Upload
daniel-chambers
View
217
Download
0
Embed Size (px)
Citation preview
Stochastic Processes and their Applications 32 (1989) 93- 107
North-Holland 93
NECESSARY CONDITIONS FOR NONLINEAR FUNCTIONALS
OF GAUSSIAN PROCESSES TO SATISFY
CENTRAL LIMIT THEOREMS
Daniel CHAMBERS”
Mathematic.s Drpartment, Boston College, Chestnut Hill, MA 02167, USA
Eric SLUD**
Received 28 July 1987
Revised 26 October 1988
Let X=(X,, t CR) be a stationary Gaussian process on (R, .F, P) with time-shift operators
( CJ,, s E R) and let H(X) = L’(f), u(X), P) denote the space of square-integrable functionals of
X. Say that Y t H(X) with EY =O satisfies the Central Limit Theorem (CLT) if
A family of martingales (Z,(t), t 2 0) is exhibited for which 2, (“c) = Z,, and martingale tech-
niques and results are used to provide suficient conditions on X and Y for the CLT. These
conditions are then shown to be necessary for slightly more restrictive central limit behavior of Y.
AMS Subject Ch\jfication.s: hOF05, 60G10, 60644
multiple Wiener-It8 integral 3 stochastic integral with respect to square-integrable
martingale * predictable variance process * martingale Functional Central Limit Theorem * Band-
weighted Central Limit Theorem
1. Introduction
Let X = (X,, t E [w) be a (progressively measurable) stationary Gaussian process on
a probability space (0, 9, P), with EX,, = 0 and E(X,,X,) = I, e”‘a(dx) for a fixed
nonatomic probability measure v on Iw, called the “spectral measure” of X. Take
9= a(X) to be the a-field generated by the variables X,, t E [w, and denote by
{U, E Iw} the family of “time-shift” linear isometries on H(X) = L”(R, 9, P) which
* Research supported by Faculty Research Grant, Boston College.
** Research supported by the Ofhce of Naval Research under contract NOOO14-X6-K-0007
0304.4149/89/S3.50 0 1989, Elsevier Science Publishers B.V. (North-Holland)
94 D. Chamher.~, E. S/UC/ / Functional central limit theorem
satisfy U, (X,) = X,, , for all s, t E Ft. Call any element Y of H(X) a “square-integrable
functional” of X. We say that Y (with E Y = 0) satisfies the Central Limit Theorem
(CLT) if
1 (V,(T))-“’ U,(Y) ds: ,V(O, 1) as T-co
where V,(T) = E{Jb U,( Y) ds}‘. Many authors have studied the problem of describ-
ing classes of functionals Y (of stationary processes X) which satisfy the CLT. Sun
(1963, 1965), Cuzick (1976), Breuer and Major (1983) and Giraitis and Surgailis
(1985) obtained progressively larger classes of functionals Y of the form G(X,,)
which satisfy the CLT, and the latter two papers included examples with (r con-
tinuous-singular (and a”“’ absolutely continuous with respect to Lebesgue measure
for some m 3 2). Maruyama (1976, 1985, 1986) and Chambers and Slud (1988)
obtained general sufficient conditions for Y = Ir((X,: t E R)) E H(X) to satisfy the
CLT where Y might depend on infinitely many X,. Other authors (Taqqu, 1975,
1979; Rosenblatt, 1979, 1987; Dobrushin and Major, 1979; Fox and Taqqu, 1985)
have obtained many interesting examples and classes of Y for which the suitably
normalized random variables SC; CJ,( Y) ds converge in distribution to non-Gaussian
limits.
This paper is concerned primarily with analytical necessary conditions on (7 and
Y for the CLT to hold. The functionals Y = h(X) are allowed to depend on arbitrarily
many coordinates X,, and as in Maruyama (1976, 1985, 1986) and Chambers and
Slud (1988), conditions on Y will be expressed in terms of the well-known representa-
tion of mean-0 functionals in H(X) as sums CT_, I& (Major, 1981, Section 6;
Chambers and Slud, 1988, Section 1) of multiple complex Wiener-It6 stochastic
integrals, which have the following properties:
(1”) Each integrand ,f;(x,, . . , xk) belongs to the space L’(cr“, sym) of complex-
valued functions which are square-integrable with respect to k-fold product measure
ah on [w“, are invariant under permutations of their arguments and which satisfy
.&(x) =.fA(-x), where “ ” denotes complex conjugation.
