• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // This file is part of Eigen, a lightweight C++ template library
2 // for linear algebra.
3 //
4 // We used the "A Divide-And-Conquer Algorithm for the Bidiagonal SVD"
5 // research report written by Ming Gu and Stanley C.Eisenstat
6 // The code variable names correspond to the names they used in their
7 // report
8 //
9 // Copyright (C) 2013 Gauthier Brun <brun.gauthier@gmail.com>
10 // Copyright (C) 2013 Nicolas Carre <nicolas.carre@ensimag.fr>
11 // Copyright (C) 2013 Jean Ceccato <jean.ceccato@ensimag.fr>
12 // Copyright (C) 2013 Pierre Zoppitelli <pierre.zoppitelli@ensimag.fr>
13 // Copyright (C) 2013 Jitse Niesen <jitse@maths.leeds.ac.uk>
14 // Copyright (C) 2014-2016 Gael Guennebaud <gael.guennebaud@inria.fr>
15 //
16 // Source Code Form is subject to the terms of the Mozilla
17 // Public License v. 2.0. If a copy of the MPL was not distributed
18 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
19 
20 #ifndef EIGEN_BDCSVD_H
21 #define EIGEN_BDCSVD_H
22 // #define EIGEN_BDCSVD_DEBUG_VERBOSE
23 // #define EIGEN_BDCSVD_SANITY_CHECKS
24 
25 namespace Eigen {
26 
27 #ifdef EIGEN_BDCSVD_DEBUG_VERBOSE
28 IOFormat bdcsvdfmt(8, 0, ", ", "\n", "  [", "]");
29 #endif
30 
31 template<typename _MatrixType> class BDCSVD;
32 
33 namespace internal {
34 
35 template<typename _MatrixType>
36 struct traits<BDCSVD<_MatrixType> >
37 {
38   typedef _MatrixType MatrixType;
39 };
40 
41 } // end namespace internal
42 
43 
44 /** \ingroup SVD_Module
45  *
46  *
47  * \class BDCSVD
48  *
49  * \brief class Bidiagonal Divide and Conquer SVD
50  *
51  * \tparam _MatrixType the type of the matrix of which we are computing the SVD decomposition
52  *
53  * This class first reduces the input matrix to bi-diagonal form using class UpperBidiagonalization,
54  * and then performs a divide-and-conquer diagonalization. Small blocks are diagonalized using class JacobiSVD.
55  * You can control the switching size with the setSwitchSize() method, default is 16.
56  * For small matrice (<16), it is thus preferable to directly use JacobiSVD. For larger ones, BDCSVD is highly
57  * recommended and can several order of magnitude faster.
58  *
59  * \warning this algorithm is unlikely to provide accurate result when compiled with unsafe math optimizations.
60  * For instance, this concerns Intel's compiler (ICC), which perfroms such optimization by default unless
61  * you compile with the \c -fp-model \c precise option. Likewise, the \c -ffast-math option of GCC or clang will
62  * significantly degrade the accuracy.
63  *
64  * \sa class JacobiSVD
65  */
66 template<typename _MatrixType>
67 class BDCSVD : public SVDBase<BDCSVD<_MatrixType> >
68 {
69   typedef SVDBase<BDCSVD> Base;
70 
71 public:
72   using Base::rows;
73   using Base::cols;
74   using Base::computeU;
75   using Base::computeV;
76 
77   typedef _MatrixType MatrixType;
78   typedef typename MatrixType::Scalar Scalar;
79   typedef typename NumTraits<typename MatrixType::Scalar>::Real RealScalar;
80   typedef typename NumTraits<RealScalar>::Literal Literal;
81   enum {
82     RowsAtCompileTime = MatrixType::RowsAtCompileTime,
83     ColsAtCompileTime = MatrixType::ColsAtCompileTime,
84     DiagSizeAtCompileTime = EIGEN_SIZE_MIN_PREFER_DYNAMIC(RowsAtCompileTime, ColsAtCompileTime),
85     MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime,
86     MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime,
87     MaxDiagSizeAtCompileTime = EIGEN_SIZE_MIN_PREFER_FIXED(MaxRowsAtCompileTime, MaxColsAtCompileTime),
88     MatrixOptions = MatrixType::Options
89   };
90 
91   typedef typename Base::MatrixUType MatrixUType;
92   typedef typename Base::MatrixVType MatrixVType;
93   typedef typename Base::SingularValuesType SingularValuesType;
94 
95   typedef Matrix<Scalar, Dynamic, Dynamic, ColMajor> MatrixX;
96   typedef Matrix<RealScalar, Dynamic, Dynamic, ColMajor> MatrixXr;
97   typedef Matrix<RealScalar, Dynamic, 1> VectorType;
98   typedef Array<RealScalar, Dynamic, 1> ArrayXr;
99   typedef Array<Index,1,Dynamic> ArrayXi;
100   typedef Ref<ArrayXr> ArrayRef;
101   typedef Ref<ArrayXi> IndicesRef;
102 
103   /** \brief Default Constructor.
104    *
105    * The default constructor is useful in cases in which the user intends to
106    * perform decompositions via BDCSVD::compute(const MatrixType&).
107    */
108   BDCSVD() : m_algoswap(16), m_numIters(0)
109   {}
110 
111 
112   /** \brief Default Constructor with memory preallocation
113    *
114    * Like the default constructor but with preallocation of the internal data
115    * according to the specified problem size.
116    * \sa BDCSVD()
117    */
118   BDCSVD(Index rows, Index cols, unsigned int computationOptions = 0)
119     : m_algoswap(16), m_numIters(0)
120   {
121     allocate(rows, cols, computationOptions);
122   }
123 
124   /** \brief Constructor performing the decomposition of given matrix.
125    *
126    * \param matrix the matrix to decompose
127    * \param computationOptions optional parameter allowing to specify if you want full or thin U or V unitaries to be computed.
128    *                           By default, none is computed. This is a bit - field, the possible bits are #ComputeFullU, #ComputeThinU,
129    *                           #ComputeFullV, #ComputeThinV.
130    *
131    * Thin unitaries are only available if your matrix type has a Dynamic number of columns (for example MatrixXf). They also are not
132    * available with the (non - default) FullPivHouseholderQR preconditioner.
133    */
134   BDCSVD(const MatrixType& matrix, unsigned int computationOptions = 0)
135     : m_algoswap(16), m_numIters(0)
136   {
137     compute(matrix, computationOptions);
138   }
139 
140   ~BDCSVD()
141   {
142   }
143 
144   /** \brief Method performing the decomposition of given matrix using custom options.
145    *
146    * \param matrix the matrix to decompose
147    * \param computationOptions optional parameter allowing to specify if you want full or thin U or V unitaries to be computed.
148    *                           By default, none is computed. This is a bit - field, the possible bits are #ComputeFullU, #ComputeThinU,
149    *                           #ComputeFullV, #ComputeThinV.
150    *
151    * Thin unitaries are only available if your matrix type has a Dynamic number of columns (for example MatrixXf). They also are not
152    * available with the (non - default) FullPivHouseholderQR preconditioner.
153    */
154   BDCSVD& compute(const MatrixType& matrix, unsigned int computationOptions);
155 
156   /** \brief Method performing the decomposition of given matrix using current options.
157    *
158    * \param matrix the matrix to decompose
159    *
160    * This method uses the current \a computationOptions, as already passed to the constructor or to compute(const MatrixType&, unsigned int).
