Characterizations of distributions by generalizations of variance bounds and simple proofs of the CLT (Q1372400)

The authors begin by summarizing the following known results regarding the lower/upper bounds for functions of rv's. (In the following \(g\) is any absolutely continuous function.) 1. Let \(X\sim N(0,s^2)\). Then \(s^2E^2 [g'(X)]\leq \text{var} [g(X)]\leq s^2E [g'(X)]^2\) (Chernoff 1981; Cacoullos 1982). 2. (Multivariate version) Let \(X=(X_1, \dots,X_p) \sim N(0,S)\). Then \(E[g'(X)] SE[g(X)]\leq\text{var}[g(X)]\leq E[g'(X)Sg(X)]\). Here \(g'(x)=(g_1, \dots, g_p)\) and \(g_i\) is the partial derivative of \(g\) wrt \(x_i\) (Chen 1982; Cacoullos 1982). 3. Equality attains in the above inequalities iff \(g\) is linear. 4. Let \(X=(X_1,\dots,X_p)\) with \(X_i\) independent rv's with densities \(f_i\), means \(m_i\) and variances \(s^2_i\). Then \[ \sum^p_{i=1} \biggl\{ s^2_iE^2 \bigl[w_i(x_i) g_i(X)\bigr]\biggr\}\leq \text{var} \bigl[g(X)\bigr] \leq \sum^p_{i=1} \biggl\{s_i^2E \bigl[w_i(X_i)g^2_i(X)]\biggr\} \] where the \(w_i\) functions are defined by \(s^2_if_i(x_i)w_i(x_i)= \int(m_i-t)f_i(t)dt\) (Cacoullos 1989). 5. The above results also hold for discrete variables if they are non-negative and integer-valued and the derivative \(g'(x)\) is replaced by the difference \(g(x+1)-g(x)\) and the partial derivative \(g_i(x)\) is replaced by the partial difference \(g(x+z_i)-g(x)\). Here \(z_i\) is the 0-vector except for a 1 in the \(i\) th position. 6. Note that 1. above implies that (in 4. with \(p=1)\) \(w=1\) iff \(X\) is normal. 7. Also in 4. with \(p=1\), if \(X\) is discrete, \(w=1\) characterizes the Poisson distribution (Cacoullos 1985). 8. Further, if \(\text{var}[w_n(X_n)]\to 0\) as \(n\to\) infinity then the corresponding standardized rv's \(X_n\to N(0,1)\) weakly (Cacoullos et al. 1992). 9. (Most general case) Let \(X=(X_1,\dots,X_p)\) be a random vector with density \(f(x)\) and dispersion matrix \(S\). Let \(w(x)=(w_1(x),\dots,w_p(x))^T\), the vector \((q_1(x), q_2(x),\dots, q_p(x))=S^{-1}x\), \(m_i=E[q_i(X)]\), \(y_i=\) vector \(x\) with \(x_i\) replaced by \(t_i\) and finally \(w_i(x)f(x)= \int(m_i-q_i(y_i)) f(y_i)dt_i\). Then 4. above takes the following form: \[ E(w_1g_1, \dots,w_pg_p) SE(w_1g_1, \dots,w_p g_p)^T\leq\text{var}[g(X)]\leq??? \] (the general upper bound is not known) (Cacoullos et al. 1992). 10. Let \(U_x=\sup_g\text{var}[g(X)]/s^2 E[g'(X) ]^2\). Then \(U_x\geq 1\) with equality iff \(X\) is normal (Borovkov et al. 1983). 11. Let \(Lx=\inf_g\text{var}[g(X)]/s^2E[g'(X)]^2\). Then \(Lx\leq 1\) with equality iff \(X\) is normal (Calcoullos et al. 1989). 12. Let \(X\) be a rv with support in the interval \((a,b)\), density \(f\), mean \(m\) and variance \(s^2\). Let \(h(x)\) be a given function and \(z(x)\) be defined by \(z(x)f(x)= \int^x_a(E(h)-h(t))f(t)dt\). Assume that \(E| z(X)g'(X)|\) is finite. Then \(\text{cov} (h(X), g(X))= E(z(X)g'(X))\) (Cacoullos 1995). In this paper lower/upper bounds are obtained by using a generalization of the type indicated in 12. above and the resultss are used to obtain characterizations of distributions. Here are the typical results: A. Under the assumptions of 12., if \(h\) is absolutely continuous \((ac)\), \[ (*)\qquad\text{var} [g(X)]\geq E^2[z(X) g'(X)]/E [z(X)h'(X)]. \] Equality holds iff \(g\) is linear in \(h\;(**)\). B. If there are functions \(h\) and \(z\) that satisfy \((*)\) for every differentiable function \(g\) and equality holds under \((**)\), then \(h,z\) and \(f\) are related as in 12. C. Under the conditions of \(A\), let \(h\) be \(ac\) with \(h'(x)>0\), then \[ (***)\qquad\text{var} [g(X)]\leq E^2[z(X) (g'(X))^2/h'(X)]. \] Equality holds iff \(g\) is linear in \(h\). D. If there are functions \(h\) and \(z\) that satisfy \((***)\) for every \(ac\) function \(g\) and equality holds under \((**)\), then \(h,z\) and \(f\) are related as in 12. E. Setting \(g(x)=x\) in \(A\) and \(B\) above we obtain: Let \(X\) have a log concave density \(f\), then \(\text{var} (X)\leq E[-1/(\ln f(x))'']\) with equality iff \(X\) is normal. F. A necessary and sufficient condition for \(\text{var}(g(X)){}\leq cE[(g'(X))^2]\) to hold for every \(ac\) \(g\) with equality iff \(g\) is linear in some function \(h\) which is continuously twice differentiable is that the density \(f\) be of the form \(f(x)=c_0 \exp[-\int (ch''(x)+ h(x)-Eh(X)) dx/ch'(x)].\) G. Discrete versions of \(A\) through \(E\) are also true. H. Let \(X_1,X_2, \dots\) be a sequence of standardized rv's with corresponding \(w\)-functions \(w_1,w_2, \dots\) If \(\text{var} [w_n(X_n)]\to 0\) then \(X_n\) converges weakly to \(N(0,1)\).