RooFit Data Visualization Tutorial

Wouter Verkerke (UC Santa Barbara) David Kirkby (UC Irvine)

2000-2002 Regents of the University of California

Wouter Verkerke, UCSB

Data Visualization in RooFit - Overview

RooDataSet(x,y,z)

- Binning

- Projection (x,y,z)

(x)

- Normalization
RooAbsPdf(x,y,z)

Wouter Verkerke, UCSB

RooFit users tutorial

1-Dimensional plots

The basics

Wouter Verkerke, UCSB

1-Dimensional plots class RooPlot

1-Dimensional plots are most frequently used and have special support in RooFit via the RooPlot class: Derives from TH1 for implementation of graphics, axes etc
Container class for plotable objects: doesnt contain any data itself, TH1 member functions operating on data are non-functional Persistable with ROOT I/O (including contents)

Hold a list of objects to be plotted

Datasets (represented as histograms) PDF projections (represented as curves) Any other TObject that can be drawn (e.g. TArrow, TPaveText)

Takes care of normalization PDF projection curves

Unit-normalized curve is automatically multiplied by number of events of last plotted dataset

Facilitates automatic projection of PDFs onto plotted observable

RooPlot knows plotted observable and all observables of last plotted dataset. PDF are automatically Normalized over all known observables Projected over all known observables except the plotted observables
Wouter Verkerke, UCSB

Using RooPlot the basics

A RooPlot class is easiest created from a RooRealVar
RooRealVar x(x,magic x,-10,10); RooPlot* xframe = x.frame() ; xframe->Draw() ;

// To change title xframe->SetName(blah);

Title of RooRealVar

// Alternate frame() methods // change default range, binning RooPlot* xframe = x.frame(-5,5) ; RooPlot* xframe = x.frame(40) ;

Default plot range = limits of x

Wouter Verkerke, UCSB

Using RooPlot Adding datasets

// d contains x,y,z RooDataSet *d ; d->plotOn(frame) ; frame->Draw() ; Poisson errors

// list frame contents frame->Print(v) ;

RooPlot::frame(088aa410): "A RooPlot of "magic x"" Plotting RooRealVar::x: "magic x" Plot contains 1 object(s) (Options="P") RooHist::gData_plot__x: "Histogram of gData_plot__x"

// Adding a dataset also updates // the set of normalization observables frame->getNormVars()->Print(1) ; (x,y,z)
PDFs added after this dataset that depend on y,z will be normalized & projected over y,z
Wouter Verkerke, UCSB

Using RooPlot Datasets and binning

Default binning
RooDataSet *d ; d->plotOn(frame) ; frame->Draw() ;

Custom (non-uniform) binning

// Create binning object RooBinning b(-10,10) ; // Add single boundary b.addBoundary(0.5) ; // Add (x,-x) pairs of boundaries b.addBoundaryPair(1) ; b.addBoundaryPair(2) ; // Add uniform patterns b.addUniform(2,-10,-2) ; b.addUniform(20,2,10) ; RooDataSet *d ; d->plotOn(frame),Binning(b)) ; frame->Draw() ;
Wouter Verkerke, UCSB

Automatic bin content adjustment according to bin size

Using RooPlot Adding PDF projections

// pdf depends on x,y,z RooAbsPdf* p p->plotOn(frame) ; RooAbsReal::plotOn(f) plot on x integrates over variables (y,z) frame->Draw() ;
Adaptive spacing of curve points to achieve 1 precision

Automatic because RooPlot remembers dimensions of last plotted dataset (x,y,z)

// list frame contents frame->Print(v) ;

RooPlot::frame(088aa410): "A RooPlot of "magic x"" Plotting RooRealVar::x: "magic x" Plot contains 2 object(s) (Options="P") RooHist::gData_plot__x: "Histogram of gData_plot__x" (Options="L") RooCurve::curve_gProjected: "Projection of g"
Wouter Verkerke, UCSB