(2”) The k-fold stochastic integral operator v%!Z, is a linear isometry from
L”( d, sym) onto the subspace H,(X) of H(X), where H(X) is a real Hilbert space
with inner-product (Y”‘, Y”‘), = E( Y”’ . Y”‘).
(3”) The subspaces Hk = Hk (X) are mutually orthogonal, and H(X) = RO
@F=, H,(X). The subspace rW@@;“_, H,(X) is the closed linear span in H(X) of
the polynomials of degree m or less in the variables X,, r E [w.
(4”) I,( 1) = X,,, and if fA E L’(rr“, sym) and s E Iw, then
u,(r,(fh)) = I,(c “(‘I+ -‘+ ‘i)fA(X, ) . . . ) XL)).
(5’) If r(t)= I,(&_,,,,) and p(t)- I,(-$(xto,,,-Xr r,o,)), for t30, and if T(t)= I’( - t), p(t) = -p( - t) for t < 0, then r( . ) and p( . ) are real independent mean-0
continuous martingales on [0, m). Defining P(t)-f’(t)+ip(t), we have P(-t) =
p(t) a.s.
D. Chamherc, E. S/d / Functional central limit theorem
It follows from these properties that if Y = C,* ~, 1~ (.fi 1, then
U,(Y) ds = f hh.~), h=I
where
A,r(x) =
and
.,(-r,-,:(I,,‘II,(Y)d~)*=~~,(kl) ‘~~;l~i,,(r)l’~(dx,)-..-(d-i,)
The further properties of multiple Wiener-It; integrals which have previously
been used in proving CLT’s via the Method of Moments are formulas (the “diagram
theorem” of Dobrushin, 1979) for expectations of the form
E[(~,,(.f~,))‘l . . (h,,(.h,,)Y”l.
The approach to CLT’s in the present paper is quite different: our results here rely
on stochastic calculus and Martingale (Functional) Central Limit Theory applied
to the functional j,y U,(Y) ds expressed as a stochastic integral with respect to the
square-integrable (complex) martingale (p(t), t > 0). Our starting point is the follow-
ing pair of known properties of multiple Wiener-It8 integrals (cf. Kallianpur (1980,
Theorem 6.7.1), where the same facts are presented in different notations in the case
where (T is absolutely continuous with respect to Lebesgue measure).
(6”) For each m 2 1 and t > 0, let x:,? denote the L2(~“‘, sym) function x, ,,,,f,‘( . ),
and put xb- I and L’(cr”, sym) _= R. Then, given,f, E L7(a”‘, sym), k 2 1, the functions
,fk, and .ff, defined for t > 0 by
A(u, > . . ,Uh ,)=(fh(UI,...,UI,~I,t)+.f~(U,,...,UI,~,,-r))
xx; ,(~,,~.‘,%I)
and
.fL(u,,..., uh ,I = i(.h(u,, . . . , kI, 1) -.h(u,, . . , cl, -t))
xx;Ll(~,,..., %,I
are L’( a’-‘, sym) functions, and for kz2 the random functions Ik_,(lk,) and
Ik_,(j’&) are adapted and predictable with respect to the increasing right-continuous
a-field family
9, = a({P-null sets}, (p(s): 0s s G t)), t 2 0, (1.1)
I
1
I
x
(7”) IL(&) = I,~,(_&,)~‘(dt)+ 1~ ,(.f?,)p(dt) 0 0
where the integrals on the right are L’-sense stochastic integrals with respect to
square-integrable martingales in the sense of Kunita-Watanabe-Meyer.
Property (6”) is immediate from ( lo)-( .5”), with predictability following easily by
induction on k, and (7”)-which is easy to check for special elementary functions
,fL such as those of Major (1981, pp. 24-25)-is proved by the usual techniques of
L’ approximation.
For simplicity of expression, we extend our notation of multiple Wiener-It6
integrals: whenever g,< : [WA +C is a &square-integrable complex valued function
which is symmetric in its arguments, define
r,(gr)-i{I,(g,~(x)+g,,(-x))-iz,(ig,,(x)-ig,,(-x))}.
Taking expected modulus squared of both sides of the last equation and applying
(2”) to the right hand side, one checks easily that again
Now P(t) defined in (5”) can be expressed as I,(x,,~,,,), and formula (7”) becomes
i
i\
(7’) Ik (.L 1 = 1~ ,(.h(. Oxl;‘,(. ))P(df) > LX>
where the ssX is defined separately on (-CO, 0] and on (0, CO) as a complex stochastic
integral.
