Mathematics Division, National Center for Theoretical Sciences at Taipei NCTS/TPE-Math Technical Report 2006-004 Hausdorff and Packing Dimension Results for the Sample Paths of Self-similar Random Fields Yimin Xiao ∗ Michigan State University Narn-Rueih Shieh National Taiwan University March 25, 2006 Abstract Let X = {X(t), t ∈ RN } be a random field with values in Rd . We develop measure theoretic methods for determining the Hausdorff and packing dimensions of the image X(E) for any given closed set E ⊂ RN . We show that these results are applicable to Gaussian random fields, self-similar stable random fields with stationary increments, the (N, d)-stable sheets and the Rosenblatt process. Running head: Hausdorff and Packing Dimension Results for Self-similar Random Fields. 2000 AMS Classification numbers: 60G15, 60G17, 60J35, 28A80. Key words: Self-similar stable random fields, the stable sheets, Gaussian random fields, the Rosenblatt process, Hausdorff dimension, packing dimension, packing dimension profiles, images. 1 Introduction Fractal dimensions such as Hausdorff dimension, box-counting dimension and packing dimension are very useful in characterizing roughness or irregularity of stochastic processes and random fields. Many authors have studied the Hausdorff dimension and exact Hausdorff measure of the images of Markov processes and Gaussian random fields. We refer to Taylor (1986a) and Xiao (2004) for extensive surveys on results and techniques for Markov processes, and to Adler (1981) and Kahane (1985) for results on Gaussian random fields. It is well known [cf. ∗ Research partially supported by NSF grant DMS-0404729. 1 Kahane (1985, Chapter 18)] that if X = {X(t), t ∈ R N } is a fractional Brownian motion in Rd of index H (0 < H < 1), then for every Borel set E ⊂ R N , n o 1 dimH X(E) = min d, dimH E a.s., (1.1) H where dimH denotes the Hausdorff dimension. There have been various effort to extend (??) to other non-Markovian processes and/or random fields, but only with partial successes. See, for example, Kˆono (1986), Lin and Xiao (1994), Benassi, Cohen and Istas (2003). In order to establish a result similar to (??) for a random field X, it is standard to determine upper and lower bounds for dimH X(E) separately. While the capacity argument [based on Frostman’s theorem] is very useful for determining the lower bound, the methods based on the classical covering argument for obtaining an upper bound for dim H X(E) are quite restrictive and require to impose strong conditions on X. As such, the aforementioned authors have to assume that the random fields either have H¨older continuous sample functions or have first moments. Hence their results are not applicable to stable random fields in general. The first objective of this paper is to develop a measure theoretic method for studying upper bound for the Hausdorff dimension of the image set X(E) of a random field X under minimal conditions. Our method is new and is more powerful than those mentioned above. The second objective of the present paper is to study the packing dimension of the image set X(E) by using again the measure theoretic methods. Packing dimension and packing measure were introduced by Tricot (1982) and Tricot and Taylor (1985) as a dual concept to Hausdorff dimension and Hausdorff measure. It has been known that the packing dimension of a set E characterizes different aspects of the geometry of E than the Hausdorff dimension does; and more information on the fractal structure of E is available when both the Hausdorff and packing dimensions of E are known. Furthermore, packing dimension may even be needed for solving problems that only involve Hausdorff dimension, see Xiao (2004, Sections 7 and 12) for further information. Because of these reasons, packing dimension has become a very useful tool in analyzing fractal sets and sample paths of stochastic processes. See Tricot (1982), Tricot and Taylor (1985), Saint Raymond and Tricot (1988), Taylor (1986a, 1986b), Perkins and Taylor (1987), Falconer and Howroyd (1997), Pruitt and Taylor (1996), Talagrand and Xiao (1996), Xiao (1997, 2004) and the references therein for more information. Let X = {X(t), t ∈ RN } be a fractional Brownian motion with values in R d of index H ∈ (0, 1). Talagrand and Xiao (1996) proved that, if E ⊂ R N is a Borel set, then the packing dimension analogue of (??) may fail and the packing dimension of E alone is not enough to determine dimP X(E), where dimP denotes packing dimension. Later Xiao (1997) proved that dimP X(E) = 1 DimHd E H a.s., (1.2) where Dims E is the packing dimension profile of E defined by Falconer and Howroyd (1997) [see §2 for its definition]. The arguments in Talagrand and Xiao (1996) and Xiao (1997) rely on the modulus of continuity of fractional Brownnian motion, thus can not be applied to processes with discontinuous sample paths. Recently, Schilling and Xiao (2005) have extended (??) to a large class of Markov processes with jumps including stable L´evy processes, certain Feller processes associated to pseudodifferential operators and stable-like processes on fractals. In this paper, we show that their 2 arguments can be modified to establish packing dimension results for non-Markov processes and random fields. Our results are applicable to Gaussian random fields, self-similar stable fields with stationary increments, the stable sheets and the Rosenblatt process. Let X = {X(t), t ∈ RN } be a random field with values in Rd , which will simply be called an (N, d)-random field. Recall that X = {X(t), t ∈ R N } is said to be H-self-similar if for every constant c > 0, we have d X(c t), t ∈ RN = cH X(t), t ∈ RN ; (1.3) and X is said to have stationary increments if for every h ∈ R N , d d X(t + h) − X(h), t ∈ RN = X(t) − X(0), t ∈ RN . (1.4) In the above, X = Y denotes the two processes have the same finite dimensional distributions. Fractional Brownian motion is a very important example of self-similar Gaussian random fields with stationary increments. A systematic account on non-Gaussian self-similar stable processes can be found in Samorodnitsky and Taqqu (1994). The rest of this paper is organized as follows. In Section 2, we recall briefly the definitions and some basic properties of Hausdorff dimension, packing dimension and packing dimension profiles of sets and Borel measures. Let X = {X(t), t ∈ R N } be an (N, d)-random field and let µ be a finite Borel measure on RN . In Section 3, we establish upper and lower bounds for the Hausdorff and packing dimensions of the image measure µ X := µ ◦ X −1 under certain general conditions. In Section 4, we derive analogous results for the image set X(E), where E ⊂ R N is a closed set. Finally in Section 5, we show that the general theorems in Sections 3 and 4 are applicable to several classes of random fields. Throughout this paper we will use λn to denote the Lebesgue measure on Rn , no matter the value of n. When n ≥ 2, we use hx, yi to denote the inner product and k · k to denote the Euclidean norm in Rn . We will use K to denote an unspecified positive constant which may differ from line to line. Some specific constants in Section i will be denoted by K i,1 , Ki,2 and so on. 2 Hausdorff dimension, packing dimension and packing dimension profiles In this section we recall briefly the definitions and some basic properties of Hausdorff dimension, packing dimension and packing dimension profiles for Borel measures and sets. More detailed information can be found in Falconer (1990), Mattila (1995) and Edgar (1998). 2.1 Hausdorff dimension of sets and measures Let Φ be the class of functions φ : (0, δ) → (0, 1) which are right continuous, increasing with φ(0+) = 0 such that there exists a finite constant K > 0 for which φ(2s) ≤K φ(s) for 0 < s < 3 1 δ. 2 (2.1) For φ ∈ Φ, the φ-Hausdorff measure of E ⊂ R N is defined by φ-m(E) = lim inf ε→0 X φ(2ri ) : E ⊂ i ∞ [ B(xi , ri ), ri < ε , i=1 (2.2) where B(x, r) denotes the open ball of radius r centered at x. The Hausdorff dimension of E is defined by dimH E = inf α > 0 : sα -m(E) = 0 . For a finite Borel measure in RN , its Hausdofff dimension (or lower Hausdorff dimension as it is sometimes called) is defined by dimH µ = inf dimH E : µ(E) > 0 and E ⊂ RN is a Borel set , (2.3) and its upper Hausdorff dimension is defined by dim∗H µ = inf dimH E : µ(RN \E) = 0 and E ⊂ RN is a Borel set . (2.4) We will make use of the following equivalent characterizations for dim H µ and dim∗H µ proved by Hu and Taylor (1994, Corollary 4.2): −β N dimH µ = sup β > 0 : lim sup r µ B(x, r) = 0 for µ-a.a. x ∈ R (2.5) r→0 and ∗ dimH µ = inf β > 0 : lim sup r −β µ B(x, r) > 0 for µ-a.a. x ∈ r→0 RN . (2.6) The Hausdorff dimensions of an analytic set E ⊂ R N and finite Borel measures on E are related by the following identity: dimH E = sup dimH µ : µ ∈ M+ (2.7) c (E) , where M+ c (E) denotes the family of finite Borel measures with compact support in E. This can be verified by (??) and Frostman’s lemma. 2.2 Packing dimension and packing dimension profile For φ ∈ Φ, Taylor and Tricot (1985) defined the set function φ-P (E) on R N by X φ(2ri ) : B(xi , ri ) are disjoint, xi ∈ E, ri < ε . φ-P (E) = lim sup ε→0 (2.8) i The set function φ-P is not an outer measure because it fails to be countably subadditive. But it gives rise to a metric outer measure φ-p on R N as follows X [ φ-P (En ) : E ⊂ En . (2.9) φ-p (E) = inf n n 4 φ-p(E) is called the φ-packing measure of E. If φ(s) = s α , then sα -p(E) is called the αdimensional packing measure of E. The packing dimension of E is defined by dimP E = inf α > 0 : sα -p (E) = 0 . (2.10) The packing dimension can also be defined through the upper box-counting dimension, which is more convenient to use in many applications. For any ε > 0 and any bounded set E ⊂ R N , let N (E, ε) be the smallest number of balls of radius ε needed to cover E. The upper box-counting dimension of E are defined as dimB E = lim sup ε→0 log N (E, ε) . − log ε Tricot (1982) proved that dimP E = inf sup dimB En : E ⊂ n ∞ [ n=1 En , (2.11) see also Falconer (1990, p.45). It is well known that 0 ≤ dim H E ≤ dimP E ≤ N for every set E ⊂ RN . Similar to (??) and (??), the packing dimension of a Borel measure µ on R N (or lower packing dimension as it is sometimes called) is defined by dimP µ = inf{dimP E : µ(E) > 0 and E ⊂ RN is a Borel set}, (2.12) and the upper packing dimension of µ is defined by dim∗P µ = inf{dimP E : µ(RN \E) = 0 and E ⊂ RN is a Borel set}. For a finite Borel measure µ on RN and for any s > 0, let Z µ Fs (x, r) = min{1, r s ky − xk−s } dµ(y) (2.13) (2.14) RN be the s-dimensional potential. Falconer and Howroyd (1997) defined the packing dimension profile and the upper packing dimension profile of µ as o n (2.15) Dims µ = sup β ≥ 0 : lim inf r −β Fsµ (x, r) = 0 for µ-a.a. x ∈ RN r→0 and n o Dim∗s µ = inf β > 0 : lim inf r −β Fsµ (x, r) > 0 for µ-a.a. x ∈ RN , (2.16) 0 ≤ Dims µ ≤ Dim∗s µ ≤ s (2.17) r→0 respectively. From these definitions it is easy to show that and if s ≥ N , then Dim∗s µ = dim∗P µ, Dims µ = dimP µ, 5 (2.18) see Falconer and Howroyd (1997) for a proof. Note that the identities in (??) give equivalent characterizations of dimP µ and dim∗P µ in terms of the potential FNµ (x, r). Similar to (??), it can be easily proved that dimP E = sup dimP µ : µ ∈ M+ (2.19) c (E) , see Falconer and Howroyd (1997, p.277). Motivated by this, Falconer and Howroyd (1997) defined the s-dimensional packing dimension profile of E ⊂ R N by Dims E = sup Dims µ : µ ∈ M+ (2.20) c (E) . It follows from (??) and (??) that 0 ≤ Dims E ≤ s and Dims E = dimP E if s ≥ N. (2.21) By the above definition, one can also verify that for any Borel set E ⊂ R N with dimH E = dimP E, we have Dims E = min s, dimP E . (2.22) Packing dimension profiles for finite Borel measures and sets carry more information about local and global geometric properties of the measures and sets than packing dimension does. By using packing dimension profiles, Falconer and Howroyd (1997) determined the packing dimension of projections. It turns out they are very useful in studying various random fractals as well. In Section 3 we will make use of the following lemma, which is a consequence of Proposition 18 in Falconer and Howroyd (1997). Lemma 2.1 Let σ : R+ → [0, N ] be any one of Dims µ, Dim∗s µ, Dims E. Then σ(s) is nondecreasing and continuous. 3 Hausdorff and packing dimensions of the image measures Let X = {X(t), t ∈ RN } be an (N, d)-random field defined on some probability space (Ω, F, P). We assume throughout this paper that (t, ω) 7→ X(t, ω) is B(R N )×F-measurable, where B(RN ) is the family of Borel subsets of RN . For a general random field X which is neither continuous nor Markovian, the classical covering methods for determining the Hausdorff and packing dimensions of X(E) are not applicable [It should be pointed out that dimH X(E) and dimP X(E) may not be determined by the uniform H¨older exponent even when X is continuous; see Example ?? below]. Our approach to these problems is measure theoretic. In this section, we derive upper and lower bounds for the Hausdorff and packing dimensions of the image measures of X and, in Section ??, we prove analogous results for the image set X(E). We will make use of the following two conditions: (C1) There exist positive and finite constants H 1 λ, β, h0 and K3,1 such that λβ > 1 and for all t ∈ RN and h ∈ (0, h0 ), λ H1 ≤ K3,1 | log h|−λβ . (3.1) P sup kX(s) − X(t)k ≥ h | log h| ks−tk≤h 6 (C2) There exist positive constants H 2 and K3,2 such that for all s, t ∈ RN and r > 0, P kX(s) − X(t)k ≤ ks − tkH2 r ≤ K3,2 min 1, r d . (3.2) Remark 3.1 Both Conditions (C1) and (C2) implicitly require the random field X to have certain approximate scaling property (or self-similarity). When X is H 1 -self-similar, Condition (C1) is satisfied whenever the tail probability of sup ks−tk≤h kX(s) − X(t)k decays no slower than a polynomial rate. Hence it is satisfied by many random fields including Gaussian and stable random fields. See Proposition ?? below and Section ??. On the other hand, Condition (C2) is easily satisfied if the random vector X(s) − X(t) /ks − tkH2 has a bounded density function, which is the case for all the random fields considered in this paper. The following proposition gives a simple sufficient condition for a self-similar process X = {X(t), t ∈ R} to satisfy Condition (C1). More precise information can be obtained if further distributional properties of X are known; see Section 5. Proposition 3.2 Let X = {X(t), t ∈ R} be a separable, H-self-similar process in R d with stationary increments. If there exist positive constants β > 