• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // This file is part of Eigen, a lightweight C++ template library
2 // for linear algebra.
3 //
4 // Copyright (C) 2008 Gael Guennebaud <gael.guennebaud@inria.fr>
5 // Copyright (C) 2010 Jitse Niesen <jitse@maths.leeds.ac.uk>
6 //
7 // This Source Code Form is subject to the terms of the Mozilla
8 // Public License v. 2.0. If a copy of the MPL was not distributed
9 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
10 
11 #ifndef EIGEN_REAL_SCHUR_H
12 #define EIGEN_REAL_SCHUR_H
13 
14 #include "./HessenbergDecomposition.h"
15 
16 namespace Eigen {
17 
18 /** \eigenvalues_module \ingroup Eigenvalues_Module
19   *
20   *
21   * \class RealSchur
22   *
23   * \brief Performs a real Schur decomposition of a square matrix
24   *
25   * \tparam _MatrixType the type of the matrix of which we are computing the
26   * real Schur decomposition; this is expected to be an instantiation of the
27   * Matrix class template.
28   *
29   * Given a real square matrix A, this class computes the real Schur
30   * decomposition: \f$ A = U T U^T \f$ where U is a real orthogonal matrix and
31   * T is a real quasi-triangular matrix. An orthogonal matrix is a matrix whose
32   * inverse is equal to its transpose, \f$ U^{-1} = U^T \f$. A quasi-triangular
33   * matrix is a block-triangular matrix whose diagonal consists of 1-by-1
34   * blocks and 2-by-2 blocks with complex eigenvalues. The eigenvalues of the
35   * blocks on the diagonal of T are the same as the eigenvalues of the matrix
36   * A, and thus the real Schur decomposition is used in EigenSolver to compute
37   * the eigendecomposition of a matrix.
38   *
39   * Call the function compute() to compute the real Schur decomposition of a
40   * given matrix. Alternatively, you can use the RealSchur(const MatrixType&, bool)
41   * constructor which computes the real Schur decomposition at construction
42   * time. Once the decomposition is computed, you can use the matrixU() and
43   * matrixT() functions to retrieve the matrices U and T in the decomposition.
44   *
45   * The documentation of RealSchur(const MatrixType&, bool) contains an example
46   * of the typical use of this class.
47   *
48   * \note The implementation is adapted from
49   * <a href="http://math.nist.gov/javanumerics/jama/">JAMA</a> (public domain).
50   * Their code is based on EISPACK.
51   *
52   * \sa class ComplexSchur, class EigenSolver, class ComplexEigenSolver
53   */
54 template<typename _MatrixType> class RealSchur
55 {
56   public:
57     typedef _MatrixType MatrixType;
58     enum {
59       RowsAtCompileTime = MatrixType::RowsAtCompileTime,
60       ColsAtCompileTime = MatrixType::ColsAtCompileTime,
61       Options = MatrixType::Options,
62       MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime,
63       MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime
64     };
65     typedef typename MatrixType::Scalar Scalar;
66     typedef std::complex<typename NumTraits<Scalar>::Real> ComplexScalar;
67     typedef typename MatrixType::Index Index;
68 
69     typedef Matrix<ComplexScalar, ColsAtCompileTime, 1, Options & ~RowMajor, MaxColsAtCompileTime, 1> EigenvalueType;
70     typedef Matrix<Scalar, ColsAtCompileTime, 1, Options & ~RowMajor, MaxColsAtCompileTime, 1> ColumnVectorType;
71 
72     /** \brief Default constructor.
73       *
74       * \param [in] size  Positive integer, size of the matrix whose Schur decomposition will be computed.
75       *
76       * The default constructor is useful in cases in which the user intends to
77       * perform decompositions via compute().  The \p size parameter is only
78       * used as a hint. It is not an error to give a wrong \p size, but it may
79       * impair performance.
80       *
81       * \sa compute() for an example.
82       */
83     RealSchur(Index size = RowsAtCompileTime==Dynamic ? 1 : RowsAtCompileTime)
m_matT(size,size)84             : m_matT(size, size),
85               m_matU(size, size),
86               m_workspaceVector(size),
87               m_hess(size),
88               m_isInitialized(false),
89               m_matUisUptodate(false)
90     { }
91 
92     /** \brief Constructor; computes real Schur decomposition of given matrix.
93       *
94       * \param[in]  matrix    Square matrix whose Schur decomposition is to be computed.
95       * \param[in]  computeU  If true, both T and U are computed; if false, only T is computed.
96       *
97       * This constructor calls compute() to compute the Schur decomposition.
98       *
99       * Example: \include RealSchur_RealSchur_MatrixType.cpp
100       * Output: \verbinclude RealSchur_RealSchur_MatrixType.out
101       */
102     RealSchur(const MatrixType& matrix, bool computeU = true)
103             : m_matT(matrix.rows(),matrix.cols()),
104               m_matU(matrix.rows(),matrix.cols()),
105               m_workspaceVector(matrix.rows()),
106               m_hess(matrix.rows()),
107               m_isInitialized(false),
108               m_matUisUptodate(false)
109     {
110       compute(matrix, computeU);
111     }
112 
113     /** \brief Returns the orthogonal matrix in the Schur decomposition.
114       *
115       * \returns A const reference to the matrix U.
116       *
117       * \pre Either the constructor RealSchur(const MatrixType&, bool) or the
118       * member function compute(const MatrixType&, bool) has been called before
119       * to compute the Schur decomposition of a matrix, and \p computeU was set
120       * to true (the default value).
121       *
122       * \sa RealSchur(const MatrixType&, bool) for an example
123       */
matrixU()124     const MatrixType& matrixU() const
125     {
126       eigen_assert(m_isInitialized && "RealSchur is not initialized.");
127       eigen_assert(m_matUisUptodate && "The matrix U has not been computed during the RealSchur decomposition.");
128       return m_matU;
129     }
130 
131     /** \brief Returns the quasi-triangular matrix in the Schur decomposition.
132       *
133       * \returns A const reference to the matrix T.
134       *
135       * \pre Either the constructor RealSchur(const MatrixType&, bool) or the
136       * member function compute(const MatrixType&, bool) has been called before
137       * to compute the Schur decomposition of a matrix.
138       *
139       * \sa RealSchur(const MatrixType&, bool) for an example
140       */
matrixT()141     const MatrixType& matrixT() const
142     {
143       eigen_assert(m_isInitialized && "RealSchur is not initialized.");
144       return m_matT;
145     }
146 
147     /** \brief Computes Schur decomposition of given matrix.
148       *
149       * \param[in]  matrix    Square matrix whose Schur decomposition is to be computed.
150       * \param[in]  computeU  If true, both T and U are computed; if false, only T is computed.
151       * \returns    Reference to \c *this
152       *
153       * The Schur decomposition is computed by first reducing the matrix to
154       * Hessenberg form using the class HessenbergDecomposition. The Hessenberg
155       * matrix is then reduced to triangular form by performing Francis QR
156       * iterations with implicit double shift. The cost of computing the Schur
157       * decomposition depends on the number of iterations; as a rough guide, it
158       * may be taken to be \f$25n^3\f$ flops if \a computeU is true and
159       * \f$10n^3\f$ flops if \a computeU is false.
160       *
161       * Example: \include RealSchur_compute.cpp
162       * Output: \verbinclude RealSchur_compute.out
163       */
164     RealSchur& compute(const MatrixType& matrix, bool computeU = true);
165 
166     /** \brief Reports whether previous computation was successful.
167       *
168       * \returns \c Success if computation was succesful, \c NoConvergence otherwise.
169       */
info()170     ComputationInfo info() const
171     {
172       eigen_assert(m_isInitialized && "RealSchur is not initialized.");
173       return m_info;
174     }
175 
176     /** \brief Maximum number of iterations.
177       *
178       * Maximum number of iterations allowed for an eigenvalue to converge.
179       */
180     static const int m_maxIterations = 40;
181 
182   private:
183 
184     MatrixType m_matT;
185     MatrixType m_matU;
186     ColumnVectorType m_workspaceVector;
187     HessenbergDecomposition<MatrixType> m_hess;
188     ComputationInfo m_info;
189     bool m_isInitialized;
190     bool m_matUisUptodate;
191 
192     typedef Matrix<Scalar,3,1> Vector3s;
193 
194     Scalar computeNormOfT();
195     Index findSmallSubdiagEntry(Index iu, Scalar norm);
196     void splitOffTwoRows(Index iu, bool computeU, Scalar exshift);
197     void computeShift(Index iu, Index iter, Scalar& exshift, Vector3s& shiftInfo);
198     void initFrancisQRStep(Index il, Index iu, const Vector3s& shiftInfo, Index& im, Vector3s& firstHouseholderVector);
199     void performFrancisQRStep(Index il, Index im, Index iu, bool computeU, const Vector3s& firstHouseholderVector, Scalar* workspace);
200 };
201 
202 
203 template<typename MatrixType>
compute(const MatrixType & matrix,bool computeU)204 RealSchur<MatrixType>& RealSchur<MatrixType>::compute(const MatrixType& matrix, bool computeU)
205 {
206   assert(matrix.cols() == matrix.rows());
207 
208   // Step 1. Reduce to Hessenberg form
209   m_hess.compute(matrix);
210   m_matT = m_hess.matrixH();
211   if (computeU)
212     m_matU = m_hess.matrixQ();
213 
214   // Step 2. Reduce to real Schur form
215   m_workspaceVector.resize(m_matT.cols());
216   Scalar* workspace = &m_workspaceVector.coeffRef(0);
217 
218   // The matrix m_matT is divided in three parts.
219   // Rows 0,...,il-1 are decoupled from the rest because m_matT(il,il-1) is zero.
220   // Rows il,...,iu is the part we are working on (the active window).
221   // Rows iu+1,...,end are already brought in triangular form.
222   Index iu = m_matT.cols() - 1;
223   Index iter = 0; // iteration count
224   Scalar exshift(0); // sum of exceptional shifts
225   Scalar norm = computeNormOfT();
226 
227   if(norm!=0)
228   {
229     while (iu >= 0)
230     {
231       Index il = findSmallSubdiagEntry(iu, norm);
232 
233       // Check for convergence
234       if (il == iu) // One root found
235       {
236         m_matT.coeffRef(iu,iu) = m_matT.coeff(iu,iu) + exshift;
237         if (iu > 0)
238           m_matT.coeffRef(iu, iu-1) = Scalar(0);
239         iu--;
240         iter = 0;
241       }
242       else if (il == iu-1) // Two roots found
243       {
244         splitOffTwoRows(iu, computeU, exshift);
245         iu -= 2;
246         iter = 0;
247       }
248       else // No convergence yet
249       {
250         // The firstHouseholderVector vector has to be initialized to something to get rid of a silly GCC warning (-O1 -Wall -DNDEBUG )
251         Vector3s firstHouseholderVector(0,0,0), shiftInfo;
252         computeShift(iu, iter, exshift, shiftInfo);
253         iter = iter + 1;
254         if (iter > m_maxIterations * m_matT.cols()) break;
255         Index im;
256         initFrancisQRStep(il, iu, shiftInfo, im, firstHouseholderVector);
257         performFrancisQRStep(il, im, iu, computeU, firstHouseholderVector, workspace);
258       }
259     }
260   }
261   if(iter <= m_maxIterations * m_matT.cols())
262     m_info = Success;
263   else
264     m_info = NoConvergence;
265 
266   m_isInitialized = true;
267   m_matUisUptodate = computeU;
268   return *this;
269 }
270 
271 /** \internal Computes and returns vector L1 norm of T */
272 template<typename MatrixType>
computeNormOfT()273 inline typename MatrixType::Scalar RealSchur<MatrixType>::computeNormOfT()
274 {
275   const Index size = m_matT.cols();
276   // FIXME to be efficient the following would requires a triangular reduxion code
277   // Scalar norm = m_matT.upper().cwiseAbs().sum()
278   //               + m_matT.bottomLeftCorner(size-1,size-1).diagonal().cwiseAbs().sum();
279   Scalar norm(0);
280   for (Index j = 0; j < size; ++j)
281     norm += m_matT.row(j).segment((std::max)(j-1,Index(0)), size-(std::max)(j-1,Index(0))).cwiseAbs().sum();
282   return norm;
283 }
284 
285 /** \internal Look for single small sub-diagonal element and returns its index */
286 template<typename MatrixType>
findSmallSubdiagEntry(Index iu,Scalar norm)287 inline typename MatrixType::Index RealSchur<MatrixType>::findSmallSubdiagEntry(Index iu, Scalar norm)
288 {
289   Index res = iu;
290   while (res > 0)
291   {
292     Scalar s = internal::abs(m_matT.coeff(res-1,res-1)) + internal::abs(m_matT.coeff(res,res));
293     if (s == 0.0)
294       s = norm;
295     if (internal::abs(m_matT.coeff(res,res-1)) < NumTraits<Scalar>::epsilon() * s)
296       break;
297     res--;
298   }
299   return res;
300 }
301 
302 /** \internal Update T given that rows iu-1 and iu decouple from the rest. */
303 template<typename MatrixType>
