protected:
virtual Double_t EvalControl(const Int_t* powers) virtual Double_t EvalFactor(Int_t p, Double_t x) virtual void MakeCandidates() virtual void MakeCoefficientErrors() virtual void MakeCoefficients() virtual void MakeCorrelation() virtual Double_t MakeGramSchmidt(Int_t function) virtual void MakeNormalized() virtual void MakeParameterization() virtual void MakeRealCode(const char* filename, const char* classname, Option_t* option) virtual Bool_t Select(const Int_t* iv) virtual Bool_t TestFunction(Double_t squareResidual, Double_t dResidur) public:
TMultiDimFit TMultiDimFit() TMultiDimFit TMultiDimFit(Int_t dimension, TMultiDimFit::EMDFPolyType type = kMonomials, Option_t* option) TMultiDimFit TMultiDimFit(TMultiDimFit&) virtual void ~TMultiDimFit() virtual void AddRow(const Double_t* x, Double_t D, Double_t E = 0) virtual void AddTestRow(const Double_t* x, Double_t D, Double_t E = 0) virtual void Browse(TBrowser* b) static TClass* Class() virtual void Clear(Option_t* option) virtual void Draw(Option_t* option = "d") virtual Double_t Eval(const Double_t* x, const Double_t* coeff = 0) virtual void FindParameterization(Option_t* option) virtual void Fit(Option_t* option) Double_t GetChi2() const Double_t GetError() const Int_t* GetFunctionCodes() const const TMatrixD* GetFunctions() const virtual TList* GetHistograms() const Double_t GetMaxAngle() const Int_t GetMaxFunctions() const Int_t* GetMaxPowers() const Double_t GetMaxQuantity() const Int_t GetMaxStudy() const Int_t GetMaxTerms() const const TVectorD* GetMaxVariables() const Double_t GetMeanQuantity() const const TVectorD* GetMeanVariables() const Double_t GetMinAngle() const Double_t GetMinQuantity() const Double_t GetMinRelativeError() const const TVectorD* GetMinVariables() const Int_t GetNCoefficients() const Int_t GetNVariables() const Int_t* GetPowerIndex() const Double_t GetPowerLimit() const const Int_t* GetPowers() const Double_t GetPrecision() const const TVectorD* GetQuantity() const Double_t GetResidualMax() const Int_t GetResidualMaxRow() const Double_t GetResidualMin() const Int_t GetResidualMinRow() const Double_t GetResidualSumSq() const Double_t GetRMS() const Int_t GetSampleSize() const const TVectorD* GetSqError() const Double_t GetSumSqAvgQuantity() const Double_t GetSumSqQuantity() const Double_t GetTestError() const Double_t GetTestPrecision() const const TVectorD* GetTestQuantity() const Int_t GetTestSampleSize() const const TVectorD* GetTestSqError() const const TVectorD* GetTestVariables() const const TVectorD* GetVariables() const static TMultiDimFit* Instance() virtual TClass* IsA() const virtual Bool_t IsFolder() const virtual Double_t MakeChi2(const Double_t* coeff = 0) virtual void MakeCode(const char* functionName = "MDF", Option_t* option) virtual void MakeHistograms(Option_t* option = "A") virtual void MakeMethod(const Char_t* className = "MDF", Option_t* option) virtual void Print(Option_t* option = "ps") const void SetMaxAngle(Double_t angle = 0) void SetMaxFunctions(Int_t n) void SetMaxPowers(const Int_t* powers) void SetMaxStudy(Int_t n) void SetMaxTerms(Int_t terms) void SetMinAngle(Double_t angle = 1) void SetMinRelativeError(Double_t error) void SetPowerLimit(Double_t limit = 1e-3) virtual void SetPowers(const Int_t* powers, Int_t terms) virtual void ShowMembers(TMemberInspector& insp, char* parent) virtual void Streamer(TBuffer& b) void StreamerNVirtual(TBuffer& b)
private:
static TMultiDimFit* fgInstance Static instance protected:
TVectorD fQuantity Training sample, dependent quantity TVectorD fSqError Training sample, error in quantity Double_t fMeanQuantity Mean of dependent quantity Double_t fMaxQuantity Max value of dependent quantity Double_t fMinQuantity Min value of dependent quantity Double_t fSumSqQuantity SumSquare of dependent quantity Double_t fSumSqAvgQuantity Sum of squares away from mean TVectorD fVariables Training sample, independent variables Int_t fNVariables Number of independent variables TVectorD fMeanVariables mean value of independent variables TVectorD fMaxVariables max value of independent variables TVectorD fMinVariables min value of independent variables Int_t fSampleSize Size of training sample TVectorD fTestQuantity Test sample, dependent quantity TVectorD fTestSqError Test sample, Error in quantity TVectorD fTestVariables Test sample, independent variables Int_t fTestSampleSize Size of test sample Double_t fMinAngle Min