Using RooPlot Adding PDF projections

Change draw option (e.g. Fill) p->plotOn(xframe,DrawOption(F)) // Correction w.r.t default normalization p->plotOn(xframe,Norm(0.7)) ; // Override number of events for PDF normalization p->plotOn(xframe,Norm(RooAbsReal::NumEvent,10000)) ; // Use expected number of events of extended PDF p->plotOn(xframe,Norm(RooAbsReal::RelativeExpected,1.0)) // Raw scale factor (no bin width correction is applied) p->plotOn(xframe,Norm(RooAbsReal::Raw,5.27)); // No variables are projected by default when PDF // is plotted on an empty frame

Modify the default normalization in various ways

(Re)define manually which of the // Enter custom definition of observables PDF variables xframe->updateNormVars(RooArgSet(x,y,z)) ; are observables p->plotOn(xframe) ;

Wouter Verkerke, UCSB

Other RooPlot features

Change attributes of last added plot elements
frame->getAttLine()->setLineColor(kRed) frame->getAttMarker()->setMarkerType(22)

Change the plotting order of contained objects

frame->drawAfter(objectName1,objectName2) ;

Add non-RooFit objects

TArrow *a = new TArrow(0,0,5,7) ; frame->addObject(a) ;

Use frame->Print(v) to see list of object names

Merge contents from another RooPlot

frame->merge(frame2) ;

Curve/histogram 2 calculation
frame->chiSquare() ; frame->chiSquare(curveName,histName) ;
Wouter Verkerke, UCSB

RooFit users tutorial

Projecting out dimensions

Projection via Integration Projecting discrete vs real observables Projection via data averaging Mixing projection methods

Wouter Verkerke, UCSB

Projecting out hidden dimensions - Integration

PDF is always normalized over all observables
Normalization set n = variables PDF and dataset have in common

PDF is projected over all unplotted observables

The plot variable set x = the plotted dimensions of the PDF (for a 1-D RooPlot this is always 1 variable) The projection set p is n x The projected PDF function is Projected observables

Pf ( x) =

f ( x, p) dp

f ( x, p) dxdp

Plotted observables
Wouter Verkerke, UCSB

Projecting out hidden dimensions

Example in 2 dimensions
2-dim dataset D(x,y) 2-dim PDF P(x,y)=gauss(x)*gauss(y)

1-dim plot versus x

Pp ( x) =

p( x, y)dxdy p( x, y)dx

p( x, y)dy

1-dim plot versus y

Pp ( y ) =

p( x, y)dxdy
Wouter Verkerke, UCSB

RooProdPdf automatic optimization

Example in 2 dimensions
2-dim dataset D(x,y) 2-dim PDF P(x,y)=gaus(x)*gauss(y)

1-dim plot versus x

Pp ( x ) =

g ( x) g ( y )dxdy g ( x )dx g ( y)dy

g ( x) g ( y )dy

g ( x) g ( y )dy

g ( x) g ( x )dx

1-dim plot versus y

g ( x )dx g ( y ) = g ( y ) P ( y) = = g ( x) g ( y )dxdy g ( x)dx g ( y )dy g ( y )dy

p
Wouter Verkerke, UCSB

g ( x) g ( y )dx

Projecting out discrete observables

Works the same way as for real observables
Projected discrete dimension is summed over all its states

Example: B-Decay with mixing

dataset(dt,mixState) & PDF(dt,mixState) 1-dim plot versus dt:

Pp (t ) =
Use summation instead of integration for discrete states

= =

Expand summation

p(t , M )dM p(t , M )dtdM p(t , M ) p(t , M )dt

mS

pmixed (t ) + punmixed (t )
mixed

mS

(t )dt + punmixed (t )dt

Wouter Verkerke, UCSB

Projection works universally for real and discrete observables

Projecting out observables Data averaging

An alternative method to project out observables is to construct a data weighted average function:
Integrate over y Sum over all yi in dataset D

Pp ( x) =

p( x, y)dxdy

p( x, y) dy

1 Pp ( x) = N

i =1, N

p( x, y )dx
i

p ( x, yi )

The summed variable (y) is treated as a parameter

PDF is not normalized over y in above example

Can be used to cancel the effect of a disagreement between data and PDF in a projected observable
Example: per-event errors: PDF is usually flat in dtErr, distribution in data is usually peaked.

