acceptodds
Under review as a conference paper at ICLR 2027

When Does Reconstruction Identify BM Tensor Representations?

Abstract

The Bhattacharya–Mesner (BM) decomposition represents a third-order tensor through factors indexed by complementary pairs of data indices. We study identifiability of these factors from a fully observed tensor. On the entrywise-nonzero stratum, we characterize all term-preserving transformations and construct a canonical normalization. A full-column-rank Jacobian in the normalized coordinates certifies local identifiability, determines reconstruction-loss curvature, and yields local perturbation bounds. For cubic tensors, parameter counting gives a necessary term-count ceiling of for generic local identifiability. Below this ceiling, separable template ratios generate continuous families of inequivalent decompositions with identical reconstructions. Under explicit full-column-rank assumptions, the slice-wise mixing families have dimension in one mode, with corresponding formulas in the other modes; these ambiguities also occur for positive factors. At rank one, we establish global identifiability, closed-form recovery, and a membership test expressed entirely in tensor entries. Exact finite-field witnesses and synthetic experiments examine gauge freedom, structural ambiguity, numerical conditioning, and optimization behavior.

Primary area:
learning theory
Keywords:
Bhattacharya–Mesner decomposition, Tensor identifiability, Essential uniqueness, Gauge symmetry, Jacobian criterion, Tensor decomposition

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_mJ1QOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

This paper studies identifiability of the so-called Bhattacharya–Mesner (BM) decomposition of a tensor. Specifically, local identifiability is characterized by studying the kernel of the Jacobian of the map that sends a parameter vector into its corresponding tensor. This kernel contains directions that are due to inherent scaling ambiguities, which are not harmful for essential uniqueness, but may contain others if the decomposition is not locally unique. The authors then capitalize on this analysis to derive error bounds on decomposition factors due to a perturbation. They also give a template of cases in which local essential uniqueness is violated, and a proof of global essential uniqueness in the rank-one case.

Despite the relevance of the presented results, the manuscript suffers from various significant drawbacks, as discussed below. First, in my view it fails to describe exactly how it fits in the literature, how it connects to prior results, and which gaps it fulfills. Second, there are substantial presentation issues, related to the clarity and proper formalization of its contents.

Critical strengths
  1. The paper's results represent one non-negligible (albeit moderate) step towards understanding the identifiability of the BM model.
Critical weaknesses
  1. The paper suffers from severe lack of clarity. Many important elements and information are left implicit, or are discussed using informal or otherwise undefined expressions, leaving the reader to try and connect the dots too often. For instance:
  • What are separable template ratios?
  • What is the dimension ceiling?
  • Is X^{(j) a tensor slice?
  • What is the definition of Θr\Theta_r in Definition 1?
  • What is the meaning of the symbol that appears in (4) before SrS_r?
  • The identity UikVijWjk=1U_{ik}V_{ij}W_{jk}=1 appears before giving any definition or explanation of what the factors in the left-hand side mean.
  • What are anchor entries?
  • What is exactly the definition of F(z)F(z)?
  • What is a locally constant-rank point?
  • What is the "stated obstruction" of line 140?
  • What is the "fixed-template mixing family" of Theorem 7?
  • Why is Gaussian elimination performed in 6.1 modulo the prime q=1,000,003?
  • The experiment whose results are presented in Section 6.2 has to be explained carefully. What is being done exactly?
  • What does the phrase "The template-transition experiment perturbs a separable ratio" mean?
  1. Citations and proper discussion on relation to prior literature are often missing as well. To give an important example, no reference is given for the BM decomposition (the main object of this paper) in the Introduction.

  2. There is insufficient historical contextualization and discussion on prior contributions to tensor decomposition identifiability, and on how the currently presented results related to them. The authors have certainly been inspired or otherwise informed by existing papers on CP identifiability, for instance. But this discussion is completely absent from the text, apart from a very succinct mention of some references in Section 7. Even this part misses a very important and definitely related reference, namely "Generic and typical ranks of multi-way arrays" by P. Comon et al., who have measured the dimension of the tangent space at a generic point, for the CP parametrization of a tensor with a given rank.

  3. The paper falls short of giving conditions for global essential uniqueness, apart from the rank-one case.

Questions for authors
  1. Certain results seem to follow quite standard arguments, which are not specific to the studied BM decomposition. For instance, wouldn't Proposition 5 also hold say, for the CP decomposition, with the appropriate changes in the involved objects? Has some similar result been stated in the literature for other tensor decompositions? If so, they have to be cited and discussed here as inspiration for your own developments.

