3. L ´ EVY PROCESSES
3.3 Additive L´ evy Processes
Assume Ft and Fs are conditionally independent of Ftfs for all s, t ∈ RN+, where t f s = (min(ti, si))1≤i≤N. Then
E
sup
04s4t
X(t)
≤
p p − 1
N p
E[X(t)p] ∀ t ∈ RN+ and p > 1. (3.29) Proof. We refer to Chapter 7 of [15], Theorem 2.3.2 for a proof.
3.3 Additive L´ evy Processes
An N -parameter random field X := {X(t) : t ∈ RN+} on Rd is an additive L´evy process if each sample path of X satisfies:
X(t) = X1(t1) + · · · + XN(tN) ∀ t = (t1, . . . , tN) ∈ RN+, (3.30) where X1, . . . , XN are independent one-parameter L´evy processes on Rd. Let Ψ1, . . . , ΨN
denote the characteristic exponent of X1, . . . , XN, respectively. Then the characteristic function of X is E[eiξ·X(t)] = exp(−PN
j=1tjΨj(ξ)) for all ξ ∈ Rd. The associated process X = { ee X(t) : t ∈ RN} of X is defined by
X(t) :=e
N
X
j=1
sgn(tj)Xj(|tj|) ∀ t ∈ RN. (3.31)
Note that eX is indexed by all of RN. Moreover,
∀ s, t ∈ RN+ : eX(t − s) has the same distribution as X(t) − X(s), (3.32)
33 and the Fourier transform of eX(t) is given by
Eeiξ· eX(t) = exp
−
N
X
j=1
|tj|Ψj(sgn(tj)ξ)
∀ξ ∈ Rd. (3.33)
We will investigate the sample paths properties of additive L´evy processes. The potential theory developed in Khoshnevisan, Xiao, and, Zhong [23] is crucial in our investigation. For the sake of completeness, we reproduce them here. In the remaining part of this section, we will fix an additive L´evy process X.
Let (Ω, F , P) be the underlying probability space and let λd be the Lebesgue measure on Rd. We will first extend the probability measure P to a σ-finite measure Pλd such that the additive L´evy process X is “uniformly” distributed under Pλd. For each x ∈ Rd, define Px to be the law of x + X, in particular,
Px{X(t) ∈ A} = P{x + X(t) ∈ A} ∀ t ∈ RN+, A ∈ B(Rd). (3.34) Furthermore let Ex be the corresponding expectation operator. Finally, we define the sigma-finite measure Pλd, and a corresponding expectation operator, Eλd, via
Pλd{Ω0} = Z
Rd
Px{Ω0} dx ∀ Ω0∈ F (3.35)
Eλd[Z] = Z
Rd
Ex[Z] dx ∀ Z : Ω → R+ bounded measurable. (3.36) Remark 3.24. For each A ∈ B(Rd), by applying Fubini’s theorem and a change of variable under the Lebesgue measure λd, we obtain
Pλd{X(t) ∈ A} = E
Z
Rd
1{x+X(t)∈A}(x) dx
= λd(A), (3.37) for all t ∈ RN+. This shows that the distribution of X under Pλd is λd. Furthermore, since λdis σ-finite, Pλd is also σ-finite.
Remark 3.25. Let G be a sub-σ-algebra of F . For each random variable Y with Eλd[|Y |] <
∞, let Eλd[Y |G] denote the conditional expectation of Y given G under the measure Pλd, that is, Eλd[Y |G] is G-measurable, Eλd[Y |G] ∈ L1(Pλd), and Eλd[Y 1Ω0] = Eλd[Eλd[Y |G]1Ω0] for all Ω0 ∈ G. By standard integration theory (the Radon-Nikod´ym theorem), Eλd[Y |G]
exists and is unique up to a Pλd-zero event.
Remark 3.26. The conditional expectations under Pλd have all the nice properties of con-ditional expectations under probability measures. For example, we show here a concon-ditional
Fubini’s theorem. Let G be a sub-σ-algebra and Y (t) be a random function from RN to Rd. Then for a fixed nonrandom Borel probability measure µ on RN and all Ω0 ∈ G,
Eλd filtration is the inclusion property with respect to the ordering in the index set. Thus FA := {FA(t)}04(A)t4(A)∞ is a valid filtration for each fixed A. In fact, this generalize the multiparameter filtration defined in the previous section as the multiparameter filtration defined there is simply FΠ. Moreover the filtration FA has the conditional independence property.
