Pith. sign in

REVIEW 3 major objections 4 minor 43 references

Signature Reconstruction from Randomized Signatures

T0 review · 3 major / 4 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read Depth-two randomized signatures let the signature features of a path be reconstructed from the flow of a random controlled differential equation, with the number of recoverable features exponential in the hidden dimension.

desk verdict The algebraic core on tree-like vector fields is worth knowing, but the advertised reconstruction theorem rests on a scaling identity that is false as written. read the letter →

arxiv 2502.03163 v1 pith:HNSBS3RZ submitted 2025-02-05 math.CA cs.LGmath.PRstat.ML

Mie Glückstad , Nicola Muca Cirone , Josef Teichmann This is my paper · ORCID
classification math.CAcs.LGmath.PRstat.ML MSC 60L1017B6668T07
keywords randomizedsignaturescontrolleddifferentialequationssignaturereconstructiontree-likevectorfieldslinearindependencedepth-twoneuralexponentialactivationhiddendimension
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

The paper is after a proof of a conjecture about randomized signatures: that the signature features of a path, which are dense and useful in machine learning, can be recovered from the outputs of a controlled differential equation with random neural vector fields. The authors show this is possible up to order $L$ when the hidden dimension satisfies $N \ge L-1$ and the vector fields are depth-two neural networks with exponential activation. The proof works by expanding the solution in a Taylor series whose coefficients are iterated vector fields, organizing those fields into tree-like components, and proving linear independence for the exponential family. The payoff is the exponential growth of reconstructible features in the hidden dimension, and a general linear-independence criterion that applies to arbitrary vector fields.

What carries the argument

The load-bearing object is the collection of tree-like vector fields $V_\tau$, indexed by letter-labelled recursive trees, which organize the iterated derivatives appearing in the Taylor expansion of $g(Y_t)$. Linear independence of these $V_\tau$ up to order $L$ is shown for $V_i(x)=\exp(A_i\exp(D_i x))$ with algebraically independent random coefficients when $N\ge m-1$, and this independence turns the $m$-th $r$-derivative of the scaled solution into a uniquely solvable linear system for the signature components of order $m$.

What would settle it

A direct calculation settles the proof's core step: take $N=1$, $V(x)=x^2$, and $X_t=t$. The scaled equation $\dot Y^{\eta,r}=r(Y^{\eta,r})^2$ has explicit solution $Y^{\eta,r}_t=\eta/(1-r\eta t)$, while the asserted identity $Y^{\eta,r}_t=rY^{\eta/r}_t$ gives a different function of $r$; consequently the $m$-th $r$-derivative at $r=0$ computed from the explicit solution differs from the value the proof's formula assigns, so Theorem 4.1's reconstruction mechanism fails at this step for this nonlinear vector field.

Watch

Extended reading notes

Core claim

The paper's central claim is that the signature of a bounded-variation path can be read off from the flow of a randomly initialized controlled differential equation. Under a linear-independence condition on the tree-like vector fields generated by $V_1,\dots,V_d$, Theorem 4.1 reconstructs all signature components of order $m\le L$ by differentiating the $r$-scaled solutions at $r=0$ and solving a linear system whose coefficient matrix is invertible exactly when the tree-like fields are independent. The constructive example is the depth-two exponential field $V_i(x)=\exp(A_i\exp(D_i x))$ with algebraically independent random coefficients, where $N\ge m-1$ is shown to guarantee the required independence; hence $N$ can grow linearly while the number $d^m$ of recovered features grows exponentially.

Load-bearing premise

The argument requires that scaling the vector fields by $r$ merely rescales the initial value of an unscaled solution, an identity that is exact for linear vector fields; for the nonlinear fields the paper uses, this scaling relation is asserted rather than established, and the exponential vector fields are also assumed to be globally bounded so that all solutions exist.