Wouter Verkerke, UCSB

Selecting data averaging as the projection method

// PDF and data defined elsewhere, // observables:dt,dtErr,mixState RooAbsData* data ; RooAbsPdf* bmixPdf ; // Create frame and plot data as usual RooPlot* dtframe = dt.frame() ; data->plotOn(dtframe) ;
The ProjWData() modifier overrides the projection method of selected variables: Observable dterr will be averaged over the values in dataset projData

// Plot bmixPdf, projecting dterr with data bmixPdf->plotOn(dtframe,ProjWData(dterr,projData)) ; RooAbsReal::plotOn(bmixPdf) plot on dt integrates over variables (mixState) RooAbsReal::plotOn(bmixPdf) plot on dt averages using data variables (dtErr)

ProjWData only controls how observables are projected. It does not override which observables are projected
Wouter Verkerke, UCSB

Example: integration vs. data averaging on per-event errors

Special property of per-event errors: Distribution in data and PDF do not agree OK Integrating out per-event errors RooPlot* dtFrame = dt.frame() ; data->plotOn(dtFrame) ; bmixPdf.plotOn(dtFrame) ; dtFrame->Draw() ;

Projecting per-event errors with data RooPlot* dtFrame = dt.frame() ; data->plotOn(dtFrame) ; bmixPdf.plotOn(dtFrame, ProjWData(dterr,projData)); dtFrame->Draw() ; OK

Wouter Verkerke, UCSB

Selecting data averaging as the projection method

Projection via data averaging may be applied to any observable
Also discrete valued observables

Choosing data averaging instead of integration changes the meaning of the projected function
The theoretical model / experimental data distinction is blurred: the plotted curve takes part of its behavior from the dataset Often applied to non-physics observables (e.g. per-event errors)
Shape of per-event error distribution irrelevant to physics and may be hard to model correctly in a PDF

Can also to applied to well-modeled physics observable: Example: plot t distribution of B-mixing PDF while
projecting the mix state via integration True model/experimental data comparison projecting the mix state with data averaging Compare only dt shape aspect of model with data Any effects that purely arise from PDF/data discrepancy in B 0/B0 Wouter Verkerke, UCSB counter are taken out

Data average projection - Performance

Data-averaged projections can be computationally expensive
Effectively the sum of N curves is plotted (N = #evts in projection dataset)

Projections with large datasets can be accelerated enormously by using binned projection data sets
Works the same way, just provide a binned dataset RooPlot* dtFrame = dt.frame() ; data->plotOn(dtFrame) ; dterr.setFitBins(50) ; RooDataHist projData(projData,projData,dtErr,data) ; bmixPdf.plotOn(dtFrame,ProjWData(projData)); Minor loss of precision may occur, but with sufficient data and a prudent binning net loss may be less than plotting precision Example: unbinned projection with 20K events: 51.2 sec binned projection with 100 bins: 0.2 sec

Also possible when projecting >1 dimensions, and/or discrete dimensions

Simply create a multi-dimensional binned dataset
Wouter Verkerke, UCSB

Integration vs. data averaging - Summary

Default projection method for all observables is integration To override integration method with data averaging method, provide a projection dataset with observables to be averaged
Projection dataset only controls method of projection, not which variables are projection Projection dataset may contain both discrete and real observables Projection dataset may be binned (speed vs accuracy tradeoff)

Any projected PDF observable may be averaged with data instead of integrated Final projection may be combination of data-averaging & integration

Wouter Verkerke, UCSB

RooFit users tutorial

Slicing & Cutting

Plotting a slice in real & discrete dimensions Understanding normalization in slicing

Wouter Verkerke, UCSB

Plotting a slice of a dataset

Use the optional cut string expression
// Mixing dataset defines dt,mixState RooDataSet* data ; // Plot the entire dataset RooPlot* frame = dt.frame() ; data->plotOn(frame) ; // Plot the mixed part of the data RooPlot* frame_mix = dt.frame() ; data->plotOn(frame, Cut(mixState==mixState::mixed)) ; Works the same for binned data sets The target RooPlot will retain the total number of events for future PDF normalizations (not the number of events in the slice)
More about this later

Wouter Verkerke, UCSB

Plotting a slice of a PDF plotSliceOn()

To plot a (projection of a) slice of a PDF use Slice()
RooAbsReal::plotOn(frame,Slice(sliceSet),) overrides default set of observables to project out Argument sliceSet specifies the set of observables that should not be projected out Position of slice is determined by the current value of slice observable

Slice in y

Slice in x

y = y.getVal()

x = x.getVal()

Slicing can be done in real and discrete dimensions Slice set can have an any number of dimensions
Wouter Verkerke, UCSB

Example: plotting mixed-only slice of data and PDF

RooPlot* dtframe = dt.frame() ; data->plotOn(dtframe,Cut(mixState==mixState::mixed)) ; mixState = "mixed" ; bmix.plotOn(dtframe,Slice(mixState)) ; dtframe->Draw() ;

Wouter Verkerke, UCSB

Understanding the normalization for PDF/data slices

fmixed A PDF plotted with plotSliceOn() is normalized to all observables, including the sliced observables, therefore

Pf (t , M )dt 1

p (t , M ) = dt p (t, M ) dt M p (t , M ) dt = f = M p (t, M )dt

M

The integral of a PDF slice projection is not 1!

Integral of PDF projection =

Fraction of mixed events predicted by PDF

Ntotal

The RooAbsData::plotOn() function with cut gives the full (uncut) number of events to the RooPlot so that the final normalization comes out as

r r total PDF C f ( x ) = Pf ( x ) N data f mixed Vbin r slice = Pf ( x ) N PDF Vbin

slice N data

The normalization of the PDF slice curve reflects the PDFs prediction of the slice fraction Wouter Verkerke, UCSB

Understanding the normalization for PDF/data slices

Data has large fraction of mixed events than PDF predicts

PDF and data agree on fraction of mixed events

Wouter Verkerke, UCSB

Slices in a real-valued observable

Real-valued slices have no width
Usually not that useful (equivalent slices in data are usually empty) Finite width slices can be made with different technique (see later)

Example plot:

effect of per-event error

RooPlot* dtframe = dt.frame() ; data->plotOn(dtframe) ; // not a slice dtErr=0.1 ; bmix.plotOn(dtframe,Slice(dtErr)) ; dtErr=0.5 ; bmix.plotOn(dtframe,Slice(dtErr)) ; dtErr=1.0 ; bmix.plotOn(dtframe,Slice(dtErr)) ; dtframe->Draw() ;

Wouter Verkerke, UCSB

Plotting slices with finite width - Introduction

Problem: analytic calculation of the projection of a band of a PDF often very hard or impossible Solution: Numeric solution via ToyMC approach
Construct finite width slice as weighted average of no-width slices:

1) 2) 3)

Generate a sufficiently large ToyMC sample to be plotted Reduce the ToyMC data to the band to be plotted Plot the PDF the usual way, projecting out all unplotted observables via data averaging. Use the reduce ToyMC set as weighting dataset
Wouter Verkerke, UCSB

Plotting slices with finite width - Example

Example setup: (background) Argus(mB)*Decay(dt) + Gauss(mB)*BMixDecay(dt) (signal)
// Plot projection on mB RooPlot* mbframe = mb.frame(40) ; data->plotOn(mbframe) ; model.plotOn(mbframe) ; // Plot mixed slice projection on deltat RooPlot* dtframe = dt.frame(40) ; data>plotOn(dtframe, Cut(mixState==mixState::mixed)) ; mixState=mixed ; model.plotOn(dtframe,Slice(mixState)) ;

