For those interested, this question arises because in markov random fields one encounteres a lot of algorithms which look vor variables that minimize the potential of multiple cliques. So if a variable $X$ gives the clique potentials $a$ and $b$ these algorithms would choose this over a variable $Y$ that gives clique potentials $c$ and $d$, if $a+b<c+d$.
From a probabilistic point of view, it would make more sense to choose $Y$ if $ {\rm e}^{-a} + {\rm e}^{-b} <{\rm e}^{-c} + {\rm e}^{-d} $