Necessary and suﬃcient condition for the functional centrallimittheoreminH¨olderspaces ByAlfredasRacˇkauskas 1 , 3 , 4 and Charles Suquet 2 , 3 Revised version September 10, 2003

Abstract Let ( X i ) i ≥ 1 be an i.i.d. sequence of random elements in the Banach space B , S n := X 1 + ∙ ∙ ∙ + X n and ξ n be the random polygonal line with vertices ( k/n, S k ), k = 0 , 1 , . . . , n . Put ρ ( h ) = h α L (1 /h ), 0 ≤ h ≤ 1 with 0 < α ≤ 1 / 2 and L slowly varying at inﬁnity. Let H oρ ( B ) be the H¨olderspaceoffunctions x : [0 , 1] 7→ B , such that || x ( t + h ) − x ( t ) || = o ( ρ ( h )), uniformly in t . We characterize the weak convergence in H ρo ( B ) of n − 1 / 2 ξ n to a Brownian motion. In the special case where B = R and ρ ( h ) = h α , our necessary and suﬃcient conditions for such convergence are E X 1 = 0 and P ( | X 1 | > t ) = o ( t − p ( α ) ) where p ( α ) = 1 / (1 / 2 − α ). This completes Lamperti (1962) invariance principle. MSC 2000 subject classiﬁcations . Primary-60F17; secondary-60B12. Key words and phrases .CentrallimittheoreminBanachspaces,Ho¨lder space, invariance principle, partial sums process.

1 Department of Mathematics, Vilnius University, Naugarduko 24, Lt-2006 Vilnius, Lithuania. E-mail: Alfredas.Rackauskas@maf.vu.lt 2 CNRSFRE2222,LaboratoiredeMath´ematiquesApplique´es,Bˆat.M2, Universit´eLilleI,F-59655Villeneuved’AscqCedex,France. E-mail: Charles.Suquet@univ-lille1.fr 3 Research supported by a cooperation agreement CNRS/LITHUANIA (4714). 4 Partially supported by Vilnius Institute of Mathematics and Informatics. 1

A.RacˇkauskasandCh.Suquet

1 Introduction Let ( B, k k ) be a separable Banach space and X 1 , . . . , X n , . . . be i.i.d. random elements in B . Set S 0 = 0, S k = X 1 + ∙ ∙ ∙ + X k , for k = 1 , 2 , . . . and consider the partial sums processes ξ n ( t ) = S [ nt ] + ( nt − [ nt ]) X [ nt ]+1 , t ∈ [0 , 1] and ξ s n r := n − 1 / 2 ξ n . In the familiar case where B is the real line R , classical Donsker-Prohorov invariance principle states, that if E X 1 = 0 and E X 12 = 1, then sr D ξ n → W, (1) − in C [0 , 1], where ( W ( t ) , t ∈ R ) is a standard Wiener process and − D → denotes convergence in distribution. The ﬁniteness of the second moment of X 1 is clearly necessary here, since (1) yields that ξ n sr (1) satisﬁes the central limit theorem. Replacing C [0 , 1] in (1) by a stronger topological framework provides more continuous functionals of paths. With this initial motivation, Lamperti [7] considered the convergence (1)withrespecttosomeHo¨lderiantopology.Letusrecallhisresult. For 0 < α < 1, let H oα be the vector space of continuous functions x : [0 , 1] → R such that l δ i → m 0 ω α ( x, δ ) = 0, where ω α ( x, δ ) = sup | x ( | tt ) −− sx | ( α s ) | . s,t ∈ [0 , 1] , 0 <t − s<δ H oα is a separable Banach space when endowed with the norm k x k α := | x (0) | + ω α ( x, 1) . Lamperti [7] proved that if 0 < α < 1 / 2 and E | X 1 | p < ∞ , where p > p ( α ) := 1 / (1 / 2 − α ), then (1) takes place in H oα . This result was derived again by Kerkyacharian and Roynette [5]byanothermethodbasedonCiesielski[2]analysisofHo¨lderspacesbytriangular functions. Further generalizations were given by Erickson [3] (partial sums processes indexed by [0 , 1] d ), Hamadouche [4] (weakly dependent sequence ( X n )),Racˇkauskasand Suquet [10] (Banach space valued X i ’sandH¨olderspacesbuiltonthemoduli ρ ( h ) = h α ln β (1 /h )). Considering a symmetric random variable X 1 such that P { X 1 ≥ u } = cu − p ( α ) , u ≥ 1, Lamperti [7] noticed that the sequence ( ξ s n r ) is not tight in H oα . This gives some hint that thecostoftheextensionoftheinvarianceprincipletotheHo¨lderiansettingisbeyondthe square integrability of X 1 . The simplest case of our general result provides a full answer to this question for the space H oα . Theorem 1. Let 0 < α < 1 / 2 and p ( α ) = 1 / (1 / 2 − α ) . Then ξ n sr − − D −→ W in the space H oα n →∞ if and only if E X 1 = 0 and lim t p ( α ) P {| X 1 | ≥ t } = 0 . t →∞ 2

