Deep Koopman operator framework for causal discovery in nonlinear dynamical systems

Juan Nathaniel, Carla Roesch, Jatan Buch, Derek DeSantis, Adam Rupe, Kara Lamb, Pierre Gentine · 2025

Read Deep Koopman operator framework for causal discovery in nonlinear dynamical systems by Juan Nathaniel, Carla Roesch, Jatan Buch, Derek DeSantis, Adam Rupe, Kara Lamb, Pierre Gentine (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-2505-14828.txt.

Introduction

Juan Nathaniel and co-authors present a method to find cause and effect links in nonlinear systems that change over time, in a paper from 2025. Classic tools struggle when feedback loops, mixed time scales, and drift are present, as in the climate system. Their method builds on a theory that rewrites nonlinear motion in a linear space.

Deep learning finds the best variables for that rewrite. Cause strength is then measured as the gap between the motion of the effect alone and the joint motion of cause and effect. Tests beat prior methods at finding and sizing causal signals. The team also applied the method to the El Nino pattern in the Pacific, a major climate driver.

Catalog record

CREATORJuan Nathaniel, Carla Roesch, Jatan Buch, Derek DeSantis, Adam Rupe, Kara Lamb, Pierre Gentine
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE58 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2505-14828

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Deep Koopman operator framework for causal discovery in nonlinear dynamical systems Juan Nathaniel, Carla Roesch, Jatan Buch, Derek DeSantis, Adam Rupe, Kara Lamb, Pierre Gentine Abstract We use a deep K oopman operator-theoretic formalism to develop a novel c ausal discovery algorithm, Kausal . Causal discovery aims to identify cause-effect mechanisms for better scientific understanding, explainable decision-making, and more accurate modeling. Standard statistical frameworks, such as Granger causality, lack the ability to quantify causal relationships in nonlinear dynamics due to the presence of complex feedback mechanisms, timescale mixing, and nonstationarity. This presents a challenge in studying many real-world systems, such as the Earth’s climate. Meanwhile, Koopman operator methods have emerged as a promising tool for approximating nonlinear dynamics in a linear space of observables. In Kausal , we propose to leverage this powerful idea for causal analysis where optimal observables are inferred using deep learning. Causal estimates are then evaluated in a reproducing kernel Hilbert space, and defined as the distance between the marginal dynamics of the effect and the joint dynamics of the cause-effect observables. Our numerical experiments demonstrate Kausal ’s superior ability in discovering and characterizing causal signals compared to existing approaches of prescribed observables. Lastly, we extend our analysis to observations of El Niño-Southern Oscillation highlighting our algorithm’s applicability to real-world phenomena. Our code is available at https://github.com/juannat7/kausal . † † affiliationnotice: Equal contribution 1 Introduction Causal discovery seeks to disentangle the underlying cause-effect mechanisms of physical processes, which is crucial for improved scientific understanding ( Aloisi et al. 2022 ; Cattry et al. 2025 ; Roesch et al. 2025 ; Williams et al. 2025 ) . Since performing interventions to detect causality is often infeasible in many domains, there is a great interest to develop data-driven causal discovery methods ( Granger 1969 ; Hyvärinen et al. 2010 ; Runge 2020 , e.g.,) . However, they tend to operate under certain assumptions, such as time-invariance, stationarity, and linearity ( Granger 1969 ; Runge et al. 2019 ; Bareinboim et al. 2022 ; Camps-Valls et al. 2023 ) , which limit their applicability. 1 1 1 Causal discovery is different from causal effect estimation, where the latter aims to quantify the causal effect between variables for an existing (either discovered or assumed) causal graph. For instance, these simplifications fail while examining nonlinear systems due to the presence of complex feedback mechanisms, spatial autocorrelation, and timescale mixing ( Lorenz 1963 ; Rupe & Crutchfield 2024 ) . In general, there are two categories of approaches for inferring causal relationships in nonlinear dynamical systems: i) using information flow to identify causal influence among variables and subspaces ( Granger 1969 ; Liang & Kleeman 2005 ; Majda & Harlim 2007 ; James et al. 2016 ; Tank et al. 2021 ) ; and ii) performing a lifting transformation to project the system of interest to a higher-dimensional space where causal effects can be investigated more easily ( Lorenz 1991 ; Sugihara et al. 