GitHub-Proxy/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-04-21 09:09:36 +08:00
Code Issues Packages Projects Releases Wiki Activity

bug #899: remove "rank-revealing" qualifier for SparseQR and warn that it is not always rank-revealing.

Browse Source
This commit is contained in:
Gael Guennebaud 2019-02-19 22:52:15 +01:00
parent 9ac1634fdf
commit 482c5fb321
1 changed files with 17 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

22
Eigen/src/SparseQR/SparseQR.h
View File

@ -41,15 +41,16 @@ namespace internal {
/** /**
* \ingroup SparseQR_Module * \ingroup SparseQR_Module
* \class SparseQR * \class SparseQR
* \brief Sparse left-looking rank-revealing QR factorization * \brief Sparse left-looking QR factorization with numerical column pivoting
* *
* This class implements a left-looking rank-revealing QR decomposition * This class implements a left-looking QR decomposition of sparse matrices
* of sparse matrices. When a column has a norm less than a given tolerance * with numerical column pivoting.
* When a column has a norm less than a given tolerance
* it is implicitly permuted to the end. The QR factorization thus obtained is * it is implicitly permuted to the end. The QR factorization thus obtained is
* given by A*P = Q*R where R is upper triangular or trapezoidal. * given by A*P = Q*R where R is upper triangular or trapezoidal.
* *
* P is the column permutation which is the product of the fill-reducing and the * P is the column permutation which is the product of the fill-reducing and the
* rank-revealing permutations. Use colsPermutation() to get it. * numerical permutations. Use colsPermutation() to get it.
* *
* Q is the orthogonal matrix represented as products of Householder reflectors. * Q is the orthogonal matrix represented as products of Householder reflectors.
* Use matrixQ() to get an expression and matrixQ().adjoint() to get the adjoint. * Use matrixQ() to get an expression and matrixQ().adjoint() to get the adjoint.
@ -64,6 +65,17 @@ namespace internal {
* *
* \implsparsesolverconcept * \implsparsesolverconcept
* *
* The numerical pivoting strategy and default threshold are the same as in SuiteSparse QR, and
* detailed in the following paper:
* <i>
* Tim Davis, "Algorithm 915, SuiteSparseQR: Multifrontal Multithreaded Rank-Revealing
* Sparse QR Factorization, ACM Trans. on Math. Soft. 38(1), 2011.
* </i>
* Even though it is qualified as "rank-revealing", this strategy might fail for some
* rank deficient problems. When this class is used to solve linear or least-square problems
* it is thus strongly recommended to check the accuracy of the computed solution. If it
* failed, it usually helps to increase the threshold with setPivotThreshold.
*
* \warning The input sparse matrix A must be in compressed mode (see SparseMatrix::makeCompressed()). * \warning The input sparse matrix A must be in compressed mode (see SparseMatrix::makeCompressed()).
* \warning For complex matrices matrixQ().transpose() will actually return the adjoint matrix. * \warning For complex matrices matrixQ().transpose() will actually return the adjoint matrix.
* *
@ -331,7 +343,7 @@ void SparseQR<MatrixType,OrderingType>::analyzePattern(const MatrixType& mat)
m_R.resize(m, n); m_R.resize(m, n);
m_Q.resize(m, diagSize); m_Q.resize(m, diagSize);
// Allocate space for nonzero elements : rough estimation // Allocate space for nonzero elements: rough estimation
m_R.reserve(2*mat.nonZeros()); //FIXME Get a more accurate estimation through symbolic factorization with the etree m_R.reserve(2*mat.nonZeros()); //FIXME Get a more accurate estimation through symbolic factorization with the etree
m_Q.reserve(2*mat.nonZeros()); m_Q.reserve(2*mat.nonZeros());
m_hcoeffs.resize(diagSize); m_hcoeffs.resize(diagSize);