(2)

FCLTinH¨olderspaces

WewouldliketopointherethatTheorem1contrastsstronglywiththeH¨olderian invariance principle for the adaptive self-normalized partial sums processes ζ s n e . These are deﬁned as random polygonal lines of interpolation between the vertices ( V k 2 /V n 2 , S k /V n ), k = 0 , 1 , . . . , n , where V 02 = 0 and V k 2 = X 12 + . . . X k 2 . It is shown in [11] that ( ζ s n e ) converges in distribution to W in any H oα (0 < α < 1 / 2) provided that E | X 1 | 2+ ε is ﬁnite for some arbitrary small ε > 0. This condition can even be relaxed into “ X 1 is in the domain of attraction of the normal distribution” in the case of symmetric X i ’s (this last condition is also necessary). To describe our general result, some notations are needed here. We write C( B ) for the Banach space of continuous functions x : [0 , 1] → B endowed with the supremum norm k x k ∞ := sup {k x ( t ) k ; t ∈ [0 , 1] } . Let ρ be a real valued non decreasing function on [0 , 1], null and right continuous at 0, positive on (0 , 1]. Put ω ρ ( x, δ ) := sup k x ( ρt () t −− xs () s ) k . s,t ∈ [0 , 1] , 0 <t − s<δ We associate to ρ theHo¨lderspace H oρ ( B ) := { x ∈ C( B ); l δ im 0 ω ρ ( x, δ ) = 0 } , → equiped with the norm k x k ρ := k x (0) k + ω ρ ( x, 1) . We say that X 1 satisﬁes the central limit theorem in B , which we denote by X 1 ∈ CLT( B ), if n − 1 / 2 S n converges in distribution in B . This implies that E X 1 = 0 and X 1 is pregaussian . It is well known (e.g. [8]), that the central limit theorem for X 1 cannot be characterized in general in terms of integrability of X 1 and involves the geometry of the Banach space B . Of course some integrability of X 1 and the partial sums is necessary for the CLT. More precisely, e.g. [8, Corollary 10.2], if X 1 ∈ CLT( B ), then lim t 2 sup P k S n k > t √ n = 0 . t →∞ n ≥ 1 The space CLT( B ) may be endowed with the norm clt ( X 1 ) := sup E k n − 1 / 2 S n k . (3) n ≥ 1 Let us recall that a B valued Brownian motion W with parameter µ ( µ being the distribu-tion of a Gaussian random element Y on B ) is a Gaussian process indexed by [0 , 1], with independent increments such that W ( t ) − W ( s ) has the same distribution as | t − s | 1 / 2 Y . The extension of the classical Donsker-Prohorov invariance principle to the case of B -valued partial sums is due to Kuelbs [6] who established that ξ n sr converges in distribution in C( B ) to some Brownian motion W if and only if X 1 ∈ CLT( B ). This convergence of ξ s n r will be referred to as the functional central limit theorem in C( B ) and denoted by X 1 ∈ FCLT( B ). Of course in Kuelbs FCLT, the parameter µ of W is the Gaussian distribution on B with same expectation and covariance structure as X 1 . The stronger property of convergence in distribution of ξ s n r in H oρ ( B ) will be denoted by X 1 ∈ FCLT( B, ρ ). An obvious preliminary requirement for the FCLT in H ρo ( B ) is that the B -valued Brownian motion has a version in H oρ ( B ). From this point of view, the critical ρ is ρ c ( h ) = p h ln(1 /h )duetoLe´vy’sTheoremonthemodulusofuniformcontinuityofthe Brownian motion (see e.g. [12] and Proposition 4 below). So our interest will be restricted 3

