Judd and Slotnick, 1994, Numerical Dynamic Programming with Shape-Preserving Splines, here.
Parametric approximations of the value function are a critical feature of the value function iteration method for solving dynamic programming problems with continuous states. Simple approximation methods such as polynomial or spline interpolation may cause value function iteration to diverge. We show that shape-preserving splines can avoid divergence problems while producing a smooth approximation to the value function.
Ruszczynski et al., Lectures on Stochastic Programming, here. I remember seeing this before.