update to roofit/roostats release notes for 5.24

git-svn-id: http://root.cern.ch/svn/root/trunk@29199 27541ba8-7e3a-0410-8455-c3a389f83636

roofit/doc/v524/index.html

@@ -3,6 +3,115 @@
 <a name="roofit"></a> 
 <h3>RooFit</h3>
 
+This release of ROOT contains RooFit version 3.00. A summary of new features is listed below
+
+<h4>Error visualization</h4>
+It is now possible to visualize the effect of the uncertainties on parameters from a fit
+on any p.d.f. or function projection. To do so use the new <tt>VisualizeError()</tt> argument
+in a plotOn call
+
+<pre>
+    RooFitResult* fr = pdf->fitTo(*data,Save(),...) ;
+    pdf->plotOn(frame,VisualizeError(*fr),...) ;
+</pre>
+
+Two techniques for error visualization are implemented. The default is
+linear error progation, and results in an error band that is by
+construction symmetric. The linear error is calculated as 
+<pre>
+
+   error(x) = Z* F_a(x) * Corr(a,a') F_a'(x)
+  
+   where     F_a(x) = [ f(x,a+da) - f(x,a-da) ] / 2,
+  
+           with f(x) = the plotted curve
+                'da' = error taken from the fit result
+          Corr(a,a') = the correlation matrix from the fit result
+                  Z = requested significance 'Z sigma band'
+  
+</pre>
+The linear method is fast (requires 2*N evaluations of the curve,
+where N is the number of parameters), but may not be accurate in the
+presence of strong correlations (~>0.9) and at Z>2 due to linear and
+Gaussian approximations made.
+<br><br>
+Alternatively, errors can be visualized using a sampling method. In
+this method a number of curves is calculated with variations of the
+parameter values, as sampled from a multi-variate Gaussian p.d.f. that
+is constructed from the fit results covariance matrix.  The error(x)
+is determined by calculating a central interval that capture N% of the
+variations for each valye of x, where N% is controlled by Z (i.e. Z=1
+gives N=68%). The number of sampling curves is chosen to be such that
+at least 100 curves are expected to be outside the N% interval. Intervals from
+the sampling method can be asymmetric, and may perform better in the
+presence of strong correlations, but may take (much) longer to
+calculate. The sampling method also assumes that the uncertainty on the
+parameters can modeled by a multi-variate Gaussian distribution.
+<br><br>
+A complete example is provided in a new tutorial macro <tt>rf610_visualerror.C</tt>,
+the output of which is shown below. 
+<br><br>
+   <img src="roofit_524_visualerror.gif" alt="Graphics Output of macro rf610_visualerror">
+<br><br>
+It is also possible to visualize partial errors (from a subset of the parameters),
+as shown above.
+
+
+<h4>Binned dataset generation</h4>
+
+A new method <tt>RooAbsPdf::generateBinned()</tt> has been implemented
+that samples binned datasets (<tt>RooDataHist</tt>) from any
+p.d.f. 
+<pre>
+   RooDataHist* data = pdf.generateBinned(x,10000) ; 
+</pre>
+This binned generation interface samples the p.d.f. at each bin
+center and applies a Poisson fluctuation to each sampled value. 
+The binning of the returned <tt>RooDataHist</tt> is controlled by the default
+binning associated with the observables generated. To set the number
+of bins in x to 200, do e.g. <tt>x.setBins(200)</tt> prior to the call
+to <tt>generateBinned()</tt>
+<br><br>
+The binned dataset generation method does not (yet) support the concept of
+prototype datasets.
+
+
+<h4>New minimizer interface to Minuit2, GSLMinimizer etc...</h4>
+
+A new minimizer interface, <tt>RooMinimizer</tt> has been added (contribution
+from Alfio Lazarro). The new minimizer is similar in functionality to
+the existing class <tt>RooMinuit</tt>, but supports the new ROOT abstract
+minimizer interface and supports multiple minimizer packages and algorithms
+through that interface.
+<br><br>
+The present interface of <tt>RooMinimizer</tt> is identical to that of <tt>RooMinuit</tt>
+with two extensions
+<ul>
+<li> The <tt>setMinimizer(const char*)</tt> method allows to choose between "minuit" and "minuit2") 
+     as implementation for <tt>migrad(),hesse(),minos()</tt> etc...
+<li> The <tt>minimizer(const char* package, const char* alg)</tt> provides a completely generic interface
+     to all minimizers, where <tt>package</tt> is the package (minuit,GSLminimizer) and <tt>alg</tt> is the algorithm (migrad) to be used
+</ul>
+By default, <tt>RooMinuit</tt> is still used when <tt>RooAbsPdf::fitTo()</tt> is called, but can be overridden with
+a <tt>Minimizer()</tt> named argument
+<pre>
+   // Minimization with MINUIT/MIGRAD through RooMinuit
+   pdf->fitTo(data) ; 
+
+   // Minimization with MINUIT/MIGRAD through RooMinimizer 
+   pdf->fitTo(data,Minimizer("minuit")) ; 
+
+   // Minimization with MINUIT2/MIGRAD through RooMinimizer 
+   pdf->fitTo(data,Minimizer("minuit2")) ; 
+
+   // Minimization with GSLMultiMin/conjugatefr through RooMinimizer
+   pdf->fitTo(data,Minimizer("GSLMultiMin","conjugatefr")) ; 
+</pre>
+
+Note that installation of GSL and the ROOT MathMore package is needed to access the GSL Minimizers and that the GSL
+Minimizer do not implement error analysis.
+
+
 <h4>New numeric integration algorithms available</h4>
 
