Introduction
Here is a draft of a method to construct precision matrices based on correlated Gaussian samples with mean zero and variance one.
Method
Let be a random vector of size and , where is the number of temporal replicates. We assume that and that the marginal variance of the -th element of , , is . This means that the diagonal of is a vector of ones, and that
Furthermore, it is assumed that Q is a sparse precision matrix. Using the properties of Gaussian conditional distributions, we have
where is the set containing the neighbors of site , i.e. the sites that are such that if .
Assume that we have realizations of that can be used to infer the precision matrix . We set up a regression model to estimate the non-zero elements of . Here, we consider as a realization, i.e. as an observation. The regression model for each site, , will be
At each site , we estimate the parameter vector with
where
and is the -th neighbor oy at time . The variance of is and it is estimated with
The next step is to transform and such that they give estimates of the elements of , namely
where is the -th element og . Let be a matrix with -th element . Note that , and thus . Furthermore let be a diagonal matrix such that
An estimate of Q can now be presented as
where is an identity matrix of size .
We have to make sure that is symmetric. This can be achieved by setting
and defining new regression parameters that are such that
which gives
and let and be the matrices containing the ’s and the ’s.
We can not be sure of being positive definite. One way to check whether the matrix is positive definite or not, is to compute the Cholesky decomposition of , that is, , and check whether all the diagonal elements of L are positive. If the matrix is invertible then it is more likely that it is positive definite, while if is not invertible then it is not positive definite. The estimated precision matrix, , is invertible if is invertible, where . Strictly diagonally dominant matrices are invertible. In general, the A, with elements , is strictly diagonally dominant if
The matrix is strictly diagonally dominant if
for all . Alternatively, is found for each to tune such taht it is strictly diagonally dominant, using