Convergence to stable laws in relative entropy (Q376268): Difference between revisions

Let \(X\) and \(Z\) be random variables with distributions \(\mu\) and \(\nu\), respectively. The relative entropy distance \(D(X||Z)\) is defined by \(D(X||Z)=\int_{R}\log(d\mu/d\nu)d\mu\) provided that \(\mu < \nu\); otherwise \(D(X||Z)=\infty\). Let \((X_{n})_{n\geq 1}\) be i.i.d. random variables, and assume that the normalized sums \(Z_{n}=(X_1+\dots +X_{n})/b_{n}-a_{n}\), where \(a_{n} \in R\) and \(b_{n}>0\), converge weakly to a random variable \(Z\) with a non-extremal stable distribution. The main result states that \(D(Z_{n}|| Z) \to 0\) as \(n \to \infty\) if and only if \(D(Z_{n_0}|| Z) \leq \infty\) for some \(n_0\). Reviewer's remark: The wrong statement ``By convexity of the function \(u \to \log(1+u)\), \(u \geq 0\)'' in the proof of Corollary 2.6 casts doubt about its validity.