NEWS

spexvb 0.1.0.9001

spexvb() now returns sigma, the slab posterior standard deviation of each coefficient (length p). The marginal posterior SD of beta_j is sqrt(omega*sigma^2 + omega*(1 - omega)*mu^2).

get.initials() now respects a user-supplied tau_e instead of always re-estimating it, so a known error precision is held fixed (as the documentation already stated). The algorithm is unchanged.

Hoisted iteration-constant computations (sigma, log-odds, sqrt(tau_b * tau_e)) out of coordinate-ascent inner loops in all three C++ backends (fit_linear_alpha_remap, fit_linear_gamma_hierarchy, fit_logistic_alpha_remap).
Replaced O(np) post-remapping matrix-vector product (W = X * approx_mean) with O(n) scalar multiply (W *= alpha).
Vectorized entropy(), replaced std::pow(x, 2) with x * x, reused precomputed X_2 in the logistic backend, and removed dead allocations.
R side: direct C++ function calls instead of do.call() string lookup, eliminated redundant lapply copy in return list, reduced standardization from two scale() passes to one via Var = E[X^2] - E[X]^2.
Measured 24--33% speedup on moderate to high-p settings with identical coefficients (within tolerance 1e-8) and zero iteration count changes.

Initial CRAN release.
Implements spexvb() (linear PX-VB), hspvb() (hierarchical Gamma prior), and spexvb.logistic() (logistic PX-VB) for high-dimensional variable selection with spike-and-slab priors.
Cross-validation via cv.spexvb() and cv.spexvb.fit() with parallel support through foreach.