161    */
162   BDCSVD& compute(const MatrixType& matrix)
163   {
164     return compute(matrix, this->m_computationOptions);
165   }
166 
167   void setSwitchSize(int s)
168   {
169     eigen_assert(s>3 && "BDCSVD the size of the algo switch has to be greater than 3");
170     m_algoswap = s;
171   }
172 
173 private:
174   void allocate(Index rows, Index cols, unsigned int computationOptions);
175   void divide(Index firstCol, Index lastCol, Index firstRowW, Index firstColW, Index shift);
176   void computeSVDofM(Index firstCol, Index n, MatrixXr& U, VectorType& singVals, MatrixXr& V);
177   void computeSingVals(const ArrayRef& col0, const ArrayRef& diag, const IndicesRef& perm, VectorType& singVals, ArrayRef shifts, ArrayRef mus);
178   void perturbCol0(const ArrayRef& col0, const ArrayRef& diag, const IndicesRef& perm, const VectorType& singVals, const ArrayRef& shifts, const ArrayRef& mus, ArrayRef zhat);
179   void computeSingVecs(const ArrayRef& zhat, const ArrayRef& diag, const IndicesRef& perm, const VectorType& singVals, const ArrayRef& shifts, const ArrayRef& mus, MatrixXr& U, MatrixXr& V);
180   void deflation43(Index firstCol, Index shift, Index i, Index size);
181   void deflation44(Index firstColu , Index firstColm, Index firstRowW, Index firstColW, Index i, Index j, Index size);
182   void deflation(Index firstCol, Index lastCol, Index k, Index firstRowW, Index firstColW, Index shift);
183   template<typename HouseholderU, typename HouseholderV, typename NaiveU, typename NaiveV>
184   void copyUV(const HouseholderU &householderU, const HouseholderV &householderV, const NaiveU &naiveU, const NaiveV &naivev);
185   void structured_update(Block<MatrixXr,Dynamic,Dynamic> A, const MatrixXr &B, Index n1);
186   static RealScalar secularEq(RealScalar x, const ArrayRef& col0, const ArrayRef& diag, const IndicesRef &perm, const ArrayRef& diagShifted, RealScalar shift);
187 
188 protected:
189   MatrixXr m_naiveU, m_naiveV;
190   MatrixXr m_computed;
191   Index m_nRec;
192   ArrayXr m_workspace;
193   ArrayXi m_workspaceI;
194   int m_algoswap;
195   bool m_isTranspose, m_compU, m_compV;
196 
197   using Base::m_singularValues;
198   using Base::m_diagSize;
199   using Base::m_computeFullU;
200   using Base::m_computeFullV;
201   using Base::m_computeThinU;
202   using Base::m_computeThinV;
203   using Base::m_matrixU;
204   using Base::m_matrixV;
205   using Base::m_isInitialized;
206   using Base::m_nonzeroSingularValues;
207 
208 public:
209   int m_numIters;
210 }; //end class BDCSVD
211 
212 
213 // Method to allocate and initialize matrix and attributes
214 template<typename MatrixType>
215 void BDCSVD<MatrixType>::allocate(Index rows, Index cols, unsigned int computationOptions)
216 {
217   m_isTranspose = (cols > rows);
218 
219   if (Base::allocate(rows, cols, computationOptions))
220     return;
221 
222   m_computed = MatrixXr::Zero(m_diagSize + 1, m_diagSize );
223   m_compU = computeV();
224   m_compV = computeU();
225   if (m_isTranspose)
226     std::swap(m_compU, m_compV);
227 
228   if (m_compU) m_naiveU = MatrixXr::Zero(m_diagSize + 1, m_diagSize + 1 );
229   else         m_naiveU = MatrixXr::Zero(2, m_diagSize + 1 );
230 
231   if (m_compV) m_naiveV = MatrixXr::Zero(m_diagSize, m_diagSize);
232 
233   m_workspace.resize((m_diagSize+1)*(m_diagSize+1)*3);
234   m_workspaceI.resize(3*m_diagSize);
235 }// end allocate
236 
237 template<typename MatrixType>
238 BDCSVD<MatrixType>& BDCSVD<MatrixType>::compute(const MatrixType& matrix, unsigned int computationOptions)
239 {
240 #ifdef EIGEN_BDCSVD_DEBUG_VERBOSE
241   std::cout << "\n\n\n======================================================================================================================\n\n\n";
242 #endif
243   allocate(matrix.rows(), matrix.cols(), computationOptions);
244   using std::abs;
245 
246   const RealScalar considerZero = (std::numeric_limits<RealScalar>::min)();
247 
248   //**** step -1 - If the problem is too small, directly falls back to JacobiSVD and return
249   if(matrix.cols() < m_algoswap)
250   {
251     // FIXME this line involves temporaries
252     JacobiSVD<MatrixType> jsvd(matrix,computationOptions);
253     if(computeU()) m_matrixU = jsvd.matrixU();
254     if(computeV()) m_matrixV = jsvd.matrixV();
255     m_singularValues = jsvd.singularValues();
256     m_nonzeroSingularValues = jsvd.nonzeroSingularValues();
257     m_isInitialized = true;
258     return *this;
259   }
260 
261   //**** step 0 - Copy the input matrix and apply scaling to reduce over/under-flows
262   RealScalar scale = matrix.cwiseAbs().maxCoeff();
263   if(scale==Literal(0)) scale = Literal(1);
264   MatrixX copy;
265   if (m_isTranspose) copy = matrix.adjoint()/scale;
266   else               copy = matrix/scale;
267 
268   //**** step 1 - Bidiagonalization
269   // FIXME this line involves temporaries
270   internal::UpperBidiagonalization<MatrixX> bid(copy);
271 
272   //**** step 2 - Divide & Conquer
273   m_naiveU.setZero();
274   m_naiveV.setZero();
275   // FIXME this line involves a temporary matrix
276   m_computed.topRows(m_diagSize) = bid.bidiagonal().toDenseMatrix().transpose();
277   m_computed.template bottomRows<1>().setZero();
278   divide(0, m_diagSize - 1, 0, 0, 0);
279 
280   //**** step 3 - Copy singular values and vectors
281   for (int i=0; i<m_diagSize; i++)
282   {
283     RealScalar a = abs(m_computed.coeff(i, i));
284     m_singularValues.coeffRef(i) = a * scale;
285     if (a<considerZero)
286     {
287       m_nonzeroSingularValues = i;
288       m_singularValues.tail(m_diagSize - i - 1).setZero();
289       break;
290     }
291     else if (i == m_diagSize - 1)
292     {
293       m_nonzeroSingularValues = i + 1;
294       break;
295     }
296   }
297 
298 #ifdef EIGEN_BDCSVD_DEBUG_VERBOSE
299 //   std::cout << "m_naiveU\n" << m_naiveU << "\n\n";
300 //   std::cout << "m_naiveV\n" << m_naiveV << "\n\n";
301 #endif
302   if(m_isTranspose) copyUV(bid.householderV(), bid.householderU(), m_naiveV, m_naiveU);
303   else              copyUV(bid.householderU(), bid.householderV(), m_naiveU, m_naiveV);
304 
305   m_isInitialized = true;
306   return *this;
307 }// end compute
308 
309 
310 template<typename MatrixType>
311 template<typename HouseholderU, typename HouseholderV, typename NaiveU, typename NaiveV>
312 void BDCSVD<MatrixType>::copyUV(const HouseholderU &householderU, const HouseholderV &householderV, const NaiveU &naiveU, const NaiveV &naiveV)
313 {
314   // Note exchange of U and V: m_matrixU is set from m_naiveV and vice versa
315   if (computeU())
316   {
317     Index Ucols = m_computeThinU ? m_diagSize : householderU.cols();
318     m_matrixU = MatrixX::Identity(householderU.cols(), Ucols);
319     m_matrixU.topLeftCorner(m_diagSize, m_diagSize) = naiveV.template cast<Scalar>().topLeftCorner(m_diagSize, m_diagSize);
320     householderU.applyThisOnTheLeft(m_matrixU); // FIXME this line involves a temporary buffer
321   }
322   if (computeV())
323   {
324     Index Vcols = m_computeThinV ? m_diagSize : householderV.cols();
325     m_matrixV = MatrixX::Identity(householderV.cols(), Vcols);
326     m_matrixV.topLeftCorner(m_diagSize, m_diagSize) = naiveU.template cast<Scalar>().topLeftCorner(m_diagSize, m_diagSize);
327     householderV.applyThisOnTheLeft(m_matrixV); // FIXME this line involves a temporary buffer
328   }
329 }
330 
331 /** \internal
332   * Performs A = A * B exploiting the special structure of the matrix A. Splitting A as:
333   *  A = [A1]
334   *      [A2]
335   * such that A1.rows()==n1, then we assume that at least half of the columns of A1 and A2 are zeros.
