[root] Log of /trunk/math/matrix/inc/TMatrixDEigen.h

Revision 15134 - (view) (download) (as text) (annotate) - [select for diffs]
Modified Tue May 23 04:47:42 2006 UTC (8 years, 8 months ago) by brun
Original Path: trunk/matrix/inc/TMatrixDEigen.h
File length: 2803 byte(s)
Diff to previous 15068

From Federico Carminati:

"I have implemented all copy and equal operators needed to silence all
warnings in AliRoot, as requested. I have implemented shallow copies as
would do the default operators synthetized by the compiler.
Most operators are protected. If users complain, you just have to move
them into the public area, but class derivation is of course supported.
It has been a terrible job, I have modified 278 files, but the changes
are backward compabile, and this goes a long way to permitting user to
use the effc++ flag with root headers."

Revision 10329 - (view) (download) (as text) (annotate) - [select for diffs]
Modified Sat Oct 16 18:09:17 2004 UTC (10 years, 3 months ago) by brun
Original Path: trunk/matrix/inc/TMatrixDEigen.h
File length: 2347 byte(s)
Diff to previous 9893

From Eddy Offermann:
- added new decomposition class TDecompBK: Bunch-Kaufman algorithm
  this class is designed to decompose real symmetric matrices . Now
  we can guarantee that inverse of a symmetric matrix is again
  symmetric .
- As a consequence of the introduction of TDecompBK, we added
  Invert(..) functionality to TMatrixD/FSym
  Also added InvertFast to TMatrixD/FSym which uses Cramer
  inversion for matrices <= (6 x 6) , resulting in large speedup
  but less accuracy . The necessary routines are encapsulated
  in the new TMatrixDSymCramerInv class .

  Note of caution :
  Unlike the other decompositions, Bunch-Kaufmann does not result
  in a triangular matrix of which the determinant is trivially calculated
  by multiplying the diagonal elements . As a consequence,
  we have to invoke yet another decomposition when the user
  requests the determinant .
  So Invert() will be roughly twice faster than Double_t *det; Invert(det)

- Changed the algoritm for the Cholesky decomposition so that we
  do not need to have a copy of the original matrix around !
  After removal of class member fA, increased the version to 2
- Added sorting of eigen -values/-vectors to TMatrixDEigen .
  They are now sorted according to Re^2+Im^2 of the (possible)
  complex eigenvalues in descending order . TMatrixDSymEigen
  already sorted . Since most users use TMatrixD instead of
  TMatrixDSym, this will ensure backward compatiblity since
  old root contained the symmetric algorithms in TMatrixD .
- Added rank 1 update algorihms : A += alpha * x y^T
  (x and y are vectors)
  to the non-sparse matrix classes (adding this to sparse
  would make this matrix immediately dense :)
- Moved Invert fom TDecompBase to the different decomposition
  classes . Reason is that some classes can decompose A(m,n)
  where m > n . The result of Invert() is then a "pseudo-invert"
  : A_pseudo = (A^T A)^-1 A^T .
- Removed an unfortunate bug in the TMatrixD/FSub class :
  both operators
      void operator+=(Double_t val);
      void operator*=(Double_t val);
  performed actually
      void operator=(Double_t val);
~

Revision 9893 - (view) (download) (as text) (annotate) - [select for diffs]
Modified Fri Sep 3 13:41:35 2004 UTC (10 years, 4 months ago) by brun
Original Path: trunk/matrix/inc/TMatrixDEigen.h
File length: 2274 byte(s)
Diff to previous 8892

From Eddy Offermann
The matrix includes contained circular dependencies (which were of
course
stopped by the #ifdef's) . Situation has been cleaned up by inserting
several forward declarations in the header files and moving some
includes
to the source files . A change in a header file now causes less
compilation
in other directories like hist.
This cleanup made adding an include file in quadp/src/TQpVar necessary
and some additional include files in the tutorials and test files

Revision 8892 - (view) (download) (as text) (annotate) - [select for diffs]
Modified Wed May 12 10:39:29 2004 UTC (10 years, 8 months ago) by brun
Original Path: trunk/matrix/inc/TMatrixDEigen.h
File length: 2222 byte(s)
Diff to previous 8558

From Eddy Offermann:
1) patch for the matrix directory (cvs diff -a matrix)
2) tar file containing new modules:
  matrix/inc/TDecompSparse.h
  matrix/inc/TMatrixDSparse.h
  matrix/src/TDecompSparse.cxx
  matrix/src/TMatrixDSparse.cxx
3) new stressLinear.cxx

