15.11.17

15.11.17

HPTPC Sensitivities:

Since the last group meeting I've run spectral function fake data fits for ND280 in Pmu - cosTheta and Pp - cos Theta.

HPTPC Pmu-Tmu Flux (L), ND280 Pmu-Tmu Flux (R)

HPTPC Pp-Tmu Flux (L), ND280 Pp-Tmu Flux (R)

There's a difference in energy dependence (the flux bins go up in energy cyclically), so there's a difference in energy dependence of the results for the two detectors.

I then made posterior postfit flux plots for the HPTPC and ND280 Spectral Function fake data fits in P_{mu}-cos Theta_{mu} and P_{p}-cos Theta_{mu}.

Difference between HPTPC and ND280 for these plots: Thresholds.pdf

My fit in Pmu-cos dAlpha has finished, and I will make the plots from that and start running in cos phi. I will start running i paralellised chains to speed up these processes.

2018 OA:

Have had some issues with psyche but now that's sorted I'm going to run checks on my changes, adding in new normalisation parameters.

Have also been reading through each change Clarence has made to try and make sure I understand what he's doing.

Will check I can produce all the MaCh3 plots from TechNote 324 using the preexisting executables and scripts.

CCD Characterisation:

At the end of last week I made plots of measured noise vs predicted noise for the 4shooter CCDs. The results were surprising in that the noise was almost the same for each CCD. I am going to take more data on the new vessel and talk again with Gabriela about how I am calculating the predicted noise, and whether we need to take new dark current measurements.

DAQ Expert:

Call outs now much more frequent. I will be giving update at PN meeting on Friday on the status of the experiment.

08.11.17

I'm now the on-call 'DAQ Expert'. Wednesday's are the beam maintenance day so I've been taking the different subsystems in and out of global/local so that they can do their individual tests today. I've had no major issues to deal with so far

Single Transverse Variables - ND280/HPTPC Comparison:

Have previously shown results from HPTPC fits in p-cosTheta and in STVs, and ND280 fits in p-cosTheta. I've now run an ND280 fit in STVs to see if the gain is due to the combination of HPTPC and STVs and not just the STVs.

For the STV fits, the combination of variables which changed the most under cross section parameter variations us used for each sample.

HPTPC:

ND280:

Spectral Function Fake Data:

Last week I'd run Fake Data fits of SF 'data' to RFG 'MC' for P_mu - cosTheta_mu and P_p - cosTheta_mu. This is the overlay of the two flux plots for comparison. The pull on the parameters for the lepton momentum fit seems to vary with the energy more than for the proton momentum fit.

I have started running the same fits (SF Fake Data - RFG Nominal MC, and RFG Asimov for both PmuTmu and PpTmu) for ND280 energy thresholds, and will start doing the same for other kinematic variables.

2018 OA:

I've got Clarence's updated branch of MaCh3 now with psyche3. The issues with compiling last week were in a function called that we don't use so that has been commented out for now. I've started adding in new normalisation parameters being applied for the next oscillation analysis. I am doing this for the near detector fit, and Artur is doing it for SK so we have been confiring over names etc.

So far I've added the parameters to the xsec covariance maker, and made the new covariance matrix. I've also added the new cross section model to the enumeration of models. Artur and I have done exactly the same for those bits.

Where we differ is how they are applied because we have different codes (samplePDF). I have added the parameters where the normalisation bins are initially located in the covariance file, and then in the two functions where they are applied, one at the start when the MC is produced, and one where it gets updated at each step in the MCMC.

I still need to finish off adding in the energy dependence, and will then start running validations.

CCD Characterisation:

I've had troubles using some of the compiled code for calculating the rms for each run so now using non-compiled code. I've got it working now I think so should have the results soon. Also was delayed by linappserv issues

I spoke to Jen briefly about the Vessel data - TREx interface. She said it's best to get it into the format that Zack has been sending them simulations in, but that they can do some processing if necessary. So we just need to remove pedestal/bias/bad pixels so the background is 0 I think.

01.11.17

Made distributions of events in P_mu - cosTheta and P_p - cosTheta, for both SF and RFG nuclear models. I've fitted the SF as data to the RFG as MC for both sets of variables, and also did RFG Asimov fits.

Lepton Momentum - cos Theta:

SF 'Data', RFG Nominal MC RFG Asimov

SF 'Data', RFG Nominal MC RFG Asimov

Proton Momentum - cos Theta:

SF 'Data', RFG Nominal MC RFG Asimov

SF 'Data', RFG Nominal MC RFG Asimov

As in TN-285, see significant change in flux parameters from priors for SF to RFG fit, and smaller change to cross section parameters (for the SF Fake Data fits, MAQE was set to 1.33 and the MEC parameters were set to 0).

Importantly there is a definite change for the proton momentum fit, so the change in RFG weighting is affecting the hadronic variables.

I will now run the same fits for ND280 thresholds to see how much better HPTPC can be at distinguishing between these models. Could also look at how each of the STV distributions is changed by varying the nuclear model, rather than cross section model parameters.

Also have run a (RFG Asimov) fit of single transverse variables to compare to the HPTPC/ND280 fits I've done, to see how much the advantage of the STV + HPTPC fit is due to the STVs rather than the HPTPC. Haven't produced the plots from that fit yet.

2018 OA:

I've got Clarence's branch which is now compatible with Psyche v3. However, I am unable to compile it as it seems like it's not finding my ROOT TXMLEngine class. I get the same error on both Linappserv and Guillimin. I've been messaging back and forth with Clarence and at the moment he's looking back into his setup scripts.

But I can still look at the code and see where/how the coavariance matrices are produced and how we'd input new normalisation parameters.

ND280:

Still going into control room each day so that whenever something needs to be done for the DAQ I can see what to do.

Also, this week the TPC was repaired and I was volunteered to help with the removal of the TPC panel. This involved unscrewing the panel, attaching it to the crane, and holding it steady as it was lifted away. I then helped replace the cover today in a process pretty much the reversal of the removal.

25.10.17

HPTPC Sensitivity:

Previously I showed the ratio of SF to RFG events in P-cosTheta space. This is just showing the same plots but for q^2 - cos theta. After posting I realise the titles don't distinguish SF/RFG. The pattern for each sample is SF, RFG, ratio.

