A practical computational framework for the multidimensional moment-constrained maximum entropy principle (Q2572793)

The author considers the problem of minimization of the Shannon entropy \[ S(p)=\int_{\mathbb{R}^N} p(\lambda)\log p(\lambda)d\lambda \] or the relative entropy \[ P(p)=\int_{\mathbb{R}^N} p(\lambda)\log (p(\lambda)/\Pi(\lambda))d\lambda \] over all probability densities \(p\) on \(R^N\) which satisfy ``moment conditions'' (\(i=1,\dots,L\)) \[ \int_{\mathbb{R}^N}C_i(\lambda)p(\lambda)d\lambda=f_i \] for some fixed \(f_i\) and \(C_i\) being products of powers. The proposed numerical method to this problem is based on the Lagrange multipliers technique and an explicit Shannon entropy representation for a class of functions. Newton's method is used for the numerical optimization of the Lagrangian. Application to fitting a full-scale atmospheric general circulation model is considered.