HAL Id: jpa-00246751
https://hal.archives-ouvertes.fr/jpa-00246751
Submitted on 1 Jan 1993
HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.
equations
G. Mahoux, M. Mehta
To cite this version:
G. Mahoux, M. Mehta. Level spacing functions and non linear differential equations. Journal de
Physique I, EDP Sciences, 1993, 3 (3), pp.697-715. �10.1051/jp1:1993100�. �jpa-00246751�
Classification Physics Abstracts
2.30 2.90 5.90
Level spacing functions and
nonlinear differential equations
G. Mahoux and M-L-
Mehta(*)
S-Ph.T-, C-E-
Saclay,(**)
F-91191 Gif-sur,Yvette Cedex, FranceAbs tract The Fredholm determinant of the kernel sin K(z
-v)/K(z -v)
on the finite interval(-t, t),
appears in the theory of random matrices and in some other problems of mathematical physics. In a previous article we studied the functionsS(t), A(t)
andB(t)
related to thisFredholm determinant, and derived relations among and diJferential equations for them. Here
we exploit these relations to deduce the power series at t = 0, and the asymptotic behaviour at t = cc, of the various level spacing functions of the random matrix theory.
1 Introduction.
The
theory
of random matricesprovides
a reasonableunderstanding
of the statisticalproperties
of energy spectra of
heavy
nuclei and ofclassically
chaotic systems. Three ensembles of randommatrices, orthogonal, unitary
andsymplectic,
areextensively studied; they
are characterizedby
a parameterfl taking
values 1, 2 and 4respectively. Among
variousquantities
of interest,one considers
Ep(n, t),
theprobability
that arandomly
chosen interval oflength 2t,
measured in units of the local meanspacing,
containsexactly
neigenvalues [I].
The mainobject
of thispaper is the
decay
of theseprobabilities
when thelength
of the interval goes toinfinity.
The
Ep(n,tl's
areessentially
linear combinations of the derivatives of certain Fredholm determinants [2].Thus,
for theunitary
ensemble(fl
=2),
one hasE2(n,t)
=( (-))
~F(z,t)lz=i, (I.I)
where
~
F(z,t)
=
fl (I zli(t)), (1.2)
I=o
is the Fredholm determinant of the
integral equation lf(x)
=
j~ Il(x,Y)f(Y)
dY.(1.3)
(*)
Member of C-N-R-S,, France(**)
Laboratoire de la Direction des Sciences de la Matibre du Commissariat hl'Energie
Atomiquewith the kernel K
~~~~~~~
~~i~~ ~~
~~'~~
For the
orthogonal
orsymplectic
ensembles, it is more convenient [2] to work with thefollowing
linear combinations of El
(n, t)
~+~~'~~ ~~)~~~i)
+
El (2n
1,t),
>~
~~'~~E-
(n, t)
=
Ei(2n, t)
+Ei(2n
+ 1,t),
n 2 0,(1.6)
so that
E4(n,t)
=iE+(n, 2t)
+E-(n, 2t)1, (1.7)
E+(n,t)
=( (-))~ F+(z,t)lz=i, (1.8)
where
F+(z,t)
are the Fredholm determinants of theintegral equation (1.3)
with the kernelsK+:
1<+(z,
Y) =iK(x,
Y) +K(-z,
Y)i(1.9)
The
integral equation (1.3)
has been known for along
time [3], its solutions are calledspheroidal functions; they
are either even orodd;
the even(odd)
solutionscorrespond
to the kernelK+
(I<-).
Suppressing
thedependence
on z, let us setA(t)
=( (log F+(t)
+
log
F-(t)), (1.10)
B(t)
= ~(log F+(t) log F-(t)) (I-11)
In a
previous
paper [4] we studied these functions and derived thefollowing
relations(A(t)
= 2B~(t), (1.12)
((tA)
= z(S(t)(2, (1.13)
((tB)
= z lie(S(t)~)
,
(l.14) 2~(tB)
= z Im
(S(t)~)
,
(l.15)
where the
complex
valued functionS(t)
=S(z,t)
iscompletely
determinedby
the non-linear differentialequation
~~
~ ~~2)t~~
~~~ ~~~~ ' ~~'~~~(S*
is thecomplex conjugate
ofS), along
with the initial conditionS(0)
= 1.We also
proved that,
as a consequence of the aboveequations, A(t)
andB(t) satisfy
the non linear second order diRerentialequations
(t ~j
+
2~)
+16x~t~ (~) 8~ (t~
+
A)
=0, (1.17)
~ ~ ~
(t
+2@
t +x~tB) 16B~ 1(t
+B)
+4x~t~B2
= 0.(1.18)
~
t
~
Eq.(1.17)
was first derivedby limbo, Miwa,
Mori and Sato [5].Note
that,
as a consequence ofequations (1.12)
and(1.13)
A =
z(S(~ 2tB~. (1.19)
Then,
a careful examination ofequations (1.12)-(1.15)
and(1.19)
shows that any one of the three functionsS,
A and B,completely
determines the two others.Moreover,
each of thesefunctions
satisfy
a non linear diRerentialequation, equations (1.16)-(1.18) respectively.
