Scaling limits for Hawkes processes and application

to nancial statistics
E. Bacry

, S. Delattre

, M. Homann

and J.F. Muzy

Abstract
We prove a law of large numbers and a functional central limit
theorem for multivariate Hawkes processes observed over a time in-
terval [0, T] in the limit T . We further exhibit the asymptotic
behaviour of the covariation of the increments of the components of
a multivariate Hawkes process, when the observations are imposed by
a discrete scheme with mesh over [0, T] up to some further time
shift . The behaviour of this functional depends on the relative size
of and with respect to T and enables to give a full account of
the second-order structure. As an application, we develop our results
in the context of nancial statistics. We introduced in [17] a micro-
scopic stochastic model for the variations of a multivariate nancial
asset, based on Hawkes processes and that is conned to live on a tick
grid. We derive and characterise the exact macroscopic diusion limit
of this model and show in particular its ability to reproduce important
empirical stylised fact such as the Epps eect and the lead-lag eect.
Moreover, our approach enable to track these eects across scales in
rigorous mathematical terms.
Keywords: Point processes. Hawkes processes. Limit theorems. Discretisa-
tion of stochastic processes.
Mathematical Subject Classication: 60F05, 60G55, 62M10.

CMAP CNRS-UMR 7641 and

Ecole Polytechnique, 91128 Palaiseau, France

Universite Paris Diderot and LPMA CNRS-UMR 7599, Bote courrier 7012 75251
Paris, France

ENSAE-CREST and LAMA CNRS-UMR 8050, 3, avenue Pierre Larousse, 92245

Malako, France

SPE CNRS-UMR 6134 and Universite de Corte, 20250 Corte, France

1
1 Introduction
1.1 Motivation and setting
Point processes have long served as a representative model for event-time
based stochastic phenomena that evolve in continuous time. A compre-
hensive mathematical development of the theory of point processes can be
found in the celebrated textbook of Daley and Vere-Jones [8], see also the
references therein. In this context, mutually exciting processes form a spe-
cic but quite important class of point processes that are mathematically
tractable and widely used in practice. They were rst described by Hawkes
in 1971 [12, 13] and according to [8]: Hawkes processes gure widely in
applications of point processes to seismology, neurophysiology, epidemiol-
ogy, and reliability... One reason for their versatility and popularity is that
they combine in one model both a cluster process representation and a sim-
ple conditional intensity representation, which is moreover linear. It comes
closest to fullling for point processes, the kind of role that the autoregres-
sive model plays for conventional time series. Informally, to a multivariate
d-dimensional counting process N = (N
1
, . . . , N
d
) with values in N
d
is asso-
ciated an intensity function (
1
, . . . ,
d
) dened by
P
_
N
i
has a jump in [t, t +dt]

F
t
_
=
i,t
dt, i = 1, . . . , d
where P stands for probability and F
t
is the sigma-eld generated by N up
to present time t. A multivariate Hawkes process has intensity

i,t
=
i
+
_
(0,t)
d

j=1

ij
(t s)dN
j,s
, i = 1, . . . , d (1)
and is specied by
i
R
+
= [0, ) and for i = 1, . . . , d, the
ij
are
functions from R
+
to R
+
. More in Section 2 below for rigorous denitions.
The properties of Hawkes processes are fairly well known: from a prob-
abilistic point a view, the aforementioned book of Daley and Vere-Jones [8]
gives a concise synthesis of earlier results published by Hawkes [1214] that
focus on spectral analysis following Bartlett [4] and cluster representation,
see Hawkes and Oakes [15]. From a statistical perspective, Ogata studied
in [25] the maximum likelihood estimator and showed that parametric ex-
periments generated by the observation of Hawkes processes are regular in
the sense of Fisher information (see for instance [9] for a modern formu-
lation of statistical regularity). Ogata also studied the numerical issue of
exact simulation of Hawkes processes [26], and this topic has known several
2
developments since, see [22] for a review. Recently, the nonparametric esti-
mation of the intensity functions has been investigated by Reynaud-Bouret
and Schbath [27] and Al Dayri et al. [3] in dimension d = 1.
However, in all these papers and the references therein, the focus is on
the microscopic properties of Hawkes processes, i.e. their innitesimal
evolution, possibly under a stationary regime or for a large time horizon
[0, T] in order to guarantee a large number of jumps for statistical inference
purposes. In the present work, we are rather interested in the macroscopic
properties of Hawkes processes, in the sense of obtaining a limit behaviour
for the multivariate process (N
Tv
)
v[0,1]
as T , for a suitable normal-
isation. Our interest originates in nancial data modelling: in [17], we in-
troduced a stochastic model for the price S = (S
1
, . . . , S
n
) of a multivariate
asset, based on a d = 2n dimensional Hawkes process of the form (1), with
representation
S
1
= N
1
N
2
, S
2
= N
3
N
4
, . . . , S
n
= N
d1
N
d
. (2)
In this context, the fact that the S
i
take values on Z accounts for the discret-
ness of the price formation under a limit order book. If we take
2i1
=
2i
for every 1 i d, the process S is centred and the mutually exciting
properties of the intensity processes
i
under (1) allow to reproduce em-
pirically microstructure noise and the Epps eect, as demonstrated in [17].
Microstructure noise a major stylised fact in high frequency nancial data
(see e.g. [10, 21, 23, 24, 28, 29]) is characterised by the property that in
microscopic scales
1
, an upward change of price for the component S
i
is more
likely to be followed by a downward jumps of S
i
and vice versa. Similarly,
a jump of a component N
2i1
or N
2i
of S
i
at time t will increase the con-
ditional probability, given the history up to time t, that a jump occurs for
the component N
2j1
or N
2j
of S
j
through the exciting eect of the kernels

2i1,2j1
,
2i1,2j
,
2i,2j1
or
2i,2i
, thus creating a certain dependence
structure and the Epps eect, another well-document stylised fact implying
that the empirical correlation bewteen two assets vanishes on microscopic
scales [11]. However, both microstructure noise and the Epps eect vanish
at coarser scales where a standard diusion regime dominates. Thus, a cer-
tain macroscopic stability property is desirable, namely the property that
S behaves like a continuous diusion on coarse scales. Also, one would like
to be able to track the aforementioned microscopic eects produced by the
kernels
ij
in the diusion limit of S. This is the main topic of the paper
1
when the data are sampled every few seconds or less.
3
and also its novelty, as far as the modelling angle is concerned.
The importance of mutually exciting processes has expanded in nancial
econometrics over the last years, see [5, 7, 16, 20] among others. From a
wider perspective however, a mathematical analysis of the limiting behaviour
of Hawkes processes has a relevance beyond nance: apart from the interest
of such a study for its own sake, one could also presumably apply asymptotic
results in other application elds. These include seismology of course, for
which Hawkes processes were originally introduced (see [31] for instance) and
also the citation we used above from the monograph [8]; in a more speculative
or prospective way, one could mention recent related areas such as trac
networks [6] or genomic analysis [27] where mutually exciting processes take
a growing importance.
1.2 Main results and organisation of the paper
Because of this versatility in the use of Hawkes processes, we present our
results in a fairly general setting in Sections 2, 3 and 4 with relatively mini-
mal assumptions. It is only in the subsequent Section 5 that we consistently
apply our results to the multivariate price model S described by (1)-(2) for
nancial data modelling.
In Section 2, we recall the construction of Hawkes processes in a rigourous
and general setting. We state in particular the fact that as soon as the
kernels
ij
that dene the intensity process are locally integrable, the ex-
istence of N with uniqueness in law is guaranteed on a rich enough space,
as follows from the theory of predictable projection of integer-valued ran-
dom measures of Jacod [18]. We state in Section 3 a law of large numbers
(Theorem 1) and a functional central limit theorem (Theorem 2). The law
of large numbers takes the form
sup
v[0,1]
_
_
T
1
N
Tv
v (Id K)
1

_
_
0 as T (3)
almost surely and in L
2
(P). We set for the Euclidean norm on R
d
. The
limit is described by = (
1
, . . . ,
d
) identied with a column vector and
the d d matrix K =
_