CC0Pi0P:

CC0Pi1P:

CC0PiNP:

CC1Pi0P:

CC1Pi1P:

CC1PiNP:

We see quite a big change, like for p-theta, when there's no pions. I'm trying to understand if this is expected, when there is a pion the nuclear model has less of an effect.

Don't think this really tells us much more than what we'd already seen. The fits Fake Data fits of SF to RFG are running, but had linappserv/farm issues over the weekend.

I also showed the overlay plot of ND280/HPTPC/HPTPC with STVs before. I've started running a fit with ND280 and STVs, to show that anything we gain is due to the combination of HTPTC and STVs, not just the STVs.

I've also made the pion+proton momentum a variable to be able to fit in. I'd wanted to make all protons and all pions as variables to but this will involve a lot more fiddling with psyche than I'd anticipated so I'm not certain it's worth the effort.

Next OA:

Looking through Clarence's latest branch. In conversation with him about changes to the code + trying to read as much as I can on cross-section/past analyses

Now up and running on Compute Canada so going to run some practice fits on there

Also lots of DAQ expert training/reading - and surviving the typhoon!

18.10.17

HPTPC Single Transverse Variables

Gave a talk on HPTPC sensitivity studies at the Near Detector Upgrade workshop before the T2K collaboration meeting. In doing so I realised I'd never really made a full overlay plot comparing HPTPC, ND280 and HPTPC with STV comparison plot. These are shown below. For the STV fit (yellow), the kinematic variable is for each sample is whichever changed the most under xsec parameter variations. These are shown in the table below the images. For the other fits the variables are lepton momentum and angle for all samples.

I think this just further shows us what we thought from the variations of xsec parameters and changing between the RFG and SF nuclear models, that we're not getting the full potential with what I'm doing. You can maybe convince yourself that the STV fit gives the best constraint, but it is not a large or significant effect.

I am going to fit the SF HPTPC P_mu-Theta_mu to the nominal RFG HPTPC P_mu Theta_mu data, but this taking a little bit of time. At the moment the input for both the data and MC for a fit is the splines file, so am going 'above that' and just inserting histograms for the MC but haven't finished that yet. I talked through with Clarence where abouts to do that.

Also on my to do list I still have finishing off making new kinematic variables for the all protons, all pions, and all hadrons kinematic variables rather than just the highest momentum ones.

Near Detector/BANFF Fit

Have got more of an idea about the schedule for the BANFF/MaCh3 validations for the next OA.

From now onwards I'll be helping Clarence make the changes in the cross section and nuclear models to the MaCh3 code. Or at the very least making sure I can do and understand what he's done. By mid November we should have the final inputs and will run the first asimov fits to MC.

At the moment I'm reading through the changes that are going to be made/likely to be made to the models.

HPTPC Prototype

Am taking data on the old vessel to finish off characterising the cameras. After Mark solved the light leak, I never got round to comparing the noise to the predictions. I'm now taking more data to do that.

I am also looking at the detector-TREx interface. This will involve masking, removing hot pixels, subtracting the pedestal etc for data from the vessel so that it can be inputted into TREx. A lot of this code already exists from the DMTPC but I need to get it running for HPTPC. I haven't made much progress with this yet.

DAQ Shifts

I have done most of the training for DAQ shifts over the last two days. I still need to do more training for my DAQ expert shifts but am going into the control room each day so I can shadow the shifters there and be shown DAQ things when the DAQ expert is around.

04.10.17

I've taken out the RFG weighting, and also changed the value of the MAQE and MEC parameters to reproduce the ratio plots for the 2D distributions (but for HPTPC events). There's a much bigger difference than when just changing individual cross section parameters, as expected. There seems to be much less change when there is a pion involved.

The z axis is events per 100 MeV per '0.1 cos theta', apart from the ratio plots, which are Spectra Function to Nominal.

Lepton Momentum and Angle:

CC0Pi0P:

CC0Pi1P:

CC0PiNP:

CC1Pi10:

CC1Pi1P:

CC1PiNP:

Proton Momentum and Lepton Angle:

CC0Pi1P:

CC0PiNP:

CC1Pi1P:

CC1PiNP:

I'm also making different variables to fit in: momentum of highest momentum pion/proton, momentum of all pions/protons, momentum of all hadrons, as particularly for the single transverse variables I think the complete hadronic momentum vector would be useful. And make similar plots as those above for q^2.

I've now made the plots for Data Fit of my reproduction of this year's OA. These all look like those in TN-324, and the event rates match up.

Also packing!

27.09.17

New variables were not giving the change in distributions that we expected.

The number of events changes for different variables, so something is going wrong for a small number of events. When printing out the variables I didn't see any unphysical values, so this I haven't got to the bottom of that yet. (Patrick showed me how to do integrals including the over/under-fill bins so it's definitely not leptons with >30GeV)

To check how the hadronic variables are correlated to the leptonic ones, I plotted Lepton Momentum against Proton Momentum, and Lepton Angle against Proton Angle.

There is quite a bit of degeneracy, so the weighting may not be affecting the hadronic side in the way we need. To take a step back and see if making a big change to the cross section model affected the distributions, I took out the weighting for the Relativistic Fermi Gas. So this has completely changed how the nucleons are modelled within the nucleus, but there is still not much change in the distribution for both the proton and lepton momentum.

I've started making plots from my reproduction of this years near detector OA. The event rates match up with those in the technical note. These are the constraints on the Asimov fit for xsec and beam parameters. I will do the same for the data fit this afternoon.

Also helped move stuff about in the lab and reading up on cross section models

20.09.17

I'd been changing each parameter of the cross section model 1 by 1, by 1 sigma, to see how this affected the spectra of each kinematic variable.

I found that the difference in the spectra (the ratio of events in each bin before and after the 1 sigma variation) was smaller for cos Alpha and cos Phi.

To investigate if this was because the angles are 'cos-ed' I made the non cos-ed kinematic variables and fitted in these. I did not find that the change increased substantially.

dAlpha 1Pi0P:

CA5 RES:

dAlpha 1Pi1P:

CA5 RES:

Ma RES:

dAlpha 1PiNP:

CA5 RES:

dPhi 0Pi1P:

MAQE:

dPhi 0PiNP:

MEC(C):

dPhi 1Pi0P:

FSI Pion Abs:

dPhi 1Pi1P:

CA5 RES:

Ma RES:

dPhi1PiNP:

FSI Inel Low E:

So the alpha and phi variables are changing less than the other variables (delta P, P_proton, P_pion) when cross section parameters are changing. To investigate whether this is physical or a bug somewhere I've looked at the total number of events before and after the change.

