making scans faster

[philippeller]

Just putting this idea here not to forget:

For scans (e.g. 1-d nutau norm, 2-d contours, ...) we always start the minimizer from nominal, but it would converge much faster if started from a neighboring best-fit point.
Example: when we scan nutau norm from 0 to 2 in steps of 0.1, then we can do a first fit at 0 starting from nominal settings, and then use this bestfit value to seed the next fit at 0.1, and so forth.....could save quite some time

[thehrh]

Nice. I see two limitations of this idea, however:

• Scan has to be run sequentially
• One has to trust the first best-fit point to reflect the global optimum, otherwise one could systematically end up in neighbouring local minima and possibly not even notice since the scan could look smooth (while random/nominal seeds would likely lead to the recovery of the global optimum for some scan points, but end up in local minima for others)

[philippeller]

I agree with these points. The first one is at least the case when we do nutau trials, we throw one pseudodata and then scan the whole range from left to right in one job.
The second point is of course the one that would need validation, and might also work for one kind of analysis and not the other...so for your NSI maybe not (?)

[LeanderFischer]

Closing, since we've been doing ok with the current setup..