 <p>RooFit can now interface all MathCore numeric intergration
@@ -19,22 +128,6 @@ causing MINUIT converge problems due to numeric precision issues.
 <p>In future release some more numeric integrators will be migrated to
 a MathCore implementation.
 
-<h4>Optional persistent caching of numeric integrals</h4>
-
-<p>For p.d.f.s with numeric integrals that remain difficult or very time consuming,
-a new persistent caching technique is now available that allows to precalculate
-these integrals and store their values for future use. This technique works transparently
-for any p.d.f. stored in a RooWorkspace.
-
-<p>One can store numeric integral values for problems with zero, one or two floating parameters.
-In the first case, the value is simply stored. In cases with one or two floating parameters
-a grid (histogram) of integral values is stored, which are interpolated to return integral
-values for each value of the parameters.
-
-<p>A new tutorial macro <tt>rf903_numintcache.C</tt> has been added to <tt>$ROOTSYS/tutorials/roofit</tt>
-to illustrate the use of this feature.
-
-
 <h4>Interface to TFoam adaptive MC sampler added</h4>
 
 <p>RooFit can now use the <tt>TFoam</tt> adaptive MC sampler for event generation of p.d.f.s that
@@ -51,6 +144,25 @@ numeric integration algorithm.
 <p>A new tutorial macro <tt>rf902_numgenconfig.C</tt> has been added to <tt>$ROOTSYS/tutorials/roofit</tt>
 to illustrate the use of the steering.
 
+<p>A macro that demonstrates of the power of these newly interface numeric algorithms is provided at the
+end of the RooFit section of the release notes.
+
+<h4>Optional persistent caching of numeric integrals</h4>
+
+<p>For p.d.f.s with numeric integrals that remain difficult or very time consuming,
+a new persistent caching technique is now available that allows to precalculate
+these integrals and store their values for future use. This technique works transparently
+for any p.d.f. stored in a <tt>RooWorkspace</tt>.
+
+<p>One can store numeric integral values for problems with zero, one or two floating parameters.
+In the first case, the value is simply stored. In cases with one or two floating parameters
+a grid (histogram) of integral values is stored, which are interpolated to return integral
+values for each value of the parameters.
+
+<p>A new tutorial macro <tt>rf903_numintcache.C</tt> has been added to <tt>$ROOTSYS/tutorials/roofit</tt>
+to illustrate the use of this feature.
+
+
 <h4>Representation of function and p.d.f. derivatives</h4>
 
