Local time as an element of the Sobolev space

For a centered Gaussian random ?eld taking its values in R^d, we investigate the existence of a local time as a generalized functional, i.e an element of some Sobolev space. We give the sfficient condition for such an existence in terms of the field covariation and apply it in several examples: the...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Datum:2007
1. Verfasser: Rudenko, A.V.
Format: Artikel
Sprache:Englisch
Veröffentlicht: Інститут математики НАН України 2007
Online Zugang:https://nasplib.isofts.kiev.ua/handle/123456789/4508
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Назва журналу:Digital Library of Periodicals of National Academy of Sciences of Ukraine
Zitieren:Local time as an element of the Sobolev space / A.V. Rudenko // Theory of Stochastic Processes. — 2007. — Т. 13 (29), № 3. — С. 65–79. — Бібліогр.:12 назв.— англ.

Institution

Digital Library of Periodicals of National Academy of Sciences of Ukraine
_version_ 1859861041501962240
author Rudenko, A.V.
author_facet Rudenko, A.V.
citation_txt Local time as an element of the Sobolev space / A.V. Rudenko // Theory of Stochastic Processes. — 2007. — Т. 13 (29), № 3. — С. 65–79. — Бібліогр.:12 назв.— англ.
collection DSpace DC
description For a centered Gaussian random ?eld taking its values in R^d, we investigate the existence of a local time as a generalized functional, i.e an element of some Sobolev space. We give the sfficient condition for such an existence in terms of the field covariation and apply it in several examples: the self-intersection local time for a fractional Brownian motion and the intersection local time for two Brownian motions.
first_indexed 2025-12-07T15:45:37Z
format Article
fulltext Theory of Stochastic Processes Vol. 13 (29), no. 3, 2007, pp. 65–79 UDC 519.21 ALEXEY V. RUDENKO LOCAL TIME AS AN ELEMENT OF THE SOBOLEV SPACE For a centered Gaussian random field taking its values in d, we investigate the existence of a local time as a generalized functional, i.e an element of some Sobolev space. We give the sufficient condition for such an existence in terms of the field covariation and apply it in several examples: the self-intersection local time for a fractional Brownian motion and the intersection local time for two Brownian motions. Introduction Let (T, B) be a measurable space with finite measure ν on it. Let ξ(t), t ∈ T , be a centered Gaussian random field with values in R d. Consider the integrals Lε =∫ T fε(ξ(t))ν(dt), where fε approximates, in a sense, δ0 as ε → 0+ (δ0 denotes the mea- sure with weight 1 concentrated in 0 ∈ R d). If the limit of Lε exists, we can say that the local time exists (in some sense) and we call this limit as the local time of ξ at 0 ∈ R d. We say that Lε is the approximation for the local time of ξ at 0 ∈ R d. Our aim is to investigate the existence of local time so the question is: When does Lε converge as ε → 0+? The condition for L2-convergence in terms of covariation is relatively easy to obtain (see [2],[9]). We consider the convergence in some Sobolev space D2,α defined in the book of Watanabe [10]. The interest in this topic arises from the possibility to define the local time as an element of the Sobolev space with negative α, while it does not exist in the usual sense. It is possible to determine the exact smoothness of a local time in the sense of spaces D2,α. Sometimes the local time does not exist as an element of any Sobolev space, but if we modify the approximation by subtracting its mathematical ex- pectation, we obtain the convergence to the so-called renormalized local time. A classical example of renormalization is the renormalization of the self-intersection local time for a two-dimensional Brownian motion. This result can be found, for instance, in [7]. We introduce some kind of renormalization, and our definition includes this classical setup. In [5], Imkeller, Perez-Abreu, and Vives found an explicit form of the Itô–Wiener expansion for an approximation of the k-fold self-intersection local time of a Brown- ian motion. They estimated Hermite polynomials in this expansion and obtained the convergence of the approximation in some Sobolev space. Albeverio, Hu and Zhou [8] employed a similar approach, but their result is opposite. They proved that the renor- malized self-intersection local time of a planar Brownian motion is not differentiable in the Watanabe–Meyer sense, i.e. it does not belong to the space D2,1. Another question was considered in [11] where Nualart and Hu deal with the renormalization of the 2-fold self-intersection local time for a fractional Brownian motion. They proved the conver- gence of this renormalization in L2 (which is same as in D2,0) under some conditions. 2000 AMS Mathematics Subject Classification. Primary 60H07. Key words and phrases. Local time, Itô–Wiener expansion, Sobolev spaces, Gaussian random field, fractional Brownian motion. Research was partially supported by the Ministry of Education and Science of Ukraine, project GP/F13/0095. 