For a fixed token sequence scored by a model, cross-entropy is H and perplexity is defined as exp(H), using natural logarithms. If H decreases, what happens to perplexity?