No label defined (Q3976294)

[For the entire collection see Zbl 0732.00019.] The solution of the following equation for a \(p\times p\) matrix \(A: (1/n)\sum^ n_{i=1} u(| Ax_ i|)Ax_ i(Ax_ i)^ T=I_ p\) is considered. Here \(x_ i=(x_{i1},x_{i2},\ldots,x_{ip})^ T\) are \(p\)-variate observation vectors, \(u\) is a given weight function, \(I_ p\) is the \(p\times p\) identity matrix. The problem arises within the affinely invariant approach for the robust estimation of a covariance matrix. The known algorithms for solving this problem include Newton-like and conjugate gradient procedures. But these algorithms do not work properly if the columns of the matrix \(Y=\hbox{diag}(\sqrt{u(| Ax_ i| ))}X\) become nearly collinear during the iterations. A simple modification of the fixed point algorithm is presented which determines the pseudorank \(k\leq p\) of \(Y\) by means of the QR decomposition with column pivoting. A numerical example is given.