Pith. sign in

REVIEW 4 major objections 6 minor 15 references

Path Signatures for Feature Extraction. An Introduction to the Mathematics Underpinning an Efficient Machine Learning Technique

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The signature of a path captures its shape up to tree-like equivalence.

desk verdict A genuinely accessible intro to path signatures, but the stated uniqueness theorem is false as printed and needs a fixed-start condition (or 'up to translation'). read the letter →

arxiv 2506.01815 v1 pith:7NXWARU2 submitted 2025-06-02 cs.LG math.PR

Stephan Sturm This is my paper
classification cs.LGmath.PR MSC 60L1062H30
keywords pathsignaturesiteratedintegralsfeatureextractiontimeseriesclassificationRiemann-Stieltjesintegraltree-likeequivalenceroughpathsChen'sidentity
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

Path signatures are introduced as a way to extract features from data streams for classification and prediction. The paper's central assertion is that a sequence of iterated integrals attached to a bounded-variation path—its signature—characterizes the path almost completely, up to reparameterization and retraced excursions, and that the truncated signature supplies a finite vector of real numbers a classifier can use. The mathematical claim stands on a theorem the paper states as Theorem 4.2, and the practical claim rests on the pipeline of interpolating a discrete stream into a path, truncating the signature at a chosen level, and handing the result to a standard classifier. Unlike moments or volatility, these features record the order in which events happen. The paper is deliberately an introduction rather than a fully rigorous treatment, and it assumes the reader has basic real analysis but not rough path theory.

What carries the argument

The central object is the signature of a bounded-variation path, the infinite sequence of iterated Riemann--Stieltjes integrals $$S_{a,b}(X) = (1, $S^{1}$_{a,b}(X), $S^{{1,1}}$_{a,b}(X), \ldots),\qquad $S^{{i_1,\ldots,i_n}}$_{a,b}(X) = \int_a^b \$int_a^{{t_n}}$ \cdots \$int_a^{{t_2}}$ $dX^{{i_1}}$_{t_1}\cdots $dX^{{i_n}}$_{t_n}.$$ Each entry is an iterated integral over the path components in a prescribed order, so the level-$n$ block records signed volumes associated with the order in which coordinates move. The practical object is the truncated signature, which keeps only levels $0$ through $N$ and has $1+d+\cdots+d^N$ entries for a $d$-dimensional path. Two supporting mechanisms carry the argument: time augmentation, which adds the coordinate $t$ so that the path is parameterized by its own time coordinate and the one-dimensional degeneracy disappears, and Chen's identity, which lets a signature on a long interval be assembled from signatures on adjacent subintervals and gives the recursive computation used in applications.

What would settle it

Construct two paths that are tree-like equivalent but carry different class labels; a signature-based classifier must assign identical features to both, so if any classifier using the complete stream can separate them, the practical claim that signatures are an adequate feature set fails. This is directly testable with synthetic streams.

Watch

Extended reading notes

Core claim

The central claim this expository paper makes is that the signature $S_{a,b}(X)$ of a $d$-dimensional bounded-variation path $X:[a,b]\to\mathbb{R}^d$ characterizes the path uniquely up to tree-like equivalence, and that the level-$N$ truncation of this signature provides a usable feature set for machine learning. Tree-like equivalence means two paths have the same signature when they differ only by the speed at which they are traversed or by portions that run out and come back, since those parts make no net contribution to the iterated integrals. The paper first develops the scalar analogue, where iterated Riemann integrals reproduce Taylor coefficients, and then shows why the one-dimensional case degenerates to endpoint information; that failure motivates going to higher-dimensional paths, and adding time as an explicit coordinate makes the graph $(t,X_t)$ uniquely recoverable. The practical claim is the production pipeline: interpolate the data stream to a path, truncate the signature at depth $N$, and feed the resulting real numbers to a classifier such as a random forest.