A.RacˇkauskasandCh.Suquet

to functions ρ generatingaweakerHo¨ldertopologythan ρ c . More precisely, we consider the functions ρ of the form ρ ( h ) = h α L (1 /h ) where 0 < α ≤ 1 / 2 and L is slowly varying at inﬁnity. Moreover when α = 1 / 2, we assume that L ( t ) increases faster at inﬁnity than ln β t for some β > 1 / 2. Throughout the paper we use the notation θ ( t ) = t 1 / 2 ρ t 1 , t ≥ 1 . (4) OurcharacterizationoftheFCLTintheHo¨lderspaceH oρ ( B ) reads now simply: X 1 ∈ FCLT( B, ρ ) if and only if X 1 ∈ CLT( B ) and for every A > 0, t li → m ∞ t P k X 1 k ≥ Aθ ( t ) = 0 . Moreover when α < 1 / 2, it is enough to take A = 1 in the above condition. Clearly in the special case B = R and ρ ( h ) = h α , this characterization is exactly Theorem 1. It is also worth noticing that like in Kuelbs FCLT, all the inﬂuence of the geometry of the Banach space B is absorbed by the condition X 1 ∈ CLT( B ). The paper is organized as follows. Section 2 presents some background on the sequen-tialnormequivalenttotheinitialH¨oldernormofH ρo ( B ), the tightness in H oρ ( B ) and the admissibleHo¨ldertopologiesfortheFCLT.Section3givesageneralnecessarycondition which holds even for more general ρ . Section 4 contains the proof of the suﬃcient part in thecharacterizationofHo¨lderianFCLT.Sometechnicalauxilliaryresultsaredeferredin Section 5 to avoid overweighting of the exposition. 2 Preliminaries 2.1 Analytical background With the aim to use a sequential norm equivalent to k x k ρ , we require, following Ciesielski (see e.g. [13, p.67]), that the modulus of smoothness ρ satisﬁes the conditions: ρ (0) = 0 , ρ ( h ) > 0 , 0 < h ≤ 1; (5) ρ is non decreasing on [0 , 1]; (6) ρ (2 h ) ≤ c 1 ρ ( h ) , 0 ≤ h ≤ 1 / 2; (7) Z h ρ ( uu )d u ≤ c 2 ρ ( h ) , 0 < h ≤ 1; (8) 0 h Z h 1 ρ ( uu 2 )d u ≤ c 3 ρ ( h ) , 0 < h ≤ 1; (9) where c 1 , c 2 and c 3 are positive constants. Let us observe in passing, that (5), (6) and (8) together imply the right continuity of ρ at 0. The class of functions ρ satisfying these requirements is rich enough according to the following. Proposition 2. For any 0 < α < 1 , consider the function ρ ( h ) = h α L (1 /h ) where L is normalized slowly varying at inﬁnity, continuous and positive on [1 , ∞ ) . Then ρ fulﬁlls conditions (5) to (9) up to a change of scale. 4

¨ FCLT in Holder spaces

Proof. Let us recall that L ( t ) is a positive continuous normalized slowly varying at inﬁnity if it has a representation L ( t ) = c exp Z bt ε ( u ) d uu with 0 < c < ∞ constant and ε ( u ) → 0 when u → ∞ . By a theorem of Bojanic and Karamata [1, Th. 1.5.5], the class of normalized slowly varying functions is exactly the Zygmund class i.e. the class of functions f ( t ) such that for every δ > 0, t δ f ( t ) is ultimately increasing and t − δ f ( t ) is ultimately decreasing. It follows that for some 0 < a ≤ 1, ρ is non decreasing on [0 , a ]. Then (6) is satisﬁed by ρ ˜( h ) := ρ ( ah ). Due to the continuity and positivity of ρ ˜ on (0 , 1], each inequality (7) to (9) will be fulﬁlled if its left hand side divided by ρ ˜( h ) has a positive limit when h goes to 0. For (7), α this limit is clearly 2 . For (8), we have by [1, Prop. 1.5.10], ˜ ∞ − 1 − Z 0 h ρ ( uu )d u = a α Z v α L ( v/a ) d v ∼ α 1 ρ ˜( h ) . 1 /h Similarly for (9), we obtain by [1, Prop. 1.5.8], h Z h 1 ρ ˜( uu 2 )d u = a α h Z 11 /h v − α L ( v/a ) d v ∼ 1 ρ ˜( − h ) α .

