Memory limits affect speed in changing graph models

Limited workspace requires polynomial update times for connectivity in a new dynamic model.

A new model for dynamic algorithms uses a bounded-space workspace. The algorithm has read-only oracle access to a changing input.

The study looks at dynamic undirected ss-tt connectivity. This means reporting if ss and tt are connected after every edge update to an nn-vertex graph.

The graph is accessed via adjacency list queries or adjacency matrix queries. An adjacency list query uses degree and the jj-th neighbor. An adjacency matrix query is a different way to access the graph.

With adjacency lists, incremental algorithms need Θ~(n2/S2)\tilde{\Theta}(n^2/S^2) worst-case time per insertion. With an adjacency matrix, Ω~(n2/S)\tilde{\Omega}(n^2/S) time is necessary for incremental algorithms.

Fully dynamic algorithms need Ω~(n/S)\tilde{\Omega}(n/S) amortized time per update with adjacency lists. They need Ω~(n2/S)\tilde{\Omega}(n^2/S) with an adjacency matrix, even on forests.

Polynomial update time is necessary whenever S≤n1−εS \le n^{1-\varepsilon}. This holds even for a single pair of vertices and even for incremental algorithms.

Key facts

  • Incremental algorithms with adjacency lists need Θ~(n2/S2)\tilde{\Theta}(n^2/S^2) worst-case time per insertion.
  • For adjacency matrices, Ω~(n2/S)\tilde{\Omega}(n^2/S) time is necessary for incremental algorithms.
  • Fully dynamic algorithms need Ω~(n/S)\tilde{\Omega}(n/S) amortized time per update with adjacency lists.
  • Polynomial update time is necessary when S≤n1−εS \le n^{1-\varepsilon}.

All sections

Edition