You can not select more than 25 topics
Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
284 lines
11 KiB
284 lines
11 KiB
3 months ago
|
// David Eberly, Geometric Tools, Redmond WA 98052
|
||
|
// Copyright (c) 1998-2021
|
||
|
// Distributed under the Boost Software License, Version 1.0.
|
||
|
// https://www.boost.org/LICENSE_1_0.txt
|
||
|
// https://www.geometrictools.com/License/Boost/LICENSE_1_0.txt
|
||
|
// Version: 4.0.2019.08.13
|
||
|
|
||
|
#pragma once
|
||
|
|
||
|
#include <Mathematics/CubicRootsQR.h>
|
||
|
|
||
|
// An implementation of the QR algorithm described in "Matrix Computations,
|
||
|
// 2nd edition" by G. H. Golub and C. F. Van Loan, The Johns Hopkins
|
||
|
// University Press, Baltimore MD, Fourth Printing 1993. In particular,
|
||
|
// the implementation is based on Chapter 7 (The Unsymmetric Eigenvalue
|
||
|
// Problem), Section 7.5 (The Practical QR Algorithm). The algorithm is
|
||
|
// specialized for the companion matrix associated with a quartic polynomial.
|
||
|
|
||
|
namespace gte
|
||
|
{
|
||
|
template <typename Real>
|
||
|
class QuarticRootsQR
|
||
|
{
|
||
|
public:
|
||
|
typedef std::array<std::array<Real, 4>, 4> Matrix;
|
||
|
|
||
|
// Solve p(x) = c0 + c1 * x + c2 * x^2 + c3 * x^3 + x^4 = 0.
|
||
|
uint32_t operator() (uint32_t maxIterations, Real c0, Real c1, Real c2, Real c3,
|
||
|
uint32_t& numRoots, std::array<Real, 4>& roots) const
|
||
|
{
|
||
|
// Create the companion matrix for the polynomial. The matrix is
|
||
|
// in upper Hessenberg form.
|
||
|
Matrix A;
|
||
|
A[0][0] = (Real)0;
|
||
|
A[0][1] = (Real)0;
|
||
|
A[0][2] = (Real)0;
|
||
|
A[0][3] = -c0;
|
||
|
A[1][0] = (Real)1;
|
||
|
A[1][1] = (Real)0;
|
||
|
A[1][2] = (Real)0;
|
||
|
A[1][3] = -c1;
|
||
|
A[2][0] = (Real)0;
|
||
|
A[2][1] = (Real)1;
|
||
|
A[2][2] = (Real)0;
|
||
|
A[2][3] = -c2;
|
||
|
A[3][0] = (Real)0;
|
||
|
A[3][1] = (Real)0;
|
||
|
A[3][2] = (Real)1;
|
||
|
A[3][3] = -c3;
|
||
|
|
||
|
// Avoid the QR-cycle when c1 = c2 = 0 and avoid the slow
|
||
|
// convergence when c1 and c2 are nearly zero.
|
||
|
std::array<Real, 3> V{
|
||
|
(Real)1,
|
||
|
(Real)0.36602540378443865,
|
||
|
(Real)0.36602540378443865 };
|
||
|
DoIteration(V, A);
|
||
|
|
||
|
return operator()(maxIterations, A, numRoots, roots);
|
||
|
}
|
||
|
|
||
|
// Compute the real eigenvalues of the upper Hessenberg matrix A. The
|
||
|
// matrix is modified by in-place operations, so if you need to
|
||
|
// remember A, you must make your own copy before calling this
|
||
|
// function.
|
||
|
uint32_t operator() (uint32_t maxIterations, Matrix& A,
|
||
|
uint32_t& numRoots, std::array<Real, 4>& roots) const
|
||
|
{
|
||
|
numRoots = 0;
|
||
|
std::fill(roots.begin(), roots.end(), (Real)0);
|
||
|
|
||
|
for (uint32_t numIterations = 0; numIterations < maxIterations; ++numIterations)
|
||
|
{
|
||
|
// Apply a Francis QR iteration.
|
||
|
Real tr = A[2][2] + A[3][3];
|
||
|
Real det = A[2][2] * A[3][3] - A[2][3] * A[3][2];
|
||
|
std::array<Real, 3> X{
|
||
|
A[0][0] * A[0][0] + A[0][1] * A[1][0] - tr * A[0][0] + det,
|
||
|
A[1][0] * (A[0][0] + A[1][1] - tr),
|
||
|
A[1][0] * A[2][1] };
|
||
|
std::array<Real, 3> V = House<3>(X);
|
||
|
DoIteration(V, A);
|
||
|
|
||
|
// Test for uncoupling of A.
