Minimality of Tree Tensor Network Ranks

Jana Jovcheva; Nick Vannieuwenhoven; Tim Seynnaeve

REVIEW 2 major objections 5 minor 1 cited by

Minimality of Tree Tensor Network Ranks

T0 review · 2 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read For tree tensor networks, bond dimensions are minimal exactly when local inequalities hold at every vertex.

desk verdict A clean proof of the tree tensor network minimality characterization, solid over infinite fields, with an unstated field assumption that makes the genericity claim false over finite fields. read the letter →

arxiv 2509.09463 v1 pith:FRFIKWC6 submitted 2025-09-11 math.NA cs.NA

Jana Jovcheva , Tim Seynnaeve , Nick Vannieuwenhoven This is my paper · ORCID

classification math.NAcs.NA MSC 15A6965F99

keywords treetensornetworkbonddimensionsminimalrankadmissiblerankseffectivemultilinearZariskiopendenseH-Tuckerdecomposition

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper settles when a tuple of bond dimensions in a tree tensor network is genuinely minimal—that is, when some tensor can be represented with exactly those bond dimensions and no strictly smaller ones. The answer is a set of local inequalities: at every vertex, each incident bond dimension must be no larger than the physical dimension at that vertex times the product of the other incident bond dimensions. If these inequalities hold, the tensors that need exactly that bond dimension form a Zariski open and dense subset of the representable tensors, so minimality is generic; if they fail, no tensor requires exactly that tuple. The proof works by showing that a representation is minimal precisely when every local core tensor has full effective multilinear rank along its incident edges. This gives a practical test for overparameterization in tensor-network compression and generalizes the known star-graph (Tucker) characterization to arbitrary tree topologies.

What carries the argument

The key object is the effective multilinear rank of a local core tensor: the tuple of ranks of its flattenings, one per incident bond edge, obtained by treating that bond space as rows and all other bond spaces together with the physical space as columns. Lemma 3.2 shows that if every local flattening has full rank equal to the bond dimension, then contracting an entire subtree preserves injectivity, so the flattening of the full tensor along any cut has rank equal to the bond dimension across the cut. Combined with the admissibility inequalities, this makes full effective multilinear rank a generic condition, and the paper uses this to prove both minimality and openness/density.

What would settle it

Over a finite field such as $\mathbb{F}_2$, construct a tree network and an admissible tuple $r$ for which no choice of local tensors achieves full effective multilinear rank everywhere—for instance, a vertex with two neighbors, physical dimension 2, and both bond dimensions 2, where the two rank-2 flattening conditions may have empty intersection over $\mathbb{F}_2$. If such a configuration exists, Theorem 3.6 fails as stated, showing the infinite-field hypothesis is essential.

Watch

Extended reading notes

Core claim

The central theorem (Theorem 3.6) characterizes minimal tree tensor network ranks: a tuple $r$ of bond dimensions is admissible—satisfying $r_{ij} \leq \dim V_i \prod_{k\in\operatorname{nb}(i)\setminus\{j\}} r_{ik}$ for every edge $(i,j)$—if and only if $\operatorname{TN}^\circ(G,r)$ is nonempty. In that case $\operatorname{TN}^\circ(G,r)$ is a Zariski open and dense subset of $\operatorname{TN}(G,r)$, so a generic tensor representable with bond dimensions at most $r$ is actually representable with exactly $r$ and no smaller tuple. If $r$ is not admissible, then $\operatorname{TN}^\circ(G,r) = \emptyset$. Along the way the paper proves (Theorem 3.5) that $r$ is minimal for a particular tensor $T$ if and only if each local core tensor $T_i$ has effective multilinear rank equal to the bond dimensions on its incident edges.

Load-bearing premise

The proof that an admissible tuple is actually attained relies on intersecting finitely many nonempty Zariski open sets being nonempty, which holds for infinite fields but not for finite fields; the paper does not state the field is infinite.

