130 23 11MB
English Pages 212 [206] Year 1984
Lecture Notes in Mathematics Edited by A. Dold and B. Eckmann
1095
Stochastic Analysis and Applications Proceedings of the International Conference held in Swansea, April 11-15, 1983
Edited by A. Truman and D. Williams
Springer-Verlag Berlin Heidelberg New York Tokyo 1984
Editors Aubrey Truman David Williams Department of Mathematics and Computer Science University College of Swansea Singleton Park, Swansea SA2 8PR Wales
AMS Subject Classification (1980): 60H05, 60H10 ISBN 3-540-13891-9 Springer-Verlag Berlin Heidelberg New York Tokyo ISBN 0-387-13891-9 Springer-Verlag New York Heidelberg Berlin Tokyo
This work is subject to copyright. All rights are reserved, whether the whole or part of the material is concerned, specifically those of translation, reprinting, re-use of illustrations, broadcasting, reproduction by photocopying machine or similar means, and storage in data banks. Under § 54 of the German Copyright Law where copies are made for other than private use, a fee is payable to "Verwertungsgesellschaft Wort", Munich.
© by Springer-Verlag Berlin Heidelberg 1984 Printed in Germany Printing and binding: Beltz Offsetdruck, Hemsbach / Bergstr. 2146/3140-543210
PREFACE
This volume contains a number of papers presented at the Workshop on Stochastic Analysis and its Applications, held in Swansea from 11 April to 15 April 1983, together with some more recent research papers by the Swansea school.
The applications include such diverse
topics as stochastic mechanics and the Titius-Bode law, non-standard Dirichlet forms and polymers, statistical mechanics, quantum stochastic processes, the applications of local-time to proving path-wise uniqueness of solutions of stochastic differential equations and its application to excursion theory, Bessel processes and pole-seeking Brownian motion, queues, potential theory and Wiener-Hopf theory. Some new results for Brownian motion on how one process determines another are also given.
The applications to Mathematical Physics
appear first, followed by the papers on local-time, Bessel processes and queues.
The papers of the Swansea school are collected together
at the end of the volume.
We are grateful to SERC for financial support
through research grant GR/C52162 and we are especially indebted to James Taylor for invaluable help and advice during the conference. Finally we should like to record our thanks to Mrs E. Williams, Mrs M. Prowse and Mrs M. Brook for making such an excellent job of typing the Swansea contributions.
A. Truman D. Williams Swansea April, 1983
TABLE OF CONTENTS S. ALBEVERIO, PH. BLANCHARD, R. H0EGH-KROHN, 'Newtonian diffusions and planets, with a remark on non-standard Dirichlet forms and polymers'. J.T. LEWIS, J.V. PULE, mechanics' •
'The equivalence of ensembles in statistical
· 25
F. PAPANGELOU, 'The uniqueness of regular DLR measures for certain one-dimensional spin systems. .
36
R.L. HUDSON, K.R. PARTHASARATHY,
45
'Generalised Weyl operators'.
J.F. LE GALL, 'One-dimensional stochastic differential equations involving the local-times of unknown processes'.
• 51
P. McGILL, 'Time changes of Brownian motion and the conditional excursion theorem' • .
· 83
M. YOR, 'On square-root boundaries for Bessel processes and poleseeking Brownian motion' • .
.100
P.K. POLLETT, 'Distributional approximations for networks of quasireversible queues' • .
.108
J. HAWKES,
.130
'Some geometric aspects of potential theory'.
G.C. PRICE, L.C.G. ROGERS, D. WILLIAMS,
ISAdS '.
G.C. PRICE,
'BM(JR3)
and its area integral .155
'The unique factorisation of Brownian products'.
N. BAKER, 'Some integral equalities in Wiener-Hopf theory'. L.C.G. ROGERS, D. WILLIAMS, theory' •
'A differential equation in Wiener-Hopf
.166 .169
. .187
NEWTONIAN DIFFUSIONS AND PLANETS, WITH A REMARK ON NON-STANDARD DIRICHLET
AND POLYMERS
by
S. Albeverio Mathematisches Institut Ruhr-Universitat Bochum
Ph. Blanchard Theoretische Physik Universittit Bielefeld D-4800 Bielefeld
R. H¢egh-Krohn Universite de Provence Centre de Physique Theorique, CNRS F-13288 Marseille and Matematisk Institutt Universitetet i Oslo Blindern, Oslo
Abstract We discuss diffusion processes on Riemannian manifolds, for which a Newton law holos
(in the stochastic sense). We
the existence
of a general mechanism for the formation of impenetrable barriers for these processes, corresponding to the nodes of the density of their cistribution. We discuss some applications to natural phenomena like the formation of planetary systems; the morphology of galaxies, the formation of zones of winds in the atmosphere and
the formation
of spokes in the rings of Saturn. \1e also relate the recent hyperfinite theory of Dirichlet forms with the theory of local times of Brownian ')
motion, polymer measures and thp. (¢l theory.
')
cf quantum field
2
1. Introduction
In this lecture we shall discuss two topics, which are connected by the theory of diffusion processes. In the first part, consisting of Sections 2, 3 and 4, we shall discuss a class of diffusion processes, which we call "Newtonian diffusion processes", which show the remarkable phenomenon of "barrier formation", leading to a possible explanation of a large class of natural phenomena. In the second part we shall briefly discuss the concept of local time of Brownian motion and a new hyperfinite version of the theory of Dirichlet forms and apply this to the study of polymer measures associated with certain quantum field theoretical models. In Section 2 we give the definition and basic properties of Newtonian diffusions on manifolds. This theory has been introduced by E. Nelson in connection with stochastic mechanics [ 3 ], [ 4 l . [29] and developed further particularly by Dankel [ 1], Dohrn and Guerra [ 2] and for the case of manifolds recently by Meyer [ 5] and Morato [19]. We review the basic formalism and discuss the stationary case in particular (this case has been discussed previously by ourselves in nection with Dirichlet forms, in [ 7],
[13], [28] and, in con-
[10] and by Nagasawa [16]).
In Section 3 we discuss the general mechanism in the symmetric case for the barrier formation in Newtonian processes, using previous results obtained in ref.
[7] and [10], in the context of the "Dirichlet ap-
proach" to quantum mechanics
(see e.g. [42],
[31], [34],
[50], [52], [57]).
In Section 4 we discuss the mentioned applications to natural phenomena like the formation of planetary systems, the morphology of galaxies, zones of winds circulation and the formation of spokes in the rings of Saturn. In Section 5 we briefly discuss some problems in connection with the socalled polymer measures. This involves the study of "times spent at intersections of Brownian motions", quantities that also have arisen in the very stimulating lecture of Prof. E.B. Dynkin. We mention a couple of central problems in this area and can partially
indicate how in dimension 4 we
solve these by using a non standard
theory of Dirichlet forms.
2. Newtonian Diffusion Processes In this section we shall briefly describe how an important class of Markov diffusion processes, called "Newtonian processes", shows an
3
interesting phenomenon of formation of barriers on the nodes of the solution of a linear equation of elliptic or Schrodinger type, and how this remarkable property can be used to describe situations in nature, in which regular patterns of "confinement" arise. Let
M be a smooth oriented Riemannian manifold of dimension
Let
X t ETc JR+ ' be a diffusion process with values in M t, analytic description of X is by its infinitesimal generator t which we assume of the form
..:LA 2'-' + where
SiD.
S.D
0 fJ· i
l
d The L
D
'
(2.1)
1r
=
t
•••
is a (non random) COO vector
d,
field (the "drift"), which might depend explicitly on the is the covariant derivative and
D
M
tor on S(Xt,t)
The connection between
6
time t . is the Laplace-Beltrami oper a-:
S
and the process
is the (mean) forward derivative of
X t
sense that i
S (x,t) where
E[. IX
t
6t from
X t
to
=
x]
lim
x]
(L,t)
means conditional expectation with respect to
is the vector attached to Xt +lI t
X t
6tl
' wi th length
X + lIt and X S(Xt,t) t t. derivative in the sense of [5 ] . dx
(2.2)
6t+o
distance of
Let
X is that t at time t in the
tangent to geodesics
equals to the geodesics
is also the forward stochastic
be the Riemannian volume element on
M. Due to the assumpp (x,t) of the law
tions we know that there exists a smooth density of
Xt
Let
f
with respect to
f E
, then
dx, i.e. E[foX
t]
f
=
t
E dx)
f(x)p(x,t)dx M
ap
M
dP(X
p (x,t)dx. and
dt
E[f
0
] =
f(x)3t(x,t)dx . On the other hand, by the definition of
left hand side is equal to
f
(L M
t
f) (x,t) p(x,t)dx . By partial integra-
tions we arrive at the Kolmogorov forward equation (Fokker-Planck equation) 1
26p - div(Sp)
(2.3)
4
- div(B· ) being the adjoint of ator on vector fields on
L
M
and
t
div
means divergence oper-
Let us now denote by i.e. that
X , t E -T, the time reversed process to X f t t It is well known, see e.g. [27], has the same law as is again a Markov process with infinitesimal generator
X t
- i3 . with
S. D
(2.4)
D
being the "backward drift" defined, for t E T,
Si
by lim (H)
x]
lIt+o i
(2.5)
i
with Y -lit defined as Y lit with -lit replacing lit. Then B is the backward stochastic derivative of X . By the same procedure as t above, one arrives at the Fokker-Planck equation for the reversed process,
t E T : d
ItP Set now
u
ity and
v
1
2 (B-S) A
and
P
V"
+ di v (i3
1
2(
).
u
(2.6)
p)
is called the osmotic veloc--
is called the current velocity. Inserting these expressions
in the Fokker-Planck equations (2.3),
(2.6) we get the "continuity
equation" -div(pv)
P
(2.7)
and the "osmotic equation" 1
2'''; P
=
As remarked first by Nelson in the case u
(2.8)
div(Pu) M
d :IR
we have also
21 v log P .
