Chapter 5

Self diffusion.

5.1 Motion of a tagged particle.

Let us look at the simple exclusion process in equilibrium on Zd at density . The distri-
bution is the Bernoulli distribution defined by [(x) = 1] = with {(x) : x Zd }
being independent. Let us suppose that at time 0, there is a particle at 0 which is tagged
and observed. It is convenient to move the origin with that particle. The simple exclusion
process now acts only on Zd {0} and is the environment as seen by the particle. The
environment changes in two different ways. When one of the other particles currently at x
moves to y. The generator for this part is
1 !
A1 = (y x)[f ( x,y ) f ()] (5.1)
2
x,y!=0

Or the tagged particle moves from 0 to z and then the origin is shifted to z. This is a
transformation Tz that acts when (z) = 0 and the new configuration on Zd {0} is given
by
(Tz )(x) = (x + z) if x #= z, 0; (Tz )(z) = 0
contributing to the generator the term
!
A2 = (z)(1 (z))[f (Tz ) f ()] (5.2)
z

The full generator is therefore

1 !
(Af )() = (y x)[f ( x,y ) f ()]
2
x,y!=0
!
+ (z)(1 (z))[f (Tz ) f ()] (5.3)
z
= (A1 f )() + (A2 f )()

77
78 CHAPTER 5. SELF DIFFUSION.

It is not difficult to check that the probability distribution on Zd {0} is a reversible

invariant distribution for A given by 5.3. The jumps x y and y x as well as Tz and
Tz provide pairs with detail balance. The rates are same in either direction and is
invariant under the transitions.
Our main tool is a central limit theorem for additive functions of a reversible Markov
process. Given a real valued function f on a space X , a Markov process on that space with
generator A and a reversible ergodic invariant measure for A satisfying with E [f (x)] = 0
under suitable conditions we will show that
" t
f (x(s))ds = M (t) + a(t)
0

where M (t) is a square integrable Martingale with stationary increments and a(t) is neg-
ligible. # If A is the self adjoint generator of the process A has a spectral resolution

A = 0 E(d). We have the Dirichlet form D(f ) = %Af, f &L2 () associated with A.
The space H1 is the abstract Hilbert space obtained by completing the space of square inte-
grals functions with respect to the Dirichlet inner product. One might start with functions
u in the domain of A, ensuring the finiteness of D(f ). The completion will be an abstract
space H1 . There will be a dual H1 to H1 relative to the inner product of H0 = L2 ().
Formally 'u'1 = %(A)1 u, u& and

'u'21 = sup 2%u, f & D(f )

f

We have the following theorem.

Theorem 5.1.1. If f is in L2 with spectral resolution %E(d)f, f &, and

"
1
%(A) f, f & = 1 %E(d)f, f & < = 2 <
0

there is a square integrable Martingale M (t) with stationary increments such that
E P [M (t)2 ] = 2 2 t and
" t
f (x(s))ds = M (t) + a(t)
0

with E[|a(t)|2 ] = o(t) as t . The central limi theorem follows. Moreover

P [ sup |a(t) c T ] 0
0tT

for every c > 0 implying the functional CLT.

For the proof of the theorem we need two lemmas.

5.1. MOTION OF A TAGGED PARTICLE. 79

Lemma 5.1.2. Let P be a reversible stationary Markov process with invariant measure
and generator A. Let u L2 [] with D(u) = %Au, u& < . Then
$
e
P [ sup |u(x(t))| %] T D(u) + 'u'22
0tT %

Proof. Since D(|u|) D(u) we can assume that u 0. If x(t) is a Markov process and
is the exit time from G, then Ex [e ] = v(, x) is the solution of

v(x) (Av)(x) = 0 for x G; v = 1 on Gc

The function v is also the minimizer of

'v'22 + D(v)

over v such that v = 1 on Gc . Therefore the solution v satisfies

'v '22 inf ['v'22 + D(v)]

v:v=1 on Gc

If we take for G the set u(x) < %, the function v = u$ $ is an admissible choice for v.
Therefore with = T ,1
$
1
'v '1 'u'22 + T D(u)
%
We obtain the estimate
" " $
e
Px [ < T ]d eT Ex [e ]d = e'v '1 'u'22 + T D(u)
%

This lemma quantifies the statement that set of singularities of a function u on Rd that
is in the Sobolev space W21 (Rd ) has capacity 0. In other words even if u has singularities,
a Brownian path will not see it, i.e. u((t)) is almost surely continuous.
Lemma 5.1.3. Let 'u'2 and D(u) be finite. Then for any c > 0

lim sup P [ sup |u(x(t)| c T |] = 0
T 0tT

Proof. For any given > 0 find u' L such that 'u u' '22 and D(u u' ) . Clearly

lim sup P [ sup |u' (x(t)| c T |] = 0
T 0tT

and
'
e
lim sup P [ sup |(u u )(x(t))| c T |]
T 0tT c
and can be made arbitrarily small the lemma is proved.
80 CHAPTER 5. SELF DIFFUSION.