Write D j for the set of dyadic numbers of level j in [0 , 1], i.e. D 0 = { 0 , 1 } and for j ≥ 1, D j = { (2 k + 1)2 − j ; 0 ≤ k < 2 j − 1 } . For any continuous function x : [0 , 1] → B , deﬁne λ 0 ,t ( x ) := x ( t ) , t ∈ D 0 and for j ≥ 1, λ j,t ( x ) := x ( t ) − 21 x ( t + 2 − j ) + x ( t − 2 − j ) , t ∈ D j . The λ j,t ( x ) are the B -valued coeﬃcients of the expansion of x in a series of triangular functions. The j -th partial sum E j x of this series is exactly the polygonal line interpolating x between the dyadic points k 2 − j (0 ≤ k ≤ 2 j ). Under (5) to (9), the norm k x k ρ is equivalent (see e.g. [12]) to the sequence norm k x k ρ seq := s j u ≥ 0 p ρ (21 − j ) t m ∈ a D x j k λ j,t ( x ) k . It is easy to check that k x − E j x k s ρ eq = s i> u j p ρ (21 − i ) t m ∈ a D x i k λ i,t ( x ) k . 5

(10)

A.RaˇckauskasandCh.Suquet

2.2 Tightness The dyadic aﬃne interpolation which is behind the sequential norm is also useful to investigate the tightness in H ρo ( B ). Indeed it is not diﬃcult to check that H oρ ( B ) can be expressed as a topological direct sum of closed subspaces (a Schauder decomposition) by ∞ H o ( B ) = M W i . ρ i =0 Here W 0 is the space of B -valued functions deﬁned and aﬃne on [0 , 1] and for i ≥ 1, W i is the space of B -valued polygonal lines with vertices at the dyadics of level at most i and vanishing at each dyadic of level less than i . It may be helpful to note here that each W i has inﬁnite dimension with B . This Schauder decomposition of H oρ ( B ) allows us to apply Theorem 3 in [14] and obtain the following tightness criterion. Theorem 3. The sequence ( Y n ) of random elements in H oρ ( B ) is tight if and only if the following two conditions are satisﬁed: i) For each dyadic t ∈ [0 , 1] , the sequence ( Y n ( t )) n ≥ 1 is tight on B . ii) For each ε > 0 , j li → m ∞ lim sup P {k Y n − E j Y n k ρ seq > ε } = 0 . (11) n →∞ 2.3AdmissibleH¨oldernorms Let us discuss now the choice of the functions ρ for wich it is reasonable to investigate aHo¨lderianFCLT.If X 1 ∈ FCLT( B, ρ ) and ` is a linear continuous functional on B then clearly ` ( X 1 ) ∈ FCLT( R , ρ ). So we may as well assume B = R in looking for a r necessary condition on ρ . As polygonal lines, the paths of ξ n s belong to H oρ for any ρ such that h/ρ ( h ) → 0, when h → 0. The weaker smoothness of the limit process W and the necessity of its membership in H ρo put a more restrictive condition on ρ . Proposition 4. Assume that for some X 1 , the corresponding process ξ s n r converges weakly to W in H ρo . Then lim θ ( t ) t →∞ ln 1 / 2 t = ∞ . (12) Proof. Let ω ( W, δ ) denote the modulus of uniform continuity of W . Since W has nec-essarily a version in H oρ , we see that ω ( W, δ ) /ρ ( δ ) goes a.s. to zero when δ → 0. This convergence may be recast as ( W, δ ) p δ ln 1 = δ li → m 0 p ωδ ln(1 /δ ) ρ ( δ () /δ ) 0 a.s. ByL´evy’sresult[9,Th.52,2]onthemodulusofuniformcontinuityof W , we have with positive probability lim inf δ → 0 ω ( W, δ ) / p δ ln(1 /δ ) > 0, so the above convergence implies δ li → m 0 p δ ln δ ()1 /δ )0 , = ρ (

which is the same as (12).