This follows, see [29], by computing for
(2.9)
f,g E C: (T
x
M) :
5
where and
a 11 + B.V (the operator of mean derivative on functions) 0+ - at + .1. 2 a D - at - .1.11 + B. V . Using partial integrations to bring 0 + to 2
act on
gp
and using Fokker-Planck's equation (2.3), we arrive easily
a
1
-0_ ""-at - B.V+V(logP)V + 2 11 · we then get (2.9). FroD. this equation (2.9) we have,
f r om this to the conclusion that
B "" D_Xt taking the time derivative and using the continuity equation (2.7)
Using
- grad div v - grad (v.u)
(2.10)
.
We shall now define the mean acceleration associated with the process
X . To do this we would like to have the concept of mean forward and t backward derivatives of vectors on M . The appropriate definition has i been given by Dohrn and Guerra ( 2 ], (5]. Let F "" F (x,t) be a vec-
M, then the mean forward derivative of
tor field on
P
is defined by (2.11)
0+ F (x , t) '" lim ( 11 t) - 1E ( T X lItto t' where T
+' P is the vector at T , M obtained from the vector y,y "y y+"y PET r1 by Oohrn-Guerra's stochastic parallel transport along the geoY
desics from
y
to
y + lIy . We recall briefly the definition of this
transport, for more details see (2 ], ( 4] and [ 5 l : Let So
s
s1
' be a segment of geodesics on
Ty(So)M . Let
hIt)
M. Let
, t E (0,1], be a curve on
M
Let us transport in a Levi-Civita parallel way getting a vector field Yt(so) {Yt(s)
h(t) , So
B(s) '" : t
G(t)
Yr(so) s
s1
r
Yt(s)
• Let
Yt(S)
G(t), s < s1
s . This then gives, for
Ty,y+lIy F
y
"" P. h tt )
o
. The family of geodesics
=
B(s)
differs from the Levi-Civita displacement of in
such that
y(s )"'G(t) along
s"" So
this is a vector field along ) ,y(s)F
Y(s) , be a vector in
be the geodesic such that
t E (o,1J) is parallel for
B(So) = F . By definition
pi
and F
Let
Y(s) T
with
y(so)'Y(s)
F
by second order terms
y(so)' y + 11y = y(s)
the transport
needed in (2.11). One computes easily ( 2], (19] 0+
+ S.V + 21 £l DR (2.12)
D £lOR '" £l-R
+
13.\7
1
2 £l OR
being the Laplace-de Rham-Kodaira Laplacian on
the Ricci tensor, acting on vectors.
M, R
being
6
Let us now define the
a(xt,t)
a(xt,t)
by
- 1(0 D + 2 +,t)
we get
(2.13)
hence dV
a + U'vu - V'Vv +
at
21
6
DR
(2.14)
u .
Let us also remark that a purely probabilistic description of the process is given by the solution of the stochastic differential equation (in
Ito's sense)
with
W the standard Brownian motion on M . Of course this is not t an intrinsic description, for such see e.g. [5], [53]. Given S we X and get P, hence i3 t satisfying (2.10), (2.14). Moreover, given X t P we can get Sand i3 as mean forward resp.
can, under suitable assumptions, construct and hence
u
and
v
and its distribution
backward derivatives and then get with
a
u, v
defined as mean acceleration of
satisfying (2.10),
(2.14),
X t.
For a class of diffusions diffusions"
X which we shall call here "Newtonian t' (they coincide, under regularity assumptions, with Nelson's
conservative diffusions [4 ]), we shall show that one can recover X t from (2.10), (2.14) and the initial conditions u(x,o), v(x,o). Definition:
A Markov process
X is called a t satisfies "Newton's law in the mean", in the sense
if X t that there exists a positive constant m and a real--valued function V
on
Mx T
such that
7
and such that in addition the corresponding current velocity is a gradient field. Ne then say that toni an diffusion.
a
!ve shall see below that there exist
is the acceleration of the New-:
conservative Newtonian diffusions.
We shall first discuss the forms of their distributions. Let be any function on
Mx T
such that
is only defined modulo functions of
v(x,t) '" grad S (x,t) t
alone). Let
S(x,t)
(clearly
P (x,t)
S
be the
density of the distribution of
X , as above. Under our assumptions t P is smooth and strictly positive; hence log P is well defined. R+'S 1 Setting we have easily e 1 , so that = P, R = 2 log 0 from m a -'ilV and (2.14):
div u (because by taking grad div and
v
from
'
on this and using
grad R
grad lul
2
=
2u.'ilu,
div grad
grad Ivl
2
=
(2.15 ) 6
both u R being gradients, we see that this is just (2.14». Moreover, 60
=
u
grad
o
+
=
2v''ilv,
we have
e
2R
- div (e 2 R 'ilS)
hence
o
R + } div grad S + grad R.grad S From (2.15),
(2.16 )
(2.16) we see that
2.± _ _ .l2 3t -
A ,I,
V
,"
(2.17)
Thus we see that for conservative Newtonian diffusions the probability distribution and all
p(x,t)
t ET
by
of the process p
=
II/!
i 2
,
where
X t
is given, for all
is a solution of the Schro-
dinger equation (2.17), with initial condition 11jJ(x,o)
1
2
and
v
such that
u
1jJ
is a solution of (2.17) and if we write II' = e and v by u = grad R , resp. v grad S ; then
satisfy (2.10) resp.
particular
1jJ(x,o)
gives the initial distribution of the process.
Vice versa if and define
x E M
B
(2.14). From
u
and
v
R+iS u
we can get in
and thus the stochastic equation for a proces
Xt , II/!(x,t)
2 the distribution of which is then for all times 0 (x,t) = if at time t 0 it has distribution 11jJ(x ,o) [2 Moreover, the 1
process statisfies Newton's equation in the mean.
,
8
Remark:
It has recently been shown by Nelson that the stochastic New-
ton law in the definition of conservative Newtonian diffusion can be replaced by a variational principle [ 4]
(see also [30],
[58]).0
We are particularly interested in the case where there exists a station' ary distribution
P(x,t)
=
P(x)
for the process, i.e.
this case we have from the above, that
11ji(x,t)
1
2
,
=0
and hence
In R,
is
independent of t . Assume 1ji satisfies (2.17) with initial condition "'(x,o) (x)+iS(x,o) • Th en we h ave, u s i nq 33R '¥ t 0 e that 31ji lll1jJ + V 1jJ Le . }ll1ji + V 1ji is equivalent with . 1ji t i 3t 2
e
R+iS
[
Hence for all
x
+ (vS)2 + 2ivR.vS + V] e
(VR)2 + such that
eR(x)
R+i S
(2.18)
*0 (2.19 )
llS + 2 vR.vS
o .
(2.20)
This system of coupled partial differential equations with initial con' dition
=
S{x,o)
So(x)
has a solution of the form
S(x,t) = iff
1 '2ll1jJ + V 1jJ
E1jJ
has a solution of the form
R( x j iEt iSo(x) e e e
1ji (x,t)
(2.21)
Et + So(x)
.
Splitting in real and imaginary parts
we get
II R
(vR) 2 + (VS ) 2 + V o
(2.22)
E
o . We set
P
= e 2R
and remark that (2.23) can be written in the form llSO +
i.e.
