pith. sign in

arxiv: 2606.23472 · v1 · pith:QQN55PPUnew · submitted 2026-06-22 · 📊 stat.ML · cs.LG

Time Series Classification through Diffeomorphic Time Warping (DiffTW)

Vicky Geneva Haney , Kamel Lahouel , Victor Rielly , Bruno M. Jedynak This is my paper

Pith reviewed 2026-06-26 06:11 UTC · model grok-4.3

classification 📊 stat.ML cs.LG
keywords time series classificationdiffeomorphic mappingdynamic time warpingtransport equationnearest neighbor
0
0 comments X

The pith

Diffeomorphic mappings from transport equations supply a continuous dissimilarity measure that beats dynamic time warping for nearest-neighbor classification on most time series benchmarks.

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

The paper introduces DiffTW to replace the discrete point matching of DTW with learned smooth invertible transformations between entire time series functions. These transformations are obtained by approximating the flow along characteristic curves of a linear transport partial differential equation whose velocity field depends on position. The partial differential equation is converted to ordinary differential equations by the method of characteristics, and the objective follows from the fundamental theorem of calculus. A reproducing kernel Hilbert space representation of the velocity field together with optimal control allows flexible optimization of the mapping. Inserted as a distance into a 1-nearest-neighbor classifier, the resulting measure improves accuracy over DTW on 60 of 86 standard datasets.

Core claim

DiffTW supplies a dissimilarity measure by learning diffeomorphic transformations that approximate the flows of characteristic curves of a linear transport equation with space-dependent velocity, obtained by reducing the equation to ordinary differential equations and optimizing the velocity field in a reproducing kernel Hilbert space.

What carries the argument

Diffeomorphic mapping produced by integrating the velocity field along the characteristic curves of the transport equation.

If this is right

  • Time series similarity is measured by a continuous deformation rather than discrete point correspondences.
  • The dissimilarity can be inserted directly into any distance-based classifier such as 1-nearest neighbor.
  • The construction links time-series alignment to the theory of linear transport equations.
  • The method yields higher classification accuracy than DTW on 60 of the 86 tested datasets.

Where Pith is reading between the lines

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

  • The same transport-based construction could be tested on other sequence problems such as clustering or anomaly detection.
  • Alternative partial differential equations or different function spaces for the velocity field might produce further gains on particular data domains.

Load-bearing premise

The learned mappings can be treated as approximations to the flows associated with the characteristic curves of a linear transport equation that has a space-dependent velocity field.

What would settle it

A head-to-head run of 1-nearest-neighbor classification on the same 86 datasets in which the DiffTW distance fails to exceed DTW accuracy on more than 43 datasets would falsify the performance claim.

Figures

Figures reproduced from arXiv: 2606.23472 by Bruno M. Jedynak, Kamel Lahouel, Vicky Geneva Haney, Victor Rielly.

Figure 1
Figure 1. Figure 1: Visualization of continuous time warping via the method of characteristics. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Top Left: The original function ϕ0 and the target ϕ1. Top Right: The warping evolution and characteristic curves, illustrating how individual time points migrate from the source ϕ0 to the offset estimated target ϕ1,est across the integration steps. Middle Left: The warping path progression mapping the trajectory dynamics of x(t) across optimiza￾tion steps, shown for every 5th iteration against the gray das… view at source ↗
Figure 3
Figure 3. Figure 3: Example of an ECG signal capturing a full cardiac cycle, featuring the P wave, [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Top Left: The ECG input signals ϕ0 (N) and the target ϕ1 (R). Top Right: The warping evolution and characteristic curves. Middle Left: Warping path progression mapping the trajectory dynamics of x(t) across 11 optimization iterations. Middle Right: Evolution of the learned velocity field αest over 11 iterations. Bottom Left: Objective value across iterations, showing decreasing values. Bottom Right: Norm o… view at source ↗
Figure 5
Figure 5. Figure 5: Accuracy comparison across 86 datasets. DiffTW achieves a higher median [PITH_FULL_IMAGE:figures/full_fig_p023_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Performance on the BME dataset, highlighting DiffTW’s ability to precisely [PITH_FULL_IMAGE:figures/full_fig_p023_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: ACSFOne dataset signatures, characterized by sparse high-energy bursts that [PITH_FULL_IMAGE:figures/full_fig_p024_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Ham dataset spectrographic measurements, where both methods achieved iden [PITH_FULL_IMAGE:figures/full_fig_p025_8.png] view at source ↗
read the original abstract

