Likelihood transform: making optimization and parameter estimation easier

Yan Wang

DOI: https://doi.org/10.48550/arXiv.1402.6211

2014-02-25

Abstract:Parameterized optimization and parameter estimation is of great importance in almost every branch of modern science, technology and engineering. A practical issue in the problem is that when the parameter space is large and the available data is noisy, the geometry of the likelihood surface in the parameter space will be complicated. This makes searching and optimization algorithms computationally expensive, sometimes even beyond reach. In this paper, we define a likelihood transform which can make the structure of the likelihood surface much simpler, hence reducing the intrinsic complexity and easing optimization significantly. We demonstrate the properties of likelihood transform by apply it to a simplified gravitational wave chirp signal search. For the signal with an signal-to-noise ratio 20, likelihood transform has made a deterministic template-based search possible for the first time, which turns out to be 1000 times more efficient than an exhaustive grid- based search. The method in principle can be applied to other problems in other fields as the spirit of parameterized optimization and parameter estimation problem is the same.

General Relativity and Quantum Cosmology