A Variational Solver for Belief Aggregation on Static Epistemic Graphs: Convex Equilibrium, Canonical Regularization, and Bounded-Delay Convergence
Year: 2026
Type: article
Abstract: We present a variational framework for aggregating continuous beliefs on static epistemic graphs. The framework defines a strictly convex energy functional that uniquely characterizes equilibrium beliefs under support, contradiction, and derivation dependencies. We extend prior formulations to admit directed cycles in the derivation subgraph through auxiliary consensus variables (θC ), thereby resolving the orphaned epistemic ambiguity of mutual derivation without sacrificing convexity. We prove linear convergence of projected cyclic Gauss–Seidel via the Ostrowski–Reich theorem, removing the need for restrictive diagonal dominance. For large-scale acceleration, we introduce a hybrid priority scheduler and establish formal bounded-delay convergence under the Bertsekas–Tsitsiklis asynchronous framework. All regularization hyperparameters are canonically derived from a Gaussian graphical model and calibrated offline via a practical grid-search surrogate; at runtime the energy functional remains parameter-fixed, ensuring deterministic behavior. Experimental validation on sparse scale-free networks demonstrates substantial acceleration over standard cyclic sweeps; we explicitly note that dense-graph regimes remain the domain of Krylovsubspace methods. The dynamic-graph and continuous-time extensions are identified as conjectural heuristics reserved for future work.
Author: Naing, Myo Sett
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225891601