(2.23)
12 22.· vS P 0
o
(2.24)
*
V (PVS
= 0 for all x such that P 0 0'1e set v VS o o) is just the continuity equation div (Pv). Hence if, l O R iEt lSo 1ji is such that'2ll1jJ + V1jJ = E1jJ has a solution of the form 1jJ = e e e
then (2.24)
then (2.22),
(2.23) are satisfied. Setting
u =
P
in this case
we then have that the system of coupled equations is satisfied. From
9
u
and
v
we can compute
This process has stationary case
for all times
S
and hence the process
p as invariant distribution. Note that in the S
d08s not depend explicitly on
Remark: The case where
VS
= 0
ving for the process
(J o at =
0,
hence
P
t
is the only case where
situation (2.23) is c Le ar Ly satisfied. Setting
X t.
u =
v
= 0
log
. In this
P and sol--
K we find from the continuity equation that t, is stationary. v 0 is equivalent with the sym-·
metry in
(M, pdx) of the Markov semigroup P giving the distribut tion of X [12]. In fact, this is equivalent with time reversal int variance [11].
3. Barriers for Newtonian Diffusion Processes In the case of a symmetric diffusion process a general theory of formation of barriers for the process (non trivial decomposition into time ergodic components) has been given in [13], [16], [7], [10], [28], [51]. Let us take the opportunity to recall here the main facts of this theory. Let
E
L 2 . Let
be a regular Dirichlet form on
M be a locally com-
pact space with a countable base for the topology and let Radon measure on
m
be a
M, strictly positive on every non void open subset.
It is well known that there exists a one-to-one correspondence between submarkovian semigroups on L 2(M; m) and Dirichlet forms (i.e. positive symmetric closed bilinear forms E on L 2(M; m) with the contraction property with
f
#
_
(f v
0)
A
1
Vf for which E(f,f) < + 00). The correspondence is such that if -L is the infinitesimal generator of P t then E (f, f) (L1/2 f, L1/2 f) , where on the right hand side we have the scalar product in L 2(M; m). Moreover, diffusion semigroups are in one-to-one correspondence with local Dirichlet forms, local meaning
E(f,g)
=
0
whenever
f,g
have
disjoint supports. On the other hand, to any regular Dirichlet form (regular meaning that the continuous functions with compact support in the domain of the form i.e.
D(E) n Co(M)
are dense both in
and in (Co(M) ,I: 011=), where El (f,f) '" E(f,f) + (f,f) and l) 11.11 00 denotes the supremum norm) by a construction of Fukushima and Silverstein there is a Hunt process with m-symmetric transition
(D(E) ,E
function precisely Pt ' Pt being the symnetric sernigroup associated with E . X can be taken to be a diffusion (in the sense of having continuous t
10
(X continuous in t E [o,!; J) 1) iff E is local in t addition to being regular. One can show [54] that any regular local
paths,
pX
Dirichlet form can be written as closed extension of the form E(f,f) for some Hilbert space (Jf"
ae,
f being smooth cylinder functions on
v , V o being positive Radon measures with values in the cone of positive selfadjoint operators resp. in JR+ • If
M is an oriented Riemannian manifold and
on
M then the closure of
2
2
L (M;m) xL (M;m),
E(f,f)
=
f
m
is a Radon measure
df(x).df(x) m(dx)
in
M
1 o
first defined on C (M)functions, is a local Dirich
let form. Sufficient conditions for the existence of the closure can be extracted from work by Fukushima
8 J, Albeverio, H¢eghKrohn and
Streit [31] and Rockner and Wielens
9
(the latter reference gives
also a survey of this type of results). In particular, if a density
P
with respect to the volume measure
dx
m(dx)
and if
p
has is
strictly positive on compacts and locally Lipschitz, then there exists only one closed extension of E [33J. Other results yielding closability (but in general not uniqueness) involve e.g. conditions of the type pl/2
p > 0 , dx a.e and that the distributional derivatives of with respect to local coordinates are locally in L 2(dX) , on
any open subset of [9J, [46J.
M whose complement has mmeasure zero ([31], see also E on L 2(M;m) , (M being
For any regular Dirichlet form
again a locally compact second countable Hausdorff space and
m
a
Radon measure strictly positive on non void open sets) one defines the capacity of U by Cap U
inf [ E (f r f) + (f, f) J
the infimum being taken over the set L u of functions f in the domain of E, which satisfy f 1 mae on U (the infimum is taken to be
+=
if
L
u
¢). By standard methods one can then extend the
definition to any subset quet capacity. L unique element
A c M as an outer capacity yielding a Cho-
u being closed and convex for open U there exists a in L u which minimizes E(f,f) + (f,f) , this el-
eU
ement being the equilibrium potential of Cap (U)
U. One has
0 s e
U
s 1
and
11
of
is actually a version of the hitting probability ,
.{"'" is identified with a ?/A/jt-o//.d= pha-se-Laan.si.t.con.; from (2.9) it is clear that, in terms of the free energy densi ty f >fI, there is a non-empty phase-transi 'ti on segment [Xl, X2] . From our present point of view the interest lIes in his method of proof. He used the f'o l.Low i.ng result of Berezin and Sinai [2]: Berezin-Sinai Lemma: In onden. .tha.: a non empty pha-1e-uan.ji.t.Lon -1egment 1lJi.;th ch.emcccu. potenti.a1 '" 0= exi..j;/: to//. -some (/> , it .L.j -1utf-i-uent ;that tu//. -some. I> > 0 and 1';> 0 and all -1utf-i-uentJ..i/t. , (3.3)
In other words, a first-order phase-transItion can be detected as a violation of the law of large numbers in the grand canonical ensemble. (Griffiths [5] showed that for sufficiently large and .(. the mean-value of X is less than S from which
t
1-
28 (3.3) certainly follows.) Dobrushin gave a proof of the Berezin-Sinai Lemma which is simpler than the one given in [2]. He deduced it from the following Dobrushin Lemma:
we tiov e
Fo». S;'O and
Wlc:lX
-
mllx
\
/
(3.4)
»he».« The function .x: I-t is convex; in the particular c as e of the La t t i.ce-cgas model it satisfies the symmetry Gondition (3.5)
i t follows that the right-hand side of (3.4) is equal
implies that 14.
to is a phase-transition segment.
- gli) .
Then (3.3)
The Bose-Einstein Phase-Transition
The traditional description of Bose-Einstein condensation is this: in a system of non-interacting hosons in thermal equilibrium the excited states saturate at a critical value Pc. of the density; when the density P is increased beyond this value the excess p - Pc. goes into the zero-energy state. The phenomenon is sometimes described as 'condensation in momentum spaGe'. The condensate has zero entropy as well as zero energy, and so makes no contribution to the pressure. Consequently, the pressuredensity isotherm has a flat part: the pressure increases with increasinp, density for densities below p... and thereafter remai n s constant. There is a basi c d i.f'f icu I ty which we have to face if we attempt a rigorous proof of these statements: a phasetransition manifests itself sharply in the mathematical behaviour of thermodynamic functions onl yin the bulk 1 i mi t, but in thi s 1 i mi t there is no uni que preci se formulation of the zero-energy state. For non-interacting particles in a box of finite volur.1e, the single-particle energy-levels are well-defined and there is a unique ground state; as the volume increases, every energy-level tends to zero; for the infinite system, the si.ngle-parti c Ie energy-spectrum is a continuum fi 11 ing the ha If-line but there are no eigenstates. There are two good candidates for the concept of macroscopic occupa t i.on of the zero-energy state: mac.aoocopi.c. occupation of- the ;ywund -di;a;te is said to occur when the number of particles in the ground state becomes proportional to the volume; conden-dai;,Lon is said to occur when the number of particles whose energy levels lie in an arbitrarily small band above zero becomes proportional to the volume. ObViously, the first implies the second. However, the second can occur without the first; this is called conden-dat,Lon. These matters are discussed in [6] where it is proved that there are, in general, two critical densities: there is (>, whi.ch is the density at which singularities in the t.her-modynam ic functions occur, there is Pm which is the rai n i Dum dens i ty for macroscopi c occupation of the ground state. Generali.zed condensation oC'curs whenever p is greater than Pc. macroscopic occupation of the ground state occurs if and only if the weak law of large numbers for the particle number density is violated. As far as we know, the first rigorous proof of the macroscopiC' occupation of the ground state of the Laplacian when the bulk-limit is taken by dilating an arbi.trary star-shaped regi.on was sketched by Kac in 1971; his manuscript remained unpublished until 1977 when it was incorporated in the review by Ziff, Uhlenbeck and Kac [7]. The mathemaUcal details were ied in the thesis of PULE [8] and in the papers of Cannon [9] and LEWIS and [ 10 ] ; the connect i on wi th the work of Araki and Vioods [lJ_] was di.scussed by LEWIS [12]. Kac obtained the limiting distribuUon Kl7C;P> (now known
29
as the Kac d.is t.r-Lbu't lon ) of the particle number- density density p by conputing its Laplace transform:
Xt
= N/IALI at :fixed mean
He found that, when P exceeds P.. , the distribution is exponential; details may be found in [6] where it is shown that, in genera], the distri buti on is infinitely divisible. In the mean-field model of a s y s t.era of interacti.ng bosons, the interaction energy is represented by a tern 11. N7./'2.ll\t l wh i.ch is added to the hamiltonian of the free boson gas, where Q. is a strictly positive constant representing the strength of the interaction. This crude model of a system of interacting bosons is conmonly called the .i..m.pVl.f-ee..t bO-1on [}a-1; it is of interest because the patholOGical aspects of the free boson gas are removed by the mean-field interaction: the grand canonical parti ti.on function converges for aII real values of the chemical potential [4]; the weak law of large numbers holds for the particle nUr7lber density for all values of the chemi.cal potenti.al [13] (see also [14] and [15]). However, it is proved in [16] that generalized condensation persists in the imperfect boson gas: generali.zed condensatIon i.s stable with respect to a r7lean-field perturbation of the free-particle hamiltonIan. §5.
