REVIEW 1 major objections 3 minor 23 references
Tensor decompositions on simplicial complexes with invariance
T0 review · 1 major / 3 minor · reviewed 2026-08-14 · deepseek-v4-flash
Pith's one-line read Every tensor fixed by a group action can be rewritten so that the symmetry is explicit, after one enriches the index structure enough to make the action free.
desk verdict Solid framework and correct main theorem, but the abstract overclaims by dropping connectedness; patch that and it's a good paper. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The carrying object is the $(\Omega,G)$-decomposition: a finite index set $I$ with local vectors $v^{(i)}_{\beta}$ indexed by functions $\beta:\widetilde{\mathcal{F}}_i\to I$ from the multiset of facets incident to vertex $i$, combined as $\sum_{\alpha\in I^{\widetilde{\mathcal{F}}}} v^{(0)}_{\alpha|_0}\otimes\cdots\otimes v^{(n)}_{\alpha|_n}$, subject to the equivariance condition $v^{(i)}_{\beta}=v^{(gi)}_{g\beta}$. The proof mechanism is freeness of the action on the facet multiset: a $G$-linear map $z:\widetilde{\mathcal{F}}\to G$ lets the construction average over the group so that the total sum becomes $g\cdot v$, which equals $v$ by invariance, while freeness makes the averaging unambiguous. Weighted complexes enter because raising facet weights and re-labelling copies by group elements is what produces a free refinement of an arbitrary action.
What would settle it
Take the 3-vertex simplex with the cyclic group acting by permuting the three tensor factors, and choose a tensor invariant under that cycle but not under all permutations. Before refinement no $(\Sigma_2,C_3)$-decomposition exists because the decomposition would force full symmetry; after tripling the facet weight the theorem's construction must produce one. Computing both for a small concrete tensor separates the necessity of freeness from the sufficiency of the refined construction.
Extended reading notes
Core claim
The central discovery is an existence theorem: for a connected weighted simplicial complex $\Omega$ with a free action of the group $G$, every $G$-invariant element $v$ of the algebraic tensor product has an $(\Omega,G)$-decomposition, meaning $v = \sum_{\alpha\in I^{\widetilde{\mathcal{F}}}} v^{(0)}_{\alpha|_0}\otimes\cdots\otimes v^{(n)}_{\alpha|_n}$ with local vectors satisfying $v^{(i)}_{\beta}=v^{(gi)}_{g\beta}$. Since any finite group action on a weighted complex can be refined to a free action by multiplying all facet weights by $|G|$ (Proposition 7), the consequence is unconditional: every invariant tensor admits an invariant decomposition after suitable enrichment, using only nonnegative multiples of the vectors of any initial tensor decomposition. The paper also proves an alternative existence criterion for blending actions (Theorem 17), establishes separable and purification analogues under the same hypotheses, and shows that the nonnegative and positive semidefinite decompositions of nonnegative tensors correspond exactly to the separable and purification decompositions of an associated diagonal state (Theorem 43).
Load-bearing premise
The proofs start from a finite sum of elementary tensors, so the framework applies only to tensors in the algebraic tensor product; in infinite dimensions this excludes states that are not finite-rank, and the existence theorems do not by themselves cover them.
Editorial extensions
If this is right
- Translationally invariant matrix product operator forms and symmetric tensor decompositions are recovered as instantiations, so the existence theorems apply directly to them.
- The cost of explicit invariance is controlled: for a free action of a finite group, $\operatorname{rank}_{(\Omega,G)}(v)\le |G|\operatorname{rank}_{\Omega}(v)$, and the same bound holds for the separable rank; for a normal subgroup $H$, $\operatorname{rank}_{(\Omega,G)}(v)\le |G/H|\operatorname{rank}_{(\Omega,H)}(v)$.
- The ordinary tensor rank is the largest of all ranks considered: on any connected complex, $\operatorname{rank}_{\Omega}(v)\le\operatorname{rank}_{\Sigma_n}(v)$, and the same transfer holds for separable and purification ranks.
