88 Contiguity

Before answering the question of interest, we give two characterizations of contiguity

in terms of the asymptotic behavior of the likelihood ratios of Pn and Qn. The likelihood
ratios dQnfdPn and dPn/dQn are nonnegative and satisfy
dQn dPn
E" - - < I and EQ - - < 1.
'• d Pn - " d Qn -
Thus, the sequences oflikelihood ratios d Qn/ d Pn and d Pn/ d Qn are uniformly tight under
Pn and Qn. respectively. By Prohorov's theorem, every subsequence has a further weakly
converging subsequence. The next lemma shows that the properties of the limit points
determine contiguity. This can be understood in analogy with the nonasymptotic situation.
For probability measures P and Q, the following three statements are equivalent by (iii) of
Lemma6.2:
dQ
Q « P, Ep dP = 1.

This equivalence persists if the three statements are replaced by their asymptotic counter-
parts: Sequences Pn and Qn satisfy Qn<JPn, ifandonlyifthe weak limit points of dPnfdQn
under Qn give mass 0 to 0, if and only if the weak limit points of dQn/dPn under Pn have
mean 1.

6.4 Lemma (Le Cam'sfirst lemma). Let Pn and Qn besequencesofprobabilitymeasures

on measurable spaces (Qn, An). Then the following statements are equivalent:
(i) Qn <I Pn.
(ii) IfdPnfdQn .& U along a subsequence, then P(U > 0) = 1.
2,3: prove contiguous;
(iii) IfdQn/dPn V along a subsequence, then EV = 1.
4: prove not contiguous
( iv) For any statistics Tn : Qn 1--+ :Ilk: If Tn 0, then Tn 0.

Proof. The equivalence of (i) and (iv) follows directly from the definition of contiguity:
Given statistics Tn. consider the sets An = {IITn II > e }; given sets An. consider the statistics
Tn = lA.·
(i) =} (ii). For simplicity of notation, we write just {n} for the given subsequence
along which dPnfdQn .& U. For given n, we define the function gn(e) = Qn(dPnfdQn <
e) - P(U < e). By the portmanteau lemma, liminf gn(e) 2: 0 for every e > 0. Then, for
en .J.. 0 at a sufficiently slow rate, also lim inf gn (en) 2: 0. Thus,

P(U = 0) = limP(U <en) :::: liminf Qn <en).

On the other hand,

If Qn is contiguous with respect to Pn, then the Qn-probability of the set on the left goes
to zero also. But this is the probability on the right in the first display. Combination shows
that P(U = 0) = 0.
(iii)=} (i). If Pn(An) --* 0, then the sequence lo.-A. converges to 1 in Pn-probability.
By Prohorov's theorem, every subsequence of {n} has a further subsequence along which