Pith. sign in

REVIEW 3 major objections 4 minor 60 references

For linear differential equations, the cheapest path encoding uses a genuinely intermediate signature degree and interval count.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 00:16 UTC pith:6SAVDTPW

load-bearing objection Worth reading: real new bounds for signature compression, but the headline optimality claim is proved only for a relaxed proxy, not the exact discrete problem. the 3 major comments →

arxiv 2607.26281 v1 pith:6SAVDTPW submitted 2026-07-28 math.NA cs.NAmath.PR

Concise (varepsilon,r)-representations of a path

Emilio Ferrucci , Oliver Perr\'ee , Terry Lyons This is my paper
classification math.NA cs.NAmath.PR MSC 60L1065L20
keywords log-signaturepath compressionEuler schemelinear controlled differential equationsrough pathsfree Lie algebra dimensionstorage complexitysignature methods
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper asks how to store a path in the least memory while still being able to solve every linear controlled differential equation dY = A Y dX with |A| ≤ r to accuracy ε. Its answer is that the best encoding is generally neither of the two naive extremes — keeping only first-order increments on many intervals (N=1) nor storing one high-degree log-signature on the whole path (m=1) — but a middle point that balances the truncation degree N and the number of intervals m. For the relaxed version of this storage problem, the paper proves a unique optimum and derives its asymptotics: as r grows or ε shrinks, N* ~ sqrt(log(r e^r/ε)/log d) and m* ~ r sqrt(log d/λ) exp(sqrt(λ log d)) with λ = log(r e^r/ε). It further proves error bounds for degree-N Euler schemes that decay factorially in N, both for p-variation rough paths and for Itô SDEs, and shows numerically that the intermediate optimum beats either naive scheme in storage cost.

Core claim

The central claim is Theorem 1.5: for the continuously relaxed storage problem (1.14), the unique optimal representation satisfies N* ∼ sqrt(λ/log d) and m* ∼ r sqrt(log d/λ) exp(sqrt(λ log d)) as r→∞ or ε→0+, where λ = log(r e^r/ε). Because λ diverges in both limits, the optimal truncation degree and the optimal number of intervals both tend to infinity, so the optimum is asymptotically interior to the two naive choices. The paper also proves that this interior optimum beats the storage cost of the naive scheme with N=1 (Proposition 1.6). Around this optimization result, the paper establishes sharp Euler error bounds for linear controlled equations: factorial decay in N with polynomial deca

What carries the argument

The central object is the two-parameter family of representations of a path as truncated log-signatures of degree N on m subintervals. Storage cost is counted as mΛ_d(N), where Λ_d(N) is the dimension of the free Lie algebra over d generators truncated at degree N, asymptotically d^{N+1}/((d-1)N). The argument is carried by the relaxed constrained problem (1.14), which replaces the exact discrete cost with this asymptotic formula and the worst-case Euler error with the sharp bound g_r(N,m)=r^{N+1} e^r/((N+1)! m^N) set equal to ε. The function φ(N)=f_d(N,m*(N)) is shown to be strictly log-convex, which yields the unique optimum and its asymptotics.

Load-bearing premise

The load-bearing premise is that the continuously relaxed problem (1.14) — which uses the asymptotic dimension formula for Λ_d(N) and the sharp Euler error bound g_r set equal to ε — has the same asymptotics as the true discrete integer optimization problem (1.10); the paper gives numerical evidence but no theorem for this equivalence.