Editorial extensions

If this is right

  • For any smooth vector fields whose tree-like fields are linearly independent up to order $L$, the signature components of $X$ up to order $L$ are uniquely determined by the family of terminal solutions $(Y^y_T)_{y\in\mathbb{R}^N}$.
  • For the depth-two exponential randomized signature with $N\ge L-1$, the reconstruction holds almost surely under i.i.d. absolutely continuous random initialization.
  • The number of reconstructible features of order $m$ is $d^m$, so the hidden dimension needs to grow only logarithmically in the number of recovered features.
  • The classical one-layer randomized signature $V_i(x)=\sigma(A_i x)$ does not admit the required independence: ladder trees produce proportional tree-like vector fields, so depth is essential for the result.
  • On a Lie group $G$, the same Taylor-expansion scheme is outlined, contingent on an analogous linear-independence statement for vector fields on $G$.

Reading between the lines

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

  • Beyond the theorem, an untrained depth-two exponential reservoir could serve as a fixed nonlinear featurizer whose features provably span the same space as signature features up to order $L$, at a state cost that grows linearly in $L$; this is an editorial projection, not a claim of the paper.
  • The linear-independence threshold $N\ge m-1$ suggests a sharp capacity transition: below roughly $m$ hidden units, word-algebra relations force dependencies no matter how the vector fields are chosen, so random feature richness is bounded; this is an extrapolation of Remark 3.7.
  • The proof mechanism could be tested by replacing $\exp$ with other activations, since the exponential case is what makes the factorisation $\sigma^{(n)}=\sigma$ clean; a generic analytic activation would require reworking the independence argument for its Taylor coefficients.
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 / 4 minor

Summary. The paper studies reconstruction of the signature components of a bounded-variation path X from the family of terminal values (Y^y)_{y∈R^N} of a controlled differential equation Y_t = y + Σ_i ∫_0^t V_i(Y_s) dX^i_s, where the vector fields are random neural-network type fields. It develops a recursive-tree representation of iterated vector fields, proves a linear-independence criterion for those tree-like vector fields, and constructs explicit depth-two exponential (nested-exponential) vector fields satisfying the criterion when the hidden dimension satisfies N ≥ m−1. On this basis it claims in Theorem 4.1 that the signature components up to order L can be uniquely reconstructed from the unscaled family (Y^y)_{y∈R^N}; Theorem 4.2 states an analogous Lie-group reconstruction from the r-scaled family (Z^{ζ,r})_{ζ∈G,r∈R}. The abstract concludes that the number of reconstructible signature features is exponential in the hidden dimension for neural depth-two vector fields.

Significance. If Theorem 4.1 were valid, the paper would provide a useful quantitative complement to known Lie-algebra independence results and a theoretical justification for randomized signatures in reservoir computing. The tree calculus of Section 3, especially Lemma 3.4 and Proposition 3.6, is a natural and potentially valuable contribution, and the nested-exponential construction in Section 3.3 is an interesting explicit example of linearly independent iterated vector fields. However, the proof of the central reconstruction theorem rests on an elementary false scaling identity, and the advertised reconstruction claim is therefore not established. The Lie-group theorem is also explicitly only sketched. Because the main result fails at its decisive step, the paper in its present form does not support its headline claim.

