A shorter proof of Kanter's Bessel function concentration bound (Q2641903)

The main result is Theorem 2.1: \(Q(\{k,k+1\})<S_\lambda (\{0,1\})=G(\lambda)=e^{-\lambda}(I_0(\lambda)+I_1(\lambda))\). Here \(Q\) is the convolution of \(n\) symmetric three-point distributions with masses \(p_j/2\), \(1-p_j\) \(p_j/2\) in \(-1,0\) and \(1,j=1,\dots,n\), and \(S_\lambda\) is the symmetrized Poisson distribution with parameter \(\lambda=p_1+\cdots+p_n>0\), i.e., the distribution of \(U-V\) where independent \(U\) and \(V\) are Poisson \((\lambda/2)\). This distribution has mass \(e^{-\lambda}I_h(\lambda)\) with \(h=|j|\) in \(j\in\mathbb{Z}\). Here \(I_k\) is the modified Bessel function of order \(k\). This theorem is slightly more general than Lemma 4.3 in the paper on concentration functions by \textit{M. Kanter} [J. multivariate Analysis 6, 222--236 (1976; Zbl 0347.60043)]. The proof is claimed to be shorter than Kanter's. By choosing suitable \(p_1,\dots, p_n\) a set of inequalities for the concentration function of the sum of \(n\) independent real-valued random variables is derived. The inequality 2.1 is sharp. The proof of 2.1 starts by Fourier inversion of the characteristic function of \(Q\) and uses a special analytic inequality proved in section 3.1 by hard analysis. A number of bounds, some asymptotic, for \(G(\lambda)\) is proved.