Positive-part moments via characteristic functions, and more general expressions (Q1745276): Difference between revisions

The paper under review gives a unifying and generalizing approach to represent the positive part and absolute moments of a random variable for any real number exponent in terms of its characteristic function of the random variable. The characteristic function is a natural way to express Lévy processes and infinitely divisible distributions, the Spitzer identity to relate the distribution of the cumulative maxima of the partial sums of independent identically distributed (i.i.d.) random variables with the distributions of the positive parts of those sums that are examples to utilize the properties of the characteristic function. Section 2 introduces general results for any locally integrable Borel-measurable function \(g\) and any real number \(p\) in terms of \[ c_p^+(g) = \lim_{\varepsilon \to 0^+, T\to \infty}\frac{g(t)}{t^{p+1}}dt = \int_{0^+}^{\infty} \frac{g(t)}{t^{p+1}}dt, c_p^-(g) \int_{0^+}^{\infty} \frac{g(-t)}{t^{p+1}}dt. \] Hence, Proposition 2.1 expresses, from simple changing variables and definitions, \[ \int_{0^+}^{\infty} \frac{g(tx)}{t^{p+1}}dt = c_p^+(g) x_+^p + c_p^-(g)x_-^p, \] for \(x_+^p = x^p\) if \(x>0\), zero otherwise, and \(x_-^p = (-x)_+^p\). So \(|x|^p = x_+^p + x_-^p\) and \(x^[p] = x_+^p - x_-^p\). Theorem 2.3 states that for certain conditions on a real-valued random variable \(X\) and a complex-valued random variable \(Y\) \[ \int_{0^+}^{\infty} \frac{EYg(tX)}{t^{p+1}}dt = c_p^+(g) EYX_+^p + c_p^-(g)EYX_-^p, \] Hence for two functions \(g_1\) and \(g_2\) one can solve the linear equation to express \(EYX_+^p\) and \(EYX_-^p\) in Corollary 2.4. Taking \(Y=1\) and \(g(t)\) to be a polynomial in t, \(e^{ivt}\) and \(e^{-ivt}\), we get \(EYg(tX)\) as a linear combination of the derivatives of various orders of the characteristic function of \(X\) at points \(vZ\) for all integer multiples of v. If \(f\) is the characteristic function of a random variable X with \(E|X|^p<\infty\), then \(f^{(p)}(t) = i^p EX^pe^{itX}\) as a definition for any real number \(p\) derivative of characteristic functions. Theorem 2.5 further extends to \(g(t) =\sum_{j\in J}a_j\overline{t^{q_j}}e^{iv_jt}\) for a finite set J, by verifying Theorem 2.3 can be applied for this function \(g\). This is an omnibus result, specifying the parameters \(p, r, u, J, a_j, q_j\) and \(v_j\). Section 3 is to illustrate various special cases of Theorem 2.5. In Subsection 3.1 this is done for \(p=r=0\) and \(g(t) = \frac{e^{it} - e^{-it}}{2i}\) and \(f\) is the characteristic function of a random variable X. Using the identity \(2e^{iuX_+} = 1+e^{iuX} +e^{iuX}\mathrm{sgn}X - \mathrm{sgn} X\) and special choices of those parameters, one obtains the characteristic function of the positive part of X. Similarly, it works for \(X_-^p, |X|^p\) and \(X^{[p]}\) in Subsection 3.2. The case of \(|X|^p\) is mentioned in Subsection 3.3 for Brown's and Laue's different approaches with Lindeberg-type condition and the convergence of the absolute moments in the central limit theorem. Another different approach by Zolotarev is given in Subsection 3.4 with Gamma functions. Subsection 3.5 extends some results of Wolfe and Gurland for symmetric differences. Section 4 discusses computational issues involved in the integrals for those the small divisors may cause the instability, in particularly that the integral converges too slow near the infinity. By breaking the integral from zero to a finite number and another from the finite number to infinity, it reduces the degree of the near-cancellation effect in the numerator and the degree of the smallness of the denominator. Appendix 1 devotes to the understanding of constants \(c_p^{\pm }(g_n)\) used in the dominated convergence. Appendix 2 devotes to the fractional derivatives of the characteristic function.