[G8b_run] Tagger Sag
Michael Dugger
dugger at jlab.org
Fri Aug 21 22:11:39 EDT 2009
Volker,
Thank you for your prompt reply. I apologize for not being up to date
regarding the announcement of Chucks momentum corrections.
I would like to state that my only concern is that we can get an energy
correction that allows Ken to move forward with the new determination of
the polarizations. To that effect, I hope that Chuck's energy corrections
are correct, in that the correction is so small as to be nearly
negligible. We just have to be extra sure since none of us (Stuart,
Chuck, or I) have been able to obtain the same results. As I stated
earlier, I believe that the kinematic fitter will give us the best
results. A kinematic fitter is best suited for this kind of measurement.
Yet, we just want to make sure that everything is golden. Otherwise, we
end up coming back to the data some time later. No fun.
With regards to your email:
I was not asking about a a cut on the tail of a pull distribution.
Chuck wrote:
"Therefore I had to "zoom in" on the y-axis in order carry this process
out and the result of this "zooming" is the
EtrueMinusEmeas_v_Emeas_ForSlicing.gif plot. The projection of this plot
onto the y-axis is indeed a Gaussian as the confidence level cut only cuts
off the tails of the distribution and does not affect the peak position."
Chuck writes about the Gaussian of the [Etrue - Emeasured] distribution.
If this were divided by the uncertainty, you would be correct, in that
that would then be the pull distribution that he would be describing.
However, this is not the case.
I appreciate that getting the pulls centered, with variance =1
is not an easy task. That being said, it is paramount that any results
coming from a kinematic fitter come close to having a variance equal to
one. Otherwise the iterations will converge too slowly, or not at all.
This is not easy and is also the reason why it was a big deal that CMU got
their fitter to work to the satisfaction of the Collaboration.
When dealing with the tagger energies life is a bit easier. The tagger
energy is decoupled from CLAS, in that there are no off-diagonal elements
in the covariance matrix to worry about. This means that you should be
able to obtain a variance close to one for the tagger energy pull by
simply adding in percentages of uncertainty to the tagger uncertainty
values until you get a pull of one.
To my eye, the variance of the tagger energy looks too large. This
suggests that the uncertainties are under-represented. This means that the
contribution of the covariance matrix due to tagger energy is
under-represented. This further implies that this variable gets changed
more slowly in each passing iteration than would be optimal had the
uncertainty been correct in the first place.
When it comes to the confidence levels I should rephrase:
It would be good to make a subset of events that are kinematically
compatible with the reaction:
gamma p -> p pi+ pi- AND no other reaction. This means that confidence
level cut is only screening out events that are not as perfect as one
might like, yet can not be ruled out as being anything other than from the
stated reaction. Once this subset has been made, you can set the
confidence level to zero. It should not matter in the study of the energy
correction.
That is my two cents.
Have a good weekend.
Sincerely,
Michael
On Fri, 21 Aug 2009, Volker Crede wrote:
> Barry, Mike,
>
> I think Chuck has shared his corrections already with the run group, but
> probably did not communicate things properly.
>
> It's not easy to tweak the pulls such that they all have mean zero and
> width one, but I agree it should be the goal. As a matter of fact, the
> whole pull and covariance matrix business is very difficult. Chuck used
> the proton pulls (all three of them) to determine the momentum corrections
> and this also improved the photon energy pull substantially. We never
> touched the energy directly, but you can see the improvement in the plots
> he sent out. We have discovered that the pulls are extremely sensitive to
> changes in the initial 4-vecs. A proton momentum change of less than 1%
> can sometimes move the pull mean from positive to negative. On the other
> hand, mean and width can be studied and adjusted fairly independently. The
> width is mainly based on the error of a variable (and thus determined by
> the covariance matrix), while the mean is dominated by the value of the
> variable itself, more or less independent of its error.
>
> Mike, a cut on the tails of the pull distribution is indeed equivalent to
> a confidence level cut. The difference we are looking at here is just a
> pull that is not properly normalized. However, I agree we should study the
> dependence of CL cuts on the tagger sag. My prediction is that the result
> remains more or less unchanged.