65 66 ALEXEY V. RUDENKO Later on, they gave a sufficient condition for the existence of the same renormalized local time as an element of D2,α for positive α [12]. In [2], we have found a suitable integral representation for the expectation of Itô– Wiener expansion terms for local time approximations and use it here to derive the condition for the convergence via the covariation of a field. This condition is both neces- sary and sufficient if the underlying Gaussian field is centered and we consider the local time at zero. We also get a similar condition for the convergence of a renormalized local time. In our context, the renormalization is the subtraction of several leading terms in the Itô–Wiener expansion. We also consider some applications including the 2-fold self-intersection local time for a fractional Brownian motion. We have to mention that our main ideas were already present in [2]. Let us mention that there are other ways to define the local time as an element of some generalized space of random variables when it does not exist in the usual sense. In [1,4], the authors consider the Hida distributions instead of Sobolev spaces. The structrure of the exposition is the following. In the first section, the necessary and sufficient condition for the convergence is given in terms of a sequence of integrals. In the second section, this condition is rewritten in terms of one integral. In the third section, we introduce the renormalization using a generalization of results from the previous sections (with an additional condition on approximations). The last section is devoted to applications. 1. Necessary and sufficient condition for the existence of local time Let’s describe briefly the construction of Sobolev spaces and related objects (as it appears in [10]). To do this, we need to define a probability space more precisely and impose some conditions on ξ. Let B be a Banach space with Gaussian probability measure on it, which plays the role of a probability space for ξ. Denote the covariation matrix function for ξ as K(s, t), s, t ∈ T . Suppose that each ξ(t) is linear as a function of ω ∈ B, and σ({ξ(t), t ∈ T }) coincides with B(B), a Borel σ-algebra on B. Suppose also that ξ is jointly measurable (as a function of both t ∈ T and ω ∈ B) and that ξ is not degenerate ν-a.e.: ν({t| detK(t, t) = 0}) = 0. Let Hn, n = 0, 1, . . . , be a subspace of L2(B, μ) generated by all polynomials on B with degree less or equal n. By polynomials on B, we mean functions of the form P (l1(x), l2(x), . . . , lk(x)), where P is a polynomial in k variables, and l1, . . . , lk ∈ B∗. By Gn, we denote an orthogonal supplement of Hn−1 in Hn for n ∈ N. Let Pn be a projector on Gn. Then we have ∞⊕ n=0 Gn = L2(B, μ) (polynomials are dense in L2(B, μ)). The corresponding decomposition h = ∞∑ n=0 Pnh, h ∈ L2(B, μ) is called the Itô–Wiener expansion (or the chaos decomposition [6]). We can define a set of norms for each h ∈ Hn as ‖h‖2 2,α = ∞∑ n=0 (1 + n)α ‖h‖2 2 where ‖ · ‖2 is the norm in L2(B, μ). Then D2,α is the completion of ∞∪ n=0 Hn by the norm ‖·‖2,α. We define an approximation for δ0 as fε(x) = 1 εd f( x ε ), f ∈ S(Rd), f � 0, ∫ Rd f(x)dx = 1. LOCAL TIME AS AN ELEMENT OF THE SOBOLEV SPACE 67 We denote G(s, t) = K−1/2(t, t)K(s, t)K−1/2(s, s), In = ∫ T ∫ T ∫ Sd ‖G(s, t)x‖n σ(dx) ν(ds)√ detK(s, s) ν(dt)√ detK(t, t) , where Sd is a unit sphere in R d, and σ is the uniform surface measure on Sd. The matrix G is defined as ν × ν-a.e., thus the integral is well defined. Theorem 1. Fix α ∈ R. The following statements are equivalent: 1. Lε → L, ε → 0+ in D2,α, 2. lim ε→0+ ‖Lε‖2,α < +∞, 3. ∞∑ n=0 I2n(2n + 1)α+d/2−1 < +∞. Proof. We will prove the following implications 1 ⇒ 2 ⇒ 3 ⇒ 1. 1 ⇒ 2 is obvious. 2 ⇒ 3. Let’s write a formula for ‖Lε‖2,α using the Itô–Wiener expansion. By def- inition, Lε is bounded so Lε ∈ L2(Ω) and we may write the Itô–Wiener expansion Lε = ∞∑ n=0 an(ε). We have ‖Lε‖2,α = ∞∑ n=0 (n + 1)αE(an(ε))2. We know that (from [2]) E(an(ε))2 = = ∫ T ∫ T ∫ Rd ∫ Rd f(x)f(y) ∫ Rd ∫ Rd eiε((x,u)+(y,w))qn(s, t, u, w)dudwdxdyν(ds)ν(dt), where qn(s, t, u, w) = (2π)−2d (−1)n n! (K(s, t)u, w)ne− 1 2 (K(s,s)u,u)e− 1 2 (K(t,t)w,w). From (2), we get lim ε→0+ E(a0(ε))2 < +∞. On the other hand, E(a0(ε))2 = (∫ T ∫ Rd f(x)e− 1 2 ε(K−1(t,t)x,x) (2π)−d/2√ detK(t, t) dxν(dt) )2 . We can see that E(a0(ε))2 is a monotonous function of ε with maximum at ε = 0. We conclude that ∫ T ν(dt)√ det K(t, t) < +∞. Let’s estimate E(an(ε))2: (1) E(an(ε))2 � ∫ T ∫ T ∫ Rd ∫ Rd f(x)f(y) ∫ Rd ∫ Rd |qn(s, t, u, w)| dudwdxdyν(ds)ν(dt) = = ∫ T ∫ T ∫ Rd ∫ Rd |qn(s, t, u, w)| dudwν(ds)ν(dt) = Jn. 68 ALEXEY V. RUDENKO We want to transform Jn. ∫ Rd ∫ Rd |qn(s, t, u, w)| dudw = (2π)−2d 1 n! ∫ Rd ∫ Rd |(K(s, t)u, w)|ne− 1 2 (K(s,s)u,u)e− 1 2 (K(t,t)w,w)dudw = = (2π)−2d n! √ detK(t, t) detK(s, s) ∫ Rd ∫ Rd |(G(s, t)u, w)|ne− 1 2 ‖u‖2− 1 2‖w‖2 dudw = = (2π)−2d n! √ det K(t, t) detK(s, s) ∫ Rd e− 1 2 ‖u‖2 (∫ Rd |(G(s, t)u, w)|ne− 1 2‖w‖2 dw ) du = = (2π)−2d n! √ detK(t, t) detK(s, s) ∫ Rd e− 1 2‖u‖2 ‖G(s, t)u‖n (∫ Rd |w1|n e− 1 2‖w‖2 dw ) du = = (2π)−2d n! √ detK(t, t) detK(s, s) ∫ Rd |w1|n e− 1 2‖w‖2 dw· · ∫ R+ e− 1 2 r2 rn+d−1 (∫ Sd ‖G(s, t)x‖n σ(dx) ) dr = = (2π)−2d n! √ detK(t, t) detK(s, s) ∫ Rd |w1|n e− 1 2‖w‖2 dw· · ∫ R+ e− 1 2 r2 rn+d−1dr ∫ Sd ‖G(s, t)x‖n σ(dx) = = C(n, d)√ detK(t, t) detK(s, s) ∫ Sd ‖G(s, t)x‖n σ(dx). As we can see, Jn = C(n, d)In, where C(n, d) = (2π)−2d n! ∫ Rd |w1|n e− 1 2‖w‖2 dw ∫ R+ e− 1 2 r2 rn+d−1dr = = 2 n! (2π)−(3d+1)/2 ∫ ∞ 0 vn 1 e−v2 1/2dv ∫ ∞ 0 rn+d−1e−r2/2dr = = 2n+(d−1)/2 n! (2π)−(3d+1)/2 ∫ ∞ 0 x(n+d−2)/2e−xdx ∫ ∞ 0 x(n−1)/2e−xdx = = 2n+(d−1)/2 n! (2π)−(3d+1)/2Γ( n + d 2 )Γ( n + 1 2 ). We need to find the asymptotics of C(n, d) for large n. We can use the asymptotics of the gamma function (the Stirling formula), Γ(x) ∼ √ 2π/x (x e )x , x → +∞, LOCAL TIME AS AN ELEMENT OF THE SOBOLEV SPACE 69 and the equality n! = Γ(n + 1). We have C(n, d) = 2n+(d−1)/2 Γ(n + 1) (2π)−(3d+1)/2Γ( n + d 2 )Γ( n + 1 2 ) ∼ ∼ 2(d−1)/2(2π)−(3d+1)/22n √ 8π n + d · · ( n + d 2e )(n+d)/2( n + 1 2e )(n+1)/2( e n + 1 )n+1 = = 2(d−1)/2(2π)−(3d+1)/22−(d+1)/2(8π)1/2e−(d+1)/2· · √ 1 n + d (1 + d + 1 n + 1 )(n+1)/2(n + d)(d−1)/2 ∼ ∼ (2π)−3d/2nd/2−1, n → +∞. The integral ∫ Sd ‖G(s, t)x‖n σ(dx) can be calculated explicitly in terms of eigenvalues of G(s, t), but it is more convenient for us to leave it in the form of integral. Indeed, we can use that the expression taken to the n-th power under integral, which is ‖G(s, t)x‖, is less or equal than 1 when ‖x‖ = 1 (from the definition of G). Therefore, the condition I0 < +∞ (this is our assumption) gives us In < +∞, n ∈ N. Using the Lebesgue theorem of dominated convergence, we get lim ε→0+ E(an(ε))2 = Jn for even n and lim ε→0+ E(an(ε))2 = 0 for odd n. We can see now that +∞∑ n=0 J2n(2n + 1)α ≤ lim ε→0+ ‖Lε‖2,α < +∞. Using the asymptotics of C(n, d), it is easy to complete the proof of this implication. 3 ⇒ 1. Using the Hölder inequality, we get I2n+1 � M √ I2nI2n+2, n = 0, 1, . . . (the constant is independent of n). So if we include I2n+1 into the infinite sum, then ∞∑ n=0 In(n + 1)α+d/2−1 < +∞. Using the formula for Ean(ε1)an(ε2) (see [2]) and the condition In < +∞, we conclude that Ean(ε1)an(ε2) is convergent to Jn for even n and to 0 for odd n. Thus, an(ε) is convergent in L2(Ω). Define an(0) = L2 − lim ε→0+ an(ε). From inequality (1), we get E(an(ε))2 � Jn = C(n, d)In Consequently, ∞∑ n=0 (n + 1)αE(an(0))2 � ∞∑ n=0 (n + 1)αC(n, d)In < +∞, and there exists L = ∞∑ n=0 an(0) ∈ D2,α. Now all left to do is to use a uniform bound for tails of the Itô–Wiener expansion for approximations. We get ‖L − Lε‖2,α = ∞∑ n=0 (n + 1)αE(an(0) − an(ε))2 → 0, ε → 0 + . Note that if statement (3) in the theorem is not fulfilled, then it is not possible to construct the local time as an element of D2,α using approximations of this kind and some Sobolev space with smaller α. Indeed, suppose that we have convergence to some 70 ALEXEY V. RUDENKO L in D2,α0 with α0 < α (while there is no convergence for our fixed α). Then using same arguments as in theorem’s proof, we can prove ‖L‖2,α = +∞∑ n=0 E(a2n(0))2(2n + 1)α = +∞∑ n=0 J2n(2n + 1)α = ∞. Thus L can not be an element of D2,α. For some values of α, it is possible to simplify the condition of the theorem. Corollary 1. If α < − d 2 , then statements in the theorem above are equivalent to∫ T ν(dt)√ det K(t, t) < +∞. Proof. Note that I0 = σ(Sd) (∫ T ν(dt)√ detK(t,t) )2 . Obviously, statement (3) of the theorem yields I0 < +∞. This proves one side of equivalence. Now suppose that I0 < +∞. As we have seen already, In � I0, n ∈ N. Taking into account that α + d 2 − 1 < −1, we have ∞∑ n=0 I2n(2n + 1)α+d/2−1 < +∞. 2. Integral form of the condition for local time existence We want to give a different form of the condition for the local time existence from Theorem 1, using the following lemma. Lemma 1. Let (E, F) be ameasurable space with measure μ on it and let f be a mea- surable function on E with values in [0, 1]. Denote: Jn = ∫ E (f(x))nμ(dx), pβ : [0, 1] → [1,∞], β ∈ R, pβ(z) = ⎧⎨ ⎩ zβ, β < 0 1 − ln z, β = 0 1, β > 0. For any γ ∈ R. the following statements are equivalent: 1. ∞∑ n=0 Jn(n + 1)γ < +∞, 2. ∫ E p−γ−1(1 − f(x))μ(dx) < +∞. Proof. Note that if γ < −1, then both the sum and the integral converge when μ is finite or diverge when μ(E) = +∞. So it is enough to consider the case γ � −1. It is obvious that the sequence (n +1)γ may be replaced by the sequence of positive numbers cn, n ∈ N, if cn ∼ (n + 1)γ . We use this and take cn to be coefficients near qn in the Taylor series of p−γ−1(1 − q) at the point q = 0. We can write an explicit form for cn: γ = −1 : p−γ−1(1 − q) = 1 − ln(1 − q) = 1 + ∞∑ n=1 1 n qn = ∞∑ n=0 cnqn , γ > −1 : p−γ−1(1 − q) = (1 − q)−γ−1 = ∞∑ n=0 ( n∏ k=1 ( γ k + 1) ) qn = ∞∑ n=0 cnqn. LOCAL TIME AS AN ELEMENT OF THE SOBOLEV SPACE 71 All sums converge if |q| < 1. It can be verified that cn satisfy the needed property. The condition f ∈ [0, 1] allows us to write ∞∑ n=0 Jncn = ∫ E ∞∑ n=0 (f(x))ncnμ(dx) = ∫ E p−γ−1(1 − f(x))μ(dx). Lemma 1 is proved. Now we apply Lemma 1 to get a different form of the condition of Theorem 1. Note that, for α < − d 2 , we have already done this above. Corollary 2. Fix α ∈ R. The following statements are equivalent: 1. Lε → L, ε → 0+ in D2,α, 2. ∫ T ∫ T ∫ Rd p−α−d/2(1 − ‖G(s, t)x‖)σ(dx) ν(ds)√ detK(s, s) ν(dt)√ detK(t, t) < +∞. Proof. We choose μ(dx, ds, dt) = σ(dx) ν(ds)√ det K(s,s) ν(dt)√ detK(t,t) to be the measure and ‖G(s, t)x‖ to be the function f from Lemma 1. Now Corollary 2 is a direct conse- quence of Theorem 1 and Lemma 1 (we also need the fact that we can include terms with odd numbers into sum (3) from Theorem 1, but we showed it in the proof of this Theorem). 