[Halld-tagger] [EXTERNAL] Fwd: Accidental subtraction

[Shepherd, Matthew]

Hi all,

From Peter:

> My initial thought was that this is wrong and not recommended. Did the guidance by the beam line group change regarding this issue? Given that the current a2 cross-section analysis is using this method and there is a big push to publish it asap I am concerned that this might not be resolved properly and might set a bad precedent going forward.


> On Feb 20, 2023, at 5:33 AM, Richard Jones via Halld-tagger <halld-tagger at jlab.org> wrote:
> 
> In PWA, i understand that this avoids the pain of negative weights and so improves the statistical error from the fits (or at least it gives that feeling). In fact, it introduces a set of new systematic errors of its own that will probably drive us back to the more rigorous approach before we are done. For the moment I am not speaking up about this because we just need to get our first results out. But eventually this needs to be given a critical review.

I'm not sure if Malte and Lawrence or Ryan are on this mailing list.  There has been quite a bit of discussion on this topic.  In the case of the eta pi0 analysis, at the early stages it shown that both methods produced the same result.  I believe Ryan has also checked this in the omega pi pi case (to obtain the pi1 limit) as well as a number of other cases. 

The issues Richard notes are certainly real.  I think the question is:  do they pose practical problems to the analysis?  Also, it is important to remember that efficiency in analysis and many other things is anchored to signal MC.  So, to make a mistake, one has to come up with problems that are not modeled in signal MC.  I believe our general experience is that once there are reasonable cuts made there are typically few events (few percent) with multiple combinations in the signal RF bin.  

It is not so much about improving statistical error in the fit, it is about doing the analysis in a simpler way.  It is nice to look at plots where all the entries in a histogram have a weight of 1.  We had some early cases where I think people were making plots (perhaps with no chi^2 cut) and the subtractions were huge and they become invisible in the subsequent analysis.  It is a pain to maintain this other dimension in backgrounds throughout the analysis.

I could be wrong, but my feeling is that the level of potential systematic error introduced by using the best chi^2 is at least an order of magnitude (if not more) below the other challenging systematic issues in amplitude analysis.  I'm not sure we will ever get to the stage where we care about it.  So for these analyses we should focus our brain cycles on the harder problems and do anything we can to make addressing those harder problems easier and more efficient.

It would be good to keep all of these things in mind and keep a lookout for potential failure points of this simpler approach in the context of ongoing analyses, but I don't think I've seen one yet. 

Matt

-------------- next part --------------
A non-text attachment was scrubbed...
Name: smime.p7s
Type: application/pkcs7-signature
Size: 4075 bytes
Desc: not available
URL: <https://mailman.jlab.org/pipermail/halld-tagger/attachments/20230220/34467b88/attachment.p7s>