Relative versus absolute values of TolX and TolFun

In using solvers such as fminsearch and lsqnonlin the on-line help documentation state that these tolerances are generally absolute tolerances but can be used as relative tolerances. What is not clear to me is: 1. when they are used in each of these ways; 2. whether it is possible to specify how you would like them to be used (i.e., in an absolute versus relative sense).

Thank, David

[Matt J]

I agree. The Optimization Options Reference could use a lot more detail on how the stopping criteria are defined. I often see questions about whether TolX is an L2-bound or an max-norm bound, for example.

[Amro]

related: How do the tolerances work in the Optimization Toolbox?

Answer 1

[Matt J]

1. Dunno. See also my comment above.

2. I don't think you can manually alter the meaning of the built-in stopping criteria. However, you can impose your own customized stopping criteria using the OutputFcn option. You can also quasi-deactivate built-in criteria e.g., by setting TolX=0. Unfortunately, it doesn't let you set TolX<0, which would completely deactivate it.