Skip to content

Scalable correlated inference using LatentKron

Anarya Ray requested to merge scalable_inference into main

Created by: AnaryaRay1

For m1m2z inference, the GP's covariance matrix is a Kroneker product of individual matrices corresponding to m1m2 and z. We can thus exploit the fact that the Cholesky decomposition of a Kroneker product is the Kroneker product of individual Cholesky decompositions. This changes the computational scalability of the problem O(\prod_{\theta}n_{bins,\theta}^3) to O(\sum_{\theta} n_{bins,\theta}^3). This is implemented by using Pymc GP's LatentKron instead of Latent.

Merge request reports