The primary consequence which we draw from (6”)-(7”) is that the square-
integrable martingale
II
T Zr(r)= E U,(Y) dsl9, JKVI (1.2)
0 II
has an explicit stochastic-integral form for each fixed Y.
Theorem 1. Let Y =I;‘=, IA(_fk)~ H(X), and let {S,} be given 6-v (5”) and (1.1).
Then jbr t E [0, 11, the stochastic process Z,(t) qf (1.2) is up to equiualence an a.s.
continuous 9,martingale with the stochastic-integral representation
I
Z,(f) = I’
G,(s)P(ds) =2 (I
Re( G,(s))T(ds) ~ I
’ Im(G(s))pu(ds) -I 0 0 1
for t 30, where, in the notation qf (7’), the predictable complex random integrand
GT( . ) is dcjined by
L,h,A~, s)xL,(. )) (a a.e.)
and satisfies G,( -s) = G,(s).
In the next section we apply the structural information of Theorem 1 to the
question of asymptotic normality for T +CD of the random variables Z,(a).
2. Statement of results
Necessary and sufficient conditions are known (Rootzen, 1977; Ganssler and
Hausler, 1979) for the Functional Central Limit Theorem to hold for uniformly
square-integrable sequences of continuous martingales. Of course, sufficient condi-
tions for the CLT are available (McLeish, 1974; Helland, 1982), but no good general
necessary conditions for the martingale CLT are known. For this reason, and because
our main interest is in necessary conditions, we introduce a slightly more restrictive
limit-theorem property for (mean-O) functionals Y =I’;=, I,,(.h) E H(X):
Say Y satisfies the Band-weighteti Central Limit Theorem (BCLT) if the family
of %[O, CO) functions of t is uniformly equicontinuous, and if for every positive
integer m, vector (Y E I?“‘, and 0 < t, <. . . < t,,, - CO, the random variable
Y”J= : IL ; cy,~~(X),f~(X) h-l ( I 1 >
either satisfies the CLT or is such that 7- 2 E iJ U,( Ye,‘) ds I/ Vv(T)+O as T-tacl.
0
There now follows easily from McLeish (1974) and Rootzen (1977)
Theorem 2. Suppose that Y = C, I,(,fk) E H(X) and G,( .) are us in Theorem 1.
(a) 1. I lGAs)I’~(d ) s s 1 as T + ~0 then Y satisfies the CLT.
(b) [f Y satisjies the BCLT and for some E > 0, lim sup,,, EZF’(m) COO, then
I IG,(s)l’o(ds) s 1 as T+ cc.
To clarify the relationship between the CLT and BCLT for Y, we next reproduce
the conditions shown by Chambers and Slud (1988, Theorem 1) to be sufficient for
CLT, indicating also how slight is the additional condition needed to imply BCLT.
Theorem 3. Suppose that Y = CT=, Zk(fk) E H(X) ,forfixed m 2 1, and also
(i) u is absolutely continuous with respect to Lebesgue measure on R, and its
density f is such that on compact sets
(min{.f( . ), kf})““’ -,f *n’ un@rmly as M + ~8;
(ii) O<c (k!) -’ lim inf km’ I iwA
I.I&)l’x~,,, t...+,i,l- ,,,o”(dx), I, 11-o+
1 (k!).‘limsup h ~’ I
w” I~(~)l~x~,,,+...+,~~,- ,,,ck(d-~) < 0~; !. I,-01
(iii) for each k 3 m,
lim sup lim sup h -’ M-oc h-O+ I
rm“ lf,(~)12xc,.~,+...+.~,,-~ /t,~/i(x)~ -.v&W) = 0.
Then 0 < lim inf,,, T-’ V,(T) G lim sup,,, T-’ V,(T) <co, and Y satisjies the
CLT. Suppose in addition that Y satisjes the following two conditions:
(ii’) ,for each 0 s s < t, if
O<c (k!)-’ lim sup h-’ I
I.fi(412xc,., + . ..+T.J- ,,,~k(W, k I, -0 + [-‘,f]~~\r-\,,]l~
then also
Then y satisJies the BCLT.
The necessary conditions of Theorem 2(b) for BCLT are still probabilistic rather
than purely analytical. We provide equivalent analytical conditions in the special
case of homogeneous second order stochastic integrals. Here the uniform-integra-
bility hypothesis of Theorem 2(b) is automatically satisfied, according to the
following proposition.