major comments (3)
  1. [§4.1, proof of Theorem 4.1] The proof introduces the r-scaled CDE and asserts that for every η ∈ R^N and r ≠ 0 one has Y^{η,r}_t = rY^{η/r}_t. This identity is false for nonlinear vector fields and already fails for linear fields. If Z_t := rY^{η/r}_t, then dZ_t = Σ_i rV_i(Z_t/r)dX^i_t, whereas the r-scaled equation for Y^{η,r} has vector fields rV_i(Z_t). Equality would require V_i(Z/r)=V_i(Z) for all r, which is a degree-one homogeneity condition not satisfied by general smooth fields. For V(x)=x and X_t=t, the true solution is Y^{η,r}_t = ηe^{rt}, while the claimed expression gives rY^{η/r}_t = ηe^t. Consequently the quantity d^m/dr^m g(Y^{η,r}_t)|_{r=0}, which the proof equates with Σ_{w∈W_m} V_wg(η)∫_{∆^m}dX^w, is not computable from the collection (Y^y)_{y∈R^N} of solutions of the unscaled equation (4.1.1). The linear system from which the signature components are to be solved therefore involves data not available under the theorem's hypothesis. The central claim of Theorem 4.1 is unsupported by the proof. Note that Theorem 4.2 explicitly includes r in its data, which is what the r-derivative argument actually requires.
  2. [§4.2, Theorem 4.2] Theorem 4.2 is not proven. The proof relies on a 'similar claim of linear independence' of the operators {V_w} on the Lie group, and the paper explicitly says that this 'should be provable in a similar fashion' and that the details will not be given. Since this linear-independence assertion is the load-bearing step of the reconstruction argument, the theorem is at best a conjecture. Additionally, the Taylor expansion on Lie groups (Theorem 2.4) is stated without proof, and Remark 2.5 does not address the differentiability-in-r and global-existence issues needed for the derivative step.
  3. [§3.3.2 and §4.1, application to nested exponentials] The advertised application uses vector fields V_i(x)=exp(A_i exp(D_i x)), which are smooth but not globally Lipschitz and not globally bounded. Theorem 4.1 assumes only that V_1,…,V_d are smooth, and the paper does not state or prove a global existence hypothesis. For an arbitrary bounded-variation driver X, the CDE solution may fail to exist on the whole interval [0,T] for some initial values, so the collection (Y^y)_{y∈R^N} is not a priori well defined. The proof also needs the r-scaled solutions to be sufficiently differentiable in r in a neighborhood of r=0 and defined up to time T; this requires global existence and is not addressed.
minor comments (4)
  1. [§4.2, proof of Theorem 4.2] In the displayed equation after taking the m-th derivative, the iterated integral is written as ∫_{∆^k} dX^w_r even though the summation is over w ∈ W_m; the index should be m.
  2. [§2.2, Remarks 2.3 and 2.5] The remarks say convergence of the Taylor expansion is not needed because remainder terms disappear after differentiating in r and evaluating at r=0, but differentiability of r ↦ Y^{η,r}_t in a neighborhood of r=0 and existence of the solution up to time T are hypotheses that are neither stated nor proved.
  3. [§3.2, proof of Proposition 3.6] The proof separates first-order tree-like operators from higher-order ones using linear independence of differential operators of different orders, but this step is asserted rather than justified; a short justification by applying the operators to coordinate functions and monomials would make the proof self-contained.
  4. [§3.3.2] The key linear-independence argument for the exponential functions with algebraically independent coefficients is stated in words rather than proved as a lemma. Since this is the point where algebraic independence is actually used, a precise statement and proof of the relevant exponential-function independence lemma would strengthen the paper.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: Theorem 4.1's reconstruction is derived from the linear-independence assumption and the Taylor expansion, not assumed; the notable flaw in the proof is a false scaling identity, which is a correctness gap rather than a circular reduction.