Lemma 3.27. Let FA be the filtration with fixed A ⊂ Π. Then for all s, t ∈ RN+, FA(s) and FA(t) are conditionally independent given FA(sf(A)t), where
(sf(A)t)i =
(min(si, ti), if i ∈ A,
max(si, ti), if i ∈ AC. (3.41) Proof. We refer to Chapter 7, Theorem 2.4.1 of [15] for a proof. Although Theorem 2.4.1 of [15] is a result for a Brownian sheet, the same method works for an additive L´evy process.
In fact, the structure of an additive L´evy process is simpler than a Brownian sheet and so is the proof.
35 Now we can combine Pλd and FA to generalize the multiparameter martingales defined in the previous section. Since the filtration FA and the conditional expectation under Eλd are well-defined, the generalization is straightforward. In particular, a multiparameter process M := {M (t)}04(A)t4(A)∞is a submartingale with respect to the filtration FAunder the σ-finite measure Pλd if
1. M is adapted to FA;
2. M (t) ∈ L1(Pλd), that is, Eλd[|M (t)|] < ∞, for all 0 4(A)t 4(A)∞; and 3. Eλd[M (t)|FA(s)] ≥ M (s), Pλd-almost surely for all 0 4(A) s 4(A) t.
In the classical potential theory for one-parameter L´evy processes, the semigroup {Pt}t≥0 plays an important role ([14], Lecture 9). In the case of additive L´evy processes, we define a similar family of operators P = {Pt}t∈RN on L∞(Rd) by
(Ptf )(x) := E[f ( eX(t) + x)] ∀ f ∈ L∞(Rd) or f : Rd→ R+, for all x ∈ Rd. (3.42) Our first observation is the Markov property of X in terms of Pt.
Lemma 3.28 ([23], Proposition 3.2). Suppose that A ⊆ Π and that s 4(A) t, where s and t are both in RN+. Then, for every measurable function f : Rd→ R+,
Eλd[f (X(t))|FA(s)] = (Pt−sf )(X(s)), (3.43)
Pλd-a.s., where Ptis defined in (3.42).
Proof. For an arbitrary positive integer m, consider measurable functions f, g, h1, . . . , hm: Rd→ R+ and t, s, s1, . . . , sm ∈ RN+ such that t <(A) s <(A)sk for all 1 ≤ k ≤ m. Then,
The first equality follows from Fubini’s theorem; we apply a change of variable y = x + X(s) to obtain the second one; and the third equality follows from Fubini’s theorem and independent increments of X. Furthermore, according to (3.42), we have
Eλd
where we apply Fubini’s theorem to obtain the second equality, change of variable x = y − X(s) to obtain the third one; and (3.36) for the last one.
We are interested in the “local time” that an additive L´evy process spent in a given set.
In order to quantify these “local times,” we first define the occupation measure Oµ for any given Borel probability measure µ on RN+:
Oµ(A) = Z
RN+
1A(X(s)) µ(ds) ∀ A ∈ B(Rd), (3.46)
where 1A(·) is the indicator function of A and B(Rd) denotes the Borel σ-algebra of Rd. By standard integration theory, we can extend the occupation measure to a random linear functional on L1(µ):
Oµ(f ) = Z
RN+
f (X(s)) µ(ds) ∀ f ∈ L1(µ). (3.47)
In particular, the Fourier transform of the occupation measure cOµexists, and it is given by
Ocµ(ξ) = Z
RN+
eiξ·X(s)µ(ds) ∀ ξ ∈ Rd. (3.48)
The following lemma computes the second moment of the Fourier transform of cOµ.
37 Lemma 3.29 ([23], Lemma 2.4). For all Borel probability measure µ on RN+,
E and sgn(·) is the usual sign function.
Proof. Since
and |eiξ·(X(s)−X(t))| = 1, we can apply Fubini’s theorem twice to obtain
E[k cOµk2L2(Rd)] =
Finally (3.49) follows from this and (3.33).
The following two technical lemmas are useful in estimating moments of random vari-ables under Pλd.
where we used Fubini’s theorem to derive the third equality and a change of variable y = X(s) + x for the last one. We can apply Fubini’s theorem again to compute the Fourier transform of Pt−sg:
P\t−sg(ξ) = E desired result by applying Plancherel’s theorem to the last term of (3.54) and combining it with (3.55).
Proof. First assume f ≥ 0. Then Fubini’s theorem shows that Eλd[|Oµ(f )|2]
where we used Lemma 3.30 to derive the second equality. Finally for a general f ∈ L1(Rd) ∩ L2(Rd), we note that Eλd[|Oµ(f )|2] < ∞. Therefore we can use Fubini’s theorem to derive the same result.
