class RooNumConvPdf: public RooAbsPdf

Numeric 1-dimensional convolution operator PDF. This class can convolve any PDF with any other PDF using a straightforward numeric calculation of the convolution integral

This class should be used as last resort as numeric convolution calculated this way is computationally intensive and prone to stability fitting problems. The preferred way to compute numeric convolutions is RooFFTConvPdf, which calculates convolutions using Fourier Transforms (requires external free FFTW3 package)

RooNumConvPdf implements reasonable defaults that should convolve most functions reasonably well, but results strongly depend on the shape of your input PDFS so always check your result.

The default integration engine for the numeric convolution is the adaptive Gauss-Kronrod method, which empirically seems the most robust for this task. You can override the convolution integration settings via the RooNumIntConfig object reference returned by the convIntConfig() member function

By default the numeric convolution is integrated from -infinity to +infinity through a

x -> 1/x

coordinate transformation of the tails. For convolution with a very small bandwidth it may be advantageous (for both CPU consumption and stability) if the integration domain is limited to a finite range. The function setConvolutionWindow(mean,width,scale) allows to set a sliding window around the x value to be calculated taking a RooAbsReal expression for an offset and a width to be taken around the x value. These input expression can be RooFormulaVars or other function objects although the 3d 'scale' argument 'scale' multiplies the width RooAbsReal expression given in the 2nd argument, allowing for an appropriate window definition for most cases without need for a RooFormulaVar object: e.g. a Gaussian resolution PDF do setConvolutionWindow(gaussMean,gaussSigma,5) Note that for a 'wide' Gaussian the -inf to +inf integration may converge more quickly than that over a finite range!

The default numeric precision is 1e-7, i.e. the global default for numeric integration but you should experiment with this value to see if it is sufficient for example by studying the number of function calls that MINUIT needs to fit your function as function of the convolution precision.