Note also that if S is a solution of
equation (1.16),
theniS,
-S and -iS are others. This four-folddegeneracy
is of course removedby
the initial condition at t = 0.Changing
S into iSchanges
thesign
of B, andinterchanges F+
and F-.Now,
in order to get rid of as manyclumsy
factors x aspossible
incoming formulae,
we decide from now on to shift from the variables t and z to the new variablesr = xi and
(
=
~~
(l.20)
x
We also define new functions
A(r)
=?A(t)
andB(r)
=
?B(t), (1.21)
and we rename
F+((, r)
andS((, r)
the functionsF+(z,t)
andS(z,t),
which we also writeF+(r)
andS(r)
for short.With these
definitions, equations (1.10)-(1.19)
becomeA(r)
=-) (1°g J~+(r) +1°g J~-(r)) (1.1°')
~(T)
=) (1°g J~+(r)
1°g F-
(r))
,
(I.ll')
)A(r)
=
B~(r), (1.12')
~A(r)
=
( (S(r)(~ r~~(r), (1.19')
) (rB(r))
=
(
Re(S(r)~)
,
(l.14')
~
2rB(r)
=(
Im(S(r)~
,
(l.15')
(r ~~
+ 2~~
+4rB) 482 1(r ~~
+)
+
4r~B2
= 0.
(1.18')
T T
~
T
~
Power series of
F((, r)
andF+((, r)
near r = 0 are known [2].They
are of the form~~~'T)
# 1~f
~~T)~~
~~~
nj ajn
(-4r~)~
,
~
(l.22)
with the coefficients ajn
given
as finite sums(see appendix B).
The
asymptotic
behaviours ofF(2/x, r)
andF+(2/x, r) (I.e.
ofEp(0, t), fl
= 1,2 and4)
forr » I are also known [6]:
1°g?
~r)
=log E~(0 t)
=log
J~+~
r)
+logy (~ r)
,
~i
~~~'~ '
" K'
~°~~~ ~~'~)
~ ~°~(~
~()
~ ~+ ~~(~l)"Cn (r
+()
~,
(l.24)
n=3
1)
~~ 24 ~
4
~°~~
~ ~'(l.25)
Here, C is
3/2
times the derivative of the Riemann zeta function evaluated at -I; its approx- imate value is C = -0.248131716.Two new facts came up
recently. Firstly,
theasymptotic
behaviours ofE+ (n, t)
andE2(n, t)
for a few small values of n were derived
by
a combination of theproperties
ofToeplitz
determi- nants and the diRerentialequation (1.17) ii]. Secondly,
athermodynamic
model of continuous Coulombcharges
wasexactly
solved in the limit oft » I; and a combination of it with equa- tions(1.12)-(1.16)
was used to express the Fredholm determinants as aproduct
of a smooth factor and anoscillating factor,
the laterdepending
on Jacobianelliptic
functions [8]. Here we revisit thesequestions by exploiting equations (1.10')-(1.19').
The
present
paper isorganised
as follows:In section
2,
we derive theTaylor expansions
of the Fredholm determinants at r= 0. In
section
3,
we express the real andimaginary
parts ofS(r)
as well asJt(r)
andB(r)
in terms of the fifth Painlevd transcendents. The functionB(r)
can also beexpressed
in terms of a thirdPainlevd transcendent. In section 4, we derive an
expansion
of various functions in powers of(( 2/x).
In the
asymptotic region
r » I, this method issimpler
and faster than the earlier ones, except for one defect: it leaves undetermined the constants c+, which appear inlog F+((, r)
asintegration
constants. To fixthem,
one must use the initial conditionsF+((, 0)
= 1, a difficultexercise which we leave out here.
One more undetermined constant, in
S((, r),
is calculable once one knows the solution of the standardproblem
ofconnecting
the behaviour at r = 0 of a Painlevd transcendent to itsbehaviour at r = cc, a
problem
which can be solvedby using
the method of "isomonodromic deformations" devisedby
thejapanese
school of M. Sato.Unfortunately,
it has not yet been solved for all solutions of the fifth Painlevdequation,
andparticularly
not for those of interest here. Wehope
to come back to thisquestion
in a near future. Here we deduce the value of the above mentioned undetermined constant from other sources.In
appendix A,
we find that the dominant term ofB(r)
for r » I is a Jacobianelliptic
sine function.
However,
since the initial conditions at r= 0 are not taken into account, two constants remain undetermined. In
appendix
B, we list a few of theexpansion
coefficients ajn ofequation (1.22).