39 is separable, thanks to Theorem 3.17. The following lemma gives an upper bound for the second moment of the martingale MAµf under the measure Pλd. and Corollary 3.23 show that,
Eλd
Since the above inequality is true for all n ≥ 1, we have
Eλd
Finally, (3.60) follows from Lemma 3.31.
PACKING DIMENSION AND ADDITIVE L´ EVY PROCESSES
For a nonrandom Borel set F ⊆ RN+, the random image set X(F ) usually exhibits fractal structure, and it is natural to calculate the Hausdorff dimension and packing dimension of X(F ). When N = 1, there is a rich literature regarding the Hausdorff dimension dimH(X(F )) of X(F ), see [20] and its extensive reference. However, there have been few results regarding the packing dimension dimP(X(F )) of X(F ). In a recent paper, Khoshnevisan, Schilling, and Xiao [19] defined a new family of packing dimension profile Dimκ and found that
dimP(X(F )) = Dimκ(F ) a.s. (4.1)
When N > 1, a formula regarding the Hausdorff dimension dimH(X(F )) is given by Yang [40]. However, there is little known about the packing dimension dimP(X(F )).
The packing dimension profile defined in [19] is probabilistic but has analytic significance as well. In fact it can be regarded as an extension of the packing dimension profile defined by Falconer and Howroyd [6] and Howroyd [10]. However this generalization only works for subsets of R, since the packing dimension profile in [19] is defined on the “time domain,”
that is, the half real line R+ = [0, ∞). It is natural to ask whether we can generalize it to RN for all N ≥ 1 and still keep the connection to the packing dimension profile of Falconer and Howroyd [6] and Howroyd [10].
In this chapter, we extend the packing dimension profile defined in [19] to higher dimensions and derive a multiparameter version of (4.1). We also show that our definition gives probabilistic interpretations of the packing dimension profile of Falconer and Howroyd [6] and Howroyd [10]. The extension from N = 1 to N ≥ 1 is not straightforward since the structure of RN is richer than R if N > 1. In particular, the result of [19] relies on a stopping time argument while we can not define stopping times for multiparameter processes in general. In order to overcome this difficulty, we employ the potential theory developed in Chapter 3 for additive L´evy processes.
41
4.1 Box-dimension Profiles
Let X = {X(t)}t∈RN
+ be an additive L´evy process on Rd. Its associated process eX = { eX(t) : t ∈ RN} is defined by
X(t) :=e
N
X
j=1
sgn(tj)Xj(|tj|) ∀ t ∈ RN. (4.2)
For every x ∈ Rd, let |x| := max1≤k≤d|xk| denote the l∞ norm of x. Then B(x, ε) := {y ∈ Rd: |x − y| < ε} is the l∞open ball centered at x with radius ε > 0. For every ε > 0, define κε(t) := P{| eX(t)| ∈ B(0, ε)}. (4.3) It follows that κε(t) is nondecreasing in ε for every fixed t. Moreover, Lemma 3.12 ensures that {Xj(tj)}tj≥0 is continuous in probability for all 1 ≤ j ≤ N . Thus, κε(t) is continuous in t for every fixed ε. We consider the family of functions κ := {κε}ε>0.
Definition 4.1. For every bounded Borel set F ⊆ RN, the box-dimension profile dimκ(F ) of F with respect to the family κ is defined as
dimκ(F ) := sup (
η > 0 : lim
ε↓0
ε−η inf
µ∈P(F )
Z Z
κε(t − s)µ(ds)µ(dt) = 0 )
, (4.4)
whereP(F ) denotes the collection of all Borel probability measures supported on F . Remark 4.2. When N = 1, we have
κε(t) = P{|X(|t|)| ∈ B(0, ε)} = P{|X(t)| ∈ B(0, ε)}, (4.5) for all ε > 0 and t ≥ 0. This shows that our definition of box dimension profile generalizes that of Khoshnevisan, Schilling, and Xiao [19]. For this reason we will use the same notations for dimension profiles as those in [19].
A routine check shows that dimκ has the following properties:
(i) dimκ is monotone; i.e., dimκ(F ) ≤ dimκ(G) if F ⊆ G;
(ii) dimκ is finitely stable; i.e., dimκ(∪nl=1Gl) = max1≤l≤ndimκ(Gl);
(iii) dimκ is translation invariant; i.e., dimκ(F ) = dimκ(t + F ) for all t ∈ RN, where t + F := {t + s : s ∈ F }.
We may also express dimκ(F ) in potential-theoretic terms: The following technical lemma simplifies the calculation of Zκ(ε; F ).
Lemma 4.3. For every bounded Borel set F ⊆ RN, let F denote the closure of F . Then wherePf(F ) denotes the collection of all finitely-supported [discrete] probability measures on F .
Proof. It suffices to prove
ν∈infPf(F ) as the converse inequality is trivial. Choose and fix any µ ∈P(F ). For every n ≥ 1, define
In(¯k) := k1
43 It follows immediately that νn ∈Pf(F ). By the uniform continuity of κε(t) and (4.13), we have
n→∞lim
Z Z
κε(t − s)νn(ds)νn(dt) − Z Z
κε(t − s)µ(ds)µ(dt)
= 0. (4.15)
This further implies that inf
ν∈Pf(F )
Z Z
κε(t − s)ν(ds)ν(dt) ≤ Z Z
κε(t − s)µ(ds)µ(dt). (4.16)
Since µ ∈P(F ) is arbitrary, we can take inf over P(F ) on the right hand side of the above inequality to complete the proof.
Lemma 4.4. For every bounded Borel set F ⊆ RN, let F denote the closure of F . Then
dimκ(F ) = dimκ(F ). (4.17)
Proof. SincePf(F ) ⊆P(F ) ⊆ P(F ), Lemma 4.3 implies that inf
µ∈P(F )
Z Z
κε(t − s)µ(ds)µ(dt) = inf
µ∈P(F )
Z Z
κε(t − s)µ(ds)µ(dt). (4.18)
Then (4.17) follows from Definition 4.1.
Remark 4.5. The above lemma implies that dimκ is not σ-stable in general. For example, consider F = QN+∩[0, 1]N. On one hand, for every t ∈ RN the only Borel probability measure supported on {t} is the Dirac measure δt. Then we can compute dimκ({t}) by definition and obtain dimκ({t}) = 0. On the other hand dimκ(F ) = dimκ(F ) = dimκ([0, 1]N) according to Lemma 4.4. We will show in Corollary 4.17 that if κ is determined by a d-dimensional isotropic α-stable process X with 0 < α ≤ 2, then dimκ([0, 1]N) = DimFHd/α([0, 1]N), where DimFHs denotes the s-dimensional packing dimension profile of Falconer and Howroyd [6]. Since DimFHd/α([0, 1]N) = min(d/α, N ) (Falconer and Howroyd [6], P286), we see that dimκ(F ) > supt∈F dimκ({t}).
4.1.1 An Equivalent Definition of Box-dimension Profiles
In general it is hard to use Definition 4.1 to compute the box dimension profile of a given set since the hitting probability P{| eX(t)| ∈ B(0, ε)} is only explicitly computable when X has nice properties. However, we may express the box-dimension profile in terms of the characteristic exponent Ψ of the underlying additive L´evy process X.
For every Borel probability measure µ on RN, and for all ξ ∈ Rd, define the “energy An application of Fubini’s theorem shows
Eµ(ξ) =
The following proposition uses the characteristic exponent of X to compute box-dimension profiles. It is a generalization of Theorem 2.6 of Khoshnevisan, Schilling, and Xiao [19] to the multiparameter setting. The proof of Theorem 2.6 in [19] works here and needs only minor changes. See also Khoshnevisan and Xiao [21], proof of Theorem 1.1.
Proposition 4.6. For every compact set F ⊆ RN, dimκ(F ) = sup Proof. For all ε > 0, let fε denote the [scaled] P´olya distribution:
fε(ξ) :=
Then its Fourier transform is given by fˆε(x) := an application of Plancherel’s theorem to the right-hand side of (4.24) shows that the following holds for all ε > 0 and t ∈ RN:
45 Then the following elementary inequality,
1 − cos(2u)
This shows that the right hand side of (4.21) is less than or equal to dimκ(F ).
In order to establish the converse estimate we introduce a Cauchy process {C(λ)}λ≥0, independent of X, whose coordinate processes C1, . . . , Cd are i.i.d. standard symmetric Cauchy processes in R. Thus the characteristic function of C(λ) is E[exp{ix · C(λ)}] = e−λPdl=1|xl| for all x ∈ Rd. Since
e−(k/ε)Pdl=1|xl|≤ 1B(0,ε)(x) + e−k1B(0,ε)c(x) ∀ ε, k > 0 and x ∈ Rd, by replacing x with eX(t) and taking expectation on both sides, we get
κε(t) ≥ Eexp iX(t) · C(k/ε) − ee −k
|t − s|, and integrate with respect to µ(ds)µ(dt) to find that Z Z be arbitrary small, we get the desired estimate.
4.2 Packing Dimension Profiles
We can regularize the box-dimension profile dimκ defined in (4.4) to produce a family of packing-type dimension profiles.
Definition 4.7. For every Borel set F ⊆ RN, the packing dimension profile Dimκ(F ) of F with respect to the family κ is defined as
Dimκ(F ) := inf
One can verify that Dimκ has the following properties:
(i) Dimκ is monotone; i.e., Dimκ(F ) ≤ Dimκ(G) if F ⊆ G;
(ii) Dimκ is σ-stable; i.e., Dimκ(∪∞n=1Gn) = supn≥1Dimκ(Gn).
We can also associate Dimκ to a “packing measure with respect to the family κ.” In order to do that, we borrow some ideas from Howroyd [10].
Definition 4.8. For every Borel set F ⊆ RN and real number δ > 0, we say that a sequence of triples (ωi, ti, εi)i≥1 is a (κ, δ)-packing of F if for all i ≥ 1: (i) ωi ≥ 0; (ii) ti ∈ F ; (iii) 0 < εi≤ δ; and (iv)P
j≥1ωjκεj(ti− tj) ≤ 1.
Definition 4.9. For every constant s > 0, the s-dimensional packing measure Ps,κ(F ) of F ⊆ RN with respect to the family κ is defined as
where P0s,κ is a premeasure defined by
P0s,κ(F ) := lim
Definition 4.10. For every Borel set F ⊆ RN, the packing dimension profile P-dimκ(F ) of F with respect to the family κ is defined as
P-dimκ(F ) := inf{s > 0 : Ps,κ(F ) = 0}. (4.35)
47 The two packing dimension profiles defined in (4.31) and (4.35) coincide. This can be shown by adapting the proof of Theorem 26 of Howroyd [10]. For completeness we record it here.
First we introduce an equivalent formulation of Zκ(ε; F ) (see (4.7) for definition) for fixed ε > 0 and bounded Borel set F ⊆ RN. A sequence of pairs (ωi, ti)ki=1is a size ε weighted
with the convention that sup ∅ := 0. Then the following technical lemma holds.
Lemma 4.11. For every bounded Borel set F ⊆ RN and real number ε > 0,
Nκ(ε; F ) = 1/Zκ(ε; F ). (4.37)
Proof. On one hand, for all η < Nκ(ε; F ), (4.36) implies the existence of a size ε weighted κ-packing (ωi, ti)ki=1 of F such that Pk
i=1ωi > η. Let W = Pk
i=1ωi and define a discrete probability measure µ on {ti}ki=1 by µ(ti) = ωi/W . It follows immediately that µ ∈P(F ).
Then elementary calculation shows that
Zκ(ε; F ) ≤
(vj, sj)lj=1 by removing from the sequence (u∗i, ti)ki=1 all those pairs (u∗i, ti) with u∗i = 0, then we see that (vj/J, sj)lj=1 is a size ε weighted κ-packing of F . It follows
Nκ(ε; F ) ≥
l
X
j=1
vj J = 1
J. (4.40)
Since J ≤Pk
i,j=1κε(ti− tj)ωiωj < η, we have Nκ(ε; F ) ≥ 1/Zκ(ε; F ). This completes the proof.
Proposition 4.12. For every Borel set F ⊆ RN, we have
Dimκ(F ) = P-dimκ(F ). (4.41)
Proof. Consider an arbitrary bounded Borel set E ⊆ RN. For each ε > 0, let (ωi, ti)ki=1 be a size ε weighted κ-packing of E. Then (ωi, ti, ε)ki=1 is trivially a (κ, ε)-packing of E. It follows from this fact, (4.34), and (4.36) that for all s > 0,
Pεs,κ(E) ≥ εsNκ(ε; E) = εs Zκ(ε; E)−1
, (4.42)
where we used Lemma 4.11 to derive the last equality. Now if 0 < η < Dimκ(F ), then for every sequence of bounded Borel sets {Fk}k≥1 with F ⊆ ∪k≥1Fk, we can find some Fn such that dimκ(Fn) > η. Combining with (4.42) and (4.6), we get P0η,κ(Fn) > 0. Therefore P-dimκ(F ) ≥ η. Let η ↑ Dimκ(F ) to obtain Dimκ(F ) ≤ P-dimκ(F ).
On the other hand, if 0 < η < P-dimκ(F ), then for every sequence of bounded Borel sets {Fk}k≥1 with F ⊆ ∪k≥1Fk, we can find some Fn such that P0η,κ(Fn) > 0. For every fixed θ satisfying 0 < θ < η, we can find some a such that 0 < a < P0η,κ(Fn). Since Pεη,κ(Fn) decreases to P0η,κ(Fn) as ε ↓ 0, we have Pεη,κ(Fn) > a for each ε > 0. Then (4.34) implies the existence of a (κ, ε)-packing (ωi, ti, ri)∞i=1 of Fn such that P∞
i=1ωiriη > a. For each m ≥ 1, define
Km = X
{i:2−m<ri≤21−m}
ωi. (4.43)
Then there must be a positive integer M (ε) such that
KM (ε)> 2(M (ε)−1)θ(1 − 2θ−η)a, (4.44) since otherwise P∞
m=1Km2(1−m)η ≤ P∞
m=12(m−1)θ(1 − 2θ−η)a2(1−m)η = a. Let {i(j)}Jj=1 be a subsequence such that 2−M (ε) < ri(j) ≤ 21−M (ε) for all 1 ≤ j ≤ J and PJ
j=1ωi(j) ≥ 2(M (ε)−1)θ(1 − 2θ−η)a. The fact that (ωi, ti, ri)∞i=1 is a (κ, ε)-packing of Fn implies that
49 P
j≥1ωjκrj(ti− tj) ≤ 1 for all i ≥ 1. Since κε(t) is nondecreasing in ε for fixed t and ri(j) ≤ 21−M (ε) for all 1 ≤ j ≤ J , we see that (ωi(j), ti(j))Jj=1 is a size 2−M (ε) weighted κ-packing of Fn. Therefore,
Nκ(2−M (ε); Fn) ≥ 2(M (ε)−1)θ(1 − 2θ−η)a. (4.45) Since M (ε) ↑ ∞ as ε ↓ 0, by taking logarithms and then limsup as ε ↓ 0, we obtain
dimκ(Fn) = lim
ε↓0
log Zκ(2−M (ε); Fn)
log 2−M (ε) = lim
M →∞
log Nκ(2−M; Fn)
− log 2−M ≥ θ. (4.46) Let θ ↑ η to find Dimκ(F ) ≥ P-dimκ(F ). This completes the proof.
The connection between packing dimension profile Dimκ and the family of measures {Ps,κ}s≥0 can be used to prove the following technical lemma, which will be used later.
Lemma 4.13. If a Borel set F ⊆ RN satisfies Dimκ(F ) > s for some s ≥ 0, then there exists a compact set K ⊆ F such that Dimκ(K ∩ G) ≥ s for all open sets G ⊆ RN with K ∩ G 6= ∅.
Proof. Since Dimκ(F ) > s, for every collection of bounded Borel sets {Fn}n≥1 with F ⊆
∪n≥1Fn, we can find some Fn such that Dimκ(Fn) > s. This implies that Ps,κ(Fn) = ∞.
Then we can use the proof of Theorem 22 of Howroyd [10] to complete the argument.
4.2.1 Main Result
With the box-dimension profile and packing dimension profile defined so far, we are able to present our main result regarding the packing dimension of images of additive L´evy processes.
Theorem 4.14. Let X := {X(t) : t ∈ RN+} be an additive L´evy process in Rd and let dimM denote the upper box-counting dimension. Then for all nonrandom bounded Borel sets F ⊆ RN+:
dimM(X(F )) = dimκ(F ) a.s.; and (4.47)
dimP(X(F )) = Dimκ(F ) a.s. (4.48)
We will prove this theorem in the next section.
4.3 Proof of the Main Theorem
In this section we strive to prove Theorem 4.14. The proof relies on finding upper bounds for the probabilities of the events that the underlying additive L´evy process hits small balls.
In order to estimate such probabilities, we adapt the methods of section 4 of Khoshnevisan, Xiao, and Zhong [23].
4.3.1 Hitting Probabilities
For every δ > 0 and Borel set F ⊆ RN+, let Fδ denote the closed δ-enlargement of F , that is, Fδ := {y ∈ RN+ : dist(y, F ) ≤ δ}, where dist(y, F ) := infx∈F|x − y|. Let λd denote the Lebesgue measure on Rd. For fixed ε > 0 and Borel set F ⊆ RN+, we want to estimate Eλd (X(F ))ε. It turns out that this quantity is equal to the probability of the event that the underlying additive L´evy process hits small balls.
Proposition 4.15. For every bounded Borel set F ⊆ RN+ and small number ε > 0, we have
Eλd (X(F ))ε ≤ Cεd
Zκ(ε; F ), (4.49)
where Zκ(ε; F ) is defined in (4.7) and the constant C depends only on N and d.