|
||
|
Real tr12 = A[1][1] + A[2][2];
|
||
|
if (tr12 + A[2][1] == tr12)
|
||
|
{
|
||
|
GetQuadraticRoots(0, 1, A, numRoots, roots);
|
||
|
GetQuadraticRoots(2, 3, A, numRoots, roots);
|
||
|
return numIterations;
|
||
|
}
|
||
|
|
||
|
Real tr01 = A[0][0] + A[1][1];
|
||
|
if (tr01 + A[1][0] == tr01)
|
||
|
{
|
||
|
numRoots = 1;
|
||
|
roots[0] = A[0][0];
|
||
|
|
||
|
// TODO: The cubic solver is not designed to process 3x3
|
||
|
// submatrices of an NxN matrix, so the copy of a
|
||
|
// submatrix of A to B is a simple workaround for running
|
||
|
// the solver. Write general root-finding/ code that
|
||
|
// avoids such copying.
|
||
|
uint32_t subMaxIterations = maxIterations - numIterations;
|
||
|
typename CubicRootsQR<Real>::Matrix B;
|
||
|
for (int r = 0; r < 3; ++r)
|
||
|
{
|
||
|
for (int c = 0; c < 3; ++c)
|
||
|
{
|
||
|
B[r][c] = A[r + 1][c + 1];
|
||
|
}
|
||
|
}
|
||
|
|
||
|
uint32_t numSubroots = 0;
|
||
|
std::array<Real, 3> subroots;
|
||
|
uint32_t numSubiterations = CubicRootsQR<Real>()(subMaxIterations, B,
|
||
|
numSubroots, subroots);
|
||
|
for (uint32_t i = 0; i < numSubroots; ++i)
|
||
|
{
|
||
|
roots[numRoots++] = subroots[i];
|
||
|
}
|
||
|
return numIterations + numSubiterations;
|
||
|
}
|
||
|
|
||
|
Real tr23 = A[2][2] + A[3][3];
|
||
|
if (tr23 + A[3][2] == tr23)
|
||
|
{
|
||
|
numRoots = 1;
|
||
|
roots[0] = A[3][3];
|
||
|
|
||
|
// TODO: The cubic solver is not designed to process 3x3
|
||
|
// submatrices of an NxN matrix, so the copy of a
|
||
|
// submatrix of A to B is a simple workaround for running
|
||
|
// the solver. Write general root-finding/ code that
|
||
|
// avoids such copying.
|
||
|
uint32_t subMaxIterations = maxIterations - numIterations;
|
||
|
typename CubicRootsQR<Real>::Matrix B;
|
||
|
for (int r = 0; r < 3; ++r)
|
||
|
{
|
||
|
for (int c = 0; c < 3; ++c)
|
||
|
{
|
||
|
B[r][c] = A[r][c];
|
||
|
}
|
||
|
}
|
||
|
|
||
|
uint32_t numSubroots = 0;
|
||
|
std::array<Real, 3> subroots;
|
||
|
uint32_t numSubiterations = CubicRootsQR<Real>()(subMaxIterations, B,
|
||
|
numSubroots, subroots);
|
||
|
for (uint32_t i = 0; i < numSubroots; ++i)
|
||
|
{
|
||
|
roots[numRoots++] = subroots[i];
|
||
|
}
|
||
|
return numIterations + numSubiterations;
|
||
|
}
|
||
|
}
|
||
|
return maxIterations;
|
||
|
}
|
||
|
|
||
|
private:
|
||
|
void DoIteration(std::array<Real, 3> const& V, Matrix& A) const
|
||
|
{
|
||
|
Real multV = (Real)-2 / (V[0] * V[0] + V[1] * V[1] + V[2] * V[2]);
|
||
|
std::array<Real, 3> MV{ multV * V[0], multV * V[1], multV * V[2] };
|
||
|
RowHouse<3>(0, 2, 0, 3, V, MV, A);
|
||
|
ColHouse<3>(0, 3, 0, 2, V, MV, A);
|
||
|
|
||
|
std::array<Real, 3> X{ A[1][0], A[2][0], A[3][0] };
|
||
|
std::array<Real, 3> locV = House<3>(X);
|
||
|
multV = (Real)-2 / (locV[0] * locV[0] + locV[1] * locV[1] + locV[2] * locV[2]);
|
||
|
MV = { multV * locV[0], multV * locV[1], multV * locV[2] };
|
||
|
RowHouse<3>(1, 3, 0, 3, locV, MV, A);
|
||
|
ColHouse<3>(0, 3, 1, 3, locV, MV, A);
|
||
|
|
||
|
std::array<Real, 2> Y{ A[2][1], A[3][1] };
|
||
|
std::array<Real, 2> W = House<2>(Y);
|
||
|
Real multW = (Real)-2 / (W[0] * W[0] + W[1] * W[1]);
