Eigen: Is it possible to create LeastSquareDiagonalPreconditioner-like conditioner if i only can compute Aty and Ax?

I want to solve least-squares like system A^t * A * x = -A^t * x. (I'm implementing Gauss-Newton method for special problem).

I wrote special routines which allow me to compute A * x and A^t * y products. With such routines it's easy to use matrix-free solvers thanks to Eigen.

But my approach converges not as good as Eigen::LeastSquaresConjugateGradient. I made a small test and it's looks like LeastSquareDiagonalPreconditioner speed ups convergence a lot.

My question is - how i can use LeastSquareDiagonalPreconditioner or implement own Preconditioner if i can only compute matrix products? I'm not very good with understanding preconditioning/conjugate gradient stuff btw.

EDIT

For clarity - i want to use Matrix-free solvers from Eigen with my product routines.

EDIT 2

Matrix-vector products was obtained by using forward and reverse mode autodiff on some objective functions.

1 answer

  • answered 2018-11-08 14:51 ggael

    The easiest might be to implement your own preconditioner class inheriting DiagonalPreconditioner and implementing something like LeastSquareDiagonalPreconditioner ::factorize() but adapted to your type. Basically you need to compute:

     m_invdiag(j) = 1./mat.col(j).squaredNorm();
    

    for all column j using a strategy to what you already implemented for the product operators.