NEWS

---

bpgmm 1.3.4

• Implemented model 3 (CUC) with the corrected $\Psi_k$ rate term $C_{\tilde{Y}\tilde{Y}k} + A_k/m$, consistent with the other shared-loading models and the model-3 log posterior in Lu, Li, and Love (2021, Supplement A.2).
• Fixed the covariance-structure RJMCMC loading proposal for models UCC, UCU, UUC, and UUU so the proposal covariance uses the per-cluster precision $(C_{YYk} + \alpha_2 I_{q_k})^{-1}$ rather than a pooled inverse.
• Fixed the $\alpha_2$ hyperparameter Gibbs update so the Gamma shape uses $q p/2 + d_{\alpha_2}$ with $q = \sum_k q_k$ computed from the active loading matrices, rather than treating a single q_new value as the total latent dimension.
• Expanded the vignette set with the corrected Gibbs updates, hyperparameter formulas, eight PGMM covariance structures, and the reversible split/combine move.

bpgmm 1.3.3

• Made the cluster-mean prior and its Gibbs update consistent. pgmm_rjmcmc() now centers the data internally so the cluster-mean prior mean xbar is 0 (the value implied by the augmented loading posterior in Lu, Li, and Love (2021, Supplement A.1)), matching the zero-mean conditional already used by the sampler. Previously the prior mean was set to a single random observation while the update used 0, so the two were inconsistent. Sampled means are returned on the original (uncentered) data scale.
• Corrected the split/combine acceptance ratio. The split proposes one independent sign per coordinate for the new component means, so its proposal density carries a 2^-p factor (p = nrow(X)) that evaluate_split_clusters() had dropped. This factor is required for the split/combine move to be reversible when the two merged components are not uniformly ordered across coordinates. The corresponding deterministic split in the paper omits this sign randomization; see the package errata notes.
• Fixed an error (q_vec entries must be positive integers) that aborted the cluster-number RJMCMC (m_step = 1, and especially split_combine = 1) whenever the active clusters became non-contiguous, for example after an interior empty cluster was removed. The prior evaluators now compact to the active clusters, the native latent-score update validates only active clusters, and the combine evaluator uses the merged cluster's own factor count.

bpgmm 1.3.2

• Fixed the beta hyperparameter Gibbs update so the Gamma rate uses the rate hyperprior s_vec[3] (s_beta) instead of the shape hyperprior d_vec[3], matching the conditional posterior in Lu, Li, and Love (2021, Appendix A.1.3). Results are unchanged under the default symmetric hyperpriors d_vec = s_vec = c(1, 1, 1) but are corrected for user-supplied asymmetric beta hyperpriors.
• Fixed the native latent-score update so unallocated observations draw from N(0, I_q) as specified in the paper's conditional for the scores, rather than from the component score covariance. These scores are not used in any downstream sufficient statistic, so posterior results are unaffected.

bpgmm 1.3.1 (2026-05-28)

• Revised the vignette set to reduce repeated examples across vignettes.
• Made the model-and-sampler, data-preparation, model-selection, variable-prioritization, and diagnostics vignettes more formula-focused and closer to the notation in Lu, Li, and Love (2021).
• Changed the variable-prioritization simulation so it no longer duplicates the larger MFA model-selection example.

bpgmm 1.3.0

• Standardized the native Rcpp interface to snake_case function names.
• Renamed C++ source files to snake_case and removed stale compiled artifacts from src/.
• Cleaned internal native wrapper documentation while preserving sampler behavior.

bpgmm 1.2.9

• Added data-preparation and posterior-diagnostics vignettes.
• Expanded the vignette collection so each vignette covers a distinct analysis task: input preparation, sampler details, small worked examples, larger model selection, exploratory variable prioritization, and multi-chain diagnostics.

bpgmm 1.2.8

• Added larger model-selection and exploratory variable-prioritization vignettes with runnable examples and plots.
• Documented how to use posterior allocations and loading matrices for exploratory variable prioritization without presenting it as formal Bayesian variable selection.

bpgmm 1.2.7

• Moved latent-factor score updates from R loops to native C++.
• Moved prior density evaluation for lambda and psi to native C++ while preserving the existing internal R API.
• Added strict native tests for latent-score sampling, input validation, and closed-form prior density checks.

bpgmm 1.2.6

• Added pgmm_rjmcmc_chains() for CRAN-safe multi-core execution of independent RJMCMC chains.
• Added deterministic per-chain seeding and unit tests for the independent-chain wrapper.

bpgmm 1.2.5

• Cleaned native C++ readability without changing sampler behavior.
• Replaced stale include guards and prototypes, removed selected namespace pollution, and clarified covariance-constraint branch checks with model-label comments.
• Regenerated Rcpp exports after tightening native helper signatures.

