A new conic method for unconstrained minimization (Q1197422): Difference between revisions

A conic function has the form \[ c(x)={1\over 2}{x^ T Qx\over (1+p^ T x)^ 2}+{b^ T x\over 1+p^ T x}+a, \] where \(Q\) is the \(n\times n\) symmetric conjugacy matrix and \(p\in R^ n\) is the gauge vector defining the horizon of \(c(x)\). This is related to the quadratic form \(q(y)=c(x(y))={1\over 2} y^ T Qy+b^ T y+a\) through the scaling \(y=x/(1+p^ T x)\), \(x=y/(1-p^ T y)\). The function \(c(x)\) has a unique minimum only if \(Q\) is positive definite. In this case, provided that a solution does exist, the minimizer \(\beta\) is given by \(\beta=Q^{- 1}b/(1+p^ T Q^{-1}b)\), \(1+p^ T Q^{-1} B\neq 0\). Most of the conic methods model the objective function \(f(x)\) around \(x\) by a model of the form \[ c(x_ i+d)=f_ i+{g^ T_ i d\over 1+p^ T_ i d}+{{\textstyle{1\over 2}}d^ T qd\over(1+p^ T_ id)^ 2}, \] where \(f_ i\), \(g_ i\), and \(F_ i\) are respectively, the function value, the gradient, and the Hessian of \(f(x)\) at \(x_ i\). However, the authors of this paper use a model of the form \[ c(x)={1\over 2}\cdot{1+p^ T x\over 1+p^ T\beta}\nabla c(x)(x-\beta)+c(\beta), \] which does not involve an estimate of the conjugacy matrix \(Q\) on the Hessian matrix of the objective function. An algorithm is given which is based on this, and it converges in \(n+1\) iterations on conic functions of \(n\) variables. Numerical results for many standard general test functions are given which indicate that the algorithm is very robust and superior in function evaluations and number of iterations to the corresponding quadratic method.