For 0Pi1P the number of events is the same for each variable, but after varying a parameter (I've been looking at MAQE) this changes. There are then more events in the P_pion-cosTheta, P_proton-cosTheta and deltaP-cosTheta distributions.

The only difference I can think of is how the events are binned. For each momentum variable, I go up to 30GeV. Is it possible there are events, with the varied MAQE parameter, where the lepton has momentum >30GeV so is not included in the spectra with x axis P_lepton, but with proton/pion momentum <30GeV so there are more events in those distributions. For the angles all possible values are plotted so this would not be an issue. (The pion momentum is initially set to 0 so the total number of events is still applicable for the pion momentum distribution, even in the 0Pi1P sample).

Perhaps pointing towards a bug, for 1Pi1P, the P_lepton-cosTheta distribution has a different number of total events to P_lepton-cosdAlpha and P_lepton-cosdPhi before any parameters are changed. These distributions have the same x axis, and the y axis should include all possible values, so I would expect them to include the same events. I will now look back at how the variables are defined to try see if something is going wrong here.

Also helped remove/reattach the clean tent leg and tidy up in the lab

30.08.17

30.08.17

Light Leak Tests:

I covered each of the cameras with black felt, one by one, and then all together, to try and see where the light is getting in. The results are shown in this pdf: LightTests.pdf

When each camera is covered, the structure in that camera's image is reduced. The others are reduced as well but by less. I don't understand why when camera 3 is covered, the structure is larger in the other images. I would have thought it was something wrong with the run, but the structure is less in the camera 3 image for it.

For each camera, when all the cameras were covered the light leak is reduced, but not gone entirely. This could be because my putting of the felt on the cameras wasn't light tight. Or could mean there's light coming from elsewhere as well.

But we see the light leak with the shutter closed (right), identical to having the shutter open (left):

However, as they're so similar it seems the shutter has had no effect. So the shutter may not be closing as I expected and will email Gabriela about this.

Transverse Variables:

When changing the parameters of the cross section model by +/- 1 sigma, I found that the change in the distribution of events in cos alpha and cos phi was less than for the other variables. Because I'm just taking the ratio of events in bins, and that for most events alpha and phi are close to 0 or pi, I thought this could be because they're cos-ed. I've put in new kinematic variables that can be fitted in, alpha and phi. I haven't finished analysing these yet to see which combinations of variable parameter and sample give the biggest change in spectra.

When I've done that, the next steps are:

-Run 5 fits, changing the kinematic variable for one sample in each

-Look at change in distributions when changing parameters by 3 sigma

-Look at where different event types are in the kinematic variable space

This Year's Near Detector Analysis:

The fit reporducing this summer's near detector analysis has finished. The MC fit seems ok and I will talk to Clarence tomorrow about making plots and analysing the results to check they're the same. But the data fit crashed so is rerunning.

I've now got a Compute Canada account so Clarence will show me tomorrow how to use that, so I can run fits using GPUs.

23.08.17

This Year's Near Detector Analysis:

My fit which should reproduce this summer's near detecter analysis is now running.

Initially had crashes due to root/gsl/psyche but these now sorted.

When it finishes will see how to make the relevant plots and see if results are the same

Transverse Variables:

I've made pdfs with plots showing how the spectra of the new variables vary when each parameter of the cross section model is varied by +/- 1 sigma.

dP_cosTheta_var_Spectra.pdf, P_cosAlpha_var_Spectra.pdf, P_cosPhi_var_Spectra.pdf, P_cosTheta_var_Spectra.pdf, Pproton_cosTheta_var_Spectra.pdf, Ppion_cosTheta_var_Spectra.pdf

I went through and looked at which spectra varied the most for which samples, and which cross section parameter changes caused them.

Definition of the variables:

There was less change for the delta alpha and delta phi spectra. I think this may be at least partly to do with them being cos-ed. For a small momentum imbalance, their values are near 0 or pi, where a small change in alpha or phi means an even smaller fractional change in the cos. The way I was determining change in spectra was using the ratio of the number of events in each bin to the orignal spectra.

These are the highlights:

CC0PI0P:

MaRES:

FSI Inel Low E:

CC0PI1P:

MaQE:

MEC(C):

CC0PiNP:

MaQE:

MEC(C):

CC1Pi0P:

MaRES:

BgRES:

CC Coh Norm (C):

FSI Inel Low E:

CC1PI1P:

Ca5RES:

MaRES:

FSI Inel Low E:

CC1PINP:

Ca5RES:

MaRES:

FSI Inel Low E:

09.08.17

Transverse Variables:

Have re-run fits for all the new variables. Had crashes over the weekend for Pion and Proton momentum, but increasing the required momentum on the Farm script seems to solve that. Those two fits will finish today, the results of the rest are shown below:

This is for better binning of the spectra, but still get less constraints than for original values. The spectra are here: P_cosAlphaFakeData.pdf, P_cosPhiFakeData.pdf, dP_cosThetaFakeData.pdf, Ppion_cosThetaFakeData.pdf, Pproton_cosThetaFakeData.pdf

So now looking more at how changing individual parameters in the cross section changes the spectra.

This Year's OA:

Am reproducing this year's summer near detector oscillation analysis. Clarence talked me through it on Thursday.

I have the data and MC files required, and have got the required branch of MaCh3.

But I'm getting a crash at the moment. Think it's to do with where I've got root/cmt/psyche installed.

If I don't get to the bottom of it today, I'm going to Imperial tomorrow so will be able to talk to Clarence then and hopefully get it running by the end of the week.

Light Leak:

Helped focus the cameras, and took more data to find light leak (see Mark's page)

The mean pixel intensity changes a lot within the run, in both bias and raw images:

02.08.17

CUDA:

Last week I was at a CUDA course in Oxford. CUDA is a parallel computing platform, that multi-threads NVIDIA GPUs to run programs/algorithms quicker.

MaCh3 is written so that it can be run with or without CUDA, I just need to get access to an NVIDIA GPU.

At the course there were both lectures and practical sessions, so I was able to write my own basic kernels, the functions that multiple threads on the GPU run in parallel.