angle for acepting new function Double_t fMaxAngle Max angle for acepting new function Int_t fMaxTerms Max terms expected in final expr. Double_t fMinRelativeError Min relative error accepted Int_t* fMaxPowers [fNVariables] maximum powers Double_t fPowerLimit Control parameter TMatrixD fFunctions Functions evaluated over sample Int_t fMaxFunctions max number of functions Int_t* fFunctionCodes [fMaxFunctions] acceptance code Int_t fMaxStudy max functions to study TMatrixD fOrthFunctions As above, but orthogonalised TVectorD fOrthFunctionNorms Norm of the evaluated functions Int_t* fMaxPowersFinal [fNVariables] maximum powers from fit; Int_t* fPowers [fMaxFunctions*fNVariables] Int_t* fPowerIndex [fMaxTerms] Index of accepted powers TVectorD fResiduals Vector of the final residuals Double_t fMaxResidual Max redsidual value Double_t fMinResidual Min redsidual value Int_t fMaxResidualRow Row giving max residual Int_t fMinResidualRow Row giving min residual Double_t fSumSqResidual Sum of Square residuals Int_t fNCoefficients Dimension of model coefficients TVectorD fOrthCoefficients The model coefficients TMatrixD fOrthCurvatureMatrix Model matrix TVectorD fCoefficients Vector of the final coefficients TVectorD fCoefficientsRMS Vector of RMS of coefficients Double_t fRMS Root mean square of fit Double_t fChi2 Chi square of fit Int_t fParameterisationCode Exit code of parameterisation Double_t fError Error from parameterization Double_t fTestError Error from test Double_t fPrecision Relative precision of param Double_t fTestPrecision Relative precision of test Double_t fCorrelationCoeff Multi Correlation coefficient TMatrixD fCorrelationMatrix Correlation matrix Double_t fTestCorrelationCoeff Multi Correlation coefficient TList* fHistograms List of histograms Byte_t fHistogramMask Bit pattern of hisograms used TVirtualFitter* fFitter ! Fit object (MINUIT) TMultiDimFit::EMDFPolyType fPolyType Type of polynomials to use Bool_t fShowCorrelation print correlation matrix Bool_t fIsUserFunction Flag for user defined function Bool_t fIsVerbose public:
static const TMultiDimFit::EMDFPolyType kMonomials static const TMultiDimFit::EMDFPolyType kChebyshev static const TMultiDimFit::EMDFPolyType kLegendre
/*
A common problem encountered in different fields of applied science is to find an expression for one physical quantity in terms of several others, which are directly measurable.
An example in high energy physics is the evaluation of the momentum of a charged particle from the observation of its trajectory in a magnetic field. The problem is to relate the momentum of the particle to the observations, which may consists of of positional measurements at intervals along the particle trajectory.
The exact functional relationship between the measured quantities (e.g., the space-points) and the dependent quantity (e.g., the momentum) is in general not known, but one possible way of solving the problem, is to find an expression which reliably approximates the dependence of the momentum on the observations.
This explicit function of the observations can be obtained by a least squares fitting procedure applied to a representive sample of the data, for which the dependent quantity (e.g., momentum) and the independent observations are known. The function can then be used to compute the quantity of interest for new observations of the independent variables.
This class TMultiDimFit implements such a procedure in ROOT. It is largely based on the CERNLIB MUDIFI package [2]. Though the basic concepts are still sound, and therefore kept, a few implementation details have changed, and this class can take advantage of MINUIT [4] to improve the errors of the fitting, thanks to the class TMinuit.
In [5] and [6] H. Wind demonstrates the utility of this procedure in the context of tracking, magnetic field parameterisation, and so on. The outline of the method used in this class is based on Winds discussion, and I refer these two excellents text for more information.
And example of usage is given in $ROOTSYS/tutorials/multidimfit.C.
Let by the dependent quantity of interest, which depends smoothly on the observable quantities , which we'll denote by . Given a training sample of tuples of the form, (TMultiDimFit::AddRow)
So what TMultiDimFit does, is to determine the number of terms , and then terms (or functions) , and the coefficients , so that is minimal (TMultiDimFit::FindParameterization).