0
(t) dt, where = (
ij
) is a matrix of functions
from R
+
to R
+
. The convergence (3) holds under the assumption that the
spectral radius of K is strictly less than 1, which guarantees in particular all
the desirable integrability properties for N. As for a central limit theorem,
under the additional condition that t
1/2
(t) is integrable componentwise, a
4
consequence of Theorem 2 is the convergence of the processes

T
_
T
1
N
Tv
v(Id K)
1

_
, v [0, 1]
to
(Id K)
1

1/2
W
v
, v [0, 1].
in law for the Skorokod topology. The limit is described by a standard d-
dimensional Brownian motion (W
v
)
v[0,1]
and the d d diagonal matrix
dened by
ii
=
_
(Id K)
1

_
i
. The proof of Theorems 1 and 2 relies
on the intensity theory of point processes: using that N
t

_
t
0

s
ds is a
d-dimensional martingale, if we set X
t
= N
t
E(N
t
) with E() denoting
expectation, we then have the fundamental representation
X
t
= M
t
+
_
t
0
(t s)X
s
ds
of X as a sum of a martingale and a convolution product of with X itself.
Separating the two components, we can then take advantage of the powerful
theory of limit theorems of semimartingales, as exposed in the comprehen-
sive book of Jacod and Shiryaev [19].
In Section 4, we consider yet another angle of study that is useful for
applications. Given two square integrable point processes N and N
t
with
value in N
d
, setting X
t
= N
t
E(N
t
) and X
t
t
= N
t
t
E(N
t
t
), one can dene
the empirical cross-correlation of N and N
t
at scale over [0, T] as
V
,T
(N, N
t
) =
1
T
]T/|

i=1
_
X
i
X
(i1)
__
X
t
i
X
t
(i1)
_

,
where X
t
and X
t
t
are identied as column vectors and ()

denotes trans-
position in R
d
. In rigorous terms, V
,T
(N, N
t
) shall rather be called the
empirical correlation matrix of the increments of X and X
t
sampled at scale
over [0, T]. This object is of major importance in practice: viewed as a
function of , it reveals the correlation structure across scales between N
and N
t
. It is crucial for the understanding of the transition of (N
Tv
)
v[0,1]
from a microscopic regime of a point process to the macroscopic behaviour of
a Brownian diusion, up to appropriate normalisation, as T . More pre-
cisely, given a time shift =
T
R, we focus in the paper on the -shifted
empirical correlation matrix of N at scale , namely
2
V
,T
(N, N
+
). We
2
This is actually a more general object since we can formally recover V,T (N, N

) from
V,T (N

, N

+
), where N

= (N, N

) is a 2d-dimensional point process by letting 0

in V,T (N

, N

+
).
5
prove in Theorem 3 that V
,T
(N, N
+
) is close in L
2
(P) to a deterministic
explicit counterpart
v
,
=
_
R
2
+
_
1
[ts[

_
+
_
Id
0
(ds) +(s)ds
_

_
Id
0
(dt) +(t)

dt
_
under appropriate conditions on and relative to T as T . The
approximation v
,
is described by that already appears in Theorems 1
and 2 and =

n1

n
, where
n
is the matrix of n-fold convolution prod-
uct of . This apparently cumbersome formula paves the way to explicit
computations that yield crucial information about the dependence structure
of functionals of N, as later developed in the paper.
Section 5 is devoted to some applications of Theorems 1, 2 and 3 to the
study of the multivariate price model S described by (1)-(2) for nancial data
modelling. We elaborate on particular examples the insight given by v
,
.
Concerning macroscopic limits of a univariate price S = N
1
N
2
obtained
for d = 2, we derive in Proposition 1 the limiting variance of the rescaled
process T
1/2
(N
1,Tv
N
2,Tv
) in the special case where
1,2
=
2,1
=
and
1,1
=
2,2
= 0. This oversimplication enables to focus on the phe-
nomenon of microstructure noise and reveals the macroscopic trace of in
the variance of the limiting Brownian motion. We discuss the theoretical
implications of the obtained formulas in statistical nance. In the same way,
we explicit in Proposition 2 the macroscopic correlation structure obtained
with a bivariate price process S = (S
1
, S
2
) = (N
1
N
2
, N
3
N
4
) when only
cross-excitations
1,3
=
3,1
and
2,4
=
2,4
are considered, the other com-
ponents
ij
being set to 0 for simplicity. Again, we obtain an illuminating
formula that characterises the inuence of the microscopic dynamics on the
macroscopic correlation of the Brownian limit. This enables in particular in
Proposition 3 to demonstrate how Hawkes process can account for the Epps
eect [11] and also the lead-lag eect between two nancial assets that has
been given some attention in the literature recently [1, 2, 30].
Section 6 develops some preliminary tools for the proofs. Section 7, 8
and 9 are devoted to the proof of Theorems 1, 2 and 3 respectively. Some
technical results and computations of the application Section 5 are delayed
until an appendix.
6
2 Multivariate Hawkes processes
Consider a measurable space (, F) on which is dened a non-decreasing
sequence of random variables (T
n
)
n1
taking their values in (0, ], and such
that T
n
< T
n+1
on the event {T
n
< }, for all n 1. Let (Z
n
)
n1
be a
sequence of discrete random variables taking their values in {1, . . . , d} for
some positive integer d. Dene, for t 0
N
i,t
=

n1
1
TntZn=i
.
Remark that N
i,0
= 0 by construction. We endow with the ltration
(F
t
)
t0
where F
t
is the -algebra generated by the random variables N
i,s
,
s t, 1 i d. According to Jacod [18], for any progressively measurable
non-negative processes (
1,t
)
t0
, . . . , (
d,t
)
t0
satisfying
_
Tn
0

i,s
ds < almost-surely,
there exists at most one probability measure P on (, F

) such that the

compensator (or predictable projection) of the integer-valued random mea-
sure
N(dt, dx) =

n1
1
Tn<

(Tn,Zn)
(dt, dx)
on (0, ) {1, . . . , d} is
(dt, dx) =
d

i=1

i,t
dt
i
(dx),
where is the Dirac mass. In other words, for all n 1, for all i {1, . . . , d},
the process
N
i,tTn

_
tTn
0

i,s
ds
is a (F
t
)-martingale. This implies that the law of the d-dimensional process
(N
1
, . . . , N
d
) is characterised by (
1
, . . . ,
d
). Moreover if is rich enough
we have the existence of such a probability measure P.
Denition 1. We say that N = (N
1
, . . . , N
d
) is a multivariate Hawkes pro-
cess when

i,t
=
i
+
_
(0,t)
d

j=1

ij
(t s)dN
j,s
(4)
where
i
R
+
and
i,j
is a function from R
+
to R
+
.
7
We have a non-explosion criterion, as a consequence for instance of equal-
ity (10) in Lemma 2, Section 6 below.
Lemma 1 (Non-explosion criterion). Set T

= lim
n
T
n
. Assume that the
following holds:
_
t
0

ij
(s)ds < for all i, j, t. (5)
Then T

= almost surely.
3 Law of large numbers and functional central limit
theorem
On a rich enough probability space (, F, P), we consider a Hawkes process
N = (N
t
)
t0
according to Denition 1, satisfying (5) and specied by the
vector
= (
1
, . . . ,
d
)
and the d d-matrix valued function
= (
i,j
)
1i,jd
.
Note that in this setting, we do not assume a stationary regime for N.
Consider the assumption
For all i, j we have
_

0

ij
(t)dt < and the spectral radius
(K) of the matrix K =
_

0
(t) dt satises (K) < 1.
(A1)
First we have a law of large numbers in the following sense:
Theorem 1. Assume that (A1) holds. Then N
t
L
2
(P) for all t 0 and
we have
sup
v[0,1]
_
_
T
1
N
Tv
v (Id K)
1

_
_
0 as T
almost-surely and in L
2
(P).
Next we have an associated functional central-limit theorem. Introduce
the functions
n
dened on R
+
and with values in the set of d d-matrices
with entries in [0, ] by

1
= ,
n+1
(t) =
_
t
0
(t s)
n
(s) ds, n 1. (6)
8
Under (A1) we have
_

0

n
(t) dt = K
n
hence the series

n1

n
converges
in L
1
(dt). We set
=

n1

n
(7)
Theorem 2. Assume that (A1) holds. We have
E(N
t
) = t +
_
_
t
0
(t s)s ds
_

where L
1
(dt) is given by (7). Moreover, the processes
1

T
_
N
Tv
E(N
Tv
)
_
, v [0, 1]
converge in law for the Skorokod topology to
(Id K)
1

1/2
W
v
, v [0, 1]
as T , where (W
v
)
v[0,1]
is a standard d-dimensional Brownian motion
and is the diagonal matrix such that
ii
= ((Id K)
1
)
i
.
Consider now the following restriction on :
_

0
(t) t
1/2
dt < componentwise. (A2)
Using Theorem 1 and Assumption (A2), we may replace T
1
E(N
Tv
) by its
limit in Theorem 2 and obtain the following corollary.
Corollary 1. Assume that (A1) and (A2) hold. Then the processes

T
_
1
T
N
Tv
v(Id K)
1

_
, v [0, 1]
converge in law for the Skorokod topology to
(Id K)
1

1/2
W
v
, v [0, 1]
as T .
9
4 Empirical covariation across time scales
For two square integrable counting processes N and N
t
with values in N
d
,
set
X
t
= N
t
E(N
t
), X
t
t
= N
t
t
E(N
t
t
).
The empirical covariation across time scales of N and N
t
is the process
V
,T
(N, N
t
), T > 0, > 0, taking values in the set of d d matrices and
dened as
V
,T
(N, N
t
) =
1
T
]T/|

i=1
_
X
i
X
(i1)
__
X
t
i
X
t
(i1)
_

where
_
X
i
X
(i1)
_
is identied as a column vector and
_
X
i
X
(i1)
_

denotes its transpose, and we set X

t
= 0 for t 0. More precisely, given a
time shift R, we are interested in the behaviour of V
,T
(N, N
+
). In
essence, V
,T
(N, N
+
) can be viewed as a multivariate cross-correlogram
across scales of N: it can be consistently measured from empirical data
and its limiting behaviour as T plays a key tool in understanding the
second-order structure of linear functions of N across scales .
Theorem 3. In the same setting as in Theorem 2, let (
T
)
T>0
and (
T
)
T>0
be two families of real numbers such that
T
> 0. If
T
/T 0 and

T
/T 0 as T , we have
V

T
,T
(X, X

T
+
) v

T
,
T
0 as T in L
2
(P)
where
v
,
=
_
1
[[

_
+
+
_
R
2
+
ds dt
_
1
[ts[

_
+
(s)(t)

+
+
_

0
ds (1
[s+[

)
+
(s)+
_

0
ds (1
[s[

)
+
(s)

, (8)
or equivalently,
v
,
=
_
R
2
+
_
1
[ts[

_
+
_
Id
0
(ds) +(s)ds
_

_
Id
0
(dt) +(t)

dt
_
,
where is the diagonal matrix such that
ii
=
_
(Id K)
1

_
i
and the
function is given by (7).
10
Remark 1. On can check that v
,
= E
_
(N

N
0
)(N
+
N

_
where N
is a counting process of the (unique in law) stationary multivariate Hawkes
process on R associated to and . Thus, another way to obtain v
,
is to
compute E
_
(N

N
0
)(N
+
N

_
in the stationary regime, as in [1214]
by means of the Bartlett spectrum of N, see [4]. However, the stationary
restriction is superuous and moreover, only very specic parametric form
of
ij
like exponential functions enable to carry such computations.
Remark 2. Obviously, we have v
,
= v

,
.
Remark 3. For xed R, we retrieve the macroscopic variance of Theorem
2 by letting . More precisely, we have
v
,

as
where = (Id K)
1
, and the eect of vanishes as . For all
= 0 we have
v
,
0 as 0.
This convergence simply expresses the fact that the two processes N and
N
+
cannot jump at the same time, producing a at autocorrelogram for
suciently small sampling mesh .
In the same way as Corollary 1 is obtained from Theorem 2, we have the
following renement of Theorem 3.
Corollary 2. Finally, in the same setting as in Theorem 3, assume more-
over that (A2) holds. Dene

X
t
=
_
N
t
t(Id K)
1
if t 0,
0 if t < 0.
We have
V

T
,T
(

X,

X

T
+
) v

T
,
T
0 as T in L
2
(P),
where v
,
is given by (8).
5 Application to nancial statistics
5.1 The macroscopic trace of microstructure noise
Following [17], we introduce a univariate price process S = (S
t
)
t0
by setting
S = N
1
N
2
,
11
where (N
1
, N
2
) is a Hawkes process in the case d = 2, with
_

1,t
= +
_
(0,t)
(t s)dN
2,s
,

2,t
= +
_
(0,t)
(t s)dN
1,s
for some R
+
and : R
+
R
+
. With our notation, this corresponds to
having = (, ) and
=
_
0
0
_
If = 0, we nd back a compund Poisson process with intensity and
symmetric Bernoulli jumps. This corresponds to the simplest model for a
random walk in continuous time, constrained to live on a lattice, the tick-grid
in nancial statistics accounting for the discreteness of price at n scales.
Microstructure noise corresponds to the property that an upward jump of
S will be more likely followed by a downward jump and vice versa. This
phenomenon lays its roots in microeconomic analysis of price manipulation
of agents [10, 21, 23, 24, 28, 29]. In our simple phenomenological setting,
it will be reproduced by the introduction of the kernel , as empirically
demonstrated in [17]. The question we can now address is the macroscopic
stability of the model. Do we retrieve a standard diusion in the limit
T for an approriate scaling of S and how does the eect of inuence
the macroscopic volatility? By Theorems 1 and 2, we readily obtain an
explicit anwser:
Proposition 1 (Macroscopic trace of microstructure noise). Assume that

L
1 =
_

0
(t)dt < 1. Then
_
T
1/2
S
Tv
, v [0, 1]
_

_
W
v
, v [0, 1]
_
as T ,
in law for the Skorokod topology, where (W
v
)
v[0,1]
is a standard Brownian
motion and

2
=
2
(1
L
1)(1 +
L
1)
2
.
Remark 4. Note that if we take = 0, we retrieve the standard convergence
of a compound Poisson process with symmetric jump to a Brownian motion.
Remark 5. By assumption, 0
L
1 < 1 and a closer inspection of the
function
x (, x)
2
=
2
(1 x)(1 +x)
2
for x [0, 1)
12
reveals an interesting feature: for small microstructure eect (namely if
x =
L
1 less than 1/3) the eect of microstructure tends to stabilise the
macroscopic variance in the sense that (, x)
2
(, 0)
2
, whereas beyond
a critical value x 0.61, we have (, x)
2
(, 0)
2
and even (, x)
2

as x 1.
5.2 Macroscopic correlations for bivariate assets
We now turn to a bivariate price model S = (S
1
, S
2
) obtained from a Hawkes
process in dimension d = 4, of the form
(S
1
, S
2
) = (N
1
N
2
, N
3
N
4
)
with = (
1
,
2
,
3
,
4
) such that

1
=
2
and
3
=
4
together with
=
_
_
_
_
0 0 h 0
0 0 0 h
g 0 0 0
0 g 0 0.
_
_
_
_
The upward jumps of S
1
excite the intensity of the upward jumps of S
2
and,
in a symmetric way, the downward jumps of S
1
excite the downward jumps
of S
2
via the kernel g : R
+
R
+
. Likewise, the upward jumps of S
2
excite
the upward jumps of S
1
and the downward jumps of S
2
excite the downward
jumps of S
1
via the kernel h : R
+
R
+
. For simplicity we ignore other
cross terms that could produce microstructure noise within the inner jumps
of S
1
and S
2
. This relativeley simple dependence structure at a microscopic
level enables to obtain a non-trivial form of the macroscopic correlation of
the diusion limits of S
1
and S
2
.
Proposition 2. Assume that h
L
1g
L
1 < 1. The 2-dimensional processes
T
1/2
_
S
1,Tv
, S
2,Tv
_
v[0,1]
converge in law as T for the Skorokod topol-
ogy to
_
X
1
X
2
_
=

2
(1 h
L
1g
L
1)
3/2
_

1/2
1
W
1
+
1/2
2
h
L
1 W
2

1/2
1
g
L
1 W
1
+
1/2
2
W
2
_
with

1
=
1
+h
L
1
3
,
2
=
3
+g
L
1
1
. (9)
and where (W
1
, W
2
) = (W
1,t
, W
2,t
)
t[0,1]
is a standard Brownian motion.
13
The proof is a consequence of Theorems 1 and 2 and is given in appendix.
Remark 6. The macroscopic correlation between S
1
and S
2
is thus equal to
the cosine of the angle of the two vectors
_

1/2
1
,
1/2
2
h
L
1
_
and
_

1/2
1
g
L
1,
1/2
2
_
.
Obviously, it is always nonnegative and strictly less than 1 since the deter-
minant
1/2
1

1/2
2
_
1 h
L
1g
L
1
_
of the two above vectors is positive unless
the
i
are all 0.
5.3 Lead-lag and Epps eect through the cross correlations
across-scales
We keep up with the model and the notation of Section 5.2 but we now
study the quantities
V
,T
(S
1
, S
1,+
), V
,T
(S
2
, S
2,+
), V
,T
(S
1
, S
2,+
).
In particular, the quantity V
,T
(S
1
, S
2,+
) is a powerful tool for the statisti-
cal study of lead-lag eects, i.e. the fact that the jumps of S
1
can anticipate
on those of S
2
and vice-versa, see for instance [1, 30]. The Epps eect i.e.
the stylised fact statement that the correlation between the increments of
two assets vanishes at ne scales can be tracked down likewise. Theorem
3 enables to characterise in principle the limiting behaviour of these func-
tionals. This is described in details in Proposition 3 below.
We consider g and h as functions dened on R by setting g(t) = h(t) = 0
is t < 0. Assume that h
L
1g
L
1 < 1. Then the series
F :=

n1
(h g)
n
converges in L
1
(R, dt). If f is a function on R we dene

f by

f(t) = f(t).
We have
Proposition 3. Assume that h
L
1g
L
1 < 1. Let (
T
)
T>0
and (
T
)
T>0
be
two families of real numbers such that
T
> 0. If
T
/T 0 and
T
/T 0
as T we have
V

T
,T
(S
1
, S
1,
T
+
) C
11
(
T
,
T
) 0 as T in L
2
(P),
where
C
11
(, ) =
2
1|h|
L
1
|g|
L
1

0
+F +

F +F

F
_

0
+
2
, h

h
_
(),
14
with

(x) =
_
1 |x|/
_
+
.
We also have
V

T
,T
(S
1
, S
2,
T
+
) C
12
(
T
,
T
) 0 as T in L
2
(P)
and
V

T
,T
(S
2
, S
1,
T
+
) C
12
(
T
,
T
) 0 as T in L
2
(P),
with
C
12
(, ) =
2
1|h|
L
1
|g|
L
1

0
+F +

F +F

F
_

h +
1
g
_
().
Remark 7 (The Epps eect). For f L
1
we have

f 0 pointwise as
0. Therefore we obtain
C
12
(, ) 0 as 0 for every R
and this characterises the Epps eect. The same argument, together with

(0) = 1, yields the convergence

C
11
(, )
2
1
1 h
L
1g
L
1
1
=0
as 0.
Remark 8 (The Lead-Lag eect). Following [17] and as a consensus in the
literature, we say that a lead-lag eect is present between S
1
and S
2
if there
exists > 0 and = 0 such that
C
12
(, ) = C
12
(, ).
Therefore, an absence of any lead-lag eect is obtained if and only if the
function
_

0
+ F +

F + F

F
_

h +
1
g
_
is even. This is the case if

1
g =
2
h. Now, let > 0. If g = h

and
1
=
3
, then
C
12
(, ) = C
12
(, +) for every > 0, R.
This particular choice for h and g models in particular the property that S
1
acts on S
2
in the same manner as S
2
acts on S
1
with an extra temporal shift
of . Since we always have
lim

C
12
(, ) = 0,
there exists
0
such that C
12
(,
0
+ ) = C
12
(,
0
), or in other words, we
have a lead-lag eect.
15
Remark 9 (Macroscopic correlations). Since

1 as , we obtain
the convergence
C
11
(, )
2
1|h|
L
1
|g|
L
1
_
_

0
+F +

F +F

F
_

0
+
2
h

h
_
=
2
1|h|
L
1
|g|
L
1
_
1 + 2 F + ( F)
2
_
_

1
+
2
( h)
2
_
= Var(X
1
).
Likewise, we have C
12
(, ) Cov(X
1
, X
2
) as .
Remark 10. Finally, note that if we use the convenient parametrisation
h(t) =
1
exp(
1
t) and g(t) =
2
exp(
2
t), assuming further h
L
1g
L
1 =

2
/(
1

2
) < 1, then standard computations yield the explicit form
F(t) =

1

1

2
_
exp(
2
t) exp(
1
t)
_
1
R
+
(t)
with

1
=
1
2
_

1
+
2
+
_
(
1

2
)
2
+ 4
1

2
_
,

2
=
1
2
_

1
+
2

_
(
1

2
)
2
+ 4
1

2
_
and we have 0 <
2
<
1
. This allows to obtain a close formula for C
11
and
C
12
. We do not pursue these computations here.
6 Preparation for the proofs
In the sequel, we work in the setting of Sections 2 and 3 under Assumption
(A1).
Lemma 2. For all nite stopping time S one has:
E(N
S
) = E(S) +E
_
_
S
0
(S t)N
t
dt
_
(10)
E(N
S
) (Id K)
1
E(S) componentwise. (11)
Proof. Recall that (T
p
)
p1
denote the successive jump times of N and set
S
p
= S T
p
. Since the stochastic intensities
i
are given by (4) one has
E(N
Sp
) = E(S
p
) +E
_
_
Sp
0
dt
_
(0,t)
(t s)dN
s
_
.
16
Moreover, by Fubini theorem
_
Sp
0
dt
_
(0,t)
(t s)dN
s
=
_
[0,Sp)
_
_
Sp
s
(t s)dt
_
dN
s
=
_
[0,Sp)
_
_
Sps
0
(t)dt
_
dN
s
.
Now, integrating by part with (t) =
_
t
0
(s)ds, we can write
0 = (0)N
Sp
(S
p
)N
0
=
_
(0,Sp]
(S
p
t)dN
t

_
Sp
0
(S
p
t)N
t
dt.
Remark that both sides of the above equality are nite since

d
i=1
N
i,Sp
p.
We obtain
E(N
Sp
) = E(S
p
) +E
_
_
Sp
0
(S
p
t)N
t
dt
_
and derive (10) using that N
Sp
N
S
as p and
_
Sp
0
(S
p
t)N
t
dt =
_
Sp
0
(t)N
Spt
dt
_
S
0
(t)N
St
dt =
_
S
0
(St)N
t
dt.
We next prove (11). We have
E(N
Sp
) = E
_
S
p
+
_
Sp
0
(S
p
t)N
t
dt
_
E(S
p
) +E
_
_

0
(S
p
t)N
t
dt
_
componentwise,
= E(S
p
) +KE(N
Sp
).
By induction
E(N
Sp
)
_
Id +K+ +K
n1
_
E(S
p
) +K
n
E(N
Sp
)
componentwise for all integer n. On the one hand,

d
i=1
N
i,Sp
p. On the
other hand, since (K) < 1 we have K
n
0 as n and

n=0
K
n
= (Id K)
1
,
therefore
E(N
Sp
) (Id K)
1
E(S).
This readily yields (11) since E(N
S
) = lim
p
E(N
Sp
).
17
Let
n
, n 1, and =

n1

n
be dened as in Theorem 2. By
induction it is easily shown that
_

0

n
(t)dt = K
n
for all n. Therefore
_

0
(t)dt =

n1
K
n
is nite componentwise by assumption (A1). We
next state a multivariate version of the well known renewal equation, which
proof we recall for sake of completeness.
Lemma 3. Let h be a Borel and locally bounded function from R
+
to R
d
.
Then there exists a unique locally bounded function f : R
+
R
d
solution to
f(t) = h(t) +
_
t
0
(t s)f(s)ds t 0, (12)
given by
f
h
(t) = h(t) +
_
t
0
(t s)h(s)ds.
Proof. Since L
1
(dt) and h is locally bounded, the function f
h
is locally
bounded. Moreover f
h
satises (12). It follows that
_
t
0
(t s)f
h
(s)ds =
_
t
0
(t s)h(s)ds +
_
t
0
ds(t s)
_
s
0
(s r)h(r)dr
=
_
t
0
(t s)h(s)ds +
_
t
0
dr
_
t
r
ds(t s)(s r)h(r)
=
_
t
0
(t r)h(r)dr
since
_
t
0
(t s)(s)ds = (t) (t). As for the uniqueness, if f satises
(12) then
f
h
(t) f(t) =
_
t
0
(t s)(f
h
(s) f(s))ds
thus if g
i
(t) = |f
h,i
(t) f
i
(t)|, 1 i d, one has
g(t)
_
t
0
(t s)g(s)ds componentwise,
which yields
_

