grasps 0.1.3

• Fixed within-group local linear approximation (LLA) weights so that diagonal entries are excluded when diag.include = FALSE.
• Revised adaptive group weights so that gamma is applied consistently when constructing Frobenius-norm-based weights.
• Added optional iterative LLA for non-convex penalties through lla, tol.lla, and maxiter.lla; one-step LLA remains the default.
• Added configurable initial estimators for adaptive and non-convex penalties through initial, including the lasso estimate (default), the inverse empirical covariance matrix, and a user-supplied symmetric matrix.
• Added iterations, converged, initial.converged, lla.iterations, and lla.converged to report inner ADMM and outer LLA convergence information.
• Added converged.grid and lla.converged.grid to retain convergence status for every tuning-parameter candidate and cross-validation fold.
• Added warnings when the selected estimate, its selected cross-validation fits, or the iterative LLA procedure does not satisfy its convergence criterion.
• Strengthened validation of data, covariance matrices, group labels, penalty parameters, initial estimators, and solver controls before calling the native solvers.
• Required an explicit sample size for covariance-matrix input and strengthened validation of cross-validation folds and training-fold initial estimators.
• Limited validation of model-selection and automatic-grid controls to cases where those controls are active.
• Added positive-definiteness and finite-score checks during model selection. Invalid candidates receive infinite scores, and selection stops when no candidate has a finite score.
• Strengthened validation in gen_prec_sbm() and corrected its reported sparsity to use undirected off-diagonal edges.
• Strengthened validation in sparsify_block_banded(), corrected its reported sparsity, and added explicit handling of non-positive-definite results.
• Strengthened validation in performance(), retained computable metrics for non-positive-definite inputs, and returned NA for undefined metrics.
• Strengthened validation in prec_to_adj() and clarified the handling of diagonal entries, thresholding, and non-positive-definite inputs.
• Added vectorized input validation and penalty-specific gamma checks to compute_penalty() and compute_derivative().
• Expanded standalone regression tests for input validation, convergence reporting, invalid candidate handling, simulation utilities, performance measures, adjacency conversion, and penalty calculations.

grasps 0.1.2

• Fixed the tuning-parameter grid so that all combinations of user-supplied alpha and lambda values are evaluated.
• Fixed automatic tuning-parameter generation so that a separate lambda sequence is constructed for each value of alpha using its corresponding lambda.max.
• Validation of nlambda and lambda.min.ratio is now performed only when lambda = NULL.
• Added a check for a non-positive or non-finite lambda.max to prevent the generation of invalid lambda sequences.
• Fixed within-group weight generation in gen_prec_sbm() so that each edge weight is sampled once and assigned symmetrically to the corresponding entries of the precision matrix.
• Added input validation for cond.target and explicit handling of empty graphs in gen_prec_sbm().
• Clarified the roles of lambda.safe and lambda.max in the documentation of the automatic lambda-grid construction.
• Expanded the documentation of the positive-definiteness and condition-number adjustment in gen_prec_sbm() and added relevant references.

grasps 0.1.1

• Added function plot.adjmat().
• Added function prec_to_adj().
• Use d instead of p to denote the dimension.
• Revised the estimator expression in the vignette.
• Added NEWS file to record the changelog.

grasps 0.1.0

• Initial release of the grasps package.
• The grasps is a toolbox for precision matrix estimation with group structure.
• Key features:
  • Unified regularization framework for sparse network learning.
  • Supports element-wise sparsity and group-wise shrinkage.
  • Includes both convex and non-convex penalties (e.g., adaptive lasso, SCAD, MCP).
  • Provides model selection tools (e.g., CV, EBIC, HBIC).
• Designed for applications where variables exhibit grouped or modular relationships (e.g., brain networks, biological pathways).