Of course it's more than a little unlikely that will ever become exact zero as a result of the procedure outlined below. Therefore, the user is asked to provide a minimum relative error (TMultiDimFit::SetMinRelativeError), and will be considered minimized when
Optionally, the user may impose a functional expression by specifying the powers of each variable in specified functions (TMultiDimFit::SetPowers). In that case, only the coefficients is calculated by the class.
As always when dealing with fits, there's a real chance of over fitting. As is well-known, it's always possible to fit an polynomial in to points with , but the polynomial is not likely to fit new data at all [1]. Therefore, the user is asked to provide an upper limit, to the number of terms in (TMultiDimFit::SetMaxTerms).
However, since there's an infinite number of to choose from, the user is asked to give the maximum power. , of each variable to be considered in the minimization of (TMultiDimFit::SetMaxPowers).
One way of obtaining values for the maximum power in variable , is to perform a regular fit to the dependent quantity , using a polynomial only in . The maximum power is is then the power that does not significantly improve the one-dimensional least-square fit over to [5].
There are still a huge amount of possible choices for ; in fact there are possible choices. Obviously we need to limit this. To this end, the user is asked to set a power control limit, (TMultiDimFit::SetPowerLimit), and a function is only accepted if
To further reduce the number of functions in the final expression, only those functions that significantly reduce is chosen. What `significant' means, is chosen by the user, and will be discussed below (see 2.3).
The functions are generally not orthogonal, which means one will have to evaluate all possible 's over all data-points before finding the most significant [1]. We can, however, do better then that. By applying the modified Gram-Schmidt orthogonalisation algorithm [5] [3] to the functions , we can evaluate the contribution to the reduction of from each function in turn, and we may delay the actual inversion of the curvature-matrix (TMultiDimFit::MakeGramSchmidt).
So we are let to consider an matrix , an element of which is given by
We now take as a new model . We thus want to minimize
So for each new function included in the model, we get a reduction of the sum of squares of residuals of , where is given by (4) and by (9). Thus, using the Gram-Schmidt orthogonalisation, we can decide if we want to include this function in the final model, before the matrix inversion.
Supposing that steps of the procedure have been performed, the problem now is to consider the function.
The sum of squares of residuals can be written as
Two test are now applied to decide whether this function is to be included in the final expression, or not.
Denoting by the subspace spanned by the function is by construction (see (4)) the projection of the function onto the direction perpendicular to . Now, if the length of (given by ) is very small compared to the length of this new function can not contribute much to the reduction of the sum of squares of residuals. The test consists then in calculating the angle between the two vectors and (see also figure 1) and requiring that it's greater then a threshold value which the user must set (TMultiDimFit::SetMinAngle).
Let be the data vector to be fitted. As illustrated in figure 1, the function will contribute significantly to the reduction of , if the angle between and is smaller than an upper limit , defined by the user (TMultiDimFit::SetMaxAngle)
However, the method automatically readjusts the value of this angle while fitting is in progress, in order to make the selection criteria less and less difficult to be fulfilled. The result is that the functions contributing most to the reduction of are chosen first (TMultiDimFit::TestFunction).
In case isn't defined, an alternative method of performing this second test is used: The function is accepted if (refer also to equation (13))
From this we see, that by restricting -- the number of terms in the final model -- the fit is more difficult to perform, since the above selection criteria is more limiting.
The more coefficients we evaluate, the more the sum of squares of residuals will be reduced. We can evaluate before inverting as shown below.
Having found a parameterization, that is the 's and , that minimizes , we still need to determine the coefficients . However, it's a feature of how we choose the significant functions, that the evaluation of the 's becomes trivial [5]. To derive , we first note that equation (4) can be written as
It's important to realize that the training sample should be representive of the problem at hand, in particular along the borders of the region of interest. This is because the algorithm presented here, is a interpolation, rahter then a extrapolation [5].
Also, the independent variables need to be linear independent, since the procedure will perform poorly if they are not [5]. One can find an linear transformation from ones original variables to a set of linear independent variables , using a Principal Components Analysis (see TPrincipal), and then use the transformed variable as input to this class [5] [6].
