optional smoothness penalty along the wavelength axis and GPU support. The authors include a short demonstration on stellar spectra that recovers components and lines them up against PCA directions and template spectra. The code is open under MIT at the linked repo. That is the core of what they deliver. The package fills a practical gap for R users who want non-negative decompositions with some built-in regularization and do not want to roll their own torch code. The demo shows the output looks reasonable and distinct from PCA, which is the minimum one expects from such a tool. The implementation choices follow standard practice for this class of problem, so nothing in the method itself is surprising. The main limitation is that the work stays at the level of a software announcement. The demonstration is compact and does not report reconstruction errors, sensitivity to the regularization strength, timing comparisons, or tests on larger or noisier datasets. Without those numbers it is difficult to judge how much the smoothness term helps in practice versus plain NMF. The paper does not claim new theory or performance guarantees, so the absence of quantitative validation is consistent with the scope but still leaves the reader wanting more evidence that the package improves on existing options. Readers who already work with stellar or galaxy spectra in R and need a ready-made non-negative factorization routine will find this immediately usable. People outside that niche or looking for methodological novelty will not. The package is public and the description is clear enough that a referee could check the code and ask for expanded validation if needed. I would send it to peer review rather than desk-reject it.

Referee Report

0 major / 2 minor

Summary. The manuscript introduces SpectralUnmix, an R package implementing regularized non-negative matrix factorization via proximal-gradient updates on the torch framework, with optional GPU acceleration and smoothness regularization along the spectral axis. A compact demonstration applies the method to a subset of stellar spectra and compares the recovered low-rank components against principal-component directions and representative stellar template spectra. The package is released under the MIT license.

Significance. If the implementation and demonstration hold, the work supplies the astroinformatics community with an open-source, GPU-capable tool for non-negative spectral decomposition that incorporates a standard smoothness prior. The explicit release of reproducible code under an MIT license is a clear strength that lowers the barrier for adoption in large spectroscopic surveys.

minor comments (2)

[Abstract] The demonstration paragraph should report the number of spectra, the chosen factorization rank, and the numerical value of the smoothness regularization strength so that the comparison with PCA and templates can be reproduced exactly.
A brief table or sentence quantifying reconstruction error (e.g., Frobenius norm or mean squared residual) for the NMF solution versus PCA would make the practical advantage of the non-negativity and smoothness constraints more concrete without altering the manuscript’s scope.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for their positive evaluation of the SpectralUnmix package and for recommending minor revision. The report does not enumerate specific major comments, so we address the overall assessment below and confirm that the manuscript has been updated to incorporate the suggested improvements in clarity and reproducibility.

Circularity Check

0 steps flagged

No significant circularity

full rationale

The manuscript is a software package announcement describing an R implementation of regularized NMF using proximal-gradient updates with optional smoothness regularization. No derivation chain, theoretical prediction, or first-principles result is presented that could reduce to its inputs by construction. The central content is the availability of the package and a compact demonstration on stellar spectra, which follows standard NMF techniques without any self-definitional, fitted-input, or self-citation load-bearing steps. The work is self-contained as an implementation release.

Axiom & Free-Parameter Ledger

1 free parameters · 0 axioms · 0 invented entities

The work relies on standard NMF assumptions (non-negativity, low-rank structure) and the choice of smoothness regularization; no new free parameters, axioms, or invented entities are introduced beyond those implicit in regularized NMF.

free parameters (1)

smoothness regularization strength
The regularization parameter along the spectral axis must be chosen; its specific value is not stated in the abstract.

pith-pipeline@v0.9.0 · 5399 in / 1012 out tokens · 34566 ms · 2026-05-15T12:52:52.736915+00:00 · methodology

SpectralUnmix: A Torch-Based Regularized Non-negative Matrix Factorization

Core claim

What carries the argument

Load-bearing premise

What would settle it

discussion (0)