REVIEW 4 major objections 3 minor 22 references
Path signatures of ODE solutions
T0 review · 4 major / 3 minor · reviewed 2026-08-15 · deepseek-v4-flash
Pith's one-line read The signature of a path—its infinite collection of iterated integrals—determines, through linear equations, whether the path is the jet extension of a solution to a polynomial Cauchy problem.
desk verdict Theorem 5.9's converse is false as stated—condition (3) only forces F_i(X(t)) to be constant, not zero—and Lemma 3.4 has a separate proof gap, though the holonomic characterization in Section 4 is a genuine contribution. 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 key machinery is the signature $\sigma(X)=(\sigma^{(k)}(X))_{k\ge 0}$ of iterated integrals, which by the signature uniqueness theorem determines a reduced path up to translation. Around it the paper builds the shuffle identity, the half-shuffle homomorphism $M_{\tilde p}$ that tracks how a polynomial map transforms signatures and produces equations (3), and the holonomic equations (7), which encode that the path can be read as a jet extension. Theorem 5.9 puts these three pieces together so that the ODE membership test is linear in $\sigma(X)$.
What would settle it
Take the system $F(x_1,x_2,x_3)=x_3-1$, the Cauchy problem $x'=1$, $x(0)=0$, and the path $X(t)=(t,2t,2)$ on $[0,1]$. All three conditions of Theorem 5.9 hold: $X(0)\in\Sigma$, the holonomic equation holds, and $\langle\sigma(X),3w\rangle=0$ for every word $w$ because the third coordinate is constant. The claimed solution $f(t)=X_2(t)=2t$ has $f'=2\ne 1$, so the stated converse fails.
Extended reading notes
Core claim
Theorem 5.9 is the central claim: for a polynomial Cauchy problem with $d=1+r(l+1)$ coordinates, a reduced path $X$ is the $l$-jet extension of a generalized solution if and only if $X(0)\in\Sigma$, equation (7) holds for every admissible index and every word $w$, and equation (3) holds for each defining polynomial $F_i$ and every word $w$. Equation (7) encodes holonomicity—each derivative of a component is matched to the next jet component—and equation (3) encodes containment in the variety $E=\{F_1=\cdots=F_m=0\}$ after translating the path to start at $X(a)$. The force of the theorem is that both families of conditions are linear in the signature entries, so the solution set of the ODE system is cut out by linear equations in signature space.
Load-bearing premise
The converse of the main theorem assumes, without checking, that the starting point of the path already satisfies the polynomial equations defining the ODE system, whereas the signature conditions alone only force those expressions to be constant along the path—so a path starting outside the solution variety can pass all three tests without being a solution.
Editorial extensions
If this is right
- If the characterization is correct, checking whether a reduced path is a generalized solution of a polynomial Cauchy problem reduces to an infinite system of linear equations in the signature entries together with an initial-point membership test.
- Corollary 6.4 spells this out for linear vector fields: $X(t)=(t,f(t),f'(t))$ is an integral curve of $x'=Ax$ with $f(0)=p$ exactly when $X(0)=(0,p,Ap)$ and $\sum_{j=1}^r a_{ij}\langle\sigma(X),(1+j)w\rangle=\langle\sigma(X),(1+r+i)w\rangle$ for all $i$ and all words $w$.
- Corollary 6.6 gives the analogous signature test for integral curves of a Hamiltonian vector field with Hamiltonian $H(x,p)=v\cdot x+\frac12 p^T A p$.
- The variety-containment results (Theorem 3.8 and Corollary 3.9) yield a signature criterion for lying on an algebraic variety up to a constant, and the ODE application is the special case where the variety is the solution set.
- Because the signature determines a path only up to tree-like excursions, the criterion is inherently about generalized solutions modulo harmless back-and-forth reparametrization.
Reading between the lines
- A minimal repair of the stated converse is to add the hypothesis $F_i(X(0))=0$ for every $i$; with that hypothesis, the constants forced by condition (3) are zero and the paper's own Corollary 3.9 yields the missing step.
- The infinite system suggests truncating the equations at words of bounded length as a computational proxy; Remark 3.12 shows this can never be a certificate, so truncated tests should be read as approximate-membership evidence.
- The same mechanism should transfer to non-polynomial, analytic vector fields if an analytic analogue of the signature transformation is available; the paper lists this as future work.