 <p>A new class has been added that can represent the derivative of any p.d.f or function w.r.t. any
@@ -59,6 +171,11 @@ parameter or observable. To construct e.g. a first order derivative of a Gaussia
     RooAbsReal* dgdx = gauss.derivative(x,1) ; 
 </pre>
 <p>A more complete example is available in the new tutorial macro <tt>rf111_derivatives.C</tt>
+<br><br>
+   <img src="roofit_524_derivatives.gif" alt="Graphics Output of macro rf111_derivatives">
+<br><br>
+
+
 
 <h4>Improved handling of chi-squared fits</h4>
 
@@ -100,7 +217,7 @@ new functionality.
 create a likelihood from a p.d.f and a dataset rather than constructing a <tt>RooNLLVar</tt>
 object directly. This is because part of the likelihood construction functionality such a using
 multiple <tt>Range()</tt>s, or the inclusion for constraint terms are only available through
-<tt>createNLL</tt>.
+<tt>createNLL()</tt>.
 
 <p>To promote the consistency of this interface, a similar method <tt>RooAbsReal::createChi2()</tt>
 has been added to construct chi-squared functions of a dataset and a function or p.d.f.
@@ -112,9 +229,9 @@ recast a profile of a profile into a single profile object.
 <h4>Multivariate Gaussian modeling of parameters estimates from a fit</h4>
 
 <p>You can now construct a multivariate Gaussian p.d.f on the parameters of a model that 
-represents the result of a fit, from any RooFitResult object.
+represents the result of a fit, from any <tt>RooFitResult</tt> object.
 <pre>
-   RooAbsPdf* paramPdf = fitresult->createPdf(RooArgSet(a,b)) ;
+   RooAbsPdf* paramPdf = fitresult->createHessePdf(RooArgSet(a,b)) ;
 </pre>
 <p>The returned object is an instance of the newly added class <tt>RooMultiVarGaussian</tt>, that can
 model correlated Gaussian distributions in an arbitrary number of dimensions, given a 
@@ -129,7 +246,8 @@ that many) and implements in effect internal generation strategy for its observa
 <p>The FFT convolution operator p.d.f. class <tt>RooFFTConvPdf</tt> has been substantially upgraded
 for improved performance has several new options
 
-<p>For the overflow buffering, which aims to reduce cylical spillover from the FFT convolution, 
+<ul>
+<li>For the overflow buffering, which aims to reduce cylical spillover from the FFT convolution, 
 a choice of three algorithms is now provided:
 <ol>
 <li> Extend the p.d.f. somewhat beyond its original domain (the new default)
@@ -138,13 +256,24 @@ a choice of three algorithms is now provided:
 </ol>
 
 The new default algorithm provides a more sensible result for p.d.f.s with significant
-spillover issues, provided that the p.d.f. can be continuated beyond its original domain.
+spillover issues, provided that the p.d.f. can be continuated beyond its original domain.<br><br>
 
-<p>It is now also possible to express FFT convolutions in terms of other observables than the
+<li>Convolution in non-observables is also explicitly supported now. One can e.g. construct a p.d.f 
+of the form G(x) = Int[dy] ( F(x,y) (*) H(y) ). A new tutorial macro <tt>rf211_paramconv</tt> illustrates 
+how such convolutions can be constructed<br><br>
+
+<li>It is now also possible to express FFT convolutions in terms of other observables than the
 convolution observable itself. A common occurrence of that situation is a (circular) convolution a polar
 angle theta, for a p.d.f. that is ultimately expressed in terms of cos(theta). 
 A new tutorial macro <tt>rf210_angularconv</tt> illustrates how to convolutions of angular observable
 with or without an optional cosine transformation for the final observable.
+</ul>
+
+
+<br><br>
+   <img src="roofit_524_convolution.gif" alt="Graphics Output of macro rf210_angularconv">
+<br><br>
+
 