Description of changes:
1) many updates of comments in code
2) possibility to construct random matrices, general
   symmetric and positive definite
3) Introduction of new matrix class: TMatrixDSparse .
   It defines a general sparse matrix in
Harwell-Boeing
   format (sparse structure definition optimized for
   matrix-vector multiplications).
   Like all the other matrix classes, it derives from
   TMatrixDBase
4) Introduction of a new decomposition class:
   TDecompSparse . It allows decomposition (and
   therefore equation solving) of sparse symmetric
   matrices

Revision 8442 - (view) (download) (as text) (annotate) - [select for diffs]
Modified Fri Mar 19 14:20:40 2004 UTC (10 years, 10 months ago) by brun
Original Path: trunk/matrix/inc/TMatrixDEigen.h
File length: 2231 byte(s)
Diff to previous 8143

From Eddy Offermann:

1. Removed streaming of matrix view classes (TMatrixDRow,TMatrixDColumn,TMatrixDDiag,TMatrixDFlat)
   because a. we  want these classes to be as light as possible
           b. old implementation was wrong, some member elements were "(!)" out
2. Added a series of functions to support upcoming quadratic programming functionality:

    TVectorD/F:

    - AddElemMult(TVectorF &t,Float_t alpha,const TVectorF &s1,const TVectorF &s2);
      t += alpha * s1*s2 , where "s1*s2 is an element-wise multiplication
    - AddElemMult(TVectorF &t,Float_t alpha,const TVectorF &s1,const TVectorF &s2,const TVectorF &map);
      t += alpha * s1*s2 , where "s1*s2 is an element-wise multiplication. It is only performed
           for those elements i where map[i] != 0

    - AddElemDiv (TVectorF &t,Float_t alpha,const TVectorF &s1,const TVectorF &s2);
      t += alpha * s1\s2 , where "s1\s2 is an element-wise division
    - AddElemDiv (TVectorF &t,Float_t alpha,const TVectorF &s1,const TVectorF &s2,const TVectorF &map);
      t += alpha * s1\s2 , where "s1\s2 is an element-wise division. It is only performed
           for those elements i where map[i] != 0

    - ElementMult(TVectorF &t,const TVectorF &s,const TVectorF &map);
      t += t*s , where "t*s" is an element-wise multiplication. It is only performed
           for those elements i where map[i] != 0
    - ElementDiv (TVectorF &t,const TVectorF &s,const TVectorF &map);
      t += t/s , where "t/s" is an element-wise division. It is only performed
           for those elements i where map[i] != 0

    - TVectorD &SelectNonZeros       (const TVectorD &select);
      put elements for which select[i] == 0 to zero
    - Bool_t    MatchesNonZeroPattern(const TVectorD &select);
      verify that elements are != 0 for which select[i] != 0
    - Bool_t    SomePositive         (const TVectorD &select);
      verify that all elements > 0 for which select[i] != 0
    - void      AddSomeConstant      (Double_t val,const TVectorD &select);
      Add val to those elements for which select[i] != 0

    - TVectorF &TVectorF::Invert()
      v[i] = 1/v[i]

    TVectorD/F and TMatrixD/FBase:

    - virtual Double_t Sum () const;
    - virtual Double_t Min () const;
    - virtual Double_t Max () const;

3. Added another data member fA to TDecompChol . Unlike the other decomposition schemes, it is not
   possible to store the original input matrix in one of the decomposition members

4. Added method "SetMatrix(const TMatrix.. &)" to all decomposition classes . It allows to
   change the matrix that should be decomposed. Of course it resets the class status to kInit
   and triggers a new factorization.

5. Added private methods to TMatrixD and TMatrixDSym

  const TMatrixD EigenVectors(TVectorD &eigenValues) const;  // This function is now obsolete (and not implemented), you
                                                             // should use TMatrixDSymEigen or TMatrixDEigen .

6. Changed the name "Adopt" to "Use" to be in agreement with the convention that the memory management
   of adopted members is now the reponsibility of the class

Revision 8013 - (view) (download) (as text) (annotate) - [select for diffs]
Added Sun Jan 25 20:33:32 2004 UTC (11 years ago) by brun
Original Path: trunk/matrix/inc/TMatrixDEigen.h
File length: 2184 byte(s)

New Linear Algebra package from Eddy Offermann.
This new package reimplements the previous classes TMatrix and TMatrixD.
The new classes should be back compatible with the previous version except
the function GetElements.
New classes have been introduced for symmetric matrices,
lazy matrices.
New algorithms (LU, SVD) have been introduced.
A new test suite test/stressLinear.cxx is introduced.
A complete description of this package will be posted in the coming days.
The classes are well documented in the implementation headers.

This form allows you to request diffs between any two revisions of this file. For each of the two "sides" of the diff, enter a numeric revision.

Links to HEAD:	(view) (download) (as text) (annotate)
Sticky Revision: