Construction of qmutilde for muhat < 0 about pyhf HOT 6 OPEN

Quick comment, the situation $\hat{\mu} &lt; 0$ is common in vanilla signal+background signal strength situations like $Pois(n | \mu s + b)$, e.g. when $n &lt; b$. The presence of a local minima / non-convex log-likelihood curve however is uncommon there (unless something really weird with signal systematics). The EFT situations are ones where you might expect some interference or something that could lead to local minima, but in those situations $\mu$ is some parameter where it's not clear to me that you would want to do an upper-limit. In that case it's not so much about $\tilde{q}\mu$ vs. ${q}\mu$ as it is about $q$ vs. $t$ (e.g. single-sided vs. double sided).

from pyhf.

I believe so, @will-cern can presumably confirm.

from pyhf.

So following up on some discussion points raised by @cranmer from our NYU/UW-Madison group meeting discussion:

• Incorrect is probably too strong a word here given that if your likelihood function has this shape then you are probably not in the regieme that using the asymptotics is a good idea and you should be using pseudoexperiments, where you are free to choose whatever test statistic you'd like.

A question that I have is if you believe that you would be in a situation where (edit) $\mu &lt; 0$, like an EFT analysis, then why would you be selecting $\tilde{q}_{\mu}$ — which is "For the case where one considers models for which $\mu \geq 0$" — and not $q\mu$?

from pyhf.

I agree that "incorrect" is not a good description, maybe "inconsistent with the definition" is better. The question about the usefulness of that approach in such a scenario stands of course.

Regarding your question: $\hat{\mu}&lt;0$ could always happen due to "unlucky" observed data even when you expect $\mu\geq0$ for physics reasons. There is no way to protect against this beyond designing your model with a bound at 0 and you would not know whether or not this will happen prior to unblinding. The method and test statistic is already fixed at that point.

from pyhf.

Sorry, you're both obviously correct, but I was typing #2352 (comment) while trying to listen to another meeting and I typoed. I meant to write

A question that I have is if you believe that you would be in a situation where $\mu &lt; 0$

(so not $\hat{\mu} &lt; 0$)

but if I had read better I would have seen that @alexander-held addressed this in the first sentence:

For models with some physically motivated bounds (like $\mu\geq0$)

so we're already starting there with the answer to my question.

from pyhf.

@alexander-held For reference, as you mention in your 2023-10-16 ATLAS Statistics Committee Meeting slides this was addressed in

xRooFit: fixed recently and deployed in StatAnalysis

I'm assuming this is https://gitlab.cern.ch/will/xroofit/-/commit/ebf6a49194c3c08dd2e12d8933fe16067b6243e8 (xRooFit doesn't use MR based workflow so going by commit message)?

from pyhf.