Average case analysis of DJ graphs

Sreedhar et al. V.C. Sreedhar, G.R. Gao, Y.-F. Lee, A new framework for elimination-based data flow analysis using DJ graphs, ACM Trans. Program. Lang. Syst. 20 (2) (1998) 388–435; V.C. Sreedhar, Efficient program analysis using DJ graphs, PhD thesis, School of Computer Science, McGill University, Montréal, Québec, Canada, 1995] have presented an elimination-based algorithm to solve data flow problems. A thorough analysis of the algorithm shows that the worst-case performance is at least quadratic in the number of nodes of the underlying graph. In contrast, Sreedhar reports a linear time behavior based on some practical applications.

In this paper we prove that for goto-free programs, the average case behavior is indeed linear. As a byproduct our result also applies to the average size of the so-called dominance frontier.

A thorough average case analysis based on a graph grammar is performed by studying properties of the j-edges in DJ graphs. It appears that this is the first time that a graph grammar is used in order to analyze an algorithm. The average linear time of the algorithm is obtained by classic techniques in the analysis of algorithms and data structures such as singularity analysis of generating functions and transfer lemmas.