Likelihood equations and scattering amplitudes (Q2076291)

The paper draws a parallel between likelihood equations in statistics and the scattering equations of particles in physics in the formulation of Cachazo-He-Yuan (CHY). Both sets of equations arise from determining the critical points of \[ \ell_s(p) := \sum_{i=0}^n s_i \log(p_i) - \left( \sum_{i=0}^n s_i \right) \log \left( \sum_{i=0}^n p_i \right), \] subject to further model constraints. In statistics, \(s\) is a fixed data vector and \(p\) are unknown probabilities. The minimization of the above \textit{log-likelihood function} yields the ``best explanation'' of the observed data on the statistical model. On the physics side, \(s\) is the collection of Mandelstam invariants and \(p\) encodes the positions of repelling particles on a line; solutions to the critical equations represent equilibria. The connection is first exemplified by the construction of a statistical model on the moduli space of \(m\) points in general position on the projective line; this mirrors the CHY setup of \(m\) particles. The result on the physics side, due to Cachazo-Mizera-Zhang, that under certain linear assumptions on \(s\), all \((m-3)!\) equilibria are real is recovered from Varchenko's theorem in algebraic statistics. The paper then goes on to compute the critical points for some higher Grassmannians in place of the moduli space of \(m\) points on a line; a statistical model consisting of symmetric distributions with low rank; and positively parametrizable models, again connecting to scattering amplitudes in physics. A central aspect of the paper is that it is also a case study of how to solve rational function equations numerically over the reals and how to certify distinctness of solutions in practice. The employed software \texttt{HomotopyContinuation.jl} achieves astonishing results, in particular concerning the runtime required to count the number of real solutions in the CHY \(m\)-particle case, confirming numbers in the physics literature. The certification feature makes the presented approach suitable for computing (lower bounds on) algebraic degrees of optimization problems, for real enumerative geometry and enumerative problems in physics. I found one remark at the bottom of page 180 easy to misinterpret, namely that ``all discrete conditional independence models'' have a parametrization by positive rational functions from the positive orthant. This would imply that all discrete CI models have a rational point and settle an open question of Matúš from 1999. After talking to the authors, I think what they had in mind are models defined by a \textit{single} conditional independence statement. Then the claim is true and a satisfactory argument is given in that paragraph of the paper.