bayespy.utils.linalg.block_banded_solve(A, B, y)[source]

Invert symmetric, banded, positive-definite matrix.

A contains the diagonal blocks.

B contains the superdiagonal blocks (their transposes are the subdiagonal blocks).

Shapes: A: (…, N, D, D) B: (…, N-1, D, D) y: (…, N, D)

The algorithm is basically LU decomposition.

Computes only the diagonal and super-diagonal blocks of the inverse. The true inverse is dense, in general.

Assume each block has the same size.

Return: * inverse blocks * solution to the system * log-determinant