An algorithm of MCMC method for solving \(f(x)=0\) (Q1430645)

An iterative stochastic algorithm for the numerical solution of a one dimensional equation \(f(x)=0\) is proposed. When an approximation \(x_n\) is obtained the guess value for the next step approximation is generated as a random variable \(xi\) with the mean \(x_n\) and the variance \(\lambda^2|f(x_n)|\). If \(|f(xi)|<c_n|f(x_n)|^2\) the guess is adopted, i.e. \(x_{n+1}=xi\), otherwise a new guess is generated with the same distribution. The author demonstrates that under two assumptions on the parameters \(\lambda\) and \(c_n\) \(n=1,2\dots\), \(f(x_n)\to 0\). (The author claims that \(x_n\) tends to some root of the equation but is is not true since there can be many roots and the sequence \(x_n\) can jump from one root to another). The second assumption is that \(\lim_{n\to\infty}\prod_{j=1}^n c_j=0\). The first assumption is formulated quite poorly and seems difficult to verify. In an example \(f(x)=x^2-ax+b+d\sin(1/x)\) is considered, \(xi\) are normally distributed, \(\lambda=300\), \(c=0.95\). The author does not show how to verify the first assumption in this case.