Inappendix C,
wegive asymptotic expansions
of a few coefficient functionsappearing
in section 4. Inappendix D,
we tabulate a few of the cnappearing
inequation (1.24),
and derivenumerically
for these coefficients alarge
nexpansion
which suggests thatthe functions
F+(2 lx, r)
are Borel summable atinfinity.
2. Power series of
F+((,r)
for r «1.In this brief section we use
equations (1.10')-(1.16')
to derive the power seriesexpansions
around r = 0 ofS((, r), B((, r), Jt((, r)
andF+((, r).
We obtainsuccessively
~~<,
~~ = i +<r
+<2 II r~
+<~ It)
~~+
~~ '~~
~~~
~~ ~~~~~
~ i
(r r~
+)(r~
+B((, T)
"(
+~~~ ~~ ~
~~ ~
~~ ~~
~~
(2.2)
+
(~
~(~
+(~)
~~ ~Jt((, r)
~=
( j)~ ~~~~ ~~~ ~)~
~~ ~~~ ~~~
~~(2.3)
J~~
(<, r)
= (r +(r~
~
(~~~
~~25 ~~~
~)~
~~~
~~~~~~
~~~
.~~
~
(( r)
=
(r~
+((T~ ~05~~~
~255~~~~
~(2.6)
~~ i
r
+
~(~r~ ~(~~~
+~05~~~~
~The
integration
constants inlog F+
are determined from the initial conditionsF+((, 0)
= 1.The last three
equations
are in agreement with the knowngeneral expressions
recalled inappendix
B.3. Fifth and third Painlev4 transcendents.
In this section we will show that
(I)
The real part ofS(r)
isexpressed
in terms of a fifth Painlevd transcendent(P5) (here
andin what follows a
prime
denotes the derivative with respect tor),
Y" "
()
+ Y'~ + ~~~~~~
~YY+
))
+7)
+b~ ) ~~~, (3.I)
witho=-fl=1/32,7"0,b=-2,andnearr=0,
(2) The part of
S(r)
is in rms ofa
P5 with thesame lues
y,(r)
= -1 2((r)~'~ 2(r 4((r)~'~ 6(~r~
+O(r~'~). (3.3)
Any
one of the above two P5's can besimply expressed
in terms of the other.(3)
The functionJt(r)
isexpressed
in terms of a P5,Eq.(3.I),
with either a =1/2, fl
= 7 = 0,
b =
8,
and near r = 0,Yai = -<r
<~r~ <~
l<I r~
+°(r~) (3.4)
orwitha=fl=0,7=-4i,b=8,andnearr=0,
y~~ = 4ir + 4
(it I) r~
+16(r~
+O(r~). (3.5)
This latter case was studied in references [5] and
ill].
(4)
The functionB(r) (which
is almost theimaginary
part ofS~ (r))
isexpressed
in terms of aP5, equation (3.I),
with o= =
0, fl
=-1/8,
7" 1; and near r =
0,
Ybi =
(2r)~'~
+ 4r<
+
(2r)~'~
4<~
1°) r~
+°(r"~). (3.6)
It is also
expressed
in terms of a P3Y" "
~ ~
+ ~~ ~
~
+ 7Y~ +
~, (3.7)
witha=fl=-7=b=-I,andnearr=0,
y~~ = + +
<2
))
r +o(r2). (3.8)
To start
with,
considerequation (1.16').
Let us prove that itsgeneral
solution is a homc-graphic
function of a P5.Writing
S =(R
+it) /vf,
where R and I are real, andseparating
the real andimaginary
parts, we getR =
~'~
~,
I =
f~ (3.9)
This
pair
ofequations
expresses R in terms of I andI',
and I in terms of R and R'.Elimination of I
gives
R"
=
~~~ R'~ ~(~
~ +~(R~ r). (3.10)
r r r r
Elimination of R
gives
the sameequation
for I.equation (3.10)
hassingular points
at+@,
andcc. To
displace
them to0,
and cc we set yr =(R+@) /(R-@),
and y;=
(I+jfl/(I- jfl,
then
yr(r)
andyi(r) satisfy
the fifth Painlevdequation (3.I)
with the parameters a =-fl
=
1/32,
7 = 0 and b= -2. Near the
origin
r = 0,S(r)
isgiven by equation (2.I),
whichimplies
that for r <
I, yr(r)
andy;(r)
aregiven by equations (3.2)
and(3.3).
We now turn to
equation (1.17').
Its solution is also ahomographic
function of a P5. Tosee
this,
note that it is almosthomogeneous
in Jt. Therefore let [9]Jt =
e~~,
u =rw', (3.ll)
and for convenience let [10]
~
v = u-1- ~"
(3.12)
u
Then
A'
=-uA/r, (rAl'
=rA'+
A=
-(u I)A, (3.13)
(rA)"
=
rA"
+2A'
=
(u'A
+(u I)A')
="~A. (3.14)
r
Disregarding
the non-zerofactors, equation (I. Ii')
now reads(~ i)2
v~ +
16r~
+ 4r A= 0.
