In the linear case, if you have the SVD of your matrix available,
computing the L-curve is simple and straightforward.
In the nonlinear case, there is no way (that I know of) to easily
compute the L-curve, other than re-solving the nonlinear problem many
times for different lambdas. So in this case its up to the user to
determine how to select lambda. You can of course generate an L-curve by
plotting the residual norm against the solution norm for different lambdas.
Patrick
On 03/07/2016 10:30 AM, viktor drobot wrote:
Hi all! I wonder why linear squares fit module has routines for determining
the optimal value of lambda which is used in Tikhonov regularization (i. e.
by L-corner approach) while nonlinear module has no such routines? How do
you find the optimal lambda value? Thank you!