Proof. Now we return to complete the proof of Theorem 5.1.1. First let us note that the
condition is natural. An elementary calculation shows that
" t " t" t
1 P 1
E [| f (x(s))ds|2 ] = E P [ f (x(s))f (x(s' ))dsds' ]
t 0 t 0 0
"
2
= %Ts! s f, f &dsds'
t 0ss! t
" t
s
= 2 (1 )%Ts f, f &ds
t
"0
,2 %Ts f, f &ds
0
= 2%(A)1 f, f &
= 2 2

Since %Tt f, f & 0, the convergence has to be absolute. Let us solve the resolvent equation

u Au = f

Then Au = u f and
" t " t
u (x(t)) u (x(0)) u (x(s))ds + f (x(s))ds = M (t)
0 0

where M (t) is a martingale with

"
1 2
E[M (t)2 ] = 2D(u ) = 2%Au , u & = 2 %E(d)f, f &
t 0 ( + )2
An easy computation shows that ( + )1 1 and is dominated by 1 which is
integrable with respect to %E(d)f, f &. The martingales M (t) have a limit in L2 (P ).
u 0 in L2 (). Therefore a (t) = u (x(0)) u (x(t)) has a limit a(t) and
" t
f (x(s))ds = M (t) + a(t)
0

We will show that E[|a(t)|2 ] = o(t). Then martingale CLT will imply our result. This is
again a spectral calculation.
" "
P 2 1 et 1 et
E [|a(t)| ] = 2 lim %E(d)f, f & = 2 %E(d)f, f &
0 0 ( + )2 0 2
t #
Since 1et and 0 1 %E(d)f, f & < , the dominated convergence theorem implies
that
1
lim E P [|a(t)|2 ] = 0
t t
5.1. MOTION OF A TAGGED PARTICLE. 81

To prove the functional CLT, we need to consider

" tT " tT
1 1 1 1
f (x(s))ds = M (T t) + u (s)ds [u (x(tT )) u (x(0))]
T 0 T T 0 T
The functional CLT holds for 1T M (tT ) and uniformly so as 0 because M (t) M (t)
in mean square. We note that with the help of the dominated convergence theorem,
"
2
'u ' = %E(d)f, f & 0
0 ( + )2

as 0. Clearly with the choice of = T 1 ,

" tT "
1 1 T
T = sup | u (s)ds| |u (x(s))|ds
0t1 T 0 T 0

and E[|T |2 ] 0 as t . To complete the proof we need to show that, with = T 1

P [ sup |u 1 (x(s))| c T ] 0
0tT T

We can represent u 1 as u + (u 1 u ). By lemma 5.1.3, we have for any > 0,

T T

lim sup P [ sup |u (x(s))| c T ] = 0
T 0sT

Moreover
1
lim 'u 1 u '22 = 0
T T T
and
lim D(u 1 u ) = 0
T T
0

They imply that

lim sup lim sup P [ sup |u 1 (x(s)) u (x(s))| c T ] = 0
T
0 T 0tT

We now return to the motion of the tagged particle. We need to keep track of its motion
as well as the changing environment seen by it. If w Zd is the location of the tagged
particle in the original reference frame, then jointly the generator for w(t) Z d and
d
() {0, 1}Z {0} is
!
)() =
(Af (z)(1 (z))[f (w + z, Tz ) f (w, )] + (Af )(w, ) (5.4)
z
82 CHAPTER 5. SELF DIFFUSION.

with A acting only on for each w. The (t, ) part is a Markov process by itself and is
in equilibrium at density , the distribution being . We are interested in establishing a
central limit theorem for w(t). We note that
" t!
w(t) w(0) = z (z)(1 (s, z))ds + M (t)
0 z

where M (t) is a Martingale with the decomposition

!
M (t) = zMz (t)
z

and " t
Mz (t) = Nz (t) (z) (1 (s, z))ds
0
The quantity !
V (()) = z (z)(1 (z))
z
has mean 0 in equilibrium and one may expect a CLT for
" t
V ((s, ))ds
0

We will prove a decomposition of the form

" t
V ((s, ))ds = N (t) + a(t)
0

where N (t) is a martingale and a(t) is negligible. Then

w(t) w(0) = M (t) + N (t) + a(t)

and since central limit theorems for martingales are automatic the result will follow. The
quantities here are vectors and the equations are for each component or they are interpreted
as
%w(t) w(0), & = %M (t), & + %N (t), & + %a(t), &
for Rd . We have now the main theorem.