|
||
|
std::array<Real, 2> MW = { multW * W[0], multW * W[1] };
|
||
|
RowHouse<2>(2, 3, 0, 3, W, MW, A);
|
||
|
ColHouse<2>(0, 3, 2, 3, W, MW, A);
|
||
|
}
|
||
|
|
||
|
template <int N>
|
||
|
std::array<Real, N> House(std::array<Real, N> const& X) const
|
||
|
{
|
||
|
std::array<Real, N> V;
|
||
|
Real length = (Real)0;
|
||
|
for (int i = 0; i < N; ++i)
|
||
|
{
|
||
|
length += X[i] * X[i];
|
||
|
}
|
||
|
length = std::sqrt(length);
|
||
|
if (length != (Real)0)
|
||
|
{
|
||
|
Real sign = (X[0] >= (Real)0 ? (Real)1 : (Real)-1);
|
||
|
Real denom = X[0] + sign * length;
|
||
|
for (int i = 1; i < N; ++i)
|
||
|
{
|
||
|
V[i] = X[i] / denom;
|
||
|
}
|
||
|
}
|
||
|
else
|
||
|
{
|
||
|
V.fill((Real)0);
|
||
|
}
|
||
|
V[0] = (Real)1;
|
||
|
return V;
|
||
|
}
|
||
|
|
||
|
template <int N>
|
||
|
void RowHouse(int rmin, int rmax, int cmin, int cmax,
|
||
|
std::array<Real, N> const& V, std::array<Real, N> const& MV, Matrix& A) const
|
||
|
{
|
||
|
// Only elements cmin through cmax are used.
|
||
|
std::array<Real, 4> W;
|
||
|
|
||
|
for (int c = cmin; c <= cmax; ++c)
|
||
|
{
|
||
|
W[c] = (Real)0;
|
||
|
for (int r = rmin, k = 0; r <= rmax; ++r, ++k)
|
||
|
{
|
||
|
W[c] += V[k] * A[r][c];
|
||
|
}
|
||
|
}
|
||
|
|
||
|
for (int r = rmin, k = 0; r <= rmax; ++r, ++k)
|
||
|
{
|
||
|
for (int c = cmin; c <= cmax; ++c)
|
||
|
{
|
||
|
A[r][c] += MV[k] * W[c];
|
||
|
}
|
||
|
}
|
||
|
}
|
||
|
|
||
|
template <int N>
|
||
|
void ColHouse(int rmin, int rmax, int cmin, int cmax,
|
||
|
std::array<Real, N> const& V, std::array<Real, N> const& MV, Matrix& A) const
|
||
|
{
|
||
|
// Only elements rmin through rmax are used.
|
||
|
std::array<Real, 4> W;
|
||
|
|
||
|
for (int r = rmin; r <= rmax; ++r)
|
||
|
{
|
||
|
W[r] = (Real)0;
|
||
|
for (int c = cmin, k = 0; c <= cmax; ++c, ++k)
|
||
|
{
|
||
|
W[r] += V[k] * A[r][c];
|
||
|
}
|
||
|
}
|
||
|
|
||
|
for (int r = rmin; r <= rmax; ++r)
|
||
|
{
|
||
|
for (int c = cmin, k = 0; c <= cmax; ++c, ++k)
|
||
|
{
|
||
|
A[r][c] += W[r] * MV[k];
|
||
|
}
|
||
|
}
|
||
|
}
|
||
|
|
||
|
void GetQuadraticRoots(int i0, int i1, Matrix const& A,
|
||
|
uint32_t& numRoots, std::array<Real, 4>& roots) const
|
||
|
{
|
||
|
// Solve x^2 - t * x + d = 0, where t is the trace and d is the
|
||
|
// determinant of the 2x2 matrix defined by indices i0 and i1.
|
||
|
// The discriminant is D = (t/2)^2 - d. When D >= 0, the roots
|
||
|
// are real values t/2 - sqrt(D) and t/2 + sqrt(D). To avoid
|
||
|
// potential numerical issues with subtractive cancellation, the
|
||
|
// roots are computed as
|
||
|
// r0 = t/2 + sign(t/2)*sqrt(D), r1 = trace - r0
|
||
|
Real trace = A[i0][i0] + A[i1][i1];
|
||
|
Real halfTrace = trace * (Real)0.5;
|
||
|
Real determinant = A[i0][i0] * A[i1][i1] - A[i0][i1] * A[i1][i0];
|
||
|
Real discriminant = halfTrace * halfTrace - determinant;
|
||
|
if (discriminant >= (Real)0)
|
||
|
{
|
||
|
Real sign = (trace >= (Real)0 ? (Real)1 : (Real)-1);
|
||
|
Real root = halfTrace + sign * std::sqrt(discriminant);
|
||
|
roots[numRoots++] = root;
|
||
|
roots[numRoots++] = trace - root;
|
||
|
}
|
||
|
}
|
||
|
};
|
||
|
}
|