What would settle it

For a fixed d and a range of ε, compute the exact integer optimum of mΛ_d(N) subject to the exact Euler error inequality for the scalar ODE dY = rY dt (whose error is (1.13)); if the optimal N does not scale as sqrt(log(r e^r/ε)/log d), or the ratio of the true optimum to the relaxed cost does not approach 1, then the relaxed problem misrepresents the discrete one.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • For a path that drives linear equations, raising the signature degree N can substitute for refining the partition: the Euler error decays factorially in N even when m is fixed, so m=1 is feasible for large N.
  • The closed-form asymptotics give a parameter-free rule for choosing (N,m) from d, r, and ε alone, independent of the path's length or geometry.
  • The optimal storage cost grows more slowly than the cost of the N=1 representation, so higher-degree log-signatures are useful in practice for high-accuracy or high-norm problems.
  • In the p-variation regime the same tradeoff holds, with error decaying polynomially in m and factorially in N; the same structural asymptotics reappear for L2 errors of stochastic paths.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • My inference: the factorial decay that drives the result is special to linear equations; for nonlinear CDEs the corresponding Euler error lacks the factorial factor, so the intermediate optimum may fail or shift, and the paper's Remark 1.7 hints at this limit.
  • My inference: the asymptotic formula for N* can be read as a practical heuristic for signature-based compression pipelines: choose N from the allowed operator norm r and tolerance ε before choosing m, and only then optimize m.
  • My inference: a natural testable extension is to verify whether the same asymptotics hold when storage is measured in bits (floating-point precision) rather than number of reals, since the paper explicitly sets aside quantization error.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper studies the memory cost of representing a path as a collection of degree-N truncated log-signatures on m intervals, optimized under the requirement that the representation can approximate solutions of any linear controlled differential equation dY=AYdX with |A|≤r to accuracy ε. For bounded-variation paths, the authors replace the exact integer optimization problem (1.10) by a continuously relaxed proxy (1.14), prove that the relaxed problem has a unique minimizer, and derive the asymptotics N*∼√(λ/log d), m*∼r√(log d/λ) exp(√(λ log d)) with λ=log(re^r/ε) as r→∞ or ε→0+. They also prove factorial-decay Euler error bounds in the 1-variation and p-variation settings (Theorems 1.3 and 2.8), derive an L2 Euler error bound for Itô SDEs (Theorem 3.1), and support the results with numerical experiments.

Significance. The paper proposes a clean, practically motivated formulation of a storage-versus-expressivity tradeoff for path representations, and the main asymptotic analysis of the relaxed problem is transparent, parameter-free, and yields explicit, falsifiable predictions. Theorem 1.3's sharp 1-variation Euler bound with factorial decay in N is a useful standalone result, and the p-variation and L2 analogues are of independent interest. The authors provide reproducible code and extensive numerical validation. The central weakness is that the main optimality claims are proved only for the relaxed proxy (1.14), not for the original discrete problem (1.10), and the abstract's strictness claim over both naive choices is stronger than what is proved. If the transfer issue is resolved, or the claims are carefully qualified, the paper would be a solid contribution.

major comments (3)
  1. [§1, (1.10) vs (1.14), Theorem 1.5] The main theorem is a statement about the relaxed problem (1.14), where Λ_d(N) is replaced by d^{N+1}/((d-1)N) and the exact ℓ∞ constraint is replaced by the sharp bound g_r(N,m)=ε. The paper states that (1.10) is 'not analytically tractable' and then treats the relaxed optima as the answer to the original storage problem. No theorem shows that the integer argmin of (1.10), or its value, is asymptotically equivalent to the relaxed argmin; Figures 3–5 and Remark 1.8 provide numerical and algorithmic support, but not a proof. Because the abstract and the practical recommendations ('N* ∼ …, m* ∼ …') concern actual storage, this gap is load-bearing. Please add a transfer or equidistribution theorem, or consistently state the optimality results as being for the relaxed proxy only.
  2. [Abstract and Proposition 1.6] The strict optimality over naive choices is only proved for the pair (1,m°); the sentence before (1.24) explicitly says 'We do not show the corresponding statement for (N°,1) but verify it experimentally'. The abstract nevertheless asserts that the optimum 'generally lies strictly in between the two naive choices N=1 or m=1', and the introduction describes 'also m=1 is asymptotically suboptimal' as an empirically validated claim. This overstates the proven content. Either prove the comparison with (N°,1) or qualify the abstract and contribution statements accordingly.
  3. [§3, Theorem 3.1] The key quasi-shuffle identity underlying the Itô signature moment estimate is cited to [CFL+26, 4.14], an unpublished paper with overlapping authorship. This identity is the mechanism that gives factorial decay and the entire L2 bound; without a proof or a precise self-contained statement, Theorem 3.1 is not verifiable from the present manuscript. Please include the identity (or a proof) or clearly mark the theorem as conditional on an external preprint.
minor comments (4)
  1. [Theorem 1.3 proof] The sharpness example uses a one-dimensional ODE dY=rY dt, which conflicts with the standing assumption d≥2 in Remark 1.1. The argument can be embedded into R^d, but the mismatch should be noted.
  2. [Lemma 1.4] The convexity proof shows θ''>0 at ε=e^r r via η(N)>0. The extension to ε<e^r r is immediate because θ'' increases with λ, but this step is not spelled out.
  3. [Theorem 1.2 proof] The asymptotic estimate for Σ_d(N) is attributed to a Mathematics Stack Exchange answer [Did]. This is acceptable, but the short argument could be made self-contained.
  4. [Figure 5 caption] The caption says the figure corroborates 'the strictness claim of Proposition 1.6 and adding a similar one for the choice (N°,1)', but the text before (1.24) only claims experimental verification for (N°,1). The caption should be aligned with the text.

Circularity Check

0 steps flagged

No circularity in the main derivation: Theorem 1.5 is a self-contained analysis of the explicitly relaxed proxy (1.14); the exact-to-relaxed transfer is a correctness gap, not a circular step.

full rationale

The derivation chain for the main claim is: exact problem (1.10); explicit admission that it is 'not analytically tractable'; substitution of the asymptotics Λ_d(N)∼d^{N+1}/((d-1)N) and the Euler bound g_r(N,m)=ε, producing relaxed problem (1.14); Lemma 1.4 proves unique solution; Theorem 1.5 derives the asymptotics of that solution by calculus (θ'=0, Stirling, digamma). No parameter is fitted to data, and the asymptotic formulas are not assumed in the constraint; they are solved for. The only substantive gap is that the paper never proves that the integer argmin of (1.10) is asymptotically close to the relaxed argmin of (1.14)—the paper calls the substitution a 'rephrase' and supports it numerically (Figures 3–5). That is a correctness/rigour concern (the abstract's 'optimal representation' wording goes beyond the proved relaxed statement), not a circular reduction by construction. Likewise, Proposition 1.6 proves the comparison against (1,m°); the m=1 comparison is explicitly experimental ('We do not show the corresponding statement for (N°,1) but verify it experimentally'), which weakens the abstract's 'strictly in between' claim but again is not circular. Self-citations are present: Theorem 3.1 uses the Itô-signature moment identity [CFL+26, 4.14] (authors include two co-authors) and Lemma 2.4 cites [LL24]; these are component identities, not the target error bounds, and the main Section 1 theorem does not rest on them. Accordingly I find no circular step; score 2 reflects the minor self-citations and the unproved exact-to-relaxed transfer, not circularity.

Axiom & Free-Parameter Ledger

0 free parameters · 6 axioms · 0 invented entities

The central derivation relies on classical rough-path and Lie-algebra results, with no fitted data parameters. Section 3 additionally leans on a self-cited signature moment identity. No new physical or mathematical entities are introduced.

axioms (6)
  • standard math Factorial decay of signature for bounded variation paths: |S^n_{s,t}| ≤ ||X||^n_{1-var}/n! [LCL07, Prop 2.2]
    Used in Theorem 1.3 to bound the one-step Euler error.
  • standard math Witt's formula for dimensions of free Lie algebra homogeneous components
    Used in Theorem 1.2 to obtain the asymptotic storage dimension Λ_d(N).
  • standard math Factorial decay of the signature of a p-rough path: |S^n_{s,t}| ≤ ω^{n/p}/(β_p (n/p)!) [Lyo98, Thm 2.2.1; LX13, Thm 2.3]
    Used in Theorem 2.8 for the p-variation Euler bound.
  • domain assumption Standing assumption d≥2 so the free Lie algebra is nontrivial; time can always be prepended (Remark 1.1)
    Reduces the storage dimension discussion to the genuinely multidimensional case.
  • domain assumption Itô signature moment identity for Brownian motion with drift from [CFL+26, (4.14)]
    Used in Theorem 3.1 for the L2 Euler estimate; a self-cited result with overlapping authors.
  • standard math Mittag-Leffler inequality expp(x) ≤ p e^x [GKMR20]
    Used in Theorem 2.8 to convert Mittag-Leffler factors into exponentials.

pith-pipeline@v1.3.0-alltime-deepseek · 28424 in / 17098 out tokens · 155969 ms · 2026-08-01T00:16:10.543210+00:00 · methodology

