On the sparsity of fitness functions and implications for learning