mirror of
https://gitlab.com/libeigen/eigen.git
synced 2025-08-04 03:00:39 +08:00
Backport numerical robustness fixes from 3.3 branch
This commit is contained in:
parent
4f7baefa81
commit
2a3680da3d
@ -359,29 +359,42 @@ struct svd_precondition_2x2_block_to_be_real<MatrixType, QRPreconditioner, false
|
||||
{
|
||||
typedef JacobiSVD<MatrixType, QRPreconditioner> SVD;
|
||||
typedef typename SVD::Index Index;
|
||||
static void run(typename SVD::WorkMatrixType&, SVD&, Index, Index) {}
|
||||
typedef typename MatrixType::RealScalar RealScalar;
|
||||
static bool run(typename SVD::WorkMatrixType&, SVD&, Index, Index, RealScalar&) { return true; }
|
||||
};
|
||||
|
||||
template<typename MatrixType, int QRPreconditioner>
|
||||
struct svd_precondition_2x2_block_to_be_real<MatrixType, QRPreconditioner, true>
|
||||
{
|
||||
typedef JacobiSVD<MatrixType, QRPreconditioner> SVD;
|
||||
typedef typename SVD::Index Index;
|
||||
typedef typename MatrixType::Scalar Scalar;
|
||||
typedef typename MatrixType::RealScalar RealScalar;
|
||||
typedef typename SVD::Index Index;
|
||||
static void run(typename SVD::WorkMatrixType& work_matrix, SVD& svd, Index p, Index q)
|
||||
static bool run(typename SVD::WorkMatrixType& work_matrix, SVD& svd, Index p, Index q, RealScalar& maxDiagEntry)
|
||||
{
|
||||
using std::sqrt;
|
||||
using std::abs;
|
||||
using std::max;
|
||||
Scalar z;
|
||||
JacobiRotation<Scalar> rot;
|
||||
RealScalar n = sqrt(numext::abs2(work_matrix.coeff(p,p)) + numext::abs2(work_matrix.coeff(q,p)));
|
||||
|
||||
const RealScalar considerAsZero = (std::numeric_limits<RealScalar>::min)();
|
||||
const RealScalar precision = NumTraits<Scalar>::epsilon();
|
||||
|
||||
if(n==0)
|
||||
{
|
||||
// make sure first column is zero
|
||||
work_matrix.coeffRef(p,p) = work_matrix.coeffRef(q,p) = Scalar(0);
|
||||
|
||||
if(abs(numext::imag(work_matrix.coeff(p,q)))>considerAsZero)
|
||||
{
|
||||
// work_matrix.coeff(p,q) can be zero if work_matrix.coeff(q,p) is not zero but small enough to underflow when computing n
|
||||
z = abs(work_matrix.coeff(p,q)) / work_matrix.coeff(p,q);
|
||||
work_matrix.row(p) *= z;
|
||||
if(svd.computeU()) svd.m_matrixU.col(p) *= conj(z);
|
||||
if(work_matrix.coeff(q,q)!=Scalar(0))
|
||||
}
|
||||
if(abs(numext::imag(work_matrix.coeff(q,q)))>considerAsZero)
|
||||
{
|
||||
z = abs(work_matrix.coeff(q,q)) / work_matrix.coeff(q,q);
|
||||
work_matrix.row(q) *= z;
|
||||
@ -395,19 +408,25 @@ struct svd_precondition_2x2_block_to_be_real<MatrixType, QRPreconditioner, true>
|
||||
rot.s() = work_matrix.coeff(q,p) / n;
|
||||
work_matrix.applyOnTheLeft(p,q,rot);
|
||||
if(svd.computeU()) svd.m_matrixU.applyOnTheRight(p,q,rot.adjoint());
|
||||
if(work_matrix.coeff(p,q) != Scalar(0))
|
||||
if(abs(numext::imag(work_matrix.coeff(p,q)))>considerAsZero)
|
||||
{
|
||||
Scalar z = abs(work_matrix.coeff(p,q)) / work_matrix.coeff(p,q);
|
||||
z = abs(work_matrix.coeff(p,q)) / work_matrix.coeff(p,q);
|
||||
work_matrix.col(q) *= z;
|
||||
if(svd.computeV()) svd.m_matrixV.col(q) *= z;
|
||||
}
|
||||
if(work_matrix.coeff(q,q) != Scalar(0))
|
||||
if(abs(numext::imag(work_matrix.coeff(q,q)))>considerAsZero)
|
||||
{
|
||||
z = abs(work_matrix.coeff(q,q)) / work_matrix.coeff(q,q);
|
||||
work_matrix.row(q) *= z;
|
||||
if(svd.computeU()) svd.m_matrixU.col(q) *= conj(z);
|
||||
}
|
||||
}
|
||||
|
||||
// update largest diagonal entry
|
||||
maxDiagEntry = max EIGEN_EMPTY (maxDiagEntry,max EIGEN_EMPTY (abs(work_matrix.coeff(p,p)), abs(work_matrix.coeff(q,q))));
|
||||
// and check whether the 2x2 block is already diagonal
|
||||
RealScalar threshold = max EIGEN_EMPTY (considerAsZero, precision * maxDiagEntry);
|
||||
return abs(work_matrix.coeff(p,q))>threshold || abs(work_matrix.coeff(q,p)) > threshold;
|
||||
}
|
||||
};
|
||||
|
||||
@ -424,18 +443,19 @@ void real_2x2_jacobi_svd(const MatrixType& matrix, Index p, Index q,
|
||||
JacobiRotation<RealScalar> rot1;
|
||||
RealScalar t = m.coeff(0,0) + m.coeff(1,1);
|
||||
RealScalar d = m.coeff(1,0) - m.coeff(0,1);
|
||||
if(t == RealScalar(0))
|
||||
if(d == RealScalar(0))
|
||||
{
|
||||
rot1.c() = RealScalar(0);
|
||||
rot1.s() = d > RealScalar(0) ? RealScalar(1) : RealScalar(-1);
|
||||
rot1.s() = RealScalar(0);
|
||||
rot1.c() = RealScalar(1);
|
||||
}
|
||||
else
|
||||
{
|
||||
RealScalar t2d2 = numext::hypot(t,d);
|
||||
rot1.c() = abs(t)/t2d2;
|
||||
rot1.s() = d/t2d2;
|
||||
if(t<RealScalar(0))
|
||||
rot1.s() = -rot1.s();
|
||||
// If d!=0, then t/d cannot overflow because the magnitude of the
|
||||
// entries forming d are not too small compared to the ones forming t.
