Machine learning the real discriminant locus (Q2674017)

Systems of polynomial equations arising in applications typically occur in \textit{parametric form}; for many real-valued choices of parameter values, we may need to solve the system over and over again. For both numerical homotopy continuation methods and the so-called algebraic methods (Gröbner bases, etc.), a weakness from the applied point of view is that they usually must compute all solutions over the complex numbers, whereas only real solutions are of interest. This paper investigates several approaches to sampling the discriminant locus of a parametric family of systems where the behavior of the real solutions changes, and using these samples to drive inference about the number of real solutions and how many solution paths to track. A key contribution is a novel sampling scheme for sampling parameter points near the boundary of the discriminant locus. Using training data sampled in this way, experimental results show that classification of the number of real solutions on several test cases than for uniformly sampled training data. This can be accomplished without ever computing the equation of the discriminant using a \textit{witness set} of parameter values whose cardinality is the degree of the discriminant. Examples involving parametrized univariate quadratic and cubic polynomials are studied, along with the multivariate equations of equillibria for the Kuramoto model with 3 or 4 oscillators. Experiments study neural networks, support vector machines, and nearest-neighbor classifiers. Experiments on the chosen examples demonstrate the ability to compute, with high rates of success, all real solutions by tracking only real paths chosen via nearest-neighbors from a sufficiently large set of pre-computed samples.