From intersection local time to the Rosenblatt process
Tomasz Bojdecki
0
Luis G. Gorostiza
0
Anna Talarczyk
0
0
L. G. Gorostiza Centro de Investigacion y de Estudios Avanzados
, Mexico City,
Mexico
The Rosenblatt process was obtained by Taqqu (Z. Wahr. Verw. Geb. 31:287302, 1975) from convergence in distribution of partial sums of strongly dependent random variables. In this paper, we give a particle picture approach to the Rosenblatt process with the help of intersection local time and white noise analysis, and discuss measuring its longrange dependence by means of a number called dependence exponent. The Rosenblatt process was introduced and studied by Taqqu [33], motivated by a counterexample of Rosenblatt regarding a strong mixing condition [26]. See Taqqu [36] for the history and an overview of the process, and related work. The Rosenblatt

process arises from convergence in distribution of normalized partial sums of strongly
dependent random variables. It lives in the second Wiener chaos, in contrast to
Gaussian processes (which belong to the first chaos) that are obtained from sums
of independent or weakly dependent random variables. The Rosenblatt process
possesses some of the main properties of the (Gaussian) fractional Brownian motion with
Hurst parameter greater than 1/2: mean zero, Hlder continuity, nondifferentiability,
selfsimilarity and stationarity of increments (hence, it has the same form of
covariance as fractional Brownian motion), infinite divisibility, longrange dependence in
the sense that the sum over k of covariances of increments on the intervals [0, 1]
and [k, k + 1] diverges, and it is not a semimartingale. The Rosenblatt process is
the simplest nonGaussian Hermite process (see Taqqu [34]). It is a counterpart of
the fractional Brownian motion, which is the most prominent longrange dependent
Gaussian process. Significant attention has been attracted by the Rosenblatt process
due to its mathematical interest, and possible applications where the Gaussian
property may not be assumed. Recent papers on the subject include [39], which develops
a related stochastic calculus and mentions areas of application (see also references
therein), [23] and [40], where new properties of the process have been found, and
[17], which provides a strong approximation for the process. Relevant information for
the present paper on the Rosenblatt process and fractional Brownian motion is given
in the next section.
Our main objective in this paper was to show a different way of obtaining the
Rosenblatt process. The method of Taqqu, which was developed for Hermite processes
generally [34], is based on limits of sums of strongly dependent random variables. The
Rosenblatt process can also be defined as a double stochastic integral [36,39]. Our
approach consists in deriving the Rosenblatt process from a specific random particle
system, which hopefully provides an intuitive physical interpretation of this process.
A useful tool here is the theory of random variables in the space of tempered
distributions S S (Rd ) (d = 1 in our case), since it permits to employ some nice
properties of this space, and the Rosenblatt process can be expressed with the help
of an S random variable. This was noted by Dobrushin [13]. Relations between
random particle systems and random elements of S have been studied by many authors,
beginning with MartinLf [24]. Our approach is in the spirit of Adler and coauthors,
e.g., [1], where a general scheme (but still not so general to cover our case) was
developed for representing a given random element of S as the limit of appropriate
functionals of some particle system (with high density in [1]). That approach was later
applied in [2,11,32] to give particle picture interpretations of the selfintersection
local times of density processes in S . We stress that our principal aim is to construct
the Rosenblatt process by means of a particle system, and to this end, an
important step is to study a suitable random element of S . Particle picture approaches
have been used to obtain fractional Brownian motion and subfractional Brownian
motion with Hurst parameter H , in different ways for H < 1/2 and H > 1/2
[8,12].
Our results can be summarized as follows.
The Rosenblatt process with parameter H , defined for H (1/2, 1), is represented
in the form t = Y, 11[0,t] , t 0, where Y is an S random variable which is obtained
from the Wick product : X X :, where X is a centered Gaussian S random variable
(see (3.1)). X is in a sense a distributional derivative of a suitable fractional Brownian
motion, and the relation between Y and X corresponds to the informal formula (43)
in [36]. Note that : X X : is a counterpart of the Hermite polynomial of order
2, H2(x ) = x 2 1, used in [33,36]. The possibility to take test functions of the
form = 11[0,t] to bring in a time parameter has been noted for example in [34]. This
formulation is in the spirit of white noise analysis.