(3.15)
u
The derivative of this last
equation gives
vv' + 16r + 2r ~'~
i
~~~(- l
+
t
+
ii
A= 0.
(3.16)
Eliminating
Jt between the last twoequations,
we getvv'+
16r (v~ +16r~) (-~
+~'~ i)
= 0.
(3.17)
'~ '~ r
Using equation (3.12),
one findsby elementary algebraic manipulations
that v factorizes in the left hand side ofequation (3.17).
The solution v = 0gives
u = I/(I cr),
with c anintegration
constant.
However,
this cannotcorrespond
to the valueofJt(r)
at r = 0. The otherpossibility
~
~'~ ~
~~
~
u
~'
l
~2r ~) ~~~
~~'
~~'~~~or
using equation (3.12)
~ ~
v~
+ii~2'
~~'~~~Equation (3.12)
expresses v in terms of ~ and~',
whileequation (3.19)
expresses ~ in terms ofv and v'.
On
writing
v in terms of ~ e yai, one obtains the Painlevdequation
of the fifthkind, equation (3.I),
with the parameters a =1/2, fl
= 7 = 0 and b= 8. And on
writing
u in terms of v, one obtains~"
v2
~16r2
~'~~~
+
~~~
v2
/)6r2
~~2~
~~~ ~ ~~~~~ ~~'~~~Setting
v=
4irf
we get<" =
fi<'~ [
+
~(<~
i)
4<(<~i). (3.21)
This last
equation
hassingular points
at I, I and cc. Asbefore,
one setsf
=(ya2+1)/(ya2-1),
then ya2 satisfies the fifth Painlevd
equation (3.I)
with the parameters a =fl
=0,
7 = -4i and b= 8. This
equation
has beeninvestigated
in references [5] and[11].
Near r
= 0, Jt is
given by equation (2.3),
whichimplies
that yai and ya2 aregiven by equations (3A)
and(3.5).
Finally,
considerequation (1.18').
We write it asb" + 4b =
@,
b+ 2rB.
(3.22)
To see that this is an
algebraic
transform of a Painlevdtranscendent,
we follow Bureau[12],
and consider thepair
ofequations
~
~b
~~r'
~~ ~' ~~'
~~~~~
This
pair
is likeequations (3.9).
If we eliminate ~b, weget equation (3.22)
and if we eliminate b, we get~"
~(
~~b
2r) ~~ ~b
~2r~
~~~~ ~~~'
~~'~~~To put this last
equation
in the canonical form, let~ =
'~
j~~,
~b =-j, (3.25)
so that
~
~"' 2~a
~
~a
~
~"~
~r2j~
~ ~~" ~~'~~~
This is almost the fifth Painlevd
equation.
To put it in standardform,
taker~/2
as thenew
independent
variable.Eq.(3.26)
transforms into the fifth Painlevdequation (3.I)
witha = b
=
0, fl
=-1/8
and 7= 1.
As b
=
0,
we can transform it into a third Painlevdequation (3.7).
To doit,
we follow Gromak[13],
and consider thepair
ofequations
~~
q'+q~+1'
~~T~2'-~2+1'
~~'~~~Eliminating
q, we getequation (3.26)
for ~a; andeliminating
~a, we get forJ~ the third Painlevd
equation
in its standard form(3.7),
with a =fl
= -7 = b
= -1.
4.
Asymptotic expansions
forlarge positive
r.Any
one of the functionsS((, r), Jt((, r)
orB((, r)
iscompletely
determinedby
the diRerentialequation
itsatisfies,
and initial conditions at r = 0. Inparticular,
itsasymptitic
behaviour at r = cc is alsocompletely
determined. Butderiving
thisasymptotic
behaviour is a difficultproblem,
whichimplies connecting
the twopoints
r = 0 and r= cc. We do not know yet how
to solve it for the one
parameter
classes of solutions definedby
the above functions.So,
what we do is topostulate
aparticular
seriesexpansion
forS((, r)
around r = cc(see (4A) below),
and we find that theresulting
series forlog F+
coincides withDyson's asymptotic
expansion
oflog F+
at(
=
2/x. Then,
weexpand S((, r)
around(
=2/x.
