Statistical Parameter Calibration with the Generalized Fluctuation Dissipation Theorem and Generative Modeling

Parameter calibration in complex dynamical systems often relies on costly optimization routines or ad hoc tuning to match statistical properties of observations. In this work, we develop a principled framework for statistical calibration grounded in the Generalized Fluctuation-Dissipation Theorem (GFDT). This approach provides exact linear response formulas that relate infinitesimal changes in internal model parameters to infinitesimal changes in statistics of arbitrary observables. In other words, the GFDT yields parameter Jacobians of system statistics without requiring adjoint models, ensemble perturbations, or repeated simulations. We demonstrate the framework's utility across a hierarchy of systems, including analytically tractable linear models, nonlinear double-well potentials, and multiscale stochastic models relevant to climate dynamics. We show that these Jacobians can be embedded within classical optimization schemes - such as Newton-type updates or regularized least squares - to guide parameter updates. The method is further extended to handle perturbations in both drift and diffusion terms, enabling unified treatment of deterministic and stochastic calibration. Our results establish the GFDT as a rigorous and interpretable foundation for parameter tuning in non-equilibrium systems.