Predrag Pilipović: Parameter Estimation from Steady-State Observations via Generator Symbols
Time: Tue 2026-12-01 13.15 - 14.15
Location: KTH 3721 (Lindstedtsvägen 25)
Participating: Predrag Pilipović (University of Copenhagen)
Abstract: We consider parameter estimation from independent steady-state observations. For example, we may want to estimate how bacterial division rate depends on cell length from a single snapshot of a culture. Steady-state density is rarely available in closed form, making likelihood estimation difficult. Existing methods use the fact that, at steady state, the instantaneous change of a test function has mean zero. Some select a finite set of test functions, while others use kernels to combine such conditions. We model the symbol, the function that specifies how an operator acts on Fourier characters. For a process with drift, diffusion, or jumps, the symbol describes these local dynamics, and stationarity implies that the corresponding moment vanishes at every frequency. More generally, this moment condition defines a statistical model without requiring an underlying process. We estimate the parameters by the continuum generalized method of moments and derive asymptotically optimal weights. We also show when a growing sample of frequencies gives the same asymptotic precision as the full continuum.