 
 <h4>Option for improved calculation of errors in weighted likelihood fits</h4>
@@ -156,11 +285,66 @@ scale with the sum of the weights, rather than the number of the events in the d
 if you double all event weights, all parameter errors will go down with sqrt(2)). In chi-squared
 fits event weights can processed correctly by using both the sum of the weights and the 
 sum of the weights-squared for each bin. The newly added option <tt>SumW2Error()</tt> implements a similar
-strategy for (unbinned) weighted ML fits (more documentation to follow for the next release)
+strategy for (unbinned) weighted ML fits by applying a correction to the covariance matrix
+as follows
+<pre>
+   V' = V C-1 V
+</pre>
+where <tt>V</tt> is the covariance matrix from the fit to weighted data, and <tt>C-1</tt> is the inverse of the
+covariance matrix calculated from a similar likelihood that constructed with the event weights applied squared
+
+<h4>Redesign of RooFit dataset class structure</h4>
+
+The original class structure of RooFit featured an abstract dataset
+class <tt>RooAbsData</tt>.  Inheriting from that was a single class
+<tt>RooTreeData</tt>, which implemented datasets with a ROOT
+<tt>TTree</tt>-based storage implementation, and inheriting from that
+two classes <tt>RooDataSet</tt> , representing unbinned data, and
+<tt>RooDataHist</tt>, representing binned data. A main problem with
+this structure was that the implementation of the storage technology
+(<tt>TTree</tt>) and the data representation (binned vs unbinned) were
+intertwined. 
+<br><br>
+Starting with version 3.00, the class structure has been
+rearranged: Now classes <tt>RooDataSet</tt> and <tt>RooDataHist</tt> inherit directly
+from class <tt>RooAbsData</tt>, and class <tt>RooAbsData</tt> now owns an object that
+inherits from <tt>RooAbsDataStore</tt> that implements the storage of the
+data. This new class structure allows multiple data storage implementations to
+be applied efficiently to both <tt>RooDataSet</tt> and <tt>RooDataHist</tt>
+At present a single implementation of <tt>RooAbsDataStore</tt> exists,
+class <tt>RooTreeDataStore</tt>, that contains the storage implementation
+formerly implement in class <tt>RooTreeData</tt>. Methods in class <tt>RooTreeData</tt>
+that were not specific to the storage technology have been moved to
+class <tt>RooAbsData</tt>.
+<br><br>
+If your user code only uses the classes <tt>RooDataSet</tt>,<tt>RooDataHist</tt> and <tt>RooAbsData</tt>
+nothing will change: Existing <tt>RooDataSet</tt>s and <tt>RooDataHist</tt>s
+(that inherit from <tt>RooTreeData</tt>) can be read in <i>without problems</i> in
+RooFit 3.00 and will be converted on the fly to the new dataset structure
+in memory.
+<br><br>
+<i>User code that explicitly uses <tt>RooTreeData</tt> pointers should
+be changed to <tt>RooAbsData</tt> pointers</i>. This change should be transparent
+for all uses, with the exception of the <tt>RooTreeData::tree()</tt> method. 
+Explicit access to tree implementation can still be obtained
+through the <tt>RooTreeDataStore::tree()</tt> method. (A pointer to the datastore
+can be obtained through the <tt>RooAbsData::store()</tt> method.)
+Note that in future releases it is no longer guaranteed that all datasets are implemented
+with a plain <tt>TTree</tt> implementation, so any user code that uses the tree
+implementation directly should implement checks that the implementation
+is indeed tree-based (<tt>data->store()->InheritsFrom(RooTreeDataStore::Class())==kTRUE)</tt>).
+<br><br>
+In future release additional implementations of <tt>RooAbsDataStore</tt> will
+be provided that will support new dataset functionality such as the
+ability to construct 'joint' dataset from two input datasets without
+the need to copy the input data and 'filtered' datasets that represent
+a reduced view (in dimensions or by selecting events) of a dataset
+without the need to copy content.
+
 
 <h4>Various workspace improvements</h4>
 
-A number of smaller and larger improvements has been made to the RooWorkspace class.
+A number of smaller and larger improvements has been made to the <tt>RooWorkspace</tt> class.
 
 <ul>
 <li><p><b>Direct interactive access to contents from CINT</b> - 
@@ -303,18 +487,87 @@ various features of the object factory
 the workspace. If an error is detected in the expression, no objects
 will be committed to the workspace, thus leaving no 'partial builds'.
 