Wouter Verkerke, UCSB

Plotting slices with finite width - Example

Example setup: (background) Argus(mB)*Decay(dt) + Gauss(mB)*BMixDecay(dt) (signal)

Reduce dataset before plotting

RooDataSet* mbSliceData = data->reduce("mb>5.27") ; mbSliceData->plotOn(dtframe2, "mixState==mixState::mixed) RooDataSet *toyMC = model.generate( RooArgSet(dt,mixState,tagFlav,mB), 80000); RooDataSet* mbSliceToyMC = toyMC->reduce(mb>5.27); model.plotOn(dtframe2,Slice(mixState), ProjWData(mb,mbSliceToyMC))

Generate a sufficiently large ToyMC sample to be plotted

Reduce the toyMC data to the band to be plotted Plot the PDF the usual way, projecting out all unplotted observables via data averaging.

Wouter Verkerke, UCSB

Plotting non-rectangular PDF regions

The ToyMC projection technique makes no assumptions on the shape of the selected region
Regions of arbitrary size, shape and dimension can be selected

Example: Likelihood projection plot

Common technique in rare decay analyses PDF typically consist of N-dimensional event selection PDF, where N is large (e.g. 6.) Projection of data & PDF in any of the N dimensions doesnt show a significant excess of signal events To demonstrate purity of selected signal, plot data distribution (with overlaid PDF) in one dimension, while selecting events with a cut on the likelihood in the remaining N-1 dimensions

Wouter Verkerke, UCSB

Plotting data & PDF with a likelihood cut

Simple example
3 observables (x,y,z) Signal shape: gauss(x)gauss(y)gauss(z) Background shape: (1+ax)(1+by)(1+cz) Plot distribution in x with cut on likelihood in (y,z)

// Plot x distribution of all events RooPlot* xframe1 = x.frame(40) ; data->plotOn(xframe1) ; sum.plotOn(xframe1) ;

Integrated projection of data/PDF on X doesnt reflect signal/background discrimination power of PDF in y,z
Wouter Verkerke, UCSB

Plotting data & PDF with a likelihood cut

RooDataSet* data = sum.generate(RooArgSet(x,y,z),50000) ; RooAbsReal* pdfProj = sum.createProjection(RooArgSet(y,z),x) ; RooFormulaVar nllFunc("nll","-log(likelihood)","-log(@0)",*pdfProj) ; RooRealVar* nll = data->addColumn(nllFunc) ;

The createProjection() method create a RooPlot* pframe = nll->frame(4.5,7.5,100) ; data->plotOn(pframe) ; projection of sum over x, with (y,z) as observables:
RooDataSet* sliceData =r data->reduce(RooArgSet(x,y,z),"nll<5.2") ;

x y , z , p )dx r f (= , x.frame(40) ; Pf ( y , z, p xframe2 RooPlot* ) = r f ( x , y, z , p)dxdydz sliceData->plotOn(xframe2) ;

sum.plotOn(xframe2,"L",1.0, RooAbsReal::Relative,sliceData) ;

Automatic optimization: If f factorizes as g(x)*h(y,z):

r Pf ( y , z , p ) = = r ) h( y , z , ph ) dx r g ( x, p g )h( y, z,r ph )dxdydz h( y, z , ph ) r h( y, z, ph )dydz

g ( x, p r

Wouter Verkerke, UCSB

Plotting data & PDF with a likelihood cut

Construct per-event likelihood and add as RooDataSet* data = sum.generate(RooArgSet(x,y,z),50000) ; pre-calculated column to the dataset
RooAbsReal* pdfProj = sum.createProjection(RooArgSet(y,z),x) ; RooFormulaVar nllFunc("nll","-log(likelihood)","-log(@0)",*pdfProj) ; RooRealVar* nll = data->addColumn(nllFunc) ; RooPlot* pframe = nll->frame(4.5,7.5,100) ; data->plotOn(pframe) ; RooDataSet* sliceData = data->reduce(RooArgSet(x,y,z),"nll<5.2") ; RooPlot* xframe2 = x.frame(40) ; sliceData->plotOn(xframe2) ; sum.plotOn(xframe2,"L",1.0, RooAbsReal::Relative,sliceData) ;

