Unfolding & regularization

This section details how to perform an unfolded cross-section measurement, including regularization, inside Combine.

There are many resources available that describe unfolding, including when to use it (or not), and what are the usual issues around it. A useful summary is available at the CMS Statistics Committee pages on unfolding. You can also find a nice overview of unfolding and its usage in combine in these slides.

The basic idea behind the unfolding technique to describe smearing introduced through the reconstruction (eg of the particle energy) in a given truth level bin x_{i} through a linear relationship to the effects in the nearby truth-bins. We can make statements about the probability p_{j} that the event falling in the truth bin x_{i} is reconstructed in the bin y_{i} via the linear relationship,

y_{obs} = \tilde{\boldsymbol{R}}\cdot x_{true} + b

or, if the truth bins are expressed relative to some particular model, we use the usual signal strength terminology,

y_{obs} = \boldsymbol{R}\cdot \mu + b

Unfolding aims to find the distribution at truth level x, given the observations y at reco-level.

Likelihood-based unfolding

Since Combine has access to the full likelihood for any analysis written in the usual datacard format, we will use likelihood-based unfolding throughout - for other approaches, there are many other tools available (eg RooUnfold or TUnfold), which can be used instead.

The benefits of the likelihood-based approach are that,

In practice, one must construct the response matrix and unroll it in the reconstructed bins:

The model we use for this is then just the usual PhysicsModel:multiSignalModel, where each signal refers to a particular truth level bin. The results can be extracted through a simple maximum likelihood fit with,

    text2workspace.py -m 125 --X-allow-no-background -o datacard.root datacard.txt
       -P HiggsAnalysis.CombinedLimit.PhysicsModel:multiSignalModel --PO map='.*GenBin0.*:r_Bin0[1,-1,20]' --PO map='.*GenBin1.*:r_Bin1[1,-1,20]' --PO map='.*GenBin2.*:r_Bin2[1,-1,20]' --PO map='.*GenBin3.*:r_Bin3[1,-1,20]' --PO map='.*GenBin4.*:r_Bin4[1,-1,20]'

    combine -M MultiDimFit --setParameters=r_Bin0=1,r_Bin1=1,r_Bin2=1,r_Bin3=1,r_Bin4=1 -t -1 -m 125 datacard.root
    combine -M MultiDimFit --setParameters=r_Bin0=1,r_Bin1=1,r_Bin2=1,r_Bin3=1,r_Bin4=1 -t -1 -m 125 --algo=grid --points=100 -P r_Bin1 --setParameterRanges r_Bin1=0.5,1.5 --floatOtherPOIs=1 datacard.root

Notice that one can also perform the so called bin-by-bin unfolding (though it is strongly discouraged except for testing) with,

    text2workspace.py -m 125 --X-allow-no-background -o datacard.root datacard.txt
      -P HiggsAnalysis.CombinedLimit.PhysicsModel:multiSignalModel --PO map='.*RecoBin0.*:r_Bin0[1,-1,20]' --PO map='.*RecoBin1.*:r_Bin1[1,-1,20]' --PO map='.*RecoBin2.*:r_Bin2[1,-1,20]' --PO map='.*RecoBin3.*:r_Bin3[1,-1,20]' --PO map='.*RecoBin4.*:r_Bin4[1,-1,20]'

Nuisance parameters can be added to the likelihood function and profiled in the usual way via the datacards. Theory uncertainties on the inclusive cross section are typically not included in unfolded measurements.

The figure below shows a comparison of Likelihood based unfolding and a least-squares based unfolding as implemented in RooUnfold.

Show comparison


The main difference with respect to other models with multiple signal contributions is the introduction of Regularization, which is used to stabilize the unfolding process.

An example of unfolding in combine with and without regularization, can be found under data/tutorials/regularization.

Running python createWs.py [-r] will create a simple datacard and perform a fit both with and without including regularization.

The simplest way to introduce regularization in the likelihood based approach, is to apply a penalty term in the likelihood function which depends on the values of the truth bins (so called Tickonov regularization):

-2\ln L = -2\ln L + P(\vec{x})

where P is a linear operator. There are two different approches which are supported to construct P. If instead you run python makeModel.py, you will create a more complex datacard with each the two regularization scheme implemented. You will need to uncomment the relevant sections of code to activate SVD or TUnfold type regularization.


