Bayesian Basin Tracking: Efficient Global Continuation of Multistable Dynamical Systems
Mapping the global phase space of high-dimensional multistable dynamical systems is computationally prohibitive because conventional approaches require extensive numerical integration. Here, we introduce Bayesian Basin Tracking (BBT), an adaptive Bayesian framework that exploits the persistence of basin boundaries under parameter continuation to reconstruct global phase-space structure using only a fraction of the simulations required by conventional methods. By modeling the probability that a sampled initial condition converges to a particular attractor, the method represents the phase-space geometry established at a given parameter value through a Dirichlet-multinomial model. At a nearby parameter value, these probabilities are estimated by updating the prior distribution with newly sampled data. To detect boundary crises and bifurcations autonomously, we use the log Bayes factor as an information-theoretic sensor that triggers dense resampling only when structural changes render the historical prior statistically implausible. We validate the framework using the discrete H\'enon map, the continuous-time Duffing oscillator, and a 300-dimensional network of coupled R\"ossler oscillators. BBT overcomes the restrictive dimensional scaling of deterministic grid tessellations by concentrating the most computationally demanding calculations in structurally volatile regions. In high-dimensional synchronization landscapes, it achieves an almost sixfold computational speed-up while retaining theoretically derived error bounds.