>
> With regard to the four suggestions:
>
> i) Forming "mass X^2" from the reaction "gamma p -> p pi+ pi- X" will not
> work when you use the kinematic fitter. In fact, you need constraints. The
> hypothesis "g p -> p pi+ pi-" is nothing else but energy and momentum
> conservation and requires "X = 0" for all events. All 4-vecs are tweaked
> such that "X = 0" and then we check what CL value we get. If the "X = 0"
> requirement cannot be met, then the fit does not converge and the event is
> not kinematically fitted. To cut a long story short, the "mass X^2" plot
> is a spike.
>
> ii) Well, X = 0 ...
>
> iii) This is a good point, but we need to apply the CL cut; otherwise the
> events may not be fitted. But, we can study the Gaussian 'mean' dependence
> on the CL cut.
>
> Best wishes
>
> Volker
>
>
>
> On Fri, 21 Aug 2009, Barry Ritchie wrote:
>
>> Chuck, these are all good suggestions from Mike. Would you please:
>>
>> 1. Share your momentum corrections with Mike ASAP?
>> 2. Try the four mini-tests he suggests to firm up your results.
>> 3. Things need to be tweaked until the pulls have mean zero and standard
>> deviation 1. Until then, there's still work to do.
>>
>> ---BGR
>>
>> Professor Barry G. Ritchie
>> Department of Physics
>> Arizona State University
>> Tempe, AZ 85287-1504
>>
>> Telephone: (480) 965-4707
>> Fax: (480) 965-7954
>>
>> -----Original Message-----
>> From: g8b_run-bounces at jlab.org [mailto:g8b_run-bounces at jlab.org] On
>> Behalf Of Michael Dugger
>> Sent: Friday, August 21, 2009 3:20 PM
>> To: Charles Hanretty
>> Cc: g8b_run at jlab.org
>> Subject: Re: [G8b_run] Tagger Sag
>>
>>
>> Chuck and g8b,
>>
>> Thanks for all your work.
>>
>> The problem we face is that there has been no real agreement between
>> Chuck, Stuart, or me :(
>>
>> We need some sort of convergence and then everyone needs to use the same
>>
>> momentum and energy corrections.
>>
>> My feeling is that a kinematic fitter is most likely to give acceptable
>> results for determination of the corrections needed.
>>
>> I would like to have Chucks momentum correction and see if that helps me
>>
>> get a better agreement with him on the energy correction.
>>
>> One comment:
>> I do not think you can make this statement->
>> "The projection of this plot onto the y-axis is indeed a Gaussian as the
>>
>> confidence level cut only cuts off the tails of the distribution and
>> does
>> not affect the peak position."
>> If this were true, wouldn't a cut on the tails of the distribution be
>> equivalent to a confidence level cut? Did you test to see if the center
>> changed with change in confidence level cut?
>>
>> Some questions:
>>
>> * The title on the plot
>> http://hadron.physics.fsu.edu/~hanretty/Files/TaggerSag/photonPull_sit04
>> _WithMomCorrs.gif
>> says conf cut at 0.001%. Is this true?
>>
>> * If you correct the photon energies do you get a pull at zero? What is
>> the standard deviation of the pull? It looks different than 1. Don't you
>>
>> need to have all of the pulls close to having a standard deviation = 1.
>> Isn't the kinematic fitter very sensitive to having the correct
>> covariance matrix?
>>
>> Chuck, in the name of comparing apples to apples:
>>
>> * Can you do the same analysis using only amorphous data? It should not
>> change the results, but then we would have the same data in common.
>>
>> * Can you
>> i) form massX^2 from the reaction
>>
>> gamma p -> p pi+ pi- X
>>
>> ii) restrict events to be
>>
>> -.01 GeV^2 < massX^2 < 0.00385 GeV^2
>>
>> iii) turn off confidence level cut.