Proposition 4. Jf Y=CT_, I,(fL)rHM(X)c H(X) for some jinite M, then
{ EZ’~(W): T 2 I} is uniformly bounded.
Theorem 5. If Y = Iz(,f2) E H?(X), then a condition which implies the CLT and is
implied by the BCLT is the,following:
1 IX
I I .fr(x, s).f2(4 -s) (e
iCx+t)T_ I)(eitu-t)T_ 1)
V,(T) m.~xfli*l. ldt (s-u)(x+s)
(e i(u+r)T
+.M.% -s)f2(5 s)
_ ,)(,i(.x-*IT_ 1) 2
(u+s)(s-xx)
q(ds) + 0 (2.1)
as an L’(R’, a’) function qf (x, u) as T+w.
Since V,,(T) must necessarily +CO in order that (2.1) hold (without fi vanishing
identically), the expressions (2.1) depend essentially only on the measure q and on
the behavior of f?(u,, u2) in the neighborhood of the diagonals {u: U, = *u?}. For
calculations showing that (2.1) holds under some conditions when u is absolutely
continuous, see Chambers and Slud (1988, Lemma 2.3).
,- (~PMW(~)~ =
0 I
(~Pmo~(~-&v ‘- = 0 I
(nP)d(n-)S(n)bv () !I
0 = [(SP)d!- (SP).Il(S-)~(S)~
I I
uaql .( np)d(n-)/ ’ ,,J p%~u! aql 01 pue.G’a~~! aIq”wpald _ ql!M (nP)d(n)S:‘J LKIOJ aq1 JO ~“.hIalu! 3!lsEq3ols I? sLuJoJsuml n- = n a[q”‘.l”A
-JO-a%Eq3 aq1 1eq1 11ma.l pu\? ‘(S)N E(S-)n Lq s1uauInELIe aiz@au .IOJ (.)fi
auyaa ‘(s)~p (s)~i ‘,)J+(s)wp (S)M ‘:J s! (r)~l(r)w JO md a@u!mm aql aDu!s _ _
1 [(~P)~((~-)~-(~)~)!+(~P)~((~-)8+(~)~)1(~~~ ‘)
SI atI zs
‘(“)P J(s-_)8 - (s)S) ‘I +‘(J>p J(s-)Z + (s)Z) (’ J J =,IWVl , ‘(E’Z’P rua.IoaqL ‘0861 ‘Inddn!lle)l) murua~ s6g1
JO LLLIOJ [maua8 aql 1(8 .sa[%g.mu alqwSalu!-amnbs Ivuo8oql~o OMJ JO ulns aql s!
(sp)rl((s-)8-(s)S) Cl J
!+(sP)J((s-)8+(s)S) ” = I I J
(SP)d(SF ‘- J = (r)Lw I ‘0 e 1 .IOJ Ir?ql MOUy aM ‘JOOld
.(Ylp)d((n)s(S-)s+(L)g(s)s) ‘;l= (S)l/ ‘0~s .Wj‘pWJ (S-)L/ = (S)y c%XjM
(I’E) (SP)d(S)Y ‘- J
+(~Pb$lw ‘- J I = (sp)d(s)8 ‘-
I I z I II 0-C) -q ua$y ‘cc > (sp)n,j( s)Z/ J 3 my1 ynns pus ‘0 e I rof a~qw13.‘pand-‘,s aw (I -)8 puw
(r)8 yioq iwql yms ti uo uo~rnunJuropuw.4 paywn-xa~iuon kuw aq ( . )8 la7 ‘9 eurwaq
‘([I ‘o])$ = K
[I)” JX~+Wf_ ‘1 EJ = (j),dzJ = ‘(d)
c 11
‘([I ‘0,)~:=;1(“.“-Ix;)llia = (1)JY = ‘(J)
Lq uaA!8 ale pm u~opue.~uou ale (?d pue in Jo
slolesuadmo3 laiCan-qooa aql ‘-a.!) rl pua ,I JO sassaDold awv!.uzA alq-empald ayl
‘sluatuamu! (ue!ssnef> iC[lu!o[pue pal%?[aJlo3un asnmaq) luapuadapu! ql!M (cx’ ‘01 uo
sa~t?~u~l.‘??~ luapuadapu! am rl pue J pm ‘( 1.1) JO ‘g uo!lr?~l~y aql 01 1DadsaJ ql!M
(m ‘01 uo a@u!l.mu ue!ssnt?g xa[dwo3 I? s! ( . )d! + ( . ),[ = ( . )d ‘(,g) pue (J) 1cg
‘paxy ale D ammaw [tz.wads ~!uroleuou aql put? x ‘uog3as s!ql lnoqSho_tql T/ ‘J
‘d 01 lDadsa1 ql!M sle.Salu! 3gseq3ols lnoqe suogmap!suo3 Iwaua% ql!M u$aq aM
100 D. Chambers, E. Slud / Functional ceniral limit theorem
so that
M(.s)[(g(s) + g(-s))r(ds) +iMs) -d-.~))p(d.~)l
Now using (r), =(p), = $a([O, s]) and collecting terms, we find
,M(t),‘= 1’ ,g(r),%(d.i)+[ fi(s)g(s)P(ds)+ M(s)g(.s)P(ds) PI I I’ -I
I ’ Ig(s)12g(ds)+
I ’ = [ti(.s)g(s)+M(s)g(-s)]P(ds).