An Extensi.on of Laplace's tlethod__for Integra Is
In t.h i s secti.on, we present a version of Dobrushi n ' s Lemma whi ch holds under condi ti.ons wh i ch are satIsfied by a wide c lass of conti nuous systems in stati s ti.c a l. mechanics, both classical and quantum. do this by means of a version of Laplace's method for i.ntegrals whi ch, un Li.k e the standard treatments (see Copson [17], for example), makes no hypo t.he s i s of d if'f'er-en t i ab i Hty concerni ng the integrand. Lemr7la 1
(Laplace's method)
Lei be a -1equence of- .Lowe//. »emi.s con.cinuou» Loncci.on.s, f", suppose :thai:. on each compae..t :the -1equence {of..i i-1 bounded below and conVVI.[}e-1 unif-oJUrt1!f i:.o f, and :thai:. f( 0 )=0. Let {to" lI\ '" be a of- po(t for For each
e:-'t(Xt£krex)
dktex)
t t. tp .. ,P .. i-Et) sufficiently large, so that
t
Pt.J
- A'l- J
the inequality (6.9) follows in an analagous fashion.
Putting together Lemma 5 and Lemma 6, we have Lemma 7:
U"" n( ) Suppose that " I-i Pc,1) ;: ("'0;1'( f exists, is strictly convex and differentiable except possibly at one point fJ; then there exist strictly posi.tive constants A I A1.. (given by (6.5) and (6.7)) such that
1
35
References 1.
2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. 13. 14. 15. 16. 17. 18. 19.
\L Feller, An Introduction to Pr-obab i.I ity Theory and its ApplIcations
(two vo1umes) (Wi1ey and Sons, New York, 1966). LA. Berezi.n and Ya. G. Sinai., Trudy l1osk. Mat. Obshch. 17, 197-212 (1967). R.L. Dobrushi.n, Proc. Fi.fth Berkeley Symposi.um, III, 73-87 (1967). D. Ruel.le, Stati.stical Mechani.cs: Rigorous Results (New York, Amsterdam, Benjamin, 1969). R.B. Griffiths, Phys. Rev. 136, 437-439 (1964). r1. van den Berg, J. T. Lewi s and J. V. Pul.e, J. Math. Anal. and Appl. (in press). R.!·\' Zi.ff, G.E. Uhlenbeck and M. Kac, Physics Reports 32C, 169-248 (1977). J.V. D. Phil. Thesis, (Oxford, 1972). J.T. Cannon, Commun. Math. Phys. 29, 89-104 (1973). J.T. Lewis and J.V. Commun. flath. Phys. 36, 1-18 (1974). H. Araki. and E.J. J. Math. Phys. 4, 637 (1963). J •T. Lewis, The Free Boson Gas, tn llathemati.cs of Contemporary Phy s i cs, Ee. R.F. Streater, (Academic Press, London, 1972). E.B. Davies, Commun. Math. Phys. 28, 69 (1972). M. Fannes and A. Verbeure, Phys. Lett. 76A, 31 (1980). E. Buffet and J. V. J. Ilaths. Phys. 24, 1608 (1983). 11. van den Berg, J. T. Lewis and P. de Smedt, J. Stat. Phys. (to appear i.n Dec., 1984). E.T. Copson, Asymptotic Expansions, (C.U.P., Cambridge, 1965). D. Ruelle, Lectures in Theoretical Phy s i.c s , Ed. W.E. Br-i t t.i.n and VI.R. Chappell, VI, 73, (Univ. of Colorado Press, Boulder, 1964). R.B. Griffiths, J. r,lath. Phys. 5, 1215 (1964).
THE UNIQUENESS OF REGULAR DLR MEASURES FOR CERTAIN ONE-DIMENSIONAL SPIN SYSTEMS F. Papangelou Random fields are not always uniquely determined by their specifications, i.e. their systems of conditional distributions.
A general result is presented here,
giving sufficient conditions under which a one-dimensional specification admits at most one random field (up to equivalence in distribution), within a specified class of such fields.
In a specific application this result implies that certain one-
dimensional spin systems with long range interaction admit unique regular DLR measures, regardless of "temperature". The present article has been written in the informal style of the talk given at the conference and begins with a brief introduction to some known results from the literature, selected for their immediate relevance to our subject. line of the proofs of new results is given in the last section;
A sketchy out-
for full details
the reader is referred to [9J and [lOJ. §l.
The background. Although the systems to be considered here are ordinary stochastic processes
parametrized by the integers, it is useful to think of them as one-dimensional random fields. ables
Xi
elements
By a d-dimensional random field we mean a collection of random varidefined on the same probability space
i
of the d-dimensional lattice
as "sites", then
Xi
and parametrized by the If one thinks of the elements of
is some variable (say a spin) associated with site
i.
The speeifieation of a random field is the system of conditional distributions QA(B\r;.,jiA) J
(B
E
where
P((X.)., 1
1EJl
E
slx,
J
= r;.,jiA) J
A ranges over the finite subsets of
Whenever we refer
to a specification, we will assume that the versions of all conditional distributions in it are regular probability kernels satisfying in a strict sense the obvious consistency conditions implied by the definition. In statistical mechanics and other areas, where random fields have been widely employed, it often happens that the nature of the inter-relationship between the Xi's
is best described in terms of the specification.
However, a specification
does not always uniquely determine an (unconditional) distribution for a random field and, given a particular specification, two natural questions immediately arise: (i) is there a random field admitting the specification?
(ii) if there is such a
random field, is it unique up to equivalence in distribution?
If we identify a
random field with its distribution, these questions reduce to the existence and
37 d
uniqueness of a probability measure admitting the specification.
IT,
on the product
a-field
of
,
If there are two or more probability measures admit-
ting the same specification, we say that phase tpansition occurs. This is in summary the approach adopted by Dobrushin and by Lanford and Ruelle in the late sixties, in their treatment of equilibrium states for infinite Gibbs systems. Before describing the particular class of statistical mechanical models we will be concerned with, we state three early theorems due to Dobrushin, which give sufficient conditions for uniqueness of the measure admitted.
The theorems are not
stated in full generality here but only in a form which is relevant to our subsequent discussion.
In particular, we only consider translation invariant specifica-
tions. Suppose that a translation invariant specification is given. Take d d where o is the null element of and for k E "{O} define
II-II
where
denotes total variation and the supremum is taken over all pairs J J"'O
1.1
{O},
A
that differ only at site
Theorem ([4]).
I
If
k#o
random field.
Pk
t,
and
if either
) j iA'
or I:; j
'1;j
1:;. =
J
'1;.
J
t+1,t+2, ... ,t+k
for
then
Then the specification admits a unique random field. If the true state space of a random field arising from the given specification is a compact subset of
and the conditional probabilities have continuous densiIn the unbounded case
ties, the conditions of Theorem 1.3 are not unreasonable.
however they are too severe, even for Markovian specifications. So
0,
=
a
G(n)dn
function, and analogously for
Q{O}(.
the distribution
2.1
Theorem
cr 2 J) .
where 0,
n
n
h, x, x, z, z
where«
denotes absolute continuity and
Qh(·lx,z)
and
Q\·lx,z),
r
h
(·1 x , z ;
i,z)
is the "overlap" of
i.. e. their greatest lower bound in the space of measures.
For the remaining two hypotheses IV and V we postulate the existence of certain special sets
c
v = 1,2, ..... ;
=
1,2, •••
below. IV. A
+-
T
For any and any
v
E >
1
0,
any (metrically) compact set
there exists
1
such that
with the properties described +
C c T,
any compact set
41
Q[s,tJ(lR i- s x M (j-i+l) x JR t- j !x,z) > 1u
whenever
s < i
j < t,
E Mv(k) ward sequence (The case k = 0 V. j.l z
1
z
is a forward sequence, say
for some
k
E:
>
0,
Z k
=
that for some
k
E C, E Mv(k)
X
E A).
any compact sets
there exists an integer
Cr,l'
k
x
->-
C c T
and
\1
3.1 E:
>
on
i.