3. Renormalization We want to define the renormalization for local time approximations, but before we need to prove a more generalized form of Theorem 1. If we define local time approxima- tions more precisely taking f to be a Gaussian density, f(x) = 1 (2π)d/2 e− ‖x‖2 2 , then we can prove a result similar to Theorem 1 in the case where the convergence of Lε is in the sense of some seminorm on the space of square integrable random variables on our probability space (its values may be infinite). We define this seminorm as S(η, {cn}) = ∞∑ n=0 cnEη2 n where η is a square integrable random variable, η = ∞∑ n=0 ηn is the Itô–Wiener expansion of η and cn � 0, n = 0, 1, . . . , is an arbitrary sequence of non-negative numbers. Norms in Sobolev spaces can be considered as partial cases of this seminorm. Theorem 2. The following statements are equivalent 1. Lε → L, ε → 0+ with respect to seminorm S(·, {cn}), 2. lim ε→0+ S(Lε, {cn}) < +∞, 3. ∞∑ n=0 I2ncn(2n + 1)d/2−1 < +∞. Proof. 1 ⇒ 2 is obvious. 72 ALEXEY V. RUDENKO 2 ⇒ 3. As in Theorem 1, we consider the Itô–Wiener expansion of Lε = ∞∑ n=0 an(ε). Using the explicit form of an, we get E(an(ε))2 = = ∫ T ∫ T ∫ Rd ∫ Rd ∫ Rd ∫ Rd f(x)f(y)eiε((x,u)+(y,w))qn(s, t, u, w)dudwdxdyν(ds)ν(dt) = = (2π)d ∫ T ∫ T ∫ Rd ∫ Rd qn(s, t, u, w) exp(−ε 2 (‖u‖2 + ‖w‖2))dudwν(ds)ν(dt). We note that the function f is even and qn(s, t, u,−w) = (−1)nqn(s, t, u, w). As a consequence, the expression is zero for odd n. For even n, the function qn is non- negative, so the expression above is a monotonous function of ε. We can see now that the same considerations as in Theorem 1 are valid in this case (we choose the smallest even n such that cn > 0, then all reasonings are the same with the substitution of I0 with In). 3 ⇒ 1. Because members of the Itô–Wiener expansion with odd index are zero, the proof is the same as that of Theorem 1. Such a kind of seminorms is interesting because of the following considerations. If we take cn = 0, n < 2n0; cn = (1 + n)α, n � 2n0, where α - arbitrary real number, n0 - arbitrary integer, then the convergence with respect to seminorm is equivalent to the convergence of renormalized local time approximations Lε,n0 (first 2n0 members of the Itô–Wiener expansion of Lε are subtracted) in D2,α. We can also apply Lemma 1 to this case. Corollary 3. Fix α ∈ R, n0 ∈ N, and {cn} as above. Then the following statements are equivalent: 1. Lε,n0 → L, ε → 0+ in D2,α, 2.∫ T ∫ T ∫ Rd ‖G(s, t)x‖2n0 p−α−d/2(1 − ‖G(s, t)x‖)σ(dx) ν(ds)√ detK(s, s) ν(dt)√ detK(t, t) < +∞. Proof. We let μ(dx, ds, dt) = ‖G(s, t)x‖2n0 σ(dx) ν(ds)√ det K(s,s) ν(dt)√ detK(t,t) to be the measure from Lemma 1, and let the function f be the same as that in Corollary 2. Now we apply Theorem 2 together with Lemma 1. The only thing we need to check (like in Corollary 2) is why we can add terms with odd numbers to condition (3) for the convergence in Theorem 2. But here, because of the form of cn, we can use the same arguments as we did in Theorem 1. It is interesting to observe here some immediate conclusions about this integral condi- tion. Possible singularities which come from (detK(s, s) detK(t, t))−1/2 can be refined with ‖G(s, t)x‖2n0 . We can expect all kinds of situations here because G(s, t), s, t ∈ T and K(s, s), s ∈ T are in some sense independent (i.e. we can define one independently of the other). Also we can not obtain the convergence for bigger α using renormalization (it is obvious from the definition of renormalization, so we only confirm that) because we can not refine the singularities of p−α−d/2(1 − ‖G(s, t)x‖) with ‖G(s, t)x‖2n0 . 