6

FCLTinH¨olderspaces

3AgeneralrequirementfortheHo¨lderianFCLT We prove now that a necessary condition for X 1 tosatisfytheH¨olderianFCLTinH oρ ( B ) is that for every A > 0, t li → m ∞ t P X 1 > Aθ ( t ) = 0 . In fact, the same tail condition must hold uniformly for the normalized partial sums, so the above convergence is a simple by-product of the following general result. We point out that Conditions (6) to (9) are not involved here. In this section the restriction on ρ comes from Proposition 4. Theorem 5. If the sequence ( ξ s n r ) n ≥ 1 is tight in H oρ ( B ) , then for every positive constant A , t li → m ∞ t m su ≥ p 1 P S m > m 1 / 2 Aθ ( t ) = 0 . (13) Proof. As a preliminary step, we claim and check that N li → m ∞ n sup P ω ρ ξ n sr , 1 /N ≥ A = 0 . (14) ≥ 1 From the tightness assumption, for every positive ε there is a compact subset K in H oρ ( B ) such that P ω ρ ξ s n r , 1 /N ≥ A ≤ P ω ρ ξ s n r , 1 /N ≥ A and ξ s n r ∈ K + ε. Deﬁne the functionals Φ N on H ρo ( B ) by Φ N ( f ) := ω ρ f ; 1 /N . By the deﬁnition of H ρo ( B ), the sequence (Φ N ) N ≥ 1 decreases to zero pointwise on H ρo ( B ). Moreover each Φ N is continuous in the strong topology of H oρ ( B ). By Dini’s theorem this gives the uniform convergence of (Φ N ) N ≥ 1 to zero on the compact K . Then we have sup f ∈ K Φ N ( f ) < A for every N bigger than some N 0 = N 0 ( A, K ). It follows that for N > N 0 and n ≥ 1, P ω ρ ξ s n r , 1 /N ≥ A and ξ n sr ∈ K = 0 , which leads to P ω ρ ξ sr 1 /N ≥ A < ε, N > N 0 , n ≥ 1 , n , completing the veriﬁcation of (14). In particular we get lim sup P ω ρ ξ s m r N , 1 / N →∞ m ≥ 1 N ≥ A = 0 . (15) Now we observe that sr 1 ≤ m k a ≤ x N ρ (11 /N ) ξ s m r N ( k/N ) − ξ mN (( k − 1) /N ) ≤ ω ρ ξ s m r N , 1 /N . Writing for simplicity Y k,m = m − 1 / 2 ( S mk − S m ( k − 1) ) , we have from (15) that N li → m ∞ m su ≥ p 1 P n 1 ≤ m k a ≤ x N Y k,m > Aθ ( N ) o = 0 , (16) recalling that θ ( N ) = N 1 / 2 ρ (1 /N ). By independence and identical distribution of the X ’s, i P n 1 ≤ m k a ≤ x N Y k,m > Aθ ( N ) o = 1 − 1 − P Y 1 ,m > Aθ ( N ) N . (17) 7

