SVAR-FM: Intervention-Based Time Series Causal Discovery via Simulator-Generated Distributions
Researchers have introduced SVAR-FM, a novel framework for time series causal discovery that leverages physics-based simulators to generate interventional data. By treating the simulator as a mechanical realization of Pearl's do operator, the method physically severs confounding paths through variable clamping. The framework employs Conditional Flow Matching to learn nonlinear interventional conditionals. Theoretical analysis proves that the full structural VAR becomes identifiable under specific coverage conditions, with an derived error bound decomposing into Monte Carlo, simulator fidelity, and Flow Matching terms. A key finding is the 'sign-flip corollary,' which predicts that causal effect estimates may reverse sign if simulator accuracy falls below a certain threshold. Empirical benchmarks across four scientific domains demonstrate that SVAR-FM correctly recovers causal signs where observational methods fail due to confounding. Furthermore, a case study in ultrafast laser physics validated the sign-flip prediction by varying the accuracy of a quantum solver, showing that high-accuracy settings recover the correct causal direction with zero bias and an R-squared of 0.983.
Wire timeline
SVAR-FM: Intervention-Based Time Series Causal Discovery via Simulator-Generated Distributions
Researchers have introduced SVAR-FM, a novel framework for time series causal discovery that leverages physics-based simulators to generate interventional data. By treating the simulator as a mechanical realization of Pearl's do operator, the method physically severs confounding paths through variable clamping. The framework employs Conditional Flow Matching to learn nonlinear interventional conditionals. Theoretical analysis proves that the full structural VAR becomes identifiable under specific coverage conditions, with an derived error bound decomposing into Monte Carlo, simulator fidelity, and Flow Matching terms. A key finding is the 'sign-flip corollary,' which predicts that causal effect estimates may reverse sign if simulator accuracy falls below a certain threshold. Empirical benchmarks across four scientific domains demonstrate that SVAR-FM correctly recovers causal signs where observational methods fail due to confounding. Furthermore, a case study in ultrafast laser physics validated the sign-flip prediction by varying the accuracy of a quantum solver, showing that high-accuracy settings recover the correct causal direction with zero bias and an R-squared of 0.983.
cs.AI updates on arXiv.org