-I --I
Since h(s)= M(s)g(s)+g(-s)A4(s) satisfies h(-s)=h(s) and is equal for ~20 to
d.7) I‘. g(u)P(du) + g(-s) I‘ du)P(du) -5
’ zz k(s)g(-u) + G(-.s)g(u))P(du) -\
our assertion is proved. 0
Corollary 7. If’g( .) is a complex-valued randomfunction such that g(t) is predictable
and g(-t) = S(t) for t 3 0 and such that E 1‘7 Ig(s)12rr(ds) < CO, then the martingale
M(t) = J’_, g(s)P(ds) is reul-valued with predictable variance (M)(t) =
j’, Ig(s)l’@ ) d s an with variance EM’(t) =s’, Elg(.s)\‘m(d.s).
Proof of Theorem 1. Define g(s)- I,,_,(P~,,(. , s)xk!,( .)), and verify that g(-s) =
g(s). Apply (7”), (7’) and Corollary 7 to conclude that
I
IX
WJ~,.))2= ElL,(pk,,(., s)x!~,(.))124ds) -X
=((k-l)!))’ I
X’ IIP~,T(., 4x~!,(.)l/;/~ Ia( 1-
Therefore the orthogonal sum
G(s) = ( VY( T)) 1/z? IL-,(/%.,(., 4xiw))
must exist in L2 sense in N(X) for u-a.e. S, since the corresponding sum of L”-norms
( = expected modulus-squared) is
E(G,(s)I’=( V, (T))-’ 5 ((k-l)!)m’IJp,,,(., s),&!,(.)llf,~ 1
D. Chambers, E. S/d / Funcfional c.en/ml limit theorem 101
the integral of which with respect to a(ds) is
(,,,~))~i,(;,(~~,~))‘=(v,(~))~l~~j~ WY)dr)i=l.
Now apply (7”) to pk.-r/m and sum over k to learn that
I
IX
Z,(m) = G,(s)P(ds) -\-
=I
f u
G,(.~)P(ds) + G(s)P(ds) + -I I, I
--I
G,(s)P(ds), t > 0, -Ix
from which it follows easily that E{Zr(~)l~,} =I’, G,(s)P(ds) a.s. Since the last
stochastic integral is a.s. continuous in t, Theorem 1 is proved. q
Proof of Theorem 2. (a) By Corollary 7 applied to the stochastic integral
Z,(t) = I
,
G(s)P(ds), (Z,)(t) = ’ IGT(S)]‘g(d.s). -I I -I
If (Z,)(a) s 1 as T + a, then the CLT for Y is an easy consequence of the Rebolledo
Martingale CLT (Helland, 1982, Theorem 5.1(a)).
(b) Suppose Y satisfies the BCLT. Then Vy ( T) + 00 as T + 00, so that for 0 < s < t,
(V,(T))-’ I’ b,,Ax)12dd4 \
’ s (2/s”)( V,( T))-’ I,f,(x)]‘v(dx) + 0 as T+KJ.
The assumption of uniform equicontinuity in the BCLT implies by the Arzela-Ascoli
Theorem that for every sequence of numbers T increasing to ~0, there exists a
subsequence of numbers T’ along which
f ’ k-2 k! V,( T’)
IPl;r,(x)(“x:(x)~(dx,) . . q(dxk) + v( 1) as T’-+ cc (3.2)
uniformly for t E [0, an], where u(t) is some nondecreasing uniformly continuous
function on [0, M), and where
{I
T u(a)= lim E u,( Y- I,(_f,)) ds
I’/ V,( T’)s 1.
