libgenua
Basic Geometry, Numerical Algorithms and Interfaces
 All Classes Namespaces Files Functions Variables Typedefs Enumerations Groups Pages
Public Member Functions | Private Member Functions | Private Attributes | List of all members
LsqrSolver< Scalar > Class Template Reference

Detailed Description

template<typename Scalar>
class LsqrSolver< Scalar >

LSQR: Iterative solver for least-squares problems.

The operator 'A' passed to solve() needs to provide the interface

  void muladd(const DVector<Scalar> &x, DVector<Scalar> &y);
  void muladdTransposed(const DVector<Scalar> &x, DVector<Scalar> &y);

which compute y = A*x and y = transpose(A)*x.

This implementation differs from the LSQR paper in that it includes an additional termination criterion based on the scaled length of x. In trust- region methods for least-squares problems, a solution |s*x| < delta is often needed, and in that case, it may be advantageous to truncate LSQR early based on the scaled radius criterion.

See Also
CraigSolver, LsmrSolver

#include <lsqr.h>

Inheritance diagram for LsqrSolver< Scalar >:
[legend]
Collaboration diagram for LsqrSolver< Scalar >:
[legend]

Public Member Functions

void maxRadius (const DVector< Scalar > &s, Scalar r)
 optional termination: stop if |s*x| >= r
 
void maxRadius (Scalar r)
 optional termination: stop if |x| >= r
 
template<class Operator >
ExitCode solve (const Operator &A, const DVector< Scalar > &b, DVector< Scalar > &x, Scalar lambda=0)
 minimize |Ax - b|^2 + lambda |x|^2
 
- Public Member Functions inherited from SolIterativeSolver
void tolerance (int maxiter, double atoler, double btoler, double conlim=1e8)
 set convergence tolerances and maximum number of iterations
 

Private Member Functions

bool radiusExceeded (const DVector< Scalar > &x) const
 check radius criterion if enabled
 
void steihaugPoint (const DVector< Scalar > &xp, DVector< Scalar > &x)
 blend xp and x such that |x| = delta
 

Private Attributes

DVector< Scalar > m_work [3]
 work space, to avoid re-allocation when solve() is called often
 
DVector< Scalar > m_xscale
 scaling vector used to constrain |s*x|^2 < m_xrsq
 
Scalar m_xrsq = 0
 permitted squared length of x (0 means no constraint)
 

Additional Inherited Members

- Static Public Member Functions inherited from SolIterativeSolver
static const char * statusMessage (int code)
 access text corresponding to error code
 
static bool success (int code)
 does exit code indicate succes (in some sense, at least)?
 
- Protected Attributes inherited from SolIterativeSolver
double m_atol = 1e-6
 convergence criteria
 
double m_btol = 1e-6
 convergence criteria
 
double m_conlim = 1e8
 condition number limit (actually, limit for the estimated condition)
 
size_t m_maxiter = 128
 permitted number of iteration
 

The documentation for this class was generated from the following file: