Package jdistlib.inference
package jdistlib.inference
-
ClassDescriptionDevice-resident batched logistic-regression posterior with a spherical normal prior.Warmup-only candidate-specific Gaussian birth adaptation with checkpointable moments.Rebuilds and draws a caller-certified log-concave full conditional.Per-model isotropic random walk with warmup-only Robbins-Monro scale adaptation.Coordinated many-chain static HMC with ChEES or SNAPER trajectory adaptation.Controls coordinated ChEES/SNAPER trajectory-length adaptation across chains.Draws plus the shared trajectory adaptation selected by ChEES or SNAPER.Automatic adjusted-MCLMC pilot search for integrator step and decorrelation length.Pilot-search controls for adjusted MCLMC step size and decorrelation length.Final adjusted-MCLMC chain and auditable pilot objective values.Metropolis-adjusted microcanonical Langevin sampler (MHMCHMC).Exact gradient-informed Barker proposal with symmetric Gaussian magnitudes.Optional accelerator-facing contract for evaluating independent states in one call.Compiled named model evaluated on an unconstrained state space.In-memory state-and-stream restart point including a cloned random engine.Versioned JSON and tidy CSV interchange for retained chain draws.Immutable retained samples, sampler statistics, adaptation and restart state.Deterministic multi-chain execution and checkpoint continuation helpers.Chart-neutral immutable dataset suitable for SVG, JSON, CSV, or UI adapters.Checksummed, versioned binary checkpoint envelope with compatibility fingerprints.Selected-coordinate streaming sink with independently compressed, recoverable chunks.Selected-coordinate columns read from a chunked draw store.Gaussian Metropolis sweeps with independently adapted coordinate scales.A differentiable target whose numerical evaluation is bound to a compute backend.Controls whether NUTS target evaluation may or must use an accelerator.Standard parameter constraints and their Jacobian-aware transforms.Gaussian random-walk update of a continuous block conditional on all other coordinates.Unit-Jacobian birth/death mapping that inserts or removes one parameter coordinate.Split/merge map x,u to x+u,x-u with forward log-Jacobian log(2).Typed support for one coordinate in a mixed continuous/discrete state.Factories for trace, rank, autocorrelation, energy, and pair-plot datasets.A log density capable of adding its gradient to caller-owned storage.A factor that adds derivatives with respect to all constrained coordinates.One side of a dimension-matching map: state plus complementary auxiliaries.Reversible mapping between parameter/auxiliary pairs of equal total dimension.Executable round-trip, dimension, and reciprocal-Jacobian validation for RJ maps.Symmetric Metropolis update for one bounded or unbounded discrete coordinate.Location and energy error for one divergent retained transition.Extracts sampler pathologies with coordinates suitable for plotting.Streaming destination for retained draws; implementations must copy if needed later.Tuning-free elliptical slice sampler; target is the likelihood-only log density.Thread-confined log-density wrapper reporting value and gradient work.Immutable factor timing and numerical-stability snapshot.Wraps factors without changing whether analytic gradients are available.Immutable factor metadata and evaluator.Exact full-conditional update for one finite integer or categorical coordinate.First-class inference result combining chains, diagnostics, and provenance.Adapts one frozen ordinary JDistlib sampler transition for use within an RJ model.Gaussian prior/reference measure used by elliptical slice and pCN updates.Candidate-specific or common Gaussian birth proposal.Independent Gaussian sparse-coefficient birth proposal.Named scalar generated from one retained unconstrained state.Appends generated quantities to each state before forwarding it to another sink.Coordinate associated with divergences and a scale/reparameterization suggestion.Ranks coordinates whose divergent and non-divergent locations are most separated.One exact or MCMC-within-Gibbs state update.Composes exact, adaptive-rejection, Metropolis, or blocked Gibbs kernels.Immutable comparison between supplied and finite-difference gradients.Reports whether a differentiable target uses analytic rather than fallback gradients.Finite-difference adapters and gradient validation utilities.Fixed-trajectory HMC with dual-averaged step size and covariance adaptation.Diagnostic code, quantitative evidence, and an actionable remediation.Machine-readable inference