Functional approximations via Stein's method of exchangeable pairs (Q2028947)

Stein's characterization of a standard normal law has found many applications since its statement about fifty years ago. Combined with the method of exchangeable pairs, it was successfully used to obtain bounds on the rate of convergence of various sequences of processes to some Gaussian limits. The present paper follows this way in the framework of functional approximation. An abstract approximation theorem is proved stating that if a stochastic process satisfies a linear regression condition, its distance from a Gaussian mixture process may be explicitly bounded. A first application concerns the sum of processes obtained from random permutations, a technique used in non-parametric statistics. A second example provides a bound on the distance from a continuous Gaussian process of the process representing the rescaled total number of edges in a Bernoulli random graph.