161 lines
4.8 KiB
C++
161 lines
4.8 KiB
C++
/* ----------------------------------------------------------------------------
|
|
|
|
* GTSAM Copyright 2010, Georgia Tech Research Corporation,
|
|
* Atlanta, Georgia 30332-0415
|
|
* All Rights Reserved
|
|
* Authors: Frank Dellaert, et al. (see THANKS for the full author list)
|
|
|
|
* See LICENSE for the license information
|
|
|
|
* -------------------------------------------------------------------------- */
|
|
|
|
/**
|
|
* @file cholesky.cpp
|
|
* @brief Efficient incomplete Cholesky on rank-deficient matrices, todo: constrained Cholesky
|
|
* @author Richard Roberts
|
|
* @author Frank Dellaert
|
|
* @date Nov 5, 2010
|
|
*/
|
|
|
|
#include <gtsam/base/cholesky.h>
|
|
#include <gtsam/base/timing.h>
|
|
|
|
#include <boost/format.hpp>
|
|
#include <cmath>
|
|
|
|
using namespace std;
|
|
|
|
namespace gtsam {
|
|
|
|
static const double negativePivotThreshold = -1e-1;
|
|
static const double zeroPivotThreshold = 1e-6;
|
|
static const double underconstrainedPrior = 1e-5;
|
|
static const int underconstrainedExponentDifference = 12;
|
|
|
|
/* ************************************************************************* */
|
|
static inline int choleskyStep(Matrix& ATA, size_t k, size_t order) {
|
|
// Get pivot value
|
|
double alpha = ATA(k, k);
|
|
|
|
// Correct negative pivots from round-off error
|
|
if (alpha < negativePivotThreshold) {
|
|
return -1;
|
|
} else if (alpha < 0.0)
|
|
alpha = 0.0;
|
|
|
|
const double beta = sqrt(alpha);
|
|
|
|
if (beta > zeroPivotThreshold) {
|
|
const double betainv = 1.0 / beta;
|
|
|
|
// Update k,k
|
|
ATA(k, k) = beta;
|
|
|
|
if (k < (order - 1)) {
|
|
// Update A(k,k+1:end) <- A(k,k+1:end) / beta
|
|
typedef Matrix::RowXpr::SegmentReturnType BlockRow;
|
|
BlockRow V = ATA.row(k).segment(k + 1, order - (k + 1));
|
|
V *= betainv;
|
|
|
|
// Update A(k+1:end, k+1:end) <- A(k+1:end, k+1:end) - v*v' / alpha
|
|
ATA.block(k + 1, k + 1, order - (k + 1), order - (k + 1)) -= V.transpose() * V;
|
|
// ATA.bottomRightCorner(order-(k+1), order-(k+1)).selfadjointView<Eigen::Upper>()
|
|
// .rankUpdate(V.adjoint(), -1);
|
|
}
|
|
return 1;
|
|
} else {
|
|
// For zero pivots, add the underconstrained variable prior
|
|
ATA(k, k) = underconstrainedPrior;
|
|
for (size_t j = k + 1; j < order; ++j)
|
|
ATA(k, j) = 0.0;
|
|
return 0;
|
|
}
|
|
}
|
|
|
|
/* ************************************************************************* */
|
|
pair<size_t, bool> choleskyCareful(Matrix& ATA, int order) {
|
|
// Check that the matrix is square (we do not check for symmetry)
|
|
assert(ATA.rows() == ATA.cols());
|
|
|
|
// Number of rows/columns
|
|
const size_t n = ATA.rows();
|
|
|
|
// Negative order means factor the entire matrix
|
|
if (order < 0)
|
|
order = int(n);
|
|
|
|
assert(size_t(order) <= n);
|
|
|
|
// The index of the row after the last non-zero row of the square-root factor
|
|
size_t maxrank = 0;
|
|
bool success = true;
|
|
|
|
// Factor row-by-row
|
|
for (size_t k = 0; k < size_t(order); ++k) {
|
|
int stepResult = choleskyStep(ATA, k, size_t(order));
|
|
if (stepResult == 1) {
|
|
maxrank = k + 1;
|
|
} else if (stepResult == -1) {
|
|
success = false;
|
|
break;
|
|
} /* else if(stepResult == 0) Found zero pivot */
|
|
}
|
|
|
|
return make_pair(maxrank, success);
|
|
}
|
|
|
|
/* ************************************************************************* */
|
|
bool choleskyPartial(Matrix& ABC, size_t nFrontal, size_t topleft) {
|
|
gttic(choleskyPartial);
|
|
if (nFrontal == 0)
|
|
return true;
|
|
|
|
assert(ABC.cols() == ABC.rows());
|
|
assert(size_t(ABC.rows()) >= topleft);
|
|
const size_t n = static_cast<size_t>(ABC.rows() - topleft);
|
|
assert(nFrontal <= size_t(n));
|
|
|
|
// Create views on blocks
|
|
auto A = ABC.block(topleft, topleft, nFrontal, nFrontal);
|
|
auto B = ABC.block(topleft, topleft + nFrontal, nFrontal, n - nFrontal);
|
|
auto C = ABC.block(topleft + nFrontal, topleft + nFrontal, n - nFrontal, n - nFrontal);
|
|
|
|
// Compute Cholesky factorization A = R'*R, overwrites A.
|
|
gttic(LLT);
|
|
Eigen::LLT<Matrix, Eigen::Upper> llt(A);
|
|
Eigen::ComputationInfo lltResult = llt.info();
|
|
if (lltResult != Eigen::Success)
|
|
return false;
|
|
auto R = A.triangularView<Eigen::Upper>();
|
|
R = llt.matrixU();
|
|
gttoc(LLT);
|
|
|
|
// Compute S = inv(R') * B
|
|
gttic(compute_S);
|
|
if (nFrontal < n)
|
|
R.transpose().solveInPlace(B);
|
|
gttoc(compute_S);
|
|
|
|
// Compute L = C - S' * S
|
|
gttic(compute_L);
|
|
if (nFrontal < n)
|
|
C.selfadjointView<Eigen::Upper>().rankUpdate(B.transpose(), -1.0);
|
|
gttoc(compute_L);
|
|
|
|
// Check last diagonal element - Eigen does not check it
|
|
if (nFrontal >= 2) {
|
|
int exp2, exp1;
|
|
// NOTE(gareth): R is already the size of A, so we don't need to add topleft here.
|
|
(void)frexp(R(nFrontal - 2, nFrontal - 2), &exp2);
|
|
(void)frexp(R(nFrontal - 1, nFrontal - 1), &exp1);
|
|
return (exp2 - exp1 < underconstrainedExponentDifference);
|
|
} else if (nFrontal == 1) {
|
|
int exp1;
|
|
(void)frexp(R(0, 0), &exp1);
|
|
return (exp1 > -underconstrainedExponentDifference);
|
|
} else {
|
|
return true;
|
|
}
|
|
}
|
|
} // namespace gtsam
|