REVIEW 2 major objections 5 minor 1 cited by
An introduction to tensors for path signatures
T0 review · 2 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash
Pith's one-line read Minimal factorization of tensor product expressions is exactly tensor rank decomposition, so order-2 cases are easy and higher orders are hard.
desk verdict A clear, honest expository chapter for a book, not a research contribution: the factorization/rank content is explicitly the known tensor-rank problem, and the main real defect is a sloppy definition of the tensor product space that a referee should make the authors fix. 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 mechanism is the rank-decomposition identity: after choosing bases, an order-2 tensor product expression $X=\sum_{ij} X_{ij}\, a_i\otimes b_j$ factors into $r$ terms exactly when the coefficient matrix satisfies $X=D_1^{\mathsf T}D_2$, so the minimal term count is the matrix rank. Reduced row echelon form and singular value decomposition provide the polynomial-time algorithms, while the universal property of the tensor product is what allows bilinear problems to be represented linearly. For higher orders, the same coefficient matching turns factoring into tensor rank, and the known complexity and ground-field dependence of tensor rank are imported directly.
What would settle it
Try to find a two-term real decomposition of the $2\times2\times2$ tensor $Z$ in equation (41), which the paper says has real rank 3; exhibiting such a decomposition would refute the claimed ground-field dependence and the translation of rank into minimal factorization.
Extended reading notes
Core claim
In equations (31) through (35), the paper equates the coefficients of an order-2 tensor product expression with a sum of products of coefficients, which is precisely a rank decomposition of the coefficient matrix. Hence the minimal number of terms in a factorization is the tensor rank, equal to the usual matrix rank for order 2, and it can be computed in polynomial time using reduced row echelon form or singular value decomposition. For order 3 and above, the same identification shows minimal factorization is NP-hard, and the rank can depend on whether real or complex coefficients are allowed, as demonstrated by a concrete $2\times2\times2$ example. The paper presents these as established results of the tensor-rank literature and contributes an expository route together with accompanying computational code rather than a new theorem.
Load-bearing premise
The argument assumes that the tensor product of two vector spaces has been properly defined as a vector space; as written, the addition rule in Definition 2.7 only covers pairs sharing a component, so the construction is not fully specified without the quotient-space repair mentioned in a footnote.
Editorial extensions
If this is right
- For order-2 tensor product expressions, one tensor product per term, minimal factorizations can be found in polynomial time using reduced row echelon form or singular value decomposition.
- For expressions with three or more factors, minimal factorization is NP-hard, so no polynomial-time algorithm exists unless P equals NP.
- A real expression may require more terms when factored over the reals than over the complex numbers, so the coefficient field must be specified when asking for a minimal factorization.
- Level-2 blocks of path signatures are order-2 tensors, so their minimal factorizations inherit the matrix-rank algorithms presented here.
- A greedy grouping procedure can get stuck on a nonminimal factorization; the rank-based criterion is what certifies minimality.
Reading between the lines
- The paper's identification suggests a natural workflow for truncated signatures: compress level-2 signature blocks with singular value decomposition and keep only the dominant singular vectors, an extension the text does not develop.
- A rigorous repair of Definition 2.7 via quotient spaces, which the paper defers in a footnote, would put the universal property and hence the rank identification on fully solid ground without changing the examples.
- The same coefficient-matching argument suggests approximate factorization: for order 2, singular value decomposition gives the best low-rank approximation, so near-minimal factorization is numerically tractable, whereas the paper implies no analogous guarantee for higher orders.
- The real-versus-complex example could become a classroom check: asking students to attempt a two-term real decomposition with symbolic algebra illustrates ground-field dependence even without a proof of impossibility.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. This expository paper introduces vector spaces, direct sums, tensor products, and the tensor algebra, aiming to build intuition for path-signature applications. Sections 5 and 6 (the extended part) define tensor rank and rank decompositions, give two polynomial-time algorithms for order-2 tensors (reduced row echelon form and SVD, with Mathematica implementations), state the NP-hardness of higher-order tensor rank, and argue that the problem of minimally factoring tensor-product expressions is exactly the tensor-rank problem. The paper contains exercises with solutions and a GitHub repository. The core mathematical content is standard; the pedagogical value depends on the correctness of the initial construction of the tensor product.
Significance. If the construction issue in Definition 2.7 is repaired, this paper would be a useful and well-targeted introduction for the signatures community. Its strengths are the explicit connection to path signatures, the worked examples (including the R-vs-C rank Example 5.9), and the reproducible Mathematica code that verifies the decompositions. The NP-hardness and polynomial-time facts cited are standard and correctly attributed. The paper does not claim a new theorem; its value is expository.
major comments (2)
- [§2.2, Definition 2.7] The claim that V×W 'equipped with the bilinearity operations' is a vector space is not well-defined. Addition is specified only for pairs sharing a first or second component, so it is not a total binary operation on V×W; and Solution 4.2 gives the additive inverse of (u,v) as both (−u,v) and (u,−v), which are distinct elements of V×W unless a quotient is taken. This is not a harmless informality: Proposition 2.11 (the basis of V⊗W), Theorem 2.12 (the universal property), Definition 3.5 (tensor algebra), and the coefficient comparison in §6.3, equations (34)–(35), all rely on a bona fide vector space structure. Footnote 4 defers the quotient construction, but the deferral leaves the central construction without a valid definition. The paper should either present the standard quotient construction, or explicitly say that the 'tensor product space' is introduced informally and that all results are conditional on the quotient construction in the cited references.
- [§6.4, equations (38)–(43)] The text states that the expression Z in (38) 'can be minimally factored to three terms over R' and 'to two terms over C,' but no proof of minimality is given and no caveat is offered. This is in tension with the paper's own emphasis (Remark 5.3, footnote 10) that verifying minimality is the nontrivial part. Please add a proof or reference for the real rank of this specific Z, or rephrase the statement as an upper bound with a warning that minimality over R is not shown here (as was done for Example 5.9).
minor comments (5)
- [Example 2.13] 'The map φ is injective' should read 'The map \hat{φ} is injective,' since φ is a bilinear map on the Cartesian product, not a linear map on V⊗W.
- [Theorem 3.7] In the universal property, the displayed identity uses f(u⊗v) = f(u)·_A f(v); since f is defined only on V, the map on T(V) should be \hat{f} throughout.
- [Definition 3.11] In the Cauchy product formula, w_n is said to lie in (V*)⊗n; the dual should be V^{⊗n}.
- [§5.2] 'For tensors of order 2 or greater it is much more difficult, in fact, it is NP-hard' should read 'order greater than 2,' otherwise it contradicts the preceding sentence, which states that order-2 tensors are straightforward.
- [§6.4, first paragraph] The phrase 'The number of terms of terms in a minimal factorization' contains a duplicated 'of terms'.
Circularity Check
No significant circularity: the Section 6.3 identification of minimal factorization with tensor rank is a direct coordinate translation, and every load-bearing external result is cited to independent literature.
full rationale
The paper contains no fitted parameters, no self-citation chain, and no load-bearing claim that reduces by construction to its own input. The central identification in Section 6.3 (equations (31)-(35)) compares the defining expression of a tensor product expression (equation (26)) with the defining expression of a rank decomposition (Definition 5.2), expands the candidate factors in bases, and uses the basis property of elementary tensors (Proposition 2.11) to equate coefficients; this is a genuine translation between two independently introduced notions, not a prediction of a fitted quantity. Complexity and uniqueness facts are attributed to external benchmarks: Håstad [4] for NP-hardness of higher-order tensor rank, Kruskal [7] for real-versus-complex rank differences and uniqueness, and Horn and Johnson [5] for matrix rank factorizations. There are no self-citations among the load-bearing references. The one notable rigor gap is Definition 2.7, where V⊗W is presented with only component-sharing addition rules and with a footnote conceding that a quotient-space construction is needed but is "not really necessary at the moment"; this is a correctness and construction gap, not a circularity, because the missing quotient argument would not presuppose the rank-factorization conclusions. No step satisfies the quoted-reduction standard for circularity.
Assumptions & free parameters
assumptions (4)
- standard math The tensor product V⊗W exists and is uniquely characterized by the universal property (Theorem 2.12); the proof is omitted.
- ad hoc to paper V×W equipped with the partial bilinearity rules of Definition 2.7 is a vector space; the rigorous quotient-space construction is deferred.
- standard math Given bases of V and W, {u⊗v : u, v basis vectors} is a basis of V⊗W and dim(V⊗W) = dim(V)·dim(W) (Proposition 2.11); stated without proof.
- domain assumption Tensor rank computation for order ≥ 3 is NP-hard, read under the standard assumption P ≠ NP.
Cite this review
Pith. "Pith review of An introduction to tensors for path signatures." pith.science (2026). https://pith.science/paper/XQHXWBKF
@misc{pith2026250215703,
author = {Pith},
title = {Pith review of: An introduction to tensors for path signatures},
year = {2026},
howpublished = {\url{https://pith.science/paper/XQHXWBKF}},
note = {Machine review of arXiv:2502.15703}
}
read the original abstract
We present a fit-for-purpose introduction to tensors and their operations. It is envisaged to help the reader become acquainted with its underpinning concepts for the study of path signatures. The text includes exercises, solutions and many intuitive explanations. The material discusses direct sums and tensor products as two possible operations that make the Cartesian product of vectors spaces a vector space. The difference lies in linear Vs. multilinear structures -- the latter being the suitable one to deal with path signatures. The presentation is offered to understand tensors in a deeper sense than just a multidimensional array. The text concludes with the prime example of an algebra in relation to path signatures: the 'tensor algebra'. This manuscript is the extended version (with two extra sections) of a chapter to appear in Open Access in a forthcoming Springer volume ``Signatures Methods in Finance: An Introduction with Computational Applications". The two additional sections here discuss the factoring of tensor product expressions to a minimal number of terms. This problem is relevant for the path signatures theory but not necessary for what is presented in the book. Tensor factorization is an elegant way of becoming familiar with the language of tensors and tensor products. A GitHub repository is attached.
Forward citations
Cited by 1 Pith paper
-
Dynamic Universal Approximation via Signature Controlled Differential Equations
Any well-posed path-dependent controlled differential equation can be uniformly approximated, over bounded control and initial-history sets, by signature-controlled equations with a single monotone activation.
Reference graph
Works this paper leans on
-
[1]
I. Chevyrev and A. Kormilitzin. A primer on the signature m ethod in machine learning. arXiv preprint arXiv:1603.03788v2, 2016
arXiv 2016
-
[2]
J. Diestel, J. H. Fourie, and J. Swart. The metric theory of tensor products . American Mathe- matical Society, Providence, RI, 2008. Grothendieck’s r ´esum´e revisited
work page 2008
- [3]
- [4]
-
[5]
R. A. Horn and C. R. Johnson. Matrix Analysis. Cambridge University Press, Cambridge, 2nd edition, 2012
work page 2012
-
[6]
T. G. Kolda and B. W. Bader. Tensor decompositions and appl ications. SIAM Rev., 51(3):455– 500, 2009
work page 2009
-
[7]
J. B. Kruskal. Rank, decomposition, and uniqueness for 3- way and /u1D441-way arrays. In Multiway data analysis (Rome, 1988) , pages 7–18. North-Holland, Amsterdam, 1989
work page 1988
-
[8]
S. Rabanser, O. Shchur, and S. G¨ unnemann. Introduction t o tensor decompositions and their applications in machine learning. arXiv:1711.10781, 2017
arXiv 2017
Show all 9 references
-
[9]
the homogeneous component of degree /u1D45B
M. Reed and B. Simon. Methods of modern mathematical physics. I . Academic Press, Inc. [Harcourt Brace Jovanovich, Publishers], New Y ork, secondedition, 1980. Functional analysis. 5 Extended Section: Tensor rank The goal of Sections 5 and 6 is to introduce an applicable prob ...
1980
Reviewed August 10, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.