Plot per-event likelihood distribution to tune cut

Wouter Verkerke, UCSB

Plotting data & PDF with a likelihood cut

RooDataSet* data = sum.generate(RooArgSet(x,y,z),50000) ; RooAbsReal* pdfProj = sum.createProjection(RooArgSet(y,z),x) ; RooFormulaVar nllFunc("nll","-log(likelihood)","-log(@0)",*pdfProj) ; RooRealVar* nll = data->addColumn(nllFunc) ;

Reduce ToyMC projection dataset with RooPlot* pframe = nll->frame(4.5,7.5,100) ; cut on per-event likelihood
data->plotOn(pframe) ; RooDataSet* sliceData = data->reduce(RooArgSet(x,y,z),"nll<5.2") ; RooPlot* xframe2 = x.frame(40) ; sliceData->plotOn(xframe2) ; sum.plotOn(xframe2,ProjWData(sliceData)) ;

Plot PDF with selected ToyMC events

Wouter Verkerke, UCSB

Summary of slice plotting

To project category slices (or no-width real slices) use
RooAbsData::plotOn(frame,Cut(slice_cut_expr),) RooAbsPdf::plotOn(frame,Slice(sliceSet),) ; Normalization of PDF slice projection will reflect the PDF information on fslice, not the fslice of the data

To plot bands, likelihood slices or arbitrarily shaped regions

Use ToyMC projection technique If the number of projected observables is low (<=2) binning the ToyMC projection dataset can speed up the plotting process. Can be used in combination with Slice() to slice in observables no participating in the region cut

Wouter Verkerke, UCSB

RooFit users tutorial

Component plotting

Selecting components to be plotted Slices vs components

Wouter Verkerke, UCSB

Component plotting - Introduction

A PDF that is explicitly constructed as a sum of components via RooAddPdf can plot its components separately
Use Method Components()

Example: Argus + Gaussian PDF

// Plot data and full PDF first // Now plot only argus component sum->plotOn(xframe, Components(argus), LineStyle(kDashed)) ;
Wouter Verkerke, UCSB

Component plotting Selecting components

There are various ways to select single or multiple components to plot // Single component selection pdf->plotOn(frame,Components(argus)) ; pdf->plotOn(frame,Components(gauss)) ; // Multiple component selection pdf->plotOn(frame,Components(RooArgSet(pdfA,pdfB))) ; pdf->plotOn(frame,Components(pdfA,pdfB)) ; // Wild card expression allowed pdf->plotOn(frame,Components(bkgA*,bkgB*)) ; Wildcard option particularly useful for simultaneous PDFs built by RooSimPdfBuilder. Example: simultaneous Gauss+Argus fit over 4 tagging categories
Plots sum of all background PDFs // plot data and full PDF Syntax independent of number and data->plotOn(frame) ; names of index category states pdf->plotOn(frame) ; pdf->plotOn(frame,Components(Argus_*)) ;
Wouter Verkerke, UCSB

Component plotting Multi layer selection

Method plotCompOn() can be called on any PDF, and also works for nested RooAddPdf structures
Selection mechanism works recursively Final component selection is two-step process: 1 - Explicit selection RooAddPdf model

+
RooProdPdf sig RooProdPdf bkg

Components listed by user

*
RooGaussian sigDE RooGaussian sigMB

*RooAddPdf RooGaussian
bkgDE bkgMB

+
RooGaussian bkgMBpeak RooGaussian bkgMBcont
Wouter Verkerke, UCSB

Component plotting Implicit selection

All nodes in the path between each selected node and the top-level node is implicitly selected

