Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

Learning Barycenters from Signature Matrices

T0 review · 3 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 At matrix truncation level, this paper determines the exact minimal segment count for recovering a path from the barycenter of path signatures, and shows every group element is a barycenter of a sample with all but one element fixed.

desk verdict Solid signature-recovery paper with a real proof gap in the odd-case simultaneous normalization; the formula likely holds, but the proof needs a patch before publication. read the letter →

arxiv 2509.07815 v1 pith:IOBQSOXA submitted 2025-09-09 math.RA math.AGmath.STstat.TH

Carlos Améndola , Leonard Schmitz This is my paper · ORCID
classification math.RAmath.AGmath.STstat.TH MSC 60L1022E2515A2114Q15
keywords Liegroupbarycentersfreenilpotentgroupsiterated-integralsignaturespiecewiselinearpathspathrecoverymatrixcongruencenormalformsnon-commutativepolynomialssignaturematrices
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 asks how to average a collection of paths so that the average is still a path signature. The standard expected signature of a random path generally leaves the signature space, so the authors work with the Lie-group barycenter, a group-valued mean, and study when that barycenter is the signature of a piecewise linear path and with how few segments. They prove that every element of the free nilpotent Lie group is a barycenter of some N-tuple of group elements, with any N−1 of them fixed in advance, and they give a closed-form expression for the barycenter at matrix truncation level. The central result is an exact formula for the minimal segment count: for segment counts α, B_{d,2}(α) = min(d, α_+) if every α_i is even, and B_{d,2}(α) = min(d, α_+ − #odd_α + 1) otherwise. If correct, this makes barycenter recovery constructive and low-dimensional: a matrix rank decides the segment count, and explicit matrix congruence transformations recover a realizing path.

What carries the argument

The load-bearing object is W_α, the matrix-level (quadratic) projection of the barycenter of the sample signatures after reducing to axis subpaths. It is a block-direct sum of skew-symmetric matrices U_{α_i} − U_{α_i}^T, where U_{α_i} is the strictly upper triangular matrix with ones above the diagonal, plus a rank-one all-ones term. The argument turns on the congruence normal form of W_α: it is congruent to Γ2⊕H2(−1)^{⊕...} when all α_i are even, and to 1⊕H2(−1)^{⊕...}⊕0^{⊕...} otherwise, so its rank is m in the first case and m−#odd_α+1 in the second. The proof uses these normal forms to construct explicit matrices P that send W_α to C_r = 1/2 I_r + U_r, the signature matrix of the r-segme

What would settle it

Take d=4 and α=(2,1), where the formula predicts B_{4,2}(2,1)=3. Let X_1 be the two-segment axis path in R^4 and X_2 a one-segment path along e_1, compute the quadratic barycenter with the paper's closed form (Theorem 4.11), and write its matrix projection W_α. Then apply the explicit congruence transform P constructed in the alternative proof of Theorem 7.10 and check whether P W_α P^T = C_3 and P(1_3/2) = e_1 simultaneously. If no simultaneous transform exists, the true minimal count is 4 and the formula fails; the check is a finite symbolic computation.

Watch

Extended reading notes

Core claim

On its own terms, the paper's claim is twofold. First, the barycenter map on the free nilpotent Lie group is an epimorphism of algebraic varieties: every group element occurs as the barycenter of some sample, and fixing N−1 of the N sample elements makes the map an isomorphism (Theorem 4.7). Second, at quadratic truncation the minimal number of segments needed to realize the barycenter of signatures of piecewise linear paths is exactly the value given in Theorem 7.2. The proof exhibits the matrix projection of the barycenter, W_α = (1/2N) ⊕_i (U_{α_i} − U_{α_i}^T) + (1/2N^2) 1_{m×m}, computes its rank and congruence normal form, and gives explicit transformations sending W_α to the signature

Load-bearing premise

The exact formula rests on the premise that one and the same linear change of coordinates can simultaneously put the matrix part of the averaged signature into its standard normal form and align its vector part with the first axis; the paper asserts this compatibility in a single sentence without a detailed verification. If that premise fails, the theorem's value would be only a lower bound on the true segment count.

