Speaker
Description
Separating periodic, quasi-periodic, and stochastic signals out of noisy, irregularly sampled time series is a central problem of time-domain astronomy. It is of particular interest in the problem of extra solar planets. In the analysis of observational data the three regimes are easily confused once time gaps and colored noise distort their spectra. This talk focuses on the quasi-periodic case modelled as a second-order autoregressive process. We approach the component separation as a problem of generative modelling and parameter estimation, and compare two complementary approaches: Gaussian-process regression with a damped-oscillator kernel, and a state-space/Kalman-filter formulation — both of which accommodate gaps and uneven sampling natively. Using synthetic time series with controlled period, coherence, noise level, and cadence (both evenly and unevenly spaced), we assess the accuracy and bias of the recovered parameters and the calibration of the inferred uncertainties.