getdoce91c. 275KB Jun 04 2011 12:05:08 AM

J
ni
o
r
Elect

c

o

u

a
rn l

o

f
P

r

ob
abil
ity

Vol. 6 (2001) Paper No. 25, pages 1–33.
Journal URL
http://www.math.washington.edu/~ejpecp/
Paper URL
http://www.math.washington.edu/~ejpecp/EjpVol6/paper25.abs.html
SUPERPROCESSES WITH DEPENDENT SPATIAL MOTION
AND GENERAL BRANCHING DENSITIES
Donald A. Dawson
Zenghu Li
School of Mathematics and Statistics,
Department of Mathematics,
Carleton University,
Beijing Normal University,
1125 Colonel By Drive, Ottawa, Canada K1S 5B6 Beijing 100875, P.R. China
ddawson@math.carleton.ca

lizh@bnu.edu.cn
Hao Wang
Department of Mathematics, University of Oregon,
Eugene OR 97403-1222, U.S.A.
haowang@darkwing.uoregon.edu
Abstract We construct a class of superprocesses by taking the high density limit of a
sequence of interacting-branching particle systems. The spatial motion of the superprocess
is determined by a system of interacting diffusions, the branching density is given by an
arbitrary bounded non-negative Borel function, and the superprocess is characterized by a
martingale problem as a diffusion process with state space M(R ), improving and extending
considerably the construction of Wang (1997, 1998). It is then proved in a special case
that a suitable rescaled process of the superprocess converges to the usual super Brownian
motion. An extension to measure-valued branching catalysts is also discussed.
Keywords superprocess, interacting-branching particle system, diffusion process, martingale problem, dual process, rescaled limit, measure-valued catalyst
AMS Subject Classifications Primary 60J80, 60G57; Secondary 60J35.
Research supported by Supported by NSERC operating grant (D. D.), NNSF grant
19361060 (Z. L.), and the research grant of UO (H. W.).
Submitted to EJP on January 3, 2001. Accepted May 25, 2001.

1

Introduction

For a given topological space E, let B(E) denote the totality of all bounded Borel functions
on E and let C(E) denote its subset comprising of continuous functions. Let M(E) denote
the space of finite
R Borel measures on E endowed with the topology of weak convergence.
Write hf, µi for f dµ. For F ∈ B(M(E)) let
δF (µ)
1
= lim+ [F (µ + rδx ) − F (µ)],
r→0 r
δµ(x)

x ∈ E,

(1.1)

if the limit exists. Let δ 2 F (µ)/δµ(x)δµ(y) be defined in the same way with F replaced
by (δF/δµ(y)) on the right hand side. For example, if Fm,f (µ) = hf, µm i for f ∈ B(E m )

and µ ∈ M(E), then
m

δFm,f (µ) X
=
hΨ i (x)f, µm−1 i,
δµ(x)
i=1

x ∈ E,

(1.2)

where Ψi (x) is the operator from B(E m ) to B(E m−1 ) defined by
Ψi (x)f (x1 , · · · , xm−1 ) = f (x1 , · · · , xi−1 , x, xi , · · · , xm−1 ),

xj ∈ E,

(1.3)

where x ∈ E is the ith variable of f on the right hand side.

Now we consider the case where E = R , the one-dimensional Euclidean space. Suppose
that c ∈ C(R ) is Lipschitz and h ∈ C(R ) is square-integrable. Let
Z
ρ(x) =
h(y − x)h(y)dy,
(1.4)
R
and a(x) = c(x)2 + ρ(0) for x ∈ R . We assume in addition that ρ is twice continuously differentiable with ρ′ and ρ′′ bounded, which is satisfied if h is integrable and twice
continuously differentiable with h′ and h′′ bounded. Then
Z
1
d2 δF (µ)
AF (µ) =
µ(dx)
a(x) 2
2 R
dx δµ(x)
Z

1
d2
δ 2 F (µ)
+
ρ(x − y)
µ(dx)µ(dy)
(1.5)
2 R2
dxdy δµ(x)δµ(y)
defines an operator A which acts on a subset of B(M(R )) and generates a diffusion process
with state space M(R ). Suppose that {W (x, t) : x ∈ R , t ≥ 0} is a Brownian sheet and
{Bi (t) : t ≥ 0}, i = 1, 2, · · ·, is a family of independent standard Brownian motions which
are independent of {W (x, t) : x ∈ R , t ≥ 0}. By Lemma 3.1, for any initial conditions
xi (0) = xi , the stochastic equations
Z
dxi (t) = c(xi (t))dBi (t) + h(y − xi (t))W (dy, dt),
t ≥ 0, i = 1, 2, · · · ,
(1.6)
R
2

have unique solutions {xi (t) : t ≥ 0} and, for each integer m ≥ 1, {(x1 (t), · · · , xm (t)) : t ≥
0} is an m-dimensional diffusion process which is generated by the differential operator
m
m
1X
∂2 1 X
∂2
G :=
a(xi ) 2 +
ρ(xi − xj )
.
2 i=1
∂xi 2 i,j=1,i6=j
∂xi ∂xj
m

(1.7)

In particular, {xi (t) : t ≥ 0} is a one-dimensional diffusion process with generator G :=

(a(x)/2)∆. Because of the exchangebility, a diffusion process generated by Gm can be
regarded as an interacting particle system or a measure-valued process. Heuristically,
a(·) represents the speed of the particles and ρ(·) describes the interaction between them.
The diffusion process generated by A arises as the high density limit of a sequence of
interacting particle systems described by (1.6); see Wang (1997, 1998) and section 4 of
this paper. For σ ∈ B(R )+ , we may also define the operator B by
Z
δ 2 F (µ)
1
µ(dx).
(1.8)
σ(x)
BF (µ) =
2 R
δµ(x)2
A Markov process generated by L := A + B is naturally called a superprocess with dependent spatial motion (SDSM) with parameters (a, ρ, σ), where σ represents the branching
density of the process. In the special case where both c and σ are constants, the SDSM
ˆ = R ∪ {∂}
was constructed in Wang (1997, 1998) as a diffusion process in M(Rˆ ), where R
is the one-point compactification of R . It was also assumed in Wang (1997, 1998) that h

is a symmetric function and that the initial state of the SDSM has compact support in R .
Stochastic partial differential equations and local times associated with the SDSM were
studied in Dawson et al (2000a, b).
The SDSM contains as special cases several models arising in different circumstances such
as the one-dimensional super Brownian motion, the molecular diffusion with turbulent
transport and some interacting diffusion systems of McKean-Vlasov type; see e.g. Chow
(1976), Dawson (1994), Dawson and Vaillancourt (1995) and Kotelenez (1992, 1995). It
is thus of interest to construct the SDSM under reasonably more general conditions and
formulate it as a diffusion processes in M(R ). This is the main purpose of the present
paper. The rest of this paragraph describes the main results of the paper and gives
some unsolved problems in the subject. In section 2, we define some function-valued
dual process and investigate its connection to the solution of the martingale problem of
a SDSM. Duality method plays an important role in the investigation. Although the
SDSM could arise as high density limit of a sequence of interacting-branching particle
systems with location-dependent killing density σ and binary branching distribution, the
construction of such systems seems rather sophisticated and is thus avoided in this work.
In section 3, we construct the interacting-branching particle system with uniform killing
density and location-dependent branching distribution, which is comparatively easier to
treat. The arguments are similar to those in Wang (1998). The high density limit of the
interacting-branching particle system is considered in section 4, which gives a solution