0 and K 3,3 such that Hβ > 1 and n o P kX(1)k ≥ u ≤ K3,3 u−β , ∀ u ≥ 1. (3.3) Then there exists a positive constant K 3,4 such that for all u ≥ 1, P sup kX(t)k ≥ u ≤ K3,4 u−β . (3.4) t∈[0,1] In particular, Condition (C1) is satisfied with H 1 = H and any λ satisfying λβ > 1. Proof Without loss of generality, we can assume d = 1. Since the self-similarity index H > 0, we have X(0) = 0 a.s. Let T ∗ = {tn , n ≥ 0} be a separant for X = {X(t), t ∈ [0, 1]}. We assume 0 = t 0 < t1 < t2 · · · < tn < · · · . For any n ≥ 2, consider the random variables Y k (1 ≤ k ≤ n) defined by Yk = X(tk ) − X(tk−1 ). For 1 ≤ i < j ≤ n, let Si,j = u ≥ 1, Pj k=i Yk . By the properties of X and (??), we have that for any X j n o Yk ≥ u = P X(tj ) − X(ti−1 ) ≥ u P k=i o u (tj − ti−1 )H Hβ tj − ti−1 . n = P X(1) ≥ ≤ K3,3 u−β (3.5) Thus, Condition (3.4) of Theorem 3.2 of Moricz, Serfling and Stout (1982) is satisfied with g(i, j) = tj − ti−1 , α = Hβ and φ(t) = tβ . It is easy to see the function g(i, j) satisfies their 7 condition (1.2) with Q = 1. Therefore, it follows from Theorem 3.2 of Moricz, Serfling and Stout (1982) that there exists a constant K 3,4 (independent of n) such that for all u ≥ 1, P j X Yk ≥ u ≤ K3,4 u−β . max X(tj ) ≥ u = P max 1≤j≤n 1≤j≤n (3.6) k=1 Letting n → ∞ yields (??). Remark 3.3 • Theorem 3.2 of Moricz, Serfling and Stout (1982) requires Hβ > 1 in our context. The question whether their Theorem 3.2 holds under the condition Hβ = 1 is still open; see Moricz, Serfling and Stout (1982, p.1039) for details. Even though Proposition ?? holds for any strictly stable L´evy process X with index α ∈ (0, 2] [in this case we have H = 1/α and β = α], it is not known whether Proposition ?? holds in general if Hβ = 1. • If Hβ < 1, generally (??) does not imply (??) as shown by the linear fractional stable processes with index α ∈ (0, 1). See Maejima (1983) or Samorodnitsky and Taqqu (1994). For any Borel measure µ on RN , the image measure µX of µ under the mapping t 7→ X(t) is defined by µX (B) := µ t ∈ RN : X(t) ∈ B for all Borel sets B ⊂ Rd . 3.1 Upper bounds First we consider the upper bound for the Hausdorff dimension of the image measure µ X . Proposition 3.4 Let X = {X(t), t ∈ RN } be a random field with values in Rd . If Condition (C1) is satisfied, then for any finite Borel measures µ on R N , n 1 o dimH µX ≤ min d, dimH µ , a.s. (3.7) H1 and o n 1 dim∗H µ , dim∗H µX ≤ min d, H1 a.s. (3.8) Proof First note that for any fixed s ∈ R N and the sequence hn = 2−n (n ≥ 1), it follows from (??) that for every integer n ≥ log(1/h 0 ), −H1 n n λ X(t) − X(s) ≥ 2 (log 2 ) P sup ≤ K n−βλ . (3.9) kt−sk≤2−n Since P∞ n=1 n−βλ < ∞, the Borel-Cantelli lemma implies that almost surely, sup X(t) − X(s) ≤ 2−H1 n nλ , ∀n ≥ n0 , kt−sk≤2−n 8 (3.10) where n0 = n0 (ω, s) depends on ω and s. By Fubini’s theorem, we derive that, for any finite Borel measure µ on RN , almost surely (??) holds for µ-a.a. s ∈ R N . Now we fix an ω ∈ Ω such that (??) is valid µ-a.a. s ∈ R N and prove that (??) holds. Since dimH µX ≤ d holds trivially, we only need to prove dim H µX ≤ H11 dimH µ. Without loss of generality, we assume dimH µX > 0 and take any γ ∈ (0, dimH µX ). Then by (??) we have Z −γ lim sup r 1l{ky−xk≤r} dµX (y) = 0 (3.11) r→0 Rd for µX -a.a. x ∈ Rd . Here and in the sequel, 1lA denotes the indicator function (or random variable) of the set (or event) A. Equivalent to (??), we have Z −γ lim sup r 1l{kX(t)−X(s)k≤r} dµ(t) = 0 (3.12) r→0 RN for µ-a.a. s ∈ RN . Let us fix s ∈ RN such that both (??) and (??) hold. For any ε > 0, we choose n1 ≥ n0 such that nλ ≤ 2εn for all n ≥ n1 . By (??) we can write Z 1l{kX(t)−X(s)k≤r} dµ(t) RN ≥ ≥ ∞ Z X n=n1 2 ∞ Z X n=n1 = Z −n−1 ≤kt−sk<2−n 1l{kX(t)−X(s)k≤r} dµ(t) (3.13) 2−n−1 ≤kt−sk<2−n kt−sk<2−n1 1l{kt−sk≤r1/(H1 −ε) } dµ(t) 1l{kt−sk≤r1/(H1 −ε) } dµ(t). Hence, we have Z RN 1l{kt−sk≤r1/(H1 −ε) } dµ(t) ≤ Z + Z RN 1l{kX(t)−X(s)k≤r} dµ(t) kt−sk≥2−n1 Note that lim r −γ r→0 Z kt−sk≥2−n1 (3.14) 1l{kt−sk≤r1/(H1 −ε) } dµ(t). 1l{kt−sk≤r1/(H1 −ε) } dµ(t) = 0 (3.15) because the indicator function takes value 0 when r > 0 is sufficiently small. It follows from (??), (??) and (??) that with r = ρH1 −ε , Z lim sup ρ−(H1 −ε)γ 1l{kt−sk≤ρ} dµ(t) ρ→0 RN Z −γ (3.16) = lim sup r 1l{kt−sk≤r1/(H1 −ε) } dµ(t) r→0 RN Z ≤ lim sup r −γ 1l{kX(t)−X(s)k≤r} dµ(t) = 0. r→0 RN 9 Thus we have proven (??) holds almost surely for µ-a.a. s ∈ R N . This implies dimH µ ≥ (H1 − ε)γ a.s. Since ε > 0 and γ < dimH µX are arbitrary, (??) follows. In order to prove (??), it is sufficient to show dim ∗H µX ≤ H11 dim∗H µ a.s. Let ω ∈ Ω be fixed as above. We take an arbitrary β > dim∗H µ. By (??) we have Z −β 1l{kt−sk≤ρ} dµ(t) > 0 for µ-a.a. s ∈ RN . (3.17) lim sup ρ ρ→0 RN By using (??), we derive that for x = X(s) Z Z 1l{ky−xk≤r} dµX (y) ≥ 1l{kt−sk≤r1/(H1 −ε) } dµ(t) Rd RZN − 1l{kt−sk≤r1/(H1 −ε) } dµ(t). (3.18) kt−sk≥2−n1 It follows from (??), (??) and (??) that lim sup r −β/(H1 −ε) Z 1l{ky−xk≤r} dµX (y) Z 1l{kt−sk≤r1/(H1 −ε) } dµ(t) > 0 Rd r→0 ≥ lim sup r −β/(H1 −ε) r→0 (3.19) RN for all s ∈ RN satisfying (??). This implies dim∗H µX ≤ β/(H1 − ε) a.s. Letting ε ↓ 0 and β ↓ dim∗H µ yields (??). This finishes the proof of Proposition ??. Now we turn to the packing dimensions of µ X . The following result is proved by Schilling and Xiao (2005). Proposition 3.5 Let X = {X(t), t ∈ RN } be a random field with values in Rd . If Condition (C1) is satisfied, then for any finite Borel measures µ on R N , dimP µX ≤ 1 DimH1 d µ, H1 a.s. (3.20) dim∗P µX ≤ 1 Dim∗H d µ, 1 H1 a.s. (3.21) and 3.2 Lower bounds As we will see in next section, it is quite standard to determine the lower bound for dim H X(E) using Condition (C2) and a capacity argument based on Frostman’s theorem. However, it is more difficult to obtain a sharp lower bound for dim P X(E) directly. To overcome this difficulty, we first consider the lower bounds for the packing dimensions of the image measures under an (N, d)-random field. Proposition 3.6 Let X = {X(t), t ∈ RN } be an (N, d)-random field satisfying condition (C2). Then for every finite Borel measure µ on R N , dimP µX ≥ 1 DimH2 d µ, H2 10 a.s. (3.22) and dim∗P µX ≥ 1 Dim∗H d µ, 2 H2 a.s. (3.23) Proof We only prove (??); the proof of (??) is similar and is omitted. We will assume DimH2 d µ > 0. Otherwise, there is nothing to prove. Note that, for any fixed s ∈ RN , Fubini’s theorem imples Z µ EFd X X(s), r = E min 1, r d kv − X(s)k−d dµX (v) d Z R (3.24) d −d = E min 1, r kX(t) − X(s)k dµ(t). RN The last integrand in (??) can be written as E min 1, r d kX(t) − X(s)k−d = P kX(t) − X(s)k ≤ r + E r d kX(t) − X(s)k−d · 1l{kX(t)−X(s)k≥r} . By Condition (C2), we obtain that for all s, t ∈ R N and r > 0, rd . P kX(t) − X(s)k ≤ r ≤ K3,2 min 1, kt − skH2 d (3.25) (3.26) Denote the distribution of X(t) − X(s) by Γ s,t (·). Let ν be the image measure of Γs,t (·) under the mapping T : z 7→ kzk from Rd to R+ . Then the second term in (??) can be written as Z Z ∞ d rd r 1 l Γ (dz) = ν(dρ) d {kzk≥r} s,t d kzk ρ d R r (3.27) Z ∞ d r P kX(t) − X(s)k ≤ ρ dρ, ≤d ρd+1 r where the last inequality follows from an integration-by-parts formula. Hence, by (??) and (??) we derive that the second term in (??) can be bounded by Z ∞ d 1 ρ d Kr min 1, dρ ρd+1 kt − skH2 r ( (3.28) 1 if r ≥ kt − skH2 , d H2 ≤K r log kt−sk if r < kt − skH2 . r kt−skH2 It follows from (??), (??), (??) and (??) that for any 0 < ε < 1 and s, t ∈ R N , r d−ε E min{1, r d kX(t) − X(s)k−d } ≤ K3,5 min 1, . kt − skH2 (d−ε) (3.29) For any γ ∈ (0, DimH2 d µ), by Lemma ??, there exists ε > 0 such that γ < Dim H2 (d−ε) µ. It follows from (??) that Z r d−ε −γ/H2 min 1, lim inf r dµ(t) = 0 for µ-a.a. s ∈ RN . (3.30) r→0 |t − s|H2 (d−ε) R+ 11 By (??), (??), (??) and Fatou’s lemma we have that for µ-a.a. s ∈ R N −γ/H2 µX E lim inf r Fd X(s), r r→0 Z r d−ε −γ/H2 dµ(t) = 0. ≤ K3,5 lim inf r min 1, r→0 |t − s|H2 (d−ε) RN (3.31) By using Fubini’s theorem again, we see that almost surely, µ lim inf r −γ/H2 Fd X X(s), r = 0 for µ-a.a. s ∈ RN . r→0 Hence dimP µX ≥ γ H2 a.s. Since γ can be arbitrarily close to Dim H2 d µ, we have dimP µX ≥ 1 DimH2 d µ, H2 a.s. (3.32) This proves (??). Combining Propositions ?? and ??, we have the following result on the packing dimensions of the image measures. Theorem 3.7 Let X = {X(t), t ∈ RN } be an (N, d)-random field and let H be a positive constant. If for every ε > 0, X satisfies conditions (C1) and (C2) with H 1 = H − ε and H2 = H + ε, respectively. Then for every finite Borel measure µ on R N , Proof dimP µX = 1 DimHd µ, H a.s. (3.33) dim∗P µX = 1 Dim∗Hd µ, H a.s. (3.34) For any ε > 0, Propositions ?? and ?? imply that 1 1 Dim(H+ε)d µ ≤ dimP µX ≤ Dim(H−ε)d µ H +ε H −ε a.s. Letting ε → 0 along rational numbers and using Lemma ??, we derive (??). The proof of (??) is similar. 4 Hausdorff and packing dimensions of the image sets In order to determine the upper bounds for the Hausdorff and packing dimensions of the image set X(E), we will need some lemmas. Lemma ?? is a consequence of Theorem 2.1.11 in Edgar (1998, p.73), which is more general than Theorem 1.20 in Mattila (1995) that was used in Xiao (1997). Lemma 4.1 Let E ⊂ RN be a closed set and let f : RN → Rd be a Borel function. If ν is a finite Borel measure on Rd with support in f (E), then there exists a Borel measure µ on R N such that ν = µf and the support of µ is contained in E. 12 Lemma 4.2 Let E ⊂ RN be a closed set. Then for any Borel measurable function f : R N → Rd , dimH f (E) = sup dimH µf : µ ∈ M+ (4.1) c (E) and dimP f (E) = sup dimP µf : µ ∈ M+ c (E) . (4.2) Proof Denote the right hand side of (??) by γ E . It follows from (??) that dimH f (E) ≥ γE . + On the other hand, for any ν ∈ M+ c (f (E)), Lemma ?? implies that there exists an µ ∈ M c (E) such that ν = µf . This and (??) imply dimH f (E) ≤ γE . Hence we have proved (??). The proof of (??) is similar [when N = 1 it is also proved in Schilling and Xiao (2005)]. Remark 4.3 If we assume f is continuous, then (??) and (??) hold for all analytic sets E ⊂ RN , as shown by Xiao (1997). The following proposition gives upper bounds for the Hausdorff and packing dimension of the image sets. Proposition 4.4 Let X = {X(t), t ∈ RN } be an (N, d)-random field satisfying Condition (C1). Then for all closed sets E ⊂ RN , 1 dimH E , a.s. (4.3) dimH X(E) ≤ min d, H1 and dimP X(E) ≤ 1 DimH1 d E, H1 a.s. (4.4) Proof It is clear that (??) follows from (??) and Proposition ??. On the other hand, (??) follows from (??) and (??). Now we consider the lower bounds for the Hausdorff and packing dimensions of X(E). Proposition 4.5 Let X = {X(t), t ∈ RN } be an (N, d)-random field satisfying condition (C2). Then for every closed set E ⊂ RN , 1 dimH X(E) ≥ min d, dimH E , a.s. (4.5) H2 and dimP X(E) ≥ 1 DimH2 d E, H2 a.s. (4.6) Proof First we prove (??). For any 0 < γ < min d, H12 dimH E , we have γH2 < dimH E. Hence there exists a Borel probability measure on E such that Z Z 1 µ(ds)µ(dt) < ∞. (4.7) H2 γ E E ks − tk 13 Let µX be the image measure of µ under X. This is a Borel probability measure on X(E). In order to prove dimH X(E) ≥ γ a.s., by Frostman’s theorem [cf. Kahane (1985)], it is sufficient to show Z Z 1 µ (dx)µX (dy) < ∞ a.s. (4.8) γ X Rd Rd kx − yk By Fubini’s theorem, we have Z Z Z Z 1 1 µ(ds)µ(dt). µ (dx)µX (dy) = E E γ X kX(s) − X(t)kγ E E Rd Rd kx − yk It follows from Condition (C2) that Z ∞ n o 1 −1/γ P kX(s) − X(t)k ≤ ρ E = dρ kX(s) − X(t)kγ 0 Z ∞ 1 dρ ≤ K3,2 min 1, 1/γ (ρ kt − skH2 )d 0 (4.9) (4.10) ≤ K4,1 ks − tk−H2 γ , where in deriving the last inequality, we have used the fact that γ < d. Putting (??) into (??), we see that (??) follows from (??). Hence we have proven dim H X(E) ≥ γ almost surely, which, in turn, implies (??). Next we prove (??). Note that for any γ < H12 DimH2 d E, by (??) there exists a Borel measure µ ∈ M+ c (E) such that H2 γ < DimH2 d µ. It follows from (??) that dim P µX > γ a.s. Hence by Lemma ?? we have dimP X(E) > γ a.s. Since γ < H12 DimH2 d E is arbitrary, we obtain dimP X(E) ≥ H12 DimH2 d E a.s. The proof is completed. The following is the main result of this section. Theorem 4.6 Let X = {X(t), t ∈ RN } be an (N, d)-random field and let H be a positive constant. If for every ε > 0, X satisfies conditions (C1) and (C2) with H 1 = H − ε and H2 = H + ε, respectively. Then for every closed set E ⊂ R N , 1 a.s. (4.11) dimH X(E) = min d, dimH E , H and dimP X(E) = Proof 1 DimHd E, H a.s. Clearly (??) and (??) follow from Propositions ??, ?? and Lemma ??. (4.12) Remark 4.7 By using Remark ??, one can prove the following somewhat stronger result: if an (N, d)-random field X satisfies the assumptions of Theorem ?? and has continuous sample paths, then for every analytic set E ⊂ R N , the formulaes (??) and (??) hold almost surely. It is often desirable to compute dimP X(E) in terms of dimP E. The following is the packing dimension analogue of (??). Note that when N > Hd, then (??) does not hold in general; see Talagrand and Xiao (1996). In this sense, it is the best possible result of this kind. 14 Corollary 4.8 Let X = {X(t), t ∈ RN } be an (N, d)-random field satisfying the conditions of Theorem ?? and let E ⊂ RN be a closed set. If either N ≤ Hd or E satisfies dimH E = dimP E, then 1 dimP E , a.s. (4.13) dimP X(E) = min d, H Proof If N ≤ Hd, then (??) implies that for every Borel set E ⊂ R N , DimHd E = dimP E. Hence Theorem ?? yields dimP X(E) = H1 dimP E a.s. On the other hand, if a Borel set E ⊂ R N satisfies dimH E = dimP E, then (??) implies that DimHd E = min{Hd, dimP E}. Hence (??) follows again from (??). Remark 4.9 Note that in both (??) and (??) the exceptional null probability events depend on E. It would be interesting to investigate whether or when there is a single exceptional null probability event Ω0 such that, for all ω ∈ / Ω0 , (??) and (??) hold simutaneously for all closed (or Borel) sets E ⊂ RN . Such a “uniform dimension result” is applicable even when E is random, hence it is useful in studying intersections and collisions of stochastic processes. Uniform dimension results have been established for Brownian motion, stable L´evy processes and fractional Brownian motion; see the survey papers of Pruitt (1975), Taylor (1986a) and Xiao (2004) for further information. 5 Applications The general results in Sections 3 and 4 can be applied to various classes of random fields. We discuss some of the applications in this section. 5.1 Gaussian random fields First we consider a Gaussian random field X = {X(t), t ∈ R N } in Rd defined by X(t) = X1 (t), . . . , Xd (t) , ∀t ∈ RN , (5.1) where X1 , . . . , Xd are independent copies of a real-valued, centered Gaussian random field Y = {Y (t), t ∈ RN }, which satisfies the following condition: There exist positive constants δ0 , K5,1 , K5,2 and a non-decreasing, right continuous function φ : [0, δ 0 ) → [0, ∞) such that for all t ∈ RN and h ∈ RN with |h| ≤ δ0 , 2 K5,1 φ(|h|) ≤ E Y (t + h) − Y (t) ≤ K5,2 φ(|h|). (5.2) Define the upper index of φ at 0 by o n φ(r) α∗ = inf β ≥ 0 : lim 2β = ∞ r→0 r (5.3) o n φ(r) α∗ = sup β ≥ 0 : lim 2β = 0 . r→0 r (5.4) with the convention inf ∅ = ∞. Analogously, we define the lower index of φ at 0 by 15 Clearly, 0 ≤ α∗ ≤ α∗ ≤ ∞. Xiao (2005) showed that it is possible to have α ∗ = ∞. However, if Y is Gaussian random fields with stationary increments and its spectral measure has an absolutely continuous part, then α ∗ ≤ 1; see Lemma 3.6 in Xiao (2005). A typical example of such Gaussian random fields is the (N, d)-fractional Brownian motion of index H. In this case, α∗ = α∗ = H. Many more examples can be found in Xiao (2005). While it is well known that for every analytic set E ⊂ R N , 1 1 dimH E a.s. (5.5) min d, ∗ dimH E ≤ dimH X(E) ≤ min d, α α∗ the following packing dimension result is new, which extends Theorem 4.1 of Xiao (1997). Theorem 5.1 Let X = {X(t), t ∈ RN } be an (N, d)-Gaussian random field defined by (??) with Y satisfying (??). We assume that 0 < α ∗ ≤ α∗ ≤ 1. Then for every analytic set E ⊂ RN , 1 1 Dimα∗ d E ≤ dimP X(E) ≤ Dimα∗ d E ∗ α α∗ a.s. (5.6) In particular, if α∗ = α∗ = H ∈ (0, 1), then for every analytic set E ⊂ R N , dimP X(E) = 1 DimHd E H a.s. (5.7) Proof To verify that X satisfies Condition (C1), we make use of the following result, which can be proved by using standard Gaussian principles [see, e.g., Xiao (2005, Lemma 3.10)]: If (??) holds, then X has continuous sample paths and there exist positive constants K 5,3 and K5,4 such that for all t ∈ RN , r > 0 small enough and u ≥ K5,3 φ1/2 (r), we have P sup kX(s) − X(t)k ≥ u ks−tk≤r u2 . ≤ exp − K5,4 φ(r) (5.8) It follows from (??) and (??) that, for any ε > 0, X satisfies Conditions (C1) with H 1 = α∗ −ε. On the other hand, for every ε > 0, condition (C2) holds with H 2 = α∗ + ε. Therefore, (??) follows from Propositions ??, ?? and Remark ??. 5.2 Self-similar stable random fields If X = {X(t), t ∈ R+ } is a strictly stable L´evy process in R d , the Hausdorff dimensions of its image sets have been well studied. See Taylor (1986a) and Xiao (2004). The packing dimension results similar to those in Sections 3 and 4 have also been obtained by Schilling and Xiao (2005). Hence in this subsection, we will concentrate on non-Markov stable processes and stable random fields. Let X0 = {X0 (t), t ∈ RN } be an α-stable random field in R with the representation Z f (t, x) M (dx), (5.9) X0 (t) = F 16 where M is an α-stable random measure on a measurable space (F, F) with control measure m and skewness intensity β, f (t, ·) : F → R (t ∈ R N ) is a family of functions on F satisfying Z |f (t, x)|α m(dx) < ∞, ∀ t ∈ RN (5.10) F and also the condition Z f (t, x)β(x) ln |f (t, x)| m(dx) < ∞, ∀ t ∈ RN (5.11) F if α = 1. For any integer n ≥ 1 and t1 , . . . , tn ∈ RN , the characteristic function of the joint distribution of X0 (t1 ), . . . , X0 (tn ) is given by X X n n α ξj f (tj , ·) , ξj X0 (tj ) = exp − E exp i j=1 j=1 α,m where ξj ∈ R (1 ≤ j ≤ n) and k · kα,m is the Lα (F, F, m) norm. The class of α-stable random fields with representation (??) is very broad. In particular, if a random field X0 = {X0 (t), t ∈ RN } is α-stable with α 6= 1 or symmetric α-stable, and is separable in probability [that is, there is a countable subset T 0 ⊂ RN such that for every t ∈ RN , there exists a sequence {tk } ⊂ T0 such that X0 (tk ) → X0 (t) in probability], then X0 has a representation (??); see Theorems 13.2.1 and 13.2.2. in Samorodnitsky and Taqqu (1994) for details. For a separable α-stable random field in R given by (??), Rosinski and Samorodnitsky (1993) investigated the asymptotic behavior of P supt∈[0,1]N |X0 (t)| ≥ u as u → ∞ [see also Theorem 10.5.1 in Samorodnitsky and Taqqu (1994)]. The following tail probability estimate is a consequence of their result. Lemma 5.2 Let X0 = {X0 (t), t ∈ RN } be a separable α-stable random field in R given in the form (??). Assume that X0 has a.s. bounded sample paths. Then there exists a positive and finite constant K5,5 , depending on α, f and m only, such that for all u > 0, lim uα P sup |X0 (t)| ≥ u = K5,5 . (5.12) u→∞ t∈[0,1]N Remark 5.3 In the above lemma, it is crucial to assume X 0 has bounded sample paths almost surely. Otherwise (??) may not hold as shown by the linear fractional stable process X 0 with 0 < α < 1 [see Example ?? below]. For an α-stable random field X 0 with representation (??), some necessary conditions and sufficient conditions for X 0 to have a.s. bounded sample paths can be found in Nolan (1989), Samorodnitsky and Taqqu (1994, Chapters 10 and 12), Marcus and Rosinski (2005). We define an α-stable random field X = {X(t), t ∈ R N } in Rd by X(t) = X1 (t), . . . , Xd (t) , (5.13) 17 where X1 , . . . , Xd are independent copies of X0 . Of course, this definition is more restrictive than that of a stable process in Rd . This is not essential and we have chosen this setting for the sake of simpler presentation of the results. The following theorem gives the Hausdorff and packing dimensions of the images of selfsimilar stable random fields. Theorem 5.4 Let X = {X(t), t ∈ RN } be a separable α-stable random field in R d defined by (??), where X0 is given in the form (??). Suppose that X 0 is H-self-similar, has stationary increments, and its sample path is a.s. bounded. Then for every closed set E ⊂ R N , 1 dimH X(E) = min d, dimH E , a.s. (5.14) H and dimP X(E) = 1 DimHd E, H a.s. (5.15) Proof It follows from the self-similarity and Lemma ?? that Condition (C1) with H 1 = H is satisfied. On the other hand, Condition (C2) with H 2 = H is satisfied because X is H-selfsimilar, has stationary increments and the α-stable variable X(1) has a bounded continuous density function. Therefore, both (??) and (??) follow from Theorem ??. Remark 5.5 Note that, similar to the Gaussian case, it is the self-similarity index H that plays an explicit rˆole in Theorem ??, not the stability index α. Of course, for any H-selfsimilar, α-stable processes with stationary increments, the two indices α and H are intimately related; see Corollary 7.1.11 in Samorodnitsky and Taqqu (1994). In the following we show that Theorem ?? can be easily applied to several important types self-similar stable processes. Example 5.6 [linear fractional stable processes] Let 0 < α < 2 and H ∈ (0, 1) be given constants. we define an α-stable process X 0 = {X0 (t), t ∈ R+ } with values in R by Z X0 (t) = hH (t, s) Mα (ds), (5.16) R where Mα is a strictly α-stable random measure on R with Lebesgue measure as its control measure and where n o H−1/α H−1/α hH (t, s) = a (t − s)+ − (−s)+ n o (5.17) H−1/α H−1/α − (−s)− + b (t − s)− . In the above, a, b ∈ R are constants with |a| + |b| 6= 0, t + = max{t, 0} and t− = max{−t, 0}. Then the α-stable process X0 is H-self-similar with stationary increments, which is called an (α, H)-linear fractional stable process. If H = α1 , then the integral in (??) is understood as aM ([0, t]) for t ≥ 0 and as bM ([t, 0]) if t < 0. Hence X 0 is a strictly α-stable L´evy process. Maejima (1983) proved that if αH < 1, then X 0 is a.s. unbounded on