We define a particle system on R with initial distribution given by a Poisson random
field with Lebesgue intensity measure, and particles evolving independently
according to the standard symmetric stable Lvy process. The particles are independently
assigned charges +1 and 1 with probabilities 1/2. A crucial element of the
construction is the intersection local time of two independent stable processes, which is
known to exist for > 1/2 (Proposition 5.1 of [11]). Consider the process T defined
by
j k (x j + j , xk + k ; T ), 11[0,t] , t 0,
where the x j are the points of the initial Poisson configuration, the j are the
stable processes corresponding to those points ( j (0) = 0), the j are the respective
charges, and (x j + j , xk + k ; T ) is the intersection local time of the processes
x j + j and xk + k on the interval [0, T ]. This local time is defined as a process in
S (Definition 2.2 and Proposition 2.3). T is given a precise meaning and shown to
have a continuous modification (Lemma 3.4). The result is that for (1/2, 1), T
converges as T , in the sense of weak functional convergence, to the Rosenblatt
process with parameter H = , up to a multiplicative constant (Theorem 3.5).
Our second result concerns the analysis of longrange dependence of the Rosenblatt
process. We give a precise measure of longrange dependence by means of a number
called dependence exponent (Theorem 4.1). We show that the increments are
asymptotically independent (not just uncorrelated) as the distance between the intervals tends
to infinity (Corollary 4.2).
In Sect. 2, we give background on the Rosenblatt process, on the S random
variable X from which fractional Brownian motion is derived, on how the
abovementioned particle system produces X , and on intersection local time. In Sect. 3, we
construct the Rosenblatt process by means of the S random variable Y obtained from
: X X :, we show that the process T is well defined, and we prove convergence
to the Rosenblatt process. In Sect. 4, we discuss the longrange dependence of the
process. Section 5 contains some comments of related interest. The proofs are given in
Sect. 6.
We use the following notation:
S: the Schwartz space of smooth rapidly decreasing functions on R,
S : the space of tempered distributions (dual of S),
: complex conjugate,
(z) = R ei xz (x )dx : Fourier transform of a function ,
: convergence in law in an appropriate space,
f : convergence of finitedimensional distributions,
: convolution,
C ([0, ]): the space of real continuous functions on [0, ],
C, Ci : generic positive constants with possible dependencies in parentheses.
2 Background
2.1 The Rosenblatt process
We recall some facts on the Rosenblatt process = (t )t0 with parameter
H (1/2, 1), which are found in [33,36]. The characteristic function of the
finitedimensional distributions of the process in a small neighborhood of 0 has the form
E exp i
j=1
= exp
k=2
j=1
x1 x2H1x2 x3H1 . . . xk1 xk H1xk x1H1d x1 . . . d xk ,
p
H (2H 1)
The value of is chosen so that E 12 = 1. The series in the exponent converges for
1, . . . , p in a (small) neighborhood of 0 depending on t1, . . . , t p, and (2.1) defined
in this neighborhood determines the distribution. The process is also characterized by
the cumulants of the random variable pj=1 j t j , which are 1 = 0,
k = 2k1(k 1)! k RH,k (1, . . . , p; t1, . . . , t p), k 2.
The process arises from a (Donskertype) limit in distribution as n of the
processes
j=1
X j , t [0, T ],
r j = E Y0Y j = (1 + j 2)(H1)/2 j H1 as j .
The spectral representation of the process is
( =d means equality in distribution), where
[H (2H 1)/2]1/2
A(H ) = 2 ( 1 H ) sin(H /2) ,
B is a complex Gaussian measure on R such that B = B(1) + i B(2), B(1)( A) =
B(1)( A), B(2)( A) = B(2)( A), A is a Borel set of R with finite Lebesgue measure
 A, B(1) and B(2) are independent, and E (B(1)( A))2 = E (B(2)( A))2 = 21  A. (B can
be viewed as a complexvalued Fourier transform of white noise). The double prime
on the integral means that the diagonals 1 = 2 are excluded in the integration.
The process also has a time representation as a double integral on R2 with respect to
Brownian motion, and a finite interval integral representation obtained in [39].