full rationale

The paper's claimed derivation chain is not circular in the sense defined by the review: no signature component is assumed in the statement of Theorem 4.1, no fitted parameter is renamed as a prediction, and no load-bearing step is justified solely by a self-citation. The central reconstruction argument proceeds by taking the CDE with vector fields scaled by r, applying the Taylor expansion, differentiating in r at 0, and then solving a linear system whose unknowns are the signature components. The solvability of that system is ensured by the explicit hypothesis of linear independence of tree-like vector fields, which is an algebraic condition on the vector fields, not a restatement of the desired reconstructive conclusion. The paper proves the required linear-independence result for depth-two exponential vector fields from algebraic independence of the random matrix entries and linear independence of exponential functions, again without relying on the theorem being proved. Citations to the authors' prior work, especially the conjecture in Akyildirim et al. (2022), serve only as motivation and context; the conjecture is not used as evidence for any mathematical claim. The tree-like representation is attributed to McLachlan et al. and Gubinelli, and the Taylor expansion to Baudoin and Zhang, both independent external sources. The known weakness in the paper is not circularity: the asserted identity Y^{eta,r}_t = r Y^{eta/r}_t in the proof of Theorem 4.1 is false for general nonlinear vector fields, and the associated r-derivatives may not be computable from the collection (Y^y). That is a mathematical gap in the proof, not a reduction of the conclusion to the assumptions. Likewise, the missing global boundedness of the exponential vector fields is a regularity gap, not a circularity. The exponential-feature counting statement is just the combinatorial observation that d^m features are recovered with hidden dimension N >= m-1, a comparison that is transparent rather than imported. Overall, the paper contains no step that is equivalent to its inputs by construction, so the circularity score is 0.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

No new physical or mathematical entities are postulated. The only modeling choices are the depth-two exponential vector fields and the algebraic independence assumption on random coefficients.

assumptions (4)
  • standard math Picard iteration Taylor expansion for CDEs driven by bounded variation paths (Theorem 2.2)
    Used to expand g(Y_t) into iterated integrals; stated with reference to Baudoin and Zhang, not re-derived.
  • domain assumption Random matrix entries with absolutely continuous i.i.d. distributions are almost surely algebraically independent over Q
    Section 3.3.2 relies on this to separate exponential terms; it is a standard probabilistic fact but is not proved in the paper.
  • domain assumption Global existence of CDE solutions requires vector fields with globally bounded derivatives
    Section 2.2 assumes this, but the nested exponential fields exp(A exp(Dx)) do not satisfy it; the paper does not justify global solutions for the main example.
  • ad hoc to paper Linear independence of tree-like vector fields extends to finite-dimensional Lie groups by analogous arguments
    Theorem 4.2 depends on this unproved extension; Section 4.2 explicitly says the details are omitted.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Signature Reconstruction from Randomized Signatures." pith.science (2026). https://pith.science/paper/HNSBS3RZ

@misc{pith2026250203163,
  author       = {Pith},
  title        = {Pith review of: Signature Reconstruction from Randomized Signatures},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HNSBS3RZ}},
  note         = {Machine review of arXiv:2502.03163}
}
read the original abstract

Controlled ordinary differential equations driven by continuous bounded variation curves can be considered a continuous time analogue of recurrent neural networks for the construction of expressive features of the input curves. We ask up to which extent well known signature features of such curves can be reconstructed from controlled ordinary differential equations with (untrained) random vector fields. The answer turns out to be algebraically involved, but essentially the number of signature features, which can be reconstructed from the non-linear flow of the controlled ordinary differential equation, is exponential in its hidden dimension, when the vector fields are chosen to be neural with depth two. Moreover, we characterize a general linear independence condition on arbitrary vector fields, under which the signature features up to some fixed order can always be reconstructed. Algebraically speaking this complements in a quantitative manner several well known results from the theory of Lie algebras of vector fields and puts them in a context of machine learning.

Figures

Figures reproduced from arXiv: 2502.03163 by the authors.