splitOffTwoRows(Index iu,bool computeU,Scalar exshift)304 inline void RealSchur<MatrixType>::splitOffTwoRows(Index iu, bool computeU, Scalar exshift)
305 {
306   const Index size = m_matT.cols();
307 
308   // The eigenvalues of the 2x2 matrix [a b; c d] are
309   // trace +/- sqrt(discr/4) where discr = tr^2 - 4*det, tr = a + d, det = ad - bc
310   Scalar p = Scalar(0.5) * (m_matT.coeff(iu-1,iu-1) - m_matT.coeff(iu,iu));
311   Scalar q = p * p + m_matT.coeff(iu,iu-1) * m_matT.coeff(iu-1,iu);   // q = tr^2 / 4 - det = discr/4
312   m_matT.coeffRef(iu,iu) += exshift;
313   m_matT.coeffRef(iu-1,iu-1) += exshift;
314 
315   if (q >= Scalar(0)) // Two real eigenvalues
316   {
317     Scalar z = internal::sqrt(internal::abs(q));
318     JacobiRotation<Scalar> rot;
319     if (p >= Scalar(0))
320       rot.makeGivens(p + z, m_matT.coeff(iu, iu-1));
321     else
322       rot.makeGivens(p - z, m_matT.coeff(iu, iu-1));
323 
324     m_matT.rightCols(size-iu+1).applyOnTheLeft(iu-1, iu, rot.adjoint());
325     m_matT.topRows(iu+1).applyOnTheRight(iu-1, iu, rot);
326     m_matT.coeffRef(iu, iu-1) = Scalar(0);
327     if (computeU)
328       m_matU.applyOnTheRight(iu-1, iu, rot);
329   }
330 
331   if (iu > 1)
332     m_matT.coeffRef(iu-1, iu-2) = Scalar(0);
333 }
334 
335 /** \internal Form shift in shiftInfo, and update exshift if an exceptional shift is performed. */
336 template<typename MatrixType>
computeShift(Index iu,Index iter,Scalar & exshift,Vector3s & shiftInfo)337 inline void RealSchur<MatrixType>::computeShift(Index iu, Index iter, Scalar& exshift, Vector3s& shiftInfo)
338 {
339   shiftInfo.coeffRef(0) = m_matT.coeff(iu,iu);
340   shiftInfo.coeffRef(1) = m_matT.coeff(iu-1,iu-1);
341   shiftInfo.coeffRef(2) = m_matT.coeff(iu,iu-1) * m_matT.coeff(iu-1,iu);
342 
343   // Wilkinson's original ad hoc shift
344   if (iter == 10)
345   {
346     exshift += shiftInfo.coeff(0);
347     for (Index i = 0; i <= iu; ++i)
348       m_matT.coeffRef(i,i) -= shiftInfo.coeff(0);
349     Scalar s = internal::abs(m_matT.coeff(iu,iu-1)) + internal::abs(m_matT.coeff(iu-1,iu-2));
350     shiftInfo.coeffRef(0) = Scalar(0.75) * s;
351     shiftInfo.coeffRef(1) = Scalar(0.75) * s;
352     shiftInfo.coeffRef(2) = Scalar(-0.4375) * s * s;
353   }
354 
355   // MATLAB's new ad hoc shift
356   if (iter == 30)
357   {
358     Scalar s = (shiftInfo.coeff(1) - shiftInfo.coeff(0)) / Scalar(2.0);
359     s = s * s + shiftInfo.coeff(2);
360     if (s > Scalar(0))
361     {
362       s = internal::sqrt(s);
363       if (shiftInfo.coeff(1) < shiftInfo.coeff(0))
364         s = -s;
365       s = s + (shiftInfo.coeff(1) - shiftInfo.coeff(0)) / Scalar(2.0);
366       s = shiftInfo.coeff(0) - shiftInfo.coeff(2) / s;
367       exshift += s;
368       for (Index i = 0; i <= iu; ++i)
369         m_matT.coeffRef(i,i) -= s;
370       shiftInfo.setConstant(Scalar(0.964));
371     }
372   }
373 }
374 
375 /** \internal Compute index im at which Francis QR step starts and the first Householder vector. */
376 template<typename MatrixType>
initFrancisQRStep(Index il,Index iu,const Vector3s & shiftInfo,Index & im,Vector3s & firstHouseholderVector)377 inline void RealSchur<MatrixType>::initFrancisQRStep(Index il, Index iu, const Vector3s& shiftInfo, Index& im, Vector3s& firstHouseholderVector)
378 {