More
precisely,
we will show that, if one writes theTaylor expansions
in(,
around(
=2/x,
of the functions
S,
Jt and B as follows~
C° ~ ~2r "
S((> ~)
"2/~
(° (( p) ~j Sn(T)> (4.1)
~=o
A(<,
T) # 2io ~
Y< )) jj
~~ ~An(T), (4.2)
B((>
T) " 2~
Y(( )) jj
En
(T), (4.3)
~o
~~ ~
where a is a constant,
then,
forr
positive
and » I, the coefficient functionsSn(r), Jtn(r)
andBn(r)
haveasymptotic expansions
in powers of r~~ A few terms of theseexpansions
for low values of n are listed inappendix
C.To derive these
expansions,
we first look for a solution ofEq.(1.16')
of the formr'~jsjr~i,
co(4.4)
j=o
with 7 and
sj's complex
constants. It turns outthat,
up to the four-folddegeneracy
mentioned in theintroduction,
the solution isunique
andgiven by
theright
hand side ofequation (Cl).
From this
particular solution, using successively Eqs.(1.15'), (1.19'), (1.10')
and(I.ll'),
weconstruct the
corresponding
functionsJt,
B andlog
F+. The first two of them aregiven by
equations (C6)
and(Cll).
As for the last one, it coincides with theasymptotic expansion
oflog F+(2/x, r)
foundby Dyson
[6],equation (1.24). Thus,
the aboveparticular
solution isindeed the
asymptotic expansion
ofS(2/x, r).
Let us make two remarks. When
calculating log F+(I, t)
withequations (1.10')
and(I.ll'),
one has to
perform integrations,
andthen,
in order to determine theintegration
constants, onehas to use the initial conditions
log F+(1,0)
= 0. This determination is a difficultproblem,
which
we do not solve here. Its solution has been
given
firstby Dyson, using
thetheory
of inversescattering
[6].As a second remark, note that the
right
hand side ofequation (Cl),
dividedby (I
+I),
has real and
imaginary
parts which arerespectively
even and odd in r, a property which is not sharedby S(2/x, r) (see equation (2.I)).
This is asign
that in thelarge
rasymptotic expansion
ofS(2/x, r), (probably exponentially)
sub-dominant terms are not seen.Next,
letus
expand S((, r)
as followsS((,
T) =) f (( ))
~iRn(T)
+
iIn(r)1, (4.5)
where Rn and In are real functions of r, and let us define the two component column
#n(r)
=(j))() (4.6)
Then, equation (I. lo') implies
that in(r)
satisfies for n > I the linear diRerentialequation
~~~
=
fi4#n
+tin, (4.7)
where rid is the 2 x 2 matrix
fiq~~~ 2RoIj/r Rl/T
ll
Io IT -2RoIo/r
~
l 13
~
l
~ l
~
=
4r2
64r(
2r1(r3 jjr4
,
(4 8)
2r ~
l/r3
~ 16r4 ~ ~ ~4r2 ~ 64r4 ~
and where the
inhomogeneous
term ~bn is a function oflo, Ii, In-
IFor n = I, the
equation
ishomogeneous,
~bi " 0. Twolinearly independent particulir
solutions are
~
(Re Si(T)
~~'~~#t(T)
= e ~
Im
si(T)
'Im
s~(-r)
(4.1°) ii (r)
= e~~~(Re Si(-r)
where
Si(r)
isgiven by equation (C2).
The second one has a sub-dominant behaviour atinfinity
ascompared
with the first one. So we writeji(r)
= 2 ajt(r), (4.ii)
where
a is a constant
which,
once more, has to be determinedby using
the initial condition atthe
origin, Si (0)
=(see equation (2.I)). Although
theproblem
islinear,
it iscomplicated by
the fact that the
equation
involves the solution So of a non linear(Painlevd) equation.
For n
=
2,
one has2RoRiIi
+IOR(
(4 12)
'~2
i -2IoRiIi RoIi
This
inhomogeneous
term increases atinfinity
like e~~.Consequently,
any solution ofequation (4.7)
for n = 2 also increases atinfinity
like e~~ Since the solutions of thehomogeneous equation
increase at most like e~~, the dominant terms in theasymptotic expansion
of#2
are
independent
of the initial condition at theorigin.
Since we are interestedonly
in theexponentially
dominant terms, no new unknown constant appearshere,
and one obtainsj~(~)
= ~~y2~~~ l~e 1~2(T)
~ ~~)
@
Im52(r)
'where
52(r)
isgiven by equation (C3).
The same
phenomenon happens
at each further step n =3,4,..
of theprocedure.
Thisreadily
leads to theexpansion (4.I),
and theexpressions given
inequations (C2)
to(C5).
The calculation of the
asymptotic expansions (4.2), (4.3)
and those inappendix
C is thenstraightforward,
as well as that of the variousprobabilities
defined in the introduction. Note that all theintegration
constants that appear whencalculating G+n,
for n >I,
with the use ofequations (1.10')
and(I. II'),
are irrelevant sub-dominant terms.To find the unknown constant a we
proceed
as follows[14].