of the martingale problem of the SDSM in the special case where σ ∈ C(R )+ can be
3

extended into a continuous function on Rˆ . In section 5, we use the dual process to extend
the construction of the SDSM to a general bounded Borel branching density σ ∈ B(R )+ .
In both sections 4 and 5, we use martingale arguments to show that, if the processes
are initially supported by R , they always stay in M(R ), which are new results even in
the special case considered in Wang (1997, 1998). In section 6, we prove a rescaled limit
theorem of the SDSM, which states that a suitable rescaled SDSM converges to the usual
super Brownian motion if c(·) is bounded away from zero. This describes another situation
where the super Brownian motion arises universally; see also Durrett and Perkins (1998)
and Hara and Slade (2000a, b). When c(·) ≡ 0, we expect that the same rescaled limit
would lead to a measure-valued diffusion process which is the high density limit of a
sequence of coalescing-branching particle systems, but there is still a long way to reach
a rigorous proof. It suffices to mention that not only the characterization of those high
density limits but also that of the coalescing-branching particle systems themselves are
still open problems. We refer the reader to Evans and Pitman (1998) and the references
therein for some recent work on related models. In section 7, we consider an extension of
the construction of the SDSM to the case where σ is of the form σ = η˙ with η belonging
to a large class of Radon measures on R , in the lines of Dawson and Fleischmann (1991,

1992). The process is constructed only when c(·) is bounded away from zero and it
can be called a SDSM with measure-valued catalysts. The transition semigroup of the
SDSM with measure-valued catalysts is constructed and characterized using a measurevalued dual process. The derivation is based on some estimates of moments of the dual
process. However, the existence of a diffusion realization of the SDSM with measurevalued catalysts is left as another open problem in the subject.
Notation: Recall that Rˆ = R ∪ {∂} denotes the one-point compactification of R . Let
λm denote the Lebesgue measure on R m . Let C 2 (R m ) be the set of twice continuously
differentiable functions on R m and let C∂2 (R m ) be the set of functions in C 2 (R m ) which
ˆ.
together with their derivatives up to the second order can be extended continuously to R
2
m
2
m
Let C0 (R ) be the subset of C∂ (R ) of functions that together with their derivatives up
to the second order vanish rapidly at infinity. Let (Ttm )t≥0 denote the transition semigroup
of the m-dimensional standard Brownian motion and let (Ptm )t≥0 denote the transition
semigroup generated by the operator Gm . We shall omit the superscript m when it is
ˆ denote the extensions of (Pt )t≥0 and G to Rˆ with ∂ as a trap.
one. Let (Pˆt )t≥0 and G
We denote the expectation by the letter of the probability measure if this is specified and
simply by E if the measure is not specified.
We remark that, if |c(x)| ≥ ǫ > 0 for all x ∈
pm
t (x, y) which satisfies
m
pm
t (x, y) ≤ const · gǫt (x, y),

R , the semigroup (Ptm)t>0 has density
t > 0, x, y ∈ R m ,

(1.9)

where gtm (x, y) denotes the transition density of the m-dimensional standard Brownian
motion; see e.g. Friedman (1964, p.24).

4

2

Function-valued dual processes

In this section, we define a function-valued dual process and investigate its connection
to the solution of the martingale problem for the SDSM. Recall the definition of the
generator L := A + B given by (1.5) and (1.8) with σ ∈ B(R )+ . For µ ∈ M(R ) and a
subset D(L) of the domain of L, we say an M(R )-valued c´adl´ag process {Xt : t ≥ 0} is a
solution of the (L, D(L), µ)-martingale problem if X0 = µ and
Z t
F (Xt ) − F (X0 ) −
LF (Xs )ds,
t ≥ 0,
0

is a martingale for each F ∈ D(L). Observe that, if Fm,f (µ) = hf, µm i for f ∈ C 2 (R m ),
then
Z X
m
1
AFm,f (µ) =
a(xi )fii′′ (x1 , · · · , xm )µm (dx1 , · · · , dxm )
2 Rm i=1
Z
m
X
1
ρ(xi − xj )fij′′ (x1 , · · · , xm )µm (dx1 , · · · , dxm )
+
2 Rm
i,j=1,i6=j
= Fm,Gm f (µ),

(2.1)

and
Z
m
1 X
BFm,f (µ) =
Φij f (x1 , · · · , xm−1 )µm−1 (dx1 , · · · , dxm−1 )
2 i,j=1,i6=j Rm−1
1
=
2

m
X

Fm−1,Φij f (µ),

(2.2)

i,j=1,i6=j

where Φij denotes the operator from B(E m ) to B(E m−1 ) defined by
Φij f (x1 , · · · , xm−1 ) = σ(xm−1 )f (x1 , · · · , xm−1 , · · · , xm−1 , · · · , xm−2 ),

(2.3)

where xm−1 is in the places of the ith and the jth variables of f on the right hand side.
It follows that
LFm,f (µ) = F

m,Gm f

m
1 X
Fm−1,Φij f (µ).
(µ) +
2 i,j=1,i6=j

(2.4)

Let {Mt : t ≥ 0} be a nonnegative integer-valued c´adl´ag Markov process with transition
intensities {qi,j } such that qi,i−1 = −qi,i = i(i − 1)/2 and qi,j = 0 for all other pairs (i, j).
That is, {Mt : t ≥ 0} is the well-known Kingman’s coalescent process. Let τ0 = 0 and
τM0 = ∞, and let {τk : 1 ≤ k ≤ M0 − 1} be the sequence of jump times of {Mt : t ≥ 0}.
5

Let {Γk : 1 ≤ k ≤ M0 − 1} be a sequence of random operators which are conditionally
independent given {Mt : t ≥ 0} and satisfy
P {Γk = Φi,j |M(τk− ) = l} =

1
,
l(l − 1)

1 ≤ i 6= j ≤ l,

(2.5)

where Φi,j is defined by (2.3). Let B denote the topological union of {B(R m ) : m =
1, 2, · · ·} endowed with pointwise convergence on each B(R m ). Then
Mτ