Transverse Variables:

I've now run fits in each of the new transverse variables, with one of the old variables (P-cos(dAlpha), P-cos(dPhi), P_pion-cos(θ), fits in P_proton-cos(θ) and dP-cos(θ) have given me results where the post fit values are very different from the priors).

In green is the plot for proton mormentum. The log likelihood plot for this fit was a lot different to the others (less mixing, slower convergence) so something has gone wrong. I'm currently running tests on how I've binned the monte-carlo data in each variable, to see how this affects the constraints, as for the other combinations of variables that I've done so far the uncertainties are still larger than for the old variables, P-cos(θ).

Light Leak:

Have got taken more data this week to try to get to the bottom of the light leak. Had the shutter closed, but the leak is not consistent.

These two images are both -29 degrees, 2x2 binning, 30s exposure. The first was taken Monday afternoon, the second Tuesday evening:

These are just individual raw events in the run (of 100 images for each). I haven't had the chance to fully analyse and look at average images yet.

This afternoon Mark and I will go through each camera one by one to tie down the light leak and figure out what is going on.

Obviously something has gone wrong for the MEC(O) xsec parameter. This originally had a prior value of 0.27, but I'd changed it to be 1. Something has gone wrong in the way I'm trying to change it, which I am trying to figure out at the moment.

19.07.17

CCD Characterisation:

I have plotted the noise against exposure and binning for each of the 4 CCDs, with the shutters closed.

The flat exposure predictions for the first three cameras are due to the lower dark current for them:

This value gets multiplied by the exposure time, so having a lower dark current reduces the amount the noise increases with increasing exposure:

I spoke to Gabriella about this calculation, but I suspect it's an error here rather than the CCDs behaving peculiarly. I'd expect the noise to increase with exposure, and Gabriela's plots show this happening.

Transverse Variables:

I've now rebinned each spectra of the new variables, for each sample, including the pion and proton momenta.So I've begun running fits.

The fits that I'd started last week gave me nonsense results. Parameters were much larger than their priors and the uncertainties increased. This was due to how I was normalising the histograms of the spectra. The normalised histograms were being fitted to rather than the non-normalised as I'd cloned the histogram incorrectly.

This has now been rectified and the results of the first fit, in delta P and cos theta are shown below:

As these fits finish, I will begin to calculate the error on the predicted event rates at SK, and see which combination of variables gives the best constraint.

For the same camera, the average frame with the shutter closed is does not show the leak (could not take more than 50 bias frames). Exposure 10s, 360 frames, 2x2 binning, -25°C.

For the same camera, the average frame with the shutter closed is shown below (could not take more than 50 bias frames). Exposure 10s, 360 frames, 2x2 binning, -25°C.

But the run with the shutter closed was taken much later in the day than when it was open. Because I wasn't doing bias and events in the same run, we're not comparing like with like. It was much darker when the shutter-closed images were taken.

The 10 bias images taken in the run along with the shutter being open did show the light leak:

So I am now taking a run with 50 bias and 50 events, and will repeat this and add them all up so that we're comparing like with like.

12.07.17

Transverse Variables:

Have begun rebinning the spectra as normalisation pushed everything into the extreme bins. I've checked the raw data and this isn't because the data is at cos = +/- 1, just very close to it.

I've finished rebinning for alpha, so have begun a fit! This is in cos alpha - momentum of final state lepton.

The projection of the cos alpha axis in the spectrum for CC0Pi1P events is shown below, by bin number:

The bins are: {-1, -0.99999, -0.99998, -0.99995, -0.9999, -0.9998, -0.9997, -0.9994, -0.9988,-0.99, -0.95, 0.9, -0.8, -0.4, 0.0, 0.4, 0.8, 0.9, 0.95, 0.99, 0.9999, 0.99995, 0.99998, 0.99999, 1};

There are similar plots for each of the samples, and the momentum bins change for each as well.

I am finishing off doing this for delta phi, delta P, and the proton and pion momenta.

CCD Characterisation:

Investigating whether light leak is coming through the back of the cameras or from another flange.

This is the average image without the bias subtracted for the __ camera. Exposure 10s, 360 images, 2x2 binning, -25°C.

For the same camera, the average bias frame is shown below. Exposure 10s, 360 bias frames, 2x2 binning, -25°C.

The bias subtracted average frame:

Before this, I plotted the value of the noise for varying exposure and binning. These don't follow the predicted trends but can perhaps be understood in the context of the light leak.

Resolution Scan:

I compared the post-fit cross section parameter uncertainties for when using HPTPC energy thresholds and detector efficiencies to when using those for ND280.

HPTPC has biggest impact for pF(O), MaRES, and Eb(O).

The resolution is simulated by adding a random number from a gaussian distribution to the kinematic variables for each event.

For the angle, the original mean of the gaussian was 0.004 radians. For the momentum, it was a fraction of the momentum, which depended on the momentum:

When varying the means, and so varying the resolution, the uncertainties in the three aforementioned parameters are much more scattered than expected:

There seems to be a 'valley' around 1-2x original resolution but need more data point to confirm this. It could be that these variations are less than expected for running the same chain twice.

05.07.17

05.06.17

Transverse Variables:

Last week I showed how changing parameters in the cross section model by +/- 1 sigma changed the transverse variable spectra. The plots were for the MAQE parameter, for which sigma ~0.06. This is slightly higher than the value we were discussing last week, so the spectra have changed a bit less than we'd thought and would have expected, but not so much that we think they are incorrect.

I've normalised the bins of the transverse variable spectra by area, to see if some of the unexpected structure was down to binning effects.

This has got rid of the gaps in the spectra, and the double peaks.

But now everything has been condensed into the few most extreme bins so I am currently rebinning and checking that values are not exactly +/-1 but just very close.

Once we're happy with these spectra I will move onto fits, and seeing which combination variables give the best fit. I will also produce the spectra and fits with the proton and pion momenta.

CCD Characterisation:

I had a long Skype call with Gabriela to discuss her code and how she produced the plots in her thesis chapter on CCDs.

I now understand her code for calculating the predicted noise from the manufacturer's specifications and comparing this to the measured noise, and have made the changes needed to get the plots out that we want.

But when I plot the dark box data we get no increase in noise with exposure time, and linear increase in noise with binning. The latter we'd expect if binning was not done in hardware but this is not the case.

The first points on the plots are in better agreement with the predicted noise than the plot from last week, so if the exposure time and binning dependancies are corrected we should hopefully get the predicted noise values.