0 comments
read the original abstract

Paths $X \colon [0,T] \to \mathbb R^d$ are traditionally stored in finite memory as time series. Recent research has underscored the benefits of instead representing them as collections of iterated integrals $\{\int_{0 < u_1 < \ldots < u_n < T} \mathrm{d} X_{u_1} \otimes \cdots \otimes \mathrm{d} X_{u_n}\}_{n = 0}^N$. These two encodings can be viewed as the extrema on a two-parameter spectrum of representations of the path as degree-$N$ signatures on $m$ intervals in a partition of $[0,T]$. We ask the question of which such representation takes up the least amount of memory, measured as number of real values needed to store the truncated log-signature, subject to the constraint of it being able to approximate solutions to linear controlled differential equations (CDEs) $\mathrm{d} Y = AY \mathrm{d} X$ with $|A| \leq r$ at accuracy at least $\varepsilon$. Estimating the error in terms of the length of $X$, we find that the optimal representation generally lies strictly in between the two naive choices $N = 1$ or $m = 1$, and derive its asymptotics as $r \to \infty$ and $\varepsilon \to 0^+$. Similar considerations can be made when estimating the error in terms of the $p$-variation norm of $X$: in this regime we prove an error bound of the degree-$N$ Euler scheme for linear CDEs with decay in both $m$ and (factorially) in $N$ with the other arbitrarily fixed. We conclude by setting up the analogous problem for SDEs, with the error measured in $L^2$, and derive a similar $L^2$-Euler error estimate for It\^o SDEs with drift. We include an empirical study of the optimisation problem, which we demonstrate for toy examples of $p$-rough paths and for fractional Brownian motion.

Figures

Figures reproduced from arXiv: 2607.26281 by Emilio Ferrucci, Oliver Perr\'ee, Terry Lyons.

Figure 1
Figure 1. Figure 1: Four one-cycle voltage-increment/current patterns [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Growth of the dimension of the truncated free Lie algebra, dimension of the tensor algebra, asymptotics predicted [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Plot of the constrained optimisation problem (1.14). We also compute the discrete minimum of the actual storage cost [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of the discrete optimum with true cost function [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Comparison of the true storage cost of the proxy [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: The one-step (m = 1) Euler scheme vs. log-ODE method. We have computed the L 2 norm of the error over 32 paths sampled from fractional Brownian motion with Hurst paramter H = 0.7, averaged over 10 random matrices with norm r = 2.2. Here and in all plots below, errors are compared against the solution computed using Diffrax [Kid, Kid21] on a very fine partition. For high enough r = |A|, the latter can fail … view at source ↗
Figure 7
Figure 7. Figure 7: We solve linear CDEs driven by the “approximate pure figure [PITH_FULL_IMAGE:figures/full_fig_p021_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Error plots and storage cost for (ε, r)-admissible choices of (N, m) with mΛd(N) ≤ 10000, for r = 1 and (H, ε) ∈ {(0.3, 0.1),(0.5, 0.01),(0.7, 0.001)}. “‘ 0.3 0.5 0.7 0.9 H 10 50 100 500 1000 m Λd(N) N = 5 N = 1 N = 1 True storage cost vs. H, empirical optimum vs. naive choices, ε = 0.1, r = 1, d = 4 best overall best with m = 1 best with N = 1 best with N = bH−1c 0.3 0.5 0.7 0.9 H 50 100 500 1000 5000 100… view at source ↗
Figure 9
Figure 9. Figure 9: Storage cost plotted vs. H for (ε, r)-admissible choices of (N, m) with mΛd(N) ≤ 10000. We consider both d = 4 for a broader range of H and in the last two plots focus on d = 3 for H = 0.3, 0.4 (lowering the dimension with fixed storage budget allows for higher N, m). We plot both the empirical optimum vs. that of the naive choices with N = 1 or m = 1, as well as the choice N = ⌊H−1 ⌋ which is important fr… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

60 extracted references · 15 canonical work pages

  1. [1]

    Minimal complexity sinusoidal controls for path planning , year=

    Gauthier, Jean-Paul and Kawskiz, Matthias , booktitle=. Minimal complexity sinusoidal controls for path planning , year=

  2. [2]

    Asymptotic behavior of _

    User Did , HOWPUBLISHED =. Asymptotic behavior of _

  3. [3]