any interval of positive length. On the other hand, if αH > 1 [i.e., 1 < α < 2 and 1/α < H < 1], then 18 the Kolmogorov’s continuity theorem implies that X 0 is a.s. continuous. See also Example 12.2.3 in Samorodnitsky and Taqqu (1994) for a proof using the metric entropy. In the latter case, Takashima (1989) further studied the local and uniform H¨older continuity of the linear fractional stable processes. His Theorems 3.1 and 3.4 showed that the local H¨older exponent of X equals H. However, the exponent of the uniform H¨older continuity can not be bigger than H − α1 [cf. Takashima (1989, Theorem 3.4)]. Now let X = {X(t), t ∈ R+ } be the (α, H)-linear fractional stable process with values in d R defined by (??). When αH > 1, the Hausdorff dimension of X(E) has been determined by Lin and Xiao (1994) using different methods. It follows from Theorem ?? that, if αH > 1, then for every analytic set E ⊂ R+ , dimP X(E) = 1 DimHd E, H a.s. (5.18) Note that, similar to the Hausdorff dimension results of Lin and Xiao (1994), the packing dimension of X(E) does not depend on the uniform H¨older exponent of X. There are several ways to define linear fractional α-stable fields, see Kokoszka and Taqqu (1993). For example, for H ∈ (0, 1) and α ∈ (0, 2), define Z N N ∀ t ∈ RN , (5.19) Z H (t) = kt − skH− α − kskH− α Mα (ds), RN where Mα is an SαS random measure on RN with the N -dimensional Lebesgue measure as its control measure. This is the stable analogue of the N -parameter fractional Brownian motion. However, it follows from Theorem 10.2.3 in Samorodnitsky and Taqqu (1994) that, whenever N ≥ 2, the sample paths of Z H is a.s. unbounded on any intervals in R N . Thus the results of this paper do not apply to Z H when N ≥ 2. In general, little has been known about the sample path properties of Z H . Example 5.7 [harmonizable stable processes] For any given constants 0 < α < 2 and H ∈ eH = {ZeH (t), t ∈ RN } with values in R is (0, 1), the harmonizable fractional stable field Z defined by Z eiht,λi − 1 f H e Mα (dλ), (5.20) Z (t) = Re N RN |λ|H+ α fα is a complex-valued SαS random measure on R N with the N -dimensional Lebesgue where M eH is H-selfmeasure as its control measure. It is easy to see that the α-stable random field Z similar with stationary increments. It follows from Theorems 10.4.2 in Samorodnitsky and Taqqu (1994) [which covers the case eH has continuous 0 < α < 1] and Theorem 3 of Nolan (1989) [which is for 1 ≤ α < 2] that Z sample paths almost surely. When N = 1, Kˆono and Maejima (1991) proved the following eH : for any ε > 0, H¨older continuity of Z lim h→0 sup s, t ∈ [0, 1] |s − t| ≤ h |ZeH (t) − ZeH (s)| =0 1 1 |t − s|H log |t − s| 2 + α +ε a.s. (5.21) eH is different from that of the linear fractional stable Note that the H¨older continuity of Z processes and other self-similar stable processes with stationary increments [see Theorem 2 of 19 Kˆono and Maejima (1991)]. These results suggest that the harmonizable stable processes ZeH behaves much nicer than the linear fractional stable processes. Applying Theorem ?? to the harmonizable stable process in R d defined as in (??), still denoted by ZeH , we have that for all 0 < α < 2, H ∈ (0, 1) and for every analytic set E ⊂ R N , we have 1 H e dimH Z (E) = min d, dimH E , a.s. (5.22) H and 1 DimHd E, a.s. (5.23) dimP ZeH (E) = H Example 5.8 [The Telecom process] Let α ∈ (1, 2) and κ ∈ (0, 1 − α1 ). The Telecom process Zα = {Zα (t), t ∈ R+ } is an α-stable process with the following representation: Z Z 1 −(1−κ)− α Zα (t) = (t ∧ v − u)+ − (0 ∧ v − u)+ (v − u)+ Mα (du, dv) R R (5.24) Z Z Z t 1 −(1−κ)− α 1l{u < s < v}ds (v − u)+ = Mα (du, dv), R R 0 where Mα is a symmetric α-stable random measure on R 2 with the Lebesgue measure as its control measure. Pipiras and Taqqu (2000) have shown that this process arises naturally in modelling Ethernet traffic. They have also shown that Z α is H-self-similar with stationary increments, where H = α1 + κ, and it is different from the linear fractional stable motion. For further information on Telecom processes, see Levy and Taqqu (2000), Pipiras and Taqqu (2002). Since αH > 1, we see that the sample paths of Z α are a.s. continuous. Hence Theorem ?? applies to the d-dimensional Telecom process, still denoted by Z α , defined by (??). Example 5.9 [Self-similar fields of L´evy–Chentsov type] For N ≥ 2, we write R N = SN × R+ in polar coordinates, and let M be a symmetric α-stable random measure on R N with the Lebesgue measure dx = dsdr, (s, r) ∈ S N × R+ as its control measure. For t ∈ RN let V (t) = {(s, r) : 0 < r < hs, ti}. The α-stable field Y = {Y (t), t ∈ R N } of L´evy–Chentsov type is defined by Y (t) = M (V (t)), t ∈ RN . See Samorodnitsky and Taqqu (1994, §8.3) for more information. In the case α = 2, Y is L´evy’s multiparameter Brownian motion. It can be shown that Y is self-similar with index H = 1/α and has stationary increments “in the strong sense”; see Theorem 8.3.2 in Samorodnitsky and Taqqu (1994). Shieh (1996) proved that Y has a version, still denoted by Y , such that the sample paths Y (·) are locally bounded and that each projection process r → Y ((e, r)), e ∈ S N , is a oneparameter SαS L´evy process. We now define an (N, d)-stable field X as in (??), where the components X 1 , . . . , Xd are independent copies of Y . Then it is readily verified that the conditions (C1) and (C2) hold with H1 = H2 = 1/α. Therefore, for every closed set E ⊂ R N , dimH X(E) = min d, α dimH E , a.s. (5.25) 20 and dimP X(E) = α Dimd/α E, a.s. (5.26) Example 5.10 [The stable sheet] Recall from Ehm (1981) that an (N, d)-random field X = X(t), t ∈ RN + is called a stable sheet with index α if X has the following properties: (a) X has independent increments in the sense that, for every integer n ≥ 2 and disjoint intervals Qj , j = 1, · · · , n in RN + , the increments X(Qj ), j = 1, 2, . . . , n are independent; d (b) For every interval Q ⊂ RN + and ξ ∈ R , E exp ihξ, X(Q)i = exp − λN (Q)ϕ(ξ) , (5.27) where the characteristic exponent ϕ(ξ) satisfies ϕ(0) = 0 and Z α gα hξ, yikξk −1 µ(dy). ϕ(ξ) = −σkξk Sd In the above, σ > 0 is a constant, πα gα (x) = |x|α 1 − isgn(x) tan , 2 2i g1 (x) = |x| + x log |x|. π if α 6= 1, (5.28) and µ is a probability measure on the unit sphere S d = {y ∈ Rd : kyk = 1} with µ{y ∈ Sd : hθ, yi = 0} < 1 for all θ ∈ Sd . The stable sheet X is called isotropic (in space) if its characteristic exponent is of the form ϕ(ξ) = K kξkα , where K > 0 is a constant. The case α = 2 is just the Brownian sheet and this is the only case that X has continuous sample paths. It follows from (b) that the stable sheet X has the following self-similarity: for any (c1 , · · · , cN ) ∈ (0, ∞)N , X(c1 t1 , . . . , cN tN ) : t ∈ RN + d = Y N j=1 1/α cj X(t) : t ∈ RN + . Hence X is a special operator-self-similar stable field. Note that, unlike the other stable processes considered in this section, the stable sheet does not have stationary increment in the sense of (??). However, by using Properties (a) and (b) above one can show that the increments of X over intervals are stationary; see Ehm (1981) for details. Even though Theorem ?? can not be applied directly to the stable sheets, Theorems ?? and ?? [more precisely, their proofs] can. Thus we have the following corollary which extends the result of Ehm (1981) on dimH X([0, 1]N ) as well as the packing dimension result for the Brownian sheet in Xiao (1997). 