It is known [15] that,j+i
I
lj
csj2~~+~P+~'~e~~~~, (4.14)
J.
for
j
finite and t » I.Substituting
it in the dominant termsE+
(n>t)
~
fl
~2>,
E- (n>
t)
~
jf
~2>+1(~ i~~
E+(0, t)
~
l 12J
E-(0,t)
~
l
12j+1
>= >=
gives
~+(~'~)
c~
l(2j)jl,-n~2-n(3n-1)t-n(2n-1)/2~2«ni
~~ ~~~E+(0,t)
J"
~-(~,l)
~
fl(~j
~ ~)j,-n(n+1)~-n(3n+2)~-n(2n+1)/2 ~2«nt
~ ~~)E-(0,t)
~
J*
The
expansions (C16), (C17)-(C24)
andequation (1.8)
will all agree with the above if we choose the constant a as" "
~ (4.18)
With this
value,
all ourexpansions
agree with those ofBasor,
lYacy and Widom [7].From the
expansions
ofSj (r)
listed inappendix C,
we observe that the first terms aregiven by
'~~~
~~l6r~
~~~~~il~)2
~~~~ ~~~~~ ~
~~~)~~(~~~~
~~~~~
+
A (i
+
~llil~
+~~~~~~lllll~~
~°°+
(4.19)
This formula
gives
the correct coefficient of r-kup to k =
j
in the real part, and up to k =2j
in the
imaginary
part.We observe also that
A>
(r)
+ B>(r)
= 2 ReS>(r)
+ O(r~~>)
,
(4.20) Jtj(r) Bj (r)
= 2 ImSj(r)
+ O(r~~i)
,
(4.21)
We checked
equation (4.20)
forj
< 8 andequation (4.21)
forj
< 5.Formulae similar to
equation (4.19)
can be written for the coefficient ofr~k
inG+j(r),
£+j(r)
and£j(r).
5. Conclusion.
Our
objective
was to derive thelarge
tasymptotic
behaviour of theprobabilities Ep(n,t).
Sofar,
we did notexplicitely complete
this program. Toproceed,
one should use thefollowing
equations
El(o,I)
"
E+(°,1), E2(o,I)
"
E+(°>~)E-(°>1), (5'i)
E4(0, t)
=(E+(0, 2t)
+ E-(0, 2t))
,
(5.2)
E+(°,t)
=F+(I,t)
=?+(2/K, r), (5.3)
which first determine the
asymptotic
behaviours ofEp(0,t)
forfl
= 1, 2 and4,
in terms of those ofF+(2/x,r).
These latter aregiven by equation (1.24)
with the cm's tabulated inappendix
D. Then theasymptotic
behaviours ofEp(n,t)
aregiven by Eqs.(1.5)-(1.7), (1.24)
and
(C25)-(C38);
recall that r= xi and
(
=
2z/x.
As
compared
withprevious methods,
ours seems toprovide
moreeasily
successive terms inlarge
tasymptotic expansions.
Wejust
needalgebraic
calculations which can beperformed routinely by
software systems ofsymbolic programming
like MATHEMATICA or AMP. This enables us toinvestigate numerically
forlarge
n, the behaviour of then'~
coefficient of variousasymptotic expansions
in t. Astypical examples,
let us write theasymptotic expansions (Cl), (C6)
and(Cll)
as followsSo(r)
")
~~(Pn
+
i~n) j
,
(5.4)
~0
co
Ao(r)
=)
+~ @
,
(5.5)
n=o
80(r)
=~j ~ ~() (5-6)
r
Then,
forlarge
npn =
iii
"exp
j-c
+) j
+
°(j)1, (5.7)
~"
~~~~
~~~ ~In ~2
~~/3~~
'
~~'~~
~"
~~e
~
~~~ ~
In ~2
~~~ /3
'
~~'~~
bn =
iii
~exp
j-c
+)
+
o(j)), (5.io)
where the constant c is
c = 0.451582705289.
(5.ll)
The coefficients cn of
equation (1.24)
have two similarlarge
nexpansions,
one for even n and another one for odd n~~
~~
~ ~~~~~
~~'~~~k=0
The constant term in the exponential does not
depend
on theparity
of n~)
= -2.144729885849.
(5.13)
while
~)
= -2.763930691.,
~)
= -4.994470.
,
(n even),
~ ~
~/
= +0.930597358.,
Kp = -2.394586.
,
(n odd).
It is not clear
numerically
whether these~)
are rational.
It is to be noticed that the same dominant factor
(n/2e)~(l/n)
appearseverywhere.
The above formulae suggest that(at least)
the functionsSo, Jto,
So and F+ are Borel summable at r = cc. If it is so, their Borel transforms areanalytic
for(r(
> 2.The
problem
ofconnecting
the behaviour at r= 0 and
r = cc of any of the fifth or the third Painlevd transcendents encountered in section 3 remains open.
