06.06.2013 Views

Theory of Statistics - George Mason University

Theory of Statistics - George Mason University

Theory of Statistics - George Mason University

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

752 0 Statistical Mathematics<br />

0.1.13 Functionals<br />

Functionals are functions whose arguments are functions. The value <strong>of</strong> a functional<br />

may be any kind <strong>of</strong> object, a real number or another function, for<br />

example. The domain <strong>of</strong> a functional is a set <strong>of</strong> functions.<br />

If F is a linear space <strong>of</strong> functions, that is, if F is such that f ∈ F and<br />

g ∈ F implies (af + g) ∈ F for any real a, then the functional Υ defined on<br />

F is said to be linear if Υ(af + g) = aΥ(f) + Υ(g).<br />

A similar expression defines linearity <strong>of</strong> a functional over a distribution<br />

function space P: Υ defined on P is linear if Υ((1 − w)P1 + wP2) = (1 −<br />

w)Υ(P1) + wΥ(P2) for w ∈ [0, 1] and P1, P2 ∈ P.<br />

Functionals <strong>of</strong> CDFs have important uses in statistics as measures <strong>of</strong> the<br />

differences between two distributions or to define distributional measures <strong>of</strong><br />

interest. A functional applied to a ECDF is a plug-in estimator <strong>of</strong> the distributional<br />

measure defined by the same functional applied to the corresponding<br />

CDF.<br />

0.1.13.1 Derivatives <strong>of</strong> Functionals<br />

For the case in which the arguments are functions, the cardinality <strong>of</strong> the<br />

possible perturbations is greater than that <strong>of</strong> the continuum. We can be precise<br />

in discussions <strong>of</strong> continuity and differentiability <strong>of</strong> a functional Υ at a point<br />

(function) F in a domain F by defining another set D consisting <strong>of</strong> difference<br />

functions over F; that is the set the functions D = F1 − F2 for F1, F2 ∈ F.<br />

The concept <strong>of</strong> differentiability for functionals is necessarily more complicated<br />

than for functions over real domains. For a functional Υ over the<br />

domain F, we define three levels <strong>of</strong> differentiability at the function F ∈ F.<br />

All definitions are in terms <strong>of</strong> a domain D <strong>of</strong> difference functions over F, and<br />

a linear functional ΛF defined over D in a neighborhood <strong>of</strong> F. The first type<br />

<strong>of</strong> derivative is very general. The other two types depend on a metric ρ on<br />

F × F induced by a norm · on F.<br />

Definition 0.1.53 (Gâteaux differentiable)<br />

Υ is Gâteaux differentiable at F iff there exists a linear functional ΛF (D) over<br />

D such that for t ∈ IR for which F + tD ∈ F,<br />

<br />

Υ(F + tD) − Υ(F)<br />

lim<br />

− ΛF(D) = 0. (0.1.115)<br />

t→0 t<br />

In this case, the linear functional ΛF is called the Gâteaux differential <strong>of</strong><br />

Υ at F in the direction <strong>of</strong> F + D.<br />

Definition 0.1.54 (ρ-Hadamard differentiable)<br />

For a metric ρ induced by a norm, Υ is ρ-Hadamard differentiable at F iff<br />

<strong>Theory</strong> <strong>of</strong> <strong>Statistics</strong> c○2000–2013 James E. Gentle

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!