Exponentially Fading Memory Signature
Eduardo Abi Jaber, Dimitri Sotnikov · 2025
Read Exponentially Fading Memory Signature by Eduardo Abi Jaber, Dimitri Sotnikov (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2507-03700.txt.
Introduction
Eduardo Abi Jaber and Dimitri Sotnikov define a new way to summarize a whole path of data, in a paper published in two thousand twenty-five. In their summary, old information fades instead of piling up. The classical path signature compresses a curve into a tidy algebraic fingerprint, but it remembers the distant past at full strength.
Their exponentially fading memory signature is the mean-reverting cousin: contributions decay as they age. Building it needs rough path theory stretched to handle integration from the infinite past. The construction keeps the signature's prized traits, including a modified Chen identity, the linearization property, path-determinacy, and universal approximation. For a time-augmented Brownian motion, the new signature evolves as an Ornstein-Uhlenbeck process on a group. It is stationary and Markov, with exponentially fading dependence. Those traits suit time series and signal processing, where yesterday should matter less than today.
Catalog record
| CREATOR | Eduardo Abi Jaber, Dimitri Sotnikov |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 173 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2507-03700
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Exponentially Fading Memory Signature Eduardo Abi Jaber, Dimitri Sotnikov Abstract We introduce the exponentially fading memory (EFM) signature, a time-invariant transformation of an infinite (possibly rough) path that serves as a mean-reverting analogue of the classical path signature. We construct the EFM-signature via rough path theory, carefully adapted to accommodate improper integration from minus infinity. The EFM-signature retains many of the key algebraic and analytical properties of classical signatures, including a suitably modified Chen identity, the linearization property, path-determinacy, and the universal approximation property. From the probabilistic perspective, the EFM-signature provides a “stationarized” representation, making it particularly well-suited for time-series analysis and signal processing overcoming the shortcomings of the standard signature. In particular, the EFM-signature of time-augmented Brownian motion evolves as a group-valued Ornstein–Uhlenbeck process. We establish its stationarity, Markov property, and exponential ergodicity in the Wasserstein distance, and we derive an explicit formula à la Fawcett for its expected value in terms of Magnus expansions. We also study linear combinations of EFM-signature elements and the computation of associated characteristic functions in terms of a mean-reverting infinite dimensional Riccati equation. 1 Introduction Fading memory and time invariance are important properties when modeling “memory” effects of dynamical systems. Time invariance means that the output signal Y_{t} at time t depends only on the continuous input signal (X_{s})_{-\infty<s\leq t} up to time t , but not on the absolute time t . In other words, a parallel shift in time of the input signal results in an identical shift of the output signal. More precisely, we suppose that there exists a functional F\colon C(\mathbb{R}_{+},\mathbb{R}^{d})\to\mathbb{R} such that, for all t\in\mathbb{R} , Y_{t}=F((X_{t-s})_{s\geq 0}). (1.1) Time invariance postulates, in some sense, a stationarity in the relationship between input and output. This is a much weaker condition than the stationarity of the output itself, as it allows for a nonstationary input. Nevertheless, the assumption of stationary dependence is crucial for the application of machine learning and econometric methods in time-series analysis. It is worth noting that, in practice, often only a single realization of a time series or a process trajectory—referred to more generally here as a “signal”—is available, for instance, in financial data. Within this framework, it is natural to assume that, at any given time t , the input signal X is observable from the infinitely distant past -\infty up to t . Without this assumption, the notion of time invariance would not make sense. Fading memory is an additional assumption on the functional F , introduced to handle the infinite length of the past in a tractable way. Informally, it means that the influence of the distant past diminishes over time, i.e., that Y_{t} gradually “forgets” the remote history of X . This idea dates back to the works of Volterra 2005 and Wiener 1958 . Boyd and Chua 1985 formalize this concept by requiring the functional F to be continuous not with respect to the uniform topology, but with respect to a weighted uniform topology, thereby