Minuit-RooMinimizer interface which synchronizes parameter data and coordinates evaluation of likelihood (gradient) values.

This class provides an interface between RooFit and Minuit. It synchronizes parameter values from Minuit, calls calculator classes to evaluate likelihood and likelihood gradient values and returns them to Minuit. The Wrapper objects do the actual calculations. These are constructed inside the MinuitFcnGrad constructor using the RooAbsL likelihood passed in to the constructor, usually directly from RooMinimizer, with which this class is intimately coupled, being a RooAbsMinimizerFcn implementation. MinuitFcnGrad inherits from ROOT::Math::IMultiGradFunction as well, which allows it to be used as the FCN and GRAD parameters Minuit expects.

Note: The class is not intended for use by end-users. We recommend to either use RooMinimizer with a RooAbsL derived likelihood object, or to use a higher level entry point like RooAbsPdf::fitTo() or RooAbsPdf::createNLL().

Definition at line 31 of file MinuitFcnGrad.h.

Public Member Functions
	MinuitFcnGrad (const std::shared_ptr< RooFit::TestStatistics::RooAbsL > &absL, RooMinimizer *context, std::vector< ROOT::Fit::ParameterSettings > &parameters, LikelihoodMode likelihoodMode, LikelihoodGradientMode likelihoodGradientMode)

std::string	getFunctionName () const override
	RooMinimizer sometimes needs the name of the minimized function. Implement this in the derived class.

std::string	getFunctionTitle () const override
	RooMinimizer sometimes needs the title of the minimized function. Implement this in the derived class.

ROOT::Math::IMultiGenFunction *	getMultiGenFcn () override

void	Gradient (const double x, double grad) const
	IMultiGradFunction overrides necessary for Minuit.

void	GradientWithPrevResult (const double x, double grad, double previous_grad, double previous_g2, double *previous_gstep) const

double	operator() (const double *x) const

bool	returnsInMinuit2ParameterSpace () const

void	setOffsetting (bool flag) override
	Enable or disable offsetting on the function to be minimized, which enhances numerical precision.

void	setOptimizeConstOnFunction (RooAbsArg::ConstOpCode opcode, bool doAlsoTrackingOpt) override
	This function must be overridden in the derived class to pass on constant term optimization configuration to the function to be minimized.

bool	Synchronize (std::vector< ROOT::Fit::ParameterSettings > &parameter_settings) override
	Overridden from RooAbsMinimizerFcn to include gradient strategy synchronization.

Public Member Functions inherited from RooAbsMinimizerFcn
	RooAbsMinimizerFcn (RooArgList paramList, RooMinimizer *context)

virtual	~RooAbsMinimizerFcn ()=default

RooArgList const &	allParams () const

void	ApplyCovarianceMatrix (TMatrixDSym &V)
	Set different external covariance matrix.

void	BackProp ()
	Put Minuit results back into RooFit objects.

RooMinimizer::Config const &	cfg () const

RooArgList	constParams () const

Int_t	evalCounter () const

RooRealVar &	floatableParam (std::size_t i) const

RooArgList	floatParams () const

std::ofstream *	GetLogFile ()

double &	GetMaxFCN ()

unsigned int	getNDim () const

Int_t	GetNumInvalidNLL () const

double &	getOffset () const
	Return a possible offset that's applied to the function to separate invalid function values from valid ones.

RooArgList	initFloatParams () const

bool	SetLogFile (const char *inLogfile)
	Change the file name for logging of a RooMinimizer of all MINUIT steppings through the parameter space.

void	setOptimizeConst (Int_t flag)

bool	SetPdfParamVal (int index, double value) const
	Set value of parameter i.

bool	synchronizeParameterSettings (std::vector< ROOT::Fit::ParameterSettings > &parameters, bool optConst)
	Informs Minuit through its parameter_settings vector of RooFit parameter properties.

void	zeroEvalCount ()

Private Member Functions
template<class Func >
void	applyToLikelihood (Func &&func) const

void	syncOffsets () const
	Make sure the offsets are up to date.

bool	syncParameterValuesFromMinuitCalls (const double *x, bool minuit_internal) const
	Minuit calls (via FcnAdapters etc) DoEval or Gradient with a set of parameters x.

Private Attributes
bool	_calculatingGradient = false

std::shared_ptr< WrapperCalculationCleanFlags >	_calculationIsClean

std::unique_ptr< LikelihoodGradientWrapper >	_gradient

std::shared_ptr< LikelihoodWrapper >	_likelihood

std::shared_ptr< LikelihoodWrapper >	_likelihoodInGradient

std::vector< double >	_minuitExternalX

bool	_minuitInternalRooFitXMismatch = false

std::vector< double >	_minuitInternalX

std::unique_ptr< ROOT::Math::IMultiGradFunction >	_multiGenFcn

bool	offsets_reset_ = false

Additional Inherited Members
Protected Member Functions inherited from RooAbsMinimizerFcn
double	applyEvalErrorHandling (double fvalue) const
	Apply corrections on the fvalue if errors were signaled.

void	finishDoEval () const

void	optimizeConstantTerms (bool constStatChange, bool constValChange)

void	printEvalErrors () const
	Print information about why evaluation failed.

