1 // This file is part of Eigen, a lightweight C++ template library
2 // for linear algebra.
3 //
4 // Copyright (C) 2012 Alexey Korepanov <kaikaikai@yandex.ru>
5 //
6 // This Source Code Form is subject to the terms of the Mozilla
7 // Public License v. 2.0. If a copy of the MPL was not distributed
8 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
9 
10 #ifndef EIGEN_REAL_QZ_H
11 #define EIGEN_REAL_QZ_H
12 
13 namespace Eigen {
14 
15   /** \eigenvalues_module \ingroup Eigenvalues_Module
16    *
17    *
18    * \class RealQZ
19    *
20    * \brief Performs a real QZ decomposition of a pair of square matrices
21    *
22    * \tparam _MatrixType the type of the matrix of which we are computing the
23    * real QZ decomposition; this is expected to be an instantiation of the
24    * Matrix class template.
25    *
26    * Given a real square matrices A and B, this class computes the real QZ
27    * decomposition: \f$ A = Q S Z \f$, \f$ B = Q T Z \f$ where Q and Z are
28    * real orthogonal matrixes, T is upper-triangular matrix, and S is upper
29    * quasi-triangular matrix. An orthogonal matrix is a matrix whose
30    * inverse is equal to its transpose, \f$ U^{-1} = U^T \f$. A quasi-triangular
31    * matrix is a block-triangular matrix whose diagonal consists of 1-by-1
32    * blocks and 2-by-2 blocks where further reduction is impossible due to
33    * complex eigenvalues.
34    *
35    * The eigenvalues of the pencil \f$ A - z B \f$ can be obtained from
36    * 1x1 and 2x2 blocks on the diagonals of S and T.
37    *
38    * Call the function compute() to compute the real QZ decomposition of a
39    * given pair of matrices. Alternatively, you can use the
40    * RealQZ(const MatrixType& B, const MatrixType& B, bool computeQZ)
41    * constructor which computes the real QZ decomposition at construction
42    * time. Once the decomposition is computed, you can use the matrixS(),
43    * matrixT(), matrixQ() and matrixZ() functions to retrieve the matrices
44    * S, T, Q and Z in the decomposition. If computeQZ==false, some time
45    * is saved by not computing matrices Q and Z.
46    *
47    * Example: \include RealQZ_compute.cpp
48    * Output: \include RealQZ_compute.out
49    *
50    * \note The implementation is based on the algorithm in "Matrix Computations"
51    * by Gene H. Golub and Charles F. Van Loan, and a paper "An algorithm for
52    * generalized eigenvalue problems" by C.B.Moler and G.W.Stewart.
53    *
54    * \sa class RealSchur, class ComplexSchur, class EigenSolver, class ComplexEigenSolver
55    */
56 
57   template<typename _MatrixType> class RealQZ
58   {
59     public:
60       typedef _MatrixType MatrixType;
61       enum {
62         RowsAtCompileTime = MatrixType::RowsAtCompileTime,
63         ColsAtCompileTime = MatrixType::ColsAtCompileTime,
64         Options = MatrixType::Options,
65         MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime,
66         MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime
67       };
68       typedef typename MatrixType::Scalar Scalar;
69       typedef std::complex<typename NumTraits<Scalar>::Real> ComplexScalar;
70       typedef typename MatrixType::Index Index;
71 
72       typedef Matrix<ComplexScalar, ColsAtCompileTime, 1, Options & ~RowMajor, MaxColsAtCompileTime, 1> EigenvalueType;
73       typedef Matrix<Scalar, ColsAtCompileTime, 1, Options & ~RowMajor, MaxColsAtCompileTime, 1> ColumnVectorType;
74 
75       /** \brief Default constructor.
76        *
77        * \param [in] size  Positive integer, size of the matrix whose QZ decomposition will be computed.
78        *
79        * The default constructor is useful in cases in which the user intends to
80        * perform decompositions via compute().  The \p size parameter is only
81        * used as a hint. It is not an error to give a wrong \p size, but it may
82        * impair performance.
83        *
84        * \sa compute() for an example.
85        */
86       RealQZ(Index size = RowsAtCompileTime==Dynamic ? 1 : RowsAtCompileTime) :
m_S(size,size)87         m_S(size, size),
88         m_T(size, size),
89         m_Q(size, size),
90         m_Z(size, size),
91         m_workspace(size*2),
92         m_maxIters(400),
93         m_isInitialized(false)
94         { }
95 
96       /** \brief Constructor; computes real QZ decomposition of given matrices
97        *
98        * \param[in]  A          Matrix A.
99        * \param[in]  B          Matrix B.
100        * \param[in]  computeQZ  If false, A and Z are not computed.
101        *
102        * This constructor calls compute() to compute the QZ decomposition.
103        */
104       RealQZ(const MatrixType& A, const MatrixType& B, bool computeQZ = true) :
105         m_S(A.rows(),A.cols()),
106         m_T(A.rows(),A.cols()),
107         m_Q(A.rows(),A.cols()),
108         m_Z(A.rows(),A.cols()),
109         m_workspace(A.rows()*2),
110         m_maxIters(400),
111         m_isInitialized(false) {
112           compute(A, B, computeQZ);
113         }
114 
115       /** \brief Returns matrix Q in the QZ decomposition.
116        *
117        * \returns A const reference to the matrix Q.
118        */
matrixQ()119       const MatrixType& matrixQ() const {
120         eigen_assert(m_isInitialized && "RealQZ is not initialized.");
121         eigen_assert(m_computeQZ && "The matrices Q and Z have not been computed during the QZ decomposition.");
122         return m_Q;
123       }
124 
125       /** \brief Returns matrix Z in the QZ decomposition.
126        *
127        * \returns A const reference to the matrix Z.
128        */
matrixZ()129       const MatrixType& matrixZ() const {
130         eigen_assert(m_isInitialized && "RealQZ is not initialized.");
131         eigen_assert(m_computeQZ && "The matrices Q and Z have not been computed during the QZ decomposition.");
132         return m_Z;
133       }
134 
135       /** \brief Returns matrix S in the QZ decomposition.
136        *
137        * \returns A const reference to the matrix S.
138        */
matrixS()139       const MatrixType& matrixS() const {
140         eigen_assert(m_isInitialized && "RealQZ is not initialized.");
141         return m_S;
142       }
143 
144       /** \brief Returns matrix S in the QZ decomposition.
145        *
146        * \returns A const reference to the matrix S.
147        */
matrixT()148       const MatrixType& matrixT() const {
149         eigen_assert(m_isInitialized && "RealQZ is not initialized.");
150         return m_T;
151       }
152 
153       /** \brief Computes QZ decomposition of given matrix.
154        *
155        * \param[in]  A          Matrix A.
156        * \param[in]  B          Matrix B.
157        * \param[in]  computeQZ  If false, A and Z are not computed.
158        * \returns    Reference to \c *this
159        */
160       RealQZ& compute(const MatrixType& A, const MatrixType& B, bool computeQZ = true);
161 
162       /** \brief Reports whether previous computation was successful.
163        *
164        * \returns \c Success if computation was succesful, \c NoConvergence otherwise.
165        */
info()166       ComputationInfo info() const
167       {
168         eigen_assert(m_isInitialized && "RealQZ is not initialized.");
169         return m_info;
170       }
171 
172       /** \brief Returns number of performed QR-like iterations.
173       */
iterations()174       Index iterations() const
175       {
176         eigen_assert(m_isInitialized && "RealQZ is not initialized.");
177         return m_global_iter;
178       }
179 
180       /** Sets the maximal number of iterations allowed to converge to one eigenvalue
181        * or decouple the problem.
182       */
setMaxIterations(Index maxIters)183       RealQZ& setMaxIterations(Index maxIters)
184       {
185         m_maxIters = maxIters;
186         return *this;
187       }
188 
189     private:
190 
191       MatrixType m_S, m_T, m_Q, m_Z;
192       Matrix<Scalar,Dynamic,1> m_workspace;
193       ComputationInfo m_info;
194       Index m_maxIters;
195       bool m_isInitialized;
196       bool m_computeQZ;
197       Scalar m_normOfT, m_normOfS;
198       Index m_global_iter;
199 
200       typedef Matrix<Scalar,3,1> Vector3s;
201       typedef Matrix<Scalar,2,1> Vector2s;
202       typedef Matrix<Scalar,2,2> Matrix2s;
203       typedef JacobiRotation<Scalar> JRs;
204 
205       void hessenbergTriangular();
206       void computeNorms();
207       Index findSmallSubdiagEntry(Index iu);
208       Index findSmallDiagEntry(Index f, Index l);
209       void splitOffTwoRows(Index i);
210       void pushDownZero(Index z, Index f, Index l);
211       void step(Index f, Index l, Index iter);
212 
213   }; // RealQZ
214 
215   /** \internal Reduces S and T to upper Hessenberg - triangular form */
216   template<typename MatrixType>
hessenbergTriangular()217     void RealQZ<MatrixType>::hessenbergTriangular()
218     {
219 
220       const Index dim = m_S.cols();
221 
222       // perform QR decomposition of T, overwrite T with R, save Q
223       HouseholderQR<MatrixType> qrT(m_T);
224       m_T = qrT.matrixQR();
225       m_T.template triangularView<StrictlyLower>().setZero();
226       m_Q = qrT.householderQ();
227       // overwrite S with Q* S
228       m_S.applyOnTheLeft(m_Q.adjoint());
229       // init Z as Identity
230       if (m_computeQZ)
231         m_Z = MatrixType::Identity(dim,dim);
232       // reduce S to upper Hessenberg with Givens rotations
233       for (Index j=0; j<=dim-3; j++) {
234         for (Index i=dim-1; i>=j+2; i--) {
235           JRs G;
236           // kill S(i,j)
237           if(m_S.coeff(i,j) != 0)
238           {
239             G.makeGivens(m_S.coeff(i-1,j), m_S.coeff(i,j), &m_S.coeffRef(i-1, j));
240             m_S.coeffRef(i,j) = Scalar(0.0);
241             m_S.rightCols(dim-j-1).applyOnTheLeft(i-1,i,G.adjoint());
242             m_T.rightCols(dim-i+1).applyOnTheLeft(i-1,i,G.adjoint());
243             // update Q
244             if (m_computeQZ)
245               m_Q.applyOnTheRight(i-1,i,G);
246           }
247           // update Q
248           if(m_T.coeff(i,i-1)!=Scalar(0))
249           {
250             G.makeGivens(m_T.coeff(i,i), m_T.coeff(i,i-1), &m_T.coeffRef(i,i));
251             m_T.coeffRef(i,i-1) = Scalar(0.0);
252             m_S.applyOnTheRight(i,i-1,G);
253             m_T.topRows(i).applyOnTheRight(i,i-1,G);
254             // update Z
255             if (m_computeQZ)
256               m_Z.applyOnTheLeft(i,i-1,G.adjoint());
257           }
258         }
259       }
260     }
261 
262   /** \internal Computes vector L1 norms of S and T when in Hessenberg-Triangular form already */
263   template<typename MatrixType>
computeNorms()264     inline void RealQZ<MatrixType>::computeNorms()
265     {
266       const Index size = m_S.cols();
267       m_normOfS = Scalar(0.0);
268       m_normOfT = Scalar(0.0);
269       for (Index j = 0; j < size; ++j)
270       {
271         m_normOfS += m_S.col(j).segment(0, (std::min)(size,j+2)).cwiseAbs().sum();
272         m_normOfT += m_T.row(j).segment(j, size - j).cwiseAbs().sum();
273       }
274     }
275 
276 
277   /** \internal Look for single small sub-diagonal element S(res, res-1) and return res (or 0) */
278   template<typename MatrixType>
findSmallSubdiagEntry(Index iu)279     inline typename MatrixType::Index RealQZ<MatrixType>::findSmallSubdiagEntry(Index iu)
280     {
281       using std::abs;
282       Index res = iu;
283       while (res > 0)
284       {
285         Scalar s = abs(m_S.coeff(res-1,res-1)) + abs(m_S.coeff(res,res));
286         if (s == Scalar(0.0))
287           s = m_normOfS;
288         if (abs(m_S.coeff(res,res-1)) < NumTraits<Scalar>::epsilon() * s)
289           break;
290         res--;
291       }
292       return res;
293     }
294 
295   /** \internal Look for single small diagonal element T(res, res) for res between f and l, and return res (or f-1)  */
296   template<typename MatrixType>
findSmallDiagEntry(Index f,Index l)297     inline typename MatrixType::Index RealQZ<MatrixType>::findSmallDiagEntry(Index f, Index l)
298     {
299       using std::abs;
300       Index res = l;
301       while (res >= f) {
302         if (abs(m_T.coeff(res,res)) <= NumTraits<Scalar>::epsilon() * m_normOfT)
303           break;
304         res--;
305       }
306       return res;
307     }
308 
309   /** \internal decouple 2x2 diagonal block in rows i, i+1 if eigenvalues are real */
310   template<typename MatrixType>
splitOffTwoRows(Index i)311     inline void RealQZ<MatrixType>::splitOffTwoRows(Index i)
312     {
313       using std::abs;
314       using std::sqrt;
315       const Index dim=m_S.cols();
316       if (abs(m_S.coeff(i+1,i))==Scalar(0))
317         return;
318       Index z = findSmallDiagEntry(i,i+1);
319       if (z==i-1)
320       {
321         // block of (S T^{-1})
322         Matrix2s STi = m_T.template block<2,2>(i,i).template triangularView<Upper>().
323           template solve<OnTheRight>(m_S.template block<2,2>(i,i));
324         Scalar p = Scalar(0.5)*(STi(0,0)-STi(1,1));
325         Scalar q = p*p + STi(1,0)*STi(0,1);
326         if (q>=0) {
327           Scalar z = sqrt(q);
328           // one QR-like iteration for ABi - lambda I
329           // is enough - when we know exact eigenvalue in advance,
330           // convergence is immediate
331           JRs G;
332           if (p>=0)
333             G.makeGivens(p + z, STi(1,0));
334           else
335             G.makeGivens(p - z, STi(1,0));
336           m_S.rightCols(dim-i).applyOnTheLeft(i,i+1,G.adjoint());
337           m_T.rightCols(dim-i).applyOnTheLeft(i,i+1,G.adjoint());
338           // update Q
339           if (m_computeQZ)
340             m_Q.applyOnTheRight(i,i+1,G);
341 
342           G.makeGivens(m_T.coeff(i+1,i+1), m_T.coeff(i+1,i));
343           m_S.topRows(i+2).applyOnTheRight(i+1,i,G);
344           m_T.topRows(i+2).applyOnTheRight(i+1,i,G);
345           // update Z
346           if (m_computeQZ)
347             m_Z.applyOnTheLeft(i+1,i,G.adjoint());
348 
349           m_S.coeffRef(i+1,i) = Scalar(0.0);
350           m_T.coeffRef(i+1,i) = Scalar(0.0);
351         }
352       }
353       else
354       {
355         pushDownZero(z,i,i+1);
356       }
357     }
358 
359   /** \internal use zero in T(z,z) to zero S(l,l-1), working in block f..l */
360   template<typename MatrixType>
pushDownZero(Index z,Index f,Index l)361     inline void RealQZ<MatrixType>::pushDownZero(Index z, Index f, Index l)
362     {
363       JRs G;
364       const Index dim = m_S.cols();
365       for (Index zz=z; zz<l; zz++)
366       {
367         // push 0 down