We have mentioned in the Introduction some of the main properties of the Rosenblatt
process. We recall the selfsimilarity with parameter H : for any c > 0,
(s2H + t 2H t s2H ).
k=1
This and stationarity of increments imply
and hence, the covariance function of has the same form as that of the fractional
Brownian motion, i.e.,
In particular, the increments are positively correlated (since H > 1/2), and
We have not found a published proof of the nonsemimartingale property of , but
that is easy to show. By (2.9) with H > 1/2, it is obvious that the quadratic variation is
0. A deeper result is that selfsimilarity and increment stationarity imply that the paths
have infinite variation [41]. The nonsemimartingale property of fractional Brownian
motion (H = 1/2) follows for example from a general criterion for Gaussian processes
[7] (Lemma 2.1, Corollary 2.1).
Infinite divisibility was recently proved in [40,23].
There does not seem to be information in the literature on whether or not the
Rosenblatt process has the Markov property (it seems plausible that it does not, by
analogy with fractional Brownian motion).
The fractional Brownian motion is the only Gaussian process that has the properties
(2.8) and (2.9) (with = 1). On the other hand, there are many processes with
stationary increments satisfying (2.8) which belong to the second chaos [22]. The
Rosenblatt process is the simplest one of them.
2.2 An S random variable related to fractional Brownian motion
Recall that fractional Brownian motion (fBm) with Hurst parameter H (0, 1) is a
centered Gaussian process (BtH )t0 with covariance given by the righthand side of
(2.10) with = 1 (see [29] for background on fBm). This process can be represented
with the help of the centered Gaussian S valued random variable X with covariance
functional
E X, 1 X, 2 =
1(x )2(x )x d x , 1, 2 S,
X, 11[0,t] t0 = (K BtH )t0,
with H = 1+2 and K is some positive constant. ( X, 11[0,t] is defined by an
L2extension). For 0 < < 1 ( 21 < H < 1), this random variable X can be obtained
from the particle system described in the Introduction, i.e., the system of independent
standard stable processes (particle motions) starting from a Poisson random field
with Lebesgue intensity. Each particle has a charge 1 with equal probabilities, and the
charges are mutually independent and independent of the initial configuration and of
the particle motions. The motions have the form x j + j , where the x j s are the points
of the initial configuration, and the j are independent standard stable processes
j
independent of {x j } j , 0 = 0. The charges are denoted by j .
The normalized total charge occupation on the interval [0, T ] is defined by
We have the following proposition.
Proposition 2.1 If 0 < < 1 and T , then
(a) X T X in S , where X is as in (2.12).
(b) ( X T , 11[0,t] )t0 f (K BtH )t0 with H = 1+2 .
This fact is an easy consequence of Theorem 2.1(a) in [9], where the occupation
time fluctuations around the mean for the system without charges were considered. It
suffices to take two independent copies of such systems and to write the difference of
their occupation time fluctuations.
A similar procedure with a different functional of a particle system without charges
permits also to obtain fBm with H < 21 , as well as the corresponding random variable
X (see Theorem 2.9 in [12]).
2.3 Intersection local time
There are several ways to define the intersection local time (ILT) of two processes
(see, e.g., [1,14,25]. We will take the definition from [11], which is close to that of
[1]. Intuitively, ILT of real cadlag processes (t1)t0, (t2)t0 up to time T is given by
(v2 u1)(u1)du dv, S,
where is the Dirac distribution. We want to regard as a process in S .
To make this definition rigorous, one has to apply a limiting procedure.
Let F denote the class of nonnegative symmetric infinitely differentiable functions
f on R with compact support and such that R f (x )dx = 1. For f F , > 0, let
, x R.
We will frequently use
 f(x ) 1 and lim0 f(x ) = lim0 f (x ) = 1.
f(v2 u1)(u1)du dv, T 0, S.
Definition 2.2 If there exists an S process (1, 2) = ((1, 2; T ))T 0 such that
for each T 0, S and any f F , (1, 2; T ), is the mean square limit
of f (1, 2; T ), as 0, then the process (1, 2) is called the intersection
local time (ILT) of the processes 1, 2.
In [11], the following result was proved (Theorem 4.2 and Proposition 5.1 therein).
Proposition 2.3 Let 1, 2 be independent standard stable processes in R. If > 21 ,
then for any x , y R the ILT (x +1, y+2) exists. Moreover, for all T 0, f F ,
S, f ( + 1, + 2; T ), converges in L2(R2 , P), where is the
Lebesgue measure on R, and P is the probability measure on the underlying sample
space .
3 Particle picture for the Rosenblatt process
We begin with another representation of the Rosenblatt process, which is more suitable
for our purpose. From [13], it can be deduced that this construction was known to
Dobrushin, but we have not been able to find it written explicitly in the literature.
Therefore, we will describe it in detail, but the proof will be only sketched.
Let X be the centered Gaussian S random variable with covariance (2.12). Recall
that the Wick product : X X : is defined as a random variable in S (R2) such that
: X X :, 1 2 =
X, 1 X, 2 E X, 1 X, 2 , 1, 2 S.
(see, e.g., Chapter 6 of [18] or [3,5]). The Wick square of X is an S random variable
Y that can be written informally as Y, = : X X :, (x )yx . To make this
rigorous, we use approximation. Fix f F and let f be as in (2.15). For S we
denote
Yf , = : X X :, f, , S.
The following lemma is an easy consequence of the regularization theorem [20], and
the fact that, by Gaussianity,
(x , y)( (x , y) + (y, x ))x yd x d y,
Lemma 3.1 If 21 < < 1, then there exists an S random variable Y such that for
any f F ,
Y, = L2 lim0 Yf , , S.
The next theorem is an analogue of (2.13) for the Rosenblatt process.
Theorem 3.2 Let Y be as in Lemma 3.1. Then the real process
( Y, 11[0,t] )t0
We remark that this theorem gives a rigorous meaning to the informal expression
(43) in [36] relating the Rosenblatt process with parameter H and a fBm with
parameter H1 = H2+1 ( 43 , 1), which is given by
X corresponds to (B H1 ) and : X X : corresponds to ((B H1 ) )2. The relationship
between the parameters follows from Proposition 2.1(b) and Theorem 3.2.
In [13], S random variables such as Y are represented in terms of complex multiple
stochastic integrals related to (2.6).
After the first version of this paper had been submitted, the referee drew our attention
to preprint [4] which appeared in the meantime. That paper uses HidaKuo type
calculus [21] to construct the stochastic integral with respect to the Rosenblatt process,
but the representation (3.5) does not seem to be present there.
Representation (3.5) and Proposition 2.1 suggest a way to construct the Rosenblatt
process by means of a particle system. We consider the particle system as before with
21 < < 1. By Proposition 2.3, for each pair j , k , j = k, the intersection local time
(x j + j , xk + k ; T ) exists; moreover, it extends in a natural way to test function
11[0,t]. Namely, we have the following lemma.
Lemma 3.3 Let
a j 11I j , a j R, I j is a bounded interval,
For the convenience of the reader, let us recall (1.1),
j k (x j + j , xk + k ; T ), 11[0,t] , t 0.
j=1
Lemma 3.4 The process T is well defined (the series converges in L2), and it has a
continuous modification.
The main result of the paper is stated in the next theorem, which is a counterpart
of Proposition 2.1.
Theorem 3.5 Let 21 < < 1. Then T K in C ([0, ]) as T , > 0, where
is the Rosenblatt process with H = and K is a positive constant.
4 Dependence exponent
Longrange dependence is a general notion that has not been clearly defined and can be
viewed in different ways [19,28,35]. For a Gaussian process , longrange dependence
is usually described as slow (power) decay of the covariance of increments on intervals
[u, v], [s + , t + ] as , i.e.,
Cov(v u , t+ s+ ) Cu,v,s,t K ,
where K is a positive constant, and convergence and divergence of the series
k=1
are sometimes referred to as shortrange dependence and longrange dependence,
respectively. This criterion is also applied to nonGaussian processes with finite
second moments, such as the Rosenblatt process [36]. The underlying idea is that the
increments become uncorrelated (but not necessarily independent) at some rate as the
distance between the intervals tends to infinity. However, it can happen that K 0,
which should also be regarded as longrange dependence ([16] contains examples).
In order to characterize longrange dependence in some more precise way for
infinitely divisible processes (not necessarily Gaussian), the codifference (see [27,29]) can
be useful. In [10], we defined the dependence exponent of a (real) infinitely divisible
process as the number
sup{ > 0 : D(z1, z2; u, v, s, t ) = o( ) as },
D(z1, z2; u, v, s, t ) =  log E ei(z1(vu)+z2(t+ s+ )
log E ei z1(vu) log E ei z2(t+ s+ )
is the absolute value of the codifference of the random variables z1(v u ) and
z2(t+ s+ ). Note that if is Gaussian, then
D(z1, z2; u, v, s, t ) = z1z2Cov(v u , t+ s+ ).
For fractional Brownian motion, = K = 2 2H , and for subfractional Brownian
motion, = K = 3 2H [8].
It turns out that the same idea can be used to measure longrange dependence for the
Rosenblatt process, and this can be done without recourse to infinite divisibility. As
recalled in Sect. 2.1, the characteristic functions of the finitedimensional distributions
of the process are given by an explicit formula only for small values of the parameters,
which are z1 and z2 in our case (see (4.2)). We show next that it is enough to take z1
and z2 in an appropriate neighborhood of 0 to measure longrange dependence and
prove asymptotic independence of increments.
For simplicity we take u = s, v = t .
Theorem 4.1 Let be the Rosenblatt process with parameter H . For any 0 s < t
there exists a neighborhood U (s, t ) of 0 in R2 such that
D (z1, z2, s, t ) := D (z1, z2, s, t, s, t )
sup{ > 0 : D (z1, z2, s, t ) = o( ) as }, (4.4)
So we see that dependence exponent of the Rosenblatt process with parameter H
is the same as that for fBm B H .
From this theorem, by a standard tightness argument, stationarity of increments
of , and the fact that the law of t is determined by its characteristic function in an
arbitrarily small neighborhood of 0, we obtain the following corollary.
Corollary 4.2 For any 0 < s < t , the increments of the Rosenblatt process t s
and t+ s+ are asymptotically independent as , i.e., if (s,t) is the law of
t s and (s,t),(s+,t+ ) is the law of (t s , t+ s+ ), then
(s,t),(s+,t+ ) (s,t) (s,t) = (s,t) (s+,t+ ) .
5 Additional comments
5.1 SubRosenblatt process
It is known that if in the formula (2.13) we put 11[0,t] 11[t,0] instead of 11[0,t], we
obtain a subfractional Brownian motion (subfBm), i.e., a centered Gaussian process
with covariance
again with H = 1+2 . This process has been studied by several authors, e.g., [8,15,
38,42] and others. In particular, in [12], an analogue of Proposition 2.1(b) was proved
for subfBm.
We can now extend formula (3.5) and define a new process ( Y, 11[0,t] 11[t,0] )t0.
It is natural to call it subRosenblatt process, as it has the same covariance as subfBm.
Analogues of Theorems 3.5 and 4.1 also hold.
5.2 Rosenblatt process with two parameters
Maejima and Tudor in [22] define a class of selfsimilar processes with stationary
increments that live in the second Wiener chaos. These processes depend on two
parameters H1, H2, and the Rosenblatt process corresponds to the case H1 = H2.
One can ask about the possibility of extending our construction to those twoparameter
processes.
5.3 General Hermite processes
Taqqu [34] studies extensions of the Rosenblatt process living in Wiener chaos of
order k, k 2, which he calls Hermite processes. One can attempt to find a particle
picture interpretation for those processes. It seems that one should employ kth Wick
powers and work with intersection local times of ktuples of stable processes.
6 Proofs
Proof of Theorem 3.2 (outline). Let be the Rosenblatt process with parameter H .
It is known that its distributions are determined by its moments, and therefore, it is
enough to prove that for all n, p N, t1, t2, . . . , t p 0,
E Y, n = C n E
j=1
j=1
where s are the corresponding cumulants given by (2.5) and P(n) is the set of all
partitions of {1, . . . , n}, and # denotes cardinality of a set.
To compute E Y, n, we will need the formulas for moments of the Wick product
: X X :. These moments are expressed with the help of Feynman graphs (see, e.g.,
[3, p. 1,085] or [31, p. 422]). For fixed n, we consider graphs as follows. Suppose
that we have n numbered vertices. Each vertex has two legs numbered 1 and 2. Legs
are paired, forming links between vertices, in such a way that each link connects two
different vertices, and there are no unpaired legs left. The graph is a set of links.
Each link is described by an unordered pair {(i, j ), (l, m)}, i, l {1, . . . n}, j, m
{1, 2}, which means that leg j , growing from vertex i is paired with leg m growing
from vertex l. i = l since each link connects different vertices, and any (i, j ), i =
1, . . . , n, j = 1, 2, is a part of one and only one link. The set of all distinct graphs of
the above form will be denoted by Gn2. Let Gn2 denote the set of all connected graphs
2
in Gn .
By formulas (2.7) and (6.10) in [31], we have
I G (),
GGn2
x p,q (dxl,m ) x p,q H dx p,q.
(6.4)
Using Lemma 3.1 and similar arguments as in the proof of Lemma 3.3 below, it is not
difficult to see that
E Y, n
I G (f, ),
For G Gn2 we have
FG (x1, . . . , xn ) x1H . . . xnH dx1 . . . dxn,
where FG is a product of functions of the form f(xi ) or f(xi ). By (2.16) and the
dominated convergence theorem,
I G () =
In the proof of the integrability of the function under the integral in (6.6), we use
 (x1 x2) (xn x1)  (x1 x2)2 +  (xn x1)2,
j=1
2#B j 1(# B j 1)!
Note that each G Gn2 determines a partition G P(n), whose elements are the
sets of vertices of connected components of G. Hence, by (6.5) and (6.8),
E Y, n =
It is not difficult to see that if P(n) is of the form = {B1, . . . , Bk }, B j 2,
j = 1, . . . , k, then the number of different G Gn2 such that G = is equal to
Therefore, setting J1 = 0, we obtain
E Y, n =
2#B1(# B 1)! J#B .
Jk given by (6.6) can be also written as
Jk = C k
(x1) . . . (xk ) x1 x2H1 x2 x3H1 . . . xk x1H1 dx1 . . . dxk ,
hence combining (6.9) with (2.2), (2.5), and (6.2), we obtain (6.1).
ei x(z+z )eiy(w+w ) (z + w) (z + w ) f (w)g(w )
where s,u is the law of (s1, u1). To complete the proof of part (a), it suffices to show
that
I :=
To derive this, we cannot repeat the argument of [11] because / L1 for of the
form (3.6), we only have (6.7).
Fix > 0 such that 21 + 4 < . It is easy to see that
I C2(T )
 (z + w)  (z + w )
(1 + z )(1 + w ) (1 + z  )(1 + w  ) h2 (z, z )dzdz dwdw ,
and (6.13) permits to replace the denominator by (1 + z + w )(1 + z + w  ), hence
(6.11) follows by (6.7).
Note that we have also shown that
ei x(z+z )eiy(w+w )
To prove part (b), we observe that the argument above can be carried out for linear
combinations of functions of the form (3.6) and from S instead of . Hence, we see
that (6.14) holds for f, since
( f)(x ) = (x )(1 f (x ))
Then (b) follows from (6.11).
The proof of part (c) is the same as that of Proposition 4.4 in [11]. Only the fact
that L2 is needed here.
Remark 6.1 From the proof of part (c), it follows that
E ( f (x + 1, y + 2; T), (x + 1, y + 2; T), )2dxdy
(x + y)2 f (y) 12esuxervydxdydsdrdudv,
(x + y)2esuxervydxdydsdrdudv.
We need the following lemma which can be proved by repeating the argument of
the proof of Lemma 4.1 in [11]
E jkF(xj + j, xk + k) = E(F2(x + 1, y + 2)
j,k R2
j/k
Proof of lemma 3.4 From Lemma 3.3, it follows that (xj +j, xk +k; T),11[0,t]
are well defined, and (+j,+k; T),11[0,t] belongs to L2(R2, P),
hence by Lemma 6.2 the process T is well defined and the series in (1.1) converges in
L2(). Moreover, using the fact that (x +j, y +k, T) = (y +k, x +j; T)
(see Corollary 3.4 in [11]), we have
1
= 2 2T 2
= 2 E (x +1, y+2; T ), 2dxdy,
2
11(t1,t2](x + y) esux ervy dxdydsdrdudv.
ei(t2t1)(x+y)
Before we pass to the proof of Theorem 3.5, we observe that for any S random
variable Z (not necessarily Gaussian) such that S E Z , 2 is continuous, the
Wick product : Z Z : is well defined by an extension of (3.1). Moreover, we have
the following lemma.
form some continuous Hilbertian seminorm p on S. Assume that ZT Z and
E ZT , 2 E Z , 2, S, as T . Then : ZT ZT :, : Z Z : are
well defined and : ZT ZT :: Z Z : in S (R2) as T .
This lemma follows by a standard argument using properties of S [37], so we skip
the proof.
Lemma 6.3 together with Proposition 2.1(a) implies the following corollary.
Corollary 6.4 Let X T , X be as in Proposition 2.1. Then
: X T X T :: X X : in S (R2) as T .
Indeed, it suffices to observe that by (2.14) and the Poisson initial condition, we
have
E X T , 2 = T
so the assumptions of Lemma 6.3 are satisfied (we use (6.20)).
Proof of Theorem 3.5 For of the form (3.6), we denote by T the random variable
defined by (1.1) with 11[0,t] replaced by .
To prove the theorem it suffices to show that
for any of the form (3.6). Indeed, from (6.23) and Theorem 3.2, we infer convergence
of finitedimensional distributions, and from (6.21), we obtain tightness in C ([0, ])
for each > 0 (see [6], Thm. 12.3; note that the constant C1 in (6.21) does not depend
on T ).
Fix any f F and denote = f, > 0, where f is given by (2.15). Let
j k f (x j + j , xk + k ; T ), , > 0, S,
which is well defined by Lemma 6.2.
Using the estimate E ei1 E ei2  21 E 1 22, valid for centered random
variables 1, 2, it is easy to see that (6.23) will be proved if we show
lim sup sup E  T,, f 2
T  = 0,
0 T 1 0<1
lim sup E  : X T X T :, f, T,, f  = 0, > 0,
T 0<1
Using (6.18), (6.17), (6.15) and then (6.20), we have
 (x + y)21 f ((x + y))2
esux ervy dxdydsdudrdv
8
 (x + y)21 f ((x + y))2x ydxdy.
Hence, (6.25) follows by (2.16) and since 21 < < 1.
Next, we apply Lemma 6.2 to
(x + 1, y + 2; T ) f (x + 1, y + 2; T ), ,
and by (6.18) we obtain
E ( ( f )(x + 1, y + 2; T ), )2dxdy
 (x + y)2 f (y) 12esux ervy dxdydsdrdudv,
1
= T 2
A(T , , ):=E  : X T X T :, , T,, f 2
f
f f
E , (x + s , x + u ), (x + r , x + v)dxdvdrduds.
It is easy to see that since the support of is contained in a compact set which is
independent of , we have (see (3.2) and (3.6))
Let (Tt )t denote the stable semigroup and G its potential, G = 0 Tt dt .
From (6.32), (6.33), and the Markov property, we get
R2
1
+ 2
2
2
(x + y)2 f (y) 12x ydxdy
(x + y)2( f (y) 1)( f (x ) 1)x ydxdy
 (x + y)2 f (y) 12x ydxdy.
Hence, (6.29) follows by (6.7) and (2.16).
Finally, from (3.3) and Lemma 3.1, it is easy to see that
2
E  Y, Y, 2 = 2
 (x + y)2 f ((x + y)) 12x ydxdy,
E ei[z1(t s )+z2(t+ s )] = exp
k=2
where Rk (z1, z2, s, t, T ) is the righthand side of (2.2) with
and formula (6.34) holds for z1, z2 R and > 0 such that the series in (6.34)
converges.
To continue with the proof, we need the following lemma and corollary.
Rk (z1, z2, s, t, ) (z1 + z2)k (C (s, t ))k , k = 2, 3, . . . .
An immediate consequence of this lemma is
Corollary 6.6 For all 0 < s < t there exists a neighborhood U (s, t ) of 0 in R2 such
that (6.34) holds for (z1, z2) U (s, t ) and all > 0.
Proof of lemma 6.5 Consider an integral
x1x2H1. . . xk1xk H1xk x1H1dx1 . . . dxk ,
y x + H1 y x H1 for x , y [s, t ], > 0(s, t ).
Hence, similarly as in formula (13) in [36],
The same inequality holds for 0(s, t ).
By (6.35), (2.2),
Hence, (6.38) implies
which proves the lemma.
j=1
We return to the proof of the theorem.
k=2
It is not difficult to see that due to the circular form of the integrand in each Ik
in (6.40), after the substitution x j = x j j (see (6.37)), there appear at least two
factors of the form y x + H 1, y, x [s, t ], and for some (1, . . . , k ), there are
exactly two such factors. Hence, for each < 2 2 H ,
and for some z1, z2
Hence, the theorem follows by Lemma 6.5.
Open Access This article is distributed under the terms of the Creative Commons Attribution License
which permits any use, distribution, and reproduction in any medium, provided the original author(s) and
the source are credited.