Figure 1
Figure 1. Example of a planar rooted tree in T with vertices {1, . . . , 7} labeled by the word w = (w1, . . . , w7) ∈ W7. 11 [PITH_FULL_IMAGE:figures/full_fig_p011_1.png] view at source ↗
Figure 4
Figure 4. [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
Figure 2
Figure 2. Left to right, the spaces Tw1 , Tw1w2 , Tw1w2w3 , and Tw1w2w3w4 . Denote by π1, . . . , πN : R N → R the canonical coordinate projections, πi : x 7→ xi . Definition 3.2 (Tree-like vector fields). To each tree τ ∈ T we associate the vector field Vτ : R N → R N defined recursively, for i ∈ {1, . . . , d}, and τ1, . . . , τk ∈ T, as V•i (x) = Vi(x), (3.1.1) V[τ1···τk]•i (x) = X N j1,...,jk=1 ∂ k ∂xjk · · · ∂xj1 Vi(x) ·… view at source ↗
Figures from the paper (5 more)
Figure 3
Figure 3. Figure 3: The planar rooted tree from Figure 1 with assigned edge-directions [PITH_FULL_IMAGE:figures/full_fig_p013_3.png]
Figure 4
Figure 4. Figure 4: Left to right, the spaces T 0 w1 , T 0 (w1,w2) , T 0 (w1,w2,w3) . Definition 3.3 (Tree-like operators). Let τ 0 ∈ T 0 . Then τ 0 = [τ1 · · · τk]•0 for some τ1, . . . , τk ∈ T. We define an associated operator Vτ 0 : C∞(R N ) → C∞(R N ) by setting: V•0 ϕ(x) = ϕ(x) and 1…
Figure 5
Figure 5. Figure 5: Form of the trees τ 0 ∈ Π. By the recursive tree construction, the first order differential operator must always arise from first edge attached to the root, and thus have direction j0. As no more edges can be added to the root without changing the order of the differen…
Figure 6
Figure 6. Figure 6: Illustration of how the indices i = (i1, . . . , im−1) are obtained in a specific recursive tree with m = 5. Suppose that we are given a pair (i, w) = ((i1, . . . , im−1),(w1, . . . , wm)) uniquely deter￾mining a letter-labeled recursive tree τ ∈ Tw of order m, and let…
Figure 7
Figure 7. Figure 7: The ’ladder’-like trees (τ, w,j) and (τ ′ , w′ ,j ′ ) Then: V j τ = cτ,jσ ′ (Aw1 x)σ ′ (π1(Aw2 x))· · · σ ′ (πm−2(Awm−1 x))σ(πm−1(Awmx)) and V j ′ τ ′ = cτ ′ ,j ′σ ′ (Aw1 x)σ ′ (πm−2(Awm−1 x))· · · σ ′ (π1(Aw2 x))σ(πm−1(Awmx)), and so V j τ = cτ,j cτ′ ,j ′ V j ′ τ ′ , …

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

43 extracted references · 35 canonical work pages

  1. [1]

    Akyildirim, E., Gambara, M., Teichmann, J., and Zhou, S. (2022). Applications of signature methods to market anomaly detection. https://doi.org/10.48550/arXiv.2201.02441

    arXiv doi:10.48550/arxiv.2201.02441 2022

  2. [2]

    Allan, A. (2021). Rough path theory: Lecture notes. https://metaphor.ethz.ch/x/2021/fs/401-4611-21L/notes/RP_lecture_notes_Allan.pdf

    work page 2021

  3. [3]

    P., Salvi, C., and Szpruch, L

    Arribas, I. P., Salvi, C., and Szpruch, L. (2020). Sig- SDE s model for quantitative finance. ACM International Conference on AI in Finance

    work page 2020

  4. [4]

    Bahturin, Y. (2021). Identical relations in L ie algebras , volume 68 of De Gruyter Expositions in Mathematics . De Gruyter, Berlin, second edition

    work page 2021

  5. [5]

    Barancikova, B., Huang, Z., and Salvi, C. (2024). SigDiffusions : Score-based diffusion models for long time series via log-signature embeddings. https://doi.org/10.48550/arXiv.2406.10354

    work page Pith review arXiv doi:10.48550/arxiv.2406.10354 2024

  6. [6]

    and Zhang, X

    Baudoin, F. and Zhang, X. (2012). Taylor expansion for the solution of a stochastic differential equation driven by fractional brownian motions. Electronic Journal of Probability , Vol. 17, pp. 1 -- 21

    work page 2012

  7. [7]