    Clark, J. M. C. and Cameron, R. J. The maximum rate of convergence of discrete approximations for stochastic differential equations. Stochastic Differential Systems Filtering and Control. 1980

    1980

  4. [4]

    , title =

    Clark, Dean S. , title =. Discrete Appl. Math. , issn =. 1987 , language =. doi:10.1016/0166-218X(87)90064-3 , keywords =

    work page doi:10.1016/0166-218x(87)90064-3 1987

  5. [5]

    2015 , eprint=

    Pathwise approximation of SDEs by coupling piecewise abelian rough paths , author=. 2015 , eprint=

    2015

  6. [6]

    Patrick Kidger , year=

  7. [7]

    Castell, Fabienne and Gaines, Jessica , title =. Math. Comput. Simul. , issn =. 1995 , language =. doi:10.1016/0378-4754(93)E0062-A , keywords =

    work page doi:10.1016/0378-4754(93)e0062-a 1995

  8. [8]

    2026 , eprint=

    Orthogonal polynomials on path-space , author=. 2026 , eprint=

    2026

  9. [9]

    Patrick Kidger , title =

  10. [10]

    Magnus, Wilhelm , title =. Commun. Pure Appl. Math. , issn =. 1954 , language =. doi:10.1002/cpa.3160070404 , zbMATH =

    doi:10.1002/cpa.3160070404 1954

  11. [11]

    Dimension-free

    Boutaib, Youness and Gyurk. Dimension-free. Rev. Roum. Math. Pures Appl. , issn =. 2014 , language =

    2014

  12. [12]

    and Sidorova, Nadia , title =

    Lyons, Terry J. and Sidorova, Nadia , title =. Ill. J. Math. , issn =. 2006 , language =

    2006

  13. [13]

    Coutin, Laure and Qian, Zhongmin , title =. Probab. Theory Relat. Fields , issn =. 2002 , language =. doi:10.1007/s004400100158 , keywords =

    doi:10.1007/s004400100158 2002

  14. [14]

    and Caruana, Michael and L

    Lyons, Terry J. and Caruana, Michael and L. Differential equations driven by rough paths. 2007 , publisher =. doi:10.1007/978-3-540-71285-5 , keywords =

    doi:10.1007/978-3-540-71285-5 2007

  15. [15]

    and Victoir, Nicolas B

    Friz, Peter K. and Victoir, Nicolas B. , title =. 2010 , publisher =. doi:10.1017/CBO9780511845079 , keywords =

    doi:10.1017/cbo9780511845079 2010

  16. [16]

    2002 , publisher =

    Lyons, Terry and Qian, Zhongmin , title =. 2002 , publisher =. doi:10.1093/acprof:oso/9780198506485.001.0001 , keywords =

    arXiv 2002

  17. [17]

    Chang, Jiawei and Lyons, Terry and Ni, Hao , title =. C. R., Math., Acad. Sci. Paris , issn =. 2018 , language =. doi:10.1016/j.crma.2018.05.010 , keywords =

    work page doi:10.1016/j.crma.2018.05.010 2018

  18. [18]

    and Xu, Weijun , title =

    Lyons, Terry J. and Xu, Weijun , title =. Bull. Sci. Math. , issn =. 2013 , language =. doi:10.1016/j.bulsci.2013.04.004 , keywords =

    work page doi:10.1016/j.bulsci.2013.04.004 2013

  19. [19]

    Hara, Keisuke and Hino, Masanori , title =. Bull. Lond. Math. Soc. , issn =. 2010 , language =. doi:10.1112/blms/bdq013 , keywords =

    work page doi:10.1112/blms/bdq013 2010

  20. [20]

    2023 , eprint=

    An Adaptive Algorithm for Rough Differential Equations , author=. 2023 , eprint=

    2023

  21. [21]

    and Geng, X

    Boedihardjo, H. and Geng, X. , title =. Trans. Am. Math. Soc. , issn =. 2019 , language =. doi:10.1090/tran/7683 , keywords =

    work page doi:10.1090/tran/7683 2019

  22. [22]

    2026 , eprint=

    How Fast Do Signatures Learn? Statistical Theory and Applications for Path Regression , author=. 2026 , eprint=