Time series classification involves learning a mapping from a continuous, temporally ordered sequence of real-valued observations to a discrete response variable, like class labels. This task is fundamental in domains, including health monitoring, where the temporal structure of data is critical for accurate prediction. Dynamic Time Warping (DTW) is a standard technique for measuring similarity between sequences varying in time or speed. However, DTW is restricted to discrete point matching. To move beyond pairwise alignment, we propose a theoretical framework that learns mappings between real-valued functions. These mappings approximate the flow associated with the characteristic curves of a linear transport equation with a space-dependent velocity field, providing a diffeomorphic transformation between two time series. Using the method of characteristics, we transform this partial differential equation into ordinary differential equations (ODEs) modeling system dynamics. The objective function used to learn these ODEs derives from the fundamental theorem of calculus. To enable flexible, expressive representations of the velocity field, we utilize reproducing kernel Hilbert spaces and optimal control methods. Our method, Diffeomorphic Time Warping (DiffTW), provides a theoretically grounded dissimilarity measure. Using a 1-nearest neighbor classifier, DiffTW outperforms DTW on 60 of 86 datasets.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The manuscript introduces Diffeomorphic Time Warping (DiffTW) for time series classification. It constructs diffeomorphic mappings between functions by approximating flows along characteristic curves of a linear transport PDE with space-dependent velocity, converts the PDE to ODEs via the method of characteristics, obtains the learning objective from the fundamental theorem of calculus, represents the velocity field in an RKHS via optimal control, and reports that 1NN classification using the resulting dissimilarity outperforms DTW on 60 of 86 datasets.

Significance. If the PDE-to-ODE derivation and FTC-based objective are internally consistent and the empirical gains are reproducible, the work supplies a continuous, theoretically motivated dissimilarity that extends DTW while preserving diffeomorphic properties. This could be useful in domains requiring smooth temporal alignments. No machine-checked proofs, open code, or parameter-free derivations beyond the stated FTC link are described.

major comments (2)
  1. [Abstract] Abstract: the claim that the objective derives from the fundamental theorem of calculus and yields a dissimilarity independent of fitted parameters is load-bearing for the parameter-free assertion, yet the description leaves open whether the measure reduces to quantities defined by the RKHS coefficients or velocity parameters; explicit expansion of this step is required to confirm the construction does not become circular.
  2. [Abstract] Abstract: the central empirical result (1NN outperformance on 60/86 datasets) rests on the learned dissimilarity being a valid diffeomorphic warping; without verification that the characteristic-curve approximation and RKHS control preserve the diffeomorphism in the implemented solver, the claim that DiffTW is theoretically grounded cannot be assessed.
minor comments (1)
  1. [Abstract] Abstract: the number and identity of the 86 datasets should be stated explicitly (e.g., UCR archive version) to allow direct replication of the 60/86 count.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on the manuscript. We address each major comment below and will revise the paper accordingly to improve clarity and strengthen the theoretical grounding.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claim that the objective derives from the fundamental theorem of calculus and yields a dissimilarity independent of fitted parameters is load-bearing for the parameter-free assertion, yet the description leaves open whether the measure reduces to quantities defined by the RKHS coefficients or velocity parameters; explicit expansion of this step is required to confirm the construction does not become circular.

    Authors: We agree that the link from the FTC to a parameter-independent dissimilarity requires explicit expansion to eliminate any ambiguity. The objective is obtained by integrating the transport equation along characteristics, after which the dissimilarity depends only on the resulting endpoint mappings rather than the specific RKHS coefficients or velocity parameters. In the revision we will insert a step-by-step derivation (in the methods section, with a brief reference in the abstract) that shows the reduction explicitly and confirms the construction is non-circular. revision: yes

  2. Referee: [Abstract] Abstract: the central empirical result (1NN outperformance on 60/86 datasets) rests on the learned dissimilarity being a valid diffeomorphic warping; without verification that the characteristic-curve approximation and RKHS control preserve the diffeomorphism in the implemented solver, the claim that DiffTW is theoretically grounded cannot be assessed.

