Enforcing nonnegativity
by flexible Krylov subspaces
A new fast solver, based on a flexible Krylov method, for nonnegative least squares problems
Constrained
least squares problems arise in a variety of applications, and many iterative
methods are already available to compute their solutions. This page is devoted
to the presentation of a new solver for nonnegatively
constrained least squares problems, dubbed
NNFCGLS (NonNegative
Flexible CGLS) method.
The
starting points for NNFCGLS are the Karush-Kuhn-Tucker
(KKT) conditions, expressed for the least squares problem with nonnegativity constraints. The KKT conditions are leveraged
to form an adaptively preconditioned linear system, which is then solved by a modified
flexible conjugate gradient method.
NNFCGLS is described in details in the following paper:
FAST NONNEGATIVE LEAST SQUARES THROUGH
FLEXIBLE KRYLOV SUBSPACES (draft),
which I coauthored with Yves
Wiaux, and which is published in the SIAM Journal of Scientific Computing, 39(2), A655--A679 (2017)
Click
here to download a MATLAB software implementing NNFCGLS, including
data and scripts to reproduce all the tests described in the above paper. NNFCGLS
will be also included in IR Tools,
a MATLAB toolbox for iterative regularization of ill-posed linear systems,
which will be released soon.