Journal of Theoretical Probability, Vol. 9. No. 3. 1996 The Law of the Iterated Logarithm and Marcinkiewicz Law of Large Numbers for B-Valued U-Statistics 1 Zhonggen Su 2 Received December 8, 1994: revised Jcmuary 8. 1996 Suppose that B is a separable Banach space and (S, 6a, P) a probability space. H is a measurable symmetric kernel function from S" into B. In this paper we shall further study some limit theorems for B-valued U-statistics U",,H based on P and H. Special attention is paid upon the Marcinkiewicz type law of large numbers and the law of the iterated logarithm. Our results can be regarded as extensions of corresponding results for sums of independent B-valued random variables to U-statistics. KEY WORDS: B-U-statistics; decoupling inequality; law of large numbers; law of the iterated logaritlnn. 1. I N T R O D U C T I O N AND PRELIMINARIES Let (S, 6:, P ) be a p r o b a b i l i t y m e a s u r e space a n d X, X', Xi, i/> 1 be i.i.d. ( P ) S - v a l u e d r a n d o m variables. Let B be a s e p a r a b l e B a n a c h space w i t h n o r m I1"11 a n d its d u a l B' a n d let H(Xl ..... Xm): S " ~ B be a m e a s u r a b l e f u n c t i o n s y m m e t r i c in its a r g u m e n t s , t h a t is, H(x~ ..... x,,,) =H(xo~ ..... x~,,,) for a n y p e r m u t a t i o n a o f (1, 2 ..... m). As u s u a l , the B - v a l u e d U-statistic b a s e d o n X;, 1 ~< i ~< n, a n d k e r n e l f u n c t i o n H is d e f i n e d as 1 U.,H--(-;~ I1 ~ - - ,i)), I <<. i) < . . . H(X~, ..... X,,,,)< ~,,, -< n (n--m)[ n! ~ H(X~, ..... Xim) ( I . I ) I',',, " w h e r e I',:,= {(i~ ..... i,,,): i,<~n, i , ~ i , if r :/:s}. ~Research supported by National Natural Science Foundation of China and Zhejiang Province. 2 Department of Mathematics, Hangzhou University, Zhejiang 310028, China; Institute of Mathematics, Fudan University, Shanghai, 200433, China. 679 0894-9840/96/0700-0679509.50/0 9 1996 Plenum Publishing Corporation 680 Su Our main interest is in the limit theory for U~',,H. As it is well known, many classical limit theorems for sums of independent real random variables have been generalized to the real U-statistic case (see Serfling, ~11 and Gin~ and Zinn~4~). Thus, naturally, we wonder whether some beautiful limit theorems for sums of independent B-valued random variables have analogues for the B-valued U-statistics. The central limit theorem (CLT) for B-valued U-statistics U",,H has been presented in Arcones ~-'~ and completely understood in the case of Banach spaces of type 2. Arcones and Gin+ ~3) also have obtained that the law of large numbers for B-valued U-statistics U'~,H holds as long as E [[H(X l ..... X,,)[[ <o0. Their proof mainly relies on the fact that [[U',',H--EHI[--*O a.e. is equivalent to E [1U',',H- EH[[ ---,0 since [IU~',,H- EHI[ is a reversed submartingale. However, this fact is no longer true for the Marcinkiewicz law of large numbers (MLLN). The purpose of this paper is to further study some other limit theorems for U',',H, for example, the M L L N and the law of the iterated logarithms (LIL). Our main results given in details in Section 2 show that the two strong limit theorems for B-valued U-statistics are similar than for sums of independent identically B-valued random variables. An important tool in deriving the asymptotic theory of U-statistics is the Hoeffding's decomposition, i.e., the projection of a U-statistic on the basic observations of the sample. We give it here together with some notation. The operator n ke. , , -_- n k.... acts on P"'-integrable functions H: S"'--* B as follows: rck.,,H(x I ..... X k ) = ( J . , . ~ - - P ) . - . ( f x k - - P ) P .... kH, where Q l " " Q,,,H= ~ H ( x I ..... x,,) d Q l ( X l ) . . , dQ,,(Xm). Note that z~..... is a P-canonical function of k-variables, that is, PrCk.,,H(', x,_ ..... X k ) = 0 for almost all x,_ ..... Xk. Hoeffding's decomposition is as follows: for all P"'-integrable functions H: S " --* B, (1.2) k=l The first term in the right-hand side is just m/n~'i'= l nl.,,,H(X ~) so that we can readily apply various limit theorems for sums of independent identically B-valued random variables, which have been nearly thoroughly discussed in Ledoux and Talagrand's bibliography) s) Thus the key to our study is how to deal with those remainders. Randomization by Rademacher variables plays a role in B-valued U-statistics similar to the role it plays for sums of independent B-valued random variables due to the fact that B-valued U-statistics can be decoupled. We state later several pertinent results. Let now {XI k), i/> 1} ~-'=l be m independent copies of { Xg, i >/1 }, {e~, i >/1 } a Rademacher Iterated Logarithm Law and Marcinkiewicz Law of Large Numbers 681 sequence independent of {X;, i~> 1} and ""'*) ~ j , i>~l }~'_ _j m independent copies of {e;, i >I 1 }, independent of the variables { X~k~, i >1 1 } ~'= ~. Lemma 1.1. (de la Pefialg)). ing function. Then Let ~b: [0, Go) --, R be a convex increas- E~ (c, ~g;,H(X~,') ..... Xi,,,'"')) )~< E~b ( ~ H(Xi, ..... X;,,,) ) <~E~ (ca ~ H(X,~ )..... X,,:,')) ) (1.3) elleli ".e. I.o H(X r ). ,X~," ')) ) g*, / If H is P-canonical function, then E~ cI --.--i, ,,) ..... Xi,,, ) <~Ecb Ill) \ il """ Here Cl, c2 are positive constants depending on m (independent of ~, n, I1"11) and possibly varying from line to line. Take ~b(x)=x'-, the equivalence described in Lemma 1.1 allows us to use the defining type 2 and cotype 2 inequalities. Remark 1.1. It is known that if B is a separable Banach space and F is a subspace of B, then for some countable subset D c B' we have qF(X)=SUpf~n If(x)[ for all x in B, where qr(x) is the seminorm defined by qF(X) = inf{ IIx- yll, y ~ F}. Therefore, (a) and (b) remain true when the norm []-II is replaced by qF" Lemma 1.2. (Gin6 and Zinnt6)). of B. (a) Let K be a convex symmetric subset If Dj, j = 1, 2,..., rn are subsets of { 1, 2 ..... n}, then P s (il i,<_/./ .-,., ira) ~ DI ..... x 99 9 x Dm k_<..... 2"'- 1 \ ) 682 Su (b) \ 1nl < ...... -.~2 (2 - 1 ) m a x e H(X~, ..... Xg,,,)~K" ) Lemma 1.2 is an elementary but useful fact. It indicates that the tails of the original U-statistic can dominate the tails of a decoupled, randomized version of it. As noticed by Gin6 and Zinn, t6) it would be interesting to have inequalities analogous to those in L e m m a 1.2, but in the opposite direction. We refer the reader to de la Pefia and M o n t g o m e r y Smith c~~ for the comparison between the two tail probabilities. The following Hoffmann-Jogensen type inequality will help us treat integrability in some cases. L e m m a 1.3. (a) (Gin~ and Zinn(5)). There exist finite constants ct(p), c2(p) and c3(p) ~ (0, 1 ) such that /) .,(t) '("')H (I) X(,,,)) E ~ el, " " ei,,, ,--it ,'", i,,, I',',, max -< c t~' + c , E I 0 - 1 : ~ i m ~< tz ~' (I) Ira) (I) ei~ ...el,,, H ( X i , ..... m) ) Y (im P (1.7) i l ...