Editorial extensions

If this is right

  • For k=2, barycenter recovery becomes an exact, constructive procedure: the rank of an explicitly written matrix gives the number of segments, and the paper's explicit congruence transformations produce a realizing path.
  • The barycenter of any number of one-segment path signatures is itself a one-segment signature, so averaging straight-line paths in signature space preserves straight-line structure.
  • No quadratic signature barycenter ever needs more than d segments, so in fixed dimension the barycenter of paths with many segments has a bounded-complexity representation.
  • The cubic case already differs: the barycenter of two one-segment signatures in R^2 needs three segments (B_{d,3}(1,1)=3), so segment-count behavior depends essentially on truncation level.

Reading between the lines

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

  • Editorial inference: the rank formula for W_α points to a plausible higher-level analogue in which B_{d,k}(α) would be governed by ranks of higher-order tensor congruence orbits; that analogue is not established here and could be tested computationally at k=3 for simple α.
  • Editorial inference: the compatibility issue flagged by the one-sentence vector–matrix normalization suggests that for truncation level k≥3, simultaneous normalization across tensor levels will become a genuine obstruction; the cubic two-segment example B_{d,3}(1,1)=3 may be the first instance of such behavior.
  • Editorial inference: because the authors note that all results lift to weighted distributions, the explicit quadratic barycenter formula can be turned into a practical weighted-average path estimator for ensembles of piecewise linear data streams, a step the paper does not develop.
  • Editorial inference: a randomized numerical check—draw many α_i-segment paths, compute W_α, and compare its rank with the minimal segment count found by solving the recovery system—would localize any failure of the compatibility assumption to particular α.
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.

Referee Report

3 major / 5 minor

Summary. The paper studies the Buser–Karcher Lie group barycenter on free nilpotent Lie groups of path signatures. It proves that the barycenter map is an epimorphism of algebraic varieties (Theorem 4.7), is a non-commutative polynomial map (Theorem 4.9), and gives a closed-form expression at quadratic truncation (Theorem 4.11). For piecewise linear paths, it defines the barycenter recovery order B_{d,k}(α) and, for k=2, claims an exact formula B_{d,2}(α)=min(d, α+) if all α_i are even, and min(d, α+−#odd_α+1) otherwise (Theorem 7.2). The proof proceeds via explicit congruence normal forms of the signature matrix. The paper also provides an OSCAR implementation and several worked examples.

Significance. The main contribution is the exact recovery order for quadratic truncation and the explicit transformation to congruence normal forms, connecting signature varieties with classical matrix congruence theory. If the proof gap identified below is repaired, this is a valuable and citable result. The paper ships executable code and several examples that verify special cases, which strengthens reproducibility. The clean statement of Theorem 7.2 will likely be useful in further work on path recovery and signature geometry.

major comments (3)
  1. [§7, proof of Theorem 7.2 and Lemma 7.6] The simultaneous vector–matrix normalization is asserted without proof and is false for the displayed transform in Lemma 7.6. The proof states "since the matrix from Theorem 7.6 transforms 1_{α1} to e1", but the explicit P in Lemma 7.6 (with T_{12}) sends 1_m to √2 e_2 for m≥2, e.g. m=3. Example 7.13 instead uses a different, untwisted transform that sends 1_5 to e_1. The upper bound B ≤ rk(W_α) depends on this compatibility; without it, Theorem 7.2 would give only a lower bound. Please state and prove a lemma that the congruence normal-form transform can be chosen to also map 1_m to the prescribed vector (or prove the consistency of the two-step inverted transform).
  2. [§7, proof of Theorem 7.2] The proof concludes B_{d,2}(α)=rk(W_α), but the theorem states min(d, rk(W_α)). For α=(2,2) and d=2, one has rk(W_α)=4 while the theorem gives B=2. The proof does not address the dimensional cap d. A separate argument is needed for the case rk(W_α)>d: show that recovery with d segments is possible (e.g. by the general upper bound B≤d from [1, Theorem 3.4]) and that fewer than min(d,rk) segments are impossible, e.g. via the rank of the skew-symmetric part.
  3. [§7, Lemma 7.6] The proof of Lemma 7.6 says "by construction" and does not verify that the displayed P realizes the stated normal form, nor does it reconcile the block ordering with the T_{12} permutation. Since Theorem 7.2 relies on the vector action of this same matrix, the explicit expression should be checked and the version used in Example 7.13 (without T_{12}) should be reconciled with the lemma as stated.
minor comments (5)
  1. [§7, Lemma 7.6] The notation T_{m+1}^{12} is undefined; presumably T_{12}. Also check the indices in the expression E_{j,m-2(i-1)-1}.
  2. [§7, proof of Theorem 7.2] The proof writes B_{d,k}(α) where the theorem concerns B_{d,2}(α); correct the symbol.
  3. [§8, Example 8.4] Example 8.4 refers to "Theorem 6.2", but the statement being verified is Example 6.2, not a theorem.
  4. [§4, before Theorem 4.11] The text says 'In Theorem 2.2 we recall...' but the relevant statement is Example 2.2.
  5. [§7, Example 7.13] The final transformation matrix is very large and hard to verify visually; consider splitting the computation or providing the corresponding OSCAR output in the repository.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the recovery formula is derived from the defining barycenter equation and external congruence normal forms; self-citations are published external results. One unverified vector-transform assertion is a correctness gap, not circularity.

