TaoDiagonalNNLS.hpp

Go to the documentation of this file.

/** \file TaoDiagonalNNLS.hpp
 * \brief Direct nonnegative least squares for orthogonal constraint rows.
 */
 
#ifndef MOFEM_TAO_DIAGONAL_NNLS_HPP
#define MOFEM_TAO_DIAGONAL_NNLS_HPP
 
namespace MoFEM {
 
/** Registered TAOSHELL solver for a diagonal NNLS normal matrix. */
inline constexpr char TAODIAGONALNNLS[] = "diagonal";
 
/** Register the "diagonal" TAO type before selecting it. */
MoFEMErrorCode TaoRegisterDiagonalNNLS();
 
/** Set the fixed least-squares problem min_{x >= 0} ||J^T x - target||^2/2.
 *
 * J must have mutually orthogonal, nonzero rows. The solver checks the
 * diagonally normalized Gram matrix and computes
 * x_i = max(0, (J target)_i / (J J^T)_ii) in one algebraic update.
 * Coupled rows are rejected; use a general NNLS solver such as TAOBLMVM for
 * those problems. No line search is used.
 *
 * Call after selecting TAODIAGONALNNLS. The solver retains J and target and
 * installs objective/gradient callbacks for this problem. Set the solution
 * vector with TaoSetSolution using J's row layout; target uses its column
 * layout. Inputs may be updated or rebound between solves. Each solve
 * recomputes the Gram diagonal and verifies orthogonality.
 *
 * Nonnegativity is implicit. Explicit bounds, if supplied, must be zero and
 * positive infinity; additional constraints are unsupported. Standard TAO
 * monitoring and convergence checks use the true bound-projected gradient,
 * including its nonzero inward components at x_i = 0. The objective itself
 * can remain positive at the least-squares optimum.
 */
MoFEMErrorCode TaoDiagonalNNLSSetProblem(Tao tao, Mat jacobian, Vec target);
 
} // namespace MoFEM
 
#endif // MOFEM_TAO_DIAGONAL_NNLS_HPP