  2. Given that there is plenty of space in the 9th page, why not write some of the proofs in their full form? In particular, some of them are not much longer than their respective sketches given in the main text. One example is the proof of Theorem 3. This would help mitigating the clarity issues raised in the "Weaknesses" part.

2 comments

Thank you for your comments; we respond below and have revised the manuscript accordingly.

Definitions and experimental clarity. Appendix A now starts with a notation table. The 333^3, two-term example in Section 2 contrasts inverse column rescaling, which preserves each component, with a shear changing {1,1}\{1,1\} to {1−ϵ,1+ϵ}\{1-\epsilon,1+\epsilon\} at fixed sum. --Separable template ratios. This means rank⁡(At⊘A1)=1\operatorname{rank}(A_t\oslash A_1)=1, equivalently At=A1⊙ρ(t)σ(t)⊤A_t=A_1\odot\rho^{(t)}\sigma^{(t)\top} (Abstract and Section 4).

--Dimension ceiling. The necessary parameter-count ceiling is r⋆=⌊mpn/(mn+mp+pn−m−p−n+1)⌋r^\star=\lfloor mpn/(mn+mp+pn-m-p-n+1)\rfloor (Abstract and Section 3).

--Tensor slice. X(j)=X:,j,:∈Fm×nX^{(j)}=X_{:,j,:}\in\mathbb F^{m\times n} is a lateral slice. It is defined in Section 2 after (2).

--Parameter space. We now write the parameter space as Θr\Theta_r, with Θr=(F×)r(mn+mp+pn)\Theta_r=(\mathbb F^\times)^{r(mn+mp+pn)} (Section 2).

--Semidirect product. In Gr≅(F×)r(m+p+n−1)⋊Sr\mathscr G_r\cong(\mathbb F^\times)^{r(m+p+n-1)}\rtimes S_r, SrS_r permutes the component scaling blocks (Section 2).

--Factor ratios. For fixed tt, define Uik=A^itk/AitkU_{ik}=\widehat A_{itk}/A_{itk}, Vij=B^ijt/BijtV_{ij}=\widehat B_{ijt}/B_{ijt}, and Wjk=C^tjk/CtjkW_{jk}=\widehat C_{tjk}/C_{tjk} before using UikVijWjk=1U_{ik}V_{ij}W_{jk}=1.

--Anchors. We fix Ait1=A1tk=B1jt=1A_{it1}=A_{1tk}=B_{1jt}=1.

--Normalized map. Fill the anchors with ones to obtain θ(z)\theta(z); z∈(F×)dz\in(\mathbb F^\times)^d holds the remaining entries. Then F(z)=Φr(θ(z))∈Fm×p×nF(z)=\Phi_r(\theta(z))\in\mathbb F^{m\times p\times n} (Section 2).

--Locally constant-rank point. Here rank⁡DF\operatorname{rank}DF is constant in a neighborhood.

--Stated obstruction. It is d≤mpnd\le mpn, assuming local constant rank over R\mathbb R (Section 3).

--Fixed-template mixing family. Vary Sj∈GLrS_j\in GL_r independently in P(j)SjP^{(j)}S_j, Q(j)Sj−⊤Q^{(j)}S_j^{-\top} at fixed AtA_t. Dimension is modulo gauge and permutations (Section 4).

--Modular elimination. With q=1,000,003q=1{,}000{,}003, elimination is exact. A minor nonzero modulo a prime is nonzero in characteristic zero. Attaining min⁡(d,mpn)\min(d,mpn) certifies exact rank without floating-point tolerances; any prime preserving the minor suffices.

--Section 6.2 experiment. Section 6.2 applies the shear to the Appendix C factors, normalizes them, and compares factor variation with unchanged reconstruction; the rank check is reported there.

--Template-transition experiment. We perturb A2=A1⊙(ρσ⊤+ηN)A_2=A_1\odot(\rho\sigma^\top+\eta N) away from separability and track reduced-Jacobian rank. Sections 6.2 and H.1/H.4 give the sampling protocol, spectra, and tolerance-dependent thresholds; Figure 3(b) shows a separate transition experiment.

BM model attribution. The introduction now cites Mesner--Bhattacharya (1990, 1994) and Gnang--Filmus (2020) when introducing BM, and Tian et al. (2025) for video applications.