full rationale

The paper's derivation chain is not circular. B_{d,2}(α) is defined (Def. 6.3) as a minimal segment count, and Theorem 7.2 computes it from the closed-form matrix projection W_α (Lemma 7.7) together with external congruence normal forms ([15], [12]) and the cap d from [1]; no fitted parameter or target quantity is reused as an input. The barycenter is fixed by the Buser–Karcher equation (10). Theorem 4.2 cites [8] for existence/uniqueness: although [8] shares an author, it is a peer-reviewed, parameter-free result whose assumptions do not include Theorem 7.2, so it is independent support and does not raise the circularity score. The same holds for [1, Thm. 3.4]. Flagged for correctness, not circularity: the proof of Theorem 7.2 asserts 'since the matrix from Theorem 7.6 transforms 1_{α1} to e1' without verification; Lemma 7.6's explicit P actually sends 1_m to √2 e2 for m≥2, so the simultaneous vector/matrix normalization is an omitted proof and a possible gap, but it is not a reduction of the conclusion to an input. Score 2 only for the presence of minor self-citations.

Assumptions & free parameters 0 free parameters · 6 assumptions · 0 invented entities

The central claim is derived from the defining barycenter equation plus established signature and matrix congruence theorems. No free parameters are fitted and no new mathematical entities, forces, or dimensions are introduced; W_α and B_{d,k}(α) are definitions, not postulated entities.

assumptions (6)
  • standard math The barycenter equation (10) has a unique solution in G_{d,k} for every sample (Theorem 4.2, from [8]).
    Used throughout Sections 4-7 as the definition and existence guarantee of the barycenter; not reproven in this paper.
  • standard math The exponential and logarithm are isomorphisms of affine algebraic varieties between g_{d,k} and G_{d,k}, characterized by shuffle relations.
    Invoked in Sections 2-4 for the algebraic-geometric treatment and non-commutative polynomial representation; cited from [26] and [1].
  • standard math Chen-Chow: every element of G_{d,k} is the signature of some piecewise linear path.
    Theorem 6.6, used to prove well-definedness of B_{d,k}(α) in Proposition 6.5.
  • standard math Real matrix congruence classification: nonsingular real matrices A, B are congruent iff their cosquares A^{-⊤}A and B^{-⊤}B are similar; real congruence transformations then exist.
    Used in Proposition 7.10 to identify congruence normal forms of W_α and C_m; cited [15, Lemma 2.1] and [12, Theorem 1.1].
  • standard math The log-signature of a straight-line segment is nonzero only at level 1.
    Used in Proposition 7.1 to show the barycenter of segment signatures remains a straight segment; cited [13] and [2].
  • domain assumption The path class considered is exactly piecewise linear paths with a fixed composition α of segment counts.
    The recovery order B_{d,k}(α) is defined relative to this class; if the class were enlarged (for example, smooth paths or time-rescaled paths), the minimal segment count could change.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning Barycenters from Signature Matrices." pith.science (2026). https://pith.science/paper/IOBQSOXA

@misc{pith2026250907815,
  author       = {Pith},
  title        = {Pith review of: Learning Barycenters from Signature Matrices},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IOBQSOXA}},
  note         = {Machine review of arXiv:2509.07815}
}
read the original abstract

The expected signature of a family of paths need not be a signature of a path itself. Motivated by this, we consider the notion of a Lie group barycenter introduced by Buser and Karcher to propose a barycenter on path signatures. We show that every element of the free nilpotent Lie group is a barycenter of a group sample, where all but one sample element can be fixed arbitrarily. In the case of piecewise linear paths, we study the problem of recovering an underlying path corresponding to the barycenter of signatures. We determine the minimal number of segments required to learn from signature matrices, providing explicit transformations to the associated congruence normal forms.

