SAMOSA

MCMC package — ML-MCMC functionalities (e.g. MLDA) will be added soon.

Package hierarchy

The samosa/ package is organized as follows:

• core/ — State (ChainState), model protocol, and proposal base classes: ProposalBase, AdaptiveProposal, TransportProposalBase. Also the transport map interface and MLMC utilities.

• kernels/ — Transition kernels: MetropolisHastingsKernel, DelayedRejectionKernel, plus SyncE and transport variants.

• proposals/ — Base proposals (e.g. GaussianRandomWalk), adapters (HaarioAdapter, GlobalAdapter), and wiring via AdaptiveProposal.

• samplers/ — SingleChainSampler (and coupled / MLDA samplers for multi-level and coupling).

• maps/ — Transport maps (LowerTriangularMap, RealNVPMap, LinearOptimalTransportMap) used with TransportProposalBase.

• utils/ — tools (e.g. laplace_approx, log_banana) and post_processing (load_samples, get_position_from_states, scatter_matrix, plot_trace, plot_lag).

Sampling strategy progression

Examples are ordered by increasing complexity:

1. Simple — Fixed proposal (e.g. Gaussian random walk), optionally tuned with a Laplace approximation at the MAP.

2. Adaptation — Same base proposal wrapped with an adapter (e.g. Haario or Global) so covariance or scale is adapted during the run.

3. Delayed rejection — After a rejected first stage, propose again (e.g. with scaled covariance) to improve acceptance.

4. Transport maps — Propose in a reference space (e.g. N(0,I)) and map back to the target; the map is often pre-adapted using samples from a previous run.