>>
>> iv) see if your results remain unchanged
>>
>> If the result remains the same (and it should), than wouldn't this prove
>>
>> that the confidence level cut is not forcing the energy correction to be
>>
>> small?
>>
>> Please correct me if I made any stupid comments or silly questions. We
>> need to get this resolved so that we move forward.
>>
>> Thanks for your time.
>>
>> Sincerely,
>> Michael
>>
>>
>>
>> On Fri, 21 Aug 2009, Charles Hanretty wrote:
>>
>>> Hello All,
>>> During our last g8b meeting Mike showed some plots regarding the
>> tagger
>>> sag. I was asked to recreate this plot to verify what Mike saw. I've
>> been
>>> working on this for the past few days and have some plots to share. I
>> have a
>>> few more plots than what Mike showed I did this to verify my
>> verification. :)
>>>
>>> My process:
>>>
>>> STEP 1) Make 2 plots by running over one full run for each coherent
>> edge energy
>>> (a total of 121 data files).
>>>
>>>
>> http://hadron.physics.fsu.edu/~hanretty/Files/TaggerSag/EtrueDivEmeas_v_
>> Emeas_ForSlicing.gif
>>>
>> http://hadron.physics.fsu.edu/~hanretty/Files/TaggerSag/EtrueMinusEmeas_
>> v_Emeas_ForSlicing.gif
>>>
>>> Both of these histos are TH2F's, produced by running over the data
>> files. I
>>> defined Etrue as the photon energy coming out of the kinematic fitter
>> and Emeas
>>> as the photon energy as taken from either the GPID or TAGR bank. For
>> the plot
>>> to be filled with these values, the event must be a ppippim final
>> state
>>> and pass a fit to no missing particle (energy & momentum conservation)
>>> with a confidence level of >10%.
>>> EtrueDivEmeas_v_Emeas_ForSlicing: x-axis-> 0.8-5 GeV, 168 bins
>>> y-axis-> 0.998-1.002, 40 bins
>>> EtrueMinusEmeas_v_Emeas_ForSlicing: x-axis->0.8-5 GeV, 168 bins
>>> y-axis->-0.0024-0.0024, 40 bins
>>>
>>> STEP 2) Use the FitSlicesY() function to slice the *_ForSlicing histos
>>> (above) along the x-axis, one slice per bin, and fit each slice to a
>> gaussian. This
>>> function also makes a histogram containing the mean values of the
>> gaussian
>>> fits, bin by bin:
>>>
>>>
>> http://hadron.physics.fsu.edu/~hanretty/Files/TaggerSag/EtrueDivEmeas_v_
>> Emeas_SlicedFitMeanValues.gif
>>>
>> http://hadron.physics.fsu.edu/~hanretty/Files/TaggerSag/EtrueMinusEmeas_
>> v_Emeas_SlicedFitMeanValues.gif
>>>
>>> The structure that Mike showed in the meeting is clearly seen
>> again
>>> in the *_SlicedMeanFitValues.gif plots (good!). However, as you all
>> have
>>> probably already noticed, my y-scale is much much smaller. This goes
>> back
>>> to the original histograms (the ones made in STEP 1->
>> *_ForSlicing.gif).
>>> I first tried using the exact same scale and binning as Mike and I got
>> this:
>>>
>> http://hadron.physics.fsu.edu/~hanretty/Files/TaggerSag/EtrueMinusEmeas_
>> v_Emeas_MikeRange.gif
>>>
>>> You'll notice that the plot looks like a strip with sharp edges.
>> These
>>> edges are a result of my use of a confidence level cut, if I were to
>> not
>>> use a fitter, I would have a larger spread like what Mike has (this
>>> confidence level cut only allows for events where Etrue and Emeas are
>>> close). Therefore I had to "zoom in" on the y-axis in order carry this
>> process
>>> out and the result of this "zooming" is the
>> EtrueMinusEmeas_v_Emeas_ForSlicing.gif
>>> plot. The projection of this plot onto the y-axis is indeed a
>> Gaussian as the
>>> confidence level cut only cuts off the tails of the distribution and
>> does not
>>> affect the peak position.