A.RacˇkauskasandCh.Suquet

Consider the function g N ( u ) := 1 − (1 − u ) N , 0 ≤ u ≤ 1. As g N is increasing on [0 , 1], we have g N ( u ) ≥ g N (1 /N ) = 1 − (1 − 1 /N ) N > 1 − e − 1 , 1 /N ≤ u ≤ 1 . (18) By concavity of g N , we also have g N ( u ) ≥ N g N (1 /N ) u ≥ N (1 − e − 1 ) u, 0 ≤ u ≤ 1 /N. (19) Write u m,N := P Y 1 ,m > Aθ ( N ) and u N := sup m ≥ 1 u m,N . By increasingness and continuity of g N , sup m ≥ 1 g N ( u m,N ) = g N ( u N ). This together with (16) and (17) shows that lim N →∞ g N ( u N ) = 0. By (18), it follows that 0 ≤ u N ≤ 1 /N , for N large enough. In view of (19), we have then lim N →∞ N u N = 0. This last convergence can be recast more explicitly as l p P S m > m 1 / 2 Aθ ( N ) = 0 , N i → m ∞ N m su ≥ 1 which is clearly equivalent to (13). 4CharacterizingtheHo¨lderianFCLT Before proving our main result let us set assumptions for ρ ( h ). Deﬁnition 6. We denote by R the class of non decreasing functions ρ satisfying i) for some 0 < α ≤ 1 / 2, and some positive function L which is normalized slowly varying at inﬁnity, ρ ( h ) = h α L (1 /h ) , 0 < h ≤ 1; (20) ii) θ ( t ) = t 1 / 2 ρ (1 /t ) is C 1 on [1 , ∞ ); iii) there is a β > 1 / 2 and some a > 1, such that θ ( t ) ln − β ( t ) is non decreasing on [ a, ∞ ). Remark 7. Clearly L ( t ) ln − β ( t ) is normalized slowly varying for any β > 0, so when α < 1 / 2, t 1 / 2 − α L ( t ) ln − β ( t ) is ultimately non decreasing and iii) is automatically satisﬁed. The assumption ii) of C 1 regularity for θ is not a real restriction, since the function ρ (1 /t ) being α -regularly varying at inﬁnity is asymptoticaly equivalent to a C ∞ α -regularly varying function ρ ˜(1 /t )(see[1]).ThenthecorrespondingHo¨lderiannormsareequivalent. Put b := inf t ≥ 1 θ ( t ). Since by iii), θ ( t ) is ultimately increasing and lim t →∞ θ ( t ) = ∞ , we can deﬁne its generalized inverse ϕ on [ b, ∞ ) by ϕ ( u ) := sup { t ≥ 1; θ ( t ) ≤ u } . (21) With this deﬁnition, we have θ ( ϕ ( u )) = u for u ≥ b and ϕ ( θ ( t )) = t for t ≥ a . Theorem 8. Let ρ ∈ R . Then X 1 ∈ FCLT( B, ρ ) if and only if X 1 ∈ CLT( B ) and for every A > 0 , t →∞ k k ≥ Aθ ( t ) = 0 . (22) lim t P X 1 Corollary 9. Let ρ ∈ R with α < 1 / 2 in (20). Then X 1 ∈ FCLT( B, ρ ) if and only if X 1 ∈ CLT( B ) and lim t P k X 1 k ≥ θ ( t ) = 0 . (23) t →∞ 8

FCLTinHo¨lderspaces

Corollary 10. Let ρ ( h ) = h 1 / 2 ln β ( c/h ) with β > 1 / 2 . Then X 1 ∈ FCLT( B, ρ ) if and only if X 1 ∈ CLT( B ) and E exp d k X 1 k 1 /β < ∞ , for each d > 0 . (24) Corollary 11. Let ρ ∈ R and B = R . Then X 1 ∈ FCLT( R , ρ ) if and only if E X 1 = 0 and either (22) or (23) holds according to the case α = 1 / 2 or α < 1 / 2 . Remark 12. The requirement “for every A > 0” in (22) cannot be avoided in general. For instance let us choose B = R , X 1 symmetric such that P {| X 1 | ≥ u } = exp( − u/c ), ( c > 0) and ρ ( h ) = h 1 / 2 ln(1 /h ), so θ ( t ) = ln t . Clearly (22) is satisﬁed only for A > c , so X 1 / ∈ FCLT( R , ρ ). Proof of Theorem 8. The necessity of “ X 1 ∈ CLT( B )” is obvious while that of (22) is contained in Theorem 5. For the converse part, Kuelbs FCLT gives us for any m ≥ 1 and 0 ≤ s 1 < ∙ ∙ ∙ < s m ≤ 1 ξ n sr ( s 1 ) , . . . , ξ s n r ( s m ) − D → W ( s 1 ) , . . . , W ( s m ) in the space B m . In particular, Condition i) of Theorem 3 is automatically fulﬁlled. So the remaining work is to check Condition (11). Write for simplicity t k = t kj = k 2 − j , k = 0 , 1 , . . . , 2 j , j = 1 , 2 , . . . In view of (10), it is suﬃcient to prove that J →∞ li n m → s ∞ u P n s J u ≤ p j ρ (21 − j ) n − 1 / 21 ≤ m k a < x 2 j k ξ n ( t k +1 ) − ξ n ( t k ) k ≥ ε o = 0 . (25) lim p To this end, we bound the probability in the left hand side of (25) by P 1 + P 2 where P 1 := P n J ≤ j s ≤ u l p og n ρ (21 − j ) n − 1 / 21 ≤ m k a ≤ x 2 j k ξ n ( t k +1 ) − ξ n ( t k ) k ≥ ε o and P 2 := P n j> s l u o p g n ρ (21 − j ) n − 1 / 21 ≤ m k a ≤ x 2 j k ξ n ( t k +1 ) − ξ n ( t k ) k ≥ ε o . Here and throughout the paper, log denotes the logarithm with basis 2, while ln denotes the natural logarithm (log(2 x ) = x = ln(e x )). Estimation of P 2 . If j > log n , then t k +1 − t k = 2 − j < 1 /n and therefore with t k ∈ [ i/n, ( i + 1) /n ), either t k +1 is in ( i/n, ( i + 1) /n ] or belongs to ( i + 1) /n, ( i + 2) /n , where 1 ≤ i ≤ n − 2 depends on k and j . In the ﬁrst case we have k ξ n ( t k +1 ) − ξ n ( t k ) k = k X i +1 k 2 − j n ≤ 2 − j n 1 m ≤ i a ≤ x n k X i k . If t k and t k +1 are in consecutive intervals, then k ξ n ( t k +1 ) − ξ n ( t k ) k ≤ k ξ n ( t k ) − ξ n (( i + 1) /n ) k + k ξ n (( i + 1) /n ) − ξ n ( t k +1 ) k ≤ 2 − j +1 n 1 ≤ m i a ≤ x k X i k . n 9

