No label defined (Q3973211)

[For the entire collection see Zbl 0702.00018.] The following problem is considered: \[ \min\{F_ 0(x)\mid x\in R^ N\}=\min\{2^{-1}(Ax,x)-\langle f,x\rangle\mid x\in R^ N\}, \leqno(1) \] A is a positively defined nonsingular \(N*N\) matrix. For solving problem (1) the following iterative \((i,j,v)\)-method is presented: \(x_{n+1}=P_ i(x_ n)\), if \(\alpha_ i(x_ n)\geq v\alpha_ j(x_ n)\); \(x_{n+1}=P_ j(x_ n)\), if \(\alpha_ i(x_ n)<v\alpha_ j(x_ n)\); \(x_ 0=0\in R^ N\), \(0<v<1\), \(n=0,1,2,\ldots\); \(P_ n(x)=x-\alpha_ n(x)\xi(x)\), \(\xi(x)=Ax-f\), \(\alpha_ n(x)=(A^{2^ n-1}\xi(x),\xi(x))/(A^{2^ n}\xi(x),\xi(x))\). The rate of convergence is estimated for the two-step methods \(x_{n+1}=P_ j(P_ k(x_ n))\) with \(N=2\) for different \(i, j\). In particular, the following variants are considered: 1) \(j=k\); 2) \(j=0\), \(k=1\); 3) \(j=1\), k=2\(; 4) \)j=1\(, k=3\); 5) \(j=1\), \(k=4\); 6) \(j=2\), \(k=3\). The method of steepest descent and the minimal errors method are special cases of this algorithm with \(j=k=0\) and \(j=k=1\) accordingly. The proposed algorithm is supposed to accomplish the solution of the linear system \(Bx=y\) by minimizing the quadratic function \(F_ 0(x)=0.5\| Bx-y\|^ 2\). Numerical tests are reported for a \((50\times 50)\)- matrix \(B\). The results show the advantage of the \((1,0,0)\)- and \((1,0,1)\)-methods compared to methods of the steepest descent and of minimal errors.