pscicluna
commented
6 years ago I am currently implementing a very simple version, exclusively for spectra for the moment - the covariance matrix is modelled as the linear combination of an uncorrelated component (i.e. a diagonal matrix) and a single Gaussian component (i.e. rho_ij = exp( -(i - j)**2 / (2*d**2)), where d is a free parameter controlling the scale-length of the correlations). The two components are weighted such that the sum of the weights = 1, and is truncated for small values of rho_ij to try to improve computational efficiency. Tests are ongoing, but it appears to work so far!
This is the real meat of ampere, and so has to be done right. We need a general way of populating the covariance matrices for each kind of data and a way of getting the optimiser to understand how the parameters it is given relate to this. It will also need some optimisation as we want the matrices to be as sparse as possible for computational efficiency, so we may have to come up with truncation criteria or ways of chunking the matrices (e.g. splitting the data into blocks, with lower-resolution interactions with more distant data blocks).