IB(JR7L»
and any
v
=
C••• ,
1, If
are such
,1;-1)
M)k)
( .•.• (-k-2.(-k-l) E A
A = [s,tJ,
«.)
nMj.l(t-s+l) Ix,z)
1- -\1
Ql s ' t J (C.) nMj.l(t-s+l) x,z)
[
Q s t; J (M (t-s+l) Ix.i:)
J CM (t-s+l) 1x z ) u
j
j
there exist a compact set
for all
EA.
u
I
- 1
an integer
there is a set
II
is a random field admitted by the specification and let
s > 0
such that for every pair of integers -+
- 0
and
>- 2k + 2
with
and r
[s+l t.r-L] t-s-ll - ' (JR x,z; x,z) >-
e
The important point in this assertion is (x,z) EG and (x,z) EG • s,t s,t that e does not depend on t - s , One begins by restricting ( ••• ,X and (X ••• ) t, s) -n)
= 0
=
IT(D)
follows from this, once it is shown that two
tame probability measures admitted by the specification cannot be mutually singular. The translation invariance of
IT
is a consequence of its uniqueness.
Turning now to the spin systems considered in §2 above, suppose (3) holds and define
\P (i)
IJ(k) I,
i
=
1,2, ..• ,
so that
Hi)
0,
there is
n
L
i=l for arbitrary
sand
£.
i
1,2, ... ,£)
?;
l-E
This implies condition (i) of Theorem 2.2 and can also
be shown to imply condition (iii).
The implication (iii)
=>
(i) is proved similarly.
It is worth mentioning at this point De Masi's result ([3J) that translation invariant DLR measures for the spin systems considered here are tempered. assertion is contained in the implication (iii)
,=>
This
(i).
Finally the uniqueness part of Theorem 2.3 is an application of Theorem 3.2 one can show that the specification (1) satisfies hypotheses I-V. References 1.
Benfatto, G., Presutti, E., Pulvirenti, M.: DLR measures for one-dimensional harmonic systems, Z. Wahrscheinlichkeitstheorie und verw. Gebiete 305-312 (1978).
2.
Cassandro, M., Olivieri, E., Pellegrinotti, A., Presutti, E.: Existence and uniqueness of DLR measures for unbounded spin systems, Z. Wahrscheinlichkeitstheorie und verw. 313-334 (1978).
3.
De Masi, A.: One-dimensional DLR invariant measures are regular, Phys. 43-50 (1979).
4.
Dobrushin, R.L.: The description of a random field by means of conditional probabilities and conditions of its regularity (in Russian), Teor. Verojatnost. i Primenen. 13, 201-229 (1968). (English transl.: Theor. Probability Appl. 12, 197-224 (1968».
5.
Dobrushin, R.L.: Prescribing a system of random variables by conditional distributions (in Russian), Teor. Verojatnost. i Premenen. 15, 469-497 (1970). (English transl.: Theor. Probability Appl. l2, 458-486 (1970».
6.
Kesten, H.: Existence and uniqueness of countable one-dimensional Markov 557-569 (1976). random fields, Ann. Probability
Comm. Math.
44 7.
Lebowitz, J.L., Presutti, E.: Statistical mechanics of systems of unbounded spins, Comm. Math. Phys. 195-218 (1976). Erratum: ibid. 78, 151 (1980).
8.
Papangelou, F.: Stationary one-dimensional Markov random fields with a continuous state space. In "Probability, Statistics and Analysis" (ed. J.F.C. Kingman and G.E.H. Reuter), London Math. Soc. Lecture Note Series 21, 199-218. Cambridge: Cambridge University Press 1983.
9.
Papangelou, F.: On the absence of phase transition in one-dimensional random fields. (I) Sufficient conditions. Submitted.
10.
Papangelou, F.: On the absence of phase transition In one-dimensional random fields. (II) Superstable spin systems. Submitted.
11.
Ruelle, D.: Superstable interactions in classical statistical mechanics, Comm. Math. Phys. 18, 127-159 (1970).
12.
Ruelle, D.: Probability estimates for continuous spin systems, Phys. 50, 189-194 (1976).
Comm. Math.
GENERALISED WEYL OPERATORS by and
R L Hudson Mathematics Department University of Nottingham University Park Nottingham NG7 2RD Abstract
K R Parthasarathy Indian Statistical Institute 7, S.J.S. Sansanwal Marg New Delhi 110016 India
Using the quantum Ito's formula of [5J we construct operators satisfying a
generalisation of the Weyl commutation relations, in which scalar-valued test functions are replaced by operator-valued ones. §l.
Introduction Let
H denote
the Boson Fock space f(L 2[0,=)) over L2[0,=) [2] and for each
f E L2 [0, =) let I/J ( f) be the corresponding exponent ial vector [1J, I/J(f)
in If.
z:
0, f, C2!)
_;
(3!)
The Weyl operators W(fl, f
_1
E
2fQfGH, ... )
L2[0,=) are the unitary operators in H "hose actions
on exponential vectors are 0.1 )
W(f)'l'(g) They satisfy the Weyl relation W(f)W(g) ::: exp i i Im}W(ftg)(f,g EL 2[0,=)).
0.2)
v
Introducing the mutually adjoint annihilation and cr'eat ion operacors a( f),
f)
by
means of their actions on exponential vectors a(f)l/J(g)
= l/J(g)
d a t (f)l/J(g) ::: dtl/J(gttf)
I
t:::O
and noticing that at (f) - a( f) is essentially skew-self-adjoint, we may write t _ W(f) :::-- eltP {a (f) - a ( f) } . Quantum Brownian motion [1,5,6J is the family of operators (At' At t
...
t ? 0)
= a t( X[O,tJ')
The duality transformation [8J is a Hilbert space isomorphism, which we may use to identify the two spaces from H onto the Hilbert space L2(w), where w is Wiener measure, under conjugation by which the self-adjoint operators t
0
become multiplications by the canonical realisation X t ? 0 of Brownian motion. t, In [5] (see also [3,4,7 ) a stochastic calculus is developed for quantum Brownian
46
motion generalising the classical It6 calculus in which integration against dX is replaced by integration against the noncommuting independent stochastic differentials t dA and dA , and in which the integrands are operator valued processes (F(t): t 0) which, in the bounded case which concerns us here, are adapted, in the sense that F(t) EB(H )exI for t
O.
H = f(L 2 [ 0 , t ] ) ,
t
t
Here we follow the notation of [5J, setting
H = f(L 2 [ t , 00 ) ) ,
t
and making the canonical identification H
r
operator, the stochastic integral
=
M( t )
H t
rjI
t
H.
When it exists as a bounded
(dA t F + GdA + Hds )
o
r
is determined by the formula (f\s)F(s) +G(s)g(s)
ds .
(1.3)
o
Now let f:[O,oo)
( be locally square integrable and, for t
0, write
W (f) = W( f ). t t
Combining the relation ( h l > = exp< g , h >
< ( g),
with (1.2), we have t
112-+}, t
t
whence
Comparing with (1.3) we see that the Weyl operators Wt(f), t
0 satisfy the stochastic
differential equation Wo(f)
=
Let t
B(H).
(1. 4)
I,
H(t) be a strongly continuous self-adjoint valued map from [0,00) into
The Dyson expansion [7, Theorem X.59] permits the construction of a family of
unitary operators (Wt(H), t
0) in
H satisfying the (strong sense) ordinary differen-
tial equation WO(H) = I,
t t
(H) = iH(t)W (H). t
If H(.) is adapted, so too is (Wt(H): t Given two such maps Hj and H2, the map t
t H2(t) = W t(Hi)H 2(t)W t(H j)
is also strongly continuous, and we have
(1. 5)
0).
Wt(H) is strongly continuous in t.
47
(1. 6)
Our purpose can now be stated; we shall combine and generalise the constructions of the families Wt(f), Wt(H), establishing the existence, for non-anticipating operator valued functions F and H, with H self-adjoint valued, of an operator valued process (Wt(F,H), t
0) satisfying the generalisation of (1.4) and -of (1.6)
WO(F ,H)
(1. 7)
I,
together with the generalisation of (1.2) and (1.6) -
-
1
1--
-1-
W :: Wt(Fj tF 2, Hj tH 2 - 2i(F jF 2 t(Fj,Hj)Wt(F 2,H2) §2.
F2F j
».
(l.8)
ConstX'uction of Wt(F,H) Let h be a Hilbert space and let FO,HO
L
B(h)
Ho with the operators L0511 and HO!5"1 in B(llo 0)
in B( h (X H) satisfying
)
(2.1 )
dU:: U(-dA FotFodA-(lHot;;LOLo)dt ,
I,
so that the adjoint pX'ocess satisfies
t
t ,;;LOL t )d) ,t dUt = (dA t F O FodAt(iHO O r u.