- For every entrywise nonnegative multipartite tensor, its nonnegative $(\Omega,G)$-rank equals the separable rank of an associated diagonal positive semidefinite matrix, and its psd $(\Omega,G)$-rank equals a purification rank, so all the rank inequalities carry over.
- If the group action is blending, no enrichment is needed: every invariant element has an $(\Omega,G)$-decomposition, including fully symmetric tensors on the $n$-simplex, even when the local spaces are infinite-dimensional.
Reading between the lines
- The algebraic-tensor-product limitation suggests a natural extension: on Hilbert-space tensor products, the results should hold for the closure of finite-rank invariant states by approximation, but the paper does not address approximation error or convergence of the constructed decompositions.
- The minimal facet-weight inflation needed before an invariant decomposition exists can be read as a measure of how hidden the symmetry is; comparing this quantity across complexes could yield a resource theory of symmetry-explicit representations.
- Since tensor rank is the largest rank in the family, hardness results for tensor rank may transfer to the other ranks; the paper does not discuss computational complexity.
- The correspondence in Theorem 43 opens a route to constructing completely positive semidefinite decompositions with controlled rank by choosing a weighted complex whose free action is adapted to the symmetry of the tensor, rather than working on a single edge.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper develops a framework for tensor decompositions whose summation indices are arranged on a weighted simplicial complex Ω and that are manifestly invariant under a group G acting on Ω. It defines (Ω,G)-decompositions, separable decompositions, purification forms, and corresponding ranks, and proves existence theorems under free actions (Theorem 13) and blending actions (Theorem 17). It establishes inequalities among the ranks and applies the framework to nonnegative tensors, recovering several known matrix and tensor decompositions (Theorem 43 and Corollary 44). The main constructive result is that, on a connected Ω, after raising facet weights and refining the group action to be free, every G-invariant tensor admits such a decomposition.
Significance. The framework is genuinely unifying: known cases include translation-invariant matrix product operators, symmetric tensor decompositions, and nonnegative/PSD/cp/cpsd factorizations. The proofs are mostly constructive and the rank inequalities are explicit. The paper is careful to state that it works with algebraic tensor products, so infinite-dimensional Hilbert-space states that are not finite-rank are outside the scope. The main theorem, once the connectedness hypothesis is made explicit, is a useful contribution; the counterexample in the report shows only that the advertised unrestricted version needs amendment.
major comments (1)
- [Abstract; §1; Theorem 13 and Proposition 7] The main existence claim as stated in the abstract, in §1, and in §6 omits the connectedness hypothesis that is explicit in Theorem 13. Proposition 7 only raises facet weights and refines the action on the facet multiset; it does not change the facet incidence graph, so it cannot turn a disconnected complex into a connected one. The omission is load-bearing: let Ω be the disconnected complex on {0,1} with Ω({0})=Ω({1})=1 and Ω({0,1})=0, with C2 acting by swapping the two vertices. After any admissible weight raising, every (Ω,G)-decomposition is a pure tensor, namely (∑_α v^0_α)⊗(∑_β v^1_β), because the two facets are disjoint. The C2-invariant tensor e0⊗e1+e1⊗e0 is not pure, so it has no such decomposition. Thus the unrestricted claim is false; the abstract and introduction must either state the connectedness condition or specify an enrichment operation that can connect the complex, with a proof.
minor comments (3)
- [Theorem 17, proof after Eq. (2)] The step marked '∼' in the proof of Theorem 17 is compressed: after using equation (2), the sum over tuples is not literally equal to ∑_{g∈G} g·v unless one checks that every tuple satisfying {g0 0,...,gn n}=[n] gives a G-translate of v. This is true by blending and by G-invariance of v, but it should be spelled out; the notation w[g0]j in the same display is also ambiguous and should be written as w^{[g·0]}_j.