    2026

  23. [23]

    and Hairer, Martin , title =

    Friz, Peter K. and Hairer, Martin , title =. 2020 , publisher =. doi:10.1007/978-3-030-41556-3 , keywords =

    doi:10.1007/978-3-030-41556-3 2020

  24. [24]

    Cass, Thomas and Litterer, Christian and Lyons, Terry , title =. Ann. Probab. , issn =. 2013 , language =. doi:10.1214/12-AOP821 , keywords =

    work page doi:10.1214/12-aop821 2013

  25. [25]

    2015 , eprint=

    Uniform Factorial Decay Estimate for the Remainder of Rough Taylor Expansion , author=. 2015 , eprint=

    2015

  26. [26]

    Bahturin, Yuri and Olshanskii, Alexander , title =. J. Algebra , issn =. 2015 , language =. doi:10.1016/j.jalgebra.2014.07.039 , keywords =

    work page doi:10.1016/j.jalgebra.2014.07.039 2015

  27. [27]

    and Riedel, Sebastian and Schoenmakers, John , title =

    Bayer, Christian and Friz, Peter K. and Riedel, Sebastian and Schoenmakers, John , title =. SIAM J. Numer. Anal. , issn =. 2016 , language =. doi:10.1137/140995209 , keywords =

    work page doi:10.1137/140995209 2016

  28. [28]

    Hu, Yaozhong and Liu, Yanghui and Nualart, David , title =. Ann. Appl. Probab. , issn =. 2016 , language =. doi:10.1214/15-AAP1114 , keywords =

    work page doi:10.1214/15-aap1114 2016

  29. [29]

    ICASSP 2022 -- 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages =

    Moore, Paul and Iliant, Theodor-Mihai and Ion, Filip-Alexandru and Wu, Yue and Lyons, Terry , title =. ICASSP 2022 -- 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages =. 2022 , publisher =. doi:10.1109/ICASSP43922.2022.9747285 , url =

    arXiv 2022

  30. [30]

    Scientific Data , volume =

    Kelly, Jack and Knottenbelt, William , title =. Scientific Data , volume =. 2015 , doi =

    2015

  31. [31]

    2009 , howpublished =

    Gubinelli, Massimiliano and Lejay, Antoine , title =. 2009 , howpublished =

    2009

  32. [32]

    Lyons, Terry , title =. Math. Res. Lett. , issn =. 1994 , language =. doi:10.4310/MRL.1994.v1.n4.a5 , keywords =

    work page doi:10.4310/mrl.1994.v1.n4.a5 1994

  33. [33]

    , zbMATH =

    Abramowitz, Milton and Stegun, Irene A. , zbMATH =. Handbook of mathematical functions with formulas, graphs, and mathematical tables. 1984 , language =

    1984

  34. [34]

    and Mainardi, Francesco and Rogosin, Sergei V

    Gorenflo, Rudolf and Kilbas, Anatoly A. and Mainardi, Francesco and Rogosin, Sergei V. , title =. 2020 , publisher =. doi:10.1007/978-3-662-61550-8 , keywords =

    doi:10.1007/978-3-662-61550-8 2020

  35. [35]

    , journal=

    Kolmogorov, A. , journal=. On the Shannon theory of information transmission in the case of continuous signals , year=

  36. [36]

    Kolmogorov, A. N. and Tikhomirov, V. M. , title =. Transl., Ser. 2, Am. Math. Soc. , issn =. 1959 , language =

    1959

  37. [37]

    Kolmogorov, A. N. , title =. Russ. Math. Surv. , issn =. 1983 , language =. doi:10.1070/RM1983v038n04ABEH004203 , keywords =

    work page doi:10.1070/rm1983v038n04abeh004203 1983

  38. [38]

    2013 , eprint=

    Epsilon-complexity of continuous functions , author=. 2013 , eprint=

    2013

  39. [39]

    Fitting a C^m -Smooth Function to Data, III , urldate =

    Charles Fefferman , journal =. Fitting a C^m -Smooth Function to Data, III , urldate =

  40. [40]

    Hambly, Ben and Lyons, Terry , title =. Ann. Math. (2) , issn =. 2010 , language =. doi:10.4007/annals.2010.171.109 , keywords =

    doi:10.4007/annals.2010.171.109 2010

  41. [41]