Mτ

M

τ1
Yt = Pt−τkk Γk Pτk −τk−1
Γ
· · · Pτ2 −τ
Γ P M0 Y 0 ,
1 1 τ1
k−1 k−1

τk ≤ t < τk+1 , 0 ≤ k ≤ M0 − 1,

(2.6)

defines a Markov process {Yt : t ≥ 0} taking values from B. Clearly, {(Mt , Yt ) : t ≥ 0} is
also a Markov process. To simplify the presentation, we shall suppress the dependence of
{Yt : t ≥ 0} on σ and let E σm,f denote the expectation given M0 = m and Y0 = f ∈ C(R m ),
just as we are working with a canonical realization of {(Mt , Yt ) : t ≥ 0}. By (2.6) we have
h
n1 Z t
oi
σ
Mt
E m,f hYt , µ i exp
Ms (Ms − 1)ds
2 0
m
m
(2.7)
= hPt f, µ i
Z
Z
m
h
n 1 t−u
oi
t
1 X
+
E σm−1,Φij Pum f hYt−u , µMt−u i exp
Ms (Ms − 1)ds du.
2 i,j=1,i6=j 0
2 0
Lemma 2.1 For any f ∈ B(R m ) and any integer m ≥ 1,
h
n1 Z t
oi
σ
Mt
E m,f hYt , µ i exp
Ms (Ms − 1)ds
2 0
m−1
X
≤ kf k
2−k mk (m − 1)k kσkk h1, µim−k ,
k=0

where k · k denotes the supremum norm.
Pm−1
Proof. The left hand side of (2.8) can be decomposed as k=0
Ak with
h
i
n1 Z t
o
σ
Mt
Ak = E m,f hYt , µ i exp
Ms (Ms − 1)ds 1{τk ≤t 0 and i ≥ 1 and define {xki (t) : t ≥ 0} inductively by x0i (t) ≡ xi (0) and
Z t
Z tZ
k+1
k
xi (t) = xi (0) +
h(y − xki (s))W (dy, ds),
t ≥ 0.
c(xi (s))dBi (s) +
0
0
R
11

Let l(c) ≥ 0 be any Lipschitz constant for c(·). By a martingale inequality we have
Z T
n
o
k+1
k
2
E sup |xi (t) − xi (t)|
≤ 8
E{|c(xki (t)) − c(xik−1 (t))|2 }dt
0≤t≤T

0

+8

T

Z

E

0

≤ 8l(c)2
Z
+16

Z

nZ

T

R

o
|h(y − xki (t)) − h(y − xik−1 (t))|2 dy dt

E{|xki (t) − xik−1 (t)|2 }dt

0
T

E{|ρ(0) − ρ(xki (t) − xik−1 (t))|}dt
0
Z T
2
′′
≤ 8(l(c) + kρ k)
E{|xki (t) − xik−1 (t)|2 }dt.
0

Using the above inequality inductively we get
n
o
k
2
E sup |xk+1
(t)
−
x
(t)|
≤ (kck2 + ρ(0))(l(c)2 + kρ′′ k)k (8T )k /k!,
i
i
0≤t≤T

and hence
P

n

sup |xk+1
(t) − xki (t)| > 2−k
i

0≤t≤T

o

≤ const · (l(c)2 + kρ′′ k)k (8T )k /k!.

By Borel-Cantelli’s lemma, {xki (t) : 0 ≤ t ≤ T } converges in the uniform norm with
probability one. Since T > 0 was arbitrary, xi (t) = limk→∞ xki (t) defines a continuous
martingale {xi (t) : t ≥ 0} which is clearly the unique solution of (1.6). It is easy to
see that dhxi i(t) = a(xi (t))dt and dhxi , xj i(t) = ρ(xi (t) − xj (t))dt for i 6= j. Then
{(x1 (t), · · · , xm (t)) : t ≥ 0} is a diffusion process with generator Gm defined by (1.7). 
Because of the exchangebility, the Gm -diffusion can be regarded as a measure-valued
Markov process. Let N(R ) denote the space of integer-valued measures on R . For θ > 0,
m
let Mθ (R ) = {θ−1 σ : σ ∈ N(R )}. Let ζ be the mapping from ∪∞
to Mθ (R ) defined
m=1 R
by
m

1X
ζ(x1 , · · · , xm ) =
δx ,
θ i=1 i

m ≥ 1.

(3.1)

Lemma 3.2 For any integers m, n ≥ 1 and any f ∈ C 2 (R n ), we have
Gm Fn,f (ζ(x1 , · · · , xm )) =
+

n
m
1 X X
′′
(xl1 , · · · , xln )
a(xlα )fαα
2θn α=1 l ,···,l =1
1

1
2θn

1
+
2θn

n

n
X

m
X

α,β=1,α6=β l1 ,···,ln =1,lα =lβ
m
X

n
X

α,β=1,α6=β l1 ,···,ln =1

12

′′
(xl1 , · · · , xln )
c(xlα )c(xlβ )fαβ

′′
(xl1 , · · · , xln ).
ρ(xlα − xlβ )fαβ

(3.2)

Proof. By (3.1), we have
m
X

1
Fn,f (ζ(x1 , · · · , xm )) = n
θ

l1 ,···,ln =1

f (xl1 , · · · , xln ).

(3.3)

Observe that, for 1 ≤ i ≤ m,

n
d2
1 X X ′′
Fn,f (ζ(x1, · · · , xm )) = n
fαβ (xl1 , · · · , xln ),
dx2i
θ α,β=1
{···}

where {· · ·} = { for all 1 ≤ l1 , · · · , ln ≤ m with lα = lβ = i}. Then it is not hard to see
that
m
X
d2
c(xi )2 2 Fn,f (ζ(x1 , · · · , xm ))
dxi
i=1
m
X

n
1 X
= n
θ α,β=1 l

1 ,···,ln =1,lα =lβ

1
= n
θ

n
X

m
X

′′
(xl1 , · · · , xln )
c(xlα )2 fαα

α=1 l1 ,···,ln =1
n
X

1
+ n
θ

′′
(xl1 , · · · , xln )
c(xlα )c(xlβ )fαβ

m
X

α,β=1,α6=β l1 ,···,ln =1,lα =lβ

′′
(xl1 , · · · , xln ).
c(xlα )c(xlβ )fαβ

On the other hand, for 1 ≤ i 6= j ≤ m,
 d2
1
d2 
Fn,f (ζ(x1 , · · · , xm )) = n
+
dxi dxj dxi dxj
θ

n
X

X

α,β=1,α6=β {···}

(3.4)

′′
fαβ
(xl1 , · · · , xln ),

where {· · ·} = { for all 1 ≤ l1 , · · · , ln ≤ m with lα = i and lβ = j}. It follows that
m
X

ρ(xi − xj )

i,j=1,i6=j
n
X

1
= n
θ

d2
Fn,f (ζ(x1, · · · , xm ))
dxi dxj

m
X

α,β=1,α6=β l1 ,···,ln =1,lα 6=lβ

′′
(xl1 , · · · , xln ).
ρ(xlα − xlβ )fαβ

Using this and (3.4) with c(xi )2 replaced by ρ(0),
m
X

i,j=1

ρ(xi − xj )

d2
Fn,f (ζ(x1, · · · , xm ))
dxi dxj

n
m
1 X X
′′
= n
ρ(0)fαα
(xl1 , · · · , xln )
θ α=1 l ,···,l =1
1

1
+ n
θ

n

n
X

m
X

α,β=1,α6=β l1 ,···,ln =1

′′
(xl1 , · · · , xln ).
ρ(xlα − xlβ )fαβ

13

(3.5)



Then we have the desired result from (3.4) and (3.5).

Suppose that X(t) = (x1 (t), · · · , xm (t)) is a Markov process in R m generated by Gm .
Based on (1.2) and Lemma 3.2, it is easy to show that ζ(X(t)) is a Markov process in
Mθ (R ) with generator Aθ given by
Z
Z
δ 2 F (µ)
d2 δF (µ)
d2
1
1
a(x) 2
c(x)c(y)
Aθ F (µ) =
µ(dx) +
δx (dy)µ(dx)
2 R
dx δµ(x)
2θ R2
dxdy δµ(x)δµ(y)
Z
1
δ 2 F (µ)
d2
+
ρ(x − y)
µ(dx)µ(dy).
(3.6)
2 R2
dxdy δµ(x)δµ(y)
In particular, if
F (µ) = f (hφ1 , µi, · · · , hφn , µi),

µ ∈ Mθ (R ),

(3.7)

for f ∈ C 2 (R n ) and {φi } ⊂ C 2 (R ), then
n

Aθ F (µ) =

1X ′
f (hφ1 , µi, · · · , hφn , µi)haφ′′i , µi
2 i=1 i

n
1 X ′′
+
f (hφ1 , µi, · · · , hφn , µi)hc2φ′i φ′j , µi
(3.8)
2θ i,j=1 ij
Z
n
1 X ′′
+
f (hφ1 , µi, · · · , hφn , µi)
ρ(x − y)φ′i(x)φ′j (y)µ(dx)µ(dy).
2 i,j=1 ij
2
R

Now we introduce a branching mechanism to the interacting particle system. Suppose
that for each x ∈ R we have a discrete probability distribution p(x) = {pi(x) : i = 0, 1, · · ·}
such that each pi (·) is a Borel measurable function on R . This serves as the distribution
of the offspring number produced by a particle that dies at site x ∈ R . We assume that
p1 (x) = 0,

∞
X

ipi (x) = 1,

(3.9)

i=1

and
σp (x) :=

∞
X
i=1

i2 pi (x) − 1

(3.10)

is bounded in x ∈ R . Let Γθ (µ, dν) be the probability kernel on Mθ (R ) defined by
Z

θµ(1) ∞


1 XX
F (ν)Γθ (µ, dν) =
pj (xi )F µ + (j − 1)θ−1 δxi ,
θµ(1) i=1 j=0
Mθ (R)

14

(3.11)

where µ ∈ Mθ (R ) is given by
θµ(1)
1X
µ=
δx .
θ i=1 i

For a constant γ > 0, we define the bounded operator Bθ on B(Mθ (R )) by
Z
2
[F (ν) − F (µ)]Γθ (µ, dν).
Bθ F (µ) = γθ [θ ∧ µ(1)]
Mθ (R)

(3.12)

In view of (1.6), Aθ generates a Feller Markov process on Mθ (R ), then so does Lθ := Aθ +
Bθ by Ethier-Kurtz (1986, p.37). We shall call the process generated by Lθ an interactingbranching particle system with parameters (a, ρ, γ, p) and unit mass 1/θ. Heuristically,
each particle in the system has mass 1/θ, a(·) represents the migration speed of the
particles and ρ(·) describes the interaction between them. The branching times of the
system are determined by the killing density γθ2 [θ ∧µ(1)], where the truncation “θ ∧µ(1)”
is introduced to make the branching not too fast even when the total mass is large. At
each branching time, with equal probability, one particle in the system is randomly chosen,
which is killed at its site x ∈ R and the offspring are produced at x ∈ R according to the
distribution {pi (x) : i = 0, 1, · · ·}. If F is given by (3.7), then Bθ F (µ) is equal to
n
∞
γ[θ ∧ µ(1)] X X
′′
(hφ1 , µi + ξj φ1 , · · · , hφn , µi + ξj φn )φα φβ , µi (3.13)
(j − 1)2 hpj fαβ
2µ(1) α,β=1 j=1

for some constant 0 < ξj < (j − 1)/θ. This follows from (3.11) and (3.12) by Taylor’s
expansion.

4

Continuous branching density

In this section, we shall construct a solution of the martingale problem of the SDSM
with continuous branching density by using particle system approximation. Assume that
σ ∈ C(R ) can be extended continuously to Rˆ . Let A and B be given by (1.5) and (1.8),
respectively. Observe that, if
F (µ) = f (hφ1 , µi, · · · , hφn , µi),

µ ∈ M(R ),

(4.1)

for f ∈ C 2 (R n ) and {φi } ⊂ C 2 (R ), then
n

1X ′
AF (µ) =
fi (hφ1 , µi, · · · , hφn , µi)haφ′′i , µi
(4.2)
2 i=1
Z
n
1 X ′′
+
ρ(x − y)φ′i(x)φ′j (y)µ(dx)µ(dy),
f (hφ1, µi, · · · , hφn , µi)
2 i,j=1 ij
2
R
15

and
BF (µ) =

n
1 X ′′
f (hφ1 , µi, · · · , hφn , µi)hσφiφj , µi.
2 i,j=1 ij

(4.3)
(k)

Let {θk } be any sequence such that θk → ∞ as k → ∞. Suppose that {Xt : t ≥ 0} is a
sequence of c´adl´ag interacting-branching particle systems with parameters (a, ρ, γk , p(k) ),
(k)
unit mass 1/θk and initial states X0 = µk ∈ Mθk (R ). In an obvious way, we may also
(k)
regard {Xt : t ≥ 0} as a process with state space M(Rˆ ). Let σk be defined by (3.10)
(k)
with pi replaced by pi .
(k)

Lemma 4.1 Suppose that the sequences {γk σk } and {h1, µk i} are bounded. Then {Xt
t ≥ 0} form a tight sequence in D([0, ∞), M(Rˆ )).

:

(k)

Proof. By the assumption (3.9), it is easy to show that {h1, Xt i : t ≥ 0} is a martingale.
Then we have
n
o h1, µ i
k
(k)
P suph1, Xt i > η ≤
η
t≥0
(k)

for any η > 0. That is, {Xt : t ≥ 0} satisfies the compact containment condition of
(k)
Ethier and Kurtz (1986, p.142). Let Lk denote the generator of {Xt : t ≥ 0} and let F
be given by (4.1) with f ∈ C02 (R n ) and with each φi ∈ C∂2 (R ) bounded away from zero.
Then
Z t
(k)
(k)
F (Xt ) − F (X0 ) −
Lk F (Xs(k) )ds,
t ≥ 0,
0

is a martingale and the desired tightness follows from the result of Ethier and Kurtz (1986,
p.145).

In the sequel of this section, we assume {φi } ⊂ C∂2 (R ). In this case, (4.1), (4.2) and (4.3)
ˆ (µ) and BF
ˆ (µ) be defined
can be extended to continuous functions on M(Rˆ ). Let AF
ˆ (µ) = AF
ˆ (µ) + BF
ˆ (µ),
respectively by the right hand side of (4.2) and (4.3) and let LF
all defined as continuous functions on M(Rˆ ).
ˆ be the totality of all functions of the form (4.1) with f ∈ C 2 (R n )
Lemma 4.2 Let D(L)
0
and with each φi ∈ C∂2 (R ) bounded away from zero. Suppose further that γk σk → σ
uniformly and µk → µ ∈ M(Rˆ ) as k → ∞. Then any limit point Qµ of the distributions
(k)
of {Xt : t ≥ 0} is supported by C([0, ∞), M(Rˆ )) under which
Z t
ˆ (ws )ds,
F (wt ) − F (w0) −
LF
t ≥ 0,
(4.4)
0

ˆ where {wt : t ≥ 0} denotes the coordinate process of
is a martingale for each F ∈ D(L),
ˆ
C([0, ∞), M(R )).
16

Proof. We use the notation introduced in the proof of Lemma 4.1. By passing to a
(k)
subsequence if it is necessary, we may assume that the distribution of {Xt : t ≥ 0}
on D([0, ∞), M(Rˆ )) converges to Qµ . Using Skorokhod’s representation, we may assume
(k)
that the processes {Xt : t ≥ 0} are defined on the same probability space and the
sequence converges almost surely to a c´adl´ag process {Xt : t ≥ 0} with distribution
Qµ on D([0, ∞), M(Rˆ )); see e.g. Ethier and Kurtz (1986, p.102). Let K(X) = {t ≥
0 : P {Xt = Xt− } = 1}. By Ethier and Kurtz (1986, p.118), for each t ∈ K(X) we
(k)
have a.s. limk→∞ Xt = Xt . Recall that f and fij′′ are rapidly decreasing and each
φi is bounded away from zero. Since γk ak → σ uniformly, for t ∈ K(X) we have a.s.
(k)
ˆ (Xt ) boundedly by (3.8), (3.13) and the definition of L.
ˆ Suppose
limk→∞ Lk F (Xt ) = LF
n+1
n
ˆ
that {Hi }i=1 ⊂ C(M(R )) and {ti }i=1 ⊂ K(X) with 0 ≤ t1 < · · · < tn < tn+1 . By Ethier
and Kurtz (1986, p.31), the set K(X) is at most countable. Then
Z
nh
E F (Xtn+1 ) − F (Xtn ) −

tn+1

tn

n
Y

n
= E F (Xtn+1 )
−
=

Z

Hi (Xti )

i=1

o

o

i=1

n

ˆ (Xs )
E LF

n
Y

o

Hi (Xti ) ds

i=1

n
n
n
o
n
o
Y
Y
(k)
(k)
(k)
(k)
lim E F (Xtn+1 )
Hi (Xti ) − lim E F (Xtn )
Hi (Xti )

k→∞

− lim

k→∞

=

n
iY

n
o
n
Y
Hi (Xti )
Hi (Xti ) − E F (Xtn )

i=1
tn+1

tn

ˆ (Xs )ds
LF

lim E

k→∞

nh

i=1
Z tn+1
tn

(k)
F (Xtn+1 )

−

k→∞

i=1

n
o
n
Y
(k)
(k)
Hi (Xti ) ds
E Lk F (Xs )

(k)
F (Xtn )

−

Z

i=1
tn+1

tn

Lk F (Xs(k) )ds

n
iY

(k)
Hi (Xti )

i=1

o

= 0.
By the right continuity of {Xt : t ≥ 0}, the equality
Z
nh
E F (Xtn+1 ) − F (Xtn ) −

tn+1
tn

ˆ (Xs )ds
LF

n
iY

holds without the restriction {ti }n+1
i=1 ⊂ K(X). That is,
Z t
ˆ (Xs )ds,
F (Xt ) − F (X0 ) −
LF
0

o

Hi (Xti ) = 0

i=1

t ≥ 0,

is a martingale. As in Wang (1998, pp.783-784) one can show that {Xt : t ≥ 0} is in fact

a.s. continuous.
17

ˆ be as in Lemma 4.2. Then for each µ ∈ M(Rˆ ), there is a probabilLemma 4.3 Let D(L)
ˆ
ity measure Qµ on C([0, ∞), M(Rˆ )) under which (4.4) is a martingale for each F ∈ D(L).

Proof. It is easy to find µk ∈ Mθk (R ) such that µk → µ as k → ∞. Then, by Lemma
4.2, it suffices to construct a sequence (γk , p(k) ) such that γ√
→ ∞. This
k σk → σ as k√
√ is
elementary. One choice is described as follows. Let γk = 1/ k and σk = k(σ + 1/ k).
Then the system of equations

(k)
(k)
(k)

 p0 + p2 + pk = 1,
(k)
(k)
2p2 + kpk
= 1,

 4p(k) + k 2 p(k)
= σk + 1,
2
k
has the unique solution
(k)

p0 =
(k)

where each pi

σk + k − 1
,
2k

(k)

p2 =

k − 1 − σk
,
2(k − 2)

(k)

pk =

σk − 1
,
k(k − 2)

is nonnegative for sufficiently large k ≥ 3.



Lemma 4.4 Let Qµ be given by Lemma 4.3. Then for n ≥ 1, t ≥ 0 and µ ∈ M(R ) we
have
Z t
1
n
n
Qµ {h1, wti } ≤ h1, µi + n(n − 1)kσk
Qµ {h1, ws in−1 }ds.
2
0
ˆ be the union
Consequently, Qµ {h1, wtin } is a locally bounded function of t ≥ 0. Let D(L)
n
2
2
of all functions of the form (4.1) with f ∈ C0 (R ) and {φi } ⊂ C∂ (R ) and all functions of
the form Fm,f (µ) = hf, µm i with f ∈ C∂2 (R m ). Then (4.4) under Qµ is a martingale for
ˆ
each F ∈ D(L).
Proof. For any k ≥ 1, take fk ∈ C02 (R )) such that fk (z) = z n for 0 ≤ z ≤ k and
fk′′ (z) ≤ n(n − 1)z n−2 for all z ≥ 0. Let Fk (µ) = fk (h1, µi). Then AFn (µ) = 0 and
1
BFk (µ) ≤ n(n − 1)kσkh1, µin−1.
2

Since
Fk (Xt ) − Fk (X0 ) −

Z

t
0

LFk (h1, Xsi)ds,

t ≥ 0,

is a martingale, we get
Z t
1
Qµ fk (h1, Xt i ) ≤ fk (h1, µi) + n(n − 1)kσk
Qµ (h1, Xs in−1 )ds
2
0
Z t
1
n
≤ h1, µi + n(n − 1)kσk
Qµ (h1, Xs in−1)ds.
2
0
n

Then the desired estimate follows by Fatou’s Lemma. The last assertion is an immediate
consequence of Lemma 4.3.

18

Lemma 4.5 Let Qµ be given by Lemma 4.3. Then for µ ∈ M(R ) and φ ∈ C∂2 (R ),
Z
1 t
haφ′′ , ws ids,
t ≥ 0,
(4.5)
Mt (φ) := hφ, wti − hφ, µi −
2 0
is a Qµ -martingale with quadratic variation process
Z t
Z t Z
2
hM(φ)it =
hσφ , ws ids +
ds hh(z − ·)φ′ , ws i2 dz.
0
0
Rˆ

(4.6)

Proof. It is easy to check that, if Fn (µ) = hφ, µin , then
Z
n(n − 1)
n
n−1
′′
n−2
ˆ
hφ, µi haφ , µi +
hφ, µi
LFn (µ) =
hh(z − ·)φ′ , µi2dz
2
2
Rˆ
n(n − 1)
+
hφ, µin−2hσφ2 , µi.
2
It follows that both (4.5) and
Mt2 (φ)

2

2

Z

t

:= hφ, wt i − hφ, µi −
hφ, ws ihaφ′′, ws ids
Z t Z 0
Z t
′
2
−
ds hh(z − ·)φ , ws i dz −
hσφ2 , ws ids
0
0
Rˆ

are martingales. By (4.5) and Itˆo’s formula we have
Z t
Z t
2
2
′′
hφ, wt i = hφ, µi +
hφ, ws ihaφ , ws ids + 2 hφ, wsidMs (φ) + hM(φ)it .
0

0

Comparing (4.7) and (4.8) we get the conclusion.

(4.7)

(4.8)



Observe that the martingales {M(φ) : t ≥ 0} defined by (4.5) form a system which is
linear in φ ∈ C∂2 (R ). Because of the presence of the derivative φ′ in the variation process
(4.6), it seems hard to extend the definition of {M(φ) : t ≥ 0} to a general function
φ ∈ B(Rˆ ). However, following the method of Walsh (1986), one can still define the
stochastic integral
Z tZ
φ(s, x)M(ds, dx),
t ≥ 0,
0
Rˆ
ˆ . With those in
if both φ(s, x) and φ′ (s, x) can be extended continuously to [0, ∞) × R
hand, we have the following
Lemma 4.6 Let Qµ be given by Lemma 4.3. Then for any t ≥ 0 and φ ∈ C∂2 (R ) we have
a.s.
Z tZ
ˆ
hφ, wt i = hPt φ, µi +
Pˆt−s φ(x)M(ds, dx).
ˆ
0
R
19

Proof. For any partition ∆n := {0 = t0 < t1 < · · · < tn = t} of [0, t], we have
hφ, wti − hPˆt φ, µi =

n
X
i=1

hPˆt−ti φ − Pˆt−ti−1 φ, wti i

+

n
X
i=1

[hPˆt−ti−1 φ, wti i − hPˆt−ti−1 φ, wti−1 i].

Let k∆n k = max{|ti − ti−1 | : 1 ≤ i ≤ n} and assume k∆n k → 0 as n → ∞. Then
lim

n→∞

n
X
i=1

hPˆt−ti φ − Pˆt−ti−1 φ, wti i = − lim

n→∞

= −

Z

t
0

n Z
X
i=1

ti
ti−1

ˆ wt ids
hPˆt−s Gφ,
i

ˆ ws ids.
hPˆt−s Gφ,

Using Lemma 4.5 we have
lim

n→∞

n
X

[hPˆt−ti−1 φ, wti i − hPˆt−ti−1 φ, wti−1 i]

i=1
n Z ti
X

n Z ti
X
1
= lim
ha(Pˆt−ti−1 φ)′′ , ws ids
Pˆt−ti−1 φM(ds, dx) + lim
n→∞
n→∞
2 i=1 ti−1
ˆ
i=1 ti−1 R
Z tZ
Z t
1
Pˆt−s φM(ds, dx) +
=
ha(Pˆt−s φ)′′ , ws ids.
2 0
0
Rˆ

Z

Combining those we get the desired conclusion.



Theorem 4.1 Let D(L) be the union of all functions of the form (4.1) with f ∈ C 2 (R n )
and {φi } ⊂ C 2 (R ) and all functions of the form Fm,f (µ) = hf, µm i with f ∈ C 2 (R m ). Let
{wt : t ≥ 0} denote the coordinate process of C([0, ∞), M(R )). Then for each µ ∈ M(R )
there is a probability measure Qµ on C([0, ∞), M(R )) such that Qµ {h1, wtim } is locally
bounded in t ≥ 0 for every m ≥ 1 and such that {wt : t ≥ 0} under Qµ is a solution of
the (L, D(L), µ)-martingale problem.
Proof. Let Qµ be the probability measure on C([0, ∞), M(Rˆ )) provided by Lemma 4.3.
The desired result will follow once it is proved that
Qµ {wt ({∂}) = 0 for all t ∈ [0, u]} = 1,

u > 0.

For any φ ∈ C∂2 (R ), we may use Lemma 4.6 to see that
Z tZ
u
Pˆu−s φM(ds, dx),
Mt (φ) := hPˆu−t φ, wt i − hPˆt Pˆu−t φ, µi =
Rˆ
0
20

(4.9)

t ∈ [0, u],

is a continuous martingale with quadratic variation process
Z t
Z t Z
u
2
ˆ
hM (φ)it =
hσ(Pu−s φ) , ws ids +
ds hh(z − ·)Pˆu−s (φ′ ), ws i2 dz
ˆ
Z0 t
Z0 t ZR
=
hσ(Pˆu−s φ)2 , ws ids +
ds hh(z − ·)(Pˆu−s φ)′ , ws i2 dz.
0
0
Rˆ
By a martingale inequality we have
o
n
2
ˆ
ˆ
Qµ sup |hPu−t φ, wt i − hPu φ, µi|
0≤t≤u
Z u
Z u Z
2
ˆ
≤ 4
Qµ {hσ(Pu−s φ) , ws i}ds + 4
ds Qµ {hh(z − ·)Pˆu−s (φ′ ), ws i2 }dz
ˆ
0
Z0 u
Z
Z Ru
≤ 4
hσ(Pˆu−s φ)2 , µPˆs ids + 4 h(z)2 dz
Qµ {h1, ws ihPˆu−s(φ′ )2 , ws i}ds
0
Rˆ Z
Z0 u
Z u
2
′
2
2
≤ 4
hσ(Pˆu−s φ) , µPˆs ids + 4kφ k
h(z) dz
Qµ {h1, ws i2 }ds.
0
Rˆ
0
Choose a sequence {φk } ⊂ C∂2 (R ) such that φk (·) → 1{∂} (·) boundedly and kφ′k k → 0 as
k → ∞. Replacing φ by φk in the above and letting k → ∞ we obtain (4.9).

Combining Theorems 2.2 and 4.1 we get the existence of the SDSM in the case where
σ ∈ C(R )+ extends continuously to Rˆ .

5

Measurable branching density

In this section, we shall use the dual process to extend the construction of the SDSM to
a general bounded Borel branching density. Given σ ∈ B(R )+ , let {(Mt , Yt ) : t ≥ 0} be
defined as in section 2. Choose any sequence of functions {σk } ⊂ C(R )+ which extends
continuously to Rˆ and σk → σ boundedly and pointwise. Suppose that {µk } ⊂ M(R )
(k)
and µk → µ ∈ M(R ) as k → ∞. For each k ≥ 1, let {Xt : t ≥ 0} be a SDSM with
parameters (a, ρ, σk ) and initial state µk ∈ M(R ) and let Qk denote the distribution of
(k)
{Xt : t ≥ 0} on C([0, ∞), M(R )).
Lemma 5.1 Under the above hypotheses, {Qk } is a tight sequence of probability measures on C([0, ∞), M(R )).
(k)

Proof. Since {h1, Xt i : t ≥ 0} is a martingale, one can see as in the proof of Lemma
(k)
4.1 that {Xt : t ≥ 0} satisfies the compact containment condition of Ethier and Kurtz
(k)
(1986, p.142). Let Lk denote the generator of {Xt : t ≥ 0} and let F be given by (4.1)
with f ∈ C02 (R n ) and with {φi } ⊂ C∂2 (R ). Then
Z t
(k)
(k)
t ≥ 0,
F (Xt ) − F (X0 ) −
Lk F (Xs(k) )ds,
0

21

(k)

is a martingale. Since the sequence {σk } is uniformly bounded, the tightness of {Xt : t ≥
0} in C([0, ∞), M(Rˆ )) follows from Lemma 4.4 and the result of Ethier and Kurtz (1986,
p.145). We shall prove that any limit point of {Qk } is supported by C([0, ∞), M(R ))
so that {Qk } is also tight as probability measures on C([0, ∞), M(R )). Without loss
of generality, we may assume Qk converges as k → ∞ to Qµ by weak convergence of
probability measures on C([0, ∞), M(Rˆ )). Let φn ∈ C 2 (R )+ be such that φn (x) = 0 when
kxk ≤ n and φn (x) = 1 when kxk ≥ 2n and kφ′n k → 0 as n → ∞. Fix u > 0 and let
mn be such that φmn (x) ≤ 2Pt φn (x) for all 0 ≤ t ≤ u and x ∈ R . For any α > 0, the
paths w ∈ C([0, ∞), M(Rˆ )) satisfying sup0≤t≤u hφmn , wt i > α constitute an open subset
of C([0, ∞), M(Rˆ )). Then, by an equivalent condition for weak convergence,
o
o
n
n
Qµ sup wt ({(∂)}) > α ≤ Qµ sup hφmn , wt i > α
0≤t≤u

0≤t≤u

o
n
4
2
ˆ
h
P
φ
,
w
i
Q
sup
u−t n
t
k
2
k→∞
0≤t≤u
0≤t≤u
k≥1 α
o
n
8
8
≤ sup 2 Qk sup |hPˆu−t φn , wt i − hPˆu φn , µk i|2 + sup sup 2 hPˆu φn , µk i2 .
0≤t≤u
k≥1 α
k≥1 0≤t≤u α

≤ lim inf Qk

n

o

sup hφmn , wt i > α ≤ sup

As in the proof of Theorem 4.1, one can see that the right hand side goes to zero as
n → ∞. Then Qµ is supported by C([0, ∞), M(R )).

Theorem 5.1 The distribution Qt (µk , ·) of Xt on M(R ) converges as k → ∞ to a
probability measure Qt (µ, ·) on M(R ) given by
Z
n1 Z t
oi
h
σ
m
Mt
hf, ν iQt (µ, dν) = E m,f hYt, µ i exp
Ms (Ms − 1)ds .
(5.1)
2 0
M (R)
(k)

(k)

Moreover, (Qt )t≥0 is a transition semigroup on M(R ).
Proof. By Lemma 5.1, {Qt (µk , dν)} is a tight sequence of probability measures on M(R ).
(k )
Take any subsequence {ki} so that Qt i (µki , dν) converges as i → ∞ to some probability
measure Qt (µ, dν) on M(R ). By Lemma 2.1 we have
Z
(k)
1[a,∞) (h1, νi)h1, ν miQt (µk , dν)
M (R)
Z
1
(k)
≤
h1, ν m+1 iQt (µk , dν)
a M (R)
m
1 X −i
≤
2 (m + 1)i mi kσk ki h1, µk im−i+1 ,
a i=0
(k)

which goes to zero as a → ∞ uniformly in k ≥ 1. Then for f ∈ C(Rˆ )+ we may re(k)
gard {hf, ν m iQt (µk , dν)} as a tight sequence of finite measures on M(Rˆ ). By passing
22

(k )

to a smaller subsequence {ki} we may assume that hf, ν m iQt i (µki , dν) converges to a
finite measure Kt (µ, dν) on M(Rˆ ). Then we must have Kt (µ, dν) = hf, ν m iQt (µ, dν).
By Lemma 2.2 and the proof of Theorem 2.2, Qt (µ, ·) is uniquely determined by (5.1).
(k)
Therefore, Qt (µk , ·) converges to Qt (µ, ·) as k → ∞. From the calculations
Z
Z
Qr (µ, dη)
hf, ν m iQt (η, dν)
M (R)
M (R)
Z
oi
n1 Z t
h
σ
Mt
Ms (Ms − 1)ds Qr (µ, dη)
=
E m,f hYt, η i exp
2 0
M (R)
oi
hZ
n1 Z t
σ
Mt
Ms (Ms − 1)ds
= E m,f
hYt, η iQr (µ, dη) exp
2 0
M (R)

n1 Z r
h
o
n1 Z t
oi
σ
σ
Mr
= E m,f E Mt ,Yt hYr , µ i exp
Ms (Ms − 1)ds exp
Ms (Ms − 1)ds
2 0
2 0
oi
n 1 Z r+t
h
= E σm,f hYr+t, µMr+t i exp
Ms (Ms − 1)ds
2 0
Z
=
hf, ν m iQr+t (η, dν)
M (R)



we have the Chapman-Kolmogorov equation.

The existence of a SDSM with a general bounded measurable branching density function
σ ∈ B(R ) is given by the following
Theorem 5.2 The sequence Qk converges as k → ∞ to a probability measure Qµ on
C([0, ∞), M(R )) under which the coordinate process {wt : t ≥ 0} is a diffusion with
transition semigroup (Qt )t≥0 defined by (5.1). Let D(L) be the union of all functions
of the form (4.1) with f ∈ C 2 (R n ) and {φi } ⊂ C 2 (R ) and all functions of the form
Fm,f (µ) = hf, µm i with f ∈ C 2 (R m ). Then {wt : t ≥ 0} under Qµ solves the (L, D(L), µ)martingale problem.
Proof. Let Qµ be the limit point of any subsequence {Qki } of {Qk }. Using Skorokhod’s
(k )
representation, we may construct processes {Xt i : t ≥ 0} and {Xt : t ≥ 0} with
(k )
distributions Qki and Qµ on C([0, ∞), M(R )) such that {Xt i : t ≥ 0} converges to
{Xt : t ≥ 0} a.s. when i → ∞; see Ethier and Kurtz (1986, p.102). For any {Hj }n+1
j=1 ⊂
C(M(Rˆ )) and 0 ≤ t1 < · · · < tn < tn+1 we may use Theorem 5.1 and dominated
convergence to see that
E

n
nY

Hj (Xtj )Hn+1 (Xtn+1 )

j=1

= lim E
i→∞

n
nY

(k )

o
(k )

i
Hj (Xtj i )Hn+1 (Xtn+1
)

j=1

23

o

= lim E
i→∞

= E

n
nY

(k )
Hj (Xtj i )

j=1

n
nY

Hj (Xtj )

j=1

Z

M (R)

Z

(k )

M (R)

(k )

i
i
Hn+1 (ν)Qtn+1
−tn (Xtn , dν)

o

o
Hn+1 (ν)Qtn+1 −tn (Xtn , dν) .

Then {Xt : t ≥ 0} is a Markov process with transition semigroup (Qt )t≥0 and actually
Qk → Qµ as k → ∞. The strong Markov property holds since (Qt )t≥0 is Feller by (5.1).
To see the last assertion, one may simply check that (L, D(L)) is a restriction of the
generator of (Qt )t≥0 .


6

Rescaled limits

In this section, we study the rescaled limits of the SDSM constructed in the last section.
Given any θ > 0, we defined the operator Kθ on M(R ) by Kθ µ(B) = µ({θx : x ∈ B}).
For a function h ∈ B(R ) we let hθ (x) = h(θx).
Lemma 6.1 Suppose that {Xt : t ≥ 0} is a SDSM with parameters (a, ρ, σ). Let Xtθ =
θ−2 Kθ Xθ2 t . Then {Xtθ : t ≥ 0} is a SDSM with parameters (aθ , ρθ , σθ ).
Proof. We shall compute the generator of {Xtθ : t ≥ 0}. Let F (µ) = f (hφ, µi) with
f ∈ C 2 (R ) and φ ∈ C 2 (R ). Note that F ◦ Kθ (µ) = F (Kθ µ) = f (hφ1/θ , µi). By the
theory of transformations of Markov processes, {Kθ Xt : t ≥ 0} has generator Lθ such
that Lθ F (µ) = L(F ◦ Kθ )(K1/θ µ). Since
d
1
φ1/θ (x) = (φ′ )1/θ (x) and
dx
θ

d2
1
φ1/θ (x) = 2 (φ′′ )1/θ (x),
2
dx
θ

it is easy to check that
Lθ F (µ) =

1 ′
f (hφ, µi)haθ φ′′ , µi
2
2θ
Z
1 ′′
+ 2 f (hφ, µi)
ρθ (x − y)φ′(x)φ′ (y)µ(dx)µ(dy)
2θ
R2
1
+ f ′′ (hφ, µi)hσθ φ2 , µi.
2

Then one may see that {θ−2 Kθ Xt : t ≥ 0} has generator Lθ such that
Lθ F (µ) =

1 ′
f (hφ, µi)haθ φ′′ , µi
2
2θ
Z
1 ′′
+ 2 f (hφ, µi)
ρθ (x − y)φ′(x)φ′ (y)µ(dx)µ(dy)
2θ
R2
1 ′′
+ 2 f (hφ, µi)hσθ φ2 , µi,
2θ
24

and hence {Xtθ : t ≥ 0} has the right generator θ2 Lθ .



Theorem 6.1 Suppose that (Ω , Xt , Qµ ) is a realization of the SDSM with parameters
(a, ρ, σ) with |c(x)| ≥ ǫ > 0 for all x ∈ R . Then there is a λ × λ × Qµ -measurable function
Xt (ω, x) such that Qµ {ω ∈ Ω : Xt (ω, dx) is absolutely continuous with respect to the
Lebesgue measure with density Xt (ω, x) for λ-a.e. t > 0} = 1. Moreover, for λ × λ-a.e.
(t, x) ∈ [0, ∞) × R we have
Z
2
Qµ {Xt (x) } =
p2t (y, z; x, x)µ(dx)µ(dy)
2
R Z
Z
Z
t
+
ds µ(dy) σ(z)p2s (z, z; x, x)pt−s (y, z)dz.
(6.1)
0
R
R
Proof. Recall (1.9). For r1 > 0 and r2 > 0 we use (2.7) and (5.1) to see that
1
1
1
1
Qµ {hgǫr
(x, ·), Xt ihgǫr
(x, ·), Xt i} = Qµ {hgǫr
(x, ·) ⊗ gǫr
(x, ·), Xt2 i}
1
2
1
2
Z t
1
1
1
1
(x, ·) ⊗ gǫr
(x, ·), µ2 ids
(x, ·) ⊗ gǫr
(x, ·), µ2 i +
hPt−s φ12 Ps2 gǫr
= hPt2gǫr
1
2
1
2
0
Z
1
1
=
Pt2 gǫr
(x, ·) ⊗ gǫr
(x, ·)(y, z)µ(dy)µ(dz)
1
2
2
RZ
Z
Z
t
1
1
+
ds µ(dy) σ(z)Ps2 gǫr
(x, ·) ⊗ gǫr
(x, ·)(z, z)pt−s (y, z)dz.
1
2
0
R
R
Observe that
Z
1
1
2 1
1
Pt gǫr1 (x, ·) ⊗ gǫr2 (x, ·)(y, z) =
gǫr
(x, z1 )gǫr
(x, z2 )p2t (y, z; z1 , z2 )dz1 dz2
1
2
2
R
2
converges to pt (y, z; x, x) boundedly as r1 → 0 and r2 → 0. Note also that
Z
1
1
(x, ·) ⊗ gǫr
(x, ·)(z, z)pt−s (y, z)dz
σ(z)Ps2 gǫr
1
2
R
Z
1
1
1
≤ const · kσk √
Tǫs gǫr
(x; ·)(z)gǫ(t−s)
(y, z)dz
1
s R
1 1
≤ const · kσk √ gǫ(t+r
(y, x)
1)
s
1
≤ const · kσk √ .
st

By dominated convergence theorem we get
1
1
lim Qµ {hgǫr
(x, ·), Xt ihgǫr
(x, ·), Xt i}
1
2

r1 ,r2 →0

=

Z

p2t (y, z; x, x)µ(dy)µ(dz)
R2
Z t Z
Z
+
ds µ(dy) σ(z)p2t (z, z; x, x)pt−s (y, z)dz.
0
R
R
25

Then it is easy to check that
Z T Z
1
1
(x, ·) − gǫr
(x, ·), Xt i2 }dx = 0
lim
dt Qµ {hgǫr
1
2
r1 ,r2 →0 0
R
for eac