    Authors: The theoretical construction guarantees diffeomorphism because the velocity field belongs to an RKHS that yields Lipschitz-continuous flows, and the method of characteristics produces a diffeomorphic mapping under these conditions. Nevertheless, the referee correctly notes that the implemented solver requires explicit verification that the numerical approximations preserve this property. In the revision we will add a dedicated paragraph (with supporting analysis or bounds) confirming that the characteristic-curve discretization and RKHS control maintain invertibility and positive Jacobian determinants in practice. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper's core derivation transforms a linear transport PDE into ODEs via the method of characteristics and obtains the objective from the fundamental theorem of calculus; both steps invoke standard external mathematics rather than self-referential definitions or fitted parameters renamed as predictions. RKHS optimal control is introduced only as a representation tool for the velocity field. No self-citations, uniqueness theorems, or ansatzes smuggled from prior author work appear in the abstract or described framework. The 1NN empirical comparison on 86 datasets is an independent evaluation step, not part of the claimed derivation. The construction is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract alone supplies no explicit free parameters, axioms, or invented entities; ledger left empty due to insufficient information.

pith-pipeline@v0.9.1-grok · 5757 in / 1012 out tokens · 29397 ms · 2026-06-26T06:11:57.357618+00:00 · methodology

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

18 extracted references · 4 canonical work pages

  1. [1]

    A global averaging method for dynamic time warping, with applications to clustering , journal =

    Petitjean, Fran. A global averaging method for dynamic time warping, with applications to clustering , journal =. 2011 , publisher =

    2011

  2. [2]

    Advances in Neural Information Processing Systems , volume =

    Neural Ordinary Differential Equations , author =. Advances in Neural Information Processing Systems , volume =. 2018 , publisher =

    2018

  3. [3]

    and Ba, Jimmy , title =

    Kingma, Diederik P. and Ba, Jimmy , title =. International Conference on Learning Representations (ICLR) , year =

  4. [4]

    Advances in neural information processing systems , volume=

    Random features for large-scale kernel machines , author=. Advances in neural information processing systems , volume=

  5. [6]

    2024 , issn =

    A novel distance measure based on dynamic time warping to improve time series classification , journal =. 2024 , issn =. doi:https://doi.org/10.1016/j.ins.2023.119921 , url =

    work page doi:10.1016/j.ins.2023.119921 2024

  6. [7]

    2023 , eprint=

    Computing Continuous Dynamic Time Warping of Time Series in Polynomial Time , author=. 2023 , eprint=

    2023

  7. [8]

    Marco Cuturi and Mathieu Blondel , booktitle =. Soft-. 2017 , editor =

    2017

  8. [9]

    Hoang Anh Dau and Anthony Bagnall and Kaveh Kamgar and Chin-Chia Michael Yeh and Yan Zhu, Shaghayegh Gharghabi and Chotirat Ann Ratanamahatana and Eamonn Keogh , url =. The

  9. [10]

    Moody GB and Mark RG , url =. The

  10. [11]

    Bake off redux: a review and experimental evaluation of recent time series classification algorithms , volume=

    Bake off redux: a review and experimental evaluation of recent time series classification algorithms , author =. Data Mining and Knowledge Discovery , year =. doi:10.1007/s10618-024-01022-1 , url =

    work page doi:10.1007/s10618-024-01022-1

  11. [12]

    2011 , institution =

    Cardiac Arrhythmia Detection using Dynamic Time Warping of ECG Beats in E-Healthcare Systems , author =. 2011 , institution =

    2011

  12. [13]

    Journal of Statistical Software , volume =

    Computing and Visualizing Dynamic Time Warping Alignments in R: The dtw Package , author =. Journal of Statistical Software , volume =. 2009 , doi =

    2009

  13. [14]

    McOwen , title = "

    Robert C. McOwen , title = ". 2003 , address =

    2003

  14. [15]

    Brendon J Brewer, Daniel Foreman-Mackey, and David W Hogg

    Beg, M. Faisal and Miller, Michael I. and Trouv. Computing Large Deformation Metric Mappings via Geodesic Flows of Diffeomorphisms , journal =. 2005 , volume =. doi:10.1023/B:VISI.0000043755.93987.aa , url =

    work page doi:10.1023/b:visi.0000043755.93987.aa 2005

  15. [16]

    2022 , address =

    Sergei Pereverzyev , title =. 2022 , address =

    2022

  16. [17]

    Econometrica , number = 1, pages =

    Regression Quantiles , author =. Econometrica , number = 1, pages =

  17. [18]

    and Kozubowski, T

    Kotz, S. and Kozubowski, T. and Podg\'orski, K. , publisher =. The

  18. [19]

    and Elbroch, L

    Pozdnyakov, V. and Elbroch, L. M. and Labarga, A. and Meyer, T. and Yan, J. , title =. Methodology and Computing in Applied Probability , year = 2019, volume = 21, number = 3, pages =. doi:10.1007/s11009-017-9547-6 , url =

    work page doi:10.1007/s11009-017-9547-6 2019