Proof. We use the idea of the proof for Theorem 2.1 of [23]. Choose and fix some ∆ /∈ RN+. For each δ > 0, let Tδdenote some measurable, (QN+∩ Fδ) ∪ {∆}-valued function on Ω such that Tδ 6= ∆ if and only if |X(Tδ)| < ε/2. Then for each k > 0 define a set function
µδ,k(·) := Pλd{Tδ ∈ ·, Tδ6= ∆, |X(0)| ≤ k}
Pλd{Tδ 6= ∆, |X(0)| ≤ k} . (4.50) Since B(0, ε) is open and Xj is right continuous for all 1 ≤ j ≤ N , an application of Fubini’s theorem implies that
Pλd{Tδ 6= ∆, |X(0)| ≤ k} = Pλd{X(Fδ) ∩ B(0, ε/2) 6= ∅, |X(0)| ≤ k}
= Z
[−k,k]d
P{X(Fδ) ∩ B(−x, ε/2) 6= ∅} dx
= Eλd (X(Fδ) ⊕ B(0, ε/2)) ∩ [−k, k]d.
(4.51)
Since the last term is positive and finite, we see that µδ,k is a probability measure on Fδ. Now for every A ⊆ Π and nonnegative measurable or bounded measurable function f : Rd → R, consider the multiparameter martingale {MAµδ,kf (t)}t∈RN
+ defined in (3.58).
Then
51
Pλd-almost surely, where the last equality follows from Lemma 3.28. We obtain the following by taking a supremum over all positive rational numbers and noticing that Tδ∈ QN+∪ {∆}
and |X(Tδ)| < ε/2 on {Tδ 6= ∆, |X(0)| ≤ k}:
Pλd-almost surely. Summation over all A ⊆ Π gives X
Pλd-almost surely. Note that the null set is independent of the choice of ε > 0. By squaring and taking Pλd-expectations on both sides of the above inequality, and noticing the special choice of µδ,k, we obtain the following:
Eλd
where the last inequality follows from Jensen’s inequality. On one hand, the elementary inequality (Pn
k=1ak)2 ≤ nPn
k=1a2k shows that
Eλd
where we used Lemma 3.32 to derive the last inequality. On the other hand, we choose our f to be the Fourier transform of the ε-scaled P´olya distribution, that is,
fε(x) =
where a+:= max(a, 0) for all real numbers a. Then the Fourier transform of fε is given by fˆε(ξ) = particular, combining with (4.19), we get
Z application of Prohorov’s theorem ([15], Chapter 6, Theorem 2.5.1), together with the continuity of infxPt−sfε/2(x), imply that there exists some µ ∈ P( ¯F ), such that, along
53 Here ¯F denotes the closure of F . Now we combine (4.55), (4.56), (4.60), (4.61) and (4.62) to get
Eλd (X(F ))ε/2 · Z Z
|x|≤ε/2inf Pt−sfε(x)µ(dt)µ(ds) ≤ 16N(2π)−d(8C0)dεd, (4.63) where C0 =R
Ry−4(1 − cos y)2dy < ∞.
Finally, we notice that 1 − (2ε)−1|z| ≥ 12 whenever z ∈ B(0, ε). Consequently,
2−d1B(0,ε)(z) ≤ fε(z) ∀ z ∈ Rd. (4.64) Let z := eX(t − s) + x, using (3.42) and taking expectation to find
Pt−sfε(x) = E[fε( eX(t − s) + x)]
≥ 2−dE[1B(0,ε)( eX(t − s) + x)] = 2−dP{| eX(t − s) + x| ≤ ε}
≥ 2−dP{| eX(t − s)| < ε/2} · 1{|x| ≤ ε/2},
(4.65)
where the last line follows from the triangle inequality. Take an infimum over |x| ≤ ε/2 to get
inf
x∈Rd,|x|≤ε/2Pt−sfε(x) ≥ 2−dP{| eX(t − s)| < ε/2} = 2−dκε/2(t − s). (4.66) We combine the above inequality with (4.63) and change ε/2 to ε to obtain
Eλd (X( ¯F ))ε ≤ 16N(2π)−d(8C0)d2d(2ε)d
infν∈P( ¯F )RR κε(t − s)ν(dt)ν(ds). (4.67) The proposition is proved, thanks to Lemma 4.3.
4.3.2 Proof of Theorem 4.14
With the help of Proposition 4.15 we can adapt the proof of Theorem 2.7 of Khosh-nevisan, Schilling, and Xiao [19] for one-parameter L´evy processes to the current multipa-rameter setting.