4. Applications In this section, we apply our results to some Gaussian fields. First of all, we have to mention when the construction of the needed probability space for our Gaussian field is LOCAL TIME AS AN ELEMENT OF THE SOBOLEV SPACE 73 possible. If T is a separable metric space and our Gaussian random field is stochastically continuous, then we have the sequence of values {ξ(tn), n ∈ N} which generate a σ-field of the field. Let H be a subspace of L2(Ω) generated by linear combinations of {ξ(tn)}. It is well known (see [3],[6]) that we can construct (as a part of the abstract Wiener space) a Banach space B and a probability Gaussian measure μ on it such that H is the space of admissible shifts for μ (subspace of all linear functionals on B). In this setting, (B, B(B), μ) is a replacement for the probability space, and our field will be defined naturally on it as linear functionals on B. That is exactly the construction we need. In all cases, we suppose that the function f from the approximation is a standart Gaussian density and use corollaries 2,3. All our assumptions can be easily verified for processes below. So, in proofs, we need to verify the corresponding integral conditions only. We start from the case of fractional Brownian motion. Suppose {X(t) ∈ R d, t ∈ [0, 1]} is a fractional Brownian motion with the Hurst parameter H ∈ (0, 1). Coordinates are independent and each has the covariation function rH(s, t) = s2H+t2H−|s−t|2H 2 . Example. Let T = [0, 1]; ν(dt) = dt; ξ(t) = X(t). In this case, the local time does not exist in any Sobolev space if d ≥ 1 H . If d < 1 H , then the limit exists in D2,α for α < 1 2H − d 2 and doesn’t exist in other cases. The value of n0 does not affect the convergence of Lε,n0 (renormalization does not work). Example. Let T = [1/2, 1]; ν(dt) = dt; ξ(t) = X(t). The local time exists if α < 1 2H − d 2 (and doesn’t exist in other cases). Renormalization does not work. Proof. We have K(s, t) = rH(s, t)I G(s, t) = rH(s, t)(rH(s, s)rH(t, t))−1/2I = |s|2H + |t|2H − |t − s|2H 2tHsH I, so the integral with respect to x in the integral condition (2) from Corollary 3 can be dropped. This condition now has form∫ 1 0 ∫ 1 0 ( s2H + t2H − |s − t|2H 2sHtH )2n0 · · ( |s − t|2H − (sH − tH)2 2sHtH )−α−d/2 (st)−dHdtds. We use the polar coordinates: s = r sin φ, t = r cosφ. We can extend or reduce the domain of integration: {s2 + t2 < 1, s > 0, t > 0} ⊂ ⊂ {s < 1, t < 1, s > 0, t > 0} ⊂ ⊂ {s2 + t2 < 2, s > 0, t > 0}. For both cases, we get similar integrals and the same conditions for their finiteness. We consider only the case of a reduced domain (the second one can be treated in the same way):∫ 1 0 ∫ π/2 0 ( (cosφ)2H + (sin φ)2H − | cosφ − sin φ|2H 2(cosφ sin φ)H )2n0 · · ( | cosφ − sin φ|2H − (cosH φ − sinH φ)2 2 cosH φ sinH φ )−α−d/2 · · (cosφ sin φ)−dHdφr1−2dHdr < +∞ 74 ALEXEY V. RUDENKO Our integral becomes a product of two integrals. The integral with respect to r is finite, if (and only if) dH < 1. The integral with respect to φ can be infinite only near φ = 0, φ = π 2 and φ = π 4 , because the integrand is continuous and finite away from these points. Note that the expression taken to power n0 is bounded. Since |cosφ − sinφ|2H − (cosH φ − sinHφ)2 2cosHφsinHφ = = 1 + | sinφ|H 2| cosφ|H + |cosφ − sinφ|2H − | cosφ|2H 2cosHφsinHφ ∼ ∼ 1 − 2H | cosφ|2H−1 sin φ 2cosHφsinHφ → 1 as φ → 0+, then integral with respect to φ is finite in a neighbourhood of φ = 0 if dH < 1 (because only one important multiplier under the integral is (sinφ)−dH ). The behavior of the integral in a neighbourhood of φ = π 2 is similar. Now we investigate the behaviour in a neighbourhood of φ = π 4 : | cosφ − sin φ|2H − (cosH φ − sinH φ)2 ∼ 2H |φ − π 4 |2H − H22H−1(φ − π 4 )2 ∼ 2H |φ − π 4 |2H , φ → π 4 . From this relation, we get the sufficient condition for the integral to be finite 2H(−α − d 2 ) > −1. It is also necessary, because the expression taken to power n0 can not refine this singularity, as we mentioned earlier. Now let T = [1/2, 1]. In this case, the points s = 0 and t = 0 lie away from the integration domain. So, if we make same calculations as above, we get that φ = 0 and φ = π 2 are away from the integration domain. The only one condition for the finiteness here is α < 1 2H − d 2 . Now we turn to the interesting case of the self-intersection for a fractional Brownian motion. Here, we generalize the results from [5,8,11,12]. Example. Let T = [0, 1]2; ν(dt) = dt1dt2; ξ(t) = X(t1)−X(t2). The renormalized local time exists as an element of the Sobolev space D2,α if and only if α < 1 H − d 2 , d < 3 2H , n0 > dH − 1 2(1 − H) . Example. Let T = [0, 1/3] × [2/3, 1]; ν(dt) = dt1dt2; ξ(t) = X(t1) − X(t2). The local time exists in D2,α for α < 1 H − d 2 . Renormalization does not work. Proof. The covariation in this case is given by K(s, t) = 1 2 (|s1 − t2|2H + |s2 − t1|2H − |s1 − t1|2H − |s2 − t2|2H)I, and the correlation equals G(s, t) = g(s, t)I, where g(s, t) = |s1 − t2|2H + |s2 − t1|2H − |s1 − t1|2H − |s2 − t2|2H 2|s1 − s2|H |t1 − t2|H . The integral from Corollary 3 has the form∫ 1 0 ∫ 1 0 ∫ 1 0 ∫ 1 0 |g(s, t)|2n0 |1 − |g(s, t)||−α− d 2 |s1 − s2|−dH |t1 − t2|−dHdt1dt2ds1ds2. By introducing the new variables x = s1 − t1, y = s2 − s1, z = t1 − t2 and extending the integration domain, we simplify this integral to∫ 1 −1 ∫ 1 −1 ∫ 1 −1 ∣∣∣∣D2H(x, y, z) 2|yz|H ∣∣∣∣ 2n0 ∣∣∣∣1 − |D2H(x, y, z)| 2|yz|H ∣∣∣∣ −α− d 2 |y|−dH |z|−dHdxdydz, LOCAL TIME AS AN ELEMENT OF THE SOBOLEV SPACE 75 where Dγ(x, y, z) = |x + y|γ + |x + z|γ − |x|γ − |x + y + z|γ . This new integral is finite if and only if the starting integral is finite (because we can obtain a multiple of this integral by reducing the integration domain instead of extending it). Now we make the change of variables x = ru, y = r cosφ, z = r sin φ and reduce the integration domain by an additional constraint y2 + z2 < 1 (we get almost the same integral by extending this domain, so this procedure is also two-way):∫ +∞ −∞ ∫ 2π 0 ∣∣∣∣D2H(u, cosφ, sin φ) 2| cosφ sin φ|H ∣∣∣∣ 2n0 ∣∣∣∣1 − D2H(u, cosφ, sin φ) 2| cosφ sin φ|H ∣∣∣∣ −α− d 2 · · | cosφ|−dH | sin φ|−dH ∫ min(1/|u|,1) 0 r2−2dHdrdφdu. Integrating with respect to r, we obtain the necessary condition for finiteness, dH < 3 2 . To proceed, we need some inequalities describing the behavior of D2H(x, y, z). Lemma 2. 