Convex-invariant means and a pathwise central limit theorem (Q1092505): Difference between revisions

Let B(t) \((0\leq t<\infty)\) be standard Brownian motion in \({\mathbb{R}}^ 1\) starting at 0. For an interval \(I=(a,b)\) let \(Z(t)=1\) if B(t)\(\sqrt{t}\in I\) and \(=0\) else. The average time the process Z(t) spends in 1 should be \(\Phi\) (b)-\(\Phi\) (a), where \(\Phi\) is the distribution function of the N(0,1)-distribution. But the Cesàro-averages of Z(t) do not converge a.e. It is shown here that the desired convergence a.e. obtains if the Cesàro-method is replaced by stronger averaging methods. Let \(P_ 1({\mathbb{R}}^+)\) denote the set of probability densities \(\phi\) on \({\mathbb{R}}^+\). Define the operator \(E:L_{\infty}({\mathbb{R}})\to L_{\infty}({\mathbb{R}})\) by \(Ef(x)=f(\exp (x))\) and \(E^{-1}\) by \((E^{- 1}f)(x)=f(\log x)\) for \(x>\) and \(=0\) else. Let \(A^ 0_{\phi}f=\phi^*f\) denote the convolution operator. An averaging operator of order n is of the form \(A^ n_{\phi}=E^{-n}A^ 0_{\phi}E^ n\), and an averaging method of order n is \(\lim_{t\to \infty}(A^ n_{\phi}f)(t)\). If \(\phi (x)=e^{-x}\) \((x>0)\) then the usual Cesàro method is obtained for \(n=1\). For \(n\geq 2\) the order n averages of Z(t) converge a.e. to \(\phi\) (b)-\(\phi\) (a). Actually, the exponential change of scale by E may be replaced by rather general convex changes of scale. Let \({\mathcal C}_ 0\) denote the set of increasing continuous convex functions \(c: {\mathbb{R}}^+\to {\mathbb{R}}^+\) with \(c'(x)>0\) and c''(x)\(\geq 0\) for all x, and with c'(x)\(\to \infty\) (x\(\to \infty)\). It is shown that for \(c\in {\mathcal C}_ 0\) there exists a mean \(\lambda\) on \(L_{\infty}\) such that \(\lambda f=\lambda (\phi^*f)\) for each \(\phi\), and \(\lambda f=\lambda (f\circ c).\) In section III the scale change approach is combined with ideas of Mokobodzki. Under the assumption of Martin's axiom there exists for \(c\in {\mathcal C}_ 0\) a mean \(\lambda\) which is convolution-invariant, c- invariant and measure-linear. A mean defines an (additive) random choice on the line. An application yields a ``Monte Carlo method'' on the line of the following type: if t is picked with (symmetric) \(\lambda\) as above, then B(t) is again such a random choice from \({\mathbb{R}}:\) Formally: Put \({\bar \lambda}\)f\(=(\lambda f+\lambda f^*)/2\) where \(f^*(t)=f(-t)\). Almost surely \({\bar \lambda}\)f\(={\bar \lambda}(f(B(\cdot)).\) There are various other examples, counterexamples and open questions.