Decay of convolved densities via Laplace transform (Q6142945)

Let \(X=X_1+\cdots+X_n\) be a sum of independent and identically distributed vectors in \(\mathbb{R}^d\), where each \(X_k\) has density \(p_k\) satisfying \(p_k(x)\leq M_k\) for all \(x\in\mathbb{R}^d\), for some constants \(M_k\). Letting \[ V^*(x)=\sup_{t\in\mathbb{R}^d}\left[\langle t,x\rangle-\log\mathbb{E}e^{\langle t,X\rangle}\right]\,, \] the author shows that the density \(p\) of \(X\) is continuous and satisfies \[ p(x)\leq M\exp\left\{-\frac{1}{2}V^*(x)\right\} \] for \(x\in\mathbb{R}^d\), where \(M\) is a function of \(M_1,\ldots,M_n\). The constant \(M\) may be taken to be the geometric mean \(M=(M_1\cdots M_n)^{1/n}\) or, in the setting where \(\sum_{j\not=k}M_j^{-2/d}\geq M_k^{-2/d}\) for each \(k\), we may take \[ M=e^{d/2}\left(M_1^{-2/d}+\cdots+M_n^{-2/d}\right)^{-d/2}\,. \] The author shows through explicit examples that this latter choice can significantly outperform the former. Applications of these bounds are given to the central limit theorem in Rényi divergence and to Bernoulli convolutions. Particular attention is also paid to subexponential and sub-Gaussian random variables.