Appendix
A.A. Solution of
equation (1.18)
forlarge
t.By
acomparison
with the continuummodel,
Dyson showed [8] that for t » I andii
z-j « I,B(t)
isnearly
a Jacobianelliptic
function. We can recover this result as follows. Let us writeEq.(1.18)
asB"
+
4x~B
+~~')
~
=
16B~ (B'
+ ~~
+
x~B~)
(Al)
t t
If we
ignore
the terms inIll,
then~
= +48,
(A2)
or
d
(~)
= +d
(28~) (A3)
On
integration
thisgives
B'~
+ 4x~B~= 4
(B~
+c)~
,
(A4)
with c an
integration
constant. As B'equals
the square root of a fourthdegree polynomial
inB,
B is a Jacobianelliptic
function. Let us look for a solution of the form asn(u, I),
where sn is theelliptic
sinefunction,
a and I are constants, and u is a function oft to be determined.Then
B'~
=
a~u'~(I sn~B)(I l~ sn~B)
j~2 j~2
=
a~u'~
Im I
l~m
,
(A5)
a a
or
B'~
+ 4x~B~4(B~
+ c)~ =a~u'~
4c~ + B~(4x~
8cu'~(
I +l~))
~2 12
+B~
-4(A6)
a
Choose a, u and I so that the coefficients on the
right
hand side ofequation (A6)
are zero;This
gives
u'
=
)
=
~, u'~(i
+i~)
= 4K~ 8C
(A7)
or
'~'
l~l'
~
~ l~l'
~~~~Hence
~~~~
l~l
~~
l~~~l
~
'~~
~~~~
is the
general
solution ofequation (A2);
b and I areintegration
constants. Thisgives
the solution ofequation (Al)
to the doiuinant order.JOURNAL DE PHYSIQUE T 3 N' 3 MARCH '993 25
To calculate the corrections in
Ill, I/t~,
etc. theprocedure
would be to substituteB =
j
sn
()
+b,1)
+~)~~
+~~~
+..(A10)
+ +
and compare coefficients of various powers oft to find
equations
satisfiedby fi(t), f2(t),
etc.Of course, this cannot
give
us the constants I and b in terms of the initial conditionsB(0)
= zand
B'(0)
=2z~,
sinceequation (A.10)
is not valid near t= 0.
Dyson
[8] foundby
a finerstudy
of the continuum model that b = 0 and I variesslowly
with t.
B. Power series of
F+(z,t)
andF(z,t)
for small t.Eq.(1.22)
can be deduced from reference 2 withaj$~
and ajn as follows:aj$~
= 0 ifj
<n(n I), ajj~
= 0 ifj
< n~ and ajn = 0 ifj
<n(n -1)/2.
For values ofj
greater than orequal
to thesethey
aregiven by
thefollowing expressions (in
what followsb(j, k)
isequal
to I ifj
= k and is zero ifj # k)
~~~~ ~~"~ ~~
~0 ~0 ~~ ~~~
l
~'~~~~~
~~~
,l' (
~~~2m,
2t, + 2tk +
Ii 'l~
2tk ' ~~~~
~~"~ ~ ~~
~0 ~0 ~~ ~~~
l
~'~~~~
~~~
,/~ ,l~~
~~~2m;
2t;+ 2tk +
Ii fl 2tk
+
l~
' ~~~~~~" ~~"
~0 ~0 ~~ 1 ~~
"l
~'~~~~~
~' j~~
~~~~~~
2m, ;~+
tk +
~ '~~
~~~~
~~~~
~~~~
In
particular,
ajf~
=-)b(j, 0)
~[2(2j
+1)(2j
+ 1)1]~~,(84)
(+) ~
j
+ 2~i~~~
(2j
+2)! 2j
+(j
+1)(2j +1)
~
(2j
2)1j~ (2i
+)(2~
2i +1)
~~ ~
~ ' ~~~~aji "
-b(I, 0), (86)
aj2 "
-2~J+~[(j
+1)(2j
+1)(2j
+2)1]~~, (87)
~~~
(j
+)(~~~
~(2j
+1)1
~~~~3~~ j~
~~~~
The rational or numerical values of
aj$~
or of ajn canreadily
be calculated for Small value8 ofj
and n.C. Inver8e power series
expansions
ofSn(r), An (r), 8n(r),
etc..For r »
I,
the coefficient functionsSn(r), An (r)
andBn(r)
inequations (4.1)-(4.3)
can beexpanded
in powers of r~~ Asexplained
