Skip to contents

Changelog

Source: NEWS.md

quadrupen 1.1-0 (2026-09-17)

  • performance
    • active set: no reallocation of the Gram matrices when variables enter or leave the model, LAPACK triangular solves, in-place Cholesky downdates
    • several variables (or groups) activated at once in the working set algorithm (new control option maxadd, default 10 for sparse models and 5 for group models)
    • proximal solvers: adaptive restart of FISTA, safer Lipschitz constant (Lanczos), step-invariant stopping rule and safeguarded Anderson acceleration for PGD
    • exact block coordinate descent with Newton steps for the quadra method of group models
    • ridge and lava: no p x p algebra for a diagonal structuring matrix (the default)
    • faster information criteria; cross-validation folds built on demand (less memory)
    • typical speed-ups over 1.0-0 with default settings: x2 to x4 for lasso, MCP and SCAD, x5 to x15 for group models, x9 to x36 for elastic-net and sparse designs with large active sets, x8 for ridge and x30 for lava with p = 8000; all lambdas now converge on these cases
  • bug fixes (results may change)
    • robustness when the active set exceeds the rank of the design (no more failure of the Cholesky factorization or of the direct solve)
    • bounded_reg() with method = "quadra" (the default) stopped too early and could return suboptimal solutions; the optimality gap is now computed from the exact KKT conditions
    • l1/linf group penalty: wrong proximal operator when the group lies in the l1 ball (fista and pgd converged to wrong solutions)
    • quadra method for group models: exact solution of each block; sparse-group, coop and l1/linf variants now converge
    • proximal solvers could stop after one iteration on badly scaled data
    • convergence status now reflects the final optimality gap
    • the Lipschitz constant estimation no longer draws from R’s random number generator
  • minor updates
    • objects are rebuilt when a C++ header changes; ARMA_NO_DEBUG also on Windows

quadrupen 1.0-0 (2026-06-05)

CRAN release: 2026-06-05

  • major updates
    • complete rewriting of R code using R6 classes
    • complete rewriting of C++ code using template and OO style programming
    • included ‘FusedLasso’ from the archived package by Holger Hoefling (fixed CRAN’s complaints)
    • added group-lasso/group-elastic and variants (group penalty: l1/l2, l1/linf, cooperative Lasso)
    • added sparse group-lasso/group-elastic and variant (group penalty: l1/l2, l1/linf, cooperative Lasso)
    • added lava and/post-lava (combination of sparse and dense regularization, Chernozukov et al, 2017)
    • extended to group-lava (group penalty: l1/l2, l1/linf, cooperative Lasso)
    • added mcp and scad (concave penalties, Chernozukov et al, 2017)
    • added refit version of Lasso/Elastic-Net (“relaxed” Lasso/Enet)
    • changing many parameters (badly) named, do not expect backward compatibility
    • added vignettes
  • minor updates
    • Integration of changes from CRAN versions from 0.2-4 to 0.2-13
    • set up github workflow for pkgdown page
    • various fixes, more testing

quadrupen 0.2-4 (2014-01-16)

CRAN release: 2014-01-16

Minor: - memory leak corrected (sp_mat declaration) - linking to Rcpp/RcppArmadillo headers (requires R 3.0-2)

quadrupen 0.2-4 (2013-11-01)

CRAN release: 2014-01-16

  • added a ‘lasso’ function, simple wrapper to the elastic-net ‘function’
  • added computation of degrees of freedom (for elastic net and bounded regression)
  • added a method to compute penalized criteria (BIC/AIC) of a quadrupen fit, with plot

quadrupen 0.2-2 (2013-04-08)

CRAN release: 2013-04-08

  • minor fix to comply with recent ggplot2 updates.

quadrupen 0.2-1 (2013-02-27)

CRAN release: 2013-02-27

  • minor fix to pass CRAN check on Windows operating systems.

quadrupen 0.2-0 (2013-02-26)

  • Major updates
    • added bounded regression (regression penalized by infinity norm + structured l2 norm)
    • added corresponding features for cross-validation and stability path
  • Minor updates
  • corrected wrong annotations of the stability path (PFER)
  • handled normalization internally (‘normalize’ is no longer a parameter)
  • more simple internal handling of penscales and correction of the rescaling of the intercept
  • better use of multicore features
  • handled runtime error exception in RcppArmadillo when the system is singular (end of the solution path) A consequence is quadrupen is less likely to crash due to user’s “bad” parametrisation
  • simplification of the C++ code, bugs corrected, probably new ones added :-’(
  • added ‘examples’ and ‘tests’ directories

quadrupen 0.1-0 (2012-10-09)

CRAN release: 2012-10-10

  • first build: structured elastic-net with (weighted) quadratic loss, cross-validation and stability selection methods.