0

Let $X$ represent an $N \times K$ matrix, $\Gamma$ a $K\times K$ matrix, and $E$ a $N\times K$ matrix. Let $Z$ be a matrix satisfying $ZZ^T = (I_N+X\Gamma X^T)$ My goal is to quickly compute the quantity $ZE$. Assume that: $K << N$.

In this case, I am primarily concerned with the speed of the actual matrix multiplication $ZE$ as $N$ is very large and naive solutions ends up having $Z$ involving the matrix multiplication of an $N\times N$ with an $N\times K$ matrix. I am not actually concerned with the time it takes to compute any matrix decompositions of $I_N+X\Gamma X^T$.

My feeling is that there must be some matrix identity that would allow me to reduce the dimension of these operations by exploiting the fact that $X\Gamma X^T$ is low rank and that multiplication by the identity matrix leaves a quantity unchanged (essentially allowing me to ignore that multiplication).

If need be I am ok with taking a low-rank approximation of the quantity $ZE$.

Any ideas?

jds
  • 237
  • 1
  • 8
  • I feel like this https://math.stackexchange.com/questions/1747812/inverse-of-the-sum-of-a-invertible-matrix-with-known-cholesky-decomposion-and-di is particularly relevant but I can't figure out how use this. – jds May 24 '18 at 14:43
  • You must have some restrictions on $X$ and $\Gamma$ in order to get the Cholesky factor $Z$, no? – jjjjjj May 24 '18 at 15:45
  • Yes! Gamma is symmetric positive definite. X is a matrix of covariates so it’s safe to assume it is at least rank 1. – jds May 25 '18 at 16:46

0 Answers0