acceptodds
Under review as a conference paper at ICLR 2027

EB-gMCR: Energy-Based Generative Modeling for Signal Unmixing and Multivariate Curve Resolution

Abstract

A measurement of a mixture is the sum of a few component profiles, each weighted by a concentration, and multivariate curve resolution (MCR) recovers both from many samples. Classical MCR factorizes the data matrix, takes the component count as input, and leaves a rotational ambiguity. This paper instead models the generative process, generative MCR (gMCR): each sample selects a few components from a fixed set and is observed as their linear superposition plus noise. We prove that the data determine which decomposition a solver should reach. Under a spark condition on the component profiles, the decomposition of minimal usage is the true one, up to a relabeling of the components, and under noise the same holds up to a stated conjecture. Components recovered this way decode new samples of the same process. EB-gMCR, an energy-based solver, is one realization of gMCR: a gate selects components per sample and is trained toward minimal usage. With no count given, it recovers the component count to within a few percent on synthetic mixtures of up to 256 components, at a sample-to-component ratio of eight and a signal-to-noise ratio of 30 dB. On two public spectroscopy datasets it recovers the count exactly and reconstructs as well as the best method at that count. The trained model, frozen, decodes mixtures synthesized across the concentration simplex at the noise floor. gMCR applies to any sparse linear mixture of a fixed set of components. The results cover a solver whose candidate pool is at least as large as the true component count.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
multivariate curve resolution, spectral unmixing, identifiability, generative model, sparse selection, energy-based model

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_KbkoOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

In this paper, the authors address multivariate curve resolution (MCR), a classical problem in chemometrics that involves recovering component signals and their concentrations from measurements of mixtures. They also aim to reuse the recovered components to analyze new samples. Mathematically, each observed mixture signal is modeled as a weighted sum of component signals. The challenge is that both the component pool, including its size, and the corresponding weights may be unknown. The task is therefore to recover the shared component signals and their weights from multiple observed mixtures, and then infer the weights for new samples.

To achieve this, the authors introduce gMCR and theoretically investigate several questions: (1) Given the component pool, is the decomposition of a sample unique? (2) When the component pool is unknown, under what conditions can minimizing total component usage recover the true decomposition? (3) How does noise affect these recovery guarantees? The authors further propose EB-gMCR, a practical solver motivated by this formulation and theoretical analysis. Finally, experiments evaluate its ability to recover components, estimate their number, and reuse the learned pool to analyze new samples.

Critical strengths
  1. The method estimates the number of components automatically and reuses the learned components to analyze new mixtures.
  2. Experiments on synthetic and real data show that the method can recover meaningful component signals.
Critical weaknesses

1. Structural presentation problems. Although the manuscript provides background and explains some design choices, it lacks a clear narrative connecting the problem, the theoretical results, the solver, and the experiments. Definitions and theoretical statements often appear before their motivation and implications are sufficiently explained. This makes it difficult to understand why each result is needed and what the results collectively establish. Addressing these issues would require substantial reorganization and rewriting, rather than a few local clarifications.

2. Insufficient positioning and support for the technical contributions. The first two contribution statements are broad, while the differences from established sparse-representation and dictionary-identifiability results are insufficiently explained. The manuscript should distinguish established foundations from its new advances in assumptions, guarantees, or scope. Although the authors acknowledge limitations in applying the theory to the solver, the extent to which the guarantees support the implemented method remains unclear. The experiments demonstrate useful performance, but do not isolate the advantage of the energy-based gate over simpler alternatives. Appendix K.6 also shows that concentration thresholding substantially improves support recovery, leaving its practical selection and sensitivity important unresolved questions. Finally, Figure 2(c) evaluates reuse on synthesized mixtures rather than held-out real measurements, limiting the evidence for reuse under experimental variability.

Questions for authors
  1. Which parts of Theorems 1 and 2 are new relative to the cited sparse-representation and dictionary-identifiability results, particularly Donoho & Elad (2003) and Garfinkle & Hillar (2019)? Please clarify the specific differences in assumptions, guarantees, and scope.

  2. The manuscript acknowledges that the implemented usage-penalty schedule does not enforce the theoretical λ-window and that some recovery assumptions are unavailable in the experiments. Given these limitations, which guarantees can be established for the returned solutions, and which connections between the theory and the solver remain heuristic?

  3. What additional benefit does the energy-based gate provide over a simpler gate that also separates component selection from concentration estimation? Do the existing ablations isolate this benefit?

  4. Appendix K.6 reports substantially improved support recovery after concentration thresholding. How can this threshold be selected without knowing the true concentration scale or lower bound, and how sensitive is support recovery to this choice?