21 Corollary 5.11 Let X = {X(t), t ∈ RN + } be an (N, d)-stable sheet with index α. Then for every closed set E ⊂ (0, ∞)N , dimH X(E) = min d, α dimH E a.s. (5.29) and dimP X(E) = α Dimd/α E Proof a.s. (5.30) By Lemma 2.2 of Ehm (1981), we have for any interval [a, b] ⊂ R N + and u > 0, P sup kX(t) − X(a)k ≥ u ≤ K5,6 u−α , (5.31) t∈[a,b] where K5,6 is a positive and finite constant depending on N and α only. Hence the stable sheet X satisfies Condition (C1) with H1 = 1/α. On the other hand, for any constants 0 < ε < T and s, t ∈ [ε, T ]N , the random variable X(s) − X(t) is α-stable with characteristic function E eihξ,X(t)−X(s)i = exp − λN [0, s]∆[0, t] ϕ(ξ) , where [0, s]∆[0, t] denotes the symmetric difference of the intervals [0, s] and [0, t]. Since there exist positive and finite constants K 5,7 and K5,8 , depending on ε, T and N only, such that K5,7 ks − tk ≤ λN ([0, s]∆[0, t]) ≤ K5,8 ks − tk. By using the Fourier inversion formula, we derive n o P kX(s) − X(t)k ≤ ks − tk1/α r ≤ K5,9 min 1, r d , (5.32) (5.33) where the constant K5,9 depends on K5,7 , α and d only. Hence X satisfies Condition (C2) with H2 = 1/α for all s, t ∈ [ε, T ]N . It follows from Theorem ?? that for every closed set E ⊂ [ε, T ]N , (??) and (??) hold almost surely. Since ε < T are arbitrary, the σ-stability of dimH and dimP yields the desired conclusion. Example 5.12 [The additive stable processes] An N -parameter, d-dimensional random field X = {X(t), t ∈ RN evy process if X has the following pathwise + } is called an additive L´ decomposition: X(t) = X1 (t1 ) + · · · + XN (tN ), ∀t ∈ RN +, where X1 , · · · , XN are independent, classical L´evy processes on R d . Additive L´evy processes arise naturally in the studies of the L´evy sheets and intersections of classical L´evy processes. We refer to Khoshnevisan and Xiao (2004) for historical information, potential theoretic results for additive L´evy processes and their applications to fractal geometry of ordinary L´evy processes. When X1 , · · · , XN are strictly α-stable processes in R d , we call X an (N, d)-additive αstable process. It is an H-self-similar stable random field with H = 1/α. Similar to the stable sheet case, we can verify that the (N, d)-additive α-stable process X satisfies Conditions (C1) and (C2) with H 1 = H2 = 1/α, respectively. Therefore for every closed set E ⊂ (0, ∞)N , both (??) and (??) hold almost surely. 22 5.3 The Rosenblatt process Given an integer m ≥ 2 and a constant κ ∈ (1/2 − 1/(2m), 1/2), the Hermite process Y m,κ = {Y m,κ (t), t ∈ R+ } of order m is defined by Y m,κ (t) = K5,10 Z 0 Rm Z t Y m (s − 0 j=1 κ−1 u j )+ ds dB(u1 ) · · · dB(um ), (5.34) R0 where K5,10 > 0 is a normalizing constant depending on m and κ only and the integral Rm is the m-tuple Wiener-Itˆo integral with respective to the standard Brownian motion excluding the diagonals {ui = uj }, i 6= j. Of course the integral (??) is also well defined if m = 1. In this case, the process is a fractional Brownian motion for which the problem considered in this paper has been solved. So we will exclude it from this section. The Hermite process Y m,κ is an H-self-similar process with stationary increment and H = 1 + mκ − m 2 ∈ (0, 1). It is a non-Gaussian process and often appears in non-central limit theorems for certain processes defined as integrals or partial sums of nonlinear functionals of some stationary Gaussian sequences with long range dependence; see Taqqu (1975, 1979), Dobrushin and Major (1979), Major (1981) for further information. It follows from Theorem 6.3 of Taqqu (1979) that the Hermite process Y m,κ has the following equivalent representation: Y m,κ (t) = K5,11 Z 0 Rm m eit(u1 +···+um ) − 1 Y |uj |κ−1 ZG (du1 ) · · · ZG (dum ), i(u1 + · · · + um ) (5.35) j=1 where K5,11 > 0 is a normalizing constant and ZG is a centered complex Gaussian random measure on R with Lebesgue measure as its control measure; see Major (1981) for systematic account on multiple Wiener-Itˆo integrals. Mori and Oodaira (1986) have studied the functional laws of the iterated logarithms for the Hermite process Y m,κ . The following lemma is a corollary of Lemma 6.3 in Mori and Oodaira (1986). It can also be proven by using Theorem 6.6 in Major (1981) and a standard chaining argument. Lemma 5.13 There exist positive constants K 5,12 and K5,13 , depending on m only, such that for all u ≥ K5,12 , m,κ P max |Y (t)| ≥ u ≤ exp − K5,13 u2/m . (5.36) t∈[0,1] The case m = 2 has received considerable attention recently. The process Y 2,κ is called the Rosenblatt process by Taqqu (1975) [or the fractional Rosenblatt motion as called by Pipiras (2004)] in R. Its self-similarity index is given by H = 2κ. Besides its connections to non-central limit theorems, the Rosenblatt process also appears in limit theorems for some quadratic forms of random variables with long range dependence. Albin (1998a, 1998b) has discussed distributional properties and the extreme value theory of Y 2,κ . In particular, Albin (1998b, Section 16) has obtained bounds on the tail probability of max t∈[0,1] Y 2,κ (t) which are more precise than (??). Pipiras (2004) has established a wavelet-type expansion for the Rosenblatt process. 23 Now we consider the Rosenblatt process X 2,κ in Rd by letting its component processes to be independent copies of Y 2,κ . This non-Gaussian process resembles in many ways the d-dimensional fractional Brownian motion. However, since the tails of its distributions are fatter than those of fractional Brownian motion, we also expect to see interesting differences. It would be interesting to study further sample path properties of X 2,κ systematically. One of the motivations of such studies is to see to what extent the Gaussian methods can be pushed forward. A similar approach has been helpful in studying the symmetric stable processes; see Samorodnitsky and Taqqu (1994). We content ourselves with the following result on the Hausdorff and packing dimensions of the images of X 2,κ . Corollary 5.14 Let X 2,κ = {X 2,κ (t), t ∈ R} be a Rosenblatt process in R d defined above. Then for every analytic set E ⊂ R, we have 1 2,κ dimH E , a.s. (5.37) dimH X (E) = min d, 2κ and dimP X 2,κ (E) = 1 Dim2κd E, 2κ a.s. (5.38) Proof Since X 2,κ is 2κ-self-similar with stationary increments, Lemma ?? implies immediately that X 2,κ satisfies Condition (C1) with H1 = 2κ. On the other hand, Albin (1998a) has proven that the random variable Y 2,κ (1) has a bounded and continuous density. Thus X 2,κ also satisfies Condition (C2) with H 2 = 2κ. Therefore (??) and (??) follow from Theorem ?? and Remark ??. Remark 5.15 It is easily seen that if one could show that for m ≥ 3 the Hermite process Y m,κ has the property that P |Y m,κ (1)| ≤ x ≤ K5,14 x for all x > 0, (5.39) then Y m,κ would satisfy Condition (C2) and Theorem ?? would hold for the d-dimensional Hermite process X m,κ as well, where H = 1 + mκ − m 2 . It would be interesting to prove (??) and to further investigate sample path properties of the Hermite processes. Remark 5.16 By using the multiple Wiener-Itˆo integral with respect to a Gaussian random measure ZG in RN [cf. Major (1981)], one can define the N -parameter Hermite random fields in a way similar to (??) and/or (??). Little has been known about their distributional properties or sample path properties even in the special case of m = 2. Acknowledgements. Part of this work was done during the second author’s visit to Taiwan University and NCTS/Taipei. He would like to thank the staff for their support and the good working conditions. The authors thank Professor Patrik Albin for discussions on the Rosenblatt process. 