336   * We can thus pack them prior to the the matrix product. However, this is only worth the effort if the matrix is large
337   * enough.
338   */
339 template<typename MatrixType>
340 void BDCSVD<MatrixType>::structured_update(Block<MatrixXr,Dynamic,Dynamic> A, const MatrixXr &B, Index n1)
341 {
342   Index n = A.rows();
343   if(n>100)
344   {
345     // If the matrices are large enough, let's exploit the sparse structure of A by
346     // splitting it in half (wrt n1), and packing the non-zero columns.
347     Index n2 = n - n1;
348     Map<MatrixXr> A1(m_workspace.data()      , n1, n);
349     Map<MatrixXr> A2(m_workspace.data()+ n1*n, n2, n);
350     Map<MatrixXr> B1(m_workspace.data()+  n*n, n,  n);
351     Map<MatrixXr> B2(m_workspace.data()+2*n*n, n,  n);
352     Index k1=0, k2=0;
353     for(Index j=0; j<n; ++j)
354     {
355       if( (A.col(j).head(n1).array()!=Literal(0)).any() )
356       {
357         A1.col(k1) = A.col(j).head(n1);
358         B1.row(k1) = B.row(j);
359         ++k1;
360       }
361       if( (A.col(j).tail(n2).array()!=Literal(0)).any() )
362       {
363         A2.col(k2) = A.col(j).tail(n2);
364         B2.row(k2) = B.row(j);
365         ++k2;
366       }
367     }
368 
369     A.topRows(n1).noalias()    = A1.leftCols(k1) * B1.topRows(k1);
370     A.bottomRows(n2).noalias() = A2.leftCols(k2) * B2.topRows(k2);
371   }
372   else
373   {
374     Map<MatrixXr,Aligned> tmp(m_workspace.data(),n,n);
375     tmp.noalias() = A*B;
376     A = tmp;
377   }
378 }
379 
380 // The divide algorithm is done "in place", we are always working on subsets of the same matrix. The divide methods takes as argument the
381 // place of the submatrix we are currently working on.
382 
383 //@param firstCol : The Index of the first column of the submatrix of m_computed and for m_naiveU;
384 //@param lastCol : The Index of the last column of the submatrix of m_computed and for m_naiveU;
385 // lastCol + 1 - firstCol is the size of the submatrix.
386 //@param firstRowW : The Index of the first row of the matrix W that we are to change. (see the reference paper section 1 for more information on W)
387 //@param firstRowW : Same as firstRowW with the column.
388 //@param shift : Each time one takes the left submatrix, one must add 1 to the shift. Why? Because! We actually want the last column of the U submatrix
389 // to become the first column (*coeff) and to shift all the other columns to the right. There are more details on the reference paper.
390 template<typename MatrixType>
391 void BDCSVD<MatrixType>::divide (Index firstCol, Index lastCol, Index firstRowW, Index firstColW, Index shift)
392 {
393   // requires rows = cols + 1;
394   using std::pow;
395   using std::sqrt;
396   using std::abs;
397   const Index n = lastCol - firstCol + 1;
398   const Index k = n/2;
399   const RealScalar considerZero = (std::numeric_limits<RealScalar>::min)();
400   RealScalar alphaK;
401   RealScalar betaK;
402   RealScalar r0;
403   RealScalar lambda, phi, c0, s0;
404   VectorType l, f;
405   // We use the other algorithm which is more efficient for small
406   // matrices.
407   if (n < m_algoswap)
408   {
409     // FIXME this line involves temporaries
410     JacobiSVD<MatrixXr> b(m_computed.block(firstCol, firstCol, n + 1, n), ComputeFullU | (m_compV ? ComputeFullV : 0));
411     if (m_compU)
412       m_naiveU.block(firstCol, firstCol, n + 1, n + 1).real() = b.matrixU();
413     else
414     {
415       m_naiveU.row(0).segment(firstCol, n + 1).real() = b.matrixU().row(0);
416       m_naiveU.row(1).segment(firstCol, n + 1).real() = b.matrixU().row(n);
417     }
418     if (m_compV) m_naiveV.block(firstRowW, firstColW, n, n).real() = b.matrixV();
419     m_computed.block(firstCol + shift, firstCol + shift, n + 1, n).setZero();
420     m_computed.diagonal().segment(firstCol + shift, n) = b.singularValues().head(n);
421     return;
422   }
423   // We use the divide and conquer algorithm
424   alphaK =  m_computed(firstCol + k, firstCol + k);
425   betaK = m_computed(firstCol + k + 1, firstCol + k);
426   // The divide must be done in that order in order to have good results. Divide change the data inside the submatrices
427   // and the divide of the right submatrice reads one column of the left submatrice. That's why we need to treat the
428   // right submatrix before the left one.