1. For all γ ∈ (0, 1) ∪ (1, 2], there exist the positive constants C1 and C2 such that if 3|y| < |x|, 3|z| < |x|, then C1|yz||x|γ−2 ≤ |Dγ(x, y, z)| ≤ C2|yz||x|γ−2. If γ = 1, then, for the same x, y, z, we have |Dγ(x, y, z)| = 0, and so we can take C1 = C2 = 0. 2. If γ ∈ (1, 2), then |Dγ(x, y, z)| ≤ 21−γ(||y| + |z||γ − ||y| − |z||γ). Additionally, if y, z > 0, then the inequality changes into an equality if and only if x = −(y + z)/2. 3. If γ ∈ (0, 1), then |Dγ(x, y, z)| ≤ |y|γ + |z|γ − ||y| − |z||γ . Additionally, if y, z > 0, then the inequality changes into an equality if and only if x = −y,−z. 4. If γ = 1, then |Dγ(x, y, z)| ≤ |y| + |z| − ||y| − |z|| . Additionally, if y, z > 0, then the inequality changes into an equality if and only if x ∈ [−max(y, z),−min(y, z)]. Proof. The first inequality is equivalent to C1|uv| ≤ |Dγ(1, u, v)| ≤ C2|uv|; |u| < 1 3 , |v| < 1 3 . Because u, v is small enough by our assumption, it is possible to use the representation Dγ(x, y, z) = −γ(γ − 1) ∫ |y| 0 ∫ |z| 0 (x + sign(y)v + sign(z)w)γ−2dvdw which is valid for all x > max(0,−y,−z,−y − z) (here we denote sign(y) = 1 if y ≥ 0 and sign(y) = −1 if y < 0). We have |Dγ(1, u, v)| = ∣∣∣∣∣γ(γ − 1) ∫ |u| 0 ∫ |v| 0 (1 + sign(u)y + sign(v)w)γ−2dydw ∣∣∣∣∣ ≤ ≤ |γ(γ − 1)uv| sup |y|<1/3;|w|<1/3 (1 + y + w)γ−2 ≤ 32−γ |γ(γ − 1)| |uv| 76 ALEXEY V. RUDENKO and similarly |Dγ(1, u, v)| ≥ ( 5 3 )2−γ |γ(γ − 1)| |uv|. Note that, for γ = 1, we have |Dγ(1, u, v)| = 0 under the same assumptions on u, v. To prove other inequalities, it is enough to consider the case 0 < y ≤ z. Indeed, if we have both y and z negative, then we can introduce new variables x̃ = −x, ỹ = −y, z̃ = −z and obtain the same inequality for positive variables. If, for example, y < 0 and z > 0, then another change of variables helps: x̃ = x + y, ỹ = −y, z̃ = z. All other cases are similar because of the symmetry with respect to the exchange y ↔ z. Moreover, if y = 0, then all inequalities are trivial equalities, so we may suppose that y �= 0. Now we can prove these inequalities by finding the maximum and minimum of Dγ(x, y, z) for fixed y and z. Using the convexity of a power function, we can find the sign of the derivative of g(x) = Dγ(x, y, z) with respect to x. For γ ∈ (1, 2), we have g′(x) < 0, x ∈ (−∞,−(y + z)/2), g′(x) = 0, x = −(y + z)/2, g′(x) > 0, x ∈ (−(y + z)/2, +∞). For γ ∈ (0, 1) (note that, in this case, the derivative may not exist at some points): g′(x) > 0, x ∈ (−∞,−y − z), g′(x) < 0, x ∈ (−y − z,−z), g′(x) > 0, x ∈ (−z,−(y + z)/2), g′(x) = 0, x = −(y + z)/2, g′(x) < 0, x ∈ (−(y + z)/2,−y), g′(x) > 0, x ∈ (−y, 0), g′(x) < 0, x ∈ (0, +∞). For γ = 1 (here the derivative also does not exist at some points): g′(x) = 0, x < −y − z, g′(x) < 0, x ∈ (−y − z,−z), g′(x) = 0, x ∈ (−z,−y), g′(x) > 0, x ∈ (−y, 0), g′(x) = 0, x > 0. We also know that Dγ(x, y, z) is continuous and that DH(x, y, z) → 0, |x| → ∞ (from the first inequality). As we can see for γ ∈ [1, 2), the function has its minimum in x = −y−z 2 and always negative. The right side of the inequality is exactly |Dγ(−y−z 2 , y, z)|, so it is proved. For γ ∈ (0, 1] by same method, we find the maximum at x = 0; x = −y − z and the minimum at x = −y; x = −z (values at both points of each pair are the same). Comparing the moduli of the function values at these points, we get |Dγ(x, y, z)| ≤ |Dγ(−y, y, z)|. If y, z > 0, then the maximum is achieved only at the points x = −y−z 2 for γ ∈ (1, 2) and at x = −y; x = −z for γ ∈ (0, 1). For γ = 1, the maximum is achieved on the interval x ∈ [−z,−y]. Lemma 2 is proved. Now let us look at the integral ∫ +∞ −∞ ∫ 2π 0 ∣∣∣∣D2H(u, cosφ, sin φ) 2| cosφ sin φ|H ∣∣∣∣ 2n0 ∣∣∣∣1 − D2H(u, cosφ, sin φ) 2| cosφ sin φ|H ∣∣∣∣ −α− d 2 · · | cosφ|−dH | sinφ|−dH(max(1, |u|))2dH−3dφdu LOCAL TIME AS AN ELEMENT OF THE SOBOLEV SPACE 77 once more. We split the domain of integration into several parts by restricting ourself to the case φ ∈ (0, π 4 ) (other parts of the domain can be treated similarly): D1 = {|u| > M}, D2 = {|u| < M ; φ � ε, π 4 − φ � ε}, D3 = {|u| < M, |u| � ε, |1 + u| � ε; φ < ε}, D4 = {|u|2 + | sin φ|2 < 4ε2}, D5 = {|1 + u|2 + | sin φ|2 < 4ε2}, D6 = {|u| < M, |u + 1√ 2 | � ε; π 4 − φ < ε}, D7 = {| u cosφ + 1|2 + | tg φ − 1|2 < 16ε2}. Obviously, the union of these parts covers a selected part of the domain if we choose ε < 1 16 and M such that, for |u| > M , we have C2|u|2H−2 < 1 2 , where C2 is the constant from the first inequality above for H �= 1 2 . We use this inequality for D1 and immediately