NEWS

---

picasso 2.0.1 (2026-07-25)

Portability

• Updated the bundled Eigen headers from 3.3.5 to 3.4.0 so the package compiles cleanly under the C++20 default of recent R-devel toolchains (GCC 14/16), removing the -Wdeprecated-enum-enum-conversion install warnings that originated in the older Eigen sources. No package API or numerical behavior changes.

picasso 2.0.0

New models and solver controls

• Added multinomial logistic regression for L1, MCP, and SCAD penalties in the R and Python interfaces. The solver uses class-coupled active-set Proximal Newton/IRLS updates, pathwise warm starts, strong screening, full KKT checks, and deviance-tail early stopping for automatically generated paths.
• MCP/SCAD binomial, Poisson, square-root-lasso, and multinomial fits now use adaptive local linear approximation (LLA). lla.max.stages = 3L is the default maximum, including the initial L1 master; larger budgets are available when stricter stationarity is required.
• Added opt-in fast.mode. The default remains high precision. Fast mode uses benchmark-calibrated achieved-accuracy tolerances of 4e-4 for Poisson and 1e-4 for binomial, square-root-lasso, and multinomial; Gaussian retains 1e-7.
• Gaussian fits now default to type.gaussian = "auto". A calibrated, memory-bounded policy selects between residual-based naive updates and lazy covariance updates using the design shape and lambda-path reuse. Explicit "naive" and "covariance" requests retain their previous behavior.
• Added dfmax path-size stopping consistently across families.

Interfaces and diagnostics

• Native path fitting is now interruptible: every family polls for a pending user interrupt (Ctrl-C) at each lambda boundary and re-raises it as a standard R interrupt condition instead of running the full path to completion.
• Added binomial and Poisson offsets throughout fitting, prediction, assessment, and cross-validation. New-data offsets are applied on the link scale and are required whenever prediction or assessment evaluates the linear predictor; nonzero-support extraction does not need them.
• Added cv.picasso(), assess.picasso(), and confusion.picasso() support for all applicable families, including stratified categorical folds and stable class maps for factor and multinomial responses.
• Prediction methods now support response, link, class, and nonzero outputs as applicable, plus interpolation by lambda value through s.
• Versioned native solvers now report termination status and failure location. Gaussian iteration-limit failures retain only the previously converged path prefix. Non-Gaussian fits additionally report per-lambda runtime, objective, KKT residual, stationarity, and LLA stage counts; usable stage-budget termination remains distinct from a hard failure.
• Gaussian fit objects record both requested and resolved solver modes. Multinomial fits record the requested path length and whether a saturated generated-path tail was omitted.

Correctness, performance, and portability

• Categorical class prediction now compares finite link-scale scores directly, matching glmnet while avoiding probability-rounding ties and unnecessary sigmoid/softmax temporaries. Confusion tables use predicted classes in rows, observed classes in columns, and retain every fitted class level.
• Fixed binomial response/class prediction for requests containing multiple lambda values.
• Scalar coefficient and prediction defaults now select at most the first three available path points or coefficients, so short paths and low- dimensional models no longer fail with out-of-range default indices.
• Scalar prediction row selectors now use NULL for all rows. Explicit Y.pred.idx or p.pred.idx values are always applied instead of treating 1:5 as a hidden default sentinel.
• Invalid MCP/SCAD concavity parameters now fail instead of being silently replaced by 3; every family validates the documented LLA stage budget. Square-root-lasso reports its algorithm as active-set quadratic MM.
• Moved scalar and multinomial path-loss evaluation into the native solvers, avoiding repeated R-side matrix products after fitting.
• Cached fixed weighted column norms inside scalar active-set subproblems, compacted active working sets, vectorized multinomial curvature/KKT kernels, and reduced temporary allocations in prediction and path rescaling.
• Scalar-family assessment now keeps small paths on one matrix multiplication and evaluates larger paths in approximately 8 MiB link-predictor blocks, reducing peak memory without changing metric values.
• Strengthened validation for numeric matrix storage, finite inputs, decreasing lambda paths, dimensions, class maps, offsets, and no-intercept standardization. Constant-response and zero-gradient paths now have explicit finite behavior.
• Kept R-package and standalone native source mirrors under build-time parity checks and added scalar, multinomial, interface, and numerical-stability regression coverage.
• Removed the unused MASS attachment from the R package dependency surface; only Matrix is attached at load time.