429   divide(k + 1 + firstCol, lastCol, k + 1 + firstRowW, k + 1 + firstColW, shift);
430   divide(firstCol, k - 1 + firstCol, firstRowW, firstColW + 1, shift + 1);
431 
432   if (m_compU)
433   {
434     lambda = m_naiveU(firstCol + k, firstCol + k);
435     phi = m_naiveU(firstCol + k + 1, lastCol + 1);
436   }
437   else
438   {
439     lambda = m_naiveU(1, firstCol + k);
440     phi = m_naiveU(0, lastCol + 1);
441   }
442   r0 = sqrt((abs(alphaK * lambda) * abs(alphaK * lambda)) + abs(betaK * phi) * abs(betaK * phi));
443   if (m_compU)
444   {
445     l = m_naiveU.row(firstCol + k).segment(firstCol, k);
446     f = m_naiveU.row(firstCol + k + 1).segment(firstCol + k + 1, n - k - 1);
447   }
448   else
449   {
450     l = m_naiveU.row(1).segment(firstCol, k);
451     f = m_naiveU.row(0).segment(firstCol + k + 1, n - k - 1);
452   }
453   if (m_compV) m_naiveV(firstRowW+k, firstColW) = Literal(1);
454   if (r0<considerZero)
455   {
456     c0 = Literal(1);
457     s0 = Literal(0);
458   }
459   else
460   {
461     c0 = alphaK * lambda / r0;
462     s0 = betaK * phi / r0;
463   }
464 
465 #ifdef EIGEN_BDCSVD_SANITY_CHECKS
466   assert(m_naiveU.allFinite());
467   assert(m_naiveV.allFinite());
468   assert(m_computed.allFinite());
469 #endif
470 
471   if (m_compU)
472   {
473     MatrixXr q1 (m_naiveU.col(firstCol + k).segment(firstCol, k + 1));
474     // we shiftW Q1 to the right
475     for (Index i = firstCol + k - 1; i >= firstCol; i--)
476       m_naiveU.col(i + 1).segment(firstCol, k + 1) = m_naiveU.col(i).segment(firstCol, k + 1);
477     // we shift q1 at the left with a factor c0
478     m_naiveU.col(firstCol).segment( firstCol, k + 1) = (q1 * c0);
479     // last column = q1 * - s0
480     m_naiveU.col(lastCol + 1).segment(firstCol, k + 1) = (q1 * ( - s0));
481     // first column = q2 * s0
482     m_naiveU.col(firstCol).segment(firstCol + k + 1, n - k) = m_naiveU.col(lastCol + 1).segment(firstCol + k + 1, n - k) * s0;
483     // q2 *= c0
484     m_naiveU.col(lastCol + 1).segment(firstCol + k + 1, n - k) *= c0;
485   }
486   else
487   {
488     RealScalar q1 = m_naiveU(0, firstCol + k);
489     // we shift Q1 to the right
490     for (Index i = firstCol + k - 1; i >= firstCol; i--)
491       m_naiveU(0, i + 1) = m_naiveU(0, i);
492     // we shift q1 at the left with a factor c0
493     m_naiveU(0, firstCol) = (q1 * c0);
494     // last column = q1 * - s0
495     m_naiveU(0, lastCol + 1) = (q1 * ( - s0));
496     // first column = q2 * s0
497     m_naiveU(1, firstCol) = m_naiveU(1, lastCol + 1) *s0;
498     // q2 *= c0
499     m_naiveU(1, lastCol + 1) *= c0;
500     m_naiveU.row(1).segment(firstCol + 1, k).setZero();
501     m_naiveU.row(0).segment(firstCol + k + 1, n - k - 1).setZero();
502   }
503 
504 #ifdef EIGEN_BDCSVD_SANITY_CHECKS
505   assert(m_naiveU.allFinite());
506   assert(m_naiveV.allFinite());
507   assert(m_computed.allFinite());
508 #endif
509 
510   m_computed(firstCol + shift, firstCol + shift) = r0;
511   m_computed.col(firstCol + shift).segment(firstCol + shift + 1, k) = alphaK * l.transpose().real();
512   m_computed.col(firstCol + shift).segment(firstCol + shift + k + 1, n - k - 1) = betaK * f.transpose().real();
513 
514 #ifdef EIGEN_BDCSVD_DEBUG_VERBOSE
515   ArrayXr tmp1 = (m_computed.block(firstCol+shift, firstCol+shift, n, n)).jacobiSvd().singularValues();
516 #endif
517   // Second part: try to deflate singular values in combined matrix
518   deflation(firstCol, lastCol, k, firstRowW, firstColW, shift);
519 #ifdef EIGEN_BDCSVD_DEBUG_VERBOSE
520   ArrayXr tmp2 = (m_computed.block(firstCol+shift, firstCol+shift, n, n)).jacobiSvd().singularValues();
521   std::cout << "\n\nj1 = " << tmp1.transpose().format(bdcsvdfmt) << "\n";
522   std::cout << "j2 = " << tmp2.transpose().format(bdcsvdfmt) << "\n\n";
523   std::cout << "err:      " << ((tmp1-tmp2).abs()>1e-12*tmp2.abs()).transpose() << "\n";
524   static int count = 0;
525   std::cout << "# " << ++count << "\n\n";
526   assert((tmp1-tmp2).matrix().norm() < 1e-14*tmp2.matrix().norm());
527 //   assert(count<681);
528 //   assert(((tmp1-tmp2).abs()<1e-13*tmp2.abs()).all());
529 #endif
530 
531   // Third part: compute SVD of combined matrix
532   MatrixXr UofSVD, VofSVD;
533   VectorType singVals;
534   computeSVDofM(firstCol + shift, n, UofSVD, singVals, VofSVD);
535 
536 #ifdef EIGEN_BDCSVD_SANITY_CHECKS
537   assert(UofSVD.allFinite());
538   assert(VofSVD.allFinite());
539 #endif
540 
541   if (m_compU)
542     structured_update(m_naiveU.block(firstCol, firstCol, n + 1, n + 1), UofSVD, (n+2)/2);
543   else
544   {
545     Map<Matrix<RealScalar,2,Dynamic>,Aligned> tmp(m_workspace.data(),2,n+1);
546     tmp.noalias() = m_naiveU.middleCols(firstCol, n+1) * UofSVD;
547     m_naiveU.middleCols(firstCol, n + 1) = tmp;
548   }
549 
550   if (m_compV)  structured_update(m_naiveV.block(firstRowW, firstColW, n, n), VofSVD, (n+1)/2);
551 
552 #ifdef EIGEN_BDCSVD_SANITY_CHECKS
553   assert(m_naiveU.allFinite());
554   assert(m_naiveV.allFinite());
555   assert(m_computed.allFinite());
556 #endif
557 
558   m_computed.block(firstCol + shift, firstCol + shift, n, n).setZero();
559   m_computed.block(firstCol + shift, firstCol + shift, n, n).diagonal() = singVals;
560 }// end divide
561 
562 // Compute SVD of m_computed.block(firstCol, firstCol, n + 1, n); this block only has non-zeros in
563 // the first column and on the diagonal and has undergone deflation, so diagonal is in increasing
564 // order except for possibly the (0,0) entry. The computed SVD is stored U, singVals and V, except
565 // that if m_compV is false, then V is not computed. Singular values are sorted in decreasing order.
566 //
567 // TODO Opportunities for optimization: better root finding algo, better stopping criterion, better
568 // handling of round-off errors, be consistent in ordering
569 // For instance, to solve the secular equation using FMM, see http://www.stat.uchicago.edu/~lekheng/courses/302/classics/greengard-rokhlin.pdf
570 template <typename MatrixType>
571 void BDCSVD<MatrixType>::computeSVDofM(Index firstCol, Index n, MatrixXr& U, VectorType& singVals, MatrixXr& V)
572 {
573   const RealScalar considerZero = (std::numeric_limits<RealScalar>::min)();
574   using std::abs;
575   ArrayRef col0 = m_computed.col(firstCol).segment(firstCol, n);
576   m_workspace.head(n) =  m_computed.block(firstCol, firstCol, n, n).diagonal();
577   ArrayRef diag = m_workspace.head(n);
578   diag(0) = Literal(0);
579 
580   // Allocate space for singular values and vectors
581   singVals.resize(n);
582   U.resize(n+1, n+1);
583   if (m_compV) V.resize(n, n);
584 
585 #ifdef EIGEN_BDCSVD_DEBUG_VERBOSE
586   if (col0.hasNaN() || diag.hasNaN())
587     std::cout << "\n\nHAS NAN\n\n";
588 #endif
589 
590   // Many singular values might have been deflated, the zero ones have been moved to the end,
591   // but others are interleaved and we must ignore them at this stage.