H. Wind also outlines a method for parameterising a multidimensional dependence over a multidimensional set of variables. An example of the method from [5], is a follows (please refer to [5] for a full discussion):
To process data, using this parameterisation, do
The class also provides functionality for testing the, over the training sample, found parameterization (TMultiDimFit::Fit). This is done by passing the class a test sample of tuples of the form , where are the independent variables, the known, dependent quantity, and is the square error in (TMultiDimFit::AddTestRow).
The parameterization is then evaluated at every in the test sample, and
It's possible to use Minuit [4] to further improve the fit, using the test sample.
*/
Empty CTOR. Do not use
Constructor Second argument is the type of polynomials to use in parameterisation, one of: TMultiDimFit::kMonomials TMultiDimFit::kChebyshev TMultiDimFit::kLegendre Options: K Compute (k)correlation matrix V Be verbose Default is no options.
DTOR
Add a row consisting of fNVariables independent variables, the known, dependent quantity, and optionally, the square error in the dependent quantity, to the training sample to be used for the parameterization. The mean of the variables and quantity is calculated on the fly, as outlined in TPrincipal::AddRow. This sample should be representive of the problem at hand. Please note, that if no error is given Poisson statistics is assumed and the square error is set to the value of dependent quantity. See also the class description
Add a row consisting of fNVariables independent variables, the known, dependent quantity, and optionally, the square error in the dependent quantity, to the test sample to be used for the test of the parameterization. This sample needn't be representive of the problem at hand. Please note, that if no error is given Poisson statistics is assumed and the square error is set to the value of dependent quantity. See also the class description
Browse the TMultiDimFit object in the TBrowser.
Clear internal structures and variables
Evaluate parameterization at point x. Optional argument coeff is a vector of coefficients for the parameterisation, fNCoefficients elements long.
PRIVATE METHOD: Calculate the control parameter from the passed powers
PRIVATE METHOD: Evaluate function with power p at variable value x
Find the parameterization Options: None so far For detailed description of what this entails, please refer to the class description
Try to fit the found parameterisation to the test sample. Options M use Minuit to improve coefficients Also, refer to class description
PRIVATE METHOD: Create list of candidate functions for the parameterisation. See also class description
Calculate Chi square over either the test sample. The optional argument coeff is a vector of coefficients to use in the evaluation of the parameterisation. If coeff == 0, then the found coefficients is used. Used my MINUIT for fit (see TMultDimFit::Fit)
Generate the file <filename> with .C appended if argument doesn't end in .cxx or .C. The contains the implementation of the function: Double_t <funcname>(Double_t *x) which does the same as TMultiDimFit::Eval. Please refer to this method. Further, the static variables: Int_t gNVariables Int_t gNCoefficients Double_t gDMean Double_t gXMean[] Double_t gXMin[] Double_t gXMax[] Double_t gCoefficient[] Int_t gPower[] are initialized. The only ROOT header file needed is Rtypes.h See TMultiDimFit::MakeRealCode for a list of options
PRIVATE METHOD: Compute the errors on the coefficients. For this to be done, the curvature matrix of the non-orthogonal functions, is computed.
PRIVATE METHOD: Invert the model matrix B, and compute final coefficients. For a more thorough discussion of what this means, please refer to the class description First we invert the lower triangle matrix fOrthCurvatureMatrix and store the inverted matrix in the upper triangle.
PRIVATE METHOD: Compute the correlation matrix
PRIVATE METHOD: Make Gram-Schmidt orthogonalisation. The class description gives a thorough account of this algorithm, as well as references. Please refer to the class description
Make histograms of the result of the analysis. This message should be sent after having read all data points, but before finding the parameterization Options: A All the below X Original independent variables D Original dependent variables N Normalised independent variables S Shifted dependent variables R1 Residuals versus normalised independent variables R2 Residuals versus dependent variable R3 Residuals computed on training sample R4 Residuals computed on test sample For a description of these quantities, refer to class description
Generate the file <classname>MDF.cxx which contains the implementation of the method: Double_t <classname>::MDF(Double_t *x) which does the same as TMultiDimFit::Eval. Please refer to this method. Further, the public static members: Int_t <classname>::fgNVariables Int_t <classname>::fgNCoefficients Double_t <classname>::fgDMean Double_t <classname>::fgXMean[] //[fgNVariables] Double_t <classname>::fgXMin[] //[fgNVariables] Double_t <classname>::fgXMax[] //[fgNVariables] Double_t <classname>::fgCoefficient[] //[fgNCoeffficents] Int_t <classname>::fgPower[] //[fgNCoeffficents*fgNVariables] are initialized, and assumed to exist. The class declaration is assumed to be in <classname>.h and assumed to be provided by the user. See TMultiDimFit::MakeRealCode for a list of options The minimal class definition is: class <classname> { public: Int_t <classname>::fgNVariables; // Number of variables Int_t <classname>::fgNCoefficients; // Number of terms Double_t <classname>::fgDMean; // Mean from training sample Double_t <classname>::fgXMean[]; // Mean from training sample Double_t <classname>::fgXMin[]; // Min from training sample Double_t <classname>::fgXMax[]; // Max from training sample Double_t <classname>::fgCoefficient[]; // Coefficients Int_t <classname>::fgPower[]; // Function powers Double_t Eval(Double_t *x); }; Whether the method <classname>::Eval should be static or not, is up to the user.