health severity.One support-aware transition in a scheduled hybrid sampler.Outcome of one hybrid-kernel update.Scheduled support-aware sampler for fixed-dimensional mixed continuous/discrete targets.Per-kernel acceptance and support diagnostics for a hybrid schedule.Retained mixed-state chain and diagnostics from its scheduled kernels.Concise facade for reproducible multi-chain fitting.Parses reusable compute switches for command-line applications embedding JDistlib.Dependency-free JSON, tidy CSV, and SVG adapters for chart-neutral data.Actionable policy over sampler and posterior diagnostics.Self-contained headless HTML report composed from diagnostic data and SVGs.Deterministic validation, constrained initialization, and bounded retry helpers.Per-iteration sampler statistics used by convergence diagnostics.Immutable result of one reusable Markov transition.Small deterministic L-BFGS maximizer intended for initialization and MAP fits.An unnormalized log density on an unconstrained Euclidean state space.Pointwise, paired comparison of models evaluated with PSIS-LOO.Deterministic many-short-chain execution using common-initialization superchains.Results and nested convergence diagnostics from a superchain design.Fixed-capacity selected-coordinate draw sink and zero-copy memory-mapped reader.Immutable parameter and sampler diagnostics with machine-readable output.Rank-normalized R-hat, bulk/tail ESS, MCSE, and sampler diagnostics.Dependency-free versioned JSON serialization for inference diagnostics.Immutable Euclidean metric selection for HMC-family samplers.Metropolis-adjusted Langevin sampler with dual-averaged proposal scale.Gaussian random-walk update for a declared continuous or discrete state block.Coordinate-by-coordinate support declaration for hybrid MCMC.Fluent builder for named constrained parameters and model factors.Immutable named numeric data supplied to a model.Cached per-factor values for proposal algorithms that change few coordinates.Non-thread-safe allocation-free evaluator intended for one sampler chain.One prior, likelihood, or constraint contribution to a model log density.Analytic common priors and likelihood factors for the programmatic builder.Immutable bipartite parameter/factor graph for inspection and rendering.Graphviz DOT and versioned JSON export for model dependency graphs.Dispatches to a distinct within-model kernel for each declared model identifier.Read-only named view of one constrained model state and its observed data.MCSE and efficiency helpers for one stationary retained sequence.Multinomial-candidate NUTS with windowed adaptation and configurable metrics.Immutable names and grouping labels for pointwise likelihood contributions.Immutable numerical-optimization result.Accepted L-BFGS path used by Pathfinder approximation selection.Replica exchange using random-walk within-temperature transitions.Cold-chain draws and adjacent-temperature swap diagnostics.A differentiable map from unconstrained coordinates to constrained values.Posterior summary and modern multi-chain convergence diagnostics.Immutable parameter metadata in a compiled Bayesian model.Pareto-tail smoothing for log importance ratios, with a diagnostic shape estimate.Multi-path quasi-Newton Gaussian approximation with mixture scoring and PSIS resampling.Multi-path Gaussian approximation, PSIS diagnostic, and resampled draws.L-BFGS path plus local Gaussian draw for robust chain initialization.Immutable controls for multi-path Pathfinder initialization and sampling.Quasi-Newton Gaussian initialization result.Model contract required by pointwise predictive assessment.Pointwise log likelihoods with retained chain boundaries and observation metadata.Evaluates ordered observation-level log-likelihood contributions.Checkpoint plus the fingerprints and platform metadata needed to validate a resume.Restored RJ checkpoint plus the fingerprints and platform recorded with it.Restored sparse checkpoint plus its model/options fingerprints and platform.Chunked deterministic continuation based on MCSE, never on R-hat alone.Deterministically extended chain and the reason continuation stopped.Stopping goal for one posterior coordinate, guarded by minimum draws and chunk count.Optimizes simplex weights for stacking pointwise out-of-sample predictions.Model factor instrumented with low-overhead call, time, and non-finite counters.Lightweight callback invoked after a completed sampler transition.Forward projection-predictive variable selection for Gaussian linear reference models.Pareto-smoothed importance-sampling leave-one-out cross-validation.Optional exact or refitted LOO calculation used when importance sampling is unreliable.Reusable Gaussian random-walk Metropolis transition kernel.Isotropic Gaussian random-walk Metropolis with warmup scale adaptation.Locally informed candidate proposal using a prepared
