Re: [eigen] How to subtract a diagonal matrix from a matrix in one expression?

[Adrien Escande <adrien.escande@xxxxxxxxx>]

Dear Alexander,

if you are using fixed size matrices, I don't see any problem with having temporary objects as they don't induce any memory allocation.

The following one-liner ther work:
const Matrix2d M2 = M - Matrix2d(v.asDiagonal());

and I guess the compiler can be quite smart to avoid any overhead.

A a side remarks, be very mindful of the auto keyword with Eigen expressions.

If you do
const auto M2 = M - Matrix2d(v.asDiagonal());

M2 will be a CwiseBinaryOp, not a Matrix2d, meaning that you subtraction will be recomputed each time you are using M2.

Best regards,

Adrien

On Thu, Jan 12, 2017 at 12:44 AM, Alexander Voigt <alexander.voigt@xxxxxxxxxxxxxxxxxxxxxx> wrote:

Dear Eigen developers,

is there a way to subtract a diagonal matrix (which has been constructed
from a vector v using v.asDiagonal()) from a square matrix in one
_expression_?  (I'm asking for one _expression_, because I'd like to make
the result const and avoid a temporary.)

Example:

=====================================================
include <Eigen/Core>
using namespace Eigen;

int main() {
   Matrix<double,2,2> M;
   Matrix<double,2,1> v;

   const auto M2 = M - v.asDiagonal(); // does not compile
}
=====================================================

The marked line yields a compiler error with g++ 4.9.2.
I know that it is possible to write instead:

=====================================================
   Matrix<double,2,2> M2(M);
   M2 -= v.asDiagonal();     // works
=====================================================

However, this latter approach has the disadvantage that M2 cannot be
made const, which I'd like to achieve.  An approach that works with one
_expression_ would be

=====================================================
   const Matrix<double,2,2> M2 = [&M,&v]{
      Matrix<double,2,2> tmp(M);
      tmp -= v.asDiagonal();
      return tmp;
   }();
=====================================================

but this has the disadvantage that it is more complicated and needs a
temporary object.

Many thanks and best regards,
Alex

--
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Dr. Alexander Voigt
Institute for Theoretical Particle Physics and Cosmology
RWTH Aachen
Sommerfeldstr. 14
52074 Aachen
Room: 28A408
Phone: +49 (0)241 80 27049
Fax  : +49 (0)241 80 22187
E-Mail: alexander.voigt@xxxxxxxxxxx-aachen.de
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~