592   // To this end, let's compute a permutation skipping them:
593   Index actual_n = n;
594   while(actual_n>1 && diag(actual_n-1)==Literal(0)) --actual_n;
595   Index m = 0; // size of the deflated problem
596   for(Index k=0;k<actual_n;++k)
597     if(abs(col0(k))>considerZero)
598       m_workspaceI(m++) = k;
599   Map<ArrayXi> perm(m_workspaceI.data(),m);
600 
601   Map<ArrayXr> shifts(m_workspace.data()+1*n, n);
602   Map<ArrayXr> mus(m_workspace.data()+2*n, n);
603   Map<ArrayXr> zhat(m_workspace.data()+3*n, n);
604 
605 #ifdef EIGEN_BDCSVD_DEBUG_VERBOSE
606   std::cout << "computeSVDofM using:\n";
607   std::cout << "  z: " << col0.transpose() << "\n";
608   std::cout << "  d: " << diag.transpose() << "\n";
609 #endif
610 
611   // Compute singVals, shifts, and mus
612   computeSingVals(col0, diag, perm, singVals, shifts, mus);
613 
614 #ifdef EIGEN_BDCSVD_DEBUG_VERBOSE
615   std::cout << "  j:        " << (m_computed.block(firstCol, firstCol, n, n)).jacobiSvd().singularValues().transpose().reverse() << "\n\n";
616   std::cout << "  sing-val: " << singVals.transpose() << "\n";
617   std::cout << "  mu:       " << mus.transpose() << "\n";
618   std::cout << "  shift:    " << shifts.transpose() << "\n";
619 
620   {
621     Index actual_n = n;
622     while(actual_n>1 && abs(col0(actual_n-1))<considerZero) --actual_n;
623     std::cout << "\n\n    mus:    " << mus.head(actual_n).transpose() << "\n\n";
624     std::cout << "    check1 (expect0) : " << ((singVals.array()-(shifts+mus)) / singVals.array()).head(actual_n).transpose() << "\n\n";
625     std::cout << "    check2 (>0)      : " << ((singVals.array()-diag) / singVals.array()).head(actual_n).transpose() << "\n\n";
626     std::cout << "    check3 (>0)      : " << ((diag.segment(1,actual_n-1)-singVals.head(actual_n-1).array()) / singVals.head(actual_n-1).array()).transpose() << "\n\n\n";
627     std::cout << "    check4 (>0)      : " << ((singVals.segment(1,actual_n-1)-singVals.head(actual_n-1))).transpose() << "\n\n\n";
628   }
629 #endif
630 
631 #ifdef EIGEN_BDCSVD_SANITY_CHECKS
632   assert(singVals.allFinite());
633   assert(mus.allFinite());
634   assert(shifts.allFinite());
635 #endif
636 
637   // Compute zhat
638   perturbCol0(col0, diag, perm, singVals, shifts, mus, zhat);
639 #ifdef  EIGEN_BDCSVD_DEBUG_VERBOSE
640   std::cout << "  zhat: " << zhat.transpose() << "\n";
641 #endif
642 
643 #ifdef EIGEN_BDCSVD_SANITY_CHECKS
644   assert(zhat.allFinite());
645 #endif
646 
647   computeSingVecs(zhat, diag, perm, singVals, shifts, mus, U, V);
648 
649 #ifdef  EIGEN_BDCSVD_DEBUG_VERBOSE
650   std::cout << "U^T U: " << (U.transpose() * U - MatrixXr(MatrixXr::Identity(U.cols(),U.cols()))).norm() << "\n";
651   std::cout << "V^T V: " << (V.transpose() * V - MatrixXr(MatrixXr::Identity(V.cols(),V.cols()))).norm() << "\n";
652 #endif
653 
654 #ifdef EIGEN_BDCSVD_SANITY_CHECKS
655   assert(U.allFinite());
656   assert(V.allFinite());
657   assert((U.transpose() * U - MatrixXr(MatrixXr::Identity(U.cols(),U.cols()))).norm() < 1e-14 * n);
658   assert((V.transpose() * V - MatrixXr(MatrixXr::Identity(V.cols(),V.cols()))).norm() < 1e-14 * n);
659   assert(m_naiveU.allFinite());
660   assert(m_naiveV.allFinite());
661   assert(m_computed.allFinite());
662 #endif
663 
664   // Because of deflation, the singular values might not be completely sorted.
665   // Fortunately, reordering them is a O(n) problem
666   for(Index i=0; i<actual_n-1; ++i)
667   {
668     if(singVals(i)>singVals(i+1))
669     {
670       using std::swap;
671       swap(singVals(i),singVals(i+1));
672       U.col(i).swap(U.col(i+1));
673       if(m_compV) V.col(i).swap(V.col(i+1));
674     }
675   }
676 
677   // Reverse order so that singular values in increased order
678   // Because of deflation, the zeros singular-values are already at the end
679   singVals.head(actual_n).reverseInPlace();
680   U.leftCols(actual_n).rowwise().reverseInPlace();
681   if (m_compV) V.leftCols(actual_n).rowwise().reverseInPlace();
682 
683 #ifdef EIGEN_BDCSVD_DEBUG_VERBOSE
684   JacobiSVD<MatrixXr> jsvd(m_computed.block(firstCol, firstCol, n, n) );
685   std::cout << "  * j:        " << jsvd.singularValues().transpose() << "\n\n";
686   std::cout << "  * sing-val: " << singVals.transpose() << "\n";
687 //   std::cout << "  * err:      " << ((jsvd.singularValues()-singVals)>1e-13*singVals.norm()).transpose() << "\n";
688 #endif
689 }
690 
691 template <typename MatrixType>
692 typename BDCSVD<MatrixType>::RealScalar BDCSVD<MatrixType>::secularEq(RealScalar mu, const ArrayRef& col0, const ArrayRef& diag, const IndicesRef &perm, const ArrayRef& diagShifted, RealScalar shift)
693 {
694   Index m = perm.size();
695   RealScalar res = Literal(1);
696   for(Index i=0; i<m; ++i)
697   {
698     Index j = perm(i);
699     res += numext::abs2(col0(j)) / ((diagShifted(j) - mu) * (diag(j) + shift + mu));
700   }
701   return res;
702 
703 }
704 
705 template <typename MatrixType>
706 void BDCSVD<MatrixType>::computeSingVals(const ArrayRef& col0, const ArrayRef& diag, const IndicesRef &perm,
707                                          VectorType& singVals, ArrayRef shifts, ArrayRef mus)
708 {
709   using std::abs;
710   using std::swap;
711 
712   Index n = col0.size();
713   Index actual_n = n;
714   while(actual_n>1 && col0(actual_n-1)==Literal(0)) --actual_n;
715 
716   for (Index k = 0; k < n; ++k)
717   {
718     if (col0(k) == Literal(0) || actual_n==1)
719     {
720       // if col0(k) == 0, then entry is deflated, so singular value is on diagonal
721       // if actual_n==1, then the deflated problem is already diagonalized
722       singVals(k) = k==0 ? col0(0) : diag(k);
723       mus(k) = Literal(0);
724       shifts(k) = k==0 ? col0(0) : diag(k);
725       continue;
726     }
727 
728     // otherwise, use secular equation to find singular value
729     RealScalar left = diag(k);
730     RealScalar right; // was: = (k != actual_n-1) ? diag(k+1) : (diag(actual_n-1) + col0.matrix().norm());
731     if(k==actual_n-1)
732       right = (diag(actual_n-1) + col0.matrix().norm());
733     else
734     {
735       // Skip deflated singular values
736       Index l = k+1;
737       while(col0(l)==Literal(0)) { ++l; eigen_internal_assert(l<actual_n); }
738       right = diag(l);
739     }
740 
741     // first decide whether it's closer to the left end or the right end
742     RealScalar mid = left + (right-left) / Literal(2);
743     RealScalar fMid = secularEq(mid, col0, diag, perm, diag, Literal(0));
744 #ifdef EIGEN_BDCSVD_DEBUG_VERBOSE
745     std::cout << right-left << "\n";
746     std::cout << "fMid = " << fMid << " " << secularEq(mid-left, col0, diag, perm, diag-left, left) << " " << secularEq(mid-right, col0, diag, perm, diag-right, right)   << "\n";
747     std::cout << "     = " << secularEq(0.1*(left+right), col0, diag, perm, diag, 0)
748               << " "       << secularEq(0.2*(left+right), col0, diag, perm, diag, 0)
749               << " "       << secularEq(0.3*(left+right), col0, diag, perm, diag, 0)
750               << " "       << secularEq(0.4*(left+right), col0, diag, perm, diag, 0)
751               << " "       << secularEq(0.49*(left+right), col0, diag, perm, diag, 0)
752               << " "       << secularEq(0.5*(left+right), col0, diag, perm, diag, 0)
753               << " "       << secularEq(0.51*(left+right), col0, diag, perm, diag, 0)
754               << " "       << secularEq(0.6*(left+right), col0, diag, perm, diag, 0)
755               << " "       << secularEq(0.7*(left+right), col0, diag, perm, diag, 0)
756               << " "       << secularEq(0.8*(left+right), col0, diag, perm, diag, 0)
757               << " "       << secularEq(0.9*(left+right), col0, diag, perm, diag, 0) << "\n";
758 #endif
759     RealScalar shift = (k == actual_n-1 || fMid > Literal(0)) ? left : right;
760 
761     // measure everything relative to shift
762     Map<ArrayXr> diagShifted(m_workspace.data()+4*n, n);
763     diagShifted = diag - shift;
764 
765     // initial guess
766     RealScalar muPrev, muCur;
767     if (shift == left)
768     {
769       muPrev = (right - left) * RealScalar(0.1);
770       if (k == actual_n-1) muCur = right - left;
771       else                 muCur = (right - left) * RealScalar(0.5);
772     }
773     else
774     {
775       muPrev = -(right - left) * RealScalar(0.1);
776       muCur = -(right - left) * RealScalar(0.5);
777     }
778 
779     RealScalar fPrev = secularEq(muPrev, col0, diag, perm, diagShifted, shift);
780     RealScalar fCur = secularEq(muCur, col0, diag, perm, diagShifted, shift);
781     if (abs(fPrev) < abs(fCur))
782     {
783       swap(fPrev, fCur);
784       swap(muPrev, muCur);
785     }
786 
787     // rational interpolation: fit a function of the form a / mu + b through the two previous
788     // iterates and use its zero to compute the next iterate
789     bool useBisection = fPrev*fCur>Literal(0);
790     while (fCur!=Literal(0) && abs(muCur - muPrev) > Literal(8) * NumTraits<RealScalar>::epsilon() * numext::maxi<RealScalar>(abs(muCur), abs(muPrev)) && abs(fCur - fPrev)>NumTraits<RealScalar>::epsilon() && !useBisection)
791     {
792       ++m_numIters;
793 
794       // Find a and b such that the function f(mu) = a / mu + b matches the current and previous samples.
795       RealScalar a = (fCur - fPrev) / (Literal(1)/muCur - Literal(1)/muPrev);
796       RealScalar b = fCur - a / muCur;
797       // And find mu such that f(mu)==0:
798       RealScalar muZero = -a/b;
799       RealScalar fZero = secularEq(muZero, col0, diag, perm, diagShifted, shift);
800 
801       muPrev = muCur;
802       fPrev = fCur;
803       muCur = muZero;
804       fCur = fZero;
805 
806 
807       if (shift == left  && (muCur < Literal(0) || muCur > right - left)) useBisection = true;
808       if (shift == right && (muCur < -(right - left) || muCur > Literal(0))) useBisection = true;
809       if (abs(fCur)>abs(fPrev)) useBisection = true;
810     }
811 
812     // fall back on bisection method if rational interpolation did not work
813     if (useBisection)
814     {
815 #ifdef  EIGEN_BDCSVD_DEBUG_VERBOSE
816       std::cout << "useBisection for k = " << k << ", actual_n = " << actual_n << "\n";
817 #endif
818       RealScalar leftShifted, rightShifted;
819       if (shift == left)
820       {
821         leftShifted = (std::numeric_limits<RealScalar>::min)();
822         // I don't understand why the case k==0 would be special there:
823         // if (k == 0) rightShifted = right - left; else
824         rightShifted = (k==actual_n-1) ? right : ((right - left) * RealScalar(0.6)); // theoretically we can take 0.5, but let's be safe
825       }
826       else
827       {
828         leftShifted = -(right - left) * RealScalar(0.6);
829         rightShifted = -(std::numeric_limits<RealScalar>::min)();
830       }
831 
832       RealScalar fLeft = secularEq(leftShifted, col0, diag, perm, diagShifted, shift);
833 
834 #if defined EIGEN_INTERNAL_DEBUGGING || defined EIGEN_BDCSVD_DEBUG_VERBOSE
835       RealScalar fRight = secularEq(rightShifted, col0, diag, perm, diagShifted, shift);
836 #endif
837 
838 #ifdef  EIGEN_BDCSVD_DEBUG_VERBOSE
839       if(!(fLeft * fRight<0))
840       {
841         std::cout << "fLeft: " << leftShifted << " - " << diagShifted.head(10).transpose()  << "\n ; " << bool(left==shift) << " " << (left-shift) << "\n";
842         std::cout << k << " : " <<  fLeft << " * " << fRight << " == " << fLeft * fRight << "  ;  " << left << " - " << right << " -> " <<  leftShifted << " " << rightShifted << "   shift=" << shift << "\n";
843       }
844 #endif
845       eigen_internal_assert(fLeft * fRight < Literal(0));
846 
847       while (rightShifted - leftShifted > Literal(2) * NumTraits<RealScalar>::epsilon() * numext::maxi<RealScalar>(abs(leftShifted), abs(rightShifted)))
848       {
849         RealScalar midShifted = (leftShifted + rightShifted) / Literal(2);
850         fMid = secularEq(midShifted, col0, diag, perm, diagShifted, shift);
851         if (fLeft * fMid < Literal(0))
852         {
853           rightShifted = midShifted;
854         }
855         else
856         {
857           leftShifted = midShifted;
858           fLeft = fMid;
859         }
860       }
861 
862       muCur = (leftShifted + rightShifted) / Literal(2);
863     }
864 
865     singVals[k] = shift + muCur;
866     shifts[k] = shift;
867     mus[k] = muCur;
868 
869     // perturb singular value slightly if it equals diagonal entry to avoid division by zero later
870     // (deflation is supposed to avoid this from happening)
871     // - this does no seem to be necessary anymore -
872 //     if (singVals[k] == left) singVals[k] *= 1 + NumTraits<RealScalar>::epsilon();
873 //     if (singVals[k] == right) singVals[k] *= 1 - NumTraits<RealScalar>::epsilon();
874   }
875 }
876 
877 
878 // zhat is perturbation of col0 for which singular vectors can be computed stably (see Section 3.1)
879 template <typename MatrixType>
880 void BDCSVD<MatrixType>::perturbCol0
881    (const ArrayRef& col0, const ArrayRef& diag, const IndicesRef &perm, const VectorType& singVals,
882     const ArrayRef& shifts, const ArrayRef& mus, ArrayRef zhat)
883 {
884   using std::sqrt;
885   Index n = col0.size();
886   Index m = perm.size();