Earlier identifiability methods. Sections 1 and 7 relate the BM monomial parameterization to the no-three-way-interaction toric model. Over C\mathbb C, the Zariski closure of sums of rr components is its affine secant cone. Terracini gives its generic tangent space as the sum of component tangent spaces. We cite Landsberg (2012), Comon, ten Berge, De Lathauwer and Castaing (2009), and Chiantini and collaborators on CP identifiability. Full-column-rank tests certify nondefectivity and local, not global, uniqueness in the admissible range. Scaling-fiber and inverse-function tools are standard; the effective gauge, canonical section, and slice-mixing families are specific to BM. Appendix B extends existing certificates to infinite shape families. The positive 53,r=25^3,r=2 ambiguity is a special locus in a generically locally identifiable shape. CP can also have exceptional ambiguous points.

Global essential uniqueness. Our global result is confined to rank one. At higher rank, we establish local identifiability only. Section 8 now states explicitly that distinct isolated decompositions of the same tensor are not excluded. A general higher-rank global uniqueness theorem remains open; we make no claim beyond the local certificates.

Perturbation theory. Yes. Proposition 5 applies the implicit function theorem to local least squares, as for CP or any smooth map with full-column-rank derivative. We cite Breiding--Vannieuwenhoven (2018) on join-decomposition conditioning. BM supplies the normalized coordinates and structural obstructions; singular values depend on these coordinates. The local perturbation bound has a quadratic remainder. The residual-to-factor bound requires a Lipschitz radius; neither guarantees global optimization.

Complete proofs. Section 3 gives the full proof of Theorem 3 via the normalized section, finite permutation orbit, inverse function theorem, and constant-rank converse. Section 4 gives the full proof of Theorem 7, with the non-gauge shear, residual diagonal action, and quotient dimension. The notation table and example define the objects used.

I appreciate the authors' efforts in improving the readability of this paper. In consonance, I am upgrading my evaluation.

Reviewer_MHDCOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies when reconstruction determines the factors of a Bhattacharya–Mesner (BM) tensor decomposition, up to scaling and component permutations. It characterizes the complete term-preserving scaling freedom and constructs a canonical normalization. The reduced Jacobian then gives local identifiability and stability criteria, while parameter counting yields a necessary dimension bound. The paper constructs continuous ambiguity families, including a positive rank-two example below that bound. For entrywise-nonzero rank-one tensors, it proves global uniqueness up to gauge and gives explicit recovery formulas and a membership test. Validation combines exact modular Jacobian certificates with synthetic conditioning, noise, and ALS experiments.

Critical strengths

The distinction between scaling freedom and genuine component ambiguity is the strongest part of the paper. The complete gauge characterization makes this distinction precise, and the positive rank-two construction shows why a favorable parameter count does not guarantee identifiable components. Within the constructed family, reconstruction cannot distinguish the components; regularization therefore introduces a selection criterion beyond reconstruction loss. The exact modular certificates also turn the examples in Table 1 into rigorous generic statements over both the real and complex numbers. The paper carefully distinguishes local from global uniqueness and infinitesimal null directions from actual continuous ambiguity families.

Critical weaknesses
  1. The scope and positioning of the identifiability results limit the contribution. For r≥2r\geq2, Table 1 certifies generic local identifiability in three configurations; beyond these, the sufficient condition still requires a separate analysis of Jacobian rank. At rank one, the connection to existing log-linear models deserves explicit treatment. For positive tensors, log⁡Xijk=αik+βij+γjk\log X_{ijk}=\alpha_{ik}+\beta_{ij}+\gamma_{jk} is the no-three-way-interaction model [1], and taking logarithms of Eq. (16) gives exactly the vanishing of the reference-cell three-way interaction. The paper should explain what its membership, recovery, and perturbation results add to this established formulation. This connection alone does not settle the novelty of the higher-rank results.

  2. The ALS comparisons do not cleanly isolate the effect of identifiability on factor recovery. Only 4/50 truth-local runs in the regular setting pass the fit threshold, compared with 50/50 in each nonidentifiable setting. Factor distances also use unrescaled normalized coordinates across settings with different dimensions and scales. Fit quality, selection of successful runs, and coordinate scaling therefore complicate the comparison. The authors acknowledge that the results are qualitative, but Appendix H reports that seed-level arrays, complete generating distributions, full initialization specifications, and run-level audit/filtering logs are unavailable. The observation-noise normalization is also unspecified. These omissions make the empirical conclusions difficult to assess or reproduce, and does not support the exact mathematical results.

Questions for authors
  1. Which parts of the rank-one membership, recovery, and perturbation results go beyond the no-three-way-interaction formulation in [1]? A direct comparison would help assess this contribution.