-, i r a - I 9 . n ( II ,..., Iij)) E I m where to is any number satisfying P e(il)...ei,)) H(Xii ..... >to ~<c3 n (b) There exist finite positive constants c~(p), c~_(p) and c 3 ( p ) ~ (0, 1) such that E .... i 11 ~ $iH(Xi ' X(I I )..... X (.... ) I' i=l <~clt~+c~_E max IIH(X,, X(~~).......Y ( ,,,_~ .... I))llP I ~i~<n (1.8) Iterated Logarithm Law and Marcinkiewiez Law of Large Numbers 683 where to is any number satisfying c i H ( X i , x I ..... c . . . . supP .v >to ,)) ~<c3 I i with the supremum taken over all x = ( x ( ..... x .... ]) ~ S .... t E ~ ~.,(1) il ' " " -e- i .m(- " I ' - l ) H ( X (l il )~ ... ' 'Y ' ' i m. . .- .I <~C)t~+CzE R( I ) ~, max l <~ i m _ X') I, ]) ' l <~ lt . !1 . - - . -i, . . .E (m-2) i,,,_, . Iin - ] ii ( il ..,,, i m - I ) E I m _ 1 X H t~X (]) Y( . . . . I il '"'~ "'Sn- ]) ' X') p (1.9) where to is any number satisfying \ supP ~ .(I) . . . e- l- i r.... t ) H()t.'i,) (] ' . . . . ,. y ( ,i m, , -- lI) a- I Gil x) >to)~< c3 / with the supremum taken over all x ~ S. We only prove Eq. (1.9). Define Proof of (b). d U II - - y , g~(t, il " " .F~. ....I ])H(X(i.]) --tin- "'"" X ' i r.... I),X') a- 1 Z;'n- I M d I# m max ~<t~ l ~<im-I . ~. e(])..-~ (.... "-)Ht~ "(t) X I. . . . il vim-2 "'~'" il ~ ' " ~ itn-I ~) i X') II ,..., I m - 2 n ( il ..... i m - I ) E I m _ I Let us recall Theorem 3 o f Gin~ and Zinn. (5) It shows that, first conditioning on X' and then taking expectation E' with respect to X', we find finite positive constants c t and c2 such that for all t > 0, P(IIUT~II>3 .... ' t ) < ~ c . g ' [ P ( l l U T ( l l > t ) ] z " - ' / 2 " - ' - ' - 4 - c 2 e ( M ~ ( > t ) (1.10) 684 Su The constants can be taken to be c~ = 2 3~ 6 2+2/3+ and c2 = 1 + 3"-. 4 + . . . + 32''-'/~ - t~ct. Now letting U,d," ' = ~ e ~l) .~(.... l)H~y(l~ il "" V i m - I ~ ' ' il ~'"~ +2"-1/'2"-I--1) ' (.,-II, X ) Xim-I l ~n~-- I for any to > 0 we have EllU,~llt,<<31,,-I)~,to+ 3 ~.... ~lpf p(llUall>3 .... ~t) dt p "t 0 <~3(,,,-~Pto+ 3~,,-~)pc,_ ~, P(M. d > t) dt p o + 3' .... I~PcI f E ' [ P ( IIU,'~II > t)]'-"-'/~2"-'-l, d t p " to <~3' .... l lPt o + 3' . . . . I ) P c 2 E ( M ~ : ) P + 3 (.... I)p Cl s u p P ( [I --n t]'d'x .[I > Io) 1/{2"-l- 1) 9y~ P(II x U,'~ll > t ) d : o (1.11) Taking to is any number satisfying ( 1 )2,,,-,sup.,.P( IIU~)" II > to) -< 2.3' .... 'IPc, I :: C3 then Eq. (1.11) is bounded by 3' .... ')rt 0+ 89 IIU~IIp+3` .... ')Z'c2E(M~) ? ~<2 9 3~"'-')Pto+2 -a( - ' .... l)p,,t" 2 x,~t ' ~ . a~,c,t~? t1 III The required result follows. All these lemmas are of course repeatedly used in the proof of our main theorems. But in order to study strong limit behavior we need to establish the maximal inequality for B-valued U-statistics and estimate moments of sum of array of random variables. Our main ideas stem from de Acosta, ('~ which studied the MLLN for sums of independent identically random variables. Iterated Logarithm Law and Marcinkiewicz Law of Large Numbers 685 2. MAIN RESULTS A N D P R O O F S Let us start by describing what can be understood by a LIL for B-valued U-statistics. Set LLt = L(Lt), Lt = max(l, log t), t >/0, and a, = x/2nLLn. As in the case of sums of independent random variables, we will say that the bounded LIL holds for U:,H if A =: lim s u p : .... x/2LLn 11U::,H-EHII is a nonrandom finite number. Unlike in the independent sums case, A is not necessarily nonrandom even though x / ~ / ~ l l g " , , g - - E H l l is almost surely bounded. This is mainly because we cannot use the Kolmogorov's zero-one law. We might as well say that the compact LIL holds for U2,H if there is a compact symmetric K in B such that, almost surely . l. i. .m d ( ~v/~ ( U;;,H - EH), K ) = 0 and C [~(~ x//'~7 U;:,H--EH)} = K where d(x, K) = inf{ IIx - yll, y e K} and where C{ x,,} denotes the set of the cluster points of the sequence {x,,}. K is called the limit set. In addition, we say that