Phase Transitions – like death and taxes? Why we should care; what to do about it. Scott Kirkpatrick, Hebrew University, Jerusalem With thanks to: Uri Gordon, Erik Aurell, Johannes Schneider,…
Phase transitions are inevitable in Avogadro- scale engineering. System properties go from “good” to “not good enough” at the extremes of their parameter space What happens in between these phases? We have limited tools for understanding such transitions. Physics of disordered materials: Sharp vs. smeared – Harris criterion Glass transitions Combinatorics on large scales Sharp property crossovers – Friedgut’s theorem We should care because computing is HARD at phase boundaries.
SAT and 3-SAT as classic test cases Parameters: N variables, M constraining clauses, M/N constant For 3-SAT, phase diagram is known: M/N< 3.9“easy,” satisfiable (probably P) 3.9 < M/N < 4.27“hard,” but still satisfiable 4.27… < M/Nunsatisfiable, exponential cost Recent advances, applying message-passing, pushed boundary of solubility in the “hard-SAT” region from N = 300 to N = 10^7. General technique – soften the variables into beliefs or surveys
Surveys and Beliefs for the SAT problem Beliefs – probabilities that the spin is up or down Avg. over satisfying configurations, as estimated by the local tree Surveys – probabilities that the spin is up or down In all sat configurations. This leaves a third possibility – spins that do both at different times. Equations for both are nearly identical. We can define hybrid methods, and they prove useful. Use beliefs or surveys to guide decimation – this solves problem.
Propagating surveys or beliefs in a “cavity”
Depth of decimation characterizes BP, SP and mixed-P
Study how SP, BP, m-P evolve by analyzing their hydrodynamics Use movies of N = 100,000 Look at three cases, SP, BP, and the hybrid-P: (caution, large files)