This macro shows several ways to perform a linear least-squares analysis . To keep things simple we fit a straight line to 4 data points The first 4 methods use the linear algebra package to find x such that min $ (A x - b)^T (A x - b) $ where A and b are calculated with the data points and the functional expression :
Expanding the expression $ (A x - b)^T (A x - b) $ and taking the derivative wrt x leads to the "Normal Equations": $ A^T A x = A^T b $ where $ A^T A $ is a positive definite matrix. Therefore, a Cholesky decomposition scheme can be used to calculate its inverse . This leads to the solution $ x = (A^T A)^-1 A^T b $ . All this is done in routine NormalEqn . We made it a bit more complicated by giving the data weights . Numerically this is not the best way to proceed because effectively the condition number of $ A^T A $ is twice as large as that of A, making inversion more difficult
One can show that solving $ A x = b $ for x with A of size $ (m x n) $ and $ m > n $ through a Singular Value Decomposition is equivalent to minimizing $ (A x - b)^T (A x - b) $ Numerically , this is the most stable method of all 5
Here we calculate the generalized matrix inverse ("pseudo inverse") by solving $ A X = Unit $ for matrix $ X $ through an SVD . The formal expression for is $ X = (A^T A)^-1 A^T $ . Then we multiply it by $ b $ . Numerically, not as good as 2 and not as fast . In general it is not a good idea to solve a set of linear equations with a matrix inversion .
The pseudo inverse is calculated brute force through a series of matrix manipulations . It shows nicely some operations in the matrix package, but is otherwise a big "no no" .
An objective function L is minimized by Minuit, where $ L = sum_i { (y - c_0 -c_1 * x / e)^2 } $ Minuit will calculate numerically the derivative of L wrt c_0 and c_1 . It has not been told that these derivatives are linear in the parameters c_0 and c_1 . For ill-conditioned linear problems it is better to use the fact it is a linear fit as in 2 .
Another interesting thing is the way we assign data to the vectors and matrices through adoption . This allows data assignment without physically moving bytes around .
This macro can be executed via CLING or via ACLIC
root > .x solveLinear.C
root > gSystem->Load("libMatrix");
root > gSystem->Load("libGpad");
root > .x solveLinear.C+
Author: Eddy Offermann
This notebook tutorial was automatically generated with ROOTBOOK-izer from the macro found in the ROOT repository on Friday, March 24, 2023 at 10:56 AM.
Arguments are defined.
Double_t eps = 1.e-12;
cout << "Perform the fit y = c0 + c1 * x in four different ways" << endl;
const Int_t nrVar = 2;
const Int_t nrPnts = 4;
Double_t ax[] = {0.0,1.0,2.0,3.0};
Double_t ay[] = {1.4,1.5,3.7,4.1};
Double_t ae[] = {0.5,0.2,1.0,0.5};
Perform the fit y = c0 + c1 * x in four different ways
Make the vectors 'Use" the data : they are not copied, the vector data pointer is just set appropriately
TVectorD x; x.Use(nrPnts,ax);
TVectorD y; y.Use(nrPnts,ay);
TVectorD e; e.Use(nrPnts,ae);
TMatrixD A(nrPnts,nrVar);
TMatrixDColumn(A,0) = 1.0;
TMatrixDColumn(A,1) = x;
cout << " - 1. solve through Normal Equations" << endl;
const TVectorD c_norm = NormalEqn(A,y,e);
cout << " - 2. solve through SVD" << endl;
- 1. solve through Normal Equations - 2. solve through SVD
numerically preferred method
first bring the weights in place
TMatrixD Aw = A;
TVectorD yw = y;
for (Int_t irow = 0; irow < A.GetNrows(); irow++) {
TMatrixDRow(Aw,irow) *= 1/e(irow);
yw(irow) /= e(irow);
}
TDecompSVD svd(Aw);
Bool_t ok;
const TVectorD c_svd = svd.Solve(yw,ok);
cout << " - 3. solve with pseudo inverse" << endl;
const TMatrixD pseudo1 = svd.Invert();
TVectorD c_pseudo1 = yw;
c_pseudo1 *= pseudo1;
cout << " - 4. solve with pseudo inverse, calculated brute force" << endl;
TMatrixDSym AtA(TMatrixDSym::kAtA,Aw);
const TMatrixD pseudo2 = AtA.Invert() * Aw.T();
TVectorD c_pseudo2 = yw;
c_pseudo2 *= pseudo2;
cout << " - 5. Minuit through TGraph" << endl;
TGraphErrors *gr = new TGraphErrors(nrPnts,ax,ay,0,ae);
TF1 *f1 = new TF1("f1","pol1",0,5);
gr->Fit("f1","Q");
TVectorD c_graph(nrVar);
c_graph(0) = f1->GetParameter(0);
c_graph(1) = f1->GetParameter(1);
- 3. solve with pseudo inverse - 4. solve with pseudo inverse, calculated brute force - 5. Minuit through TGraph
Check that all 4 answers are identical within a certain tolerance . The 1e-12 is somewhat arbitrary . It turns out that the TGraph fit is different by a few times 1e-13.
Bool_t same = kTRUE;
same &= VerifyVectorIdentity(c_norm,c_svd,0,eps);
same &= VerifyVectorIdentity(c_norm,c_pseudo1,0,eps);
same &= VerifyVectorIdentity(c_norm,c_pseudo2,0,eps);
same &= VerifyVectorIdentity(c_norm,c_graph,0,eps);
if (same)
cout << " All solutions are the same within tolerance of " << eps << endl;
else
cout << " Some solutions differ more than the allowed tolerance of " << eps << endl;
All solutions are the same within tolerance of 1e-12