a family of sequences of { T,(x), n >>.1, x e S} of random variables tends to zero uniformly in probability if for any e > 0 lim sup P(IIT,,(x)II > ~)=0 rl~ r xES The introduction of this concept enables us to repeatedly use Lemma 1.3(b). Now we give our main results about the LIL for B-valued U-statistics. Theorem 2.1. Suppose that B is a separable Banach space and H: S"' ~ B is a measurable symmetric kernel function. Let { Am, n >i 1} is a sequence of independent copies of X. If (1) E IIH(X~ ..... X.,)II'-/LL IIH(X~ ..... Xm)ll < O0, (2) a'-=sup;~BiE(rcl..,foH(X1))'-<oo, 860/9/3-10 686 Su (3) 1 )' H(Xi, x I ..... x . . . . 1) an i=l 0 uniformly in probability 1 ~/zL ~,/z-;Lnn'(n-l'E 1 <<.i~j<~n 0 v/n H(X,,Xj, xt ..... x .... 2) uniformly in probability (n-m)! ~ " n---~. Z H(Xi, ..... X,,,,) r,', --, 0 in probability then ,lim x/n IIU;:,H-EHII=m~ a.e. (2.1) If condition (2) is replaced by (2)' {(nL,,,foH(XI)) 2, f E B ' , Ilfll ~< 1} is uniformly integrable, then the compact LIL holds for U',',,H with the limit set inK, where K=K,,,.,,n~x~ is the unit ball of the reproducing kernel Hilbert space associated to n L .,H( XI ). Proof In order to simplify our notations and not to obscure the main scheme, we only consider the case m = 2, since the general case is similar. Assume that {X i,, n >/1} is an independent copy of {X,,, n >/1}. Keep in mind Hoeffding's decomposition U~H-EH=2U'~zl,_H+ U'~2, 2H (2.2) We'll complete the proof of Theorem 2.1 in two steps as follows. Step 1. The following assertions are established ,}im_ v/~ ~/2LLn: [IU~'rcl ~HII_ =,lim 1 " --a,, ;i~ n,.2H(Xi) =a a.e. (2.3) =K (2.4) and n1.2H(Xi), K =0, n~o~ i=l ;r i=1 Iterated Logarithm Law and Marcinkiewicz Law of Large Numbers 687 According to Theorem 8.6 in Ledoux and Talagrand ~s~ for Eqs. (2.3) and (2.4) it is enough for us to check the following two conditions E II~,, 2H(XI)II 2/LL I1~,.2H(X1)I] < oO (2.5) and 1_ lrl.2H(Xi)~O an i = l in probability (2.6) Indeed, since ~ ( t ) = t Z / L L t is convex and increasing for t > t o large enough, there exist a positive number to and a nonnegative convex increasing function ~b0(t), t > 0, so that ~bo(t) <~~k(t) <~to + Cko(t) Noticing that r~l.2H(xl) = EH(xl, X2) - EH(XI, X2), we have E~k(llTr,.2H(Xl)ll ) <~to + E~o( llnt.2H(X~)ll ) ~< to + E~bo(2 Iln(xl, X_,)II) <~to+4E(~(IIH(X~, X2)II) < oo It remains to prove Eq. (2.6). Obviously Eq. (2.6) will be valid if lim --1 E ~ nl.2H(X i) = 0 n ~ c'~ s i=l On the other hand, E IlYT/=17r ~. 2 H(Xi)I1 ~<2E IlY.7= ]e~H(X~, X')II. So, we only need to show n ,lim 1 E an ~ e,H(X,,X') = 0 i (2.7) I Note that E IIH(Xl, X2)II2/LL lIB(X,, X,_)ll < oo implies lim --1 E max IIn(Xi, X')ll = 0 n~ oc a n i <~i~n by a simple integration by parts. In view of Lemma 1.3(b) this indicates that Eq. (2.7) will follow if 1_ an eiH(Xi, x) ~ 0 i = 1 uniformly in probability (2.8) 688 Su But this last condition is easily satisfied since 1/a,,Z~=t H(X+, X'~)~0 uniformly in probability and for each t > 0 sup P x tiH(Xi, x) > t ~<2 max sup i--I I<<.k<~n H(X~, x) > x i 1 Step 2. We turn to prove ,f2-LLn U ~_g,_.,_tt--, O a.e. (2.9) Let e > 0 be given. By the Borel-Cantelli lemma, it suffices for Eq. (2.9) to show that r ~P| k=, ,F max ~ - - - IIU~n, _~nl[ > e \2~-'<~"<~2Ax/2LLn <<.~ P ( max \ 1 <~n~ ~k ) - -" ~'. n2,H<X,, > ~ 2 a2, <oo (2.10) Let us first observe that since E IIH(X~, Xz)II2/LL }]H(X~, X2)I] < oo, then we have max ~ 1 ~<n~<2k I<~i:~j<~n ~ P( k=l 1 n2.2Hl,~m>,,,.~>(Xi,Xj)>e2ka2k) ~ <~- ~ 1 ~ g Iln2.znllllm>,.