Any unfolding method which makes use of regularization must perform studies of the potential bias/coverage properties introduced through the

inclusion of regularization, and how strong the associated regularization is. Advice on this can be found in the CMS Statistics Committee pages.

Singular Value Decomposition (SVD)

In the SVD approach - as described in the SVD paper - the penalty term is constructed directly based on the strengths (\vec{\mu}=\{\mu_{i}\}_{i=1}^{N}),

P = \tau\left| A\cdot \vec{\mu} \right|^{2},

where A is typically the discrete curvature matrix, with

A = \begin{bmatrix} 1 & -1 & ... \\ 1 & -2 & 1 & ... \\ ... \end{bmatrix}

Penalty terms on the derivatives can also be included. Such a penalty term is included by modifying the likelihood to include one constraint for each row of the product A\cdot\vec{\mu}, by including them as lines in the datacard of the form,

    name constr formula dependents delta

where the regularization strength \delta=\frac{1}{\sqrt{\tau}} and can either be a fixed value (eg by putting directly 0.01) or as a modifiable parameter with eg delta[0.01].

For example, for 3 bins and a regularization strength of 0.03, the first line would be

    name constr @0-2*@2+@1 r_Bin0,r_Bin1,r_Bin2 0.03

Alternative, valid syntaxes are

    constr1 constr r_bin0-r_bin1 0.01
    constr1 constr r_bin0-r_bin1 delta[0.01]
    constr1 constr r_bin0+r_bin1 r_bin0,r_bin1 0.01
    constr1 constr r_bin0+r_bin1 {r_bin0,r_bin1} delta[0.01]

The figure below shows an example unfolding using the "SVD regularization" approach with the least squares method (as implemented by RooUnfold) and implemented as a penalty term added to the likelihood using the maximum likelihood approach in Combine.

Show comparison

TUnfold method

The Tikhonov regularization as implemented in TUnfold uses the MC information, or rather the densities prediction, as a bias vector. In order to give this information to Combine, a single datacard for each reco-level bin needs to be produced, so that we have access to the proper normalization terms during the minimization. In this case the bias vector is \vec{x}_{obs}-\vec{x}_{true}

Then one can write a constraint term in the datacard via (eg.)

    constr1 constr (r_Bin0-1.)*(shapeSig_GenBin0_RecoBin0__norm+shapeSig_GenBin0_RecoBin1__norm+shapeSig_GenBin0_RecoBin2__norm+shapeSig_GenBin0_RecoBin3__norm+shapeSig_GenBin0_RecoBin4__norm)+(r_Bin2-1.)*(shapeSig_GenBin2_RecoBin0__norm+shapeSig_GenBin2_RecoBin1__norm+shapeSig_GenBin2_RecoBin2__norm+shapeSig_GenBin2_RecoBin3__norm+shapeSig_GenBin2_RecoBin4__norm)-2*(r_Bin1-1.)*(shapeSig_GenBin1_RecoBin0__norm+shapeSig_GenBin1_RecoBin1__norm+shapeSig_GenBin1_RecoBin2__norm+shapeSig_GenBin1_RecoBin3__norm+shapeSig_GenBin1_RecoBin4__norm) {r_Bin0,r_Bin1,r_Bin2,shapeSig_GenBin1_RecoBin0__norm,shapeSig_GenBin0_RecoBin0__norm,shapeSig_GenBin2_RecoBin0__norm,shapeSig_GenBin1_RecoBin1__norm,shapeSig_GenBin0_RecoBin1__norm,shapeSig_GenBin2_RecoBin1__norm,shapeSig_GenBin1_RecoBin2__norm,shapeSig_GenBin0_RecoBin2__norm,shapeSig_GenBin2_RecoBin2__norm,shapeSig_GenBin1_RecoBin3__norm,shapeSig_GenBin0_RecoBin3__norm,shapeSig_GenBin2_RecoBin3__norm,shapeSig_GenBin1_RecoBin4__norm,shapeSig_GenBin0_RecoBin4__norm,shapeSig_GenBin2_RecoBin4__norm} delta[0.03]