    Biagini, F., Gonon, L., and Walter, N. (2024). Universal randomised signatures for generative time series modelling. https://doi.org/10.48550/arXiv.2406.10214

    arXiv doi:10.48550/arxiv.2406.10214 2024

  8. [8]

    and Salvi, C

    Cass, T. and Salvi, C. (2024). Lecture notes on rough paths and applications to machine learning. https://doi.org/10.48550/arXiv.2404.06583

    arXiv doi:10.48550/arxiv.2404.06583 2024

Show all 43 references
  1. [9]

    and Lyons, T

    Chang, J. and Lyons, T. (2019). Insertion algorithm for inverting the signature of a path. https://doi.org/10.48550/arXiv.1907.08423

    2019 doi arXiv

  2. [10]

    and Kormilitzin, A

    Chevyrev, I. and Kormilitzin, A. (2016). A primer on the signature method in machine learning. https://doi.org/10.48550/arXiv.1603.03788

    2016 doi

  3. [11]

    Compagnoni, E., Scampicchio, A., Biggio, L., Orvieto, A., Hofmann, T., and Teichmann, J. (2023). On the effectiveness of randomized signatures as reservoir for learning rough dynamics. IEEE 2023 International Joint Conference on Neural Networks (IJCNN) , pp. 1 -- 8

    2023

  4. [12]

    Cuchiero, C., Gonon, L., Grigoryeva, L., Ortega, J.-P., and Teichmann, J. (2021a). Discrete-time signatures and randomness in reservoir computing. IEEE Transactions on Neural Networks and Learning Systems , Vol. 33, pp. 6321 -- 6330

    2021

  5. [13]

    Cuchiero, C., Gonon, L., Grigoryeva, L., Ortega, J.-P., and Teichmann, J. (2021b). Expressive power of randomized signature. The Symbiosis of Deep Learning and Differential Equations

    2021

  6. [14]

    Cuchiero, C., Schmocker, P., and Teichmann, J. (2023). Global universal approximation of functional input maps on weighted spaces. https://arxiv.org/abs/2306.03303

    2023

  7. [15]

    Fermanian, A., Lyons, T., Morrill, J., and Salvi, C. (2023). New directions in the applications of rough path theory. IEEE BITS the Information Theory Magazine

    2023

  8. [16]

    and Hairer, M

    Friz, P. and Hairer, M. (2014). A course on rough paths . Springer International Publishing

    2014

  9. [17]

    Gubinelli, M. (2010). Ramification of rough paths. Journal of Differential Equations , Vol. 248(4), pp. 693 -- 721

    2010

  10. [18]

    Hall, B. C. (2015). Lie groups, Lie algebras, and representations: An elementary introduction , volume 222 of Graduate Texts in Mathematics . Springer International Publishing

    2015

  11. [19]

    and Lyons, T

    Hambly, B. and Lyons, T. (2010). Uniqueness for the signature of a path of bounded variation and the reduced path group. Annals of Mathematics , Vol. 171(1), pp. 109 -- 167

    2010

  12. [20]

    Horvath, B., Lemercier, M., Liu, C., Lyons, T., and Salvi, C. (2023). Optimal stopping via distribution regression: a higher rank signature approach. https://doi.org/10.48550/arXiv.2304.01479

    2023 doi arXiv

  13. [21]

    Issa, Z., Horvath, B., Lemercier, M., and Salvi, C. (2024). Non-adversarial training of neural SDE s with signature kernel scores. Advances in Neural Information Processing Systems , Vol. 36

    2024

  14. [22]

    Kidger, P., Bonnier, P., Perez Arribas, I., Salvi, C., and Lyons, T. (2019). Deep signature transforms. Advances in Neural Information Processing Systems , Vol. 32

    2019

  15. [23]

    Krantz, S. G. and Parks, H. R. (2002). A primer of real analytic functions . Birkhäuser, Boston

    2002

  16. [24]

    Lee, J. M. (2012). Introduction to smooth manifolds , volume 218 of Graduate Texts in Mathematics . Springer, New York

    2012

  17. [25]