7--X 0
For the rest of this proof, we will restrict T without further comment to lie in such
a subsequence T’, and we will prove that (Z,)(CO) s 1 as T-CO.
Fix 8 > 0 arbitrarily small. From the previous paragraph, we learn that
E{Zr(r)-Z,(S)]’
= A:, k! V;,(T) j- (x;(x) -x:(x))lp~~.(x)l’cr(dx,) . . cr(dxk)
+v(t)-v(S) as T-,a, uniformly for tE[S,m].
102 D. Chambers, E. Slud / Functional central limit theorem
From the martingale property of Z,, it follows that, as T + 00,
whenever t, > 6. Since
c
I
I
,,I I/,( Y,,‘) ds ( V,(T))“’ = C ~,Zdf,),
J o I ,=I
the BCLT implies for each (Y, t with t, 2 6 that
; a,(Z,(,;)-Z&& ,V(O, R8(q t)) ;=I
as T+co
where the distribution ;V(O, 0) is understood as the distribution degenerate at 0. Then the Cramer-Wold device (Billingsley, 1968, p. 49) implies that the finite-
dimensional distributions of the martingale Z,( . ) - Z,( 6) on [ 6, a] converge weakly
to those of W( u( . )) - W( v( 6)) where W is a standard Wiener process.
In order to conclude for fixed 6 > 0 that
Z,(.)-Z,(S): W(TJ(.))- W(v(6)) in %‘[S,co] as T+oo (3.3)
we need only to verify tightness of the laws of JT(. ) = ZT( .) -Z,(6) on %‘[6, ~1,
which we do as follows. For each /3 > 0 and positive integer M, and each 0 < F’< F,
(Doob inequality), where v-‘(s) = inf{x: u(x) > s}. Now the finite-dimensional dis-
tributional convergence of 17( .) as T -+a, together with the uniform integrability
of {]i,‘r(~)]‘+~‘: 6 s s 4 CO}, implies that as T+ ~0 the last expression converges to
which is bounded above by a constant (not depending on M) multiplied by M mmF”2.
It follows for each p > 0 that M, T,,< cc can be chosen so that, for all Ta T,,,
P 1
sup 5, rt[fi,X]: ~I.O)~L.(I)# Iv-’
lZAr,-z,Ol;-3P} <P.
By a small extension of a well-known criterion (Billingsley 1968, Theorem 8.2),
tightness of the laws of Zr( .)-Z,(S) on %[6, ~1 follows.
D. Chamhtm, E. Slud / Functional central limit theorem 103
Since {Zr( . ) - Z,(S)}, _, is a uniformly square-integrable family of continuous
martingales, it follows from (3.3) and Rootzen (1977, Theorem 6) that
(ZT-ZT(S))(~)~ u(CC)-u(S) as T+oo.
But it is easy to see that (Z,( .) -Z,(~))(W) = (Z,)(a) -(Z,)(6), since the quadratic
variation [Z,]( .) and predictable-variance processes (Z,)( . ) necessarily coincide
by continuity of Z,( . ). Thus we have proved
for each 6>0, (Z,)(m)-(Z,)(S): u(a)-~(6) as T+co.
In addition, for each 6 > 0 we have seen that
lim I
8 IPU( s lP1,r(x)12 T-X _A V,(T)
g(dx) = lim I V,(T)
a(dx) = 1 -u(a). r-s -v
Therefore, since u(O+) = 0, there must exist 6 = 6(T) converging to 0 slowly enough
as T+co so that
and
i7’r) IPI,TbN2 -A(T) VY( 7-1
u(dx) + 1 - ZI(CO). (3.4)
But the martingale
-G(f)- ,, $f+$$ P(dx)
has variance equal to the left-hand side of (3.2), and must therefore converge to 0
as T+ ~0 if f is replaced by 6(T). Moreover, the martingale
has predictable-variance process at t = S(T) equal to
I
ii(T)
I~,,~C#~bWI VY( T).