Load-bearing premise

The load-bearing premise is that turning a discrete stream into one particular continuous path—by linear interpolation, time augmentation, lead-lag, or another recipe—preserves the information the classifier needs; the paper offers no argument that the choice of embedding is harmless.

Editorial extensions

If this is right

  • Order information enters the feature set through mixed iterated integrals: the two second-level terms $\int_a^b \int_a^{t_2} dX^1_{t_1} dX^2_{t_2}$ and $\int_a^b \int_a^{t_2} dX^2_{t_1} dX^1_{t_2}$ distinguish whether coordinate 1 moved before coordinate 2 or the reverse.
  • Adding time as a coordinate before computing the signature removes the one-dimensional degeneracy, so the graph $(t,f(t))$ has a unique signature and the method can recover a full function rather than only its endpoints.
  • Truncated signatures give a fixed-size numerical vector with $1+d+\cdots+d^N$ entries, independent of stream length, making the features compatible with off-the-shelf classifiers.
  • Chen's identity lets the signature over a long interval be assembled from signatures over shorter ones, so features can be updated incrementally as a stream grows.
  • The same feature construction applies to prediction as well as classification, because the signature step is independent of the downstream algorithm.

Reading between the lines

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

  • The paper presents several path constructions—linear interpolation, time augmentation, piecewise constant paths, and lead-lag—as interchangeable, but it leaves open whether the choice of embedding preserves the information a classifier needs; a testable consequence is that a fixed dataset may yield different classification accuracies under different embeddings.
  • The uniqueness theorem concerns the infinite signature, whereas the pipeline always truncates, and the paper does not bound how quickly discriminative information is lost as the truncation level drops; there may be datasets whose classes differ only in high-order signature entries.
  • Because tree-like excursions cancel in the signature, a raw path whose class label depends on the internal shape of an out-and-back loop is invisible to the feature set; time augmentation and lead-lag are heuristics that may partly restore such information, but the paper gives no guarantee.
  • A direct empirical extension would be to sweep truncation levels and embedding choices on benchmark time-series classification data and compare signature features against standard statistical features; the paper cites one successful commodity-futures study but reports no systematic comparison.
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

4 major / 6 minor

Summary. The paper is an expository introduction to path signatures for feature extraction from time series data. It motivates signatures via a Taylor-series analogy, defines Riemann-Stieltjes integrals and iterated integrals, states the Hambly-Lyons uniqueness theorem for signatures of bounded-variation paths, and describes how discrete data streams can be embedded into continuous paths by linear interpolation, time augmentation, or a lead-lag transform. It then outlines a classification pipeline that uses truncated signatures as features and includes a worked level-3 signature calculation, geometric interpretations of iterated integrals, and an appendix on Chen's identity. The paper does not present new research results; its empirical claim about classification performance is deferred to the authors' companion paper [IKM+25].

Significance. If the correctness issues identified below are fixed, the paper would serve as a useful pedagogical introduction to the signature method for students and practitioners. Its strengths are the accessible progression from Taylor series to iterated integrals and the explicit worked examples and geometric pictures, which are rare in more advanced treatments. The mathematical results quoted are standard, so the paper's contribution is expository rather than novel. The practical effectiveness of truncated signatures as features is asserted rather than demonstrated in this manuscript, and the theoretical support (Theorem 4.2) concerns the full signature, not its truncation; the paper should state this limitation clearly.