Static Protected Member Functions inherited from RooAbsMinimizerFcn
static bool	canBeFloating (RooAbsArg const &arg)

static bool	treatAsConstant (RooAbsArg const &arg)

Protected Attributes inherited from RooAbsMinimizerFcn
RooArgList	_allParams

RooArgList	_allParamsInit

RooMinimizer *	_context = nullptr

int	_evalCounter {0}

std::vector< std::size_t >	_floatableParamIndices

double	_funcOffset {0.}

std::ofstream *	_logfile = nullptr

double	_maxFCN = -std::numeric_limits<double>::infinity()

int	_numBadNLL = 0

bool	_optConst = false

#include </github/home/ROOT-CI/src/roofit/roofitcore/src/TestStatistics/MinuitFcnGrad.h>

Inheritance diagram for RooFit::TestStatistics::MinuitFcnGrad:

[legend]

Constructor & Destructor Documentation

◆ MinuitFcnGrad()

RooFit::TestStatistics::MinuitFcnGrad::MinuitFcnGrad	(	const std::shared_ptr< RooFit::TestStatistics::RooAbsL > &	absL,
		RooMinimizer *	context,
		std::vector< ROOT::Fit::ParameterSettings > &	parameters,
		LikelihoodMode	likelihoodMode,
		LikelihoodGradientMode	likelihoodGradientMode )

Parameters

[in]	absL	The input likelihood.
[in]	context	RooMinimizer that creates and owns this class.
[in]	parameters	The vector of ParameterSettings objects that describe the parameters used in the Minuit
[in]	likelihoodMode	Lmode
[in]	likelihoodGradientMode	Lgrad
[in]	verbose	true for verbose output Fitter. Note that these must match the set used in the Fitter used by `context!` It can be passed in from RooMinimizer with fitter()->Config().ParamsSettings().

Definition at line 83 of file MinuitFcnGrad.cxx.

Member Function Documentation

◆ applyToLikelihood()

template<class Func >

void RooFit::TestStatistics::MinuitFcnGrad::applyToLikelihood ( Func && func ) const

inlineprivate

Definition at line 73 of file MinuitFcnGrad.h.

◆ getFunctionName()

std::string RooFit::TestStatistics::MinuitFcnGrad::getFunctionName ( ) const

inlineoverridevirtual

RooMinimizer sometimes needs the name of the minimized function. Implement this in the derived class.

Implements RooAbsMinimizerFcn.

Definition at line 57 of file MinuitFcnGrad.h.

◆ getFunctionTitle()

std::string RooFit::TestStatistics::MinuitFcnGrad::getFunctionTitle ( ) const

inlineoverridevirtual

RooMinimizer sometimes needs the title of the minimized function. Implement this in the derived class.

Implements RooAbsMinimizerFcn.

Definition at line 59 of file MinuitFcnGrad.h.

◆ getMultiGenFcn()

ROOT::Math::IMultiGenFunction * RooFit::TestStatistics::MinuitFcnGrad::getMultiGenFcn ( )

inlineoverridevirtual

Implements RooAbsMinimizerFcn.

Definition at line 48 of file MinuitFcnGrad.h.

◆ Gradient()

void RooFit::TestStatistics::MinuitFcnGrad::Gradient	(	const double *	x,
		double *	grad ) const

IMultiGradFunction overrides necessary for Minuit.

Definition at line 247 of file MinuitFcnGrad.cxx.

◆ GradientWithPrevResult()

void RooFit::TestStatistics::MinuitFcnGrad::GradientWithPrevResult	(	const double *	x,
		double *	grad,
		double *	previous_grad,
		double *	previous_g2,
		double *	previous_gstep ) const

Definition at line 256 of file MinuitFcnGrad.cxx.

◆ operator()()

double RooFit::TestStatistics::MinuitFcnGrad::operator() ( const double * x ) const

Definition at line 135 of file MinuitFcnGrad.cxx.

◆ returnsInMinuit2ParameterSpace()

bool RooFit::TestStatistics::MinuitFcnGrad::returnsInMinuit2ParameterSpace ( ) const

inline

Definition at line 41 of file MinuitFcnGrad.h.

◆ setOffsetting()

void RooFit::TestStatistics::MinuitFcnGrad::setOffsetting ( bool flag )

inlineoverridevirtual

Enable or disable offsetting on the function to be minimized, which enhances numerical precision.

Implements RooAbsMinimizerFcn.

Definition at line 61 of file MinuitFcnGrad.h.

◆ setOptimizeConstOnFunction()

void RooFit::TestStatistics::MinuitFcnGrad::setOptimizeConstOnFunction	(	RooAbsArg::ConstOpCode	opcode,
		bool	doAlsoTrackingOpt )

inlineoverridevirtual

This function must be overridden in the derived class to pass on constant term optimization configuration to the function to be minimized.

