Graph algorithms expressed in functional languages often suffer from their inherited imperative, state-based style. In particular, this impedes formal program manipulation. We show how to model persistent graphs in functional languages by graph constructors. This provides a decompositional view of graphs which is very close to that of data types and leads to a "more fictional" formulation of graph algorithms. Graph constructors enable the definition of general fold operations for graphs. We present a promotion theorem for one of these folds that allows program fusion and the elimination of intermediate results. Fusion is not restricted to the elimination of tree-like structures, and we prove another theorem that facilitates the elimination of intermediate graphs. We describe an ML-implementation of persistent graphs which efficiently supports the presented fold operators. For example, depth-first-search expressed by a fold over a functional graph has the same complexity as the corresponding imperative algorithm.