Numerically stable log-mean-exp

Used to Monte Carlo marginalize a class-conditional likelihood over a reference bank of kinetic-parameter particles or bundles:

Usage

mbmm_logmeanexp(v)

Arguments

v

Numeric vector of per-particle log-likelihoods.

Value

Scalar log-mean-exp.

Details

$$ \log\left\{ \frac{1}{M} \sum_{m=1}^{M}\exp(v_m) \right\}. $$

Importantly, -Inf values remain part of the denominator \(M\); they represent valid particles with zero likelihood and must not be discarded.