  2. Were Jacobian conditioning and initialization distances measured for the actual ALS instances? If so, do those diagnostics help explain why only 4/50 truth-local runs succeed in the regular setting? Table 3 does not establish that its singular values correspond to these instances.

  3. Can the missing experimental specifications and run-level results in Appendix H be supplied, including the normalization of τ\tau? Do the existing runs permit a factor-distance comparison at comparable reconstruction accuracy and with coordinate scales accounted for?

[1] Taylor Brysiewicz and Aida Maraj. Lawrence Lifts, Matroids, and Maximum Likelihood Degrees. 2023 preprint

2 comments

Thank you for your comments; we respond below and have revised the manuscript accordingly.

Rank-one contribution and relation to prior work. Section 5 and Appendix D now compare the formulations directly. Positive BM rank one is exactly log⁡Xijk=αik+βij+γjk\log X_{ijk}=\alpha_{ik}+\beta_{ij}+\gamma_{jk}, and our membership identity is its multiplicative reference-cell test. We add explicit recovery in the canonical section shared with higher rank, at cost O(mn+mp+pn)O(mn+mp+pn) for nonzero real or complex entries without logarithms. Deterministic relative perturbation bounds quantify the effect of dynamic range. The formulas are algebraic recovery rather than maximum likelihood estimation. We cite Brysiewicz--Maraj, Algebraic Statistics 16 (2025), 217--242, DOI 10.2140/astat.2025.16.217, for the log-linear context. For higher rank, Section 7 clarifies the standard toric/secant framework and BM-specific gauge and ambiguity results. Appendix B proves slice extension of the existing exact certificates to every m,p,n≥5,r=2m,p,n\ge5,r=2 and every m,p,n≥8,r=3m,p,n\ge8,r=3.

Actual-instance conditioning and initialization. We measured reduced Jacobians at the truth, initialization, and endpoint of the actual truth-local ALS instances. The regular group has full-column-rank truth Jacobians, including instances that miss the initial fit threshold. Conditioning is associated with final fit, whereas initialization distance alone explains little. Continuing the same endpoints with unchanged updates and acceptance rules resolves the low acceptance under the initial budget. This separates slow optimization from local nonidentifiability. Section 6.3 and Figure 2 report the continuation; Appendix H.4 and Table 5 give the instance-level conditioning and initialization diagnostics. The generation and stopping rules are stated there as well.

Experimental specifications, records, and comparable recovery. Appendix H.4 specifies generation, initialization, seeds, stopping rules, refinement, and acceptance filtering. Table 2 compares regular, separable-template, and dimension-obstructed settings at the same fit threshold. It reports qualifying fractions together with gauge-invariant component distance DTD_T and permutation-aligned coordinate distance DzD_z. Accurate qualifying fits recover consistent components in the sampled regular settings; the ambiguous settings retain component differences. This comparison is conditional on unequal qualifying fractions and uses a separate protocol from the ALS stress test. All attempts, including the refinement exception, remain recorded. Original aggregate studies are distinguished where seed or noise records are unavailable. Noise-calibrated AUC results, null rejection rates, and uncertainty intervals are reported in Appendix H.2. Sections 6.2 and H.1/H.4 also state the template-transition threshold consistently at the specified rank tolerance.

Thank you for the revisions. The slice-extension result addresses my concern about the scope of the identifiability guarantees, and the ALS continuation experiment helps explain the original 4/50 acceptance rate.

I have one clarification about Table 2. In the regular and separable 5×5×55 \times 5 \times 5 settings, how many of the 20 ground-truth tensors contributed at least one qualifying fit in each group? Does the gap in component error persist if you first take the median across qualifying fits for each tensor and then take the median across tensors? This would show how much the reported gap depends on tensors with more successful initializations.

Reviewer_LGbUOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This is a solid, theoretically innovative work that develops a relatively complete framework for the identifiability of BM tensor representations. The authors systematically characterize the full gauge freedom beyond the diagonal scaling reported in prior literature, establish local identifiability criteria via the reduced Jacobian and inverse function theorem, derive necessary dimensional bounds, and construct explicit positive ambiguous families to demonstrate that the bound is not sufficient. Rigorous proofs are provided in the appendix, and carefully designed synthetic experiments verify the theoretical predictions. The results deepen the understanding of BM decomposition and offer valuable insights for its practical use.

