NEWS

oem 2.1.0

The default lambda sequence now starts at the smallest value at which every penalized coefficient is zero. It accounts for penalty.factor, group.weights, alpha, unpenalized variables, and the intercept (#8). Previously the largest value was max(abs(crossprod(x, y))) / n for every penalty, so a path could begin with several fits with all coefficients zero, or with some coefficients already nonzero.
Logistic regression with a sparse x returned wrong intercepts with standardize = TRUE and wrong coefficients with standardize = FALSE.
Logistic regression with standardize = FALSE could use uninitialized column scales.
Models with at least as many variables as observations now fit correctly for a sparse x, for logistic regression, and for big.oem() with an intercept or standardization. These cases previously returned wrong values or crashed R.
big.oem() with a group penalty and an intercept left the last variable out of its group, so its coefficient was always zero.
The ridge part of elastic.net and of the other .net penalties shrank unpenalized coefficients, including the intercept. For the elementwise penalties, the ridge part is now multiplied by penalty.factor, as in glmnet, and unpenalized coefficients are not shrunk by any penalty.
C++ code no longer calls Rf_error(); errors are raised with Rcpp exception handling (#25).
For logistic regression with hessian.type = "full", the lower bound on the working weights was applied to only one observation.

With standardize = TRUE, every solver now scales each column by its standard deviation, computed with denominator n, as in glmnet. Fits with a sparse x, logistic regression, big.oem(), and xval.oem() previously used the uncentered root mean square with denominator n - 1, so their results with standardize = TRUE change.
For Gaussian models, the ridge part of the .net penalties is divided by the standard deviation of y (its root mean square when intercept = FALSE), as in glmnet. This was already the case for a dense x with an intercept; it now holds for every solver except oem.xtx(), which does not see y.

The OEM iterations are now accelerated by default. The new accelerate argument of oem(), cv.oem(), xval.oem(), big.oem(), and oem.xtx() selects FISTA with adaptive restarting ("fista", the default except in oem.xtx()), SQUAREM ("squarem"), or plain OEM ("none" or FALSE). For the nonconvex penalties, an accelerated step is accepted only if it does not increase the objective. On correlated designs, FISTA reduced computation time along the path by a factor of 4 to 10 and fitted the path to the requested tolerance in cases where plain OEM stopped at maxit before converging.

accelerate = "fista" is the default in oem(), cv.oem(), xval.oem(), and big.oem(), so their results change slightly. Use accelerate = "none" to reproduce the plain OEM iterations of previous versions. oem.xtx(), the low-level interface for precomputed crossprod(x) and crossprod(x, y), keeps plain iterations by default (accelerate = "none").
The default maxit is now 5000 instead of 500. With 500 iterations, plain OEM often stopped before convergence for correlated variables, without a warning. Fits that converge sooner stop sooner, so the higher limit costs nothing for them.

Fits with a sparse x, logistic regression, big.oem(), and xval.oem() now center the columns implicitly and rescale the intercept column. When the columns have nonzero means, this reduces the number of OEM iterations along the path by about an order of magnitude.

Sparse matrices are mapped with Eigen::Map<Eigen::SparseMatrix>, so the package builds with Eigen 5 (#26, from Dirk Eddelbuettel).
Added a testthat suite, including comparisons with glmnet.
src/Makevars assigned PKG_CXXFLAGS twice, so -DNDEBUG was dropped and Eigen's debugging assertions were compiled in. It is now set once.