obtain that the finiteness of our integral is equivalent to the finiteness of two integrals∫∞ M u2n0(2H−2)+2dH−3du and ∫ π/4 0 |sinφ|2n0(1−H)−dHdφ. Both integrals are finite if and only if n0 > dH−1 2(1−H) . If H = 1 2 , we can choose M such that D2H(u, cosφ, sinφ) = 0 on D1. In this case, it is easy to see that the condition dH < 1 or n0 > 0 (it is true if the declared conditions for the local time existence hold) is sufficient for the integral over D1 to be finite. For D2, D3, D4, D5, D6, we have that 1 − D2H(u,cosφ,sinφ) 2|cosφsinφ|H is not equal to zero on the closure of these sets. Indeed, using the inequalities from Lemma 2, we get for H ∈ (0, 1 2 ): |D2H(u, cosφ, sin φ)| 2| cosφ sin φ|H � | cosφ|2H + | sin φ|2H − | cosφ − sinφ|2H 2| cosφ sin φ|H = = 1 + (| cosφ|H − | sin φ|H)2 − | cosφ − sin φ|2H 2| cosφ sin φ|H � = 1 − (| sin φ|H + | cosφ − sin φ|H − | cosφ|H)· · (| cos φ|H − | sinφ|H + | cosφ − sin φ|H)) 2| cosφ sin φ|H = = 1 − (1 + | ctg φ − 1|H − | ctg φ|H)· · (1 − | tg φ|H + |1 − tg φ|H) � � 1 − (1 + | ctg(π/4) − 1|H − | ctg(π/4)|H)· · (1 − | tg(π/4)|H + |1 − tg(π/4)|H) = 1. For H ∈ (1 2 , 1), |D2H(u, cosφ, sin φ)| 2| cosφ sin φ|H � 2−2H | cosφ + sin φ|2H − | cosφ − sin φ|2H | cosφ sin φ|H = = 2−2H(|tg2φ + ctg2φ + 2|H − |tg2φ + ctg2φ − 2|H) � � 2−2H(|tg2(π/4) + ctg2(π/4) + 2|H − ∣∣tg2(π/4) + ctg2(π/4) − 2 ∣∣H) = 1. For H = 1 2 , both inequalities are true. As we can see, the equality in these inequalities is possible only if φ = π/4, u = − 1√ 2 (from Lemma 2 and the inequalities above). But the function we consider is continuous on D2, D3, D4, D5, D6 and therefore is not equal to zero. So, this expression taken to the power −α− d/2 can be omitted in the integral. 78 ALEXEY V. RUDENKO For D2, the integrand is bounded, so the integral is always finite. For D3, D4, D5, we have to deal only with the singularity of |sinφ|−dH which can be refined by the renormalization term. For D3, we have the bound |D2H(u, cosφ, sinφ)| ≤ C|φ|, where C is a constant de- pending only on ε, and the condition n0 > dH−1 2(1−H) is sufficient for the integral to be finite. Note that, for D4, D5, this inequality is not true (for 2H < 1), because the derivative of D2H(u, cosφ, sinφ) with respect to φ at φ = 0 blows up if u = 0 or u = −1. Instead, we have to use a change of coordinates. For D3, we can also prove that, for H = 1/2, the con- dition n0 > dH−1 2(1−H) is necessary. It is enough to note that |D2H(u, cosφ, sin φ)| = 2 sin φ on 0 � φ < ε; u ∈ (− cosφ,− sin φ). For D4, let sinφ = rsinθ, u = rcosθ. We obtain the following integral:∫ 2ε 0 ∫ 2π 0 ||rcosθ + rsinθ|2H + |rcosθ + √ 1 − (rsinθ)2|2H − |rcosθ|2H− − |rcosθ + rsinθ + √ 1 − (rsinθ)2|2H |2n0 |sinθ|−dH−2n0Hr1−dH−2n0Hdθdr. Here, we already dropped multipliers under the integral bounded above and below. The expression taken to 2n0 power can be bounded (using derivatives) by the expres- sion Crmin(2H,1)|sinθ|. Thus, our integral is bounded by the product of two integrals∫ 2ε 0 r2n0min(2H,1)+1−dH−2n0Hdr and ∫ 2π 0 |sinθ|2n0−dH−2n0Hdθ. The conditions for their finiteness are 2n0(min(2H, 1)− H) + 2 − dH > 0 and 2n0(1 − H) − dH + 1 > 0. Recall that we have already the conditions dH < 3 2 and n0 > dH−1 2(1−H) as necessary and that n0 ≥ 0. We can see that these conditions are sufficient in this case. The case of D5 is almost similar. We have to let sinφ = rsinθ, u + 1 = rcosθ and obtain a similar bound for the integrand. For D6, the integrand is bounded as we have proved above. For D7, the only un- bounded part under the sign of integral is ∣∣∣1 − D2H (u,cosφ,sinφ) 2|cosφsinφ|H ∣∣∣−α− d 2 and 2|cosφsinφ|H − D2H(u, cosφ, sinφ) = = |cosφ|2H(2|1 + w|H − |1 + v + w|2H − |v − 1|2H + |v|2H + |v + w|2H), where w = tgφ − 1, v = u cosφ . We introduce w and v as new variables of integration and note that the expression above is equivalent to |v|2H + |v + w|2H (multiplied by a constant) when v2 + w2 → 0+. Using the polar coordinates, we obtain two integrals∫ 4ε 0 r1−2H(α+ d 2 )dr and ∫ 2π 0 (|cosθ|2H + |cosθ + sinθ|2H)−α− d 2 dθ. The second integral is always finite and first gives us the necessary and sufficient condition α + d 2 < 1 H . If we set T = [0, 1/3] × [2/3, 1], then we can get the same integral but on a different domain. This domain can be treated exactly like the union of D2, D6, D7 (for example, the singularities near s1 = s2 and t1 = t2 and, consequently, near sinφ = 0 and cosφ = 0 are outside the integration domain), and we have only the condition on α as necessary and sufficient for the integral to be finite. In all previous examples, we had the covariation matrix proportional to the identity one. In the next one, the situation is different. Suppose that {W1(t), W2(t) ∈ R d; t ∈ [0, 1]} are two Brownian motions which are not necessarily independent. Suppose that EW1(s)W2(t) = min(s, t)Q, where Q is some d× d matrix. We consider the intersection local time of these two processes. By {λi, i = 1, . . . , d}, we denote eigenvalues of the self-adjoint matrix Q+QT 2 . Here, we study only the condition from Corollary 