NEWS

psqn 0.3.2 (2024-09-22)

Do not use sourceCpp in tests where the files include [[Rcpp::plugins(openmp)]] when OpenMP is not supported.

the BFGS implementation supports an aboslute convergence treshold.
the user can provide pre-allocated memory to the C++ API for the BFGS implementation.

it is possible to mask (fix) some of the parameters. Non-linear equality constraints are also supported with an augmented Lagrangian method.
the code has been refactored. This may yield a speedup in some cases.
computation of the hessian is now supported using numerical differentiation with Richardson extrapolation to refine the estimate.
the gradient norm can be used as a convergence criterion.
a diagonal preconditioning method has been added to the psqn method. This requires that one defines PSQN_W_LAPACK prior to including the header files from the package and links with LAPACK and BLAS.
the license has been changed to the Apache license.

tests that compile files using Rcpp::sourceCpp() now do so in a temporary directory in tempdir(). It was done before in the package directory which violates CRAN policies.

added a generic method for partially separable functions.
made it possible to run a setup function prior to evaluating all the element functions. Thus, some computations may be done once if possible. See https://github.com/boennecd/psqn-va-ex for an example.
made it possible to change the unsigned integer type used in the package by defining the PSQN_SIZE_T macro prior to including the header files from the package.
added the incomplete Cholesky factorization preconditioner from Eigen to the package. This requires that one defines the PSQN_USE_EIGEN macro prior to including the header files from the package.