Why RooUnblindPrecision give wrong fitting results?

[Jianyu]

Dear experts,

Recently I tried using RooUnblindPrecision to perform a blind analysis, but will always get wrong fitting results. In the following code, RooLLbarEdmPdf is a diy PDF model and has 4 free parameters (a1, a2, Rg2, Ig2). I want to blind the parameter “a1” in this fit, and the details are as follows:

void fit_blind(){
    gSystem->Load("libPhysics");
    R__LOAD_LIBRARY(/home/jyzhang/Tools/RooLLbarEdmPdf/lib/libRBW.so);

    // Creat instance of RooLLbarEdmPdf
    // Definate kinematic variables
    RooRealVar Theta("Theta","Theta",-3.15,3.15);
    RooRealVar Theta1("Theta1","Theta1",-3.15,3.15);
    RooRealVar Theta2("Theta2","Theta2",-3.15,3.15);
    RooRealVar Phi1("Phi1","Phi1",-3.15,3.15);
    RooRealVar Phi2("Phi2","Phi2",-3.15,3.15);

    // Definate free parameters
    RooRealVar a1("a1","a1",0.7,-100,100);
    RooRealVar a2("a2","a2",-0.7,-1.0,0.0);
    RooUnblindPrecision a1_blind("a1_blind","a1_blind","TheBlindingString",0.7,10,a1);
    RooRealVar PL("PL","PL",0);                      // fix to 0
    RooRealVar Rg1("Rg1","Rg1",1.14e-3);
    RooRealVar Ig1("Ig1","Ig1",0.0);                // fix to 0
    RooRealVar Rg2("Rg2","Rg2",0.689e-3,0.0,5e-2);
    RooRealVar Ig2("Ig2","Ig2",0.645e-3,0.0,5e-2);
    RooRealVar RFA("RFA","RFA",0);             // fix to 0
    RooRealVar IFA("IFA","IFA",0);                 // fix to 0
    RooRealVar RHT("RHT","RHT",0);          // fix to 0
    RooRealVar IHT("IHT","IHT",0);               // fix to 0

    PL.setConstant();
    Rg1.setConstant();
    Ig1.setConstant();
    RFA.setConstant();
    IFA.setConstant();
    RHT.setConstant();
    IHT.setConstant();

    // Construct virtual PDF
    RooLLbarEdmPdf *pdf = new RooLLbarEdmPdf("pdf","pdf",Theta,Theta1,Theta2,Phi1,Phi2,a1_blind,a2,PL,Rg1,Ig1,Rg2,Ig2,RFA,IFA,RHT,IHT,true,1);

    // Read toy phspMC
    pdf->readMC("./toy_phsp_1e7.dat", "sig", "");
    // Read toy data
    RooRealVar Weight("Weight","Weight", -2.0, 2.0);
    RooArgSet theSet(Theta,Theta1,Theta2,Phi1,Phi2,Weight);
    RooDataSet *data6 = pdf->readData("./toy_data_2e5.dat", "sig", "", theSet);
    RooDataSet *data = new RooDataSet( data6->GetName(),
            data6->GetTitle(), 
            data6, *data6->get(), 0, Weight.GetName());

    RooFitResult *result = pdf->fitTo(*data, Save(kTRUE));
    a1.Print();
    a1_blind.Print();
    cout << a1_blind.getVal() << endl ;
    result->Print("v");
    return;
}

The input value of parameters are a1 = 0.7519, a2 = -0.7559, Rg2 = 0.689e-3, Ig2 = 0.645e-3.
However the fitting results are:

ERR MATRIX NOT POS-DEF
 PARAMETER  CORRELATION COEFFICIENTS  
       NO.  GLOBAL      1      2      3      4
        1  0.99900   1.000 -0.133 -0.999  0.101
        2  0.13276  -0.133  1.000  0.133 -0.013
        3  0.99900  -0.999  0.133  1.000 -0.101
        4  0.10137   0.101 -0.013 -0.101  1.000
 ERR MATRIX NOT POS-DEF
[#1] INFO:Minimization -- RooAbsMinimizerFcn::setOptimizeConst: deactivating const optimization
RooRealVar::a1 = 8.626 +/- 0.199235  L(-100 - 100) 
RooUnblindPrecision::a1_blind[ state=dummyBlindState value=a1 ] = (hidden)
0.268948

  RooFitResult: minimized FCN value: -1565.36, estimated distance to minimum: 555730
                covariance matrix quality: Full matrix, but forced positive-definite
                Status : MINIMIZE=-1 HESSE=4 

    Constant Parameter    Value     
  --------------------  ------------
                   IFA    0.0000e+00
                   IHT    0.0000e+00
                   Ig1    0.0000e+00
                    PL    0.0000e+00
                   RFA    0.0000e+00
                   RHT    0.0000e+00
                   Rg1    1.1400e-03

    Floating Parameter  InitialValue    FinalValue +/-  Error     GblCorr.
  --------------------  ------------  --------------------------  --------
                   Ig2    6.4500e-04    3.4122e-05 +/-  4.79e-03  <none>
                   Rg2    6.8900e-04    9.5350e-04 +/-  3.41e-07  <none>
                    a1    7.0000e-01    8.6260e+00 +/-  1.99e-01  <none>
                    a2   -7.0000e-01   -7.2559e-03 +/-  1.78e-05  <none>