    Integration of Paths -- a Faithful Representation of Paths by Noncommutative Formal Power Series , author =. Trans. Amer. Math. Soc. , volume =. doi:10.2307/1993193 , urldate =. 1993193 , eprinttype =

    work page doi:10.2307/1993193

  42. [42]

    The signature of a rough path: Uniqueness , journal =

    Horatio Boedihardjo and Xi Geng and Terry Lyons and Danyu Yang , keywords =. The signature of a rough path: Uniqueness , journal =. 2016 , issn =. doi:https://doi.org/10.1016/j.aim.2016.02.011 , url =

    doi:10.1016/j.aim.2016.02.011 2016

  43. [43]

    Treue Darstellung Liescher Ringe

    Witt, Ernst , journal =. Treue Darstellung Liescher Ringe. , url =

  44. [44]

    , title =

    Lyons, Terry J. , title =. Rev. Mat. Iberoam. , issn =. 1998 , language =. doi:10.4171/RMI/240 , keywords =

    doi:10.4171/rmi/240 1998

  45. [45]

    Iterated

    Chen, Kuo-Tsai , year = 1954, journal =. Iterated

    1954

  46. [46]

    2014 , eprint=

    Rough paths, Signatures and the modelling of functions on streams , author=. 2014 , eprint=

    2014

  47. [47]

    2021 , eprint=

    A Generalised Signature Method for Multivariate Time Series Feature Extraction , author=. 2021 , eprint=

    2021

  48. [48]

    Structured Linear

    Benjamin Walker and Lingyi Yang and Nicola Muca Cirone and Cristopher Salvi and Terry Lyons , booktitle=. Structured Linear. 2025 , url=

    2025

  49. [49]

    Ferrucci, Emilio and Perrée, Oliver and Lyons, Terry , title =

  50. [50]

    2026 , eprint=

    Advances in Neural Controlled Differential Equations , author=. 2026 , eprint=

    2026

  51. [51]

    Embedding and learning with signatures , journal =

    Adeline Fermanian , keywords =. Embedding and learning with signatures , journal =. 2021 , issn =. doi:https://doi.org/10.1016/j.csda.2020.107148 , url =

    arXiv 2021

  52. [52]

    , title =

    Trefethen, Lloyd N. , title =. Matematica , issn =. 2023 , language =. doi:10.1007/s44007-022-00039-6 , keywords =

    work page doi:10.1007/s44007-022-00039-6 2023

  53. [53]

    and Platen, Eckhard , title =

    Kloeden, Peter E. and Platen, Eckhard , title =. 2010 , publisher =

    2010

  54. [54]

    First Conference on Language Modeling , year=

    Mamba: Linear-Time Sequence Modeling with Selective State Spaces , author=. First Conference on Language Modeling , year=

  55. [55]

    Theoretical Foundations of Deep Selective State-Space Models , url =

    Muca Cirone, Nicola and Orvieto, Antonio and Walker, Benjamin and Salvi, Cristopher and Lyons, Terry , booktitle =. Theoretical Foundations of Deep Selective State-Space Models , url =. doi:10.52202/079017-4041 , editor =

    work page doi:10.52202/079017-4041

  56. [56]

    and Paycha, Sylvie and Prei

    Bellingeri, Carlo and Friz, Peter K. and Paycha, Sylvie and Prei. Smooth rough paths, their geometry and algebraic renormalization , fjournal =. Vietnam J. Math. , issn =. 2022 , language =. doi:10.1007/s10013-022-00570-7 , keywords =

    doi:10.1007/s10013-022-00570-7 2022

  57. [57]

    2024 , eprint=

    A High Order Solver for Signature Kernels , author=. 2024 , eprint=

    2024

  58. [58]

    2025 , eprint=

    Log-PDE Methods for Rough Signature Kernels , author=. 2025 , eprint=

    2025

  59. [59]

    , title =

    Boedihardjo, Horatio and Geng, Xi and Souris, Nikolaos P. , title =. Adv. Math. , issn =. 2020 , language =. doi:10.1016/j.aim.2020.107043 , keywords =

    arXiv 2020

  60. [60]

    2019 , type =

    Reizenstein, Jeremy Francis , title =. 2019 , type =

    2019