A.RaˇckauskasandCh.Suquet

With both cases taken into account we obtain P 2 ≤ P n j> s l u o p g n ρ (21 − j ) n − 1 / 2 n 2 − j +11 ≤ m i a ≤ x n k X i k ≥ ε o 1 ε ≤ P n j> s l u o p g n θ (2 j ) 1 m ≤ i a ≤ x n k X i k ≥ 2 o ≤ P n 1 m ≤ i a ≤ x n k X i k ≥ ε 2 j> m l i og n n θ (2 j ) o ≤ n P n k X 1 k ≥ 2 εθ ( n ) o , for n ≥ a (see Deﬁnition 6.iii)). Hence, due to (22), for each ε > 0, lim n →∞ P 2 = 0. Estimation of P 1 . Let u k = [ nt k ]. Then u k ≤ nt k ≤ 1 + u k and 1 + u k ≤ u k +1 ≤ nt k +1 ≤ 1 + u k +1 . Therefore k ξ n ( t k +1 ) − ξ n ( t k ) k ≤ k ξ n ( t k +1 ) − S u k +1 k + k S u k +1 − S u k k + k S u k − ξ n ( t k ) k . Since k S u k − ξ n ( t k ) k ≤ k X 1+ u k k and k ξ n ( t k +1 ) − S u k +1 k ≤ k X 1+ u k +1 k we obtain P 1 ≤ P 1 , 1 + P 1 , 2 , where P 1 , 1 := P n J ≤ j s ≤ u l p og n ρ (21 − j ) n − 1 / 21 ≤ m k a ≤ x 2 j k S u k +1 − S u k k ≥ ε 2 o 1 − 1 2 P 1 , 2 := P n J ≤ j m ≤ a lo x g n ρ (2 − j ) n / 1 m ≤ i a ≤ x n k X i k ≥ 4 ε o . In P 1 , 2 , the maximum over j is realized for j = log n , so P 1 , 2 = P n θ (1 n ) 1 ≤ m i a ≤ x n k X i k ≥ 4 ε o ≤ n P n k X 1 k ≥ ε 4 θ ( n ) o , which goes to zero by (22). To estimate P 1 , 1 , we use truncation arguments. For a positive δ , that will be precised later, deﬁne X i := X i 1 k X i k ≤ δθ ( n ) , X i 0 := X i − E X i , e e e e e where 1 ( E ) denotes the indicator function of the event E . Let S u k and P 1 , 1 be the e expressions obtained by replacing X i with X i in S u k and P 1 , 1 . Similarly we deﬁne S 0 u k and P 1 0 , 1 by replacing X i with X 0 i and ε with ε/ 2. Due to (22), the control of P 1 , 1 reduces e to that of P 1 , 1 because P 1 , 1 ≤ P e 1 , 1 + P n 1 ≤ m i a ≤ x n k X i k > δθ ( n ) o ≤ P e 1 , 1 + n P k X 1 k > δθ ( n ) . e Now to deal with centered random variables, we shall prove that P 1 , 1 ≤ P 0 . It uﬃces 1 , 1 s to prove that for n and J large enough, the following holds 1 u k +1 4 . (26) J ≤ j s ≤ u l p og n ρ (2 − j ) n − 1 / 21 ≤ m k a ≤ x 2 j X k E X e i k < ε i =1+ u k As j ≤ log n , 1 ≤ u k +1 − u k ≤ n 2 − j + 1 ≤ 2 n 2 − j , 0 ≤ k < 2 j , (27) 10