It seems that I get a wrong fitting results, but if I use parameter “a1” to construct PDF:
RooLLbarEdmPdf *pdf = new RooLLbarEdmPdf(“pdf”,“pdf”,Theta,Theta1,Theta2,Phi1,Phi2,a1,a2,PL,Rg1,Ig1,Rg2,Ig2,RFA,IFA,RHT,IHT,true,1);
Then I will get a correct fitting results:

  RooFitResult: minimized FCN value: -16454.1, estimated distance to minimum: 2.15989e-07
                covariance matrix quality: Full, accurate covariance matrix
                Status : MINIMIZE=0 HESSE=0 

    Constant Parameter    Value     
  --------------------  ------------
                   IFA    0.0000e+00
                   IHT    0.0000e+00
                   Ig1    0.0000e+00
                    PL    0.0000e+00
                   RFA    0.0000e+00
                   RHT    0.0000e+00
                   Rg1    1.1400e-03

    Floating Parameter  InitialValue    FinalValue +/-  Error     GblCorr.
  --------------------  ------------  --------------------------  --------
                   Ig2    5.0000e-05    6.4853e-04 +/-  6.05e-06  <none>
                   Rg2    5.0000e-05    6.8216e-04 +/-  9.02e-06  <none>
                    a1    7.0000e-01    7.5818e-01 +/-  9.52e-03  <none>
                    a2   -7.0000e-01   -7.4815e-01 +/-  9.40e-03  <none>

I don’t know what happend? Is there a problem with my method of using RooUnblindPrecision? I look forward to hearing from you, thank you so much for your help.

Best regards,
Jianyu

[jonas]

Hi, welcome to the ROOT forum!

All this RooFit unblinding stuff has probably not been used and evolved for 10 years, I can’t even tell what this RooUnblindPrecision is supposed to do by looking at the documentation.

Maybe you can explain what you mathematically and practically want to achieve and we can find a way to do it? Thanks!

Jonas

[Jianyu]

Hi Jonas, nice to see your reply. I want to find a proper way to perform blind fit.
The free parameters in my DIY model are a1, a2, Rg2, Ig2. And I’d like to blind the fitting result because this is a precise measurement. For example,:

Input value: a1 = 0.50, a2 = -0.50, Rg2 = 1.00, Ig2 = -1.00
Fitting results: a1 = 0.51, a2 = -0.49, Rg2 = 1.01, Ig2 = -0.99
Blinded fitting results: a1 = 3.45, a2 = -7.82, Rg2 = 4.11, Ig2 = -8.46

In this fit, I want to get the blinded fitting results before “Open Box”.
I would be very glad to hear from you again. Thank you!

Jianyu

[jonas]

Hi!

Sorry I don’t undestand from this what you want yet. What do you expect to be the difference between the “fitting results” and the “blinded fitting results”?

I would have thought that the blinded fit result is exactly the same, only that the value of the blinded parameter is not hidden.

The problem is I can’t tell you what could be wrong with your usage of RooUnblindPrecision without knowing what you actually want to do

I mean if you would expect to get the correct fit results, what do you expect the blinding to actually do?

[Jianyu]

Hi jonas,

The only thing I want to do is use RooUnblindPrecision to blind some of the fitting results , e.g. “a1”.
In the code I posted on the top, I using “a1_blind” to construct the DIY PDF, then fitting to the toy data.

Example 1 (blinded fit)
RooUnblindPrecision a1_blind(“a1_blind”,“a1_blind”,“TheBlindingString”,0.7,10,a1);
RooLLbarEdmPdf *pdf = new RooLLbarEdmPdf(“pdf”,“pdf”,Theta,Theta1,Theta2,Phi1,Phi2, a1_blind,a2,PL,Rg1,Ig1,Rg2,Ig2,RFA,IFA,RHT,IHT,true,1);

My understanding of blinded fit is same as yours: “the blinded fit result is exactly the same, only that the value of the blinded parameter is not hidden.”
Therefore the fitting results of example 1 (blinded fit) and example 2 (nomal fit) should completely consistent except “a1”. Because the printed value of “a1” in example 1 has been “blinded”.

Example 2 (nomal fit)
RooRealVar a1(“a1”,“a1”,0.7,-100,100);
RooLLbarEdmPdf *pdf = new RooLLbarEdmPdf(“pdf”,“pdf”,Theta,Theta1,Theta2,Phi1,Phi2, a1,a2,PL,Rg1,Ig1,Rg2,Ig2,RFA,IFA,RHT,IHT,true,1);

However, from the results which I posted on the top, the fitting results of example 1 (blinded fit) and
example 2 (nomal fit) are very different. The fit status of example 1 is “forced positive-definite”, and the fitted value of “a2”, “Rg2”, “Ig2” deviates from their input values. While the fitting results of example 2 are correct.

This is what I wonder. How can I use RooUnblindPrecision to make my blind fitting successful. Please forgive me if I still haven’t explained clearly.