Low-Rank Solution for Nonlinear Optimization over Graphs Javad Lavaei Department of Electrical Engineering Columbia University Acknowledgements Joint work with Somayeh Sojoudi (Caltech): S. Sojoudi and J. Lavaei, "Semidefinite Relaxation for Nonlinear Optimization over Graphs," Working draft, 2012. S. Sojoudi and J. Lavaei, "Convexification of Generalized Network Flow Problem," Working draft, 2012. Javad Lavaei, Columbia University 2 Problem of Interest Abstract optimizations are NP-hard in the worst case. Real-world optimizations are highly structured: Sparsity: Non-trivial structure: Question: How does the physical structure affect tractability of an optimization? Javad Lavaei, Columbia University 3 Example 1 Trick: SDP relaxation: Guaranteed rank-1 solution! Javad Lavaei, Columbia University 4 Example 1 Opt: Sufficient condition for exactness: Sign definite sets. What if the condition is not satisfied? Rank-2 W (but hidden) NP-hard Javad Lavaei, Columbia University 5 Example 2 Opt: Acyclic Graph Real-valued case: Rank-2 W (need regularization) Complex-valued case: Real coefficients: Exact SDP Imaginary coefficients: Exact SDP General case: Need sign definite sets Javad Lavaei, Columbia University 6 Sign Definite Set Real-valued case: “T “ is sign definite if its elements are all negative or all positive. Complex-valued case: “T “ is sign definite if T and –T are separable in R2: Javad Lavaei, Columbia University 7 Formal Definition: Optimization over Graph Optimization of interest: (real or complex) Define: SDP relaxation for y and z (replace xx* with W) . f (y , z) is increasing in z (no convexity assumption). Generalized weighted graph: weight set Javad Lavaei, Columbia University for edge (i,j). 8 Real-Valued Optimization Edge Cycle Javad Lavaei, Columbia University 9 Real-Valued Optimization Exact SDP relaxation: Acyclic graph: sign definite sets Bipartite graph: positive weight sets Arbitrary graph: negative weight sets Interplay between topology and edge signs Javad Lavaei, Columbia University 10 Low-Rank Solution Violate edge condition: Satisfy edge condition but violate cycle condition : Javad Lavaei, Columbia University 11 Computational Complexity: Acyclic Graph Number partitioning problem: Javad Lavaei, Columbia University ? 12 Complex-Valued Optimization Main requirement in complex case: Sign definite weight sets SDP relaxation for acyclic graphs: real coefficients 1-2 element sets (power grid: ~10 elements) Javad Lavaei, Columbia University 13 Complex-Valued Optimization Purely imaginary weights (lossless power grid): Consider a real matrix M: Polynomial-time solvable for weakly-cyclic bipartite graphs. Javad Lavaei, Columbia University 14 Graph Decomposition There are at least four good structural graphs. Acyclic combination of them leads to exact SDP relaxation. Opt: Sufficient conditions for {c12 , c23 , c13 }: Real with negative product Complex with one zero element Purely imaginary Javad Lavaei, Columbia University 15 Resource Allocation: Optimal Power Flow (OPF) Voltage V Current I Complex power = VI*=P + Q i OPF: Given constant-power loads, find optimal P’s subject to: Demand constraints Constraints on V’s, P’s, and Q’s. Javad Lavaei, Columbia University 16 Optimal Power Flow Cost Operation Flow Balance Express the last constraint as an inequality. Javad Lavaei, Columbia University 17 Exact Convex Relaxation OPF: DC or AC Networks: Distribution or transmission Result 1: Exact relaxation for DC/AC distribution and DC transmission. Energy-related optimization: Javad JavadLavaei, Lavaei,Columbia Stanford University University 17 18 Exact Convex Relaxation Each weight set has about 10 elements. Due to passivity, they are all in the left-half plane. Coefficients: Modes of a stable system. Weight sets are sign definite. Javad JavadLavaei, Lavaei,Columbia Stanford University University 17 19 Generalized Network Flow (GNF) injections flows limits Goal: Assumption: • fi(pi): convex and increasing • fij(pij): convex and decreasing Javad Lavaei, Columbia University 20 Convexification of GNF Feasible set without box constraint: Monotonic Non-monotonic Convexification: It finds correct injection vector but not necessarily correct flow vector. Javad Lavaei, Columbia University 21 Convexification of GNF Feasible set without box constraint: Correct injections in the feasible case. Why monotonic flow functions? Javad Lavaei, Columbia University 22 Conclusions Motivation: Real-world optimizations are highly structured. Goal: Develop theory of optimization over graph Mapped the structure of an optimization into a generalized weighted graph Obtained various classes of polynomial-time solvable optimizations Talked about Generalized Network Flow Passivity in power systems made optimizations easier Javad Lavaei, Columbia University 23