The linearizing projection, global theories (Q1099793)

A linearizing projection is a device, such as the Shapley value, which converts nonlinear information into linear. In this paper we construct a local theory of the linearizing projection or \((x_ 0,D,{\mathcal D})\)-value and give a few applications. Let D be a subset of a real Banach space E; let \(x_ 0\) be a distinguished point in D; let \({\mathcal D}\) be a group of isomorphisms from E into E such that \(Tx_ 0=x_ 0\) and T(D)\(\subset D\) for each T in \({\mathcal D}\). Let X be a linear space of real value functions (functionals) on D which includes the space \(E^*_ D\) of all continuous linear functionals on E restricted to D and has the property that whenever F is in X and T is a continuous linear mapping from E into E such that T(D)\(\subset D\) the functional defined by \(F\circ T(x)=F(T(x))\) for x in D is also in X. 1.1. Definition. An \((x_ 0,D,{\mathcal D})\)-value is a function P from X onto \(E^*_ D\) which satisfies the following conditions: \((L_ 1)\) P is linear; i.e. \(P(aF+bG)=aPF+bPF\) for all a,b in R, F,G in X; \((L_ 2)\) P is \({\mathcal D}\)-symmetric; i.e., if T is in \({\mathcal D}\), then \(P(F\circ T)=(PF)\circ T\) for all F in X; \((L_ 3)\) P is \(x_ 0\)-efficient; i.e., \((PF)(x_ 0)=F(x_ 0)\) for all F in X; \((L_ 4)\) P is idempotent; i.e., \(Pf=f\) for all f in \(E^*_ D.\) Thus, an \((x_ 0,D,{\mathcal D})\) value is a projection from X onto \(E^*_ D\) which is \({\mathcal D}\)-symmetric and \(x_ 0\)-efficient. Suppose the set D contains \(x_ 0\) and 0 and that the line joining 0 to \(x_ 0\), namely \(\{\) tx: \(0<t<1\}\), is contained in the interior of D. Let \(C^ 1_ 0(D)\) denote the space of all real functions F on D which satisfy the following conditions: (1) \(F(0)=0\); (2) F is continuous on D; (3) F is Gâteaux differentiable on the interior \(D^ 0\) of D; (4) the \(corren,...,x_{\pi (n)})\). For vectors \(x,y,u,v\in R^ n\) we write \((x,y)=(u,v)\) if there exists a permutation \(\pi\) of \(\{\) 1,...,n\(\}\) such that \(x_{\pi}=u\) and \(y_{\pi}=v\). Further, (x,y)\(\leq^{a}(u,v)\) if there exists a finite number of vectors \(z^ 1,...,z^ k\) such that \[ (i)\quad (x,y)=^{a}(x\uparrow,z^ 1)\quad and\quad (z\uparrow,z^ k)=^{a}(u,v), \] and (ii) \(z^{i-1}\) can be obtained from \(z^ i\) by interchanging two components of \(z^ i\), the first of which is less than second. A function g of two vectors arguments x and y for which g(x,y)\(\leq g(u,v)\) when (x,y)\(\leq^{a}(u,v)\) is said to be arrangement increasing (or AI). \textit{M. Hollander}, \textit{F. Proschan} and \textit{I. Sethuraman} [Ann. Stat. 5, 722-733 (1977; Zbl 0356.62043)] have shown, that the class of AI functions is closed under some basic operations. In particular, the convolution of AI functions is AI. In the present paper, the authors prove some corollaries of this result of Hollander, Proschman and Sethuraman. In section 2, the authors show that many geometric probabilities of random vectors X with exchangeable densities f(x) (that is, the distribution of \(x_{\pi}\) does not depend on the permutation \(\pi)\) are AI functions. In section 3 they reduce examples AI functions of geometric type from AI families of densities. Some moment inequalities are also obtained.