In Bayesian analysis, a normalization constant often cancels out of the numerator and denominator of posterior calculations. Since most of the analysis is first done in logs, often a choice is...

In Bayesian analysis, a normalization constant often cancels out of the numerator and denominator of posterior calculations. Since most of the analysis is first done in logs, often a choice is available to compute exp(u − v) or exp(u)/exp(v). Which is better, or does it make a difference at all? Compare the two methods for u ≈ v with values such as ±1, ±5, ±10, ±100. Can you show analytically which one should be better?