PRIVATE METHOD: Normalize data to the interval [-1;1]. This is needed for the classes method to work.
PRIVATE METHOD: Find the parameterization over the training sample. A full account of the algorithm is given in the class description
PRIVATE METHOD: This is the method that actually generates the code for the evaluation the parameterization on some point. It's called by TMultiDimFit::MakeCode and TMultiDimFit::MakeMethod. The options are: NONE so far
Print statistics etc. Options are P Parameters S Statistics C Coefficients R Result of parameterisation F Result of fit
Selection method. User can override this method for specialized selection of acceptable functions in fit. Default is to select all. This message is sent during the build-up of the function candidates table once for each set of powers in variables. Notice, that the argument array contains the powers PLUS ONE. For example, to De select the function f = x1^2 * x2^4 * x3^5, this method should return kFALSE if given the argument { 3, 4, 6 }
Set the max angle (in degrees) between the initial data vector to be fitted, and the new candidate function to be included in the fit. By default it is 0, which automatically chooses another selection criteria. See also class description
Set the min angle (in degrees) between a new candidate function and the subspace spanned by the previously accepted functions. See also class description
Define a user function. The input array must be of the form (p11, ..., p1N, ... ,pL1, ..., pLN) Where N is the dimension of the data sample, L is the number of terms (given in terms) and the first number, labels the term, the second the variable. More information is given in the class description
Set the user parameter for the function selection. The bigger the limit, the more functions are used. The meaning of this variable is defined in the class description
Set the maximum power to be considered in the fit for each variable. See also class description
Set the acceptable relative error for when sum of square residuals is considered minimized. For a full account, refer to the class description
PRIVATE METHOD: Test whether the currently considered function contributes to the fit. See also class description
void Draw(Option_t* option = "d") Double_t GetChi2() const Double_t GetError() const Int_t* GetFunctionCodes() const const TMatrixD* GetFunctions() const TList* GetHistograms() const Double_t GetMaxAngle() const Int_t GetMaxFunctions() const Int_t* GetMaxPowers() const Double_t GetMaxQuantity() const Int_t GetMaxStudy() const Int_t GetMaxTerms() const const TVectorD* GetMaxVariables() const Double_t GetMeanQuantity() const const TVectorD* GetMeanVariables() const Double_t GetMinAngle() const Double_t GetMinQuantity() const Double_t GetMinRelativeError() const const TVectorD* GetMinVariables() const Int_t GetNVariables() const Int_t GetNCoefficients() const Int_t* GetPowerIndex() const Double_t GetPowerLimit() const const Int_t* GetPowers() const Double_t GetPrecision() const const TVectorD* GetQuantity() const Double_t GetResidualMax() const Double_t GetResidualMin() const Int_t GetResidualMaxRow() const Int_t GetResidualMinRow() const Double_t GetResidualSumSq() const Double_t GetRMS() const Int_t GetSampleSize() const const TVectorD* GetSqError() const Double_t GetSumSqAvgQuantity() const Double_t GetSumSqQuantity() const Double_t GetTestError() const Double_t GetTestPrecision() const const TVectorD* GetTestQuantity() const Int_t GetTestSampleSize() const const TVectorD* GetTestSqError() const const TVectorD* GetTestVariables() const const TVectorD* GetVariables() const TMultiDimFit* Instance() Bool_t IsFolder() const void SetMaxFunctions(Int_t n) void SetMaxStudy(Int_t n) void SetMaxTerms(Int_t terms) TClass* Class() TClass* IsA() const void ShowMembers(TMemberInspector& insp, char* parent) void Streamer(TBuffer& b) void StreamerNVirtual(TBuffer& b) TMultiDimFit TMultiDimFit(TMultiDimFit&)