Critical points via monodromy and local methods (Q507163)

In this article, a new numerical method is introduced to compute all of the critical points of an objective function \(\psi(x,u):\mathbb{R}^{n} \times \mathbb{R}^{n}\to \mathbb{R}\), restricted to an algebraic variety \(X\in \mathbb{R}^{n}\), called model. This method consists of two different parts: -- firstly, the monodromy group action is used to randomly explore the variety of critical points; -- secondly, a trace test is used to determine with probability one, if the computed subset contains all of the critical points. This result is of interest to many areas of applied mathematics and statistics, since it concerns the problem to find the best representative of a model by optimizing an objective function. In particular, in this paper two cases are investigated with \(\psi\) being a quadratic norm and with \(\psi\) being a likelihood function. An illustrating example is given to the critical point problem. To conclude, this method is used to reproduce some known ED-degrees and ML-degrees. The times of those computations are compared against this new method.