• CQuIC Seminars

August 27, 2026 3:30 PM - August 27, 2026 3:30 PM
PAIS 2540

Host:

Noah Lordi

Presenter:

Sam Alperin

Zoom Link

Abstract: Many important probabilistic systems can be represented by a factor graph, from frustrated magnets to classical and quantum codes. Such a model defines an exact statistical theory containing its full correlation structure, even though in practice one usually wants only simple local quantities such as one- and two-site marginals. Computing these quantities is the problem of inference, which is typically approached with approximate message-passing algorithms such as belief propagation and its variants. In this talk I will show that there is a more physical way to think about this problem. Specifically, I will prove that every finite closure of the exact statistical theory of a factor graph into an effective theory, together with a choice of coordinates, generates a canonical deterministic inference algorithm. The familiar relation between the Bethe free energy and belief propagation is the simplest example; response-corrected effective theories similarly generate TAP/Onsager-type inference. This turns approximate-inference algorithm design into a problem of effective-theory choice: rather than choosing a generic algorithm and tuning it to a problem, one can ask which tractable effective theory preserves the important qualitative structure of the exact model and then use the canonical algorithm generated by that theory. I will demonstrate this construction on frustrated finite Ising systems where standard belief propagation and higher-order Kikuchi methods see pathological failure and critical slowing, and discuss how the same viewpoint suggests a systematic route toward inference and decoding for quantum codes.

zoom password availible upon request, email nlordi AT unm.edu