- [Definition 41(iii)] In the definition of a positive semidefinite (Ω,G)-decomposition, the condition (E[gi]_j)_{gβ,gβ'} = (E[i]_j)_{β,β'} uses the action on functions gβ defined in Section 2; adding one sentence with the functional definition would improve readability, since the matrix index convention is otherwise easy to misread.
- [Throughout] There are a few typographical issues (for example 'Hermitain squares' in Remark 26 and inconsistent hyphenation of 'purification'); these do not affect the mathematics.
Circularity Check
No significant circularity: the main existence results are proven in-paper; the abstract omits the connectedness hypothesis of Theorem 13, but that overstatement is a correctness issue rather than a circular derivation.
full rationale
The central derivation chain is self-contained. Theorem 11 constructs an Omega-decomposition from any finite elementary decomposition for connected complexes, and Theorem 13 then uses only the freeness of the G-action, via a G-linear map z: ~F -> G, to build invariant local vectors from that Omega-decomposition. No parameter is fitted, no empirical input is used, and no imported uniqueness claim is load-bearing. Proposition 7 is a constructive free-refinement argument that multiplies facet weights by |G|; it is not an assumption smuggled in from prior work. Citations to the authors' own [8] appear when recovering or comparing special cases, such as the translationally invariant matrix product operator form, the separable decomposition, and the nonnegative-rank correspondences, but they are not premises of Theorem 13 or Theorem 43. The use of [6, Lemma 4.2] in Theorem 17 is external and concerns a fixed finite-dimensional symmetric tensor in equation (2), independent of the target vector v. The only concerning passage is the abstract's unrestricted claim that the decomposition 'exists for all invariant tensors after possibly enriching the simplicial complex': Theorem 13 explicitly assumes the weighted simplicial complex is connected, and Proposition 7 only multiplies existing facet weights without changing the facet incidence graph, so weight-raising cannot make a disconnected complex connected. That is an overstatement or missing hypothesis, but it is not a circular reduction of the derivation to its own inputs. Accordingly, no circular step is identified, and the paper receives a low score reflecting only this non-circular expositional issue.
Assumptions & free parameters
assumptions (3)
- domain assumption Every element of the algebraic tensor product V is a finite sum of elementary tensors.
- standard math The specific symmetric tensor in equation (2) has a finite symmetric decomposition over C.
- domain assumption Local vector spaces at vertices in the same G-orbit coincide.
Cite this review
Pith. "Pith review of Tensor decompositions on simplicial complexes with invariance." pith.science (2026). https://pith.science/paper/6NPASZBS
@misc{pith2026190901737,
author = {Pith},
title = {Pith review of: Tensor decompositions on simplicial complexes with invariance},
year = {2026},
howpublished = {\url{https://pith.science/paper/6NPASZBS}},
note = {Machine review of arXiv:1909.01737}
}
read the original abstract
We develop a framework to analyse invariant decompositions of elements of tensor product spaces. Namely, we define an invariant decomposition with indices arranged on a simplicial complex, and which is explicitly invariant under a group action. We prove that this decomposition exists for all invariant tensors after possibly enriching the simplicial complex. As a special case we recover tensor networks with translational invariance and the symmetric tensor decomposition. We also define an invariant separable decomposition and purification form, and prove similar existence results. Associated to every decomposition there is a rank, and we prove several inequalities between them. For example, we show by how much the rank increases when imposing invariance in the decomposition, and that the tensor rank is the largest of all ranks. Finally, we apply our framework to nonnegative tensors, where we define a nonnegative and a positive semidefinite decomposition on arbitrary simplicial complexes with group action. We show a correspondence to the previous ranks, and as a very special case recover the nonnegative, the positive semidefinite, the completely positive and the completely positive semidefinite transposed decomposition.