>>>
>>> I am assuming that Mike is using ELoss and Stuart's momentum
>> corrections
>>> when he generates these plots. Since Eloss is the same for everyone
>> and I have
>>> my own momentum corrections (that I made using the kinematic fitter,
>> fitting to
>>> a ppippim() final state), I decided to produce the
>>> *_SlicedMeanFitValues.gif plots with the inclusion and exclusion of
>> (my)
>>> momentum corrections to see the effect these corrections have (also to
>> cover
>>> all my bases). The two distributions you see on these plots are with
>> and
>>> w/o momentum corrections (the upper is without, the lower is with).
>> The use of
>>> momentum corrections does not affect the scale of the y-axis, but only
>>> moves the distribution closer to zero (for subtractions) or closer to
>> one
>>> (for the division).
>>>
>>> Photon Pulls for a fit to gamma p -> p pip pim ()
>>> with MomCorrs:
>>>
>> http://hadron.physics.fsu.edu/~hanretty/Files/TaggerSag/photonPull_sit04
>> _WithMomCorrs.gif
>>> without MomCorrs:
>>>
>> http://hadron.physics.fsu.edu/~hanretty/Files/TaggerSag/photonPull_sit04
>> _NoMomCorrs.gif
>>>
>>> Although these pulls are not at zero, they are symmetric. The
>> shifts from
>>> zero are indicative of a systematic error and improve once I include
>> my
>>> momentum corrections. If you look at the effect my momCorrs have on
>> the
>>> *_SlicedMeanFitValues plots, you see that they only shift the
>> distribution and
>>> have no real effect on the y-axis. I would argue that if I were to
>> get all of
>>> my pulls to exactly zero (an "in a perfect world" senario), then the
>>> distribution would only shift down some more but the general shape
>> would
>>> be unaffected.
>>> Reminder: A pull is the difference between the true value (the value
>> from the
>>> fitter) and the measured value (the value found in the data)
>> normalized
>>> to the error of that particular measurement. A pull centered at zero
>>> indicates that the systematic errors for that particular variable are
>>> negligible.
>>>
>>> To summarize: I have seen the same structure in my plots that
>> Mike showed
>>> at our last g8b meeting using two appraoches. Where he and I differ
>> is in the
>>> scale of the y-axis. It seems that this difference in scale
>>> arises from my use of a kinematic fitter and also a cut on the
>> confidence level
>>> for the fit whereas Mike used some iterative routine (not saying that
>> Mike's
>>> routine is garbage). If I were to not use a fitter then I would
>>> have a spread in my *_ForSlicing plots similar to that seen in Mike's
>> plots.
>>> My use of a fitter forces me to "zoom in" on the y-axis if I ever hope
>> to slice
>>> the histogram and fit the slices. This shrinking of scale carries
>> through to
>>> the *_SlicedFitMeanValues plots and forces the amplitude of the
>> distribution
>>> shape to be much smaller. The important thing is I see the same
>> shape.
>>>
>>> -Chuck
>>> _______________________________________________
>>> G8b_run mailing list
>>> G8b_run at jlab.org
>>> https://mailman.jlab.org/mailman/listinfo/g8b_run
>>>
>>
>> _______________________________________________
>> G8b_run mailing list
>> G8b_run at jlab.org
>> https://mailman.jlab.org/mailman/listinfo/g8b_run
>>
>> _______________________________________________
>> G8b_run mailing list
>> G8b_run at jlab.org
>> https://mailman.jlab.org/mailman/listinfo/g8b_run
>>
>
> --
> --------------------------------
>
> Volker Crede
> Assistant Professor
> Florida State University
> Department of Physics
> 206 Keen Building
> Tallahassee, FL 32306 USA
>
> Phone: 001 850 644-2423
> Fax: 001 850 644-4478
> crede at fsu.edu
>
> _______________________________________________
> G8b_run mailing list
> G8b_run at jlab.org
> https://mailman.jlab.org/mailman/listinfo/g8b_run
>
More information about the G8b_run
mailing list