368         Index firstColS = zz>f ? (zz-1) : zz;
369         G.makeGivens(m_T.coeff(zz, zz+1), m_T.coeff(zz+1, zz+1));
370         m_S.rightCols(dim-firstColS).applyOnTheLeft(zz,zz+1,G.adjoint());
371         m_T.rightCols(dim-zz).applyOnTheLeft(zz,zz+1,G.adjoint());
372         m_T.coeffRef(zz+1,zz+1) = Scalar(0.0);
373         // update Q
374         if (m_computeQZ)
375           m_Q.applyOnTheRight(zz,zz+1,G);
376         // kill S(zz+1, zz-1)
377         if (zz>f)
378         {
379           G.makeGivens(m_S.coeff(zz+1, zz), m_S.coeff(zz+1,zz-1));
380           m_S.topRows(zz+2).applyOnTheRight(zz, zz-1,G);
381           m_T.topRows(zz+1).applyOnTheRight(zz, zz-1,G);
382           m_S.coeffRef(zz+1,zz-1) = Scalar(0.0);
383           // update Z
384           if (m_computeQZ)
385             m_Z.applyOnTheLeft(zz,zz-1,G.adjoint());
386         }
387       }
388       // finally kill S(l,l-1)
389       G.makeGivens(m_S.coeff(l,l), m_S.coeff(l,l-1));
390       m_S.applyOnTheRight(l,l-1,G);
391       m_T.applyOnTheRight(l,l-1,G);
392       m_S.coeffRef(l,l-1)=Scalar(0.0);
393       // update Z
394       if (m_computeQZ)
395         m_Z.applyOnTheLeft(l,l-1,G.adjoint());
396     }
397 
398   /** \internal QR-like iterative step for block f..l */
399   template<typename MatrixType>
step(Index f,Index l,Index iter)400     inline void RealQZ<MatrixType>::step(Index f, Index l, Index iter)
401     {
402       using std::abs;
403       const Index dim = m_S.cols();
404 
405       // x, y, z
406       Scalar x, y, z;
407       if (iter==10)
408       {
409         // Wilkinson ad hoc shift
410         const Scalar
411           a11=m_S.coeff(f+0,f+0), a12=m_S.coeff(f+0,f+1),
412           a21=m_S.coeff(f+1,f+0), a22=m_S.coeff(f+1,f+1), a32=m_S.coeff(f+2,f+1),
413           b12=m_T.coeff(f+0,f+1),
414           b11i=Scalar(1.0)/m_T.coeff(f+0,f+0),
415           b22i=Scalar(1.0)/m_T.coeff(f+1,f+1),
416           a87=m_S.coeff(l-1,l-2),
417           a98=m_S.coeff(l-0,l-1),
418           b77i=Scalar(1.0)/m_T.coeff(l-2,l-2),
419           b88i=Scalar(1.0)/m_T.coeff(l-1,l-1);
420         Scalar ss = abs(a87*b77i) + abs(a98*b88i),
421                lpl = Scalar(1.5)*ss,
422                ll = ss*ss;
423         x = ll + a11*a11*b11i*b11i - lpl*a11*b11i + a12*a21*b11i*b22i
424           - a11*a21*b12*b11i*b11i*b22i;
425         y = a11*a21*b11i*b11i - lpl*a21*b11i + a21*a22*b11i*b22i
426           - a21*a21*b12*b11i*b11i*b22i;
427         z = a21*a32*b11i*b22i;
428       }
429       else if (iter==16)
430       {
431         // another exceptional shift
432         x = m_S.coeff(f,f)/m_T.coeff(f,f)-m_S.coeff(l,l)/m_T.coeff(l,l) + m_S.coeff(l,l-1)*m_T.coeff(l-1,l) /
433           (m_T.coeff(l-1,l-1)*m_T.coeff(l,l));
434         y = m_S.coeff(f+1,f)/m_T.coeff(f,f);
435         z = 0;
436       }
437       else if (iter>23 && !(iter%8))
438       {
439         // extremely exceptional shift
440         x = internal::random<Scalar>(-1.0,1.0);
441         y = internal::random<Scalar>(-1.0,1.0);
442         z = internal::random<Scalar>(-1.0,1.0);
443       }
444       else
445       {
446         // Compute the shifts: (x,y,z,0...) = (AB^-1 - l1 I) (AB^-1 - l2 I) e1
447         // where l1 and l2 are the eigenvalues of the 2x2 matrix C = U V^-1 where
448         // U and V are 2x2 bottom right sub matrices of A and B. Thus:
449         //  = AB^-1AB^-1 + l1 l2 I - (l1+l2)(AB^-1)
450         //  = AB^-1AB^-1 + det(M) - tr(M)(AB^-1)
451         // Since we are only interested in having x, y, z with a correct ratio, we have:
452         const Scalar
453           a11 = m_S.coeff(f,f),     a12 = m_S.coeff(f,f+1),
454           a21 = m_S.coeff(f+1,f),   a22 = m_S.coeff(f+1,f+1),
455                                     a32 = m_S.coeff(f+2,f+1),
456 
457           a88 = m_S.coeff(l-1,l-1), a89 = m_S.coeff(l-1,l),
458           a98 = m_S.coeff(l,l-1),   a99 = m_S.coeff(l,l),
459 
460           b11 = m_T.coeff(f,f),     b12 = m_T.coeff(f,f+1),
461                                     b22 = m_T.coeff(f+1,f+1),
462 
463           b88 = m_T.coeff(l-1,l-1), b89 = m_T.coeff(l-1,l),
464                                     b99 = m_T.coeff(l,l);
465 
466         x = ( (a88/b88 - a11/b11)*(a99/b99 - a11/b11) - (a89/b99)*(a98/b88) + (a98/b88)*(b89/b99)*(a11/b11) ) * (b11/a21)
467           + a12/b22 - (a11/b11)*(b12/b22);
468         y = (a22/b22-a11/b11) - (a21/b11)*(b12/b22) - (a88/b88-a11/b11) - (a99/b99-a11/b11) + (a98/b88)*(b89/b99);
469         z = a32/b22;
470       }
471 
472       JRs G;
473 
474       for (Index k=f; k<=l-2; k++)
475       {
476         // variables for Householder reflections
477         Vector2s essential2;
478         Scalar tau, beta;
479 
480         Vector3s hr(x,y,z);
481 
482         // Q_k to annihilate S(k+1,k-1) and S(k+2,k-1)
483         hr.makeHouseholderInPlace(tau, beta);
484         essential2 = hr.template bottomRows<2>();
485         Index fc=(std::max)(k-1,Index(0));  // first col to update
486         m_S.template middleRows<3>(k).rightCols(dim-fc).applyHouseholderOnTheLeft(essential2, tau, m_workspace.data());
487         m_T.template middleRows<3>(k).rightCols(dim-fc).applyHouseholderOnTheLeft(essential2, tau, m_workspace.data());
488         if (m_computeQZ)
489           m_Q.template middleCols<3>(k).applyHouseholderOnTheRight(essential2, tau, m_workspace.data());
490         if (k>f)
491           m_S.coeffRef(k+2,k-1) = m_S.coeffRef(k+1,k-1) = Scalar(0.0);
492 
493         // Z_{k1} to annihilate T(k+2,k+1) and T(k+2,k)
494         hr << m_T.coeff(k+2,k+2),m_T.coeff(k+2,k),m_T.coeff(k+2,k+1);
495         hr.makeHouseholderInPlace(tau, beta);
496         essential2 = hr.template bottomRows<2>();
497         {
498           Index lr = (std::min)(k+4,dim); // last row to update
499           Map<Matrix<Scalar,Dynamic,1> > tmp(m_workspace.data(),lr);
500           // S
501           tmp = m_S.template middleCols<2>(k).topRows(lr) * essential2;
502           tmp += m_S.col(k+2).head(lr);
503           m_S.col(k+2).head(lr) -= tau*tmp;
504           m_S.template middleCols<2>(k).topRows(lr) -= (tau*tmp) * essential2.adjoint();
505           // T
506           tmp = m_T.template middleCols<2>(k).topRows(lr) * essential2;
507           tmp += m_T.col(k+2).head(lr);
508           m_T.col(k+2).head(lr) -= tau*tmp;
509           m_T.template middleCols<2>(k).topRows(lr) -= (tau*tmp) * essential2.adjoint();
510         }
511         if (m_computeQZ)
512         {
513           // Z
514           Map<Matrix<Scalar,1,Dynamic> > tmp(m_workspace.data(),dim);
515           tmp = essential2.adjoint()*(m_Z.template middleRows<2>(k));
516           tmp += m_Z.row(k+2);
517           m_Z.row(k+2) -= tau*tmp;
518           m_Z.template middleRows<2>(k) -= essential2 * (tau*tmp);
519         }
520         m_T.coeffRef(k+2,k) = m_T.coeffRef(k+2,k+1) = Scalar(0.0);
521 
522         // Z_{k2} to annihilate T(k+1,k)
523         G.makeGivens(m_T.coeff(k+1,k+1), m_T.coeff(k+1,k));
524         m_S.applyOnTheRight(k+1,k,G);