I,
U (0)
(2.2)
We say that the B(H)-valued adapted process F is simple if there exists an increasing sequence 0
::
to < t)
O.
x
in R-D
70
Hence
N(v) = V (
Log (f v ))
(4Log
=V
(f) )
1 ( "Z Log (fv ) n
n-xx> K.
The second assertion of the corollary is easy and left to the reader. o RemaY'k
a) Consider the two following kinds of stochastic equations Xt = Bt + Jt g(Xs)ds
(3.1 )
o
(3.2)
Xt = Bt +
f veda)
(X)
(v is in M(IR)).
IR
COY'ollaY'y 3.2 implies the following results. Any solution of (3.2) is the
strong limit of a sequence of solutions of equations of the form (3.1). Conversely let (gn) be a bounded sequence in L1 (1R ) and let Xn , n=1,2, ... be the solutions of the corresponding equations (3.1). Then there exist a measure v and n
n
a subsequence (X k) such that X k converges strongly to the process X solution of (3.2). b) As shown in the following simple example, there is no analogue of COY'ollaY'y 3.2 for measures v which only satisfy the weaker assumption
for all
x,
Set:
for all
Let Xn be defined by : = Bt + J vn(da)
(X
n)
n
71
Xn is well-defined
see the remarks after theorem 2.3. We can easily
prove that Xn
where X = UB t, t
X
U being a random variable independent of
B and such
that P(U=I) But
X
is not even a
1
P(U=-I) = "2 .
Narkov
process and thus it
cannot
be solution of an
equation of the form (3.2). We now use theorem 3.1 to obtain a new proof and also a slight general ization of a result due to Rosenkrantz (I 7 ] ). Corollary 3.3 :
and X be a continuous semimartingale such that
Let v be in
t = X0 + Bt +
X
f v(da) IR
La (X) • t
Assume that Xo is in LI(rl) and set Xnt
1
- X 2
n n t
Then Xn converges weakly towards the skew brownian motion of parameter a given by : I-a _ _ c l+a - exp( 2v (IR))
(l-V{{y}) ) l+v({yJ)
(recall that the skew brownian motion with parameter a
is the process
uniquely defined by
Remark :
Rosenkrantz treated the case v(da) = g(a)da. In this case we have J
I-a l+a
exp (-2
J g(x)dx) IR
72
Proof : Define, for each integer n and for any Borel set A
Then
so that Xn has the same law as the process
=
*
Xo + Bt +
J vn(da)
Yn defined by
(yn)
IR
Theorem 3.1 together with the relations
f v (x)
1
n
for all
f v (x)---t exp(-2v n
c(lR))
x< 0 for all
IT
y
x >0
complete the proof of the corollary.
o Remark Corollary 3.3 provides a result of convergence towards the skew brownian
moti on with parameter a, where a of a (a=l
or a=-l)
is in the i nterva 1 ] -1; 1 [ . The extreme va1ues
which correspond to reflecting brownian motion, cannot be
obtained through these methods. However it is possible to state a similar result of convergence towards a reflecting brownian motion. Let (f n) be a sequence in f
n
L1(R) S.t
;;;, 0
with
E
n
73
Let Xn be defined by :
x''t = Bt
+
r
ds.
fn
0
Then it is easy to prove that
xn
(weakly) ) n-> 0,
> 1
for any
r.v.
1 > 0,
(1.c) does not hold for
k
= 1,
even when
to be a stopping time ; again, this may be done by taking
L
1S assumed
'\,
1 = T and letting c"" "". c'
These two results clearly show the importance of the stopping times
'\,
{T 1n c} connection with the study of reflecting Brownian motion; in this Note, we take up
the next natural step, that is the study of = inf{t:
for
(p
t)
P t
a Bessel process, with dimension
in this set-up.
d
2, and we extend 1. Shepp's formula
101
Moreover, with the help of the mutual (local) absolute continuity of the Bessel laws, for dimensions
2 , when the processes start at
form of the total winding of complex (see
a
>
0, the Fourier trans-
'\,
BM around 0, up to
T is obtained c'
formula (2.b.2) below).
This formula (2.b.2) is very similar to D. Kendall's formula (32) in gives the Fourier transform of the total winding around complex
BM
[5J,
which
for the poleseeking
0
stopped when it first hits a circle centered at
O.
In the third paragraph below, a probabilistic explanation is given for this Slmllarity, using the time SUbstitution method, as advocated by D. Williams in the discussion following
D.
Kendall's paper
[5J
([5J,
p.414)
1.
2. An extension of Shepp's formula (l.h) (2.1 )
We consider, on the space
(pt(w)
w(t) ; t
" C(IR+,1R), the process of coordinates " o{p s ; s < t.}
0), and its natural filtration
To any couple of numbers
\)
and the distribution
>
on
0, and
a
>
0, we associate the dimension
(il,+1 !:L \> 2 dx2 2x dx'
a.
a >
0, 11, v
stopping time
>
O.
Acoording to
[13J,
and
[7J,
one has, for any
T
dpA (2.a)
a
i,
on
T+
()(T < 00)
where
(2.2) Confluent hypergeometric functions appear repeatedly ln the main formulae below. We have conformed with the notations and definitions used in Abramowitz Stegun ([lJ, p , 504 and (2.3) The main result in this paper is the following
102
Theorem
For any
[t
(2.b)
where
U > 0, v
0, a
a,c > 0, one has :
0, and
1 +
c
2)1/ 2 2 A = (U + v ,
and
A
M, if
A-U
2
1 A(a +1cIe.· 2 ' A+ ;
)
A(a +1cIe.· A+ 1 , 2 2 ' U, if
a < c
a
2
,
c.
>
In particular, 2
A(a ; u+1 ;
(2.b.1 )
) 2
A(a; u+1 ;
r
EO exp ( -
(2.b.2)
aL'
)
a
Before entering properly into the proof of the theorem, we remark that, if 2
denotes the distribution of the lli - va l ue d r.v.
[log(
+
1+Tc ) ; CT] ' c
then : (2.c) for any
'IT
,b
between
a and c, proving at once the infinite divisibility of
a,c
b
*
under
'IT
a,c
pU a'
'IT
b,c
This is a probabilistic proof (and improvement) of Hartman's result
1T
([4J,]).
a,c 271-2),
asserting that the right-hand side of (2.b.1), r e ap , (2.b.2), is the Laplace transform In
a,
resp.
v
2
of an infinitely divisible distribution on lli+.
We also note that identity (2.c) is probabilistically easier understood after time-changing the Bessel process
(p
t),
It is well-known (cf. D. Williams [12J under
with the inverse [7J)
(T ) of (C
t
t).
that:
pU a'
a real-valued BM U loga'
where
stands here for
starting at
(loga). Using (a.a) , one obtains:
BM, with constant drift
\1,
103
Formula (2.c) now appears as a consequence of the strong Markov property taken at time o(b), for BMll(l ). . oga We now proceed to the proof of the theorem, via two steps. We first prove formula (2.b.1). The following notation will be helpful : (20) =
I
Recall that, for any
e
>
II
'"u (20)
(20) ;
K
0, the processes
'"
2
(I (ep ) exp(-.L t ) t IJ 2 are two law of
t .:':. 0), and
plJ-local martingales, an assertion from which the Laplace transform of the a
T
inf{t : P
c
=
t
c}
under
is easily deduced (c r , J. Kent
a < c, and deduce from (2.e) that
We now suppose that
2
e T", )l Ell ['" I (ec 11rr:« + T ) exp(- -a II c 2 cJ Then, following Shepp's method equality with respect to
e'
de e
l'
IJ
(ea).
[8J, we integrate both 2 -e I 2 . P
sides of the previous .
.e , and obtaln, after the change of varlables
c
(2.g)
where
[t£]).
)-
[(1 +
u (c) p
=
f:
de e
We now use the expansion
J..::.p, 21
u (c) p
J
u (a), p
2 .eP.r (ee ). II
r (20) II
l:
n l r( ll+n+1)'
2
n=O 2
2c)
(2.h)
2
a-1
This proves (2.b.1), as a consequence of (2.g) same method, with
'"
KIJ
, where in the case
to obtain
2 .
a > c, we use the
now replacing
Step 2. The complete formula (2.b) now follows from (2.b.1), using the explicit Radon-Nikodym density formula (2.a) for
T
= T'"c
104
Remark: In fact, formula (2.h) has a long history; it is due to Hankel (c r , Watson
[llJ,
p , 384-394) and is a generalisation of formulae due to Lipschitz,
Weber, and Gegenbauer ; at the beginning of the century, formula frequently used by some physicists (again, see Watson [1 I] 3. Another interpretation
0
(3.1) For any
Opv a
from
>
,
(2.h) has been
p. 385).
total winding for pole-seeking
BM :
0, we introduce a new family of distributions
is the distribution of the
d _ 2(v+1)
dimensional Bessel process, starting
a , with "naive drift" 0, that is the distribution of the JR+-valued diffusion
with infinitesimal generator d
A
- 6) dx'
v
'I'he introduction of a terminology such as "naive drift" seems necessary, in
order to avoid confusion with the diffusion obtained by taking the radial part of a
-valued
BM, started at the origin, with
usually called Bessel process with drift (c r , Shiga - Watanabe
[9J ;
Watanabe
[lOJ
t
(E
; this latter diffusion is
Itt
0
[IJ) .
