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.

1.6 Stochastic Processes 133<br />

X0 has any given distribution with V(X0) > 0. The sequence {Xt : EXt =<br />

EXt−1, VXt = t−1<br />

k=0 VXk} is not a Markov chain.<br />

A Markov chain that is not a martingale, for example, is {Xt : Xt<br />

2Xt−1}, where X0 has any given distribution with E(X0) = 0.<br />

A common application <strong>of</strong> martingales is as a model for stock prices. As<br />

a concrete example, we can think <strong>of</strong> a random variable X1 as an initial sum<br />

(say, <strong>of</strong> money), and a sequence <strong>of</strong> events in which X2, X3, . . . represents a<br />

sequence <strong>of</strong> sums with the property that each event is a “fair game”; that<br />

is, E(X2|X1) = X1 a.s., E(X3|X1, X2) = X2 a.s., . . .. We can generalize this<br />

somewhat by letting Dn = σ(X1, . . ., Xn), and requiring that the sequence be<br />

such that E(Xn|Dn−1) a.s.<br />

= Xn−1.<br />

Doob’s Martingale Inequality<br />

A useful property <strong>of</strong> submartingales is Doob’s martingale inequality. This<br />

inequality is a more general case <strong>of</strong> Kolmogorov’s inequality (B.11), page 841,<br />

and the Hájek-Rènyi inequality (B.12), both <strong>of</strong> which involve partial sums<br />

that are martingales.<br />

Theorem 1.69 (Doob’s Martingale Inequality)<br />

Let {Xt : t ∈ [0, T]} be a submartingale relative to {Dt : t ∈ [0, T]} taking<br />

nonnegative real values; that is, 0 ≤ Xs ≤ E(Xt|Dt) for s, t. Then for any<br />

constant ɛ > 0 and p ≥ 1,<br />

<br />

Pr sup Xt ≥ ɛ<br />

0≤t≤T<br />

d<br />

=<br />

≤ 1<br />

ɛ pE(|XT | p ). (1.281)<br />

Pro<strong>of</strong>. ***fix<br />

Notice that Doob’s martingale inequality implies Robbins’s likelihood ratio<br />

martingale inequality (1.279).<br />

Azuma’s Inequality<br />

extension <strong>of</strong> Hoeffding’s inequality (B.10), page 840<br />

1.6.7 Empirical Processes<br />

Given a random sample X1, . . ., Xn with the order statistics X(1), . . ., X(n),<br />

we can form a conditional discrete distribution with CDF<br />

⎧<br />

⎨0<br />

x < X(1)<br />

Fn(x) = k/n X(k) ≤ x < X(k+1) for 1 ≤ k < n (1.282)<br />

⎩<br />

1 X(n) ≤ x.<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!