strengthening the notion of continuity. That is, F has fading memory if its output remains close for input paths that are close in the recent past, even if they differ in the distant past. Moreover, Boyd and Chua 1985 prove a folk theorem stating that any linear time-invariant operator with fading memory admits a convolution representation. In discrete time, this implies that the output can be written as a weighted moving average: y_{n}=F((x_{n-k})_{k\geq 0})=\sum_{k\geq 0}w_{k}x_{n-k}, (1.2) for some sequence of weights w=(w_{k})_{k\geq 0}\in\ell^{1} . A particularly popular example of such a weighted moving average is the exponential moving average (EWMA), defined as y_{n}=\sum_{k\geq 0}\rho^{k}x_{n-k},\quad\rho\in(0,1). (1.3) Unlike the general case, the EWMA yields a Markovian sequence, as it satisfies y_{n+1}=\rho y_{n}+x_{n+1} , which significantly simplifies both analysis and computation. However, linear functionals are quite limited in their ability to capture complex memory effects. Boyd and Chua 1985 address the nonlinear case and provide an approximation result for time-invariant fading memory operators using Volterra series. This result is a consequence of the Stone–Weierstrass theorem and thus relies on the algebraic structure of the Volterra series. Nevertheless, as the authors point out, identifying an explicit approximating Volterra series remains a challenging task: “While the approximations are certainly strong enough to be useful in applications like macro-modeling of complicated systems or in universal nonlinear system identifiers, we know of no general procedure, based only on input/output measurements, by which an approximation can be found.” For a review of the Volterra series method and the different approaches to estimating Volterra kernels, we refer to Cheng et al. 2017 . Path signatures offer an alternative to Volterra series. Originally introduced by Chen 1958 , signatures have played an important role in stochastic Taylor expansions (see Kloeden and Platen 1992 ; Ben Arous 1989 ) and have attracted significant attention through the work of Terry Lyons and the development of rough path theory (see Lyons et al. 2007 ), which has led to a wide range of applications, particularly in machine learning. We refer the interested reader to the comprehensive surveys by Chevyrev and Kormilitzin 2016 and Lyons and McLeod 2022 . Similar to Volterra series, path signatures possess a rich algebraic structure and enjoy a linearization property, making them powerful tools for functional approximation. A key advantage of path signatures, compared to Volterra series, lies in their computational tractability, made possible by the celebrated Chen 1958 identity. When applied to time series or stochastic processes, the standard approach treats path signatures as features that efficiently summarize the information contained in the signal; see, e.g., Levin et al. 2016 ; Morrill et al. 2020 . Recent works by Cuchiero et al. 2023a and Abi Jaber and Gérard 2025 propose an alternative perspective, where linear functionals on path signatures are studied as stochastic processes themselves. This viewpoint allows for explicit representations of path-dependent processes as linear functionals of the time-augmented Brownian motion signature, as shown in Abi Jaber et al. 2024a . However, this framework is not naturally suited for modeling time-invariant dependencies as defined in ( 1.1 ). In particular, one cannot directly compute the signature of an infinite path. To incorporate information from the entire past, a natural model would be Y_{t}=\left\langle\bm{\ell},\mathbb{\widehat{X}}_{0,t}\right\rangle,\quad t\geq 0, (1.4) where \mathbb{\widehat{X}}_{0,t} denotes the signature of the time-augmented path \widehat{X}_{t}=(t,X_{t}) over [0,t] . This model is clearly not time-invariant and can easily lead to overfitting, as illustrated in the examples below. Furthermore, since signatures play a role analogous to polynomials in path space, ( 1.4 ) can be interpreted as a kind of Taylor expansion of the output process Y ; see, e.g., Litterer and Oberhauser 2014 ; Kloeden and Platen 1992 . While such expansions offer excellent local approximation properties, they may fail globally: a model trained on an interval [0,T_{1}] may diverge when applied on a longer interval [0,T_{2}] with T_{2}>T_{1} , as reported in Abi Jaber and Gérard 2025 . This behavior is undesirable in the context of modeling time-invariant systems ( 1.1 ). One could “stationarize” the signature model ( 1.4 ) to make it time-invariant by computing the path signature over a rolling window of fixed length \Delta>0 , that is, by considering Y_{t}=\left\langle\bm{\ell},\mathbb{\widehat{X}}_{t-\Delta,t}\right\rangle,\quad t\in\mathbb{R}. (1.5) This approach serves as a