For a RooAbsArg, this would be RooAbsArg::constOptimizeTestStatistic.

Implements RooAbsMinimizerFcn.

Definition at line 43 of file MinuitFcnGrad.h.

◆ Synchronize()

bool RooFit::TestStatistics::MinuitFcnGrad::Synchronize ( std::vector< ROOT::Fit::ParameterSettings > & parameter_settings )

overridevirtual

Overridden from RooAbsMinimizerFcn to include gradient strategy synchronization.

Reimplemented from RooAbsMinimizerFcn.

Definition at line 266 of file MinuitFcnGrad.cxx.

◆ syncOffsets()

void RooFit::TestStatistics::MinuitFcnGrad::syncOffsets ( ) const

private

Make sure the offsets are up to date.

If the offsets need to be updated, this function triggers a likelihood evaluation. The likelihood will make sure the offset is set correctly in their shared_ptr offsets object, that is also shared with possible other LikelihoodWrapper members of MinuitFcnGrad and also the LikelihoodGradientWrapper member. Other necessary synchronization steps are also performed from the Wrapper child classes (e.g. sending the values to workers from MultiProcess::Jobs).

Definition at line 127 of file MinuitFcnGrad.cxx.

◆ syncParameterValuesFromMinuitCalls()

bool RooFit::TestStatistics::MinuitFcnGrad::syncParameterValuesFromMinuitCalls	(	const double *	x,
		bool	minuit_internal ) const

private

Minuit calls (via FcnAdapters etc) DoEval or Gradient with a set of parameters x.

This function syncs these values to the proper places in RooFit.

The first twist, and reason this function is more complicated than one may imagine, is that Minuit internally uses a transformed parameter space to account for parameter boundaries. Whether we receive these Minuit "internal" parameter values or "regular"/untransformed RooFit parameter space values depends on the situation.

The values that arrive here via DoEval are always "normal" parameter values, since Minuit transforms these back into regular space before passing to DoEval (see MnUserFcn::operator() which wraps the Fcn(Gradient)Base in ModularFunctionMinimizer::Minimize and is used for direct function calls from that point on in the minimizer). These can thus always be safely synced with this function's RooFit parameters using SetPdfParamVal.
The values that arrive here via Gradient will be in internal coordinates if that is what this class expects, and indeed this is the case for MinuitFcnGrad's current implementation. This is communicated to Minuit via MinuitFcnGrad::returnsInMinuit2ParameterSpace. Inside Minuit, that function determines whether this class's gradient calculator is wrapped inside a AnalyticalGradientCalculator, to which Minuit passes "external" parameter values, or as an ExternalInternalGradientCalculator, which gets "internal" parameter values. Long story short: when MinuitFcnGrad::returnsInMinuit2ParameterSpace() returns true, Minuit will pass "internal" values to Gradient. These cannot be synced with this function's RooFit parameters using SetPdfParamVal, unless a manual transformation step is performed in advance. However, they do need to be passed on to the gradient calculator, since indeed we expect values there to be in "internal" space. However, this is calculator dependent. Note that in the current MinuitFcnGrad implementation we do not actually allow for calculators in "external" (i.e. regular RooFit parameter space) values, since MinuitFcnGrad::returnsInMinuit2ParameterSpace is hardcoded to true. This should in a future version be changed so that the calculator (the wrapper) is queried for this information. Because some gradient calculators may also use the regular RooFit parameters (e.g. for calculating the likelihood's value itself), this information is also passed on to the gradient wrapper. Vice versa, when updated "internal" parameters are passed to Gradient, the likelihood may be affected as well. Even though a transformation from internal to "external" may be necessary before the values can be used, the likelihood can at least log that its parameter values are possibly no longer in sync with those of the gradient.

bool RooFit::TestStatistics::MinuitFcnGrad::offsets_reset_ = false

mutableprivate

Definition at line 94 of file MinuitFcnGrad.h.

roofit/roofitcore/src/TestStatistics/MinuitFcnGrad.h

roofit/roofitcore/src/TestStatistics/MinuitFcnGrad.cxx

Public Member Functions

Private Member Functions

Private Attributes

Additional Inherited Members

Constructor & Destructor Documentation

◆ MinuitFcnGrad()

Member Function Documentation

◆ applyToLikelihood()

◆ getFunctionName()

◆ getFunctionTitle()

◆ getMultiGenFcn()

◆ Gradient()

◆ GradientWithPrevResult()

◆ operator()()

◆ returnsInMinuit2ParameterSpace()

◆ setOffsetting()

◆ setOptimizeConstOnFunction()

◆ Synchronize()

◆ syncOffsets()

◆ syncParameterValuesFromMinuitCalls()

Member Data Documentation

◆ _calculatingGradient

◆ _calculationIsClean

◆ _gradient

◆ _likelihood

◆ _likelihoodInGradient

◆ _minuitExternalX

◆ _minuitInternalRooFitXMismatch

◆ _minuitInternalX

◆ _multiGenFcn

◆ offsets_reset_