(3.2) In the course of his mathematical study of Bird Navigation, D. Kendall obtained some remarkable formulae (see formula (32) which the following
1S
easily deduced
for
c'
b.
0, one gets : A(I,-v; 1+2A ; 20a) A(A-V; 1+2A; 20b)·
The comparison of this formula with (2.b.2) implies the following extension of (3.c): (3.e)
c
where a
2
. p2v) (d) (C , a T
40a'
c'
c
2
0pv ) a'
40c' .
The proof given In (3.3) for the identity (3.c) is still valid for (3.e), provided the process started at
(Y in (3.d) now stands for t) 0, with constant drift v.
a real-valued Brownian motion,
107
M. ABRAMOVITZ, I. STEGUN
[2J
M.T. BARLOW, S.D. JACKA, M. YOR
Handbook of Mathematical Functions. New-York - Dover - 1970. Inequalities for a couple of processes stopped at an arbitrary random time. To appear (1983). On the norms of stochastic integrals and other martingales. Duke Math. Journal, vol. 43, nO 4, 697-704
(1976) .
[4J
[5J
[8J
[lOJ
Uniqueness of principal values, complete monotonicity of logarithmic derivatives of principal solutions and oscillation theorems. Math. Ann. 241,257-281 (1979). D. KENDALL
Pole-seeking Brownian Motion and Bird Navigation. Journal of the Royal Statistical Society. Series B, 36, n? 3)P. 365-417, 1974.
J. KENT
Some probabilistic properties of Bessel functions. Ann. Prob. £, 760-770 (1978).
J. PITMAN
Bessel processes and Infinitely divisible laws. In : "Stochastic Integrals". Lecture Notes in Maths 851. Springer (1981) (ed. D. Williams).
L. SHEPP
A first passage problem for the Wiener process. Ann. Math. Stat. 38 (1967), p. 1912-1914.
T. SHIGA, S. WATANABE
Bessel diffusions as a one-parameter family of diffusion processes. £.f.W, 27 (1973), 37-46. On Time Inversion of One-Dimensional Diffusion processes. £eitschrift fUr Wahr. 2l (1975), 115-124.
A treatise on the theory of Bessel functions. Second edition. Cambridge University Press
(1966) .
Path-decomposition and continuity of local time for one-dimensional diffusions, I Proc. London Math. Soc., Ser. 3, 28, 738-768
(1974).
[13J
-
Loi de l'indice du lacet Brownien, et distribution de Hartman-Watson. £.f.W, 53, 71-95 (1980).
DISTRIBUTIONAL APPROXIMATIONS FOR NETWORKS OF QUASIREVERSIBLE QUEUES
P.K. Pollett Department of Mathematical Statistics and Operational Research University College Cardiff CFl lXL Great Britain
ABSTRACT. This paper is concerned with establishing Poisson approximations to flows in general queueing networks.
Bounds
are provided to assess the departure of a given flow from Poisson and these lead to simple criteria for good Poisson approximations.
The class of networks considered here are
those with a countable collection of customer classes and where the service requirement of a customer at a given queue has a general distribution which may depend upon the class of the customer.
KEYWORDS.
Queueing networks, Poisson Approximations.
109 1.
INTRODUCTION.
In a recent paper, Brown and Pollett (1982) exhibited a method for approximating customer flow processes in single class queueing networks with exponential service requirements and servers with state-dependent rates. The distance of customer flows from Poisson processes was estimated using formulas derived by Brown (1982) for general point processes.
It is the purpose of the current exposition to extend their results to a class of quasireversible networks with customers of different classes and associated general service requirements.
Bounds are provided to assess
the degree of deviation of arrival processes from suitably chosen Poisson processes.
Although the arithmetic values of these bounds are of doubtful
practical significance, they are of some theoretical interest and give rise to simple criteria for good Poisson approximations. to fall into three categories: buted customer routing. requirements are exponential
These criteria tend
light traffic, heavy traffic and evenly distri-
However, in contrast to the situation where service (Brown and Pollett (1982», the heavy traffic
approximation seems only to be possible if service effort is distributed evenly among all customers present in a given queue (the server sharing discipline).
In section 2 a standard notation is defined and various preliminary results on queueing networks are collected.
Sections 3 and 4 are devoted
to discussing Poisson approximations to arrival processes in both open and closed networks of symmetric queues.
2.
NOTATION AND PRELIMINARY RESULTS.
Let
N denote
a multiclass network consisting of J queues {1,2, ..•• ,J}
(with J possibly infinite) and a countable set of customer classes,
C.
customers are allowed to enter or leave the network it is said to be open;
If
110
otherwise, there is a fixed number of customers of each class and the network In the open case we suppose that arrivals from
is said to be
outside the network occur as independent Poisson streams, the class c arrival Define for each c in
stream at queue j having a bounded rate of v.(c). J
a
A(c)
=
(Ajk(C»
C
to be the collection of probabilities that
govern internal transitions from queues j to k for customers of class c, and let AJ,O(C)
J
=
1-\ A. (c)
be the probability that after completion of service at
queue j a class c customer leaves the network.
If
N is
closed, AjO(C) is
taken to be zero for each j and c.
In the open case define g(c)
=
(a (c) ,a , .•.. ,aj(c» 2(c) l
to be a vector
with non-negative entries that satisfies
g(c)
(1)
':!(c) + g(c)A(c).
In order that this vector be unique, we assume that it is possible for any class c customer to eventually leave the network either directly or indirectly via some sequence of queues.
The quantity a.(c) may be interpreted as the J
equilibrium arrival rate for class c customers at queue j and will be positive if it is possible for such customers to visit the queue.
In the closed case we suppose that A(c) is irreducible and non-null persistent.
This ensures that there exists a unique (up to a constant multiple)
vector with positive entries that satisfies (2)
-
o Ic) = a(c)A(c)
-
J
and it will be of no loss in generality to assume thatL a.(c) By Chang 1. j=lJ and Laverberg (1974) the quantity aj(c)/ak(c) is the ratio of the class c arrival rates at queues j and k.
111
We suppose that each queue in the network is
J.>ymme-tJUc.
(Kelly (1976»),
that is, each queue j in N operates as follows:
(i)
A total service effort is offered at a rate (units per second) when there are n
J
J
customers present;
j
A proportion y.(t,n.) of this effort is directed to
(ii)
J
J
the customer occupying queue position
when this
customer leaves the queue, customers in t+2, ...• ,n. move into positions t, t+l, •... ,n.-l J
J
respectively; When a customer arrives he chooses to occupy position
(iii)
t in the queue with probability y.
J
previously in
J
customers
.... ,n. move into positions J
t+l,t+2, ..•. ,n.+l respectively. J
For each j in {l,2, •.•. ,J} we assume that n
L v• (t,n)
and for n>O,
J
and
The fact that the same function Yj is used in both (ii)
1.
J
=0
J
and (iii) places some restriction upon the types of possible service discipline. However, it allows service requirements to take a quite general form without making equilibrium analysis unmanagable.
We suppose that successive service
requirements for customers of class c at queue j are i.i.d. random variables with distribution function F. (x) and mean JC
J
Thus, when there are n
j
customers present at queue j the rate at which class c customers are served is
J
(c)
. (n
J
i
J
)
Let e(t) the network
(customers per second).
=
N and
(x
l(t),x2(t),
•••• ,x
J(t»
be a Markov process that describes
that contains enough information for one to deduce the number
of customers in each queue and the classes of each of them.
In particular,
112
when queue j is symmetric we let x . = (n.; x , (1), x , (2) , .... ,x. (n.»
J
x.(t) = (c.(t), z.(t), u.(t»
J
J
J
J
J
J
J
J
J
where
describes the customer in queue position t.
Here c.(t) is the class of the customer z. (t) is his service requirement and J
J
u.(t) is the amcunt of service so far received.
In general
J
a continuous state space.
However, if each of the F. ,
C E
JC
(Cox (1955», for example if F.
JC
will have
C, admit a
is Hyperexponential
or a mixture of Gamma distributions, it is sometimes convenient to let z. (t) J
and u. (t) determine, respectively, the total number of (fictitious) stages of J
service and the number of stages reached.
In this case the state space
will be countable.
For each j in {1,2, .••• ,J} and c in C let a.(c) = J
J
J
the
average amount of service required by class c customers arriving in queue j, and let a
\
J.= !..c E
Ca,J (c),
the total average requirement.
For the closed
network let N(c) be the total number of class c customers and define = ( .... ,N(c), .... ) to be the vector which determines the number of customers of each class in the network. m. class c as N -c j
Denote the vector with m fewer customers of
Define n.(c) to be the number of class c customers at queue J
and let J
I
J
n,(c)=N(c), c e C} j=l J
1;N
denote the state space of
•
The following results summarise some of the important equilibrium properties of the network consisting of symmetric queues. consequence of Theorems 3.7(i) and 3.10 of Kelly (1979).
Lemma 1 is a direct
113
Lemma 1.
For the open multiclass network
N consisting and only
equilibrium distribution exists for
I
n n a. /{ IT
n=O J
r=l J
(r)}
of symmetric queues, an
0,
(11)
If
d
2
then
{oJ
E
Kesten ([35J, Theorem 2) gave the first proof of this fact, a proof that was later shortened by Bretagnolle ([aJ).
Kesten and Bretagnolle also
discuss the more difficult problem of when does {OJ E
{oJ
imply
E
r.
We also have the following alternative characterization of when
THEOREM 4.
if and only if
Suppose that X
A u (x), is bounded.
d
=
1
and that
A > O.
{oJ f
We have
has a strong Feller resolvent whose canonical density, In this case there is a positive constant
c
A
such
that O.
(x)
0
is regular u
Ie
,
that
In this case
= U x (-x)ju x (0).
Bretagno11e classifies the circumstances under which this situation occurs. Theorems 4 and 5 are given, with quite different proofs, in Port and Stone ([43J, pp.207-210). one sees that function.
u
Ie
(x)
+
U
Ie
Note that in the circumstances of Theorem 4 (-x)
is almost everywhere equal to a continuous
But this is not sufficient to ensure that the lower
semicontinuous function
u
x (x)
is continuous at the origin.
Consider,
for example, a Poisson process with unit drift.
(b)
The comparison problem. Our main result (Theorem 2) yields an immediate solution to the
comparison problem.
144
THEOREM 6.
processes having exponents 1 Re(A + 1/1 (z ) 2
Proof.
A > O.
Suppose that
and
1/1 1
1/1
Let
The finiteness of
X 2
respectively.
2
0(1) Re(A +
and
Xl
1
(II z
1/Il(Z»
II +
be two Levy If
00)
implies that of
so the result
follows from Theorem 2. The result has been obtained, with progressive weakening of the assumptions, by Orey ([41J), Kanda (r3lJ) and Hawkes ([18J).
The point
here is the very simplicity of the proof that results from our geometric approach. In fact in
then
!
caP
we proved slightly more, namely that if
A A (A) 5 M caP2 (A) 1
consequence that i f
o
< a < 1,
1/I(z) = Izl constants
and
Xl
for all analytic sets X 2
A.
This has the
are two linear stable processes of index
so that the exponents take the form a
{1 - i8sgn(z) t an j-n«}
M 1
and
M 2
-1 5 8 5 1,
with
then there are
such that (13)
M caP (B) 5 ca P (B) 5 M ca P (B) 1 1 2 1 2
for all analytic sets (c)
B.
This answers a quest ion due to Taylor (I 56J) .
The symmetrization problem (Orey). Let
of
a,
X.
X be a Levy process and let Then
Zl
and
Z2
be independent copies
is called the symmetrization of
X.
Orey
conjectured that (14) There are examples of varying degrees of sophistication to show that this
145
inclusion can be strict. Example 1. so that
Let
X(s)
= Pt
X t
= p(s). P(X)
=
t
where
P
is a Poisson process of rate one,
t
One can see that {B:A(B) = O}.
whilst
Thus the one class of sets is smallest possible whilst the other is largest possible. Example 2. y(X)
has
a
Pruitt
=
so that
2 and 3 1 < a < 3
cr 47 J)
has given an example of a subordinator
y(x(s» 2 5'
-
3
= 5'
X
which
I f we choose
y denoting Pruitt's index.
then an intersection argument of the type used in
Hawkes ([17J, the argument deducing Theorem 4 from Theorem 3) can be applied to show that almost all realizations of the range of a linear stable process of index
a
are in
but not in
The inclusion relation (14) follows from Theorem 6 and the observation that 1
(:\ + Re lji) where
§8.
lji
is the exponent of
X
and
ljis
that of
X(s).
ENERGY AND CAPACITY
We now return to the problem that was left unanswered in §4(d). In ([18J) we showed that for Levy processes, and open sets [4 cap(D)J
-1
S I(D) S [cap(D)J
-1
D,
one has
(15)
.
The upper inequality is, as is well known, in fact equality when the process is symmetric. in general.
In ([10J)
The answer is no!
and Rao ask whether this is true This is seen by taking
0 < a < 1
considering the symmetric and increasing stable processes of index
and a
146
and taking varies
D
I(D)
to be the unit interval. can be arbitrarily close to
in (15) cannot be replaced by any number
In 116J we showed that as [2 cap (D)]-I.
e, e
a
Thus the number 4
Thus the Kelvin
< 2.
It would be interesting
principle even fails for linear stable processes. to know the best constants in (15).
For more information on energy see Chung ([8]) and Chung and Rao ([10J).
§9.
WIENER TESTS.
In this section we mention a criterion for a point for a set. theme.
x
to be regular
This criterion is essentially in keeping with our geometric
We also indicate how the comparison results for capacity can be
applied to yield comparison results for regular points. (a)
The classical result.
In 159] Wiener gave a criterion for a point
x
to be a regular boundary point for the Dirichlet problem associated with a given domain. (b)
Brownian motion. is regular for
We shall discuss a probabilistic version of this.
Let K
B t
be the brownian motion in
md .
Then
x
if and only if
I
2 n(d-2)cap [K (x)J n
I
n cap [K (x ) n
n;:o"O
(d
;:0"
3)
(16)
or
n;:o"O where
K (x) n
=
{z
E
K: 2-(n+l) S
(d
Ilz- xii
Io N
Now,
J
0 1 h (N + iy) 1 e -
h(y)e
i8
6
o
ydy
Ydy
I
N i6 -2K8 0 e Ye h(y+2iK)dy
r3
-
r4
- niZres - i
2K 6 0 e- Yh(iy)dy .
J
But h(y+ 2iK) = i
and
where
ex y a i (sin 2K)
h(iy)
g
2 ah(y)
is real-valued.
Therefore,
say,
r
and
3
f
become
4
and Hence, applying Cauchy's Theorem and letting niZres + i
If we now multiply both sides by
(3.2)
Re[(i-
ae K6
i
-a K6 e
_ iae-K6) J:e
N -+
00
,
we get:
a+l {2K -8 0 e yg(y)dy.
and take real parts we obtain: i 6Yh(y)dj 2K sinh K6x 1TX I-a rrx (cOST) sin""2
181 i 1T().
If we bear in mind that
2
e
iCt
cos
1Ta 2
1Ta
+ i sin""2'
then we may
write (3.2) in the form
(3.3)
na
oo
fo(cos ""2 cos 8y sinh K8
+ sin
1TCt
2
sin 8y cosh K8)h(y)dy
=
K sinh K8x 1TX 1-Cl 1TX (cosT) sinT
We can also write (3.3) in the form
(3.4)
f - J_
1Ta 1Ta (cos""2cos8ysinhK8-sinTsin8ycoshK8)(-h(-y»dy
_00
K sinh K8x 1TX 1-a 1TX (cosT) sinT
Hence, to obtain (3.1) we put and add (3.3) and (3.4).
a
= 211
1T '
where
This then gives
11
is defined as above,
1To(X,y),
namely,
zlzl.
1TX a 1TX 1-a sgn(y)sin T (sinh 2k ) (cos 2)
o
1T (x,y)
where
0
czJ = J (cz,oo)/J (cy ;»)
Thus, for some constants
,
E:
and
e.
The fundamental equation (3.9) therefore takes the form:
(4.1)
=
E:
-Rlxl Iyl
e.
The Brownian scaling gives us further information: (4.2)
for
x < O. Y > 0 ,
192
Hence IT(x,y) = cIT(cx,cy), On taking
c
= 1/lxl,
where
u
=
Iy/xl,
l+a
c
p(x,y)
p(cx,cy).
we see that
p(x,Y)
(4.3)
and
Ixl
and
-I-a
=
k(u)
Ixl
p(-I,ly/xl)
-I-a
k(u) ,
p(-l,u) .
Substitution of (4.3) into (4.1) yields: d 2 [(A + U2 +a)k(U)] =
du
2
_ 2ARlxl
Since the left-hand side is a function of 2 d 2+Cl. 2 [(A+u )k(u)] = du and
-3-a-E:
6+a
(A+u
2+a
-2ARu
alone, we must have 6+a
,
o
r
and
00
x= -
°
and
b,
and
p(x,Y) Ixladx