Diagram Analysis of Iterative Algorithms

We study a general class of first-order iterative algorithms which includes power iteration, belief propagation and Approximate Message Passing (AMP), and many forms of gradient descent. When the input is a random matrix with i.i.d. entries, we present a new way to analyze these algorithms using combinatorial diagrams. Each diagram is a small graph, and the operations of the algorithm correspond to simple combinatorial operations on these graphs. We prove a fundamental property of the diagrams: asymptotically, we can discard all of the diagrams except for the trees. The mechanics of first-order algorithms simplify dramatically as the algorithmic operations have particularly simple and interpretable effects on the trees. We further show that the tree-shaped diagrams are essentially a basis of asymptotically independent Gaussian vectors. The tree approximation mirrors the assumption of the cavity method, a 40-year-old non-rigorous technique in statistical physics which has served as one of the most fundamental techniques in the field. We demonstrate the connection with the replica symmetric cavity method by "implementing" heuristic physics derivations into rigorous proofs. We rigorously establish that belief propagation is asymptotically equal to its associated AMP algorithm and we give a new simple proof of the state evolution formula for AMP. These results apply when the iterative algorithm runs for constantly many iterations. We then push the diagram analysis to a number of iterations that scales with the dimension n of the input matrix. We prove that for debiased power iteration, the tree diagram representation accurately describes the dynamic all the way up to n^Ω(1) iterations. We conjecture that this can be extended up to n^1/2 iterations but no further. Our proofs use straightforward combinatorial arguments akin to the trace method from random matrix theory.