Meeting #124

Comparison of the data to the theory:

In the last couple of months we made some progress related to the comparison of our data to the simulation. A "trustworthy" comparison is the last thing that is missing before publishing our results. The theoretical groups have provided us with calculations for 35 kinematic points distributed over the whole acceptance of the HRS spectrometer. Each kinematic point represents one combination of (E',\tehta_e) . For such pair the (e,e'd) asymmetries are calculated as functions of pmiss and phi_{dq}.

Old procedure:

To compare the theoretical calculations to the data, the following averaging procedure was first considered. Using real data we determined all important kinematic variables: E', theta_e, pmiss, phi_{dq} for each event. Using this information we checked if theory is available for every considered particle using following approach: Each of the theoretical points represents a center of a kinematic square. If events lies inside a given square, then theory at the center is assigned to it. After selecting the kinematic point we had to check if calculations exist for a given pmiss. Not all pmiss are available in all kinematic points. If the calculations exist then we accept event and calculate the theoretical asymmetry for a given set of kinematic variables. To determine the final theoretical asymmetry as a function of missing momentum, the asymmetries for all events in given pmiss bin were averaged. The experimental asymmetry was determined by using only those data for which theoretical calculations exists. Such procedure was chosen to ensure faithful comparison of the data to the calculations.

1.)

Later we realized that such procedure has many handicaps. By selecting only events that can be equipped with the theory we loose a huge part of the measured statistics. Furthermore, when calculating the theoretical asymmetry we always consider particle as it would come from the center of the bin, which bring certain problems, especially because our kinematic bins are still reasonably large. In particular, after performing a detailed analysis we learned that in each bin, the minimal available missing momentum can be up to 40MeV lower than minimal missing momentum available at the center. This is problematic at low pmiss, because we throw statistics away. Additionally, neglecting low pmiss data in some bins can lead to different final average value.

2.) 3.)

Furthermore, if an event comes from an edge of a selected bin, the theories from other three neighboring bins are equally valid as the one chosen. Hence, we need to correctly decide which theory we going to consider and why. All these questions and problems have led us to the a new approach for comparing data to the simulation.

New procedure:

The main goal of this new approach is to free the measured data of any unnecessary cuts. Hence, we do not want to adjust data to the theory but vice versa. To keep the comparison with the theory valid this means that we need to extend the pmiss range of the calculations.

4.)

This can be achieved by extrapolation existing theories to smaller pmiss. This is not trivial, because Asymmetries have very nonlinear pmiss dependence. However, we need to keep in mind that we need to extrapolate only for approx. 20MeV, which should not cause a dramatic change in the asymmetry. To get best possible guess for the asymmetries in the low miss region we exploited the fact that A(-pmiss, phi_dq) = A(miss, 180+phi_dq). Using this mirroring of the asymmetries to the negative pmiss-axis the extrapolation of the asymmetries turns to a more reliable interpolation over the y-axis. Of course, here we assume that asymmetry is stays humble, which is indeed the case. This is evident from the theoretical calculations for the bins where asymmetries at very low pmiss are a priori available.

5.) 6.)

Once having the theory available for all required pmiss, we went a step further and considered not only one theory at the center of a given kinematic bin but also the theories for the neighboring bins. In this approach we selected a epsilon-surrounding around each event. Its size was chosen such that for each event few closest theories are selected (the radius of the surrounding approximately agrees with the size of one kinematic bin). Then we check which theories are available inside selected zone, e.g., if event sits right on the edge of a bin, algorithm selects theories for all four closest bins. Once we determined which theories are inside our zone, we calculate the distance from the event to each of the considered bin centers. These distances are then used as weights in the calculation of the average of all the theories available inside the selected zone. Larger distance means smaller contributions of a particular kinematic point to the average asymmetry. Together with the average we can also determine the uncertainty(sigma), which is estimated as the difference between the mean and the closest theory. This sigma is a measure for the uncertainty of our theory. This uncertainty is not in any way related to the quality of the theory provided by the theoretical groups, but is solely consequence of our averaging procedure and the fact, that the theory is available only in few discrete points of the acceptance.

7.) 8.) 9.) 10.)
11.) 12.) 13.) 14.)

We perform this algorithm for each event. Similarly as before, we gather these asymmetries in different pmiss bin and in the end calculate the average values, to perform the final comparison of the data to the simulation.

15.)

New results:

The new approach for averaging theory over our acceptance resulted in the following comparison of the theory to the data:

16.)
Last modified: 12/03/13