Subspace Regularization
Subspace regularization methods can improve the stability of ill-posed inverse problems and significantly reduce the modelling efforts required for an optimization.
Singular Value Decomposition (SVD)
It is recommended to use SVD for any PEST setup (unless LSQR has been chosen, which tends to be faster for PEST runs involving more than 2500 parameters). FePEST activates this option by default. The following parameter settings determine the number of singular values/ super parameters considered during the optimization:
The parameter defines the variable MAXSING in the PEST Control file (*.pst).
This limits the number of singular values/super parameters to a maximum number.
The parameter defines the variable EIGTHRESH in the PEST Control file (*.pst).
This is the Eigenvalue threshold, which all super parameters lower than this value will be truncated. The default threshold value is set to 1e-6. At this value, the optimization is still regarded well-posed, which ensures stability of the optimization.
|
|
See the PEST users manual (5th Edition), section 8.4.2: Implementation of SVD with PEST for a full discussion of these settings. |