- For stochastic systems, replacing deterministic signatures with rough-path signatures is the natural test bed for extending the criterion to SDEs, an extension the authors mention but do not prove.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper develops algebraic conditions on the signature of a path that are claimed to characterize, up to tree-like excursions, when a path is a solution of a polynomial system of ODEs. Section 3 characterizes paths whose image lies in a polynomial hypersurface or algebraic variety up to level constants; Section 4 characterizes holonomic paths through equations (7); Section 5 combines these ingredients in Theorem 5.9, the main result, which claims necessary and sufficient conditions for a reduced path to be the jet of a solution of a polynomial Cauchy problem; Section 6 applies the result to integral curves of linear and Hamiltonian vector fields. The forward direction of Theorem 5.9 and the holonomic characterization of Section 4 appear plausible, but the converse direction of Theorem 5.9 is false as stated.
Significance. If Theorem 5.9 were correct, it would provide a concrete algebraic test on signature tensors for membership in the solution set of an ODE system and would be a useful complement to the path-variety framework of [Pre23] and the hyperplane results of [GS24]. The geometric proof strategy is natural, and the holonomic characterization in Theorem 4.7 is an interesting and apparently correct contribution that is of independent value. However, the central necessary-and-sufficient claim is invalid: condition (3) of Theorem 5.9 only forces the functions F_i(X(t)) to be constant, not zero, so the converse does not imply containment in the variety E. Because the advertised main theorem fails, the applications in Section 6 inherit the problem. The authors are transparent about the relationship with prior work, and the forward direction and the variety-level results are worth preserving if the statement is repaired.
major comments (4)
- [Section 5.2, Theorem 5.9] The converse direction of Theorem 5.9 is false. In the proof, the claim 'F_i(X(t))=0 for all t by Theorem 3.9' is unjustified: Corollary 3.9 states that condition (3) is equivalent to F_i(X(t)) = c_i for some constants c_i, not to c_i = 0. The zero-level statement is Corollary 3.10, which explicitly requires F_i(X(a)) = 0. Condition (1), X(0) in Sigma, fixes only the first d-r coordinates of X(0), leaving the last r coordinates free, so it does not imply F_i(X(0)) = 0. A concrete counterexample is r = l = 1, F(x1,x2,x3) = x3 - 1, Sigma = {x2 = 0}, and X(t) = (t, 2t, 2) on [0,1]. Then X is reduced, X(0) = (0,0,2) is in Sigma, X is holonomic because \dot X2 = 2 = \dot X1 X3, and condition (3) holds because \tilde F(p1,p2,p3) = p3 and all signatures containing the letter 3 vanish since X3 is constant. Yet F(X(t)) = 1 for all t, so f(t) = 2t is not a solution of x' = 1, x(0) = 0. Thus the advertised necessary-and-sufficient characterization is false as stated; an additional condition such as F_i(X(0)) = 0 for every i is needed.
- [Section 5.2, Theorem 5.9] Even if the missing condition F_i(X(0)) = 0 were added, the converse conclusion as stated is not justified: the paper concludes that f(t) = (X2(t), ..., X_{r+1}(t)) is a generalized solution of (21), but Definition 5.5 defines generalized solutions as paths X in jet space that need not be projectable. A path X satisfying the corrected hypotheses need not give a function f(t) that solves the ODE. For example, with r = l = 1, F = x3 - 1, Sigma = {x2 = 0}, and X(t) = (t^2, t^2, 1) on [0,1], the path is reduced, holonomic, X(0) = (0,0,1) is in Sigma, and im(X) lies in E = {x3 = 1}; hence the corrected conditions hold. But f(t) = t^2 does not solve x' = 1 on [0,1] (f'(t) = 2t). The object that is a generalized solution is the jet path X, and an honest solution is obtained only after reparametrization by x1(t). The final sentence of Theorem 5.9 should be reformulated accordingly.
- [Section 3, Lemma 3.4 and Example 3.7] The proof of Lemma 3.4 contains a substitution that is not valid for the generality claimed: it replaces phi_d(J_{d+1,j}) by 2(j + X_j(a)\emptyset), which is only correct when g(x1,...,xd) is the sum of squares x1^2 + ... + xd^2. For arbitrary g the Jacobian entry is phi_d(\partial g/\partial x_j (x + X(a))), and the subsequent factorizations in the proof of Lemma 3.4, as well as the computation in Example 3.7, do not follow. Since Corollary 3.6 and hence Theorem 3.8 rely on this lemma, the proof of the hypersurface characterization is incomplete as written. The argument can likely be repaired by carrying the general derivative symbol through the half-shuffle identities, but the current text is not correct.
- [Section 6.1, Corollary 6.4] The converse direction of Corollary 6.4 omits the initial-point condition X(0) = (0,p,Ap) (equivalently, F_i(X(0)) = 0). Without it the statement is false: for r = 1 and A = [1], the path X(t) = (t, e^t + 1, e^t) is (1,1)-holonomic and satisfies equation (23) for i = 1 because X2 and X3 have the same derivative, but F(X(t)) = x3 - x2 = -1, so X(0) is not of the form (0,p,Ap) for p = 1 and the associated f(t) = e^t + 1 is not the integral curve with f(0) = 1. The converse needs the initial-value condition as an explicit hypothesis.
minor comments (3)
- [Section 4, Theorem 4.7 proof] The proof concludes X = Y^red from Theorem 2.3 without mentioning the possible translation; since X(a) = Y^red(a), the translation is zero, but this step should be stated explicitly.
- [Section 5.2, proof of Theorem 5.9] The reference to 'Theorem 3.9' in the converse is ambiguous; the statement that would give F_i(X(t)) = 0 is Corollary 3.10, which requires F_i(X(a)) = 0. The numbering should be corrected and the missing hypothesis made explicit.
- [Section 6.2, Corollary 6.6] The displayed equations in Corollary 6.6 contain apparent index errors: the first equation should have a single term -⟨sigma(X), (1+r+i)w⟩ rather than a sum over h of (1+r+h) terms, and the initial point should read X(0) = (0, x0, p0, A p0, -v) rather than containing A x0.
Circularity Check
No circular reduction: the main equivalence chains are independent; the failure in Theorem 5.9's converse is a correctness error, not an input-output identity.
full rationale
No circular step is present. Theorem 3.8 invokes the hyperplane characterization via '[GS24, Proposition 6.3] or [Pre23, Lemma 3.1]'; although [GS24] is by two of the present authors, [Pre23] independently supplies the same statement, so the self-citation is redundant rather than load-bearing. Theorem 4.7 is proved internally from Chen's uniqueness (Theorem 2.3), the shuffle identity, and a direct construction of the auxiliary path Y, rather than imported from the authors' prior work. Theorem 5.9 assembles Theorem 3.9, Theorem 4.7, and the jet-space dictionary Theorem 5.7; no parameter is fitted and no predicted quantity is renamed as an input. There is, however, a genuine mathematical overclaim in the converse of Theorem 5.9: the proof says 'F_i(X(t)) = 0 for all t and for every i by Theorem 3.9', but Theorem 3.9/Corollary 3.9 only forces each F_i(X(t)) to be constant; the constant is F_i(X(0)), not necessarily 0, and condition (3) does not impose F_i(X(0))=0. A witness is X(t)=(t,2t,2) for F(x1,x2,x3)=x3-1 and x'=1, x(0)=0, which satisfies (1)-(3) but is not a solution. This is a correctness flaw, not circular reasoning: the necessary direction and the holonomic characterization are unaffected. The score 2 reflects only the minor, non-load-bearing self-citation to [GS24].
Assumptions & free parameters
assumptions (6)
- standard math Chen's uniqueness theorem: two reduced paths with the same signature coincide up to translation.
- standard math Signature transformation under polynomial maps: for a polynomial map p, M_ep is an algebra and half-shuffle homomorphism satisfying <sigma(p circ X), v> = <sigma(X), M_ep(v)>.
- standard math Hyperplane characterization: a path's image is contained in an affine hyperplane iff signature entries for all words containing the normal coordinate vanish.
- standard math Jet-space correspondence: for a system of polynomial ODEs, solutions correspond to jet extensions X satisfying E, as recalled in Theorem 5.2.
- domain assumption Path class is continuous and piecewise differentiable, with finite signature; reducedness is assumed for uniqueness.
- domain assumption The ODE systems in Theorem 5.9 must be polynomial; analytic extensions are left open in Section 7.
Cite this review
Pith. "Pith review of Path signatures of ODE solutions." pith.science (2026). https://pith.science/paper/FHRMTT5F
@misc{pith2026250513234,
author = {Pith},
title = {Pith review of: Path signatures of ODE solutions},
year = {2026},
howpublished = {\url{https://pith.science/paper/FHRMTT5F}},
note = {Machine review of arXiv:2505.13234}
}
read the original abstract
The signature of a path is a sequence of tensors which allows to uniquely reconstruct the path. By employing the geometric theory of nonlinear systems of ordinary differential equations, we find necessary and sufficient algebraic conditions on the signature tensors of a path to be a solution of a given system of ODEs. As an application, we describe in detail the systems of ODEs that describe the trajectories of a vector field, in particular a linear and Hamiltonian one.