Figures

Figures reproduced from arXiv: 2509.07815 by the authors.

Figure 1
Figure 1. Note that for d = 2 the signature stores the distance and the signed area of a path, see e.g. [7]. We provide a lower bound of B2,2(1, 1) in Theorem 6.2. The explicit verifications of B2,k from [PITH_FULL_IMAGE:figures/full_fig_p012_1.png] view at source ↗
Figure 2
Figure 2. All axis subpaths Axα,i of Axis3 with respect to α ∈ C3. Proof of Theorem 6.5. We provide an upper bound of Bd,k(α). Let m := α1 + · · · + αN . By definition, there exists for every zi ∈ Lim d,≤k,αi an Ai ∈ R d×αi such that zi = σ(Ai · Axisαi ). We align these transformations column-wise, A := A1 . . . AN [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗
Figure 3
Figure 3. Recovered paths Y parametrized by ω ∈ R \ {0} for bary (σ(X1), σ(X2)) in truncation level k = 3 according to Theorems 8.6 and 8.7. which coincide using Gauß’s area formula [17, p. 125], 1 2 [PITH_FULL_IMAGE:figures/full_fig_p025_3.png] 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

  1. An Efficient Algorithm for Path Recovery from Signature Tensors

    math.RA 2025-12 conditional novelty 7.1 of 10

    An exact algorithm recovers the linear map A from its third-order signature tensor G = A*C in expected O(d^4) time, replacing the Gröbner-basis method that previously solved this inverse problem.

Reference graph

Works this paper leans on

27 extracted references · 25 canonical work pages · cited by 1 Pith paper

  1. [1]

    Améndola, P

    C. Améndola, P. Friz, and B. Sturmfels , Varieties of signature tensors , Forum of Mathematics, Sigma, 7 (2019)

    work page 2019

  2. [8]

    Clausel, J

    M. Clausel, J. Diehl, R. Mignot, L. Schmitz, N. Sugiura, and K. Usevich , The Barycenter in Free Nilpotent Lie Groups and Its Application to Iterated-Integrals Signatures , SIAM Journal on Applied Algebra and Geometry, 8 (2024), pp. 519--552

    work page 2024

  3. [2]

    Améndola, F

    C. Améndola, F. Galuppi, Ángel David Ríos Ortiz , P. Santarsiero, and T. Seynnaeve , Decomposing tensor spaces via path signatures , Journal of Pure and Applied Algebra, 229 (2025), p. 107807

    work page 2025

  4. [3]

    Améndola, A

    C. Améndola, A. E. Saliby, F. Lotter, and O. R. Fité , Computing Path Signature Varieties in Macaulay2 , 2025. https://arxiv.org/abs/2506.01429

    arXiv 2025

  5. [4]

    Bayer, P

    C. Bayer, P. P. Hager, S. Riedel, and J. Schoenmakers , Optimal stopping with signatures , The Annals of Applied Probability, 33 (2023), pp. 238--273

    work page 2023

  6. [5]

    Bonnier and H

    P. Bonnier and H. Oberhauser , Proper Scoring Rules, Gradients, Divergences, and Entropies for Paths and Time Series , Bayesian Analysis, (2024), pp. 1 -- 32

    work page 2024

  7. [6]

    Buser and H

    P. Buser and H. Karcher , Gromov's almost flat manifolds , Ast\'erisque, MR 83m, 53070 (1981)

    work page 1981

  8. [7]

    Chevyrev and A

    I. Chevyrev and A. Kormilitzin , A primer on the signature method in machine learning , 2025. https://arxiv.org/abs/1603.03788

    arXiv 2025

Show all 27 references
  1. [9]

