Replaced the optional CPSS–MCP base-path dependency on ncvreg by singleton-group grpreg MCP. With one predictor per group, the group MCP penalty is the ordinary coordinatewise MCP penalty.
Corrected lambda_min_ratio provenance: both CPSS–LASSO and CPSS–MCP now record a lower path endpoint relative to the engine-specific maximum penalty. The earlier absolute-lambda.min label was documentation metadata, not the behavior of the MCP engine.
Retained the selected_path_locally_convex diagnostic as NA for backward compatibility because grpreg does not expose the former local-convexity index. No local-convexity claim is made in version 0.6.2.
HDMaxShrink 0.6.1
Added portable package-level LAPACK, BLAS, and Fortran linkage through src/Makevars and src/Makevars.win. This fixes unresolved Armadillo eigendecomposition and SVD symbols on Linux without changing estimators, APIs, defaults, seeds, or statistical calculations.
Extended source-provenance checks to include the package build recipes.
HDMaxShrink 0.6.0
Added backward-compatible mandatory_core and candidate_set arguments to cpss_select_core(). In every complementary half-sample, required predictors are centered and projected out by a checked economy SVD before LASSO or MCP sees the optional candidate family.
Separated mandatory variables, CPSS-selected extensions, ineligible optional directions, and predictors outside the selection universe in the returned audit trail. Empty extensions now leave a nonempty mandatory core intact without a top-k fallback.
Scoped the reported Meinshausen–Buehlmann expression to the eligible optional family; mandatory variables have no artificial CPSS frequency.
Updated the LASSO base fit for the glmnet 5.0 control interface while retaining compatibility with earlier interfaces.
HDMaxShrink 0.5.0
Added cpss_select_core() for strict complementary-pairs stability selection with a size-budgeted LASSO base selector and an optional MCP sensitivity selector. The implementation records feature identities, stability frequencies, base-fit convergence, MCP local-convexity diagnostics, and the applicability of the above-half stability-selection PFER expression. It never silently replaces an empty stable core by top-k variables.
Added fit_cpss_sm(), a CPSS-guided exact-null submodel estimator computed by an economy-SVD Moore–Penrose refit on an analysis sample independent of selection.
Added fit_cpss_ridge_shrinkage(), coupling CPSS–SM with an all-predictor dual-Ridge full-model endpoint and the existing conditional-Gaussian max-partial-t/inverse-moment PT, S, PS, and PPS family. Ridge is not used to calibrate the null test, and no OLS, p-by-p, or q-by-q inverse is formed.
HDMaxShrink 0.4.0
Added full_endpoint = "all_x_square_root_lasso" as the primary endpoint. It centers and RMS-scales all columns of X and y, fixes the standardized intercept at zero, and estimates all p coefficients jointly in one square-root-LASSO optimization with positive penalties on every coordinate.
Kept SM as a fresh exact-null SVD refit using only the prespecified low-dimensional X_A; no FM coefficient is copied into SM.
Back-transforms FM, SM, PT, S, PS, and PPS coefficients and predictions to the original data scale while recording the zero standardized intercept.
Preserved the conditional-Gaussian maximum partial-t test on the X_A-residualized tested block. The all-X fit does not change the null-test statistic or its Monte Carlo calibration.
Retained the thresholded partial-debiased and profiled partial endpoints as explicit backward-compatible research modes; they are no longer the primary analysis endpoint.
HDMaxShrink 0.3.0
Added fit_partial_sqrt_lasso_shrinkage() as the primary non-Ridge framework: a partial square-root LASSO supplies the pilot, FM is its thresholded partial-debiased and core-reprofiled endpoint, and SM is the exact-null SVD refit under beta[B] = 0.
Added an explicit marginal exceedance budget threshold_eta; the q = 990 study prespecifies eta 5 and reports eta 1, 3, and 5 sensitivity.
Profiled the prespecified low-dimensional core by an economy SVD. No OLS Gram inverse, q by q inverse, or p by p inverse is formed.
Added a Gaussian-iid diagonal one-step Tmax and multiplier-bootstrap null calibration, exposed separately through max_score_test_hd().
Synchronized the primary partial fit with the exact-null Gaussian theory: its default is now the conditional-Gaussian maximum partial-t test and the reciprocal inverse-moment max-Stein calibration. The raw score maximum, partial-t maximum, null second moment, and inverse-moment calibration are stored in separate fields. The former residual-multiplier plus second-moment construction remains available through explicit arguments.
Added reusable null geometry, full-model warm starts, and tests showing that the null endpoint is a genuine refit rather than a projection or oracle construction.
Retained Ridge–MCP and the earlier projection framework without changing their APIs.
HDMaxShrink 0.2.0
Added a dense–sparse endpoint framework with all-X dual Ridge as FM and all-X MCP as SM; the MCP solver receives no active-set information.
Added RcppArmadillo Ridge-path and MCP coordinate-descent kernels, with GCV and EBIC tuning respectively.
Kept the large-block test on a separate debiased square-root-LASSO score so that Ridge bias does not calibrate Tmax.
Retained the 0.1.2 projection estimator as a backward-compatible research mode.
HDMaxShrink 0.1.2
Replaced the reduced-design submodel fit with an exact restriction projection of the single all-X full square-root lasso estimator.
Projected the all-X debiased full endpoint instead of separately debiasing a reduced submodel.
Kept the multiplier-bootstrap Tmax construction unchanged and added projection-invariance tests.
HDMaxShrink 0.1.1
Added symmetric one-step debiasing of both full-model and submodel fits.
