The Exponentially Weighted Signature

Alexandre Bloch, Samuel N. Cohen, Terry Lyons, Joël Mouterde, Benjamin WalkerarXiv 2026 · arXiv

TL;DR

The paper introduces the Exponentially Weighted Signature (EWS), a novel mathematical framework for continuous-time time-series analysis that allows machine learning models to dynamically learn how to weight past information. By generalizing previous fading-memory approaches from simple diagonal matrices to general linear operators, EWS captures complex history dynamics—such as oscillations, growth, and cross-channel interactions—while preserving the powerful algebraic properties of the classical path signature.

What Problem Does This Paper Address?

In time-series analysis, treating data as a continuous path rather than a discrete sequence avoids issues like vanishing gradients in RNNs or quadratic memory scaling in Transformers. The path signature is a classical tool from rough path theory that represents a continuous trajectory through its iterated integrals, offering a powerful, non-parametric feature set.

However, the classical signature has a major flaw: it treats all historical information uniformly. It doesn't know that yesterday's market crash might be more relevant to today's price than a minor fluctuation from three months ago.

Prior attempts to fix this, like the Exponentially Fading Memory (EFM) signature, applied exponential decay to the past. But EFM restricts each data channel to decay independently and monotonically. It cannot capture systems that exhibit oscillatory behavior (like a pendulum or seasonal trends), growth modes, or cross-channel memory dependencies (where an event in channel A changes how long we remember channel B).

Methods

The authors introduce the Exponentially Weighted Signature (EWS). Instead of using a diagonal matrix for temporal decay, EWS utilizes a general bounded linear operator $A$ within a matrix exponential $e^{-hA}$.

Key methodological highlights include:

  • Linear CDE Solution: The authors prove that the EWS is the unique solution to a linear controlled differential equation (CDE) on the tensor algebra.
  • Mathematical Generalization: EWS natively unifies and generalizes State-Space Models (SSMs) at its first tensor level, while its higher levels introduce intrinsic non-linearities. It also acts as a causal, continuous-time analogue to the Short-Time Fourier Transform (STFT) and Laplace transform.
  • Parallel Computation: Crucially, EWS retains a "group-like" structure. This allows the use of a modified Chen's identity, meaning the EWS over a long trajectory can be computed efficiently in parallel using a parallel associative scan ($O(\log N)$ time complexity) combined with the Van Loan identity for linear segments.

Datasets

The framework was evaluated using simulated trajectories generated from Stochastic Differential Equations (SDEs), which serve as standard mathematical testbeds for continuous-time modeling:

  1. Time-augmented 2D Brownian Motion: Used to test the expressivity gap by attempting to reconstruct a known depth-2 EWS target containing complex and growth eigenvalues.
  2. Coupled Oscillatory SDE: A two-dimensional stochastic system exhibiting non-linear coupling through trigonometric interaction terms, mean-reverting drift, and additive noise.

Key Insights and Ideas

  • The Power of Non-Diagonal Operators: Moving from diagonal to general matrices means the model can mix channels during the time-propagation step. This allows the system to learn complex, coupled memory dynamics directly in the continuous-time domain.
  • The Jordan Block Advantage: If the operator $A$ has a Jordan block structure, EWS naturally generates polynomial-exponential memory kernels at depth one. For non-linear systems like the Duffing oscillator, this allows EWS to approximate the system at a constant truncation depth (depth 3), whereas EFM requires a depth that scales linearly with the desired approximation accuracy ($3(K+1)$).
  • Global Anchoring vs. Local Discounting: Unlike Volterra signatures, which discount increments relative to their immediate neighbors, EWS anchors its discounting to a fixed terminal time. This subtle distinction is exactly what preserves the group-like structure necessary for parallel computation.

Key Results

The authors compared the EWS learner against an EFM learner and the classical signature on two regression tasks:

  • Expressivity Target Task: When trying to learn a target function defined by an EWS with complex/growth eigenvalues, the EWS learner achieved a Root Mean Squared Error (RMSE) of $4.96 \times 10^{-4}$, drastically outperforming both EFM ($2.43 \times 10^{-2}$) and the classical signature ($2.42 \times 10^{-2}$).
  • Coupled Oscillatory SDE Regression: On the highly coupled, oscillatory SDE tracking task, EWS achieved a test RMSE of $2.61 \times 10^{-2}$, significantly outperforming EFM ($9.39 \times 10^{-2}$) and the classical signature ($1.268 \times 10^{-1}$). Analysis of the learned matrix $A$ confirmed that EWS successfully recovered the dominant complex conjugate pairs required to capture the system's underlying oscillations.
Method Val RMSE Test RMSE
EWS $\mathbf{2.63 \times 10^{-2} \pm 2.10 \times 10^{-3}}$ $\mathbf{2.61 \times 10^{-2} \pm 2.60 \times 10^{-3}}$
EFM $9.45 \times 10^{-2} \pm 5.30 \times 10^{-3}$ $9.39 \times 10^{-2} \pm 4.80 \times 10^{-3}$
Signature $1.275 \times 10^{-1} \pm 4.90 \times 10^{-3}$ $1.268 \times 10^{-1} \pm 5.10 \times 10^{-3}$

Limitations and Future Work

  • Finite Horizon Focus: The EWS is formulated over finite time horizons. Extending it safely to infinite horizons would require enforcing strict stability constraints on the spectrum of the operator $A$ to prevent integrals from exploding.
  • Optimization Challenges: The authors noted that recovering matrices with complex eigenvalues via gradient-based search results in a more challenging optimization landscape, leading to slightly higher error variances during training.
  • Multiple Clocks: While the theory allows for paths to be parameterized by multiple intrinsic "clocks" (e.g., time and quadratic variation simultaneously), this extension was not fully explored or implemented in the experimental section.

Summary updated

Abstract

The signature is a canonical representation of a multidimensional path over an interval. However, it treats all historical information uniformly, offering no intrinsic mechanism for contextualising the relevance of the past. To address this, we introduce the Exponentially Weighted Signature (EWS), generalising the Exponentially Fading Memory (EFM) signature from diagonal to general bounded linear operators. These operators enable cross-channel coupling at the level of temporal weighting together with richer memory dynamics including oscillatory, growth, and regime-dependent behaviour, while preserving the algebraic strengths of the classical signature. We show that the EWS is the unique solution to a linear controlled differential equation on the tensor algebra, and that it generalises both state-space models and the Laplace and Fourier transforms of the path. The group-like structure of the EWS enables efficient computation and makes the framework amenable to gradient-based learning, with the full semigroup action parametrised by and learned through its generator. We use this framework to empirically demonstrate the expressivity gap between the EWS and both the signature and EFM on two SDE-based regression tasks.
BibTeX
@misc{bloch2026exponentially,
  author = {Alexandre Bloch and Samuel N. Cohen and Terry Lyons and Joël Mouterde and Benjamin Walker},
  title = {{The Exponentially Weighted Signature}},
  year = {2026},
  eprint = {2603.19198},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2603.19198},
}
Figure