I've retaken the data on the vessel to see if we get the same results, and am waiting to analyse them now.

28.06.17

2x2 binning:

I've now got the spectra for each of the new kinematic variables: TransverseVariableSpectra.pdf

Looking into:

-Why are there a gaps in the peaks in 2D spectra

-Are the events which are all in extreme bins exactly cos = +/-1

-Normalising bins by area

-Are these physically what we expect e.g why does 0piNp favour cos alpha and cos phi =-1

I've looked at how changing each parameter of the cross section model by +/- 1 sigma affects these spectra.

To avoid showing ~5500 plots, I've just attached results for varying the MAQE parameter by -1sigma: dP.pdf, dalpha.pdf, dphi.pdf

Seems to be less of an effect when there's a pion involved (but note the z axis scales change)

CCD Characterisation:

With Yuri's help, have got Gabriela's ccd characterisation code running for analysis of dark box runs. However, the first plot I've got out for 1x1 binning shows us getting less noise than predicted using the manufacturer's specifications.

I'm currently checking through the code how the predicted noise is calculated. I've emailed Gabriela about the rms value I've used for the measured noise, and also what the error bars mean. I've just left them in from what was in the code for now. I will also check with her about how she scaled with temperature as that might be where differences have entered.

Note: The 10s data point here is at a different temperature to the others so it is not surprising that it doesn't rise as much as the prediction. I will change this so the prediction accounts for this too

28.06.17

These have Clarence and Patrick's 'blessing' that the results agree and differences are due to statistical fluctuations. The consistently higher flux parameters for 'with transverse' is likely due to high dimensional marginalisation effects.

So now have started running fit using one of the new variables: dP-cos theta.

First I rebinned the fake data in delta P:

The fit is still running at the moment.

Resolution Scan:

I've had multiple crashes while running fits scanning through different resolutions. These started when I started running fits with the transverse variables at the same time. When I'd checked out a new branch of MaCh3 for implementing the transverse variables I'd got some paths pointing to the same places in each so I think that's what the problem was. So I'm now continuing to run fits at different momentum/angle resolutions.

I've also run a fit for ND280 energy thresholds to compare to HPTPC. The constraints are better for most parameters but not all. The parameters with the lowest ratios are the ones I will focus on for presenting the resolution scan.

NB: For these fits MEC parameters are still ~0.27 to be consistent with the rest of the momentum scan I'm doing. Subsequent fits will have them set to 1.

CCD Characterisation:

Have taken lots of data in the dark box using the 4 shooter. These are at different exposure length and binning to be able to reproduce plots in Gabriella's thesis for comparison.

There was an issue with there being a limit on the number of images taken which Yuri solved.

Since the hot weather started, the camera could not get to -25 degrees, so data was taken at a warmer temperature so I'll have to scale it when comparing to manufacturer's specifications. This was done for the DMTPC but in the other direction.

I've been having trouble running the code for the noise calculation, but this does the same job as our 'my_first.cpp' so could proceed with that. The code for comparing to manufacturer spec seems to run fine.

Now have first images taken on vessel! These are on the DAQ at the moment as there's been issues getting it synced/copied now there's a large amount of data on there, and don't want to be running analysis on there.

21.06.17

Transverse Variables:

I ran a fit in p-cos theta with the events scaled by 0.1. The results from this are in agreement with the same fit but with the version of MaCh3 that doesn't take in transverse variables as well.

14.06.17

Resolution Scan:

Scan with events scaled by 0.1 is currently running.

I will plot the ratio of each xsec parameter value uncertainty for HPTPC to ND280, and will then focus on a few parameters where HPTPC has the greatest effect.

For these, I'll plot the uncertainties at different resolutions.

But haven't done an ND280 fit with events scaled by 0.1. This is now running.

In the mean time, below is plot of ratios of uncertainties to show what I mean. This is the ratio of 2x the resolution to original HPTPC fit. i.e for values <1, having better resolution has decreased the uncertainty:

The values seem quite scattered, so I will look at larger differences in resolution as well.

I also am looking into limits on resolution from the magnets and ultimate theoretical resolutions: fermi momentum for momentum, pixel size for angle.

Transverse Variables:

I've coded the modifications to MaCh3 to accommodate the new transverse variables (dAlpha, dPhi,dP).

I've used this run a fit, but with the old variables (cos theta and p) to check it still runs as expected. When I plot the ratios of uncertainties for this fit vs the original p-costheta fit they've changed quite a lot:

However, the original fit used in this plot was one of the first I got out and I've changed parts of samplePDF since then. I'm now running a fit on the transverse variables version of MaCh3 (but with p costheta) with the events scaled by 0.1, so I can compare to an up to date fit.

Once I'm sure the changes haven't affected anything they shouldn't, I will run a fit in the new variables. The goal is to see which combination of variables constrains the uncertainties the most.

4 Shooter Characterisation:

Initially took a 50 frame run to see how temperature of camera increasing with each image taken affects the counts in each pixel (error bars left off as they are quite big due to the spot in the image):

Want to see if this increased indefinitely or levelled off, so I've taken a long (100 frame) run on in the dark box. The mean counts per pixel for each image does not seem to increase though. The only change is there's now a block in the dark box, which gets rid of the spot in the images.

I now have code from Gabriella for characterising the camera noise and comparing with the company specifications and converting to physical units of electrons. Also got her thesis chapter on this to guide me.

When we know camera is working as it should we can light leak check, and want to be doing this for each flange installation.

Once 4 shooter running will do same characterisation for 1sh using windows interface. Will then try to help Yuir integrate 1sh into linux

07.06.17

Mark Scott told me how to correct the bug that was causing the cut off in HPTPC fake data. This was due to the order in when we were checking if the highest momentum negative track exists and selecting the lepton track. This solves the issue, as an example, the momentum-theta plot for CC0pi0p events is below:

Currently running a full fit to check this hasn't affected results.

I've got first results from a resolution scan for the HPTPC. I changed the resolution by a factor of 2 for both the angle and momentum and all combinations of each. The original resolution was 1-13% for momentum, 0.004 for cos theta. The resulting cross section plots are below, but I need to look into better ways of presenting this, maybe having the original resolution values on the same plot for comparison. When I've analysed these results properly I will look at what best to change next i.e going up or down by higher amounts for either momentum or angle. Here higher/lower refers to smearing, so lower is better resolution.

