Root Minimizer Configuration

The RootMinimizer uses ROOT to apply numeric minimization to the likelihood and find the best fit point.

Config options

These are the minimizerInterfaceConfig fields that are supported by when the minimizer type is set to RootMinimizer.

minimizerConfig Options	Type	Description	Default
type	string	Must be “RootMinimizer” for this configuration
checkParameterValidity	bool	Turn on parameter validity checks (in physical, in domain, etc)	false
checkFlatDirectionsBeforeMinimize	bool	Diagnose local flat LLH directions before minimization (including dry-runs), without blocking the fit. Error calculation always runs this check.	false
showParametersOnFitMonitor	bool	Display fit parameter parameter values on the monitor	false
enablePostFitErrorFit	bool	Evaluate the Hessian after the fit	true
maxNbParametersPerLineOnMonitor	int	Number of parameters on a monitor line	15
useNormalizedFitSpace	bool	use fit parameter interface to provide prior mean at 0 and stddev at 1	true
writeLlhHistory	bool	write a ttree registering all the llh evaluations	false
Root Minimizer Specific	Type	Description	Default
minimizer	string	engine name	Minuit2
algorithm	string	algorithm name	default from engine
errors	string	algorithm to run after the fit (HESSE or MINOS)	Hesse
tolerance	double	defines the required Estimated Distance from the Minimum stopping the fit	1E-4
functionPrecision	double	Relative numerical precision of the objective; must be finite and positive	ROOT (double)
maxFcnCalls	int	maximum number of function calls before stopping fit	1E9
maxIterations	int	maximum number of minimizer iterations before stopping fit	500
strategy	int	fitter strategy (sec. 1.3)	1
print_level	int	minimizer verbose level (p.23)	1
enableSimplexBeforeMinimize	bool	run SIMPLEX before the real fit (can help to find the minimum)	false
enablePostFitErrorFit	bool	Evaluate the Hessian after the fit	true
simplexMaxFcnCalls	int	stop SIMPLEX after N calls	1000
simplexToleranceLoose	int	loosen the minimizer tolerance by this factor during SIMPLEX	1000
simplexStrategy	int	strategy for SIMPLEX	1
generatedPostFitParBreakdown	bool	Generate figures showing hessian eigen decomp breakdown by parameter	false
generatedPostFitEigenBreakdown	bool	Generate figures showing parameter breakdown by hessian eigen	false
monitorRefreshRateInMs	int	Max refresh rate for the fit monitor in milliseconds	5000
monitorGradientDescent	bool
tolerancePerDegreeOfFreedome	int

Objective function precision

functionPrecision is passed to the numerical minimizer through SetPrecision(). It describes the precision of likelihood evaluations and affects numerical derivatives, including Hesse. It is distinct from tolerance, which controls convergence through the estimated distance from the minimum.

When omitted, GUNDAM uses the active backend’s recommendation if available. MPS recommends float epsilon (1.1920928955078125e-7) because its event weights and histogram outputs use float. CPU and reference propagation leave ROOT’s precision default unchanged. An explicit functionPrecision takes precedence. The selected recommendation or override is printed during initialization; the minimizer precision is also saved as fit/minimizer/functionPrecision_TNamed under the fitter output directory.

A value smaller than the actual numerical noise can produce unreliable Hesse errors. Increasing it allows larger differentiation steps, but does not restore double precision to GPU calculations. Check uncertainty stability against CPU and variations of this setting for the analysis being fitted. See the MPS Hesse investigation for measured limits.

SIMPLEX prefit

enableSimplexBeforeMinimize creates a separate Minuit2 Simplex minimizer before the configured main algorithm. It inherits the main minimizer’s options, including objective function precision, and uses simplexStrategy, simplexMaxFcnCalls and tolerance * simplexToleranceLoose.

The prefit copies all variable definitions, fixed states and limits. Its best parameter values become the main minimizer’s starting point. The main algorithm retains its configured step sizes, precision, strategy and tolerance; Simplex’s error estimates are not used as a Migrad covariance. The original minimizer is also restored if the prefit throws an exception.