X'vproduct.Sampler capable of restoring algorithm-specific adaptive state.Deterministic parallel execution with one independently adaptive RJ sampler per chain.Exact in-memory RJ restart point including ragged state, schedule, adaptation, and RNG.Checksummed portable persistence for complete reversible-jump restart state.Model occupancy, movement, inclusion, conditional-parameter, and reliability diagnostics.Multi-chain diagnostics for ragged reversible-jump output.Streaming sink for ragged retained RJ draws.Tidy CSV export for ragged model-specific parameters.Per-iteration within-model and trans-dimensional transition statistics.Named parameter schema for one model in a trans-dimensional target.One dimension-changing or structure-changing reversible proposal.Posterior summary conditional on a ragged parameter being present.Progress callback for one trans-dimensional chain.Transactional RJ proposal including all non-schedule Hastings terms.Immutable ragged RJMCMC draws, transition statistics, diagnostics inputs, and restart state.General Java-only RJMCMC acceptance engine with warmup-frozen move and within-model adaptation.Creates one independently mutable reversible-jump sampler per chain.Warmup, retention, schedule-adaptation, and streaming options for RJMCMC.Immutable model identifier and its dimension-specific parameter vector.Complete normalized joint density and schemas across a family of models.A fixed-model update scheduled between trans-dimensional proposals.Outcome of one within-model transition in an RJ schedule.Auxiliary-variable proposal used to create a newly active scalar parameter.Immutable provenance record for reproducing one inference run.Common contract for a reproducible MCMC chain.Versioned sampler-specific adaptation state for exact resumability.Cross-chain sampler health summary.Immutable common MCMC warmup, retention, adaptation, and safety options.Ranks posterior coefficient draws by practical-significance probability.Seeded simulation-based calibration rank utility.Coordinate-wise stepping-out and shrinkage slice sampler.Candidate selected by a normalized sparse birth proposal.Normalized proposal over candidates inactive in the conditioning model.Dimension-matching proposal for a coefficient born into a sparse model.Supplies the row score vector used by a locally informed sparse proposal.Exact sparse RJ restart state, adaptation, counters, RNG, and online summaries.Checksummed, forced, atomic persistence for complete sparse RJ restart state.Streaming callback for retained sparse draws.Crash-safe segment export for ragged sparse draws.Per-transition statistics for sparse subset RJMCMC.Complete normalized log joint for an arbitrary sparse candidate universe.Progress callback for one restartable sparse RJMCMC segment.One bounded sparse RJ segment plus its exact continuation checkpoint.Allocation-conscious add/drop/swap RJMCMC for very large sparse candidate universes.Global warmup target and bounded transition segment for restartable sparse RJMCMC.Immutable sparse subset, common parameters, and active coefficients.Online occupancy and conditional-coefficient summary stored in a sparse checkpoint.Sparse subset target with an arbitrary candidate count and a bounded active set.Adds one uniformly selected inactive candidate using a declared birth proposal.Drops one uniformly selected active candidate with the matching reverse birth density.Complete normalized log joint for one active-variable subset.Factory for the standard add/drop/swap subset-selection RJ schedule.Bit-mask model family for Java-only covariate, locus, or feature selection.Exchanges one active and inactive candidate while preserving model dimension.Common-start grouping required to interpret nested R-hat.Log density split into an untempered base (usually the prior) and likelihood.Reusable one-step Markov kernel used by chains and meta-samplers.Uniform proposal over currently inactive sparse candidates.Widely applicable information criterion from pointwise log-likelihood draws.Fingerprinted metric/step-size warmup reuse, treated as an initial guess by default.Immutable summary of MCMC adaptation.Stan-style fast/slow/final warmup schedule with expanding metric windows.Progress listener retaining lightweight step-size and schedule traces.