major comments (4)
  1. [Definition 4.1 (Eq. 9)] The displayed definition of the signature has the integration limits reversed: the outermost integral is written as ∫_b^a and the nested intervals as ∫_{t_n}^a, ..., whereas the correct definition (as used in the worked examples and in the standard literature) is ∫_a^b ∫_a^{t_n} ... ∫_a^{t_2} 1 dX^{i1}_{t1} ... dX^{in}_{tn}. As printed, the definition would compute the signature with reversed time orientation and is inconsistent with the example S^{2,1,2}_{0,1}(X) computed later in the same section.
  2. [Theorem 4.2] The statement that the signature S_{a,b}(X) characterizes X uniquely up to tree-like equivalence is missing the standard hypothesis that the paths have a common starting point, or the conclusion must be 'up to translation and tree-like equivalence.' The signature is invariant under translation: for X_t=(t,0) and Y_t=(t,1) on [0,1], all iterated integrals are identical, yet the paths have different endpoints and are not tree-like equivalent in the usual sense. The error propagates to the sentence in §4 that the graph path (t,f(t)) is characterized uniquely: f(t) and f(t)+c have the same signature. The paper should add the fixed-start condition or an explicit translation caveat and should advise recording the initial value as an additional feature.
  3. [Section 5, linear interpolation] The formula for piecewise linear interpolation, 'Xt = D(ti) + (t - ti)D(ti+1)', is incorrect: at t=ti+1 it does not equal D(ti+1) unless ti+1 - ti = 1. The correct formula is Xt = D(ti) + (t - ti)/(ti+1 - ti)(D(ti+1) - D(ti)). Because this is the canonical construction used to turn data streams into paths, the error should be corrected before publication.
  4. [Sections 4 and 6, truncation and embeddings] The exposition moves from the uniqueness theorem for the full signature to the use of truncated signatures as the feature set without stating that the uniqueness property is lost after truncation. Similarly, Section 5 presents several embeddings (linear, piecewise constant, lead-lag) as interchangeable without discussing whether the chosen embedding preserves discriminative information. The paper should explicitly state that truncated signatures are an approximation and that practical effectiveness is an empirical matter; the current text relies entirely on the unpublished companion paper [IKM+25] for that claim.
minor comments (6)
  1. [Section 3, near Eq. (7)] In the telescoping-sum derivation, the term g(t_i)(f(t_{i+1}) + f(t_i)) should read g(t_i)(f(t_{i+1}) - f(t_i)); the final formula (7) is correct, so this is a sign typo in the intermediate line.
  2. [Appendix A, Eq. (11)] The index expression in Chen's identity is garbled; it should be S^{i1...ik}_{a,b}(X) · S^{i_{k+1}...i_n}_{b,c}(X), with the zeroth-order terms equal to 1.
  3. [Appendix A, second-order calculation] In the expression for S^{i1,i2}_{a,c}(X), the differential 'dX^{i2}_{i2}' should be 'dX^{i2}_{t2}'.
  4. [Section 4, last paragraph] The phrase 'no forth-back excursion' should be 'no forward-backward excursion' (or 'back-and-forth').
  5. [Section 3] The word 'subjext' in the sentence 'the total amount of change the function is subjext to' should be 'subject'.
  6. [Section 6, reference [JWH+23]] The name 'Tibsherani' should be 'Tibshirani'.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the main theorem is external, the supporting identity is proved in-paper, and the sole self-citation is illustrative rather than load-bearing.

full rationale

The paper's central uniqueness claim (Theorem 4.2) is explicitly attributed to Hambly and Lyons [HL10]; it is imported as an external mathematical theorem rather than derived from the paper's own definitions. Chen's identity in Appendix A is derived directly from the definition of iterated Riemann-Stieltjes integrals by induction, so it does not presuppose the signature result it explains. The practical assertion that random forests on truncated signatures have proved successful cites the authors' own companion paper [IKM+25], but that citation is not used as a premise of any mathematical derivation; it is an empirical pointer and therefore not load-bearing. There is no fitted parameter renamed as a prediction, no ansatz smuggled in through the literature, and no known result merely renamed. The only substantive defect is a mathematical-precision issue: Theorem 4.2 as printed omits the standard requirement that the paths share a starting point (or the conclusion must be 'up to translation'); signatures are translation-invariant, so initial data must be supplied separately. That is a correctness caveat about the theorem's statement, not circularity in the paper's derivation chain.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

The note's exposition rests on the standard uniqueness theorem and on an unproven practical premise about truncated signatures. It introduces no free parameters and no new entities.