Critical strengths

This work establishes a complete and rigorous mathematical pipeline for the identifiability analysis of BM tensor representations. It proceeds systematically from the algebraic derivation of the gauge group, canonical normalization, reduced Jacobian construction, local identifiability via the inverse function theorem, necessary dimensional counting conditions, explicit analytic construction of ambiguous families together with analytical Jacobian rank verification, global uniqueness for rank‑1 BM tensors, and perturbation error bounds. Sketch or full proofs are provided in Appendices, rather than relying merely on empirical conjectures from numerical experiments.

Prior work (Tian et al., 2025) only identified simple diagonal scaling gauge ambiguities for BM decomposition, restricted to diagonal transformation matrices S. This paper fully characterizes the broader gauge freedom induced by general invertible linear transformations S, completing the classification of all gauge equivalence operations for BM factors.

Critical weaknesses

W1:The paper only derives necessary conditions for generic local identifiability, and does not provide easy‑to‑check sufficient conditions. The Jacobian rank criterion is a posteriori: one must instantiate and evaluate the Jacobian at a given point to test identifiability. It lacks simple, explicit, pre‑verifiable rank conditions analogous to the Kruskal condition for CP decomposition, which can be directly checked from factor matrices before optimization or tensor recovery. This limits its practical usability for practitioners who want to judge identifiability in advance.

W2:All experiments are conducted on synthetic tensors under idealized assumptions (full observation, all entries non‑zero). There are no validations on real‑world data such as the video datasets that originally motivated BM decomposition. Real data typically contains measurement noise, zero entries, missing observations and non‑ideal distributional features, which fall outside the theoretical assumptions adopted here. It remains unclear whether the identifiability and instability phenomena discovered in this work transfer to practical application scenarios.

Questions for authors

Q1. You show that the parameter counting bound is only necessary and not sufficient. Could you comment on whether there exists a natural, tractable sufficient condition for generic local identifiability of BM tensors, and what major obstacles prevent you from establishing such a condition?

Q2. Your local identifiability relies on the inverse function theorem and Jacobian rank. Since the Jacobian test is a posteriori, do you have any suggestions for practitioners to design tensor dimensions and ranks in advance to avoid structural ambiguity?

Q3. All experiments are synthetic. What are the main theoretical barriers that prevent you from testing your conclusions on real video tensors? Could you outline your roadmap for real‑data validation in future work?

1 comment

Thank you for your comments; we respond below and have revised the manuscript accordingly.

Sufficient conditions and theoretical obstacles. An exact full-column-rank witness certifies generic local identifiability for its shape before data are observed. The slice-extension lemma in Appendix B propagates existing certificates to all m,p,n≥5,r=2m,p,n\ge5,r=2 and m,p,n≥8,r=3m,p,n\ge8,r=3. Its block-Jacobian proof requires no additional single-slice certificate. These are generic shape-level conditions; a Kruskal-type criterion for individual fitted points remains open. The main obstacle is proving independence of component tangent spaces after quotienting the gauge. BM components form a pairwise-interaction toric model, and the separable-template construction shows why parameter counting and nonseparability alone are insufficient. Theorem 2 describes term-preserving rescalings and permutations. Theorem 7 constructs mixing that changes component tensors while preserving their sum, a distinct structural ambiguity illustrated in Section 2.

Advance design of dimensions and ranks. Section 8 gives three steps. Check r≤r⋆(m,p,n)r\le r^\star(m,p,n). Then consult Table 1 or extend a certified base shape with the lemma. After fitting, inspect the smallest reduced-Jacobian singular value and, at r=2r=2, the ratio gaps in all three modes. These diagnostics are local and depend on the anchor coordinates. Local identifiability does not imply that distant optimization restarts find the same solution.

Real-video validation. Appendix I now reports Car reconstruction, a dimension-obstructed pilot control, and local sensitivity around a fitted rank-one output, with the data and protocols in Tables 6--7 and Figure 6. The pilot shows that accurate reconstruction can coexist with different components. Its parameter-count-admissible counterpart does not reach the common fitting threshold, so it supports no matched-accuracy consistency claim. The perturbation experiment supports the local prediction at the fitted model point. These results concern processed video, not identifiability of the original-resolution data. Section 8 now summarizes the barriers posed by incomplete observations, factor zeros, and approximate low rank. Missing observations require a row-restricted Jacobian; zero pixels do not specify factor supports. Appendix I also gives an initial full-column-rank certificate under one fixed observation mask on a synthetic witness. The roadmap is to develop observation-restricted certificates, specify support models separately, and assess sensitivity across disjoint frame subsets. The masked check is a single-witness result, while approximate-fit control remains local under Corollary 6.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.