Higher Momentum, Original Theta

Original Momentum, Original Theta

Lower Momentum, Original Theta

Higher Mom, Lower Theta

Original Mom, Lower Theta

Lower Mom, Lower Theta

Higher Mom, Higher Theta

Original Mom, Higher Theta

Lower Mom, Higher Theta

Last Thursday I went to see Patrick and we got started on implementing transverse variables into MaCh3 fits. This is involving changes to samplePDFND2014 which I'm still doing but know what I need to do in the immediate future. I'll now go every Thursday.

I haven't had chance to finish off the macro to look at total noise from dark box images to see if the camera is heating up with each image taken, but is ongoing.

Particle Counts:

31.05.17

Went to Collaboration Meeting in Tokai, and gave update at ND Upgrade workshop.

Here, I spoke to Mark Scott about the cutoff in the fake data at around cos theta = 0.2. I was able to show him exactly what was happening, and he said there was likely a bug. Have left it with him for now.

Also spoke to Simon Bienstock about new splines he'd sent me. These are compatible with new cross section. I've got latest version of MaCh3 also compatible with new cross section, but getting crashes. Simon said the splines files may well not work with MaCh3. I think this is the case rather than it being an issue with new version of MaCh3.

Still running resolution scan for HPTPC momentum and angle. Haven't got on a cluster computer yet, will do this asap.

Going to Imperial tomorrow to talk about getting transverse variables involved in these fits. Will now go every Thursday so should start making progress with this now.

Yuri's shown me how to use m3DAQ for taking dark box images, and software for analysis. I'm writing a macro to get integrated noise for each image in a run. Want to see if this increases with number of images taken, to see if camera is warming and causing an effect. Here's an example of an image taken with 2 second exposure, showing large 'spot':

In the lab: Leak checking and vessel conditioning are still dependent on parts. TPC construction has started.

Last update I showed results of a fit where all events were scaled by 0.1. For the MEC (O) parameter of cross section model, the pre and post fit values differed. This is due to the gaussian distribution being cut off at 0:

The flux parameters are much the same.

Scaling Events:

I then ran a chain where each event is scaled by a factor of 0.1. This was because I hadn't been worrying about POT weighting while just trying to get a fit out that worked.

Now MEC(O) differs from its prior value. This makes me think it's down to low stats as the values jump around quite a lot but I need to look into this a bit more.

Now get less 'over-constraint' on CA5 RES, MA RES, and Bg RES, but still there for CC Nue Norm.

To compare effects of scaling events, below are the posterior plots for the first 4 flux parameters, as well as MAQE, MA RES, and NC Coherent Norm.

(Note the x axis scale change for NC Coherent Norm)

Wider gaussians so larger uncertainty for scaled events. MAQE and MARES seemed to be over constrained, but still have gaussian shape. For NC Coherent Norm, can see peak has moved slightly.

Am going to look more into what each parameter is to try to understand why some of the values change between these fits, or if its just stats.

HPTPC Resolution Scan:

I've started a chain where I've changed the momentum resolution of the detector. I'm going to scan through different resolutions for momentum and angle to see effect on constraints.

Chain running at the moment has increased momentum resolution by a factor of 2, but is still running. Will need to run lots of chains to complete this study but each is taking ~1.5 days. I spoke to MaCh3 people about changing step size but already at roughly optimal fraction of events accepted (step size 0.05, ~20% accepted). Perhaps could use less steps (running 500,00 at the moment) for this study? Or just keep going.

Latest Cross Section Model:

Have installed latest version of MaCh3 and made my changes so it is now compatible with proton multiplicity samples and psyche truth selections.

It all compiles now, I just need to update where the splines files are and then can try to run a chain.

I don't expect the chain to run first time out the box but on course to have it running by the end of the week.

Lab:

M3 Slow and DAQ hard drives are now in their original boxes from MIT. These have new IP addresses:

• m3daq.pp.rhul.ac.uk 134.219.109.68
• m3slow.pp.rhul.ac.uk 134.219.109.69

Both ethernet sockets (T/061 & T/062) in the TPC lab are connected to the PP network. Ethernet cable had been cut in wall but that is now fixed.

Waiting on orders before leak checking and vessel conditioning can take place. Will be early next week.

We've started on TPC stand, then can start TPC construction.

Also, I think my abstract has been accepted for PGR conference so need to prepare that and nearer the time request a practice talk. And need to finish 1st year report.

17.05.17

Am proceeding without the backward going HMNT as this gets rid of cut off at cos theta = 0.2.

I've rebinned the fake data so the bigger bins at higher momenta aren't filled as much. This is shown below for CC0pi0p events:

I fitted this fake data, running a 500,000 step chain:

I've included my previous results to show the differences:

Improvement in match up of pre and post fit values for NC Coherent Norm. But worse for Eb (O).

10.05.17

Now have full splines files:

• One for each configuration of ND280 upgrade
• Some of these are compatible with latest xsec model

And making the changes I'd made to my version to this newer version so that it will be compatible with the psyche truth selections and proton multiplicty samples I've been using.

I've also been looking at why there's a cut-off in the HPTPC fake data at cos\theta = 0.2:

I started truth smearing events from the Genie T2K flux to see if there was anything similar.

Didn't see this effect, but plot is for all CC events (the original cut-off was seen in all samples, but is strongest in cc0pi0p)

See quite a different spectrum for a few reasons. These events were on argon and using different selection efficiencies and detection thresholds. Also using different binning.

I didn't finish trying to recreate the effect in Genie as have been looking into the psyche selection truth code.

We have a backward going selection efficiency and a forward going selection efficiency, with the change between the two being at cos\theta = 0.2.

At the moment both these efficiencies are the same so in theory this shouldn't cause the issue.

But when the cross-over point is changed to 0.3, the cutoff in the spectrum moves to here:

So this is clearly where the effect is being introduced.

I spoke to Mark Scott about this, and he saw a bug in the code which would cause this sort cut off if the two efficiencies were different. I made the corrections but still saw the effect.

To further check this is where issue is I commented out one of the parts where backward efficiencies were used. This gives a more reasonable spectrum (below).

However, this is not correct, we're missing out on backward going leptons. This is just to show this is where the error comes in.