-ii(T)
We conclude by (3.4) that (Z,)(6( T)) converges to 1 - ~(00) in probability as T+oo,
and then that (ZT)(~) Jk 1, as was to be shown. The observation that (Z,)(a) =
JTr IG,(s)12a(ds) completes the proof. 0
Proof of Theorem 3. The assertion that (i)-(iii) imply the CLT with asymptotically
linear variance was proved as Theorem 1 of Chambers and Slud (1988). The same
theorem will apply to Y a,’ in place of Y, with assumptions (i) and (iii) and the
upper bound in (ii) unchanged, and with the lower bound (ii’) implying
i I ,,I
O<c (k!))’ lim inf h-’ I /l-(1+
~Ai ,;,, ~,x~“(x)~;(x)~~x(,,,+. .t\‘,~ /.,, ,&dx)
whenever the same expression with lim inf replaced by lim sup is positive. Thus
(i)-(iii) and (ii’) for Y imply (i)-(iii) for Ye,‘. Theorem 1 of Chambers and Slud
(1988) implies the CLT for each Y@ for which
7 ?
lim sup E (I
U,( Ye.‘) ds - I/
V,(T)>O. T-X 0
It remains to prove only the assertion of uniform equicontinuity in the BCLT. For
this, it suffices to prove that the nondecreasing, uniformly bounded, and uniformly
continuous functions
(FT(t)= $ (V,(T)k!))’ I, -2 I d
x;(x)IpcT(x)l’~(dx,) . . . ddx,\)
on [0, 001, satisfy, for all s < f,
V,(T)<2 ’ lim sup ((PAN) - (F~(s)) T-
1. y(x)ddx)
r-K (3.5)
where y( . ) is as defined in hypothesis (ii”) of the theorem. Observe first that for
each ks2, by Lemma 2.1 of Chambers and Slud (1988),
lim sup 3 7+.x I
Iw,~ (x;(x)-x;(x))lp~,,-(x)l’o(dx-,) . . d&J
Therefore
v,_(T) lim sup ((PA?) - PAS)) 7
T-u-
(x:(x)-x;(x))lp~,~(x)I’~(dx,) . . . ddx,)
I
s2 y(x)a(dx)
D. C‘hnmhrrc, E. Slud / Func/ional wntral limit theorem 105
by definition of y. Since lim inf,,, ( V,(T)/ T) > 0 under our assumptions, and
since all the functions (oT( .) increase from 0 to a limit less than or equal to 1, it is
easy to deduce uniform equicontinuity from (3.5) by considering the behavior of
(p7 ( .) along finite increasing sequences { tj}ylo for which 5:; ‘I y(x)a(dx) are all small,
where t,, = --Qc: and t,, = ~0. 0
Proof of Proposition 4. As in formula (4.2) and Lemma 3.7 of Chambers and Slud
(1988),
uniformly in T. q
Proof of Theorem 5. Now Y = Iz(,f2), G,(s) = (V,( T))m”‘I,(p2,T( ., s),$,‘( .)), and
Z,(t) =I’_, G,(s)P(ds). Now function G,r (s) takes the special form
of a first-order stochastic integral with nonrandom integrand. Then by Lemma 6,
Jrb, s)P(dx) (3.6)
where the adapted random function JT(. , s) satisfies Jr( -x, s) = .i, (x, s) for 1x1 c s
and is given by
Ill
JT(-T .s) = (V,(T)) ’ I 1 .mx, .~lh(U, -s) (e il\l\)t-_l)(ei(“~‘)r_l)
-IX (x+s)(.s-u)
+.Mx, -sY2(4 .s) (e Illl+<)t _ I)(e”“‘r_ 1)
(u+s)(s-x) P(du),
and where we have used the identity ,f2(-x, s) =,f?(x, -s) in two places.
In the present setting,
2 lp2,,-(x, s)l’cr(dx)a(ds)/ V,,( T) = E * IG,(s))“o(ds) _X
= EZ$(co) = 1.
Since Z,(t) is a continuous .F,-martingale with quadratic variation ( = predictable
variation) process J’, IG,(s)l’cr(ds), it follows from Millar’s (1968, Section 6)
extended Burkholder inequalities that for a universal constant C, for each T
E {I: ,G7 (s),%(ds)} s CEZh,(m).
106 D. Chambers, E. Slud / Functional central limit theorem
By Proposition 4, {j’“,, JG,(s)\‘rr(ds)} . . 7 , IS uniformly square-integrable. Thus the
condition which according to Theorems 2 is sufficient for CLT and necessary for
BCLT becomes
P, 1.2 &(x, s)P(dx)a(ds) - 0. (3.7)
But the left hand side of (3.7) is the stochastic integral
with expectation squared equal by Corollary 7 to
1
I_ 15 ,x 2
E J,-(x, s)a(ds) g(dx) a 1x1
Using the stochastic-integral formula
I’
J,(x, s)c(ds) IX/
(3.8)
and applying Corollary 7, we find
cc 7
E J,(x, s)dds) II
Pz,r(% S)PZ.T(% -.y) +P*,r(x, -~)Pz,,(U, 3) 7
V,(T) a(ds) q(du).