+<h4> Compact demo of several new major features</h4>
+
+The macro below demonstrates in a couple of lines a number of major new features in RooFit 3.00: Use of
+<ul>
+<li> workspace factory to quickly create and store (compiled) models
+<li> workspace CINT interface to easily access contents in a typesafe way
+<li> new adaptive ND numeric integration technique to normalize arbitrary p.d.f. in fast 
+     and reliable way
+<li> new adapative TFoam sampling technique to efficiently generate toy MC data from strongly
+     peaked datasets
+<li> parallel processing in likelihood construction and use of profile likelihood operator
+     to represent profile likelihoods as regular RooFit functions
+</ul>
+
+<pre>
+  void demo()  
+  {
+    <b>// Construct compiled 2-D model that requires numeric integration for normalization</b>
+    RooWorkspace w("w",1) ;
+    w.factory("CEXPR::model('1/((x-a)*(x-a)+0.001)+1/((y-b)*(y-b)+0.001)',x[-1,1],y[-1,1],a[-5,5],b[-5,5])") ;
+
+    <b>// Generate data from model (using TFoam adaptive sampling algorithm)</b>
+    RooDataSet* d = w::model.generate(RooArgSet(w::x,w::y),1000) ;
+    w::model.fitTo(*d) ; 
+
+    <b>// Make 2D plot on (x,y)</b>
+    TH2* hh = w::model.createHistogram("x,y",40,40) ;
+    hh->SetLineColor(kBlue) ;
+
+    <b>// Make Projection on x (integrate over y)</b>
+    RooPlot* framex = w::x.frame(Title("Data and p.d.f. projected on X")) ;
+    d->plotOn(framex) ;
+    w::model.plotOn(framex) ;
+
+    <b>// Construct likelihood, profile likelihood in a, and draw the latter</b>
+    RooAbsReal* nll = w::model.createNLL(*d,NumCPU(2)) ;
+    RooAbsReal* pll = nll->createProfile(w::a) ;
+    RooPlot* framea = w::a.frame(Title("Profile likelihood in parameter a")) ;
+    pll->plotOn(framea) ;
+  
+    <b>// Construct 2D cumulative distribution function from p.d.f.</b>
+    RooAbsReal* cdfxy = w::model.createCdf(RooArgSet(w::x,w::y),ScanNoCdf()) ;
+    TH2* hhcdf = cdfxy->createHistogram("x,y",40,40) ;
+    hhcdf->SetLineColor(kRed) ;
+
+    TCanvas* c = new TCanvas("c","c",650,650) ;  c->Divide(2,2) ;
+    c->cd(1) ; hh->Draw("surf") ;  c->cd(2) ; framex->Draw() ;
+    c->cd(3) ; framea->Draw()   ;  c->cd(4) ; hhcdf->Draw("surf") ;
+  }
+</pre>
+Plot that results from above macro
+<br><br>
+<img src="roofit_524_demo.gif" alt="Graphics Output of demo macro">
+<br><br>
+
 <h4>Miscellaneous small improvements</h4>
 <ul>
+<li>Utility functiom <tt>bindFunction()</tt> and <tt>bindPdf</tt> that can bind
+    external C++ functions as RooFit functions or p.d.f.s now can also take
+    <tt>ROOT::Math::Functor</tt> objects as input arguments, which allows the
+    binding of function methods of classes.<br><br>
+
+<li>By default datasets are no cloned for fit operations to save time and memory.
+    This change in procedure should save some time and memory, especially in toy MC
+    studies where the overhead in setting up the likelihood can dominate the total
+    time spent in fitting. The data cloning behavior of <tt>RooAbsPdf::fitTo()</tt>
+    and <tt>RooAbsPdf::createNLL()</tt> can be explicitly set through the <tt>CloneData()</tt> 
+    named argument<br><br>
+
 <li> It is now possible to construct a <tt>RooSimultaneous</tt> p.d.f. from other
-<tt>RooSimultaneous</tt> p.d.f.s, provided the constructor form is used that takes
-all input p.d.f.s. In this constructor simultaneous-of-simultaneous p.d.f.s are automatically
-recast to an equivalent top-level simultaneous p.d.f
-Sim of sim now possible<br><br>
+     <tt>RooSimultaneous</tt> p.d.f.s, provided the constructor form is used that takes
+     all input p.d.f.s. In this constructor simultaneous-of-simultaneous p.d.f.s are automatically
+     recast to an equivalent top-level simultaneous p.d.f
+     Sim of sim now possible<br><br>
 
