FR
ENOù et quand
09/09/2025 10:30 - Room B211 of the Coriolis building (CERMICS conference room) at ENPC
Description
Title:
Harnessing the Conditional Gaussian Nonlinear System Framework for Efficient Adaptive-Lag Online Smoothing and Causal Inference
Harnessing the Conditional Gaussian Nonlinear System Framework for Efficient Adaptive-Lag Online Smoothing and Causal Inference
Abstract:
This talk presents recent advances in data assimilation and causal inference through the Conditional Gaussian Nonlinear System (CGNS) framework—a class of nonlinear stochastic models where, conditioned on a subset of state variables, the unobserved components follow a Gaussian posterior. This structure enables efficient Bayesian state estimation and sampling via closed-form solutions, making CGNS particularly well suited for high-dimensional, multiscale systems with regime shifts and intermittent extreme events.
This talk presents recent advances in data assimilation and causal inference through the Conditional Gaussian Nonlinear System (CGNS) framework—a class of nonlinear stochastic models where, conditioned on a subset of state variables, the unobserved components follow a Gaussian posterior. This structure enables efficient Bayesian state estimation and sampling via closed-form solutions, making CGNS particularly well suited for high-dimensional, multiscale systems with regime shifts and intermittent extreme events.
We first introduce an adaptive-lag online smoother that exploits the analytical tractability of CGNSs to reduce storage demands of standard smoothing procedures. By adaptively adjusting the lag using information-theoretic criteria, this method is applicable to turbulent systems with time-varying temporal correlations and performs well in real-world problems such as Lagrangian data assimilation and online parameter estimation.
Building on this, we present Assimilative Causal Inference (ACI), a paradigm-shifting framework that traces causes backward from observed effects via inverse problem-solving rather than forward influence quantification. ACI uniquely identifies dynamic causal interactions without requiring observations of candidate causes, works with short datasets, scales efficiently to high dimensions, and provides online tracking of causal roles—even when they reverse intermittently. Within the CGNS framework, ACI enables explicit nil-causality principles and analytical characterization of causal influence ranges, offering rigorous tools for temporal attribution and prediction.
Applications to turbulent dynamics and nonlinear geophysical flows illustrate how CGNSs deliver scalable, real-time, and theoretically grounded solutions to online smoothing and causal inference.