1. Example. Let T = [0, 1]2; ν(dt) = dt1dt2; ξ(t) = W1(t1)−W2(t2). The local time exists in D2,α for all α < − d 2 if and only if the number of eigenvalues of the matrix Q+QT 2 which are equal to 1 is smaller than 2 (#{i : γi = 1} < 2) and d < 4. LOCAL TIME AS AN ELEMENT OF THE SOBOLEV SPACE 79 Example. Let T = [12 , 1]2; ν(dt) = dt1dt2; ξ(t) = W1(t1)−W2(t2). The local time exists in D2,α for α < − d 2 if and only if #{i : γi = 1} < 2. Example. Let T = [14 , 1 2 ]× [ 34 , 1]; ν(dt) = dt1dt2; ξ(t) = W1(t1)−W2(t2). The local time always exists in D2,α for α < − d 2 . Proof. The integral we have to study has the form∫ 1 0 ∫ 1 0 d∏ i=1 (t1 + t2 − 2min(t1, t2)λi)−1/2dt1dt2. Here, we used that detK(t, t) = det((t1 + t2)I − min(t1, t2)(Q + QT )) = d∏ i=1 (t1 + t2 − 2min(t1, t2)λi). Note that t1+t2−2min(t1, t2)λi = (t1+t2)(1−λi)+|t1−t2|λi. If λi �= 1 and consequently λi < 1, we have only one singularity near t1 = t2 = 0. Using the polar coordinates, we conclude that the necessary condition for the integral to be finite is 2 − d 2 > 0. If, for some i, we have λi = 1, we have to study the singularity near t1 = t2. Using the polar coordinates, we conclude that multipliers with λi �= 1 can be omitted in the integral with respect to φ, and we get the necessary condition 1 − k 2 > 0, where k = #{i : γi = 1}. Two necessary conditions we obtained are also sufficient, because we do not have other singularities in the integral. For T = [12 , 1]2 we do not have singularity near t1 = t2 = 0 and for T = [14 , 1 2 ] × [34 , 1] we also do not have singularity near t1 = t2 with obvious consequences. Bibliography 1. A.A. Dorogovtsev, V.V. Bakun, Random mappings and a generalized additive functional of a Wiener process, Theory of Stoch. Proc. 48 (2003), no. 1. 2. A. Rudenko, Existence of generalized local times for Gaussian random fields, Theory of Stoch. Proc. 12(28) (2006), no. 1-2, 142–154. 3. H.H. Kuo, Gaussian Measures in Banach Spaces, Springer, Berlin, 1975. 4. H. Watanabe, The local time of self-intersections of brownian motions as generalized brownian functionals, Lett. in Math. Phys. 23 (1991), 1–9. 5. P. Imkeller, V. Perez-Abreu, J. Vives, Chaos expansions of double intersection local time of brownian motion in d and renormalization, Stoch. Proc. and Appl. 56 (1995), 1–34. 6. P. Malliavin, Stochastic Analysis, Springer, Berlin, 1997. 7. J. Rosen, A renormalized local time for multiple intersections of planar brownian motion, Seminaire de Probabilities XX 20 (1986), 515–531. 8. S. Albeverio, Y. Hu, X.Y. Zhou, A remark on non-smoothness of the self-intersection local time of planar brownian motion, Stat. and Prob. Lett. 32 (1997), 57–65. 9. S. Orey, Gaussian sample functions and the Hausdorff dimension of level crossings, Wahr- scheinlichkeitstheorie verw. Geb. 15 (1970), 249–256. 10. S. Watanabe, Lectures on Stochastic Differential Equations and Malliavin Calculus, Springer, Berlin, 1984. 11. Y. Hu, D. Nualart, Renormalized self-intersection local time for fractional brownian motion, Annals of Prob. 33 (2005), no. 3, 948–983. 12. Y. Hu, D. Nualart, Regularity of renormalized self-intersection local time for fractional Brown- ian motion, J. of Commun. in Information and Systems (CIS) 7 (2007), no. 1, 21–30. E-mail : arooden@yandex.ru
id nasplib_isofts_kiev_ua-123456789-4508
institution Digital Library of Periodicals of National Academy of Sciences of Ukraine
issn 0321-3900
language English
last_indexed 2025-12-07T15:45:37Z
publishDate 2007
publisher Інститут математики НАН України
record_format dspace
spelling Rudenko, A.V.
2009-11-19T14:00:08Z
2009-11-19T14:00:08Z
2007
Local time as an element of the Sobolev space / A.V. Rudenko // Theory of Stochastic Processes. — 2007. — Т. 13 (29), № 3. — С. 65–79. — Бібліогр.:12 назв.— англ.
0321-3900
https://nasplib.isofts.kiev.ua/handle/123456789/4508
519.21
For a centered Gaussian random ?eld taking its values in R^d, we investigate the existence of a local time as a generalized functional, i.e an element of some Sobolev space. We give the sfficient condition for such an existence in terms of the field covariation and apply it in several examples: the self-intersection local time for a fractional Brownian motion and the intersection local time for two Brownian motions.
Research was partially supported by the Ministry of Education and Science of Ukraine, project GP/F13/0095.
en
Інститут математики НАН України
Local time as an element of the Sobolev space
Article
published earlier
spellingShingle Local time as an element of the Sobolev space
Rudenko, A.V.
title Local time as an element of the Sobolev space
title_full Local time as an element of the Sobolev space
title_fullStr Local time as an element of the Sobolev space
title_full_unstemmed Local time as an element of the Sobolev space
title_short Local time as an element of the Sobolev space
title_sort local time as an element of the sobolev space
url https://nasplib.isofts.kiev.ua/handle/123456789/4508
work_keys_str_mv AT rudenkoav localtimeasanelementofthesobolevspace