 <li> Several improvements were made in the internal handling of datasets that
-     will speedup certain data intensive operations
+     will speedup certain data intensive operations<br><br>
 </ul>
 
+
+<br>
 <hr>
 <a name="roostats"></a>
 <h3>RooStats</h3>
@@ -324,27 +577,30 @@ Sim of sim now possible<br><br>
 
 Several new tutorials were added for RooStats
 <ul>
+<li><tt>rs101_limitexample.C</tt> Demonstrates use of Frequentist, 
+Bayesian, and Likelihood intervals for a simple number counting experiment
+ with uncertainty on signal and background rates.</li>
 <li><tt>rs301_splot.C</tt> Demonstrates use of RooStats sPlot
 implementation</li>
-<li><tt>rs401c_FeldmanCousins.C</tt>Demonstrates use of
+<li><tt>rs401c_FeldmanCousins.C</tt> Demonstrates use of
 FeldmanCousins interval calculator with a Poisson problem, reproduces
 resulst from table IV and V of the original
 paper Phys.Rev.D57:3873-3889,1998.</li>
-<li><tt>rs401d_FeldmanCousins.C</tt>Demonstrates use of
+<li><tt>rs401d_FeldmanCousins.C</tt> Demonstrates use of
 FeldmanCousins interval calculator with the neutrino oscillation toy
 example described in the original paper Phys.Rev.D57:3873-3889,1998.
 Reproduces figure 12.</li>
-<li><tt>rs_bernsteinCorrection.C</tt>Demonstrates use of
+<li><tt>rs_bernsteinCorrection.C</tt> Demonstrates use of
 BernsteinCorrection class, which corrects a nominal PDF with a polynomial
 to agree with observed or simulated data.</li>
 </ul>
 