Editorial extensions

If this is right

Minimality of a tree tensor network becomes checkable by a purely local set of inequalities, so one can decide without computing any tensor whether a given bond-dimension tuple is a valid tree tensor network rank.
In admissible networks, the non-minimal tensors form a Zariski closed subset, meaning that a generic tensor has the full rank; numerical algorithms that see full-rank behavior are observing the generic case.
A leaves-to-root Hierarchical SVD procedure can reduce any non-minimal network to a minimal one by truncating local Tucker decompositions and absorbing factors into neighboring vertices.
The result extends the star-graph (Tucker) rank characterization to all tree topologies, so earlier conditional results that assume 'r is a tree tensor network rank' can now be replaced by explicit inequality checks.
The effective-multilinear-rank equality gives a certificate of minimality for a concrete tensor: if each core tensor flattens to full rank along every incident edge, the representation cannot be compressed without changing the topology.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The admissibility inequalities resemble a Hall-type condition for the existence of full-rank flattenings; this suggests a matroid or bipartite-graph interpretation of tree tensor network ranks that the paper does not explore.
The theorem as stated requires an infinite base field: the proof that an admissible tuple is attained uses that a finite intersection of nonempty Zariski open sets is nonempty, which fails over finite fields such as F_2. Practical implementations over real or complex arithmetic remain valid, but exact computation over finite fields would need a separate argument.
Because Zariski density implies Euclidean density, the result predicts that random sampling inside TN(G,r) will almost surely hit minimal-rank tensors; this is a testable numerical prediction for hierarchical tensor formats.
One could extend the same local-inequality criterion to more general tensor network graphs (with cycles) by asking whether the contraction map is birational, but the tree-specific proof does not transfer directly.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Colleague,

This paper gives the first complete if-and-only-if characterization of minimal bond dimensions for arbitrary tree tensor networks: a tuple r is minimal for some tensor iff the local inequalities (4) hold at every vertex. That is a genuine extension of Carlini–Kleppe (star graphs) and of the necessity part known for binary trees. The authors are honest that it will not surprise experts, but they do fill in the sufficiency argument nobody had written down.

The proof is mostly clean. Lemma 3.2 is a correct induction on root depth; Theorem 3.4 is the standard generic-rank argument; Theorem 3.5's two directions are coherent, with the sufficiency direction relying on Ye–Lim's flattening rank bounds, which is reasonable and not circular. I see no free parameters or invented entities.

The real soft spot is the field. The paper states everything over "a field k" and then uses Zariski topology and irreducibility. For the topology to behave as claimed, k must be infinite (ideally algebraically closed). Over finite fields Theorem 3.6's "Zariski open and dense" claim is false: the F_2 counterexample with a 3-leaf star and bond dims (2,2,2) gives TN equal to the whole 2×2×2 tensor space while TN^◦ is the proper nonempty set of full-multilinear-rank tensors, which cannot be dense in the discrete Zariski topology of a finite field. So the theorem as literally stated is too broad. I think this is a minor issue for the intended audience (numerical tensor networks work over R or C), but the authors need to state the field hypothesis explicitly and adjust the topological assertions accordingly.

Overall: the paper is worth serious refereeing. It resolves a natural question, is readable, and gives practitioners a concrete inequality check for overparameterization. The main theorem's algebraic core holds up. I would send it to review, with the request that the field assumption be made explicit.

My scores: reading group yes, cite yes.

Referee Report

2 major / 5 minor

Summary. The paper studies minimal bond dimensions for tree tensor networks. A tuple r of bond dimensions is called minimal for a tensor T if T can be represented with r but not with any componentwise smaller tuple. The main result, Theorem 3.6, characterizes when this happens: r is admissible, i.e. the local inequalities (4) hold at every vertex, if and only if the set TN^◦(G,r) of tensors whose tree tensor network rank equals r is nonempty. In the admissible case, TN^◦(G,r) is further claimed to be a Zariski open and dense subset of TN(G,r), so that minimality is generic. The proof is built on Theorem 3.5, which shows that r is minimal for T exactly when every core tensor has full effective multilinear rank equal to the incident bond dimensions, and on a local openness result (Theorem 3.4). The paper generalizes the Carlini–Kleppe characterization from star/Tucker graphs to arbitrary trees.

Significance. If the results are correct, this is a substantial and clean contribution: it reduces a global minimality question for tree tensor networks to finitely many local inequalities, proves that non-minimal tensors form a Zariski closed exceptional set, and provides a practical criterion for model reduction. The proof strategy is transparent and largely self-contained, using local Tucker refactoring for necessity and edge-cut flattening ranks for sufficiency. The paper also gives a clear reduction algorithm in Section 4. However, the central theorem is stated over an unspecified field k, and the algebraic-geometric arguments require k to be infinite; as stated, parts of the main theorem are false over finite fields. This is a load-bearing issue that must be fixed before the paper can be accepted.

major comments (2)

[Section 1 / Theorem 3.4 / Theorem 3.6] The base field k is introduced only as 'a field' in Section 1. The proof of Theorem 3.4 relies on the assertion that a finite intersection of nonempty Zariski open subsets of an affine space is nonempty, and Theorem 3.6 relies on irreducibility of TN(G,r) and on density of nonempty Zariski open subsets. Both statements fail over finite fields: in A^1 over F_2, {x≠0} and {x≠1} are nonempty open sets with empty intersection, and every subset of F_2^N is Zariski closed. Concretely, for the 3-leaf star graph with dim V_i=2 and r=(2,2,2), the tuple is admissible, TN(G,r)=F_2^8, and TN^◦(G,r) is nonempty but not dense. Hence Theorem 3.6 is false as stated if k may be finite. Please assume k is infinite (or algebraically closed) throughout, or state and prove a separate finite-field version.
[Theorem 3.5, Necessity] In the refactoring step, the text says to absorb the factor matrices A_i^{(j')} into adjacent vertices only for j'≠j. To actually replace the edge dimension r_ij by the smaller μ_ij, the factor matrix A_i^{(j)} must also be absorbed into the vertex on the other side of the edge; otherwise the edge space E_ij remains r_ij-dimensional and no component of r is strictly reduced. As written, the construction is incomplete. If the exclusion of j is a typo, it should be corrected; if not, the argument needs to explain how the deficient edge's dimension is reduced.

minor comments (5)

[Theorem 3.4] The sentence 'Because of (4) a generic tensor will flatten to a rank r_ij matrix' is too terse. Please spell out that the set of r_ij × C matrices of rank r_ij is Zariski open and nonempty precisely when r_ij ≤ C, which is exactly inequality (4).
[Theorem 3.5, Sufficiency] The application of Lemma 3.2 to an arbitrary edge (a,b) requires choosing a root orientation of G. It would help to say this explicitly before the 'Without loss of generality, assume a is the parent of b' sentence.
[Theorem 3.6] The complement is taken over the finite set {s ∈ N^E : s ≤ r, s ≠ r}; please state this explicitly to avoid a reader worrying about infinite unions.
[Section 4] The 'leaves-to-root Hierarchical SVD' reduction is only sketched. A sentence connecting it to Theorem 3.5's equality criterion would clarify why the resulting network is minimal.
[Throughout] Minor typographical and notation issues: the author line contains 'JANA JOVCHEV A' with a stray 'A'; some notation such as cM_a is used before being formally defined. Please proofread carefully.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation; central theorem rests on independent external results. Only a minor non-load-bearing self-citation ([VVM12]) appears; the finite-field caveat is a correctness gap, not circularity.

full rationale

The main derivation chain is not circular. Definition 3.1 ('admissible') is a local dimension inequality; Definition 1.2/TN^o formalizes minimality independently as non-representability with strictly smaller bond dimensions. Theorems 3.5 and 3.6 prove their equivalence using in-paper lemmas (Lemma 3.2, Theorem 3.4) and external results: [CK11] for the star/Tucker case, [YL19, Thm 8.3/8.8, Cor 8.9] for flattening-rank bounds and irreducibility, and [CLO15] for Zariski topology facts. None of these are authored by the present authors, and they do not assume the target minimality theorem. The only self-citation, [VVM12] in Section 4, is used for a standard (ST-)HOSVD recompression routine and is not load-bearing for the main theorem. A non-circular caveat: the proof of Theorem 3.4 invokes 'the intersection of finitely many nonempty Zariski open subsets of an affine space is again nonempty', which can fail over finite fields, and the main theorem is otherwise stated over an arbitrary field k; this is a correctness/assumption gap, not a circularity. Overall the paper's central claim has independent content and no step reduces by construction to its inputs.

Assumptions & free parameters 0 free parameters · 5 assumptions · 0 invented entities

The paper relies on standard algebraic geometry, the structural theory of tensor network varieties from [YL19], and the classical matrix rank fact. It introduces no fitted parameters and no new postulated entities. The only unstated load-bearing assumption is that the field is infinite.

assumptions (5)

standard math Zariski topology facts: images of affine spaces under polynomial maps are irreducible; nonempty Zariski open sets in irreducible spaces are dense; the set of full-rank matrices of size m x n is Zariski open when m <= n.
Used throughout Section 2 and in Theorems 3.4 and 3.6, e.g., the proof of Theorem 3.4 and the density argument in Theorem 3.6.
domain assumption TN(G,r) is an irreducible algebraic variety and membership is governed by edge flattening rank bounds (Theorems 8.3, 8.8, Corollary 8.9 of [YL19]).
Imported from Ye and Lim; used in the sufficiency proof of Theorem 3.5 to bound flattening ranks of alternative representations and in Theorem 3.6 to establish that TN(G,s) is closed.
standard math Tensor product of injective linear maps is injective (Greub, [Gre78, eq. 1.12]).
Used in the induction step of Lemma 3.2 to conclude rank(cM_a) = r_ap.
domain assumption The base field k is infinite (in practice R or C).
Theorem 3.4 requires the intersection of finitely many nonempty Zariski open subsets of an affine space to be nonempty, which fails over finite fields. The paper does not explicitly state this assumption.
domain assumption Carlini-Kleppe characterization for star graphs (equation (3)).
Used in Section 2 to set up the star graph case and in the necessity proof of Theorem 3.5 for the Tucker refactoring of a local tensor.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Minimality of Tree Tensor Network Ranks." pith.science (2026). https://pith.science/paper/FRFIKWC6

@misc{pith2026250909463,
  author       = {Pith},
  title        = {Pith review of: Minimality of Tree Tensor Network Ranks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FRFIKWC6}},
  note         = {Machine review of arXiv:2509.09463}
}

abstract

For a given tree tensor network $G$, we call a tuple of bond dimensions minimal if there exists a tensor $T$ that can be represented by this network but not on the same tree topology with strictly smaller bond dimensions. We establish necessary and sufficient conditions on the bond dimensions of a tree tensor network to be minimal, generalizing a characterization of Carlini and Kleppe about existence of tensors with a given multilinear rank. We also show that in a minimal tree tensor network, the non-minimal tensors form a Zariski closed subset, so minimality is a generic property in this sense.

Figures

Figures reproduced from arXiv: 2509.09463 by the authors.