Noting that the last expression in 1.1’ is symmetric in u and x, we find that the
double integral with respect to g(dx)a(du) over R” of the left hand side of (2.1)
is exactly twice the expression (3.8). Thus (3.7) is equivalent to the condition (2.1)
of our theorem. 0
Acknowledgement
The authors are grateful to C.Z. Wei for a helpful conversation related to the
tightness in Theorem 2, and to the referee for suggestions which improved the clarity
of the presentation.
107
References
P. Billingsley, Convergence of Probability Measures (John Wiley, New York, 1968).
P. Breuer and P. Major, Central limit theorems for non-linear functionals of Gaussian fields, J. Multivar.
Anal. 13 (1983), 42.5-441. D. Chambers and E. Slud, Central limit theorems for non-linear functionals of stationary Gaussian
processes, to appear in Probability Theory and Related Fields ( 1988). K.L. Chung and R.J. Williams, Introduction to Stochastic Integration (Birkhauser, Boston, 1983).
J. Cuzick, A central limit theorem for the number of zeros of a stationary Gaussian process, Ann. Prob.
4 (1976), 547-556. R. Dobrushin, Gaussian and their subordinated self-similar random generalized fields, Ann. Prob. 7
(1979), l-28. R. Dobrushin and P. Major, Non-central limit theorems for nonlinear functionals of Gaussian fields, Z.
Wahrsch. Verw. Geb. 50 (1979), 27-52.
R. Fox and M. Taqqu, Noncentral limit theorems for quadratic forms in random variables having
long-range dependence, Ann. Prob. 13 (19X5), 428-446. P. Ganssler and E. Hausler, Remarks on the Functional Central Limit Theorem for martingales, 2.
Wahrsch. Verw. Geb. SO (1979), 237-243. 1.1. Gihman and A.V. Skorohod, The Theory of Stochastic Processes (I, II) (Springer-Verlag, Berlin,
1974, 1979). L. Giraitis and D. Surgailis, CLT and other limit theorems for functionals of Gaussian processes, Z.
Wahrsch. Verw. Geb. 70 (19X5), 191-212. I. Helland, Central limit theorems for martingales with discrete or continuous time, Stand. J. Statist. 9
(1982), 79-94.
G. Kallianpur, Stochastic Filtering Theory (Springer-Verlag, New York, 19X0).
R.S. Lipster and A.N. Shiryaev, Statistics of Random Processes, I (Springer-Verlag, New York, 1977).
P. Major, Multiple Wiener-It; Integrals (Springer-Verlag, Berlin, 1981). G. Maruyama, Nonlinear functionals of Gaussian stationary processes and their applications, Proc.
Third Jap.-U.S.S.R. Symp. Prob. Th. (Springer-Verlag, Berlin, 1976), 375-37X.
G. Maruyama, Wiener functionals and probability limit theorems I: the central limit theorems, Osaka
J. Math. 22 (19X5), 697-732.
G. Maruyama, Wiener functionals and probability limit theorems II: term-wise multiplication and its
applications, Hokkaido Math. J. 15 (19X6), 405-451.
D. McLeish, Dependent central limit theorems and invariance principles, Ann. Prob. 2 (1974), 620-62X.
P. Millar, Martingale integrals, Trans. Amer. Math. Sot. 133 (1968), 145-166.
H. Rootren, On the Functional Central Limit Theorem for martingales, Z. Wahrsch. Verw. Geb. 3X (1977), 199-210.
M. Rosenblatt, Some limit theorems for partial sums of quadratic forms in stationary Gaussian variables,
Z. Wahrsch. Verw. Geb. 49 (1979), 125-132.
M. Rosenblatt, Remarks on limit theorems for nonlinear functionals of Gaussian sequences, Probab.
Th. Rel. Fields 75 (19X7), I-IO.
T.C. Sun, A central limit theorem for nonlinear functions of a normal stationary process, J. Math. Mech.
12 (1963), 945-97X.
T.C. Sun, Some further results on central limit theorems for nonlinear functions of a normal stationary
process, J. Math. Mech. 14 (1965), 71-85. M. Taqqu, Weak convergence to fractional Brownian motion and to the Rosenblatt process, Z. Wahrsch.
Verw. Geb. 31 (1975), 287-302. M. Taqqu, Convergence of iterated process of arbitrary Hermite rank, Z. Wahrsch. Verw. Geb. 50 (1979),
27-52.