 <h4>TestStatistic interface and implementations</h4>
-<p>We added a new interface class called TestStatistic. It defines the
-method Evaluate(data, parameterPoint), which returns a double.  This
-class can be used in conjunction with the ToyMCSampler class to generate
+<p>We added a new interface class called <tt>TestStatistic</tt>. It defines the
+method <tt>Evaluate(data, parameterPoint)</tt>, which returns a double.  This
+class can be used in conjunction with the <tt>ToyMCSampler</tt> class to generate
 sampling distributions for a user-defined test statistic.  </p>
-<p>The following concrete implementations of the TestStatistic interface
+<p>The following concrete implementations of the <tt>TestStatistic</tt> interface
 are currently available</p>
 <ul>
 <li><tt>ProfileLikelihoodTestStat</tt>Returns the log of profile
@@ -358,12 +614,12 @@ between 0,1.
 <h4>SamplingDistribution and the TestStatSampler interface and
 implementations</h4>
 <p>We introduced a ``result'' or data model class called
-SamplingDistribution, which holds the sampling distribution of an
+<tt>SamplingDistribution</tt>, which holds the sampling distribution of an
 arbitrary real valued test statistic.  The class also can return the
 inverse of the cumulative distribution function (with or without
 interpolation).  </p>
 <p>We introduced an interface for any tool that can produce a
-SamplingDistribution, called TestStatSampler.  The interface is
+<tt>SamplingDistribution</tt>, called <tt>TestStatSampler</tt>.  The interface is
 essentially GetSamplingDistribution(parameterPoint) which returns a
 SamplingDistribution based on a given probability density function.  We
 foresee a few versions of this tool based on toy Monte Carlo, importance
@@ -383,21 +639,60 @@ between 0,1.
 <p>A flexible framework for the Neyman Construction was added in this
 release. The NeymanConstruction is a concrete implementation of the
 IntervalCalculator interface, but it needs several
-additional components to be specified before use. The design
-factorizes the choice of the parameter points to be tested, the choice of
+additional components to be specified before use. The design
+factorizes the choice of the parameter points to be tested, the choice of
 the test statistic, and the generation of sampling distribution into
-separate parts (described above).  Finally, the NeymanConstruction class
+separate parts (described above).  Finally, the <tt>NeymanConstruction</tt> class
 is simply in charge of using these parts (strategies) and constructing
-the confidence belt and confidence intervals.  The ConfidenceBelt class
+the confidence belt and confidence intervals.  The <tt>ConfidenceBelt</tt> class
 is still under development, but the current version works fine for
-producing ConfidenceIntervals.  We are also working to make this class
+producing <tt>ConfidenceIntervals</tt>.  We are also working to make this class
 work with parallelization approaches, which is not yet complete.</p>
-<p>The FeldmanCousins class is a separate concrete implementation of the
-IntervalCalculator interface.  It uses the NeymanConstruction internally,
+<p>The <tt>FeldmanCousins</tt> class is a separate concrete implementation of the
+<tt>IntervalCalculator</tt> interface.  It uses the <tt>NeymanConstruction</tt> internally,
 and enforces specific choices of the test statistic and ordering
 principle to realize the Unified intervals described by Feldman and
 Cousins in their paper Phys.Rev.D57:3873-3889,1998.
 </p>
+<p>
+In an extension to the technique discussed in Feldman and Cousins paper, 
+the <tt>FeldmanCousins</tt> class also performs a "profile construction" if their are nuisance parameters.
+In this case, the parameters of interest are scanned in a regular grid.  For each point in the grid
+the calculator finds the best fit value of the nuisance parameters (given the data).  The construction
+is then only performed in this subspace of the parameters.  As a result, the number of points in the 
+construction only scales in the number of parameters of interest, not in the number of nuisance parameters.
+</p>
+
+<h4>Markov Chain Monte Carlo Interval</h4>
+<p>A flexible framework for Markov Chain Monte Carlo  was added in this
+release. The <tt>MCMCCalculator</tt> is a concrete implementation of the
+<tt>IntervalCalculator</tt> interface.  To use it one needs to specify the <tt>ProposalFunction</tt>. 
+There is a base class for ProposalFunctions and one concrete implementation: <tt>UniformProposal</tt>. 
+Support for other proposal functions will be added in the next release.
+The <tt>MCMCCalculator</tt> scans the space of the parameters of interest and nuisance parameters and
+produces a Bayesian posterior.  In this version, the prior must be added to the model initially,
+otherwise a flat prior is assumed.  The <tt>MCMCCalculator</tt> returns an <tt>MCMCInterval</tt>, which produces
+the smallest interval by taking a contour of the posterior.  This first version only supports 
+1,2, and 3 dimensional intervals, but will be generalized in the next release.
+</p>
+<p>
+In addition to the MCMC implementation in RooStats, one can export their model and dataset into a workspace,
+ and then use the Bayesian Analysis Toolkit (BAT) for the MCMC.  There is a wrapper available.
+</p>
+
+<h4>Redesigned SPlot class</h4>
+<p>The RooStats SPlot implementation works with any RooAbsPdf.  The class has been redesigned for more 
+convenient use.  It also adds some helper functions to check that the sum of sWeights over species is 1 for each event and 
+the sum over events for a given species equals the yield for that species.
+</p>
+
+<h4>Plotting classes</h4>
+<p>We have added new plotting classes: <tt>SamplingDistPlot</tt> and <tt>LikelihoodIntervalPlot</tt>.  
+In 1-d LikelihoodIntervalPlot shows the profile likelihood ratio and the upper/lower limits 
+of the interval for the parameter of interest.  In 2-d, the LikelihoodIntervalPlot shows
+the contour of the profile likelihood ratio for the parameters of interest.
+</p>
+
 
 
 <h4>Bernstein Correction</h4>
@@ -445,9 +740,7 @@ probability that the n+1-th
 </p>
 
 <h4>HybridCalculator</h4>
-<ul>
-<li> Add as a new  test statistics the profile likelihood ratio
-</li>
-</ul>
+<p> Add as a new test statistics the profile likelihood ratio.  Will be redesigned to use TestStatSampler and TestStatistic in next release.
+</p>