887   if(m==0)
888   {
889     zhat.setZero();
890     return;
891   }
892   Index last = perm(m-1);
893   // The offset permits to skip deflated entries while computing zhat
894   for (Index k = 0; k < n; ++k)
895   {
896     if (col0(k) == Literal(0)) // deflated
897       zhat(k) = Literal(0);
898     else
899     {
900       // see equation (3.6)
901       RealScalar dk = diag(k);
902       RealScalar prod = (singVals(last) + dk) * (mus(last) + (shifts(last) - dk));
903 
904       for(Index l = 0; l<m; ++l)
905       {
906         Index i = perm(l);
907         if(i!=k)
908         {
909           Index j = i<k ? i : perm(l-1);
910           prod *= ((singVals(j)+dk) / ((diag(i)+dk))) * ((mus(j)+(shifts(j)-dk)) / ((diag(i)-dk)));
911 #ifdef EIGEN_BDCSVD_DEBUG_VERBOSE
912           if(i!=k && std::abs(((singVals(j)+dk)*(mus(j)+(shifts(j)-dk)))/((diag(i)+dk)*(diag(i)-dk)) - 1) > 0.9 )
913             std::cout << "     " << ((singVals(j)+dk)*(mus(j)+(shifts(j)-dk)))/((diag(i)+dk)*(diag(i)-dk)) << " == (" << (singVals(j)+dk) << " * " << (mus(j)+(shifts(j)-dk))
914                        << ") / (" << (diag(i)+dk) << " * " << (diag(i)-dk) << ")\n";
915 #endif
916         }
917       }
918 #ifdef EIGEN_BDCSVD_DEBUG_VERBOSE
919       std::cout << "zhat(" << k << ") =  sqrt( " << prod << ")  ;  " << (singVals(last) + dk) << " * " << mus(last) + shifts(last) << " - " << dk << "\n";
920 #endif
921       RealScalar tmp = sqrt(prod);
922       zhat(k) = col0(k) > Literal(0) ? tmp : -tmp;
923     }
924   }
925 }
926 
927 // compute singular vectors
928 template <typename MatrixType>
929 void BDCSVD<MatrixType>::computeSingVecs
930    (const ArrayRef& zhat, const ArrayRef& diag, const IndicesRef &perm, const VectorType& singVals,
931     const ArrayRef& shifts, const ArrayRef& mus, MatrixXr& U, MatrixXr& V)
932 {
933   Index n = zhat.size();
934   Index m = perm.size();
935 
936   for (Index k = 0; k < n; ++k)
937   {
938     if (zhat(k) == Literal(0))
939     {
940       U.col(k) = VectorType::Unit(n+1, k);
941       if (m_compV) V.col(k) = VectorType::Unit(n, k);
942     }
943     else
944     {
945       U.col(k).setZero();
946       for(Index l=0;l<m;++l)
947       {
948         Index i = perm(l);
949         U(i,k) = zhat(i)/(((diag(i) - shifts(k)) - mus(k)) )/( (diag(i) + singVals[k]));
950       }
951       U(n,k) = Literal(0);
952       U.col(k).normalize();
953 
954       if (m_compV)
955       {
956         V.col(k).setZero();
957         for(Index l=1;l<m;++l)
958         {
959           Index i = perm(l);
960           V(i,k) = diag(i) * zhat(i) / (((diag(i) - shifts(k)) - mus(k)) )/( (diag(i) + singVals[k]));
961         }
962         V(0,k) = Literal(-1);
963         V.col(k).normalize();
964       }
965     }
966   }
967   U.col(n) = VectorType::Unit(n+1, n);
968 }
969 
970 
971 // page 12_13
972 // i >= 1, di almost null and zi non null.
973 // We use a rotation to zero out zi applied to the left of M
974 template <typename MatrixType>
975 void BDCSVD<MatrixType>::deflation43(Index firstCol, Index shift, Index i, Index size)
976 {
977   using std::abs;
978   using std::sqrt;
979   using std::pow;
980   Index start = firstCol + shift;
981   RealScalar c = m_computed(start, start);
982   RealScalar s = m_computed(start+i, start);
983   RealScalar r = sqrt(numext::abs2(c) + numext::abs2(s));
984   if (r == Literal(0))
985   {
986     m_computed(start+i, start+i) = Literal(0);
987     return;
988   }
989   m_computed(start,start) = r;
990   m_computed(start+i, start) = Literal(0);
991   m_computed(start+i, start+i) = Literal(0);
992 
993   JacobiRotation<RealScalar> J(c/r,-s/r);
994   if (m_compU)  m_naiveU.middleRows(firstCol, size+1).applyOnTheRight(firstCol, firstCol+i, J);
995   else          m_naiveU.applyOnTheRight(firstCol, firstCol+i, J);
996 }// end deflation 43
997 
998 
999 // page 13
1000 // i,j >= 1, i!=j and |di - dj| < epsilon * norm2(M)
1001 // We apply two rotations to have zj = 0;
1002 // TODO deflation44 is still broken and not properly tested
1003 template <typename MatrixType>
1004 void BDCSVD<MatrixType>::deflation44(Index firstColu , Index firstColm, Index firstRowW, Index firstColW, Index i, Index j, Index size)
1005 {
1006   using std::abs;
1007   using std::sqrt;
1008   using std::conj;
1009   using std::pow;
1010   RealScalar c = m_computed(firstColm+i, firstColm);
1011   RealScalar s = m_computed(firstColm+j, firstColm);
1012   RealScalar r = sqrt(numext::abs2(c) + numext::abs2(s));
1013 #ifdef  EIGEN_BDCSVD_DEBUG_VERBOSE
1014   std::cout << "deflation 4.4: " << i << "," << j << " -> " << c << " " << s << " " << r << " ; "
1015     << m_computed(firstColm + i-1, firstColm)  << " "
1016     << m_computed(firstColm + i, firstColm)  << " "
1017     << m_computed(firstColm + i+1, firstColm) << " "
1018     << m_computed(firstColm + i+2, firstColm) << "\n";
1019   std::cout << m_computed(firstColm + i-1, firstColm + i-1)  << " "
1020     << m_computed(firstColm + i, firstColm+i)  << " "
1021     << m_computed(firstColm + i+1, firstColm+i+1) << " "
1022     << m_computed(firstColm + i+2, firstColm+i+2) << "\n";
1023 #endif
1024   if (r==Literal(0))
1025   {
1026     m_computed(firstColm + i, firstColm + i) = m_computed(firstColm + j, firstColm + j);
1027     return;
1028   }
1029   c/=r;
1030   s/=r;
1031   m_computed(firstColm + i, firstColm) = r;
1032   m_computed(firstColm + j, firstColm + j) = m_computed(firstColm + i, firstColm + i);
1033   m_computed(firstColm + j, firstColm) = Literal(0);
1034 
1035   JacobiRotation<RealScalar> J(c,-s);
1036   if (m_compU)  m_naiveU.middleRows(firstColu, size+1).applyOnTheRight(firstColu + i, firstColu + j, J);
1037   else          m_naiveU.applyOnTheRight(firstColu+i, firstColu+j, J);
1038   if (m_compV)  m_naiveV.middleRows(firstRowW, size).applyOnTheRight(firstColW + i, firstColW + j, J);
1039 }// end deflation 44
1040 
1041 
1042 // acts on block from (firstCol+shift, firstCol+shift) to (lastCol+shift, lastCol+shift) [inclusive]
1043 template <typename MatrixType>
1044 void BDCSVD<MatrixType>::deflation(Index firstCol, Index lastCol, Index k, Index firstRowW, Index firstColW, Index shift)
1045 {
1046   using std::sqrt;
1047   using std::abs;
1048   const Index length = lastCol + 1 - firstCol;
1049 
1050   Block<MatrixXr,Dynamic,1> col0(m_computed, firstCol+shift, firstCol+shift, length, 1);
1051   Diagonal<MatrixXr> fulldiag(m_computed);
1052   VectorBlock<Diagonal<MatrixXr>,Dynamic> diag(fulldiag, firstCol+shift, length);
1053 
1054   const RealScalar considerZero = (std::numeric_limits<RealScalar>::min)();
1055   RealScalar maxDiag = diag.tail((std::max)(Index(1),length-1)).cwiseAbs().maxCoeff();