¨ FCLT in Holder spaces

so it suﬃces to have 2 n 1 / 2 k E X e 1 k J ≤ j m ≤ a l x og n 2 − j ε (28) ρ (2 − j ) < 4 . Writing 2 − j /ρ (2 − j ) = 2 − j/ 2 /θ (2 j ) and recalling that θ is non decreasing on [ a, ∞ ), we see that for J ≥ log a , (28) reduces to 2 J/ 2 2 nθ 1 ( / 2 2 J ) k E X e 1 k < ε 4 . (29) Now, as E X 1 = 0, we get t = P k X 1 k > δθ ( u ) δθ k E X e 1 k ≤ Z δθ ∞ ( n ) P k X 1 k > t ) d Z n ∞0 ( u ) d u. By (22), there is an u 0 ( δ ) ≥ 1 such that for u ≥ u 0 ( δ ), u P k X 1 k > δθ ( u ) ≤ 1, whence e δ θ ( nn )+ δ Z n ∞ θ ( u 2 )d u, n ≥ u 0 ( δ ) . k E X 1 k ≤ − u As θ ( u ) u − 1 / 2 = ρ (1 /u ) is non increasing, this last integral is dominated by n − 1 / 2 θ ( n ) R n ∞ u − 3 / 2 d u = 2 n − 1 θ ( n ). Now plugging the estimate e 1 θ ) (30) k E X 1 k ≤ δn − ( n into the left hand side of (29) gives 2 J 2 / 2 nθ 1 ( / 2 2 J ) k E X e 1 k ≤ 22 δ J θ 2( J 2 /J 2 ) θn ( 1 / n 2 ) ≤ 22 J δ. This concludes (26) provided δ/ 2 J < ε/ 8. Estimation of P 1 0 , 1 . Recalling (27), we have log n P 1 0 , 1 ≤ j = X J P n n − 1 / 21 ≤ m k a ≤ x 2 j k S 0 u k +1 − S 0 u k k ≥ 4 ερ (2 − j ) o 0 − ≤ lo X g n P n 1 ≤ m k a ≤ x 2 j k S u + k 1 +1 − u k S ) 0 u 1 k / k 2 ≥ 4 ε √ 2 θ (2 j ) o j = J ( u k ≤ l j o=g X Jnk 2= X j 1 P n ( k uS k 0 u + k 1 +1 −− u k S ) 0 u 1 k / k 2 ≥ 4 √ ε 2 θ (2 j ) o (31) At this stage we use tail estimates related to ψ γ -Orlicz norms (see (38) and (39) in Sec-tion 5). By Talagrand’s inequality (Theorem 15 below), (27), Lemmas 16 and 17, we get for 1 < γ ≤ 2, S 0 u k +1 − S 0 u k 2 ≤ K γ 2 clt ( X 1 ) + ( u k +1 − u k ) 1 / 2 − 1 /γ k X 0 1 k ψ γ ( u k +1 − u k ) 1 / ψ γ ≤ K 0 1 + n 2 − j 1 / 2 − 1 /γ ln 1 /γ δϕθ ( δnθ )( n ) , with a constant K 0 depending only on γ and of the distribution of X 1 . 11