24 References [1] Adler, R. J. (1981), The Geometry of Random Fields. Wiley, New York. [2] Albin, J. M. P. (1998a), A note on the Rosenblatt distributions. Stat. Probab. Lett. 40, 83–91. [3] Albin, J. M. P. (1998b), On extremal theory for self-similar processes. Ann. Probab. 26, 743–793. [4] Benassi, A., Cohen, S. and Istas, J. (2003), Local self-similarity and the Hausdorff dimension. C. R. Math. Acad. Sci. Paris 336, 267–272. [5] Dobrushin, R. L. and Major, P. (1979), Non-central limit theorems for non-linear functionals of Gaussian fields. Z. Wahrsch. verw Gebiete 50, 27–52. [6] Edgar, G. (1998), Integral, Probability, and Fractal Measures. Springer-Verlag, New York. [7] Ehm, W. (1981), Sample function properties of multi-parameter stable processes. Z. Wahrsch. verw Gebiete 56, 195–228. [8] Falconer, K. J. (1990), Fractal Geometry – Mathematical Foundations and Applications. Wiley & Sons, New York. [9] Falconer, K. J. and Howroyd, J. D. (1997), Packing dimensions for projections and dimension profiles. Math. Proc. Cambridge Philo. Soc. 121, 269–286. [10] Hu, X. and Taylor, S. J. (1994), Fractal properties of products and projections of measures in Rd . Math. Proc. Cambridge Philo. Soc. 115, 527–544. [11] Kahane, J.-P. (1985), Some Random Series of Functions. 2nd edition. Cambridge University Press, Cambridge. [12] Khoshnevisan, D. and Xiao, Y. (2004), Additive L´evy processes: capacity and Hausdorff dimension. Proc. of Inter. Conf. on Fractal Geometry and Stochastics III., Progress in Probability, 57, pp. 151–170, Birkhauser. [13] Kokoszka, P. S. and Taqqu, M. S. (1994), New classes of self-similar symmetric stable random fields. J. Theoret. Probab. 7, 527–549. [14] Kˆ ono, N. (1986), Hausdorff dimension of sample paths for self-similar processes. In Dependence in probability and statistics (Oberwolfach, 1985), pp. 109–117, Progr. Probab. Statist., 11, Birkh¨ auser Boston, Boston. [15] Kˆ ono, N. and Maejima, M. (1991), H¨ older continuity of sample paths of some self-similar stable processes. Tokyo J. Math. 14, 93–100. [16] Levy, J. B. and Taqqu, M. S. (2000), Renewal reward processes with heavy-tailed interrenewal times and heavy-tailed rewards. Bernoulli, 6, 23–44. [17] Lin, H. and Xiao, Y. (1994), Dimension properties of the sample paths of self-similar processes. Acta Math. Sinica 10, 289–300. [18] Maejima, M. (1983), A self-similar process with nowhere bounded sample paths. Z. Wahrsch. Verw. Gebiete 65, 115–119. [19] Major, P. (1981), Multiple Wiene-Itˆ o Integrals. Lecture Notes in Math. 849, Sringer-Verlag, Berlin. [20] Marcus, M. B. and Rosinski, J. (2005), Continuity and boundedness of infinitely divisible procesess: a Poission point process approach. J. Theoret. Probab. 18, 109–160. 25 [21] Mattila, P. (1995), Geometry of Sets and Measures in Euclidean Spaces. Cambridge University Press, Cambridge. [22] Mori, T. and Oodaira, H. (1986), The law of the iterated logarithm for self-similar processes represented by multiple Wiener integrals. Probab. Th. Rel. Fields 71, 367–391. [23] M´ oricz, F., Serfling, R. J. and Stout, W. F. (1982), Moment and probability bounds with quasi superadditive structure for the maximum partial sum. Ann. Probab. 10, 1032–1040. [24] Nolan, J. (1989), Continuity of symmetric stable processes. J. Maultivariate Anal. 29, 84–93. [25] Perkins, E. A. and Taylor, S. J. (1987), Uniform measure results for the image of subsets under Brownian motion. Probab. Th. Rel. Fields 76, 257–289. [26] Pipiras, V. (2004), Wavelet-type expansion of the Rosenblatt process. J. Fourier Anal. Appl. 10, 599–634. [27] Pipiras, V. and Taqqu, M. S. (2000), The limit of a renewal-reward process with heavy-tailed rewards is not a linear fractional stable motion. Bernoulli 6, 607–614. [28] Pipiras, V. and Taqqu, M. S. (2002), The structure of self-similar stable mixed moving averages. Ann. Probab. 30, 898–932. [29] Pruitt, W. E. and Taylor, S. J. (1996), Packing and covering indices for a general L´evy process. Ann. Probab. 24, 971–986. [30] Pruitt, W. E. (1975), Some dimension results for processes with independent increments. In: Stochastic Processes and Related Topics (Proc. Summer Res. Inst. on Statist. Inference for Stochastic Processes), pp. 133–165, Academic Press, New York. [31] Rosinski, J. and Samorodnitsky, G. (1993), Distributions of subadditive functionals of sample paths of infinitely divisible processes. Ann. Probab. 21, 996–1104. [32] Saint Raymond, X. and Tricot, C. (1988), Packing regularity of sets in n-space. Math. Proc. Cambridge Philo. Soc. 103, 133–145. [33] Samorodnitsky, G. and Taqqu, M. S. (1994), Stable non-Gaussian Random Processes: Stochastic models with infinite variance. Chapman & Hall, New York. [34] Schilling, R. L. and Xiao, Y. (2005), Packing dimension of the images of Markov processes. Preprint. [35] Shieh, N. R. (1996), Sample functions of L´evy–Chentsov random fields. In: Probab. Theory and Math Statist., Proceedings of the 7th Japan–Russia Symposium, (S. Watanabe et al, editors), pp. 450–459, World Scientific. [36] Takashima, K. (1989), Sample path properties of ergodic self-similar processes. Osaka Math. J. 26, 159–189. [37] Talagrand, M. and Xiao, Y. (1996), Fractional Brownian motion and packing dimension. J. Theoret. Probab. 9, 579–593. [38] Taqqu, M. S. (1975), Weak convergence to farctional Brownian motion and to Rosenblatt process. Z. Wahrsch. Verw. Gebiete 31, 287–302. [39] Taqqu, M. S. (1979), Convergence to integrated processes of arbitrary Hermite rank. Z. Wahrsch. Verw. Gebiete 50, 53–83. [40] Taylor, S. J. (1986a), The measure theory of random fractals. Math. Proc. Cambridge Philo. Soc. 100, 383–406. 26 [41] Taylor, S. J. (1986b), The use of packing measure in the analysis of random sets. In: Stochastic Processes and Their Applications (Nagoya, 1985), pp. 214–222, Lecture Notes in Math. 1203, Springer-Verlag, Berlin. [42] Taylor, S. J. and Tricot, C. (1985), Packing measure and its evaluation for a Brownian path. Trans. Amer. Math. Soc. 288, 679–699. [43] Tricot, C. (1982), Two definitions of fractional dimension. Math. Proc. Cambridge Philo. Soc. 91, 57–74. [44] Xiao, Y. (1997), Packing dimension of the image of fractional Brownian motion. Statist. Probab. Lett. 33, 379–387. [45] Xiao, Y. (2004), Random fractals and Markov processes. In: Fractal Geometry and Applications: A Jubilee of Benoit Mandelbrot, (Michel L. Lapidus and Machiel van Frankenhuijsen, editors), pp. 261–338, American Mathematical Society. [46] Xiao, Y. (2005), Strong local nondeterminism and the sample path properties of Gaussian random fields. Submitted. Narn-Rueih Shieh. Department of Mathematics, National Taiwan University, Taipei, 10617, Taiwan. E-mail: shiehnr@math.ntu.edu.tw Yimin Xiao. Department of Statistics and Probability, A-413 Wells Hall, Michigan State University, East Lansing, MI 48824, U.S.A. E-mail: xiao@stt.msu.edu URL: http://www.stt.msu.edu/~xiaoyimi 27
© Copyright 2024