assumptions (3)
  • standard math The Hambly-Lyons uniqueness theorem (Theorem 4.2) is taken as given.
    Invoked in Section 4 to justify that signatures characterize paths up to tree-like equivalence; the note does not prove it.
  • domain assumption Discrete data streams can be embedded into continuous bounded-variation paths without loss of information relevant to classification.
    Section 5 introduces linear interpolation, time augmentation, and lead-lag transforms as if the choice does not affect the usefulness of the feature set; no justification is provided.
  • ad hoc to paper Truncated signatures provide a good feature set for classification.
    The note asserts this throughout and cites the unpublished [IKM+25] for empirical support, but provides no data or benchmarks in the note.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Path Signatures for Feature Extraction. An Introduction to the Mathematics Underpinning an Efficient Machine Learning Technique." pith.science (2026). https://pith.science/paper/7NXWARU2

@misc{pith2026250601815,
  author       = {Pith},
  title        = {Pith review of: Path Signatures for Feature Extraction. An Introduction to the Mathematics Underpinning an Efficient Machine Learning Technique},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7NXWARU2}},
  note         = {Machine review of arXiv:2506.01815}
}
read the original abstract

We provide an introduction to the topic of path signatures as means of feature extraction for machine learning from data streams. The article stresses the mathematical theory underlying the signature methodology, highlighting the conceptual character without plunging into the technical details of rigorous proofs. These notes are based on an introductory presentation given to students of the Research Experience for Undergraduates in Industrial Mathematics and Statistics at Worcester Polytechnic Institute in June 2024.

Figures

Figures reproduced from arXiv: 2506.01815 by the authors.

Figure 1
Figure 1. How can we train a computer that it can distinguish between [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. ”Bump” function that is everywhere infinitely often differentiable, but not real an￾alytic. But overall, this approach is promising, and it seems the crucial object we have seen appear are the iterated integrals In(x) = Z x x0 Z xn x0 · · · Z x3 x0 Z x2 x0 1 dx1dx2 · · · dxn−2dxn−1dxn. So far all what these iterated integrals are doing is to calcu￾late the volume (x−x0) n n! of an n-dimensional simplex with side len… view at source ↗
Figure 3
Figure 3. i) A Riemann sum approximating the Riemann integral [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: The total variation of the function g is given by the total sum of the absolute increments which is the sum of the length of the vertical green lines. The total variation of a function g measures the total amount of change the function is subjext to. For avid hikers, a…
Figure 5
Figure 5. Figure 5: Examples of two-dimensional paths. A d-dimensional path is a continuous function X : [a, b] → R d , where we will now use the shorthand Xt for X(t). Thus Xt is a d-dimensional vector of one￾dimensional functions Xt = [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: i) Path Y . ii) Path Y˜ that has same signature as Y as it is tree-like equivalent. iii) Path even more tree-like (forest-like?) equivalent to Y , so signatures are indistinguishable We want also reconnect to the end of the previous section. There we have seen that the…
Figure 7
Figure 7. Figure 7: Geometric interpretation of the mixed second order terms of the signature. i) If the path stays [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: The geometric interpretation of the it￾erated integrals. The point (x1, x2) is above the path as the path passes first at t1 the length of x1 and only later at t2 the height of x2. To understand this better and enable us to general￾ize the understanding to higher dimen…
Figure 9
Figure 9. Figure 9: Different path constructions from the time series (3 [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]
Figure 10
Figure 10. Figure 10: The production pipeline of classification using signatures. In each step a choice has to be made: [PITH_FULL_IMAGE:figures/full_fig_p012_10.png]
Figure 11
Figure 11. Figure 11: Geometric interpretation of Chen’s identity. i) If the path is increasing in each component, areas [PITH_FULL_IMAGE:figures/full_fig_p015_11.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

15 extracted references · 10 canonical work pages

  1. [1]

