mirror/eigen
2
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2024-12-27 07:29:52 +08:00
Code Issues Packages Projects Releases Wiki Activity

LDLT is not rank-revealing, so we should not attempt to use the biggest diagonal elements as thresholds.

Browse Source
This commit is contained in:
Gael Guennebaud 2014-07-02 23:04:46 +02:00
parent b18a7ff6be
commit bbe9e22d60
2 changed files with 70 additions and 23 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

41
Eigen/src/Cholesky/LDLT.h
View File

@ -274,30 +274,13 @@ template<> struct ldlt_inplace<Lower>
return true;
}
RealScalar cutoff(0), biggest_in_corner;
for (Index k = 0; k < size; ++k)
{
// Find largest diagonal element
Index index_of_biggest_in_corner;
biggest_in_corner = mat.diagonal().tail(size-k).cwiseAbs().maxCoeff(&index_of_biggest_in_corner);
mat.diagonal().tail(size-k).cwiseAbs().maxCoeff(&index_of_biggest_in_corner);
index_of_biggest_in_corner += k;
if(k == 0)
{
// The biggest overall is the point of reference to which further diagonals
// are compared; if any diagonal is negligible compared
// to the largest overall, the algorithm bails.
cutoff = abs(NumTraits<Scalar>::epsilon() * biggest_in_corner);
}
// Finish early if the matrix is not full rank.
if(biggest_in_corner < cutoff)
{
for(Index i = k; i < size; i++) transpositions.coeffRef(i) = i;
break;
}
transpositions.coeffRef(k) = index_of_biggest_in_corner;
if(k != index_of_biggest_in_corner)
{
@ -333,7 +316,12 @@ template<> struct ldlt_inplace<Lower>
if(rs>0)
A21.noalias() -= A20 * temp.head(k);
}
if((rs>0) && (abs(mat.coeffRef(k,k)) > cutoff))
// In some previous versions of Eigen (e.g., 3.2.1), the scaling was omitted if the pivot
// was smaller than the cutoff value. However, soince LDLT is not rank-revealing
// we should only make sure we do not introduce INF or NaN values.
// LAPACK also uses 0 as the cutoff value.
if((rs>0) && (abs(mat.coeffRef(k,k)) > RealScalar(0)))
A21 /= mat.coeffRef(k,k);
RealScalar realAkk = numext::real(mat.coeffRef(k,k));
@ -517,13 +505,20 @@ struct solve_retval<LDLT<_MatrixType,_UpLo>, Rhs>
typedef typename LDLTType::Scalar Scalar;
typedef typename LDLTType::RealScalar RealScalar;
const Diagonal<const MatrixType> vectorD = dec().vectorD();
RealScalar tolerance = (max)(vectorD.array().abs().maxCoeff() * NumTraits<Scalar>::epsilon(),
RealScalar(1) / NumTraits<RealScalar>::highest()); // motivated by LAPACK's xGELSS
// In some previous versions, tolerance was set to the max of 1/highest and the maximal diagonal entry * epsilon
// as motivated by LAPACK's xGELSS:
// RealScalar tolerance = (max)(vectorD.array().abs().maxCoeff() *NumTraits<RealScalar>::epsilon(),RealScalar(1) / NumTraits<RealScalar>::highest());
// However, LDLT is not rank revealing, and so adjusting the tolerance wrt to the highest
// diagonal element is not well justified and to numerical issues in some cases.
// Moreover, Lapack's xSYTRS routines use 0 for the tolerance.
RealScalar tolerance = RealScalar(1) / NumTraits<RealScalar>::highest();
for (Index i = 0; i < vectorD.size(); ++i) {
if(abs(vectorD(i)) > tolerance)
dst.row(i) /= vectorD(i);
dst.row(i) /= vectorD(i);
else
dst.row(i).setZero();
dst.row(i).setZero();
}
// dst = L^-T (D^-1 L^-1 P b)

52
test/cholesky.cpp
View File

@ -68,6 +68,7 @@ template<typename MatrixType> void cholesky(const MatrixType& m)
Index cols = m.cols();
typedef typename MatrixType::Scalar Scalar;
typedef typename NumTraits<Scalar>::Real RealScalar;
typedef Matrix<Scalar, MatrixType::RowsAtCompileTime, MatrixType::RowsAtCompileTime> SquareMatrixType;
typedef Matrix<Scalar, MatrixType::RowsAtCompileTime, 1> VectorType;
@ -179,6 +180,57 @@ template<typename MatrixType> void cholesky(const MatrixType& m)
// restore
if(sign == -1)
symm = -symm;
// check matrices coming from linear constraints with Lagrange multipliers
if(rows>=3)
{
SquareMatrixType A = symm;
int c = internal::random<int>(0,rows-2);
A.bottomRightCorner(c,c).setZero();
// Make sure a solution exists:
vecX.setRandom();
vecB = A * vecX;
vecX.setZero();
ldltlo.compute(A);
VERIFY_IS_APPROX(A, ldltlo.reconstructedMatrix());
vecX = ldltlo.solve(vecB);
VERIFY_IS_APPROX(A * vecX, vecB);
}
// check non-full rank matrices
if(rows>=3)
{
int r = internal::random<int>(1,rows-1);
Matrix<Scalar,Dynamic,Dynamic> a = Matrix<Scalar,Dynamic,Dynamic>::Random(rows,r);
SquareMatrixType A = a * a.adjoint();
// Make sure a solution exists:
vecX.setRandom();
vecB = A * vecX;
vecX.setZero();
ldltlo.compute(A);
VERIFY_IS_APPROX(A, ldltlo.reconstructedMatrix());
vecX = ldltlo.solve(vecB);
VERIFY_IS_APPROX(A * vecX, vecB);
}
// check matrices with wide spectrum
if(rows>=3)
{
RealScalar s = (std::min)(16,std::numeric_limits<RealScalar>::max_exponent10/8);
Matrix<Scalar,Dynamic,Dynamic> a = Matrix<Scalar,Dynamic,Dynamic>::Random(rows,rows);
Matrix<RealScalar,Dynamic,1> d = Matrix<RealScalar,Dynamic,1>::Random(rows);
for(int k=0; k<rows; ++k)
d(k) = d(k)*std::pow(RealScalar(10),internal::random<RealScalar>(-s,s));
SquareMatrixType A = a * d.asDiagonal() * a.adjoint();
// Make sure a solution exists:
vecX.setRandom();
vecB = A * vecX;
vecX.setZero();
ldltlo.compute(A);
VERIFY_IS_APPROX(A, ldltlo.reconstructedMatrix());
vecX = ldltlo.solve(vecB);
VERIFY_IS_APPROX(A * vecX, vecB);
}
}
// update/downdate