Speaker
Description
Accurate and scalable noise modelling will be essential for extracting weak gravitational-wave signals and signatures of new physics from LISA data. Realistic time-delay interferometry channels may contain frequency-dependent correlations, while the detector noise may also evolve over the mission.
I will present a Bayesian framework for estimating the full spectral density matrix of multivariate LISA data. The model uses penalised splines to learn both auto-spectra and cross-spectra, while preserving a valid covariance matrix at every frequency. A differentiable, coarse-grained Whittle likelihood enables efficient inference using the No-U-Turn Hamiltonian Monte Carlo sampler, initialised with variational inference.
Applied to simulated LISA data, the multivariate method accurately recovers the full correlated noise structure. For an idealised symmetric detector, it agrees with the standard assumption of diagonal noise in the A, E, and T channels. However, under realistic instrumental asymmetries, this simpler univariate (diagonal) approximation breaks down and produces errors more than an order of magnitude larger than the full multivariate model presented here.
I will also discuss ongoing work extending the framework to non-stationary noise using the Wilson–Daubechies–Meyer time-frequency transform, enabling time-varying structure and joint inference of evolving LISA noise and gravitational-wave signals.