    Cuchiero, G

    C. Cuchiero, G. Gazzani, and S. Svaluto-Ferro , Signature-based models: Theory and calibration , SIAM journal on financial mathematics, 14 (2023), pp. 910--957

    2023

  2. [10]

    Decker, C

    W. Decker, C. Eder, C. Fieker, M. Horn, and M. Joswig , eds., The C omputer A lgebra S ystem OSCAR : A lgorithms and E xamples , vol. 32 of Algorithms and C omputation in M athematics, Springer, 1 ed., 8 2024

    2024

  3. [11]

    Diehl and R

    J. Diehl and R. Krieg , FRUITS: feature extraction using iterated sums for time series classification , Data Mining and Knowledge Discovery, 38 (2024), pp. 4122--4156

    2024

  4. [12]

    D. Z. Dokovic and K. D. Ikramov , On the congruence of square real matrices , Linear Algebra and its Applications, 353 (2002), pp. 149--158

    2002

  5. [13]

    P. K. Friz, T. Lyons, and A. Seigal , Rectifiable paths with polynomial log-signature are straight lines , Bulletin of the London Mathematical Society, (2024)

    2024

  6. [14]

    P. K. Friz and N. B. Victoir , Multidimensional Stochastic Processes as Rough Paths: Theory and Applications , Cambridge Studies in Advanced Mathematics, Cambridge University Press, 2010

    2010

  7. [15]

    R. A. Horn and V. V. Sergeichuk , Canonical forms for complex matrix congruence and *congruence , Linear Algebra and its Applications, 416 (2006), pp. 1010--1032

    2006

  8. [16]

    Kidger and T

    P. Kidger and T. Lyons , S ignatory: differentiable computations of the signature and logsignature transforms, on both CPU and GPU , in International Conference on Learning Representations, 2021. https://github.com/patrick-kidger/signatory

    2021

  9. [17]

    Koecher and A

    M. Koecher and A. Krieg , Ebene Geometrie , Springer, 2007

    2007

  10. [18]

    Laaksonen , Guide to competitive programming , Springer, 2020

    A. Laaksonen , Guide to competitive programming , Springer, 2020

    2020

  11. [19]

    Moakher , Means and averaging in the group of rotations , SIAM journal on matrix analysis and applications, 24 (2002), pp

    M. Moakher , Means and averaging in the group of rotations , SIAM journal on matrix analysis and applications, 24 (2002), pp. 1--16

    2002

  12. [20]

    Ni , The expected signature of a stochastic process , PhD thesis, Oxford University, UK, 2012

    H. Ni , The expected signature of a stochastic process , PhD thesis, Oxford University, UK, 2012

    2012

  13. [21]

    https://www.oscar-system.org

    OSCAR -- Open Source Computer Algebra Research system , 2024. https://www.oscar-system.org

    2024

  14. [22]

    Pennec and M

    X. Pennec and M. Lorenzi , Beyond Riemannian geometry: The affine connection setting for transformation groups , in Riemannian Geometric Statistics in Medical Image Analysis, Academic Press, 2020, pp. 169--229

    2020

  15. [23]

    K. B. Petersen, M. S. Pedersen, et al. , The matrix cookbook , Technical University of Denmark, 7 (2008), p. 510

    2008

  16. [24]

    Pfeffer, A

    M. Pfeffer, A. Seigal, and B. Sturmfels , Learning Paths from Signature Tensors , SIAM Journal on Matrix Analysis and Applications, (2019)

    2019

  17. [25]

    J. F. Reizenstein and B. Graham , Algorithm 1004: The Iisignature Library: Efficient Calculation of Iterated-Integral Signatures and Log Signatures , ACM Trans. Math. Softw., 46 (2020)

    2020

  18. [26]

    Reutenauer , Free Lie algebras , Oxford University press, 1993

    C. Reutenauer , Free Lie algebras , Oxford University press, 1993

    1993

  19. [27]

    Salvi, T

    C. Salvi, T. Cass, J. Foster, T. Lyons, and W. Yang , The Signature Kernel is the Solution of a Goursat PDE , SIAM Journal on Mathematics of Data Science, 3 (2021), pp. 873--899

    2021

Pith tools

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