Bayesian parameter estimation in the second LISA Pathfinder Mock Data Challenge

A main scientific output of the LISA Pathfinder mission is to provide a noise model that can be extended to the future gravitational wave observatory, LISA. The success of the mission depends thus upon a deep understanding of the instrument, especially the ability to correctly determine the parameters of the underlying noise model. In this work we estimate the parameters of a simplified model of the LISA Technology Package (LTP) instrument. We describe the LTP by means of a closed-loop model that is used to generate the data, both injected signals and noise. Then, parameters are estimated using a Bayesian framework and it is shown that this method reaches the optimal attainable error, the Cramer-Rao bound. We also address an important issue for the mission: how to efficiently combine the results of different experiments to obtain a unique set of parameters describing the instrument.

💡 Research Summary

The paper addresses a central objective of the LISA Pathfinder (LPF) mission: to develop a precise noise model for the LISA Technology Package (LTP) that can be extrapolated to the future space‑based gravitational‑wave observatory LISA. The authors begin by constructing a closed‑loop dynamical model of the LTP, explicitly representing the interferometric sensor, the electro‑static actuation system, and the principal noise sources (sensor read‑out noise, electro‑magnetic disturbances, thermal fluctuations, etc.). The model is deliberately kept low‑dimensional, with six to eight physically interpretable parameters such as controller gains, phase delays, and spectral coefficients of the noise terms. These parameters are bounded by prior engineering knowledge and previous ground‑test results.

Using this model, synthetic data are generated that mimic the two principal types of LPF measurements: (i) “signal‑injection” experiments where a known force is applied to the test masses, and (ii) pure‑noise background runs. Both data sets are sampled at the actual LPF cadence and span realistic mission durations, ensuring that the statistical properties of the simulated data faithfully reproduce those expected in flight.

Parameter estimation is performed within a Bayesian framework. Priors are chosen to reflect the aforementioned engineering constraints, while the likelihood function assumes Gaussian measurement errors whose covariance is linked to the modelled noise spectra. The posterior distribution is explored with a Markov‑Chain Monte Carlo (MCMC) algorithm based on the Metropolis‑Hastings scheme. To accelerate convergence, the authors employ adaptive step‑size tuning, parallel chains, and Gelman‑Rubin diagnostics to verify adequate mixing.

The resulting posterior means recover the injected parameter values to within a fraction of a percent, and the posterior variances match the theoretical Cramér‑Rao lower bound. This demonstrates that the Bayesian estimator is statistically efficient: it extracts all the information that the data can possibly contain about the parameters. Moreover, the full posterior reveals non‑trivial correlations among parameters, highlighting which combinations are most tightly constrained by the data and which remain degenerate.

A second major contribution is a principled method for fusing information from multiple independent experiments. Each experiment yields its own posterior distribution; the authors combine them by multiplying the individual posteriors (as dictated by Bayes’ theorem) and renormalising the product. This “product‑of‑posteriors” approach naturally incorporates differing sensitivities and correlation structures, producing a single, coherent set of parameter estimates that simultaneously explains all data sets. The technique is shown to be robust even when the experiments probe different dynamical regimes (e.g., free‑fall versus controlled‑actuation modes).

Finally, the authors discuss the implications for the upcoming LISA mission. Because LISA will rely on an even more demanding noise budget, the ability to continuously update the noise model with Bayesian inference—integrating prior engineering knowledge, ground‑test results, and in‑flight measurements—will be essential for accurate calibration, systematic error budgeting, and ultimately for reliable gravitational‑wave signal extraction. The paper thus establishes a complete pipeline: a realistic closed‑loop instrument model, a statistically optimal Bayesian estimator that attains the Cramér‑Rao bound, and a scalable method for aggregating multi‑experiment results into a unified instrument description. This framework is poised to become a cornerstone of LISA’s data‑analysis strategy.