**Figure 1.** Contraction of two tensor vertices i and j along shared edge (i, j), pairing the vector space Eij with its dual E ∗ ji, resulting in a new node k with structure inherited from Vi and V ∗ j . Note that the dangling edges are not edges in E, and merely provide a visualization of the vector space associated with each vertex. E ⊆ {(i, j) : i, j ∈ V, i ̸= j} are directed edges. The ordering (i, j) indicates that the edge… view at source ↗

**Figure 2.** Star graph for a 5th order tensor. The main result of this work is a necessary and sufficient condition (Definition 3.1) on (G, r) for TN(G, r) to be a minimal tree tensor network. To prove this condition, we show that a tensor network represents a tensor in TN◦ (G, r) if and only if each core tensor has full effective multilinear rank (Definition 3.3). When our condition is satisfied, the set of tensors TN◦ (G, r) … view at source ↗

**Figure 3.** Illustration of the refactoring procedure. (1) Local tensor at vertex i connected to neighbouring vertices j1, j2, j3 with bond dimensions rij1 , rij2 , rij3 . (2) Tucker decomposition of Ti into core tensor Ci and factor matrices A (j) i . (3) Refactored network with reduced bond dimension µij1 < rij1 achieved by absorbing factor matrices into vertices adjacent to i. Relabel the tree G as follows. Replace Ti by the… view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Numerical Algebraic Geometry for Energy Computations on Tensor Train Varieties
math.AG 2025-12 conditional novelty 6.0 of 10

Energy minimization over tensor-train states has a well-defined Rayleigh-Ritz degree; for small systems homotopy continuation enumerates all critical points, showing ALS often stops at suboptimal local minima and rank...

Reference graph

Works this paper leans on

6 extracted references · 1 linked inside Pith · cited by 1 Pith paper

[5]

Tensor Decomposition for Signal Processing and Machine Learning

12 REFERENCES [Sid+17] N. D. Sidiropoulos, L. De Lathauwer, X. Fu, K. Huang, E. E. Papalex- akis, and C. Faloutsos. “Tensor Decomposition for Signal Processing and Machine Learning”. In:IEEE Transactions on Signal Processing 65.13 (July 2017), pp. 3551–3582.doi:10.1109/tsp.2017.2690524. [Hac19] W. Hackbusch.Tensor Spaces and Numerical Tensor Calculus. Springer,

arXiv 2017
[128]

On the geometry of tensor net- work states

Grad- uate Studies in Mathematics. Providence, RI: American Mathematical Society, 2012, pp. xx+439.isbn: 978-0-8218-6907-9. [LQY12] J. M. Landsberg, Y. Qi, and K. Ye. “On the geometry of tensor net- work states”. In:Quantum Information & Computation12.3-4 (2012), pp. 346–354.doi:10.5555/2230976.2230988. [VVM12] N. Vannieuwenhoven, R. Vandebril, and K. Mee...

arXiv 2012
[136]

Origi- nally published as Band 136 of Grundlehren der mathematischen Wis- senschaften

Universitext. Origi- nally published as Band 136 of Grundlehren der mathematischen Wis- senschaften. New York, NY: Springer New York, 1978, pp. VIII,

1978
[296]

A Multilinear Sin- gular Value Decomposition

isbn: 978-0-387-90284-5.doi:10.1007/978-1-4613-9425-9. [DDV00] L. De Lathauwer, B. De Moor, and J. Vandewalle. “A Multilinear Sin- gular Value Decomposition”. In:SIAM Journal on Matrix Analysis and Applications21.4 (2000), pp. 1253–1278.doi:10.1137/S0895479896305696. [Bel+09] M. C. Beltrametti, E. Carletti, D. Gallarati, and G. M. Bragadin.Lec- tures on C...

doi:10.1007/978-1-4613-9425-9 2000
[2016]

Supervised Learning with Tensor Networks

Springer Cham, 2015, pp. XVI, 646.isbn: 978-3-319-16720-6.doi: 10.1007/978-3-319-16721-3. [SS16] E. Stoudenmire and D. J. Schwab. “Supervised Learning with Tensor Networks”. In:Advances in Neural Information Processing Systems. Ed. by D. Lee, M. Sugiyama, U. Luxburg, I. Guyon, and R. Garnett. Vol

doi:10.1007/978-3-319-16721-3 2015
[2019]

