Row-wise log-sum-exp

Compute the row-wise log-sum-exp of a numeric matrix in a numerically stable way. This is used to evaluate \(\log \sum_k \exp(x_{ik})\) without overflow or underflow by subtracting the row maximum before exponentiation.

Usage

row_log_sum_exp(mat)

Arguments

mat

A numeric matrix of size \(n \times K\).

Value

A numeric vector of length \(n\) where the i-th entry is \(\log \sum_{k=1}^K \exp(mat_{ik})\). Rows containing only -Inf return -Inf.