Handling of rAbsAccuracy_ and rRelAccuracy_ in HybridNew

[Ma128-bit]

Referring to src/HybridNew.cc.

Note: It is unclear to me whether the current behavior is intentional or a bug.

In several places, the limit uncertainty (limitErr) is compared against max(rAbsAccuracy_, rRelAccuracy_ * limit). However, it seems more natural to compare it against min(rAbsAccuracy_, rRelAccuracy_ * limit)

Example:
Default values:

• rAbsAccuracy_ = 0.1
• rRelAccuracy_ = 0.05

I'm obtaining:

Limit: r < 0.113864 +/- 0.050296 @ 95% CL

In this case:

• rRelAccuracy_ * limit = 0.0057
• rAbsAccuracy_ = 0.1
• limitErr = 0.050296

Since 0.050296 < max(0.1, 0.0057) = 0.1 the convergence criterion is satisfied and HybridNew stops, even though the relative uncertainty is ~44%, and no further fit of the grid toys is performed. This would not have happened if the limitErr had been compared to min(rAbsAccuracy_, rRelAccuracy_ * limit) = 0.0057.

Note:
Obviously, it could be easily fixed by passing a very small value of --rAbsAcc . In the previous example, the limit converges to:

Limit: r < 0.114009 +/- 0.000891322 @ 95% CL

However, it would be more natural if the minimum were used directly.

Conclusions
My suggestion would be to replace max with min when comparing limitErr to the accuracy thresholds, so that the stricter of the absolute and relative criteria is enforced.
If the current implementation is intentional, clarification on the rationale would be appreciated.

[gpetruc]

I don't recall, it's possible that the behaviour was intended back then, when r == 1 was a very important value (SM Higgs), and so it made sense for the code to relax the relative accuracy requirement once the value of r was well below 1, and one could switch off individually the absolute or relative requirement by setting it to zero.

Still, I agree that by now with combine being used for many more models and r = 1 not being particularly special it would probably make sense to change the max a min or make the default value of rAbsAcc zero (if you keep the max) or infinity (if you change it to min).

If you change it, throw an exception at startup if any of the accuracies is configured to zero, or treat zeros as infinity.