1 comment

I appreciate the authors' efforts in addressing my concerns. After carefully reading the revised version, however, I do not think my major concerns have been adequately addressed.

I encourage the authors to thoroughly reconsider how to present the work to improve its readability. Readers are not proof-checking machines that can easily digest mathematical statements and proofs one by one and automatically fill in the gaps in the reasoning that the authors have not explicitly explained. I do not believe this is merely a matter of personal preference, as reviewer fcX8 expressed similar concerns about the presentation. As I mentioned in W1, I do not think these issues can be addressed simply by adding explanatory text. Substantial reorganization and rewriting may still be necessary. I therefore maintain my original rating.

Reviewer_fcX8OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies multivariate curve resolution (MCR): decomposing a set of measured samples into component profiles and their concentrations. It proposes “generative MCR” (gMCR). In this model, each sample is a linear superposition of a few components drawn from a fixed pool, plus sub-Gaussian noise.

The theoretical part states:

  • uniqueness of the per-sample decomposition under a spark condition with margin (Prop. 1); identifiability of the minimal-usage decomposition (Thm. 1);
  • a “certified recovery” result for known and learned pools, the noisy learned-pool case being conditional on an open conjecture (Thm. 2);
  • a reuse corollary (Cor. 1);
  • an impossibility/consistency result as the number of samples grows (Thm. 3).

The solver, EB-gMCR, uses a neural network to output a pair of “selection energies” per component, relaxed with Gumbel-softmax. A second network outputs concentrations. Training minimizes reconstruction error plus a soft usage penalty and two regularizers, and checkpoints are selected by reconstruction-quality bands.

Experiments cover synthetic mixtures and two public spectroscopy datasets. The method is compared with an NMF variant, ICA and MCR-ALS.

Critical strengths
  • Estimating the number of components in MCR, rather than taking it as input, is a relevant and practically important problem.
  • The paper attempts to give a theoretical account of when the decomposition is identifiable.
Critical weaknesses

The paper is unclear, difficult to read and lacks references to prior work.

  • Positioning and missing related work. A large body of work on deep learning for spectral unmixing is not discussed. Energy-based models for source separation (e.g., ICA) are classical and also not cited, e.g., Teh, Welling, Osindero & Hinton, “Energy-based models for sparse overcomplete representations,” JMLR 4 (2003). Sparse regression methods for unmixing are likewise not discussed or compared, e.g., Iordache, Bioucas-Dias & Plaza, “Sparse unmixing of hyperspectral data,” IEEE TGRS 49(6) (2011).

  • Some theoretical results closely follow known results without attribution.

    • Proposition 1 is a variation of a classical compressive-sensing result and is presented without references. Donoho & Elad (2003) are cited only for Definition 1. The result is very similar to Donoho, Elad & Temlyakov, “Stable recovery of sparse overcomplete representations in the presence of noise,” IEEE TIT 52(1) (2006).
    • Theorem 1 characterizes minimizers of an ℓ0-type sparse fitting problem under a spark condition, which is a classical problem in compressive sensing. Section 4.1 cites Donoho & Elad (2003) for the underlying principle, but the paper does not say what Theorem 1 adds. That result covers the fixed-pool case, and Garfinkle & Hillar (2019) cover the learned-pool case; the latter is cited only in Section 2.
    • Theorem 2 is a blind ℓ0 sparse regression problem, in which the only randomness is the noise. It is not positioned with respect to that literature.
  • The generative framing is inconsistent with what is analyzed.

    • Section 3 indexes samples by a draw ω, in a measure-theoretic style, without defining the probability space. It also does not use the standard presentation of random variables.
    • Despite the section title, no generative model is specified: only a noise distribution and deterministic constraints are given, with no distribution over supports or concentrations.
    • Sections 4–5 analyze a deterministic sparse fitting problem with ℓ0-type selection (Eq. 3), and the certificate (Eq. 4) is entirely deterministic.
    • The generative language is therefore at odds with the content, especially since the introduction criticizes classical MCR for lacking a generative model.
  • Motivation.

    • The introduction states that any invertible T gives an equivalent factorization CT, T⁻¹S. This overlooks that constraints such as non-negativity restrict the admissible T.
    • The paper presents as a limitation that “a new sample from the same process requires a new factorization,” but does not explain why this is a limitation. The case for a generative model over direct solution of the factorization is not made.
  • Precision of the theoretical statements.

    • Section 4 poses the question of when the “pool” can be recovered, but the “learned pool” case is not formally specified.
    • Theorem 2 is difficult to parse from the main text. It does not say what ε_fit is, what exact problem is studied, or what the constraint set is; these quantities appear only in Appendix C.
    • Theorem 3(b) states a convergence of processes, but Ĝ_M, G* and d_G are not defined in the main text (only in Appendix D).
    • The statement of Proposition 2 is hard to follow. Training “steps” and t are used without a formal definition in the main text, and f(θ) denotes a loss and f_e(x) denotes the gate.
  • Solver design and the energy-based model.

    • The solver is described in an ad-hoc way.
    • The choice of an energy-based model instead of available sparse regression algorithms is not motivated.
    • The “energy” consists of two logits per component and a soft selection count. No energy function is identified, and the connection to generative energy-based models is not clear.
    • The notation in Section 5.2 mixes samples, reconstructions and mini-batches.
    • Section 5.3 (checkpointing) is hard to follow.
  • Assumptions and application. The spark condition (Definition 1) is strong, and the paper does not discuss how it can be ensured or checked in practice. More broadly, there is little discussion of how the assumptions and the model relate to the chemometrics application.

  • Experiments.

    • The baselines are identified only through the figure legends and a citation list at the end of Section 2. Their configurations and the reasons for choosing them are not described in the main text.
    • The baselines are mostly classical. Machine learning methods for sparse regression and MCR are neither compared nor positioned.
  • Relevance of Corollary 1. Corollary 1 concerns out-of-sample decoding once the pool is learned. MCR as defined in the paper decomposes a single data matrix, so the relevance of this result is unclear and not motivated.

  • Verifiability of the proofs. The proofs (Appendices A–E) run to 40 pages and seem written by AI. Combined with the clarity issues of the main text statements, this makes the proofs difficult to verify within a review.

  • Presentation.

    • The acronym NMF is not defined.
    • Several phrases are unclear or undefined, for example “trained toward minimal usage”, “its pool decodes new samples”, “physical reading”, and “pinning step”.
Questions for authors
  • Why is the generative model not specified fully, including the distributions over supports and concentrations, and the probability space underlying ω?
  • Why is the need to refit for a new sample a limitation of classical MCR, given that the stated task is decomposing a given data matrix?
  • How can the spark condition with margin be ensured or verified for real spectroscopic data?
  • What do Proposition 1 and Theorems 1–2 add beyond Donoho & Elad (2003), Donoho, Elad & Temlyakov (2006) and Garfinkle & Hillar (2019)?
  • What is the energy function of EB-gMCR? How does it relate to energy-based generative models, and why was this approach chosen over other sparse regression or deep sparse coding methods?
  • Please clarify which baselines were used and why. Have other deep sparse coding methods been considered?
  • Why is Corollary 1 relevant to MCR, which decomposes a single data matrix?
Reviewer_Zde6OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper reformulates multivariate curve resolution as a spike-and-slab sparse coding model. The proposed model is constrained so that each sample contains at most k active components from a fixed pool of unit-norm profiles. Under a spark condition, the paper proves that the model is identifiable up to a permutation. The authors also propose EB-gMCR, a neural solver with a softmax gate per sample and component to identify the components, and a checkpointing rule that keeps the sparsest model. On synthetic mixtures with up to 256 components the component count was recovered successfully, and on two spectroscopy datasets the count equals the known number of analytes.

Critical strengths
  1. Treating per-sample selection as an explicit latent variable, and making the component count an output, is a design well suited to chemometrics.
  2. The theoretical treatment is rigorous and targets the task well. The assumptions are stated clearly and limitations are discussed. The authors also provide a partial Lean 4 formalisation.
  3. The paper is clear about its limitations and about which parts of the theory the solver does not realise.
Critical weaknesses
  1. The work, while well suited to the task, is presented as more novel than it is. Sheikh, Shelton & Lücke (JMLR 2014) learn the dictionary, the activation probabilities, the noise level and the slab parameters by EM, and Drefs, Guiraud & Lücke (JMLR 2022) scale this to large spike-and-slab models. Those models use a signed Gaussian slab and do not constrain the coefficients, whereas gMCR requires non-negative concentrations bounded below by a floor. Under a spark margin, a cap on the number of active components per sample and this concentration floor, the paper shows identifiability with an unknown number of components, which is a narrower contribution than it appears. Prop. 1 is from Donoho & Elad (2003), and identifiability is established by Aharon et al. (2006), Spielman et al. (2012), Hillar & Sommer (2015) and Garfinkle & Hillar (2019). The paper cites some of these works but does not compare its assumptions and results with them.

  2. The solver does not provide solutions with the theoretical guarantees. In the numerical results, the estimated usage was higher than the ground truth.

Questions for authors
  1. What exactly do Thm 1–2 add beyond Hillar & Sommer (2015) and Garfinkle & Hillar (2019), apart from the unknown pool size?
  2. How does gMCR relate to spike-and-slab sparse coding with learned activation probabilities (Sheikh, Shelton & Lücke, 2014)?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.