in section 4 we get(1 +
I) II
85 l1813 14121997 7374679967S°(~)
"
2
@ @ @
223r8 228rio +~~
(~
+~
+
i
+
~j~~~
+~~$~)~~~
+)j (Cl)
3 5 469 2979 1227029 4411073 4184326389
Si(~)
" l +$
+@
+@
+@
+ 223rs + 228r6 + 232r7 +7 213 4465 501667 17524881
+ ~
p
+fi
+@
+@
+ 222rs + 226r6 +~~~
~ 5 63 771 43063 864555 37620327 1053966787
2 ~ " +
p
+fi
+@
+ 21sr4 + 218rs + 222r6 + 22sr7 ~13 387 6815 566999 13720419
+
@
+@
+@
+@
+ 221r6 + 224r? +~~~
Ii 581 14439 1702083 65467383 5738450209
~~~~~ ~
24r ~ 29r~ ~ 213r3 ~ 219r4 ~ 2~3r5 ~ 228r6 ~
+ i
()
+
)
+
$
+
~/(((~
+~~j()(~~
+ ,(C4)
3 63 823 24999 117585 1274445 33179409
54(~)
" l +p
+@
+@
+@
+ 213~s + 21sr6 + 218r7 ++ i
()
+
~i(~s
+I
+
)
+
~j(((~
+,
(C5) A°(T)
" +)
+
~
+
~
+
~
+
$
+
~)~$~
+~i~(~~~
+ '(~~)
~ 5 65 1823 163691 7266843 551317093
i(T)
" 1+ ~~~~$
+@
+@
+ 219~4 + 223~s + 228~6 + ,A~(r)
= 1+)
+
)
+
)
+
(((((
+(((((~
+~((((~~
+,
(C8)
~ Ii 581 14455 1704963 65581207 5747809249
3(T)
" +$
+@
+ 213r3 + 219r4 + 223rs + 228~6 + , ~~~~~~~~~
~IT
~~~2
~
~~3
~
~~~
~
~~$
~~~~~~~
~ '~~~~~
Bo(r)
=) ) /
~/()(~ ~((()((~ ~~((()((~~
,(Cl i) Bi(r)
= 1+)
+
)
+
$
+
((()(
+~((()(~
+~~/()i~
+,
(C12)
5 61 745 41579 839407 36598265
~~~~~
82(T)
" I +p
+fi
+@
+@
+ 218~5 + 222~6 +Ii 581 14423 1699203 65353559 5729123937
~~~~~
83(T)
" +$
+@
+@
+ 219~4 + 223~s + 228~6 +3 63 823 12499 117577 5097387
84(T)
" I +p
+@
+@
+ 211~4 + 213~s + 217~6 +(~l~)
Similarly,
let us writelog F+((, r)
as followslog J~+(<, r)
=i~ £ a < )) jj
~~ ~g+n(r), (c16)
Then
first, -G+o(r)
coincide with theright
hand-side ofEq.(1.24),
except that the constantterms remain unknown. We list the coefficients cn up to n = 25 in
appendix
D.The
G+n(r)
for n = 1, 2, 3, have theexpansions
~ 7 205 5305
682963 28971265
~
2835864953
~~~~~
+~ ~ ~
16r ~ 29r2 ~ 213r3 ~ 219r4 ~ 223r5 228r6 ~ '
~ l 7 l19 1485 90647 1773137 ~
80288703
~
~~~~
+2 T "
§
+$
+@
+@
+@
+ 219~s 223~67 287 22475 926961 35417889 9421927049
~~~~
~+~~~~
3 ~ 16r ~
29r2 ~ 3 x 213r3 ~ 219r4 ~ 223r5 ~ 3 x 228r6 '
~ l 7 21 l159 36589 344937
7512339
~~~~~
+4 T "
j
+$
+@
+@
+@
+ 216~s + 217~6 + '~ l
~
19 1069 39293 7099315 378600725~~~~~
~~ ~
8r ~ 16r ~ 29r2 ~ 213r3 ~ 2"r4 ~ 223r5 ~ '
~ l
(1
19 691 13725 122637530927109
~~~~~
~~ ~
(8~)2
~ ~ ifi~ ~~8~2 ~ ~ll~3 ~ ~16~4 ~ ~19~5 ~ '
~~~ ~
(8r)3
l(1
3 ~ 16r19 ~ 174329r2 ~ 3 236615 15023537 773134357 ~~~~~x 213r3 ~ 219r4 ~ 223r5 ~ '
~~~ ~
(8r)4
l(1
4 ~ 16r19 ~ 26r2263 ~ 1342721°r3 ~ 686667214r4 ~ 9145995216r5 ~ ~~~~~For n > I we write
Then
9 281 7443 997307 43169895 4350267853
£1(~)
" l +$
+@
+@
+ 219~4 + 223rs + 228r6 +(~27)
15 485 9327 846251 22426869
£2(~)
"p
+@
+@
+@
+ 218~s + 221~6 + ,(~28)
75 8865 515409 126126891 8338717077
£3(T)
"@
+@
+ 217rs + 221r6 + 227r7 + 231~8 + ,(~29)
3 l17 6063 271683 l1613831 248599305
£4(T)
"@
+@
+@
+ 224r9 + 227~io + 229~ii +(~~°)
£+i(r)
= i +)
+
fl
+
fl
+
)~()(~
+
~~j($~~
+~~(@i~~
+,