Reference graph
Works this paper leans on
-
[8]
Mixed states in one spatial dimension: decompositions and correspondence with nonnegative matrices
G. De las Cuevas and T. Netzer. Mixed states in one spatial dimension: decompositions and correspondence with nonnegative matrices. arXiv:1907.03664, 2019
work page Pith review arXiv 1907
- [1]
-
[2]
A. Berman and R. J. Plemmons. Nonnegative Matrices in the Mathematical Sciences . SIAM, 1994
work page 1994
-
[3]
D. Bini, G. Lotti, and F. Romani. Approximate solutions for the bilinear form computational problem. SIAM J. Comput. , 9:692, 1980
work page 1980
- [4]
-
[5]
J. Brachat, P. Comon, B. Mourrain, and E. Tsigaridas. Symmetric tensor decomposi- tion. Linear Algebra Appl., 433:1851, 2010
work page 2010
- [6]
-
[7]
R. J. M. Dawson. Homology of weighted simplicial complexes. Cahiers Topologie Geom. Differentielle Categ., 31:229, 1990
work page 1990
Show all 23 references
-
[9]
De las Cuevas, N
G. De las Cuevas, N. Schuch, D. Perez-Garcia, and J. I. Cirac. Purifications of multipartite states: limitations and constructive methods. New J. Phys. , 15:123021, 2013
2013
-
[10]
de Silva and L.-H
V. de Silva and L.-H. Lim. Tensor rank and the ill-posedness of the best low-rank approximation problem. SIAM J. Matrix Anal. Appl. , 30:1084, 2008. TENSOR DECOMPOSITIONS ON SIMPLICIAL COMPLEXES WITH INVARIANCE 33
2008
-
[11]
Fawzi, J
H. Fawzi, J. Gouveia, P. A. Parrilo, R. Z. Robinson, and R. R. Thomas. Positive semidefinite rank. Math. Program., 153:133, 2015
2015
-
[12]
Fiorini, S
S. Fiorini, S. Massar, S. Pokutta, H. R. Tiwary, and R. de Wolf. Exponential lower bounds for polytopes in combinatorial optimization. Journal of the ACM , 62:17, 2015
2015
-
[13]
Glasser, R
I. Glasser, R. Sweke, N. Pancotti, J. Eisert, and J. I. Cirac. Expressive power of tensor-network factorizations for probabilistic modeling, with applications from hidden Markov models to quantum machine learning. arXiv:1907.03741, 2019
1907 arXiv
-
[14]
Gouveia, P
J. Gouveia, P. A. Parrilo, and R. R. Thomas. Lifts of convex sets and cone factorizations. Math. Oper. Res., 38:248, 2013
2013
-
[15]
C. J. Hillar and L.-H. Lim. Most Tensor Problems Are NP-Hard. Journal of the ACM , 60:1, 2013
2013
-
[16]
S. Lang. Algebra. Graduate Texts in Mathematics 211, third edition. Springer-Verlag, New York, 2002
2002
-
[17]
Laurent and T
M. Laurent and T. Piovesan. Conic approach to quantum graph parameters using linear optimization over the completely positive semidefinite cone. SIAM J. Optim. , 25:2461, 2015
2015
-
[18]
R. Orus. Tensor networks for complex systems. Nat. Rev. Phys. , 2018
2018
-
[19]
Schuch, D
N. Schuch, D. Perez-Garcia, and J. I. Cirac. PEPS as ground states: Degeneracy and topology. Ann. Phys., 325:2153, 2010
2010
-
[20]
Y. Shitov. A counterexample to Comon’s conjecture. SIAM J. Appl. Algebra Geom. , 2:428, 2018
2018
-
[21]
Singh, R
S. Singh, R. N. C. Pfeifer, and G. Vidal. Tensor network decompositions in the presence of a global symmetry. Phys. Rev. A , 82:050301, 2010
2010
-
[22]
Yannakakis
M. Yannakakis. Expressing combinatorial optimization problems by linear programs. J. Comput. Syst. Sci. , 43:441, 1991
1991
-
[23]
Zhang, Z.-H
X. Zhang, Z.-H. Huang, and L. Qi. Comon’s conjecture, rank decomposition, and symmetric decomposition of symmetric tensors. SIAM J. Matrix Anal. Appl. , 37:1719, 2016
2016
Reviewed August 14, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.