Dimension of Tensor Network Varieties

arXiv:1801.02662. [BLG23] A. Bernardi, C. D. Lazzari, and F. Gesmundo. “Dimension of Tensor Network Varieties”. In:Communications in Contemporary Mathematics 25.10 (Dec. 2023), p. 2250059.doi:10.1142/S0219199722500596. UCLouvain, INMA, ICTEAM, 1348 Louvain-la-Neuve, Belgium Email address:jana.jovcheva@uclouvain.be KU Leuven, Department of Computer Science...

arXiv 2023

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.

[1] [5]

Tensor Decomposition for Signal Processing and Machine Learning

12 REFERENCES [Sid+17] N. D. Sidiropoulos, L. De Lathauwer, X. Fu, K. Huang, E. E. Papalex- akis, and C. Faloutsos. “Tensor Decomposition for Signal Processing and Machine Learning”. In:IEEE Transactions on Signal Processing 65.13 (July 2017), pp. 3551–3582.doi:10.1109/tsp.2017.2690524. [Hac19] W. Hackbusch.Tensor Spaces and Numerical Tensor Calculus. Springer,

arXiv 2017

[2] [128]

On the geometry of tensor net- work states

Grad- uate Studies in Mathematics. Providence, RI: American Mathematical Society, 2012, pp. xx+439.isbn: 978-0-8218-6907-9. [LQY12] J. M. Landsberg, Y. Qi, and K. Ye. “On the geometry of tensor net- work states”. In:Quantum Information & Computation12.3-4 (2012), pp. 346–354.doi:10.5555/2230976.2230988. [VVM12] N. Vannieuwenhoven, R. Vandebril, and K. Mee...

arXiv 2012

[3] [136]

Origi- nally published as Band 136 of Grundlehren der mathematischen Wis- senschaften

Universitext. Origi- nally published as Band 136 of Grundlehren der mathematischen Wis- senschaften. New York, NY: Springer New York, 1978, pp. VIII,

1978

[4] [296]

A Multilinear Sin- gular Value Decomposition

isbn: 978-0-387-90284-5.doi:10.1007/978-1-4613-9425-9. [DDV00] L. De Lathauwer, B. De Moor, and J. Vandewalle. “A Multilinear Sin- gular Value Decomposition”. In:SIAM Journal on Matrix Analysis and Applications21.4 (2000), pp. 1253–1278.doi:10.1137/S0895479896305696. [Bel+09] M. C. Beltrametti, E. Carletti, D. Gallarati, and G. M. Bragadin.Lec- tures on C...

doi:10.1007/978-1-4613-9425-9 2000

[5] [2016]

Supervised Learning with Tensor Networks

Springer Cham, 2015, pp. XVI, 646.isbn: 978-3-319-16720-6.doi: 10.1007/978-3-319-16721-3. [SS16] E. Stoudenmire and D. J. Schwab. “Supervised Learning with Tensor Networks”. In:Advances in Neural Information Processing Systems. Ed. by D. Lee, M. Sugiyama, U. Luxburg, I. Guyon, and R. Garnett. Vol

doi:10.1007/978-3-319-16721-3 2015

[6] [2019]

Dimension of Tensor Network Varieties

arXiv:1801.02662. [BLG23] A. Bernardi, C. D. Lazzari, and F. Gesmundo. “Dimension of Tensor Network Varieties”. In:Communications in Contemporary Mathematics 25.10 (Dec. 2023), p. 2250059.doi:10.1142/S0219199722500596. UCLouvain, INMA, ICTEAM, 1348 Louvain-la-Neuve, Belgium Email address:jana.jovcheva@uclouvain.be KU Leuven, Department of Computer Science...

arXiv 2023