Stein's method of normal approximation: some recollections and reflections (Q2054466)

The paper under review focus on the characterization of the normal distribution and the construction of Stein identities to work directly on spaces of random variables with the solution of a first order ordinary differential equation playing a crucial role. Section 2 starts with the characterization for a random variable \(W\) to be standard normal distribution by the necessary and sufficient condition \[ E[ f'(W) - W f(W)] =0, \ \ \ f\in \mathcal{G} \] where \(\mathcal{G}\) is the class of all absolutely continuous function with a.e. derivative \(f'\) such that \(E[|f'(Z)|] <\infty\) for \(Z\sim N(0, 1)\). The distance between \(\mathcal{L}(W)\) and \(N(0, 1)\), the Wasserstein distance \(d_W\), the Kolmogorov distance \(d_K\) and the total variation distance \(d_{TV}\) are recalled. Section 3 introduces the Stein identity \begin{align*} d_H(W, Z) &= \sup_{h\in H} |E[f_h'(W) - Wf_h(W)]|\\ &= \sum_{i=1}^n \sigma_i^2 E[f_h'(W) - f'_h(W^{(i)})] - \sum_{i=1}^n \sigma_i^2 E[f_h'(W^{(i)}+T_i) - f'_h(W^{(i)})], \end{align*} where \(W=\sum_{i=1}^nX_i\) and \(W^{(i)} = W -X_i\) for an independent sequence of random variables \(X_i\) with \(E[X_i]=0, \operatorname{Var}(X_i)=\sigma_i^2\) and \(E|X_i|^3 < \infty\). This leads to a bound for the Wasserstein distance with optimal order on the bound, and also for the Kolmogorov distance \(d_K\) by replacing the test function with the indicator function \(h_x(w) = 1_{w\le x}\). The first author et al. [Ann. Appl. Probab. 31, No. 6, 2881--2923 (2021; Zbl 1484.60025)] studied a Stein identity of the form \[ EWf(W) = E \int_R f'(W+t)\hat{K}(t) dt, \] where \(\hat{K}\) is a random function and obtained a general bound on the Kolmogorov distance in the normal approximation for \(W\). \textit{S. Chatterjee} [Probab. Theory Relat. Fields 143, No. 1--2, 1--40 (2009; Zbl 1152.60024); in: Proceedings of the International Congress of Mathematicians (ICM 2014), Seoul, Korea, August 13--21, 2014. Vol. IV: Invited lectures. Seoul: KM Kyung Moon Sa. 1--24 (2014; Zbl 1373.60052)] constructed a Stein identity, by Gaussian interpolation, \[ EWf(W) = ET(X)f'(W), \ \ \ T(x) = \int_0^1 \frac{1}{2\sqrt{t}} E \left(\nabla g (x) \cdot \nabla g(\sqrt{t}x+\sqrt{1-t} x) \right)dt. \] Using integration by parts, \textit{I. Nourdin} and \textit{G. Peccati} [Probab. Theory Relat. Fields 145, No. 1--2, 75--118 (2009; Zbl 1175.60053)] obtained the Stein identity \[ EFf(F) = E[f'(F)\langle DF, - DL^{-1}F\rangle_{L^2(R_+)} ] =E Tf'(f), \] for the Malliavin derivative \(D\) and \(T= \langle DF, - DL^{-1}F\rangle_{L^2(R_+)} \). Section 4 recalls the original idea of Stein for the limit of characteristic function. Section 5 provides a discussion on the concentration inequality which Stein used to overcome the difficulty arising from the discontinuity of the test function \(h\). Stein's approximation cannot be complete without a bound on the Kolmogorov distance depending on the third moment, and this is exactly where the concentration inequality comes to the rescue. Section 6 deals with exchangeable pairs which can be realized as a pair of successive states in a stationary reversible Markov chain, and \textit{L. H. Y. Chen} and \textit{X. Fang} [Bernoulli 21, No. 1, 335--359 (2015; Zbl 1354.60011)] used exchangeable pairs and a concentration inequality on a more general version of the combinatorial central theorem and obtained a third moment bound on the Kolmogorov distance. It is interesting to note that Stein's ideas on high dimensional data analysis are applied to extend the advantage of shrinkage estimation to non-Gaussian setups, and to estimate parameters and evaluate the costs for the violation of Gaussian assumptions in single index models and compressed sensing.