Glossary¶

Obj, Cool, Neigh, Move, Accept: The five typed components.

L1-L4: The four composition laws.

quasi-Monte Carlo (QMC): Quasi-Monte Carlo (low-discrepancy sequences, here used for pilot starts and polish multistart).
generalized Langevin equation (GLE): Generalized Langevin equation (colored-noise thermostat with optimal-sampling drift).
BSA / FSA / generalized simulated annealing (GSA): Boltzmann, Fast, and Generalized (Tsallis) simulated annealing.
maximum a posteriori (MAP): Maximum a posteriori (the point estimate returned by the Bayesian pilot).
Thompson sampling: Draw from each arm’s posterior and pick the max (used by the mixer on the per-chain Beta posteriors).
BAB (B-A-B) propagator: The B-A-B splitting integrator used by the generalized Langevin equation (GLE) move.
charged work: Objective and gradient evaluations recorded in a monotone

per-replica ledger. Aggregate ensemble work is the sum of unique replica charges, independent of how many processor elements execute them concurrently.

basin census: Immutable descriptor medoids, memberships, and exact visit

counts used for scarcity and unseen-mass estimates. Active-catalogue eviction does not change the census.

descriptor catalogue: A bounded communication set of validated low-energy,

well-separated representatives. A receiving replica freshly evaluates a representative before it can affect the chain or encounter record.

synchronous Feynman–Kac population: A fixed-size interacting population

formed when every replica submits one validated representative. Rank-based weights, systematic parent assignment, a family cap, and sibling rejuvenation divide search effort without using target morphology or reference energy.

diffusion Monte Carlo (DMC) analogy: The cooperative population reuses

walker weighting, resampling, and genealogy as algorithmic bookkeeping. It does not implement imaginary-time Green-function propagation, a trial wavefunction, fixed-node physics, or a quantum-energy estimator.

Gaussian Markov random field (GMRF): The latent Gaussian residual field on

the observed basin-transition graph. Posterior marginal variance supplies the cross-replica uncertainty rank.

integrated nested Laplace approximation (INLA): A broader latent-Gaussian

inference construction. The catalogue field needs only an exact Gaussian conditional solve, so it does not perform Laplace approximation or hyperparameter integration and is not a full INLA implementation.

nested sampling: A separate driver with a descending ceiling and

live-point volume weights. It remains a matched-budget comparison because those weights cannot be merged with Feynman–Kac genealogy without changing the meaning of both methods.