A Bayesian approach to sound source reconstruction: optimal basis, regularization, and focusing.

The reconstruction of acoustical sources from discrete field measurements is a difficult inverse problem that has been approached in different ways. Classical methods (beamforming, near-field acoustical holography, inverse boundary elements, wave superposition, equivalent sources, etc.) all consist--implicitly or explicitly--in interpolating the measurements onto some spatial functions whose propagation are known and in reconstructing the source field by retropropagation. This raises the fundamental question as whether, for a given source topology and array geometry, there exists an optimal interpolation basis which minimizes the reconstruction error. This paper provides a general answer to this question, by proceeding from a Bayesian formulation that is ideally suited to combining information of physical and probabilistic natures. The main findings are the followings: (1) The optimal basis functions are the M eigen-functions of a specific continuous-discrete propagation operator, with M being the number of microphones in the array. (2) The a priori inclusion of spatial information on the source field causes super-resolution according to a phenomenon coined "Bayesian focusing." (3) The approach is naturally endowed with an internal regularization mechanism and results in a robust regularization criterion with no more than one minimum. (4) It admits classical methods as particular cases.