Part of issue is I don't fully understand the code here. But as there was bug before hopefully there is just another one and it can be easily fixed, rather than an intrinsic problem with the geometries used in the simulation.

For the TREx-MaCh3 interface, I've emailed Jen a list of the variables needed for the inputs.

She's going to send the formats of how these are outputted.

Once I've got those and have started coding up the interface properly it should become clearer exactly what's required and can then get more information from Jen if needed.

This is pretty much as expected. The prefit and postfit values match each other closely, and the uncertainties have been constrained for all parameters.

However, in the cross section model, CA5 RES and MaRES are constrained by a lot. Probably more than would be expected.

Another unrelated concern was that one error I'd earlier run into was flux weightings being set to -999. When this happened I just set them to 1 and never got to the bottom of why it happened.

Also in the process of getting larger splines files. These have been produced by Simon Bienstock who has put them on IRODs for me and I am in the process of getting access/permissions to download these. Some of them are compatible with the latest xsec model.

Then next steps are adding anti neutrino samples, including proton and pion information and transverse variables, and using the updated xsec model when it's merged.

• Will then run a fit with MaCh3

04.05.17

Since last meeting I rebinned the HPTPC fake data being produced in MaCh3. The old binning was designed for 2\pi angular coverage, but have 4\pi for HPTPC. Below is a comparison of old and new binning for cc0pi0p sample:

There seems to be an unexpected cut-off at cos\theta = 0.2. Taking the projection to the cos\theta axis shows this more clearly:

The cut-off seems to be in all samples but is most prominent in cc0pi0p.

I've emailed Mark Scott these plots to see if he knows what is going on.

Am also in process of applying truth smearing to the Genie T2K flux file to see if there's anything similar.

12.04.17

Think I've almost got the fit behaving as expected.

I hadn't been adding the fake data in my MaCh3 executable in the right way to be fitted.

I ran a shorter (50,000) step chain and the mean values of the parameters seem to match the inputted values.

A 500,000 step chain died last night (due to linappserv rather than code crashing), but got ~130,000 steps.

Log Likelihood convergences:

Flux parameters:

Xsec parameters:

Cross-section parameters all closely follow priors now.

Half a million step chain is currently running.

05.04.17

Found issue with xsection weightings. These were also being read from splines file in a way I hadn't realised.

When this was set to read in as CC0pi0p for all events the flux weighting was more reasonable, and samplepdf integrals are no longer negative or zero.

I ran a 500,000 step chain overnight, and LogL converges:

But some of the parameters still don't look right.

Haven't had chance to fully investigate if widths have been reduced, and why flux parameters are >>1.

Could be as just setting flux weight to 1 when it was <0. Never got to bottom of why or where it was being set to -999 for some events.

Ongoing:

• TREx - MaCh3 interface, waiting for output file from Jen/Paula at Warwick, and will go to Imperial once a week to work on this with Patrick
• Proton and Pion Momentum and angle, and using transverse variables. Want initial fit out before implementing this

Also been preparing talk for T2KUK F2F and helping in lab

29.03.17

By setting flux weighting to 1 whenever it was <0 (had been -999 for certain events), have got rid of the 'curtain' of points above the convergence line for the sample log likelihood (before on the left, after on the right):

Total log likelihood still converges:

However, for individual parameters of flux and cross-section, have quite un-gaussian shapes:

When I was looking at the flux reweighting, I never tracked down exactly where the weights were being changed, and the integral of the samplepdfs are still 0 for some of the samples.

On Friday, I spoke to Patrick about getting pion and proton information in my MaCh3 fits.

What I've done has given me crashes, but should be fairly straight forward so just need to spend some time looking into that. I've commented it out for now while I'm getting original fit to work.

Also discussed smoothing the interface between TREX and MaCh3, so that information from TREX outputs can be inputted into MaCh3.

Also have been helping unpack equipment in lab, and preparing talks for symposium and T2K UK meeting.

22.03.17

Working through issues with truth studies for HPTPC with MaCh3:

• Last week had error message as the truth vertices from the splines file and psyche weren't matching up
• This was because, like the arrays in the splines file I'd had problems with in the previous week, there wasn't enough elements for the samples I was using
• This is also the case for all the other arrays in the splines file
• So I've added an if statement so that whenever I'm going into these arrays, the event is treated as being CC0pi0p
• With this was able to run a 50,000 step chain for a near detector fit
• But had two error messages

- One event wasn't being found at all in the splines file

• Was still able to complete the fit though
• LogL converges after initial burn-in phase, as shown below

• But for sample Log Likelihood there's a lot of deviation from this convergence

• Unlikely this is caused by the above errors as they were just for two individual events
• But then noticed that the integral of the samplepdfs were incorrect (0 or large negative)
• So looked into reweight function, and the fluxweighting is occassionally -999 (rest are near unity)
• This is from when event is loaded, rather than any of the reweighting multipliers (from splines, POT weights, or another float 'weights')

15.03.17

Still trying to fit HPTPC fake data with MaCh3.

• Had issues with the detector systematics covariance file. There's a TObjArray within the file with an axis for each of the original samples. I now have more samples than this (extra: CC0pi0p, CC0pi1p, CC0piNp, CC1pi0p, CC1pi1p, CC1piNp).
• But this was for ND280 anyway as don't have HPTPC detector covariances yet, so am now ignoring detector systematics.
• Then had troubles with cross-section systematics. In the psyche `makefakedata` equivalent, all events are assumed to be 0pi0p and there is only information for this sample.
• Emailed Mark Scott, from whom I got this version of psyche, and he confirmed that's what he'd done.
• So have inputted similar case into my version of MaCh3 for now.
• Now getting error as 'Chain vertex' and 'psyche vertex' are not equal.
• I think it's an issue with getting truth vertex from splines.
• Currently trying to get to the bottom of this.

Also battling blue screen of death laptop problems at the moment.

08.03.17

MaCh3 execuable now adds the correct Truth Selections. Am able to recreate the fake data histograms I produced in psyche with makefakedata.exe.

Below is an example of this, for the momentum and angle of the outgoing lepton in CC0Pi0P interactions.

So I then went about fitting this with MaCh3.

After a few issues with crashes, it ran and I produced a 500,000 step chain.

The likelihood is shown below:

Burn in phase ~10,000 steps.

However, the uncertainties on the parameters of the cross section, detector, and flux models are not constrained. Most seem to be within 5-10% of initial values (root of the diagonal elements of the covariance matrices).