1056   RealScalar epsilon_strict = numext::maxi<RealScalar>(considerZero,NumTraits<RealScalar>::epsilon() * maxDiag);
1057   RealScalar epsilon_coarse = Literal(8) * NumTraits<RealScalar>::epsilon() * numext::maxi<RealScalar>(col0.cwiseAbs().maxCoeff(), maxDiag);
1058 
1059 #ifdef EIGEN_BDCSVD_SANITY_CHECKS
1060   assert(m_naiveU.allFinite());
1061   assert(m_naiveV.allFinite());
1062   assert(m_computed.allFinite());
1063 #endif
1064 
1065 #ifdef  EIGEN_BDCSVD_DEBUG_VERBOSE
1066   std::cout << "\ndeflate:" << diag.head(k+1).transpose() << "  |  " << diag.segment(k+1,length-k-1).transpose() << "\n";
1067 #endif
1068 
1069   //condition 4.1
1070   if (diag(0) < epsilon_coarse)
1071   {
1072 #ifdef  EIGEN_BDCSVD_DEBUG_VERBOSE
1073     std::cout << "deflation 4.1, because " << diag(0) << " < " << epsilon_coarse << "\n";
1074 #endif
1075     diag(0) = epsilon_coarse;
1076   }
1077 
1078   //condition 4.2
1079   for (Index i=1;i<length;++i)
1080     if (abs(col0(i)) < epsilon_strict)
1081     {
1082 #ifdef  EIGEN_BDCSVD_DEBUG_VERBOSE
1083       std::cout << "deflation 4.2, set z(" << i << ") to zero because " << abs(col0(i)) << " < " << epsilon_strict << "  (diag(" << i << ")=" << diag(i) << ")\n";
1084 #endif
1085       col0(i) = Literal(0);
1086     }
1087 
1088   //condition 4.3
1089   for (Index i=1;i<length; i++)
1090     if (diag(i) < epsilon_coarse)
1091     {
1092 #ifdef  EIGEN_BDCSVD_DEBUG_VERBOSE
1093       std::cout << "deflation 4.3, cancel z(" << i << ")=" << col0(i) << " because diag(" << i << ")=" << diag(i) << " < " << epsilon_coarse << "\n";
1094 #endif
1095       deflation43(firstCol, shift, i, length);
1096     }
1097 
1098 #ifdef EIGEN_BDCSVD_SANITY_CHECKS
1099   assert(m_naiveU.allFinite());
1100   assert(m_naiveV.allFinite());
1101   assert(m_computed.allFinite());
1102 #endif
1103 #ifdef EIGEN_BDCSVD_DEBUG_VERBOSE
1104   std::cout << "to be sorted: " << diag.transpose() << "\n\n";
1105 #endif
1106   {
1107     // Check for total deflation
1108     // If we have a total deflation, then we have to consider col0(0)==diag(0) as a singular value during sorting
1109     bool total_deflation = (col0.tail(length-1).array()<considerZero).all();
1110 
1111     // Sort the diagonal entries, since diag(1:k-1) and diag(k:length) are already sorted, let's do a sorted merge.
1112     // First, compute the respective permutation.
1113     Index *permutation = m_workspaceI.data();
1114     {
1115       permutation[0] = 0;
1116       Index p = 1;
1117 
1118       // Move deflated diagonal entries at the end.
1119       for(Index i=1; i<length; ++i)
1120         if(abs(diag(i))<considerZero)
1121           permutation[p++] = i;
1122 
1123       Index i=1, j=k+1;
1124       for( ; p < length; ++p)
1125       {
1126              if (i > k)             permutation[p] = j++;
1127         else if (j >= length)       permutation[p] = i++;
1128         else if (diag(i) < diag(j)) permutation[p] = j++;
1129         else                        permutation[p] = i++;
1130       }
1131     }
1132 
1133     // If we have a total deflation, then we have to insert diag(0) at the right place
1134     if(total_deflation)
1135     {
1136       for(Index i=1; i<length; ++i)
1137       {
1138         Index pi = permutation[i];
1139         if(abs(diag(pi))<considerZero || diag(0)<diag(pi))
1140           permutation[i-1] = permutation[i];
1141         else
1142         {
1143           permutation[i-1] = 0;
1144           break;
1145         }
1146       }
1147     }
1148 
1149     // Current index of each col, and current column of each index
1150     Index *realInd = m_workspaceI.data()+length;
1151     Index *realCol = m_workspaceI.data()+2*length;
1152 
1153     for(int pos = 0; pos< length; pos++)
1154     {
1155       realCol[pos] = pos;
1156       realInd[pos] = pos;
1157     }
1158 
1159     for(Index i = total_deflation?0:1; i < length; i++)
1160     {
1161       const Index pi = permutation[length - (total_deflation ? i+1 : i)];
1162       const Index J = realCol[pi];
1163 
1164       using std::swap;
1165       // swap diagonal and first column entries:
1166       swap(diag(i), diag(J));
1167       if(i!=0 && J!=0) swap(col0(i), col0(J));
1168 
1169       // change columns
1170       if (m_compU) m_naiveU.col(firstCol+i).segment(firstCol, length + 1).swap(m_naiveU.col(firstCol+J).segment(firstCol, length + 1));
1171       else         m_naiveU.col(firstCol+i).segment(0, 2)                .swap(m_naiveU.col(firstCol+J).segment(0, 2));
1172       if (m_compV) m_naiveV.col(firstColW + i).segment(firstRowW, length).swap(m_naiveV.col(firstColW + J).segment(firstRowW, length));
1173 
1174       //update real pos
1175       const Index realI = realInd[i];
1176       realCol[realI] = J;
1177       realCol[pi] = i;
1178       realInd[J] = realI;
1179       realInd[i] = pi;
1180     }
1181   }
1182 #ifdef EIGEN_BDCSVD_DEBUG_VERBOSE
1183   std::cout << "sorted: " << diag.transpose().format(bdcsvdfmt) << "\n";
1184   std::cout << "      : " << col0.transpose() << "\n\n";
1185 #endif
1186 
1187   //condition 4.4
1188   {
1189     Index i = length-1;
1190     while(i>0 && (abs(diag(i))<considerZero || abs(col0(i))<considerZero)) --i;
1191     for(; i>1;--i)
1192        if( (diag(i) - diag(i-1)) < NumTraits<RealScalar>::epsilon()*maxDiag )
1193       {
1194 #ifdef EIGEN_BDCSVD_DEBUG_VERBOSE
1195         std::cout << "deflation 4.4 with i = " << i << " because " << (diag(i) - diag(i-1)) << " < " << NumTraits<RealScalar>::epsilon()*diag(i) << "\n";
1196 #endif
1197         eigen_internal_assert(abs(diag(i) - diag(i-1))<epsilon_coarse && " diagonal entries are not properly sorted");
1198         deflation44(firstCol, firstCol + shift, firstRowW, firstColW, i-1, i, length);
1199       }
1200   }
1201 
1202 #ifdef EIGEN_BDCSVD_SANITY_CHECKS
1203   for(Index j=2;j<length;++j)
1204     assert(diag(j-1)<=diag(j) || abs(diag(j))<considerZero);
1205 #endif
1206 
1207 #ifdef EIGEN_BDCSVD_SANITY_CHECKS
1208   assert(m_naiveU.allFinite());
1209   assert(m_naiveV.allFinite());
1210   assert(m_computed.allFinite());
1211 #endif
1212 }//end deflation
1213 
1214 #ifndef __CUDACC__
1215 /** \svd_module
1216   *
1217   * \return the singular value decomposition of \c *this computed by Divide & Conquer algorithm
1218   *
1219   * \sa class BDCSVD
1220   */
1221 template<typename Derived>
1222 BDCSVD<typename MatrixBase<Derived>::PlainObject>
1223 MatrixBase<Derived>::bdcSvd(unsigned int computationOptions) const
1224 {
1225   return BDCSVD<PlainObject>(*this, computationOptions);
1226 }
1227 #endif
1228 
1229 } // end namespace Eigen
1230 
1231 #endif
1232