~.l(X i, Xj)II e k= I 2kak 1 <~i~j<~2 k 4 ~ 2k ~<- ~ - - E IIHII Illlm >,,,~)(Xl, X2) Ek= I a2k 4 ~ 2k~. k~-E Ilgll I<ad<~ltnll6a_,t+,)(Sl, X,) =-~ =1 a o k l = k 4~ 2k~ <~- ~. - - a2,.,P(IIH(X,, X2)[I >a2t) 4 +~.. a2,+IP(IiH(X], Y2)ll > a2,) L -2k <~I= k=l a2k C ao <~+~= 2+P(IIH(X,, X2)II >a2,)< oo Here c denotes a positive numerical constant. (2.11) Iterated Logarithm Law and Mareinkiewicz Law of Large Numbers 689 After having Eq. (2.11), we only need to convince ourselves of the truth of the following claim: ~P k=l ~ ( max I ~</iKgk ~" ~Z.~ 2HItllnll<~a,~)(Xi,Xj) -" l<~i#j~n >e2ka2k)< oO (2.12) " For the proof of this more difficult part we provide two lemmas. Lemma 2.1. Suppose H is a measurable symmetric kernel function taking their values in a separable Banach space B and for some t > 0 sup max P( f ~B~ 1 <~n<~2k ~ f o H ( X i , Xj) >t)<<.~ I <<.i#j<~2k (2.13) Then P( Z max \ 1 ~<n~< "~k - I H(Xi, Xj) 1 ~i#j~2 P +8 >2t) <~i#j<~n k ( 2 n<j~2 H(X,,X:) ') >-~ (2.14) k Proof of Lemma 2.1. Assume that {X';, 1 <~i<~2 k} is an independent copy of {X;, 1 ~< i~<2k}. From the same reason as in Eq. (6.2) of Ledoux and Talagrand ~8) it follows / P ( max \ H(XI,Xj) > 2t) Y" / <~2P(max l~<n~<2k Y' H(X,,Xj)-H(X'i,X'j.)>t) (2.15) l<~i#j<~n Let r = inf{n ~<2k: liE, ~<,+:~<,, H(X;, xj) - H(X',, xj-)ll > t}. By definition, the event ( r = n ) only depends on the random variables Xi ..... X,, X'~ ,..., X',,; and 2k ~. max I ~n~ -") '~" I<~i#j<~n H(X~,Xj)-H(X~,X)) >t = ~ (z=n) ;1=1 690 Su Recall that e~ is a Rademacher variable independent of all other random variables. Since it is not possible, by triangular inequality, that both IIx + yll < Ilxll and I I x - Yll < IIx[I, we have P( IIx-t- e, Yll > Ilxll) ~ 89 for all x, y in B. Since I1~1 <.i+j<.,, H(Xi, P( X Xj)-H(XL (2.16) X))II > t on the event (r =n), then H(X,, Xj) - H(X',, X)) I <~i~j<~n + ~. , H(X~, Xj)--H(X~, X~.) > ~ , r = n ) .u<i~j<.2 k =P.v.x;| ( ~ H(X~, X j ) - H ( X ; , X~) I <~i~j<~n , t h H( Xi, Xj) -- H( X i, ?('i) > ~, r = 17 n<i~j<~2 k 1 >~~ P(r = II) On the other hand, P( X H(X,, Xj) - H(X',, X)) I <~i~j<<.n n<i~j~<2 k I <<.i#j<.2 k n<j<~2 k +p( z n<i~<2 k J Iterated Logarithm Law and Marcinkiewicz Law of Large Numbers <~2p ( , Y, H(X. Xj) >-~, 691 ) ~=n I <~i~:j<~2 k +4P ( ' 2 ) H(X,, Xi) >--~, ~=n (2.17) 1 ~<i<~n n<j<~2 k Summarizing the precedings, we can now arrive at the desired result. L e m m a 2.2. such that for any {Xi, Yj, l < i < , l , E liH(Xi, Yj)[I~'< (i) For every p/> 1 there exists a positive constant cp measurable function H: $2-~ B and any finite sequence l~<j~<m} of independent random variables with ~ the following inequality holds: For 1 ~<p~<2, E i lj=l i E ~=H(X~,Yj) + <cp i (ii) I j t H(X,,Yj) i .) 1 Forp>2, Ij=l H(X~, Yj) i <~cp n p/z - ' i=1 Ij=l + m p/2- ' E ~ H(Xi, j t Proof of Lemma 2.2. E E j=l 2 H(X,, Yj) --E i lj=l ,,, j=l E i 1 H(Xi, Yj) First observe ~ , ~. H(X. Yj) j=l i n ~2p-I(ErEx lj=l m n i=l j=l H(X. + Er Ex .j~l i=l m p ~= Z H(Xi, YJ) I 2 Z H(X,, rj) - e x i l j=l -e,.F.~. H(X,, r~) j 1 i=l (2.18) 692 Su where E x and E r denote expectation with respect to (X, ..... X.) and ( Y, ..... Y.,) respectively. We can use Theorem 2.1 of de Acosta ~'1 conditionally on (Y~ ..... Y,,,) to get i=I j=l ~. <Cp n, i I j=, i lj=l p E ~= H(X,, rj) i=l j l for 1 ~ p ~ 2 and i=1 j=l <~cpn pn- ' H(X,, Y]) i=l for p > 2 . We shall also I apply Yurinskii's inequality to the averages For each 1 ~<k~<m, let o,~. = a ( Yl ..... Y~-) EA. IIEf=, 52/=, H(X. Yj)ll next. and ~,~ = the trivial ~r-field. Set ( ~ ) Since by independence, E r Ex H(Xi, Yj) j=l j~k H(X~, Yj) ~'k =Er i=l ~k--I 1 i=l j~k we see that d k = E r ( f k [ ~ ' k ) - - E y ( f ~ - l ~ ' k - l ) , j=l i=l where for each I <<.k~m, j=l jr i=l It is plain that [fk[ ~ E x ][ZT=l H(Xi, Yk)[I, l<.k<<.m. Iterated Logarithm Law and Marcinkiewiez Law of Large Numbers 693 Keeping these in mind and arguing as in Theorem 2.1 of de Acosta, ") we can obtain Er Ex "' -EvEx i=I <~Cp "' j E Y/) I i=1 P ' H(Xi, i=l j=l lbr 1 ~<p ~<2 and Er H ( X , , Yj) j=l H(Xi, i=l 1 i=1 <~ct, m p/'- - l E j=l H(Xi, i=1 for p > 2 . Thus Lemma 2.2 is concluded. Next we continue our proof of Theorem 2.1 with the help of Lemmas 2.1 and 2.2. For simplicity, unless specially stated, throughout the sequel of the proof c denotes a positive constant that may depend on some parameters and vary at each occurrence. In fact we have no attempt to get its exact value. Since E [IH(Xl, Xz)II2/LL [IH(XI, X2)II < oo, then by using Markov's inequality and noting that r~, 2foHI~llm <,,,~.~ is a P-canonical kernel function for each f in B', we have sup / max P ( . f E B i 1 <~n<~2k \ Xr2.2 f o HI~uml <.,,,kl(Xi, X/) > e2'~a2k) 1 <~i#j<~n ~< , , , ,1t , , sup max E ~ x2.2foHI~llnll<<~,:~.)(Xi, Xj) 2 ~-z- ar,k t'Eai I<.<n<.2k [l<.i~j<.n r <~ ., ., E [[Hll2 IiiiHit<~,,2k~(Xl, X2) e -a ";k cLLa,k ~< ~ , - E HH(X~, X2)I[2/LL [[H(X~, X2)[[ e-a;_~. ---*0 ( k ---+~ ) 694 Su Thus by Lemma 2.1, in order to verify Eq. (2.12) it suffices to show for any e>0 k= I 1 ~i#.i~2 k and ")k < m k= I tl= I (2.20) I ~i~li n<j<~2 k To this aim we'll use Lemma 2.2. But before doing it we need further establish the following facts: a~k J lejHI~,H,<,,.k)(Xj, X') --+0, foreach p>0 (2.21) and 1 max E - ~ ~ 2HI(, m <~,,,_Ei(Xi, Xj.) --+0 (2.22) I ~<i~<n n<j~<2 k - Equation (2.21) can be proved as follows. Since, in view of the contractive principle (see Theorem 4.4 in Ledoux and Talagrand ts)) for any t>0, supP( .v ~=gjHI(llltll<~a,k}(.)(i,x ) > l I j I x ~<4 max I ~<n~<2 k j I H(Xj, x) > supP x j 1 then the assumption that l/a,, ~ = ~H(Xj, x ) ~ 0 uniformly in probability implies 1/a~'~.5Z~k=~ejHI(tm, <~,,,.E)(Xj,X) --* 0 uniformly in probability. So Eq. (2.21) follows at once as a direct consequence of Lemma 1.3(b). For Eq. (2.22) we notice, by Jensen's inequality and decoupling inequality given in Lemma 1.1, Iterated Logarithm Law and Marcinkiewicz Law of Large Numbers E 695 n 2 , HIi iim <~,,,.k,(Xi, X:) 1 ~i<~n n<j~2 k <~cE eigjT[2,2HI~llHil<~a,j,.l(Xi, Xj) ~ I <~i<~n n < j < ~ 2k <~cE ~ eiejHI~ tlHII <~,,,.k~(Xi, Xj) n<j~<2 k <~cE ~ ! <~i#j<~2 k F'iej,l, HI~IIHII <<.a,_kj(Xi, Xjl,) On the other hand, it's easily seen that 1 2ka------TkE • - 1 ~i#j<~2 k x/2k e,e) 1'HI, I,n,. > ~.,k,(X,, X)' ') E ][HI] I~llml>,,,.kl(Xl, X2) --,0 Therefore, Eq. (2.22) will be valid if I _k E 2ka, - ~ I <~i#j<~2 Ir eieJllH(Xi, X~.' ') ~ 0 (2.23) Since x / ~ / ~ UgH---,O in probability, then it follows from Lemma 1.2(b) that 1/2ka,_k ~-,t <~i~:~2k eiE~"~H(X~, X~'j I) ~ 0 in probability. This in turn together with Lemma 1.3(a) implies the required result of Eq. (2.23) whenever 2ka, k E max - I~<j<~2k i#j ~iH(Xi, Xj.) ~ 0 But by Eq. (2.7) 1 ~, g i H ( X i , X j ) -E max 2ka2k 1~j<~2k i ~ j ~1 E - ~0 ~ F,i H ( X i , i= I su 696 Now we are in a position to prove Eq. (2.19). Recall that ~.,. 2HI(itm <,_,,,~.)(Xi, Xj) 1 <~iv~j<~2 k 4 2"* ~ m " Z E ~"-. 2HI(IIH,<~,,.v')(Xi, Xy) N~{I.,...2 '('} i ~ N jq~ N Then ~ P( k= ~, I lt2.2HI, iIHil<~o,.kl(Xi,Xj) >$2ka2 k) I <~ir k < ~-=, P ~ ~ n2..HI, ,,H, <_<.,.k,(X._ ~ N~ { I,....2 k} > ~ 2 a2k (2.24) i~N jCN Thus, in view of Eq. (2.24) and the fact that the three random variables ~ i ~N n 2. 