|
||||
RealScalar u = t / d;
|
||||
RealScalar tmp = sqrt(RealScalar(1) + numext::abs2(u));
|
||||
rot1.s() = RealScalar(1) / tmp;
|
||||
rot1.c() = u / tmp;
|
||||
}
|
||||
m.applyOnTheLeft(0,1,rot1);
|
||||
j_right->makeJacobi(m,0,1);
|
||||
@ -826,6 +846,7 @@ JacobiSVD<MatrixType, QRPreconditioner>::compute(const MatrixType& matrix, unsig
|
||||
check_template_parameters();
|
||||
|
||||
using std::abs;
|
||||
using std::max;
|
||||
allocate(matrix.rows(), matrix.cols(), computationOptions);
|
||||
|
||||
// currently we stop when we reach precision 2*epsilon as the last bit of precision can require an unreasonable number of iterations,
|
||||
@ -857,6 +878,7 @@ JacobiSVD<MatrixType, QRPreconditioner>::compute(const MatrixType& matrix, unsig
|
||||
}
|
||||
|
||||
/*** step 2. The main Jacobi SVD iteration. ***/
|
||||
RealScalar maxDiagEntry = m_workMatrix.cwiseAbs().diagonal().maxCoeff();
|
||||
|
||||
bool finished = false;
|
||||
while(!finished)
|
||||
@ -872,16 +894,14 @@ JacobiSVD<MatrixType, QRPreconditioner>::compute(const MatrixType& matrix, unsig
|
||||
// if this 2x2 sub-matrix is not diagonal already...
|
||||
// notice that this comparison will evaluate to false if any NaN is involved, ensuring that NaN's don't
|
||||
// keep us iterating forever. Similarly, small denormal numbers are considered zero.
|
||||
using std::max;
|
||||
RealScalar threshold = (max)(considerAsZero, precision * (max)(abs(m_workMatrix.coeff(p,p)),
|
||||
abs(m_workMatrix.coeff(q,q))));
|
||||
// We compare both values to threshold instead of calling max to be robust to NaN (See bug 791)
|
||||
RealScalar threshold = max EIGEN_EMPTY (considerAsZero, precision * maxDiagEntry);
|
||||
if(abs(m_workMatrix.coeff(p,q))>threshold || abs(m_workMatrix.coeff(q,p)) > threshold)
|
||||
{
|
||||
finished = false;
|
||||
|
||||
// perform SVD decomposition of 2x2 sub-matrix corresponding to indices p,q to make it diagonal
|
||||
internal::svd_precondition_2x2_block_to_be_real<MatrixType, QRPreconditioner>::run(m_workMatrix, *this, p, q);
|
||||
// the complex to real operation returns true is the updated 2x2 block is not already diagonal
|
||||
if(internal::svd_precondition_2x2_block_to_be_real<MatrixType, QRPreconditioner>::run(m_workMatrix, *this, p, q, maxDiagEntry))
|
||||
{
|
||||
JacobiRotation<RealScalar> j_left, j_right;
|
||||
internal::real_2x2_jacobi_svd(m_workMatrix, p, q, &j_left, &j_right);
|
||||
|
||||
@ -891,6 +911,10 @@ JacobiSVD<MatrixType, QRPreconditioner>::compute(const MatrixType& matrix, unsig
|
||||
|
||||
m_workMatrix.applyOnTheRight(p,q,j_right);
|
||||
if(computeV()) m_matrixV.applyOnTheRight(p,q,j_right);
|
||||
|
||||
// keep track of the largest diagonal coefficient
|
||||
maxDiagEntry = max EIGEN_EMPTY (maxDiagEntry,max EIGEN_EMPTY (abs(m_workMatrix.coeff(p,p)), abs(m_workMatrix.coeff(q,q))));
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
Loading…
x
Reference in New Issue
Block a user