Theorem 5.1.4. The position w(t) of the tagged particle satisfies a functional CLT, with
positive definite covariance matrix S() given by
%" ! ! &
2 1 x,y 2
%S(), & = inf [ (z)(1(z))(z f f %, z&) + (yx)(f ( )f ()) ]d
f
z
2 x,y
5.1. MOTION OF A TAGGED PARTICLE. 83

First we need to prove, for each vector Rd , a bound of the form

" ! ' $
| %z, &(1 (z))(z)f ()d | C() D (f )
z

We can rewrite, after combining the z and z terms and symmetrizing

%! &

E %z, &(1 (z))(z)f ()
z
% &
1 !
= E %z, &[(1 (z)) (1 (z))](z)f ()
2 z
% &
1 !
= E %z, &[1 (z))](z)[f () f (Tz )]
2 z
% % && 1 % %! && 1
1 ! 2
2
2
2
E |%z, &| [1 (z))](z) E [1 (z))](z)[f () f (Tz )]
2 z z
' $
C() D (f )

with
1!
C() = |%z, &|2 (z)
4 z

This proves the validity of functional CLT for w(t) with an upper bound on the variance.
The next step is to establish the formula and a lower bound for it. Let us compute
%S(), &. The minimizer f = f may not exist. The space H1 of functions u L2 , with the
Dirichlet inner product, when completed, will admit objects that are not in L2 ( ). There
is no Poincare inequality available. Abstractly the space consists of collections of functions
{g x,y ()}, {gz }, that are the limits in H1 of {f ( x,y ) f ()}, (1 (z))[f (Tz ) f ()]. The
functions g x,y (), gz satisfy identities. gx,y is 0 unless (x) #= (y) and satisfies x,y + y,x =
0. Similarly gz is nonzero only when (z) = 0 and (1 (z))gz ()+ (1 (z))gz (Tz ) =
0. The Euler equation for the variational problem is
!
E [ (z)(1 (z))[gz () %, z&][f (z ) f ()]]
z
1 !
+ E [ (y x)g x,y ()[f ( x,y ) f ()]] = 0
2 x,y

for all f , which after a bit of calculation, takes the form

1! !
V () + (y x)gx,y + (z)(1 (z))gz = 0
2 z
84 CHAPTER 5. SELF DIFFUSION.

w(t) now has the representation

" t !" t
%, w(t)& = %, V ((s))&ds + %, z&(1 (z))dMz (t)
0 z 0

with " t
%, V ((s))&ds = a(t) + N (t)
0

and
!" t !" t
N (t) = gz ((s))dMz (s) + gx,y ((s))Mx,y (t)
z 0 x,y 0

with " t
Mx,y (t) = Nx,y (t) (y x)(s, x)(1 (y, s))ds
0

Therefore
" t!
%, w(t)& = [%z, & gz ((s))]dMz (s)
0 z
" t!
gx,y ((s))dNx,y (s) + a(t)
0 x,y

= M (t) + a(t)

Computing the quadratic variation of the martingale M (t) proves the formula. Finally we
will prove the non degeneracy of the quadratic form %S(), &. We have to exclude the one
dimensional nearest neighbor case, where S() 0. The proof depends on the following
fact. We can obtain an estimate of the form
'
E [((z) (z))f ()] C D2 (f )

in terms of the Dirichlet form

1 !
D1 (u) = %A1 u, u& = E [ (y x)[f ( x,y ) f ()]2 ]
4 x,y

It is possible to shift a particle from z to z without touching the tagged particle at 0.

Jump over it or go around it. This provides an estimate of the form
( ! )1
E [((z) (z))f ()] C(z) E [(y x)[f ( x,y ) f ()]2 ] 2 (5.5)
x,y
5.1. MOTION OF A TAGGED PARTICLE. 85

We can estimate, for any a > 0,

' '
%(I A)1 %V, &, %V, && %(A1 (I A)1 %V, &, (I A)1 %V, && %(A1 )1 %V, &, %V, &&
a 1
%(A1 (I A)1 %V, &, (I A)1 %V, && + %(A1 )1 %V, &, %V, &&
2 2a

Letting 0,
2 1
%S(), & %(A1 (A)1 %V, &, (A)1 %V, && %(A)1 %V, &, %V, && 2 %(A1 )1 %V, &, %V, &&
a a
We can obtain a lower bound for the first quadratic form on the right form the variational
formula ( )
%(A)1 g, g& = sup 2%g, f & %Af, f &
f

and an upper bound for the second one from (5.5). Picking a large will do it.