bpgmm 1.2.4

• Optimized the native allocation sampler by precomputing component covariance factorizations and sampling from normalized log probabilities directly.
• Removed avoidable dense probability/log-density matrices and repeated per-observation covariance validation in update_PostZ().
• Tightened native helper signatures to avoid unnecessary copies and removed stale debug-only C++ comments from the posterior lambda/psi update.

bpgmm 1.2.3

• Added CRAN-safe base R plots to the getting-started, worked-examples, and model-and-sampler vignettes.
• Visualized simulated clusters, posterior modal allocations, posterior model counts, and mixture-of-factor-analyzers covariance geometry.

bpgmm 1.2.2

• Optimized the native Calculate_Cxy() helper by accumulating cluster sufficient statistics directly from labels instead of building a dense allocation matrix and multiplying through zero weights.
• Removed repeated temporary vector construction inside the Cxy accumulation loop.
• Preserved the existing C++11 baseline for CRAN portability.

bpgmm 1.2.1

• Hardened Rcpp entry points with explicit validation for dimensions, finite numeric inputs, covariance constraints, and native hyperparameter vectors.
• Kept native build settings on the existing CRAN-friendly C++11 baseline while modernizing validation and headers.
• Cleaned native headers and removed stale C++ declarations and debug-only includes/comments.
• Expanded native unit tests for invalid Cxy, posterior lambda/psi, and hyperparameter-update inputs.

bpgmm 1.2.0

• Made the public API consistently snake_case now that the package has no CRAN reverse dependencies.
• Removed legacy exported wrappers pgmmRJMCMC(), summarizePgmmRJMCMC(), and summerizePgmmRJMCMC().
• Renamed public sampler arguments from camelCase to snake_case: m_init, m_range, q_new, d_vec, s_vec, m_step, v_step, and split_combine.
• Renamed summary arguments and result fields to snake_case, including true_cluster, allocation, n_clusters, n_constraints, allocation_samples, and constraint_samples.
• Updated README, pkgdown vignettes, examples, and tests for the breaking 1.2.0 API.

bpgmm 1.1.6

• Added a model-and-sampler vignette that explains the mixture-of-factor-analyzers likelihood, PGMM covariance constraints, conjugate priors, allocation update, RJMCMC moves, and package output fields with formulas and runnable examples.
• Linked the new mathematical vignette from the getting-started guide and pkgdown vignette index.

bpgmm 1.1.5

• Fixed the allocation prior contribution in RJMCMC acceptance calculations to use the log product of allocated mixture weights, matching the paper's joint posterior.

bpgmm 1.1.4

• Standardized internal R helper names to snake_case while preserving the exported compatibility wrappers.
• Added snake_case wrappers around generated Rcpp entry points and routed package internals through those wrappers.
• Cleaned the source layout by renaming R files to lowercase hyphenated names and removing the ignored duplicate R/choosem/ tree.
• Kept result-list names and legacy public arguments stable to avoid breaking existing user code.

bpgmm 1.1.3

• Fixed Mstep = 1 by loading the cluster-number proposal helpers as package code rather than leaving them in an ignored nested R/ directory.
• Fixed split/combine RJMCMC moves by passing X explicitly to split-allocation helpers and adding the missing latent-factor update helper.
• Fixed the C++ allocation update to add log(tao) to component log densities instead of adding raw mixture weights.
• Added C++ input validation for multivariate normal densities, log-ratio calculations, allocation dimensions, and mixture weights.
• Expanded unit coverage for all eight PGMM covariance constraints, C++ helpers, cluster-number RJMCMC, covariance RJMCMC, and split/combine moves.

bpgmm 1.1.2

• Modernized README with badges, installation guidance, paper citation, and model-constraint helper examples.
• Added summarizePgmmRJMCMC() as the correctly spelled summary function and retained summerizePgmmRJMCMC() for backward compatibility.
• Added tidyverse-style pgmm_rjmcmc() and summarize_pgmm_rjmcmc() as the preferred public API.
• Deprecated pgmmRJMCMC(), summarizePgmmRJMCMC(), and summerizePgmmRJMCMC(); these names remain available as compatibility wrappers.
• Added helpers model_to_constraint() and constraint_to_model() for translating between paper model labels and legacy constraint vectors.
• Improved package startup citation guidance for users publishing results from bpgmm.
• Added unit tests for the public API, covariance-constraint mapping, summary helpers, and native C++ wrappers.
• Fixed zero-iteration handling in pgmm_rjmcmc() and added validation for sampler inputs and summary result objects.
• Added a verbose argument to suppress per-iteration progress output in examples, tests, and scripted workflows.
• Updated C++ build settings and validation for current R/Rcpp best practices.