0
g(t)dt K
_

0
g(t)dt componentwise.
Since (K) < 1 it follows that f = f
h
almost everywhere. Therefore
_
t
0
(t s)f(s)ds =
_
t
0
(t s)f
h
(s)ds for all t
and thus f = f
h
since both function satises (12).
18
Dene the d-dimensional martingale (M
t
)
t0
by
M
t
= N
t

_
t
0

s
ds with = (
1
, . . . ,
d
).
Lemma 4. For all t 0:
E(N
t
) = t +
_
_
t
0
(t s)s ds
_
, (13)
N
t
E(N
t
) = M
t
+
_
t
0
(t s)M
s
ds. (14)
Proof. By (10) of Lemma 2 and Fubini theorem, we get
E(N
t
) = t +
_
t
0
(t s)E(N
s
) ds.
Besides, t E(N
t
) is locally bounded in view of (11). Applying Lemma
3 we obtain (13). The second formula follows from Lemma 3 and the fact
that, if X
t
= N
t
E(N
t
), representation (13) entails
X
t
= M
t
+
_
t
0
(t s)X
s
ds.
7 Proof of Theorem 1
Lemma 5. Let p [0, 1) and assume that
_

0
t
p
(t) dt < component-
wise. Then
1. If p < 1, we have
T
p
_
T
1
E(N
Tv
) v(Id K)
1

_
0 as T
uniformly in v [0, 1].
2. If p = 1, we have
T
_
1
T
E(N
T
) (Id K)
1

_
(Id K)
1
_
_

0
t(t) dt
_
(Id K)
1
as T .
19
Proof. Let p [0, 1] and assume that
_

0
t
p
(t) dt < componentwise.
We rst prove that
_

0
t
p
(t) dt < componentwise. For n 1, setting
A
n
=
_

0
t
p

n
(t) dt, we can write
A
n+1
=
_

0
t
p
_
_
t
0
(t s)
n
(s) ds
_
dt
=
_

0
_
_

0
(t +s)
p
(t) dt
_

n
(s) ds

_

0
t
p
(t) dt K
n
+K
_

0
s
p

n
(s) ds with equality if p = 1,
= A
1
K
n
+KA
n
.
Therefore for all integer N,
N

n=1
A
n
A
1
+A
1
N1

n=1
K
n
+K
N1

n=1
A
n
,
(Id K)
N1

n=1
A
n
+A
N
A
1
+A
1
N1

n=1
K
n
and
N1

n=1
A
n
(Id K)
1
(A
1
+A
1
N1

n=1
K
n
).
Letting N we derive
_

0
t
p
(t) dt =

n1
A
n
(Id K)
1
A
1
(Id K)
1
with equality if p = 1. From (13) it follows that for all v [0, 1]:
v(Id K)
1

1
T
E(N
Tv
) =
_
v
_

Tv
(t)dt +
1
T
_
Tv
0
t (t)dt
_
(15)
Since t
p
(t) is integrable, we have
T
p
_

Tv
(t)dt v
1p
_

Tv
t
p
(t) dt 0 as T
and this convergence is uniform in v [0, 1] in the case p < 1. It remains to
prove that if p < 1, we have
1
T
1p
_
T
0
t(t)dt 0 as T .
20
With G(t) =
_
t
0
s
p
(s)ds, integrating by part, we obtain
T
1p
G(T) =
_
T
0
t(t) dt + (1 p)
_
T
0
t
p
G(t) dt
and
1
T
1p
_
T
0
(t)tdt = G(T)
1 p
T
1p
_
T
0
t
p
G(t) dt.
Since G(t) is convergent as t we nally derive that the right-hand-side
in the above equality converges to 0 as T .
Denote by the Euclidean norm either on R
d
or on the set of d d
matrices.
Lemma 6. There exists a constant C
,
such that for all t, 0:
E
_
sup
tst+
M
s
M
t

2
_
C
,
.
Proof. Doobs inequality yields
E
_
sup
tst+
M
s
M
t

2
_
4
d

i=1
E
_
(M
i,t+
M
i,t
)
2
_
.
For each i {1, . . . , d}, the quadratic variation of the martingale (M
i,t
)
t0
is
_
M
i
, M
i

t
=

st
(M
i,s
M
i,s
)
2
= N
i,t
.
Thus we have
E((M
i,t+
M
i,t
)
2
) = E(N
i,t+
N
i,t
).
Besides, in view of Lemma 4 and the fact that
_

0
(t)dt = (Id K)
1
Id,
we obtain
E(N
t+
N
t
) (Id K)
1
componentwise.
Completion of proof of Theorem 1. Lemma 5 with p = 0 implies that it is
enough to prove the following convergence
T
1
sup
v[0,1]
_
_
N
Tv
E(N
Tv
)
_
_
0 as T (16)
21
almost surely and in L
2
(P). Thanks to (14) of Lemma 4, we have
sup
v[0,1]
N
Tv
E(N
Tv
)
_
1 +
_
T
0
(t)dt
_
sup
tT
M
t
,
C

sup
tT
M
t

since is integrable. Moreover

E
_
sup
tT
M
t

2
_
C
,
T
by Lemma 6, therefore convergence (16) holds in L
2
(P). In order to prove
the almost-sure convergence, it is enough to show that
T
1
sup
v[0,1]
_
_
M
Tv
_
_
0 as T almost-surely.
Let
Z
t
= (Z
1,t
, . . . , Z
d,t
) =
_
(0,t]
1
s + 1
dM
s
.
The quadratic variation of the martingale Z
i
satises
[Z
i
, Z
i
]
t
=

0<st
(Z
i,s
Z
i,s
)
2
=
_
(0,t]
1
(s + 1)
2
dN
i,s
and moreover, using integration by part and (13), we have
E
_
_
(0,)
1
(s + 1)
2
dN
s
_
= 2E
_
_
(0,)
N
s
(1 +s)
3
ds
_
< .
Therefore lim
t
Z
t
exists and is nite almost surely. It follows that
1
t + 1
M
t
= Z
t

1
t + 1
_
t
0
Z
s
ds 0 as T almost surely.
We deduce that almost surely, for all family v
T
[0, 1], T > 0, such
that Tv
T
the convergence M
Tv
T
/T 0 holds. Moreover, we have
M
Tv
T
/T 0 if sup
T
Tv
T
< . In other words T
1
M
Tv
0 uniformly
uniformly in v [0, 1], almost-surely. The proof of Theorem 1 is com-
plete.
22
8 Proof of Theorem 2
Let W = (W
1
, . . . , W
d
) be a standard d-dimensional Brownian motion. For
i = 1, . . . , d, put
i
= (
ii
)
1/2
.
Lemma 7. The martingales M
(T)
:= (T
1/2
M
Tv
)
v[0,1]
converge in law for
the Skorokod topology to (
1
W
1
, . . . ,
d
W
d
).
Proof. According to Theorem VIII-3.11 of [19], since the martingales M
(T)
have uniformly bounded jumps, a necessary and sucient condition to ob-
tain the lemma is: for all v [0, 1], for all 1 i < j d
[M
(T)
i
, M
(T)
i
]
v

2
i
v, [M
(T)
i
, M
(T)
j
]
v
0, as T in probability.
We have
[M
(T)
i
, M
(T)
i
]
v
=
1
T
N
i,Tv

2
i
v in L
2
(P) by Theorem 1
and
[M
(T)
i
, M
(T)
j
]
v
= 0
since the processes N
i
for 1 i d, have no common jump by construction.
Completion of proof of Theorem 2. Set
X
(T)
v
= T
1/2
_
N
Tv
E(N
Tv
)
_
.
In view of Lemma 7, it is enough to prove that
sup
v[0,1]
_
_
X
(T)
v
(Id K)
1
M
(T)
v
_
_
0 as T in probability.
By Lemma 4, we have
X
(T)
v
= M
(T)
v
+
_
v
0
T(Tu)M
(T)
vu
du,
hence we need to prove that
sup
v[0,1]
_
_
_
v
0
T(Tu)M
(T)
vu
du
_
_

0
(t)dt
_
M
(T)
v
_
_
0 as T (17)
23
in probability. We plan to use the fact that is integrable and the C-
tightness of the family (M
(T)
)
T>0
. The tightness is a consequence of Lemma
7 and reads:
> 0 limsup
T
P
_
sup
[uu

[
_
_
M
(T)
u
M
(T)
u

_
_
>
_
0 as 0. (18)
using also that M
(T)
0
= 0. For > 0 and v [0, 1] we have
_
_
_
v
v
T(Tu)M
(T)
vu
du
_
_
sup
0u1
M
(T)
u

_
1

T(Tu)du
sup
0u1
M
(T)
u

_

T
(t)dt 0
as T in probability, since sup
0u1
M
(T)
u
is bounded in probability
and
_

T
(t)dt 0 as T . Moreover
_
_
_
v
0
T(Tu)
_
M
(T)
v
M
(T)
vu
_
du
_
_
sup
[uu

[
_
_
M
(T)
u
M
(T)
u

_
_
_

0
(t)dt,
therefore, in order to prove (17) it suces to show that for all > 0, the
convergence
sup
v[0,1]
_
_
_
_

0
(t)dt
_
v
0
T(Tu)du
_
M
(T)
v
_
_
0 as T
holds in probability. It readily follows from (18) and the upper-bound
_
_
_
_

0
(t)dt
_
v
0
T(Tu)du
_
M
(T)
v
_
_

_
_

T
(t)dt sup
u
M
(T)
u
if v >
_

0
(t)dt sup
u
M
(T)
u
if v ,
with 0 < < .
Proof of Corollary 1. By (A2), Lemma 5 with p = 1/2 yields
T
1/2
_
T
1
E(N
Tv
) v(Id K)
1

_
0 as T
uniformly in v [0, 1]. Moreover, by (A1), Theorem 2 yields
T
1/2
_
T
1
N
Tv
T
1
E(N
Tv
)
_
(Id K)
1

1/2
W
in distribution as T and the result follows.
24
9 Proof of Theorem 3
Set
Y
t
=
_
t
0
(t s)M
s
ds (19)
in order that X = M + Y , see Lemma 4. For all 0 1, for all
integer 1 k
0
T/, dene
D
k
0
,,
(X)
T,
=
1
T
]T/|

k=k
0
_
X
(k1)+
X
(k1)+
_
.
Lemma 8. There exists a function T
,
(T) such that
,
(T) 0 as
T and such that for all 0 1 and all integer 1 k
0
T/,
we have
E
__
_
(1
k
0

T
)( )(Id K)
1
D
k
0
,,
(N)
T,
_
_
2
_

,
(T).
Proof. First we prove that
E
__
_
D
k
0
,,
(X)
T,

2
_
C
,
T
1
(20)
for some constant C
,
that depends on and only. Using
X
(k1)+
X
(k1)+
= M
(k1)+
M
(k1)+
+
_

0
ds (s)
_
M
(k1)+s
M
(k1)+s
_
and the fact that is integrable, it suces to prove that
E
__
_
D
k
0
,,
(M
s
)
T,

2
_
C
,
T
1
.
Set
S
n
=
n

k=1
_
M
(k1)+s
M
(k1)+s
_
.
Clearly (S
n
)
n1
is a discrete martingale thus
E
__
_
S
]T|
S
k
0
1
_
_
2
_
=
]T|

k=k
0
E
__
_
M
(k1)+s
M
(k1)+s
_
_
2
_
.
25
By Lemma 6 we have E
__
_
S
]T|
_
_
2
_
C
,
( )T and (20) follows. It
remains to prove that
_
_
(1
k
0

T
)( )(Id K)
1
E
_
D
k
0
,,
(N)
T,
__
_

,
(T)
where

,
(T) 0 as T . By Lemma 4, we have
E
_
N
(k1)+
N
(k1)+
_
= ( )
_
+
_
(k1)+
0
dr(r)
_
+
_
(k1)+
(k1)+
dr(r)((k 1) +r)
= ( )(Id K)
1
( )
_

(k1)+
dr(r)
+
_
(k1)+
(k1)+
dr(r)
_
(k 1) +r
_
.
Finally
_
(k1)+
(k1)+
dr(r)((k 1) +r) ( )
_

(k1)+
dr(r).
We conclude noting that

T
]T/|

k=k
0
_

(k1)+
dr(r) =

T
_

(k
0
1)+
dr(r)
]T/|

k=k
0
1
(k1)+<r

1
T
_

0
dr (r)(r T),
This last quantity converges to 0 as T by an argument similar to the
end of proof of Lemma, 5, using that is integrable.
Lemma 9. There exists a constant C
,
such that for all t, h 0, we have
E
_
sup
tst+h
M
s
M
t

4
_
C
,
(h +h
2
)
Proof. According to the Burkholder-Davis-Gundy inequality, we have
E
_
sup
tst+h
M
s
M
t

4
| F
t
_
C
d

i=1
E
__
[M
i
, M
i
]
t+h
[M
i
, M
i
]
t
_
2
| F
t
_
= C
d

i=1
E
__
N
i,t+h
N
i,t
_
2
| F
t
_
,
26
hence
E
_
sup
tst+h
M
s
M
t

4
_
C
_
E
__
_
X
t+h
X
t
_
_
2
_
+
_
_
E
_
N
t+h
N
t
__
_
2
_
.
By Lemma 4, we have
E
_
N
t+h
N
t
_
h(Id K)
1
componentwise
and
_
E
__
_
X
t+h
X
t
_
_
2
_
_
1/2

_

0
(s)
_
E
__
_
M
t+hs
M
t
_
_
2
_
_
1/2
ds.
The conclusion follows using Lemma 6 and the fact that is integrable.
With the notation introduced in Section 3, the quantity V
,T
(M
s
, M
t+
)
is equal to
1
T
]T/|

k=1
_
M
ks
M
(k1)s
__
M
kt+
M
(k1)t+
_

.
Lemma 10. For all s, t 0 we have
V

T
,T
(M
s
, M
t+
T
)
_
1
[ts
T
[

T
_
+
0 as T in L
2
(P)
and
E
__
_
V

T
,T
(M
s
, M
t+
T
)
_
_
2
_
C
,
.
Proof. We start with proving preliminary estimates. Let b
1
, b
2
, b
t
1
, b
t
2
be real
numbers such that b
1
b
2
, b
t
1
b
t
2
and (b
1
, b
2
) (b
t
1
, b
t
2
) = . Using that M
is a martingale, we successively obtain
E
__
_
T
1
]T/|

k=1
_
M
(k1)+b
2
M
(k1)+b
1
__
M
(k1)+b

2
M
(k1)+b

1
_

_
_
2
_
=
1
T
2
]T/|

k=1
E
__
_
_
M
(k1)+b
2
M
(k1)+b
1
__
M
(k1)+b

2
M
(k1)+b

1
_

_
_
2
_
=
1
T
2
]T/|

k=1
E
__
_
M
(k1)+b
2
M
(k1)+b
1
_
_
2
_
_
_
M
(k1)+b

2
M
(k1)+b

1
_

_
_
2
_
.
27
By Cauchy-Schwarz, this last quantity is less than
1
T
2
]T/|

k=1
_
E(M
(k1)+b
2
M
(k1)+b
1

4
)E(
_
_
_
M
(k1)+b

2
M
(k1)+b

1
_

_
_
4
)
_1
2
which in turn, using Lemma 9, is bounded by
C
,
1
T
_
b
2
b
1
+ (b
2
b
1
)
2
_
1/2
_
b
t
2
b
t
1
+ (b
t
2
b
t
1
)
2
_
1/2
. (21)
Moreover, if b
2
b
1
, using that [M, M] = diag(N) and Lemma 9, we
obtain that
E
__
_
_
1
T
]T/|

k=1
diag
_
N
(k1)+b
2
N
(k1)
T
+b
1
_

1
T
]T/|

k=1
_
M
(k1)+b
2
M
(k1)+b
1
__
M
(k1)+b
2
M
(k1)+b
1
_

_
_
_
2
_
is less than
C
,
1
T
_
b
2
b
1
+ (b
2
b
1
)
2
_
. (22)
We are ready to prove Lemma 10. It is a consequence of Lemma 8 and the
fact that there exists a
2
= a
2
(s, t,
T
,
T
) a
3
= a
3
(s, t,
T
,
T
) such that
a
3
a
2
=
_

T
|t s
T
|
_
+
,
a
2
T
0 as T
holds, together with the estimate
E
__
_
1
T
]T/
T
|

k=1
diag
_
N
(k1)
T
+a
3
N
(k1)
T
+a
2
_
V

T
,T
(M
s
, M
t+
T
)
_
_
2
_
C
,
1
T
_
1 +
T
_
.
(23)
Indeed, if t
T
+
T
s or s t
T

T
then the upper bound we
obtained in (21) entails
E
__
_
V

T
,T
(M
s
, M
t+
T
)
_
_
2
_
C
,
1
T
_
1 +
T
_
.
Let us rst consider the case t
T
s t
T
+
T
. Set
a
1
:= s a
2
:= t +
T
a
3
:=
T
s a
4
:=
T
t +
T
.
28
We use the following decomposition
_
M
k
T
s
M
(k1)
T
s
__
M
k
T
t+
T
M
(k1)
T
t+
T
_

=
_
M
(k1)
T
+a
3
M
(k1)
T
+a
2
__
M
(k1)
T
+a
3
M
(k1)
T
+a
2
_

+
_
M
(k1)
T
+a
3
M
(k1)
T
+a
1
__
M
(k1)
T
+a
4
M
(k1)
T
+a
3
_

+
_
M
(k1)
T
+a
2
M
(k1)
T
+a
1
__
M
(k1)
T
+a
3
M
(k1)
T
+a
2
_

.
On the one hand, (21) readily yields that both
E
__
_
1
T
]T/
T
|

k=1
_
M
(k1)
T
+a
3
M
(k1)
T
+a
1
__
M
(k1)
T
+a
4
M
(k1)
T
+a
3
_

_
_
2
_
and
E
__
_
1
T
]T/
T
|

k=1
_
M
(k1)
T
+a
2
M
(k1)
T
+a
1
__
M
(k1)
T
+a
3
M
(k1)
T
+a
2
_

_
_
2
_
are less than C
,
(1+
T
)/T. On the other hand, by (22), the same estimate
holds for
E
__
_
1
T
]T/
T
|

k=1
diag
_
N
(k1)
T
+a
3
N
(k1)
T
+a
2
_

1
T
]T/
T
|

k=1
_
M
(k1)
T
+a
3
M
(k1)
T
+a
2
__
M
(k1)
T
+a
3
M
(k1)
T
+a
2
_

_
_
2
_
,
therefore (23) holds in that case. If now t
T

T
s t
T
, setting
a
1
:= t +
T
a
2
:= s a
3
:=
T
t +
T
a
4
:=
T
s,
one readily checks that (23) holds using the same arguments.
Completion of proof of Theorem 3. Since X = M +Y by (19), we have the
following decomposition
V
,T
(X, X
+
)
= V
,T
(M, M
+
) +V
,T
(Y, Y
+
) +V
,T
(M, Y
+
) +V
,T
(Y, M
+
).
29
Setting M
t
= 0 for t 0, we can write Y
t
=
_

0
(s)M
ts
ds, therefore
(Y
i
Y
(i1)
)(Y
i+
Y
(i1)+
)

=
_
R
2
+
ds dt (s)
_
M
is
M
(i1)s
__
M
it+
M
(i1)t+
_

(t)

,
hence
V
,T
(Y, Y
+
) =
_
R
2
+
ds dt (s)V
,T
(M
s
, M
t+
)(t)

.
Likewise
V
,T
(M, Y
+
) =
_

0
dt V
,T
(M, M
t+
)(t)

,
V
,T
(Y, M
+
) =
_

0
dt (t)V
,T
(M
t
, M
+
).
In view of Lemma 10 and the fact that is integrable, by Lebesgue domi-
nated convergence theorem, we successively obtain
V

T
,T
(Y, Y

T
+
)
_
R
2
+
ds dt (s)
_
1 |t s
T
|/
T
_
+
(t)

0,
V

T
,T
(M, Y

T
+
)
_

0
dt
_
1 |t
T
|/
T
_
+
(t)

0,
and
V

T
,T
(Y, M

T
+
)
_

0
dt (t)
_
1 |t +
T
|/
T
_
+
0
in L
2
(P) as T . The proof is complete.
Proof of Corollary 2. In view of Theorem 3, we need to show
V

T
,T
(X, X

T
+
) V

T
,T
(

X,

X

T
+
) 0 as T in L
2
(P).
By Cauchy-Schwarz inequality, this convergence is a consequence of
sup
R
1
T
]T/
T
|

k=1
_

i
( +k
T
)
i
( + (k 1)
T
)
_
2
0, 1 i d,
where (t) =
_
t(Id K)
1
E(N
t
)
_
1
t0
. By (15), for t 0, we have
the decomposition
(t) =
_
t
_

t
(s)ds +
_
t
0
(s)sds
_

30
therefore the function is absolutely continuous and we have

t
(t) =
_

t
(s)ds 1
t0
.
We derive
1
T
]T/
T
|

k=1
_

i
( +k
T
)
i
( + (k 1)
T
)
_
2

T
T
_

0

t
i
(s)
2
ds.
It remains is to prove
_

0

t
(t)
2
dt < .
Since s s
1/2
(s) is integrable, we have

t
(t) t
1/2
_

t
s
1/2
(s) ds and
t
(t) C
,
t
1/2
,
hence
_

0

t
(t)
2
dtC
,
_

0
dt t
1/2
_

t
ds(s)
= 2 C
,
_

0
ds(s)s
1/2
<
and the result follows.
Appendix
Proof of Proposition 2
The spectral radius of K =
_
0 0 h 0
0 0 0 h
g 0 0 0
0 g 0 0
_
is equal to g h, and we have
_
Id K
_
1
=
1
1 h g
_
1 0 h 0
0 1 0 h
g 0 1 0
0 g 0 1
_
,
therefore
_
Id K
_
1
=
1
1 h g
_

2
_
31
where
1
and
2
are given by (9). Set X = N E(N). By symmetry,
we have N = (N
2
, N
1
, N
4
, N
3
) in distribution, thus E(N
1,t
) = E(N
2,t
) and
E(N
3,t
) = E(N
4,t
) for all t. Consequently
S
1
= X
1
X
2
and S
2
= X
3
X
4
. (24)
According to Theorem 2 the processes T
1/2
X
Tv
converge in law to the
process Y
v
= (Id K)
1

1/2
W
v
with
=
1
1 h g
_

1
0 0 0
0
1
0 0
0 0
2
0
0 0 0
2
_
.
Therefore the processes T
1/2
_
S
1,Tv
, S
2,Tv
_
v[0,1]
converge in distribution to
_
Y
1
Y
2
, Y
3
Y
4
_
and Proposition 2 is proved.
Proof of Proposition 3
From (24) it follows that
V

T
,T
(S
1
, S
1,
T
+
) = ( 1 1 0 0 ) V

T
,T
(X, X

T
+
)
_
1
1
0
0
_
and
V

T
,T
(S
1
, S
2,
T
+
) = ( 1 1 0 0 ) V

T
,T
(X, X

T
+
)
_
0
0
1
1
_
.
Consequently Proposition 3 follows from Theorem 3 with
C
11
(, ) = ( 1 1 0 0 ) v
,
_
1
1
0
0
_
, C
12
(, ) = ( 1 1 0 0 ) v
,
_
0
0
1
1
_
.
It remains to compute C
11
and C
12
. First we compute =

n1

n
. We
readily check that for all s, t 0: (t s)(s) = h(t s)g(s)Id. Thus

2
= (h g)Id. We derive

2n
= (h g)
n
Id,
2n+1
=
_
_
0 0 (hg)
n
h 0
0 0 0 (hg)
n
h
(hg)
n
g 0 0 0
0 (hg)
n
g 0 0
_
_
.
and we obtain
=

n1

2n
+

n0

2n+1
=
_
_
_
_
F 0 (
0
+F) h 0
0 F 0 (
0
+F) h
(
0
+F) g 0 F 0
0 (
0
+F) g 0 F
_
_
_
_
32
where F =

n1
(h g)
n
. Set

F(ds) =
0
(ds) + F(t)ds. Standard compu-
tations yield
Id
0
(ds) +(s)ds =
_
_
_
_
_

F(ds) 0

F h(s)ds 0
0

F(ds) 0

F h(s)ds

F g(s)ds 0

F(ds) 0
0

F g(s)ds 0

F(ds)
_
_
_
_
_
and
_
Id
0
(ds) +(s)ds
_

_
Id
0
(dt) +(t)

dt
_
=
1
1 h g
_
_
_
_
a
11
(ds, dt) 0 a
13
(ds, dt) 0
0 a
11
(ds, dt) 0 a
13
(ds, dt)
a
31
(ds, dt) 0 a
33
(ds, dt) 0
0 a
31
(ds, dt) 0 a
33
(ds, dt)
_
_
_
_
with:
a
11
(ds, dt) =
1

F(ds)

F(dt) +
2

F h(ds)

F h(dt),
a
13
(ds, dt) =
1

F(ds)

F g(dt) +
2

F h(ds)

F(dt),
a
31
(ds, dt) =
2

F(ds)

F h(dt) +
1

F g(ds)

F(dt),
a
33
(ds, dt) =
2

F(ds)

F(dt) +
1

F g(ds)

F g(dt).
Therefore
C
11
(, ) =
_
[0,)
2

(t s )2a
11
(ds, dt),
and
C
12
(, ) =
_
[0,)
2

(t s )2a
31
(ds, dt).
To complete the proof of Proposition 3, it suces to use that for two nite
measures and on R one has (for all > 0, R)
_
R
2
+

(t s )(ds)(dt) =

()
where is the image of by x x.
Acknowledgement
The research of E. Bacry and J.F. Muzy is supported in part by the Chair
Financial Risks of the Risk Foundation. The research of M. Homann is
supported in part by the Agence Nationale de la Recherche, Grant No. ANR-
08-BLAN-0220-01.
33
References
[1] F. Abergel and N. Huth. High frequency lead/lag relationship-empirical
facts. Arxiv preprint. arXiv:1111.7103v1, 2011.
[2] F. Abergel and F. Pomponio. Trade-throughs: Empirical facts and ap-
plication to lead-lag measures. In Econophysics of order-driven markets,
F. Abregel, B.K. Chakrabarti, A. Chakrabart and M. Mitra, editors,
New Economic Windows. Springer Berlin Heidelberg, 2011.
[3] K. Al Dayri E. Bacry and J.F. Muzy. Non-parametric kernel estimation
for symmetric Hawkes processes. application to high frequency nancial
data. http://arxiv.org/abs/1112.1838.
[4] M.S. Bartlett. The spectral analysis of point processes. Journal of the
Royal Statistical Society. Series B., 25:264296, 1963.
[5] L. Bauwens and N. Hautsch. Modelling nancial high frequency data
using point processes. In T. Mikosch, J-P. Kreiss, R. A. Davis, and
T. G. Andersen, editors, Handbook of Financial Time Series. Springer
Berlin Heidelberg, 2009.
[6] P. Bremaud and L. Massoulie. Power spectra of general shot noises and
hawkes point processes with a random excitation. Advances in Applied
Probability, 34:205222, 2002.
[7] Y. Ait-Sahalia J. Cacho-Diaz and R. Laeven. Modeling nancial con-
tagion using mutually exciting jump processes. Working paper, 2011.
[8] D.J. Daley and D. Vere-Jones. An introduction to the theory of
point processes. Vol. I. Probability and its Applications (New York).
Springer-Verlag, New York, second edition, 2003. Elementary theory
and methods.
[9] A. Van der Vaart. Asymptotic Statistics. Cambridge University Press,
1998.
[10] T. Andersen T. Bollerslev F.X. Diebold and P. Labys. (understanding,
optimizing, using and forecasting) realized volatility and correlation.
Great Realizations, Risk,, pages 105108, 2000.
[11] T. W. Epps. Comovements in stock prices in the very short run. Journal
of the American Statistical Association, 74:291298, 1979.
34
[12] A.G. Hawkes. Point spectra of some mutually exciting point processes.
Journal of the Royal Statistical Society, Series B, 33:438443, 1971.
[13] A.G. Hawkes. Spectra of some self-exciting and mutually exciting point
processes. Biometrika, 58:8390, 1971.
[14] A.G. Hawkes. Spectra of some mutually exciting point processes with
associated variables. In Stochastic Point Processes. ed. P. A. W. Lewis.
Wiley, New York. 1972.
[15] A.G. Hawkes and D. Oakes. A cluster process representation of a self-
exciting process. Journal of Applied Probability, 11:493503, 1974.
[16] P. Hewlett. Clustering of order arrivals, price impact and trade path
optimisation. Workshop on Financial Modeling with Jump processes,
Ecole Polytechnique, 2006.
[17] E. Bacry S. Delattre M. Homann and J.F. Muzy. Modelling mi-
crostructure noise with mutually exciting point processes. Quantitative
Finance, to appear, 2012.
[18] J. Jacod. Multivariate point processes: predictable projection, Radon-
Nikod ym derivatives, representation of martingales. Z. Wahrschein-
lichkeitstheorie und Verw. Gebiete, 31:235253, 1974/75.
[19] J. Jacod and A.N. Shiryaev. Limit theorems for stochastic processes,
volume 288 of Grundlehren der Mathematischen Wissenschaften [Fun-
damental Principles of Mathematical Sciences]. Springer-Verlag, Berlin,
1987.
[20] P. Embrechts J. T. Liniger and L. Lu. Multivariate Hawkes processes:
an application to nancial data. Journal of Applied Probability, 48:367
378, 2011.
[21] O. Barndor-Nielsen P. Hansen A. Lunde and N. Stephard. Designing
realised kernels to measure the ex-post variation of equity prices in the
presence of noise. Econometrica, 76(6):14811536, 2008.
[22] J. Moller and J.K. Rasmussen. Perfect simulation of hawkes processes.
Technical report, 2004.
[23] Y. Ait-Sahalia P.A. Mykland and L. Zhang. How often to sample a
continuous-time process in the presence of market microstructure noise.
The Review of Financial Studies, 18:351416, 2005.
35
[24] Y. Ait-Sahalia P.A. Mykland and L. Zhang. Ultra high frequency
volatility estimation with dependent microstructure noise. Journal of
Econometrics, 160:160175, 2011.
[25] Y. Ogata. The asymptotic behaviour of maximum likelihood estimators
for stationary point processes. Annals of the Institute of Statistical
Mathematics, 30:243261, 1978.
[26] Y. Ogata. On Lewis simulation method for point processes. IEEE
Information Theory, 27:2331, 1981.
[27] P. Reynaud-Bouret and S. Schbath. Adaptive estimation for Hawkes
processes; application to genome analysis. Ann. Statist., 38:27812822,
2010.
[28] C.Y. Robert and M. Rosenbaum. A new approach for the dynamics of
ultra high frequency data: the model with uncertainty zones. Journal
of Financial Econmetrics, 2009. In press.
[29] M. Rosenbaum. A new microstructure noise index. Quantitative Fi-
nance, 6:883899, 2011.
[30] M. Homann M. Rosenbaum and N. Yoshida. Estimation of the lead-lag
eect from nonsynchronous data. Bernoulli, to appear, 2011.
[31] J.Zhuang Y.Ogata and D.Vere-Jones. Stochastic declustering of space-
time earthquake occurrences. Journal of the American Statistical As-
sociation, 97:369380, 2002.
36