2HI, It*t,<,,,~)(Xi, Xy), Z l# N<i< # Nk n~- ' 2HI, IIHII<~a2kl(Xi ' "~j) and --j~N <j~<2 Z~<i<~N <~j<~2 k - #N n,- ' 2HI, IIm<,,,_,)(X,X) ~) have the same distributions, it is enough for us to prove for any t > 0, P zz, 2HI, IIHII~<a2k)( Xi, ~) ~ - k=l N~{I,...,2/'} i jq~ N -E ~, n2, 2HII ,,-I, <,,_,kl(Xi, Xy} > e2ka2 ~) i~N yen (2.25) <O0 We apply Lemma 2.2 for p = 2 and Eq. (2.21) to get P k=l Z - N~ {I,....2 t'} Z ~2, zHI, ILHIt<~,,,-*,(Xi,Xy) i~N ]r - E i~NTg2,2H]~llHll<<mt)(Xi, Xj) ~>~2ka2k I jCN Iterated Logarithm Law and Marcinkiewicz Law of Large Numbers f ~ EI <~f,=, ~1 --E Z ~" zHI(,It-ttl~#2k)(Xi, Xfl Z - , N c { 1,..., 2k} _ 697 iEN jCN ~ 7~2,2HI(IIHII<~a2kI(XI,XJ) 2 jCN <~ f 1 ,.)~,1 , .,,k ~.. E qZ k=lS-Z"- a~_A'L- N~{I,...,2 k} n2,:Hl(llHll <~,,.k)(Xi, ~ ) i~N jCN --E[,~NT~2,2HI(IIHII<~a2k)(Xi, XY) I 2 jCN f C k~l r a2k / - N~{I,....2 k} i 1 2/`.- # N + 2\ ~rc,_.2HI(llHil~.2k)(Xi, X.~) jft i=l C 2 j=l #N Y'. E <~ f (#__~E 2k-#Nn2, ",HI(tim <.,,,.,)(Xi, Xj.) 1 k ) - , T_k ," 2 E ~. 8jHI(tIHII<~..,?)(Xj, k=l 8-2- ask /lj=l < co (2.26) As for Eq. (2.20) the p r o o f is similar but L e m m a 2.2 is applied for p = 4 . Indeed, for any e > 0, k=l n:l l~i~<n n<j<~2k -E Z n 2. =HI~ IIHIL<<,2k)(Xi. X fl > 82ka2k I I ~<i<~n n<j~<2 k c~ k=l -E 2/`" Z 1 -8424k(/4 -E k Z re2. ,_HI( llHil~<#2k)(X i. X fl I <~i<~n n<j<~2k Z 1 ~<i~<n n<j~<2 k r~2, 2HI(llnll~<a2kj(Xi' % ' ) 4 698 Su <- Y k=l -E t=l c t/24ka~, n + (2 k -- n) E cr i=l -~t.- c E e,2ka4------~ k=l - "= n , 2HI~ IlUll~,?I(Xi, j=l .<y 1 ;g2"2HIlllHII <~a?l(Xi" )~j) 4 i=1 4 -__2ejHIiimm <.a,?l(Xj, X') j I (2.27) <O0 where the convergence of this series is due to Eq. (2.21). Up to now, summarizing the previous works we can end the proof of Theorem 2.1. Theorem 2.1 only presents some sufficient conditions for the LIL of B-valued U-statistics. Some of them are not necessary. More precisely, we have proved certain almost behaviors from statements in probability (weak statements). Theorem 2.2 shows that in a type 2 space the statements in probability can be understood well in terms of the moment. Theorem 2.2. Assume that B is of type 2 space and H: S"' ---, B is a measurable symmetric kernel function. If (1) lim sup E IIH(XI ..... X,,,)II2/LL IIa(gl ..... L,,)II I~llu~x,..,,_ ...........~tl>,~ = 0 where the supremum is taken over all x = (x2 ..... x,,,) E S . . . . (2) 1 a2=supr~Bi E(nm.,,,foH(X,))2< or, then lim x/~ IIU'~,H- EHI[ = ma a.e. If this condition (2) is replaced by (2') { ( n l . , , f ~ ', Ilfll <~ 1} is uniformly integrable, then the compact LIL holds for U',',H with the limit set inK, where K=K,,.,,,n~x,~ is the unit ball of the reproducing kernel Hilbert space associated to r~.,,,H(X~). Proof Assume m = 2. Along the line of the proof of the Theorem 2.1 it is enough for us to verify the following two statements Iterated Logarithm Law and Marcinkiewicz Law of Large Numbers L eiH(Xi, x) ---,0 699 uniformly in probability (2.28) in probability (2.29) an i= 1 1 ~. e~ejH(Xi, Xj) --* 0 lT(ln 1 <~i#j<~n We are ready to prove Eqs. (2.28) and (2.29) under the hypothesis (i) in type 2 spaces. In fact, for each n i> 1, ~ eiH(Xi, x) ~<supE supE .v i= 1 x ~ eiHI, ilull<..,,,(Xi, x) i= I + sup nE lIHIIv~lJ~>,,,(X, x)ll (2.30) x Note that the assumption lim sup E IIH(X, x)IIZ/LL lIB(X, x)ll I~tl,(x..,.) n> , ) = 0 l ~ ,:,'0 A" implies 1l - - sup E IIHI~Hm >~,,,,~(X. x)ll ~ 0 ( l tl (2.31) x By the type 2 inequality, eiHIiitz~Ii<_,,,,~(Xi,x) --supE r i = I -" 1 <~c ~ )hi2 E IIHI~It++ll<_,,,~(X, x)H 2 (2.32) where c is a constant of type 2. For each t > 0, the square of the right hand of this inequality is seen to be smaller than t2 1 + ~ 2LLn sup E ILH I . < , . , ~ ~,,>(X, x)ll 2 t 2 ~<2--~n + sup E IIH(X, x)IIZ/LL lIB(X, x)ll I, vz,v..,-)Jj >,) x letting 17, and then t, go to infinity concludes the proof Eq. (2.28). [] 700 Su Next we prove Eq. (2.29). Indeed we have 1 L nan l e,4H(Z ~, Zj.) < ~ - - E <~i#j~n ~. nan +-- e,4H(Z~, X's) 1 <~i~j<~n 1 E IIHI,, m ~,.)