2012 ; Gilpin 2024 ; Butler et al. 2024 ) . Thus, we pose our research question: Can we combine these two principles to apply linear causal analysis methods in nonlinear systems? A natural foundation for us to start answering this question is Koopman theory ( Koopman 1931 ; Mezić 2013 ; Brunton et al. 2022 ) , which states that finite nonlinear dynamics can be represented with an infinite-dimensional linear operator acting on the space of all possible measurement functions. This formalism presents a significant opportunity for the application of standard linear methods for system analysis, estimation, and control ( Brunton et al. 2025 ) . Recent work by Rupe et al. 2024 has established a theoretical foundation linking Koopman theory to causal analysis for dynamical systems using prescribed embeddings. Simultaneously, efficiently and effectively estimating optimal embeddings has been enabled by rapid advances of deep learning in high-dimensional settings ( Brunton et al. 2016 ; Lusch et al. 2018 ; Wang et al. 2022 ) . Therefore, data-driven Koopman operator methods have seen a surge in interest, especially for spatiotemporal analysis in a number of fields ( Kutz et al. 2018 ; Azencot et al. 2020 ; Rice et al. 2021 ; Froyland et al. 2021 ; Lintner et al. 2023 ; Lamb et al. 2024 ; Nathaniel & Gentine 2025 ) . Main contributions. In this paper, we leverage ideas from deep learning, Koopman theory, and causal inference. Our main contributions can be summarized as follows: • We introduce a novel deep K oopman operator-based c ausal discovery framework, Kausal , as illustrated in Figure 1 . • We show that Kausal infers a better representation of the basis functions through deep learning, as opposed to existing prescribed approaches, which leads to a more accurate and meaningful causal analysis in high-dimensional nonlinear systems. • We successfully apply Kausal to both idealized models and real-world phenomena highlighting the algorithm’s scalability and generalizability. Figure 1 : Schematic of the Kausal algorithm. We 1) estimate the embeddings using deep learning, and 2) approximate the Koopman operator with dynamic mode decomposition (DMD). Then, 3) we infer causal measures by computing the difference in prediction error between marginal (effect-only) and joint (effect-cause) models. 2 Methodology 2.1 Background In this section, we follow Rupe et al. 2024 to briefly introduce the mathematical background and the necessary notation to introduce our novel causal algorithm in Section 3 . Dynamical systems. Dynamical systems theory describes the evolution of a system over time, with the system state represented as a point in a phase space, \omega\in\Omega , which we assume to be in the Euclidean space \mathbb{R}^{N} . Throughout this paper variables are treated as vectors unless explicitly stated otherwise. Flow maps. Time-evolution of states is defined through a set of functions called flow maps \{\Phi^{t}:\Omega\rightarrow\Omega\}_{t\in T} that carries the phase space back onto itself. For an initial state \omega\in\Omega , flow maps evolve \omega through time t , given by {\omega(t)=\Phi^{t}(\omega)} . The orbit of a state \omega is the set \{\omega(t)=\Phi^{t}(\omega)\}_{t\in T} , representing the trajectory of the system over the time domain T . Although we denote the initial state as \omega , writing \omega(t,\omega_{0})=\Phi^{t}(\omega_{0}) with the initial state \omega_{0} would be equivalent. For simplicity, we also restrict our analysis to autonomous dynamical systems characterized by time-independent flow maps. Ordinary differential equations. In a deterministic setting, the evolution of \omega(t) over time t can be described through ordinary differential equations (ODEs) of the form: \frac{d\omega}{dt}=f(t,\omega) (1) where f is an arbitrary function and the solution \omega(t) depends on the initial condition \omega(t_{0})=\omega_{0} at time t_{0} . Components. To formalize the idea of causality in a dynamical system, we need to define different components of the system which can impact each other. For an N -dimensional phase space (i.e. N degrees of freedom) as described above, we partition the space into disjoint system component subspaces as \Omega=\Omega_{1}\times\Omega_{2}\times\dots\times\Omega_{L} with \omega_{i}\in\Omega_{i} and \omega=[\omega_{1}\text{ }\omega_{2}\text{ }\dots\text{ }\omega_{L}] , where L\leq N . Reproducing Kernel Hilbert Spaces (RKHS). Koopman operators (described in the next section) act on functions rather than directly on system states, mapping one function to another by evolving it under the system dynamics. To ensure the operations of the Koopman operator are mathematically consistent

Read the complete text