    Iterated integrals and exponential homomorphisms

    Kuo-Tsai Chen. Iterated integrals and exponential homomorphisms. Proc. London Math. Soc. (3) , 4:502--512, 1954

    work page 1954

  2. [2]

    Integration of paths, geometric invariants and a generalized B aker- H ausdorff formula

    Kuo-Tsai Chen. Integration of paths, geometric invariants and a generalized B aker- H ausdorff formula. Ann. of Math. (2) , 65:163--178, 1957

    work page 1957

  3. [3]

    A primer on the signature method in machine learning, 2016

    Ilya Chevyrev and Andrey Kormilitzin. A primer on the signature method in machine learning, 2016. https://arxiv.org/abs/1603.03788

    arXiv 2016

  4. [4]

    Extracting information from the signature of a financial data stream, 2014

    Lajos Gergely Gyurkó, Terry Lyons, Mark Kontkowski, and Jonathan Field. Extracting information from the signature of a financial data stream, 2014. https://arxiv.org/abs/1307.7244

    arXiv 2014

  5. [5]

    Sparse arrays of signatures for online character recognition, 2013

    Benjamin Graham. Sparse arrays of signatures for online character recognition, 2013. https://arxiv.org/abs/1308.0371

    arXiv 2013

  6. [6]

    Uniqueness for the signature of a path of bounded variation and the reduced path group

    Ben Hambly and Terry 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

    work page 2010

  7. [7]

    Krishnan, Adam Mullaney, Kathleen Shiffer, and Stephan Sturm

    Yasutora Ito, Hari P. Krishnan, Adam Mullaney, Kathleen Shiffer, and Stephan Sturm. Understanding lost civilizations and commodity term structures: The signature method. submitted , 2025

    work page 2025

  8. [8]

    Classifying commodities using the signature method

    Yasutora Ito, Adam Mullaney, and Kathleen Shiffer. Classifying commodities using the signature method. Technical report, Worcester Polytechnic Institute, 2024

    work page 2024

Show all 15 references
  1. [9]

    An introduction to statistical learning---with applications in P ython

    Gareth James, Daniela Witten, Trevor Hastie, Robert Tibshirani, and Jonathan Taylor. An introduction to statistical learning---with applications in P ython . Springer Texts in Statistics. Springer, Cham, 2023

    2023

  2. [10]

    Differential equations driven by rough signals

    Terry Lyons. Differential equations driven by rough signals. I . A n extension of an inequality of L . C . Y oung. Math. Res. Lett. , 1(4):451--464, 1994

    1994

  3. [11]

    Terry J. Lyons. Differential equations driven by rough signals. Rev. Mat. Iberoamericana , 14(2):215--310, 1998

    1998

  4. [12]

    A generalised signature method for multivariate time series feature extraction, 2021

    James Morrill, Adeline Fermanian, Patrick Kidger, and Terry Lyons. A generalised signature method for multivariate time series feature extraction, 2021. https://arxiv.org/abs/2006.00873

    2021 arXiv

  5. [13]

    The signature-based model for early detection of sepsis from electronic health records in the intensive care unit

    James Morrill, Andrey Kormilitzin, Alejo Nevado-Holgado, Sumanth Swaminathan, Sam Howison, and Terry 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) , 2019

    2019

  6. [14]

    Rough path theory and signatures applied to quantitative finance

    Imanol P \' e rez. Rough path theory and signatures applied to quantitative finance. P art I--IV , N.D. https://www.quantstart.com/articles/rough-path-theory-and-signatures-applied-to-quantitative-finance-part-1/ [Retrieved: 7. Nov. 2024]

    2024

  7. [15]

    The iisignature library: efficient calculation of iterated-integral signatures and log signatures, 2018

    Jeremy Reizenstein and Benjamin Graham. The iisignature library: efficient calculation of iterated-integral signatures and log signatures, 2018. https://arxiv.org/abs/1802.08252

    2018 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.