379   Vector3s& v = firstHouseholderVector; // alias to save typing
380 
381   for (im = iu-2; im >= il; --im)
382   {
383     const Scalar Tmm = m_matT.coeff(im,im);
384     const Scalar r = shiftInfo.coeff(0) - Tmm;
385     const Scalar s = shiftInfo.coeff(1) - Tmm;
386     v.coeffRef(0) = (r * s - shiftInfo.coeff(2)) / m_matT.coeff(im+1,im) + m_matT.coeff(im,im+1);
387     v.coeffRef(1) = m_matT.coeff(im+1,im+1) - Tmm - r - s;
388     v.coeffRef(2) = m_matT.coeff(im+2,im+1);
389     if (im == il) {
390       break;
391     }
392     const Scalar lhs = m_matT.coeff(im,im-1) * (internal::abs(v.coeff(1)) + internal::abs(v.coeff(2)));
393     const Scalar rhs = v.coeff(0) * (internal::abs(m_matT.coeff(im-1,im-1)) + internal::abs(Tmm) + internal::abs(m_matT.coeff(im+1,im+1)));
394     if (internal::abs(lhs) < NumTraits<Scalar>::epsilon() * rhs)
395     {
396       break;
397     }
398   }
399 }
400 
401 /** \internal Perform a Francis QR step involving rows il:iu and columns im:iu. */
402 template<typename MatrixType>
performFrancisQRStep(Index il,Index im,Index iu,bool computeU,const Vector3s & firstHouseholderVector,Scalar * workspace)403 inline void RealSchur<MatrixType>::performFrancisQRStep(Index il, Index im, Index iu, bool computeU, const Vector3s& firstHouseholderVector, Scalar* workspace)
404 {
405   assert(im >= il);
406   assert(im <= iu-2);
407 
408   const Index size = m_matT.cols();
409 
410   for (Index k = im; k <= iu-2; ++k)
411   {
412     bool firstIteration = (k == im);
413 
414     Vector3s v;
415     if (firstIteration)
416       v = firstHouseholderVector;
417     else
418       v = m_matT.template block<3,1>(k,k-1);
419 
420     Scalar tau, beta;
421     Matrix<Scalar, 2, 1> ess;
422     v.makeHouseholder(ess, tau, beta);
423 
424     if (beta != Scalar(0)) // if v is not zero
425     {
426       if (firstIteration && k > il)
427         m_matT.coeffRef(k,k-1) = -m_matT.coeff(k,k-1);
428       else if (!firstIteration)
429         m_matT.coeffRef(k,k-1) = beta;
430 
431       // These Householder transformations form the O(n^3) part of the algorithm
432       m_matT.block(k, k, 3, size-k).applyHouseholderOnTheLeft(ess, tau, workspace);
433       m_matT.block(0, k, (std::min)(iu,k+3) + 1, 3).applyHouseholderOnTheRight(ess, tau, workspace);
434       if (computeU)
435         m_matU.block(0, k, size, 3).applyHouseholderOnTheRight(ess, tau, workspace);
436     }
437   }
438 
439   Matrix<Scalar, 2, 1> v = m_matT.template block<2,1>(iu-1, iu-2);
440   Scalar tau, beta;
441   Matrix<Scalar, 1, 1> ess;
442   v.makeHouseholder(ess, tau, beta);
443 
444   if (beta != Scalar(0)) // if v is not zero
445   {
446     m_matT.coeffRef(iu-1, iu-2) = beta;
447     m_matT.block(iu-1, iu-1, 2, size-iu+1).applyHouseholderOnTheLeft(ess, tau, workspace);
448     m_matT.block(0, iu-1, iu+1, 2).applyHouseholderOnTheRight(ess, tau, workspace);
449     if (computeU)
450       m_matU.block(0, iu-1, size, 2).applyHouseholderOnTheRight(ess, tau, workspace);
451   }
452 
453   // clean up pollution due to round-off errors
454   for (Index i = im+2; i <= iu; ++i)
455   {
456     m_matT.coeffRef(i,i-2) = Scalar(0);
457     if (i > im+2)
458       m_matT.coeffRef(i,i-3) = Scalar(0);
459   }
460 }
461 
462 } // end namespace Eigen
463 
464 #endif // EIGEN_REAL_SCHUR_H
465