Re: [Quantlib-dev] Longstaff-Schwartz method, SVD and OLS

[Kakhkhor A. <kab...@gm...>]

Hi. Similar problem in SVD::solveFor. It uses all singular values to
compute matrix pseudo-inverse. Insignificant singular values should be
dropped. Quick fix:


Disposable<Array> SVD::solveFor(const Array& b) const{
    Matrix W(n_, n_, 0.0);
    Size numericalRank = this->rank();
    for (Size i=0; i<numericalRank; i++)
        W[i][i] = 1./s_[i];

    Matrix inverse = V()* W * transpose(U());
    Array result = inverse * b;
    return result;
}



Also, in SVD::rank  implementation we could replace
Real eps = std::pow(2.0,-52.0);
with
Real eps = QL_EPSILON.


Regards,
Kakhkhor Abdijalilov.