Figures
Reference graph
Works this paper leans on
-
[1]
C. Am \'e ndola, P. Friz, and B. Sturmfels. Varieties of signature tensors. In Forum of Mathematics, Sigma , volume 7. Cambridge University Press, 2019
work page 2019
-
[2]
D.V. Alekseevskii, R.V. Gamkrelidze, V.V. Lychagin, and A.M. Vinogradov. Geometry I: Basic Ideas and Concepts of Differential Geometry . Encyclopaedia of mathematical sciences. Springer-Verlag, 1991
work page 1991
-
[3]
A. V. Bocharov, V. N. Chetverikov, S. V. Duzhin, N. G. Khor kova, I. S. Krasil shchik, A. V. Samokhin, Yu. N. Torkhov, A. M. Verbovetsky, and A. M. Vinogradov. Symmetries and conservation laws for differential equations of mathematical physics , volume 182 of Translations of Mathematical Monographs . American Mathematical Society, Providence, RI, 1999. Ed...
work page 1999
-
[4]
Boedihardjo, X
H. Boedihardjo, X. Geng, T. Lyons, and D. Yang. The signature of a rough path: uniqueness. Adv. Math. , 293:720--737, 2016
2016
-
[5]
F. Brown. Iterated integrals in quantum field theory. In Geometric and topological methods for quantum field theory , pages 188--240. Cambridge Univ. Press, Cambridge, 2013
work page 2013
-
[6]
T. Cochrane, P. Foster, V. Chhabra, M. Lemercier, T. Lyons, and C. Salvi. Sk-tree: a systematic malware detection algorithm on streaming trees via the signature kernel. In 2021 IEEE International Conference on Cyber Security and Resilience (CSR) , pages 35--40, 2021
work page 2021
-
[7]
K.-T. Chen. Iterated integrals and exponential homomorphisms. Proc. London Math. Soc. (3) , 4:502--512, 1954
work page 1954
-
[8]
K.-T. Chen. Integration of paths---a faithful representation of paths by non-commutative formal power series. Trans. Amer. Math. Soc. , 89:395--407, 1958
work page 1958
Show all 22 references
-
[9]
Chevyrev, V
I. Chevyrev, V. Nanda, and H. Oberhauser. Persistence paths and signature features in topological data analysis. IEEE Transactions on Pattern Analysis and Machine Intelligence , 42(1):192--202, 2020
2020
-
[10]
Colmenarejo and R
L. Colmenarejo and R. Preiß. Signatures of paths transformed by polynomial maps. Beitrage zur Algebra und Geometrie , 61(4):695 – 717, 2020
2020
-
[11]
Eshkobilov, G
O. Eshkobilov, G. Manno, G. Moreno, and K. Sagerschnig. Contact manifolds, L agrangian G rassmannians and PDE s. Complex Manifolds , 5(1):26--88, 2018
2018
-
[12]
Foissy and F
L. Foissy and F. Patras. Natural endomorphisms of shuffle algebras. International Journal of Algebra and Computation , 23(04):989--1009, 2013
2013
-
[13]
Galuppi and P
F. Galuppi and P. Santarsiero. Rank and symmetries of signature tensors. preprint arXiv:2407.20405 , 2024
2024 arXiv
-
[14]
Hambly and T
B. Hambly and T. Lyons. Uniqueness for the signature of a path of bounded variation and the reduced path group. Ann. of Math. (2) , 171(1):109--167, 2010
2010
-
[15]
Kol \'a r , P
I. Kol \'a r , P. W. Michor, and J. Slov \'a k. Natural operations in differential geometry . Springer-Verlag, Berlin, 1993
1993
-
[16]
Lyons and A
T. Lyons and A. McLeod. Signature methods in machine learning. preprint arXiv:2206.14674 , 2024
2024 arXiv
-
[17]
T. Lyons. Differential equations driven by rough signals. Rev. Mat. Iberoamericana , 14(2):215--310, 1998
1998
-
[18]
T. Lyons. Rough paths, signatures and the modelling of functions on streams. In Proceedings of the International Congress of Mathematicians , volume 4 of Example Conference Proceedings , pages 163--184, Seoul, June 2014. Kyung Moon Publishers
2014
-
[19]
Morrill, A
J. Morrill, A. Kormilitzin, A. Nevado-Holgado, S. Swaminathan, S. Howison, and T. Lyons. The signature-based model for early detection of sepsis from electronic health records in the intensive care unit. In 2019 Computing in Cardiology (CinC) , pages Page 1--Page 4, 2019
2019
-
[20]
R. Prei . An algebraic geometry of paths via the iterated-integrals signature. preprint arXiv:2311.17886 , 2023
2023 arXiv
-
[21]
R. Preiß. An algebraic geometry of paths via the iterated-integrals signature. Nonlinear algebra seminar, 2024
2024
-
[22]
Reutenauer
C. Reutenauer. Free L ie algebras , volume 7 of London Mathematical Society Monographs. New Series . The Clarendon Press, Oxford University Press, New York, 1993. Oxford Science Publications
1993
Reviewed August 15, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.