525         m_T.applyOnTheRight(k+1,k,G);
526         // update Z
527         if (m_computeQZ)
528           m_Z.applyOnTheLeft(k+1,k,G.adjoint());
529         m_T.coeffRef(k+1,k) = Scalar(0.0);
530 
531         // update x,y,z
532         x = m_S.coeff(k+1,k);
533         y = m_S.coeff(k+2,k);
534         if (k < l-2)
535           z = m_S.coeff(k+3,k);
536       } // loop over k
537 
538       // Q_{n-1} to annihilate y = S(l,l-2)
539       G.makeGivens(x,y);
540       m_S.applyOnTheLeft(l-1,l,G.adjoint());
541       m_T.applyOnTheLeft(l-1,l,G.adjoint());
542       if (m_computeQZ)
543         m_Q.applyOnTheRight(l-1,l,G);
544       m_S.coeffRef(l,l-2) = Scalar(0.0);
545 
546       // Z_{n-1} to annihilate T(l,l-1)
547       G.makeGivens(m_T.coeff(l,l),m_T.coeff(l,l-1));
548       m_S.applyOnTheRight(l,l-1,G);
549       m_T.applyOnTheRight(l,l-1,G);
550       if (m_computeQZ)
551         m_Z.applyOnTheLeft(l,l-1,G.adjoint());
552       m_T.coeffRef(l,l-1) = Scalar(0.0);
553     }
554 
555 
556   template<typename MatrixType>
compute(const MatrixType & A_in,const MatrixType & B_in,bool computeQZ)557     RealQZ<MatrixType>& RealQZ<MatrixType>::compute(const MatrixType& A_in, const MatrixType& B_in, bool computeQZ)
558     {
559 
560       const Index dim = A_in.cols();
561 
562       eigen_assert (A_in.rows()==dim && A_in.cols()==dim
563           && B_in.rows()==dim && B_in.cols()==dim
564           && "Need square matrices of the same dimension");
565 
566       m_isInitialized = true;
567       m_computeQZ = computeQZ;
568       m_S = A_in; m_T = B_in;
569       m_workspace.resize(dim*2);
570       m_global_iter = 0;
571 
572       // entrance point: hessenberg triangular decomposition
573       hessenbergTriangular();
574       // compute L1 vector norms of T, S into m_normOfS, m_normOfT
575       computeNorms();
576 
577       Index l = dim-1,
578             f,
579             local_iter = 0;
580 
581       while (l>0 && local_iter<m_maxIters)
582       {
583         f = findSmallSubdiagEntry(l);
584         // now rows and columns f..l (including) decouple from the rest of the problem
585         if (f>0) m_S.coeffRef(f,f-1) = Scalar(0.0);
586         if (f == l) // One root found
587         {
588           l--;
589           local_iter = 0;
590         }
591         else if (f == l-1) // Two roots found
592         {
593           splitOffTwoRows(f);
594           l -= 2;
595           local_iter = 0;
596         }
597         else // No convergence yet
598         {
599           // if there's zero on diagonal of T, we can isolate an eigenvalue with Givens rotations
600           Index z = findSmallDiagEntry(f,l);
601           if (z>=f)
602           {
603             // zero found
604             pushDownZero(z,f,l);
605           }
606           else
607           {
608             // We are sure now that S.block(f,f, l-f+1,l-f+1) is underuced upper-Hessenberg
609             // and T.block(f,f, l-f+1,l-f+1) is invertible uper-triangular, which allows to
610             // apply a QR-like iteration to rows and columns f..l.
611             step(f,l, local_iter);
612             local_iter++;
613             m_global_iter++;
614           }
615         }
616       }
617       // check if we converged before reaching iterations limit
618       m_info = (local_iter<m_maxIters) ? Success : NoConvergence;
619       return *this;
620     } // end compute
621 
622 } // end namespace Eigen
623 
624 #endif //EIGEN_REAL_QZ
625