RooAddPdf model

2 Implicit selection

+
RooProdPdf sig

*
RooGaussian sigDE RooGaussian sigMB

RooProdPdf bkg

* RooGaussian
bkgDE

RooAddPdf bkgMB

+
RooGaussian bkgMBpeak RooGaussian bkgMBcont
Wouter Verkerke, UCSB

Component plotting Implicit selection

All nodes below each selected node is implicitly selected

RooAddPdf model

2 Implicit selection

+
RooProdPdf sig

*
RooGaussian sigDE RooGaussian sigMB

RooProdPdf bkg

*
RooGaussian bkgDE RooAddPdf bkgMB

+
RooGaussian bkgMBpeak RooGaussian bkgMBcont
Wouter Verkerke, UCSB

Component plotting Code example

Component selection gives feedback on explicit/implicit selection RooAddPdf
model

+
RooProdPdf sig RooProdPdf bkg

RooGaussian sigDE

* RooGaussian
sigMB

RooGaussian bkgDE

* RooAddPdf
bkgMB

+
pdf->plotOn(frame,Components(bkg)) ; RooAbsPdf::plotCompOn(model) directly selected PDF components: (bkg) RooAbsPdf::plotCompOn(model) indirectly selected PDF components: (bkgMBPeak,bkgMBCont,bkg,model)
RooGaussian bkgMBpeak RooGaussian bkgMBcont

Component selection in a PDF slice projection

Use plotOn(frame, Components(compList),Slice(sliceSet),) No special issues, just combine features of Slice() and Components()
Wouter Verkerke, UCSB

RooFit users tutorial

RooSimultaneous

Projecting and slicing RooSimultaneous PDFs

Wouter Verkerke, UCSB

Plotting RooSimultaneous PDFs

Plotting of RooSimultaneous PDFs is not different from any other PDF
Everything works the same as for regular PDFs, except that the index category cannot be projected out via integration Always provide a projection dataset for the index category (or its components if the index category is composite) Otherwise, treat the RooSimultaneous index category as a regular observable Simultaneous PDF for (A,B) plot sum of A,B RooAbsPdf *pdfA, *pdfB; // variables (x,p) RooCategory *cat ; // with state A,B RooDataSet* data // containing (x,cat) RooSimultaneous sim(sim,sim, RooArgList(pdfA,pdfB),*cat) ; // Plot data/PDF for A+B RooPlot *frame = x.frame() ; data->plotOn(frame) ; sim->plotOn(frame,ProjWData(*cat,data)) ;

Needed to project out cat

Wouter Verkerke, UCSB

Plotting RooSimultaneous PDFs

View of RooSimultaneous in 2D

Projection (=summation) over index category

Ro oS im

ind ex ca te go ry

(s) able serv F ob PD nent po Com

Wouter Verkerke, UCSB

Plotting a component PDF of a RooSimultaneous

A component PDF of a RooSimultaneous is a slice of the RooSimultaneous in the index category. Use Slice() not Components()!
Simultaneous PDF for (A,B) plot A only // Plot data/PDF for A only Needed to RooPlot *frame = x.frame() ; calculate fA data->plotOn(frame,Cut(cat==cat::A)) ; cat=A ; sim->plotOn(frame,Slice(cat),ProjWData(cat,data)) ;

Why does plotSliceOn() need data? Normalization works like in regular plotSliceOn()
RooAbsData::plotOn(frame,Cut(cutExpr)) stores total number of events without cut RooAbsPdf::plotOn(frame,Slice()) normalizes projection to 1 * fslice RooSimultaneous needs projection dataset to calculate fslice
Wouter Verkerke, UCSB