For example, nddet_cov[350][350] = 0.003467, so the uncertainty is (0.003467)^(1/2) = 0.05888

But RMS of posterior is 0.06121:

Getting some code from Clarence at Imperial to try and diagnose what's going wrong.

Have also run a 5,000,000 step chain but haven't fully looked at the results yet, but the few I have looked at don't show much improvement.

Last week had problems as some of the Truth Selections required a file I didn't have. This was ND280 upgrade specific so I didn't need it at the time but will be doing ND280 upgrade studies soon so have now got this code. It contains 2D histograms (cos\theta and momentum) of the efficiencies for muons, protons, and pions, reconstruction, and mis-identification of particles as muons and pions.

01.03.17

Have now produced a MaCh3 executable which adds the correct Truth Selections
Initially had issues with the syntax of the name of selections
Then selection source codes were looking for a file which I didn't have, but this file isn't needed in HPTPC truth selections

• Emailed Simon (whose directory the file is in), he could send it but it's ND280 upgrade specific and also the file needed depends on which version of truth selections I'm using
• The file is efficiencies.root, containing efficiencies for muons, pions, and protons, and misId for each of the particles wanted to be reconstructed as muon or pions

• I've hard coded an if statement to change between Truth and Reconstructed lepton candidates.

• Will then run a fit with MaCh3

22.02.17

For last few weeks I haven't been doing what I thought I'd been doing. I was making fake data with Psyche rather than MaCh3.

Am now making an MaCh3 executable for HPTPC, with the samples/selections needed.

Will then produce the fake data with MaCh3, and fit it.

When I'd fitted this fake data before I was getting the same likelihoods. This is because I was fitting it incorrectly

Prior to that, I'd had an issue with Psyche so copied in a backup version so recreated the fake data to check that hadn't affected anything.

This was identical to how it was before. Below is an example of the data for CC0Pi0P interactions:

--and also what different executables do

15.02.17

At collaboration meeting, I attended talks on the ECAL, NIWG, cross-section, BANFF, OA, ND upgrade, and plenary sessions on all aspects of the experiment.

Haven't got to the bottom of why the likelihoods for the fitting of the two fake data sets (HPTPC and ND280) are the same. But there's still a few things I need to check.

Once I'm sure the fittings have worked correctly, will look at how best to vary cross-section model and compare HPTPC/ND280.

Looks as expected, with initial burn in phase before levelling out.

Next step is to vary underlying cross-section model and fit the fake data, ultimately to see if HPTPC is able to tell us more about where the models go wrong than ND280.

01.02.17

Have now got new psyche code working. Added a string as argument of addSelection function, which says which selection is using the sample.

Using this, I recreated Fake Data for HPTPC and ND280 to check it's still the same as before changing the code.

New data matches that produced before changing the code, as expected. The same is true for the ND280 fake data.

Now doing asimov fit of this data with MaCh3, and will compare fits for ND280 and HPTPC. Will then vary underlying cross-section model and fit the fake data, to see if HPTPC is able to tell us more about where the models go wrong than ND280.

Need to look at how best to rebin the data so not all in same few bins.

Will talk to people at collaboration meeting next week about how to go about varying the cross-section model.

Will then run ND280 and HPTPC fake data through MaCh3.

10.01.17

21.12.16

Still haven't fitted ND280 and HPTPC fake data with MaCh3.

Having trouble with new Psyche code.

MaCh3 doesn't use internal psyche parameters file, but relies on 1-1 correspondence between selection and a sample.

With new psyche, multiple selections use the same sample.

In conversation with Mark Scott about how best to proceed.

14.12.16

Mark Scott sent over updated psyche code with truth based selection class, along with an example truth spline file.

Went to Imperial on Friday to see Patrick. He helped me get the new psyche code running, and I got the full spline files.

Created fake data for both the ND280 and HPTPC:

Distributions are consistent with each other, higher density for high cos \theta and low momentum.

Now going to going to fit the data to varying cross-section models.

Hope to get this done by Christmas.

Then push this to far detector.

Now attending MaCh3 meetings, Oscillation Analysis meetings, and ND BANFF meetings.

Similar to energy bins last week, can look at correlation for cross section bins:

Appear negatively correlated (?). As bin numbers get further apart, bins become less correlated:

- MaCh3

30.11.16

Playing with MaCh3, looking at the root TTree created from T2K data.

Cross section and flux covariance behave as expected for MCMC:

Burn in phase of same length as for log likelihood.

Similar to energy bins last week, can show correlation for cross section bins:

As bin numbers get further apart, bins become less correlated.

Still need to investigate other branches in tree.

-Test fit done by Christmas

The next two plots show the energy of the highest filled bin against the 10th, and 100th highest.

The correlation has decreased, and decreases further for the 99th bin:

If the energy of the highest filled bin is increased, the energy of the second highest bin is also likely to be increased, but the energy of lesser filled bins is not likely to be related.

23.11.16

Solved issue running MaCh3. Installing/accessing splines was main issue.

Have now run Near Detector data through MaCh3. The following plot shows the convergence Log Likelihood during the Markov process.

The first ~10,000 entries are the burn-in phase. These are rejected, to reduce the effect of the initial choice parameter values.

The following plots are a sanity check, making sure I understand what's going on. The first plot shows the energy of the highest filled bin against the energy of the second highest for each simulation. These are highly correlated, as expected.

The next two plots show the energy of the highest filled bin against the 10th, and 100th highest. The correlation decreases.

Currently reading various T2K technical notes and getting to grips with MaCh3.

16.11.16

16.11.16

Went to Imperial a couple of weeks ago to see Patrick and Clarence and install MaCh3:

- Needed to get Root in my directory on scratch, as well as CMT

- Had trouble accessing GSL libraries, as they're in a different place on Royal Holloway servers to Queen Mary, and changing file paths didn't work

- Eventually just pointed to where Asher had reinstalled GSL

- Also got various log ins for T2K intranet, T2KUK wiki account, Slack, EZuce, GitHub repositories etc

Now have got iRods installed to obtain Global Analysis file:

- Have obtained a sample ND280 fit to get used to the systems with

- Currently have an issue running this data with MaCh3

Will then look at getting BANF:

- Used for MC fit fake data at Near Detector, reproducing Far Detector fake data.

-- WilliamCharlesParker - 16 Nov 2016