    Lemercier, M., Salvi, C., Damoulas, T., Bonilla, E., and Lyons, T. (2021). Distribution regression for sequential data. Proceedings of Machine Learning Research , Vol. 130, pp. 3754 -- 3762

    2021

  18. [26]

    Lou, H., Li, S., and Ni, H. (2024). Path development network with finite-dimensional Lie group representation. Transactions on Machine Learning Research

    2024

  19. [27]

    Lyons, T. (2014). Rough paths, signatures and the modelling of functions on streams. Proceedings of the International Congress of Mathematicians 2014 , Vol. IV

    2014

  20. [28]

    J., Caruana, M., and Lévy, T

    Lyons, T. J., Caruana, M., and Lévy, T. (2007). Differential equations driven by rough paths , volume 1908 of Lecture Notes in Mathematics . Springer, Berlin, Heidelberg

    2007

  21. [29]

    Lyons, T. J. and Xu, W. (2017). Hyperbolic development and inversion of signature. Journal of Functional Analysis , Vol. 272(7), pp. 2933 -- 2955

    2017

  22. [30]

    Lyons, T. J. and Xu, W. (2018). Inverting the signature of a path. Journal of the European Mathematical Society , 20(7), pp. 1655 - 1687

    2018

  23. [31]

    I., Modin, K., Munthe-Kaas, H., and Verdier, O

    McLachlan, R. I., Modin, K., Munthe-Kaas, H., and Verdier, O. (2017). Butcher series: A story of rooted trees and numerical methods for evolution equations. Asia Pacific Mathematics Newsletter 2017 , Vol. 7(1), pp. 1 -- 11

    2017

  24. [32]

    Michor, P. W. (2008). Topics in differential geometry . American Mathematical Society

    2008

  25. [33]

    Molev, A. I. (1987). The algebraic structure of the L ie algebra of vector fields on the line. Matematicheski i\ Sbornik (Novaya Seriya) , Vol. 134(176)(1), pp. 82 -- 92 & 142 -- 143

    1987

  26. [34]

    Morandi, P. (1996). Field and Galois theory , volume 167 of Graduate Texts in Mathematics . Springer, New York

    1996

  27. [35]

    Morrill, J., Salvi, C., Kidger, P., and Foster, J. (2021). Neural rough differential equations for long time series. Proceedings of Machine Learning Research , Vol. 139, pp. 7829 -- 7838

    2021

  28. [36]

    Muca Cirone, N., Lemercier, M., and Salvi, C. (2023). Neural signature kernels as infinite-width-depth-limits of controlled R es N ets. Proceedings of Machine Learning Research , Vol. 202, pp. 25358 -- 25425

    2023

  29. [37]

    Muca Cirone, N., Orvieto, A., Walker, B., Salvi, C., and Lyons, T. (2024). Theoretical foundations of deep selective state-space models. The Thirty-eighth Annual Conference on Neural Information Processing Systems

    2024

  30. [38]

    and Salvi, C

    Muca Cirone, N. and Salvi, C. (2025). Rough kernel hedging. https://doi.org/10.48550/arXiv.2501.09683

    2025 doi arXiv

  31. [39]

    and Salvi, C

    Pannier, A. and Salvi, C. (2024). A path-dependent PDE solver based on signature kernels. https://doi.org/10.48550/arXiv.2403.11738

    2024 doi

  32. [40]

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

    2021

  33. [41]

    Samelson, H. (1990). Notes on Lie algebras . Springer, New York

    1990

  34. [42]

    Schäfl, B., Gruber, L., Brandstetter, J., and Hochreiter, S. (2023). G-signatures: Global graph propagation with randomized signatures. https://doi.org/10.48550/arXiv.2302.08811

    2023 doi arXiv

  35. [43]

    Warner, F. W. (1983). Foundations of differentiable manifolds and Lie groups , volume 94 of Graduate Texts in Mathematics . Springer, New York

    1983

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.