Proof of Theorem 4.14: (4.47). Recall that the upper box dimension of a bounded Borel set E ⊂ Rd is defined by
dimM(E) := lim sup
r↓0
log KE(δ)
− log δ , (4.68)
where KE(δ) is the Kolmogorov capacity of E, that is, KE(δ) is the maximal number m of points x1, . . . , xm in E such that mini6=j|xi − xj| ≥ δ. It follows immediately that δdKE(δ) ≤ λd(Eδ) for all δ > 0, where Eδ is the closed δ-enlargement of E. In fact, we can
find k := KE(δ) points x1, . . . , xk in E such that B(x1, δ/2), . . . , B(xk, δ/2) are disjoint and sufficiently small ε > 0. We combine this with (4.69) and use the Markov inequality to obtain
P{KX(F )(ε) ≥ ε−(η+θ)} = O(εθ) (ε ↓ 0), (4.70) for all θ ∈ (0, 1). We apply (4.70) with ε = 2−n and use Borel-Cantelli lemma in order to get
KX(F )(2−n) = O(2n(η+θ)) (n → ∞) a.s. (4.71) A standard monotonicity argument (see for example (3.14) of [5]) shows that with probabil-ity one dimM(X(F )) ≤ η + θ. Now we first let θ ↓ 0 and then η ↓ dimκ(F ) (along countable sequences) to deduce that
dimM(X(F )) ≤ dimκ(F ) a.s. (4.72)
In order to prove the converse inequality, we use a result from Khoshnevisan and Xiao [22]. According to Theorem 4.1 of [22], the upper box dimension of a bounded Borel set E can also be computed by
dimM(E) = sup
Then for any η < dimκ(F ), (4.4) shows that there exist a sequence of positive numbers {εn}n≥1 and a sequence of probability measures {µn}n≥1 ⊂ P(F ) such that εn ↓ 0 as This, Fatou’s lemma, and (4.74) together shows that
lim
55 Therefore, (4.73) implies that dimM(X(F )) ≥ η almost surely. Let η ↑ dimκ(F ) to see that dimM(X(F )) ≥ dimκ(F ) almost surely.
In order to prove equation (4.48) we need the following technical lemma.
Lemma 4.16. For all compact sets F ⊆ RN+, dimP(X(F )) = dimP(X(F )) almost surely, where ¯E denotes the closure of a set E.
Proof. It is straightforward to see that dimP(X(F )) ≥ dimP(X(F )) almost surely. Then we recall that X(t) =PN
j=1Xj(tj) for t ∈ RN+ and each Xj is a one-parameter L´evy process.
Notice that with probability one
{t ∈ F : X is discontinuous at t} ⊆
N
[
j=1
t ∈ F : Xj(tj) 6= Xj(tj−) . (4.77)
Furthermore, for each 1 ≤ j ≤ N , let Yj(t) := P
1≤l≤N,l6=jXl(tl) + Xj(tj−). Since there are at most countably many jumps for each Xj, and dimP is σ-stable, it follows that for all 1 ≤ j ≤ N ,
dimP {Yj(t) : t ∈ F, Xj(tj) 6= Xj(tj−)} =
dimP {X(t) : t ∈ F, Xj(tj) 6= Xj(tj−)} ≤ dimP(X(F )) a.s.
(4.78)
Since X(F ) ⊆ X(F ) ∪ (∪Nj=1{Yj(t) : t ∈ F, Xj(tj) 6= Xj(tj−)} almost surely, the σ-stability of dimP implies that dimP(X(F )) ≤ dimP(X(F )) almost surely. This completes the proof.
Proof of Theorem 4.14: (4.48). Definition 4.7 implies that for each η > Dimκ(F ) there exists a sequence of bounded Borel sets {Fn}n≥1 such that
F ⊆
∞
[
n=1
Fn and sup
n≥1
dimκ(Fn) < η. (4.79)
Since X(F ) ⊆ ∪n≥1X(Fn), (4.47) implies that dimP(X(F )) ≤ sup
n≥1
dimM(X(Fn)) = sup
n≥1
dimκ(Fn) < η a.s. (4.80) Let η ↓ Dimκ(F ) to obtain dimP(X(F )) ≤ Dimκ(F ) almost surely.
Next we prove the converse inequality, that is, dimP(X(F )) ≥ Dimκ(F ) almost surely.
Next we prove the converse inequality, that is, dimP(X(F )) ≥ Dimκ(F ) almost surely.