(Z, Z')ll na n Since E liB(X, X')II'-/LL lIB(Z, Z')II < oo, then Also, by the type 2 inequality once more, 1 E ~. nan 1~ha,, E JlHIiiim ~.,,~(X, X')II e,ejHI, jmrl~,.,(Xi, X)) i <~i~j<<.n ~ 0. ~LEIIH(X,X')II an ~0 Therefore Eq. (2.29) follows, which completes the proof of Theorem 2.2. [] To end this paper we shall state the M L L N of B-valued U-statistics. Their proofs are omitted since the basic ideas are the exactly same as that in Theorems 2.1 and 2.2. We only refer to the early paper of de Acosta (1~ for background and details. Theorem 2.3. Suppose that B is a separable Banach space, is a measurable symmetric kernel function such that g [[H(X1 ..... X,,) Hp < oo for some 1 ~<p < 2. Let {X,,, n >~ 1 } be a sequence of independent copies of X. If H:S"'-~B 1 ~ n(Xi, xl 1l I/p x,,,_l)-'O uniformly in probability, H(X i, Xj, x, ,..., x .... z) -) 0 uniformly in probability i=l ~'"' 111 - 1/p ~. n(n -- 1) t ~ , , j < ~ . 11' - ~/P(n n! - m)! ~ H(X,, ..... X,.,) ~ 0 in probability Then n 1 I/p(n ~ m ~ ~ n ! ~(x~,,..., x,,,) ~ o a.e. Iterated Logarithm Law and Marcinkiewicz Law of Large Numbers 701 Theorem 2.4. Suppose that B is of type p for some 1 ~<p<2. H: S m ~ B is a measurable symmetric kernel function such that lim sup E IIH(X~, X2,-.., Xm)ll p ICll~x,..,.,.....~,,,~ll >,1 = 0 where the supremum is taken over all x = (x2 ..... x,,) e S .... 1. Then nI I/p(I l n ! itl)[ H( Xi, ..... Xi,,, ) ~ 0 a.e. i',', ACKNOWLEDGMENTS This paper is partially based on the author's Ph.D. dissertation submitted to Fudan University at Shanghai. The author is deeply indebted to Professors Zhengyan Lin and Chuangrong Lu for their guidance, invaluable discussions and encouragement during the course of the work. The author also acknowledges many helpful suggestions and comments from the referee. REFERENCES 1. de Acosta, A. (1981). Inequalities for B-valued random variables with applications to the strong law of large numbers. Ann. Prob. 9, 157-161. 2. Arcones, M. A. (1994). Limits of canonical U-processes and B-valued U-statistics. J. Theor. Prob. 7, 339-349. 3. Arcones, M. A., and Gin~, E. (1993). Limit theorems for U-processes. Ann. Prob. 21, 1494-1542. 4. Gin~, E., and Zinn, J. (1992a). Marcinkiewicz type law of large numbers and convergence of moments for U-statistics. In Probability in Banach Spaces, Birkhfiuser, Boston, 8, 273-291. 5. Gin~, E., and Zinn, J. (1992). On Hoffmann-Jorgensen's inequality for U-processes. In Probability in Banach Spaces, Birkh~iuser, Boston, 8, 80-91. 6. Gin~, E., and Zinn, J. (1994). A remark on convergence in distribution of U-statistics. Am1. Prob. 22, 117-125. 7. Kahane, J. P. (1985). Some Random Series of Functions. 2nd ed. Cambridge University Press, Cambridge. 8. Ledoux, M., and Talagrand, M. (1991). Probability in Banach Spaces. Springer, New York. 9. de la Pefia, V. H. (1992). Decoupling and Khintchine's inequalities for U-statistics. Ann. Prob. 20, 1877-1892. 10. de la Pefia, V. H., and Montgomery-Smith, S. J. (1994). Bounds on tail probability of U-statistics and quadratic forms. Bull. Amer. Math. Soc. 31, 223-227. I1. Serfling, R. J. (1980). Approximation Theorem of Mathematical Statistics. Wiley, New York. 860/9/3-11