Plotting RooSimultaneous PDFs

A slice in the RooSimultaneous index category selects a component PDF

Ro oS im

ind ex

ca te go ry

index = Kaon

) ble(s a serv ob PDF t onen p Com

Wouter Verkerke, UCSB

RooSimultaneous - Projection a slice with data averaging

RooSimultaneous,Slice() and component data averaging
RooSimultaneous needs projection dataset for entire dataset Component PDF needs projection dataset for events in slice only
A A A A B B B B B B 0.12 0.23 0.17 0.43 0.34 0.07 0.19 0.13 0.22 1.05

Apparent problem: need 2 projection dataset with different sizes Solution: RooSimultaneous::plotOn automatically trims the dataset when passing it on to the components plotOn()
Wouter Verkerke, UCSB

RooSimultaneous - Projection a slice with data averaging

// Plot data for index A RooPlot *frame = x.frame() ; data->plotOn(frame,cat==cat::A) ; // Plot PDF slice for index A, project out per-event errors sim->plotSliceOn(frame,ProjWData(RooArgSet(cat,dterr),data)) ; RooSimultaneous::plotOn(sim) plot on x averages with data index category (cat) RooAbsReal::plotOn(sim) plot on dt averages using data variables (dterr) RooAbsReal::plotOn(sim) reducing given projection dataset to entries with cat==A RooAbsReal::plotOn(sim) only the following components of the projection data will be used: (dterr)

RooSimultaneous index projection uses entire dataset Component dterr projection uses subset of dataset with cat==A
Wouter Verkerke, UCSB

RooFit users tutorial

Miscellaneous

Asymmetry plots Likelihood plots Plots in more 1 dimension

Wouter Verkerke, UCSB

Asymmetry plots
RooFit supports generic asymmetry plotting in any RooCategory with (+1,-1) or (+1,0,-1) states
Example: mixState asymmetry of BMixing PDF & data RooPlot* dtframe = dt.frame(40) ; data->plotOn(dtframe,Asymmetry(mixState)) ; bmix->plotOn(dtframe,Asymmetry(mixState), ProjWData(dterr,data)) ; Can be combined with other plot arguments

Wouter Verkerke, UCSB

Asymmetry plots - Features

RooAbsData::plotOn(Asymmetry())
Non-uniform bin sizes OK Points have binomial errors instead of Poisson errors

RooAbsReal::plotOn(Asymmetry())
All regular PDF projection techniques work:
Projection via integration Projection with data averaging Slice plotting ToyMC region plotting

Wouter Verkerke, UCSB

Likelihood scans in 1 dimension

Plot log(L) for a PDF/dataset on a frame
// cpmixPdf and cpmixData previously defined RooPlot* frame = sin2b.frame(0,1,20) ; cpmixPdf->plotNLLOn(frame,cpmixData,1.0,kTRUE) ;

Adaptive NLL sampling used (standard for all RooPlot curves). Explicit control over resolution tunes CPU/precision tradeoff Optional automatic baseline shift to zero

Wouter Verkerke, UCSB

Likelihood contours in 2 dimensions

Interface to MINUIT contour plots

prod.plotNLLContours(data,meanx,sigmax) ;

RooFitResult* r = prod.fitTo(*data,"mhvr") ; RooPlot* frame = new RooPlot() r->plotOn(frame,meanx,sigmax,"ME12VHB") ;

Quick contours from corr. coefs

Wouter Verkerke, UCSB

Plotting in more than 2,3 dimensions

No equivalent of RooPlot for >1 dimensions
Usually >1D plots are not overlaid anyway Methods provided to produce 2/3D ROOT histograms from datasets and PDFs/functions TH2* ph2 = x.createHistogram("x vs y pdf",y,0,0,0,bins) ; prod.fillHistogram(ph2,RooArgList(x,y)) ; ph2->Draw("SURF") ; TH2* dh2 = x.createHistogram("x vs y data",y,0,0,0,bins) ; data->fillHistogram(dh2,RooArgList(x,y)) ; dh2->Draw("LEGO") ;

Wouter Verkerke, UCSB