Probability Distribution of The Free Energy of The Continuum Directed Random Polymer in 1 + 1 Dimensions - Amir - 2010 - Communications On Pure and Applied Mathematics - Wiley Online Library PDF

Probability Distribution of the Free Energy
of the Continuum Directed Random Polymer

in 1 C1 Dimensions
GIDEON AMIR
University of Toronto
IVAN CORWIN
Courant Institute
AND
JEREMY QUASTEL
Abstract
We consider the solution of the stochastic heat equation
d
T
Z =
1
2
d
2
A
Z Z

W
with delta function initial condition
Z(T = 0. X) =
A=0
whose logarithm, with appropriate normalization, is the free energy of the con-
tinuum directed polymer, or the Hopf-Cole solution of the Kardar-Parisi-Zhang
equation with narrow wedge initial conditions.
We obtain explicit formulas for the one-dimensional marginal distributions,
the crossover distributions, which interpolate between a standard Gaussian dis-
tribution (small time) and the GUE Tracy-Widom distribution (large time).
The proof is via a rigorous steepest-descent analysis of the Tracy-Widom
formula for the asymmetric simple exclusion process with antishock initial data,
which is shown to converge to the continuum equations in an appropriate weakly
asymmetric limit. The limit also describes the crossover behavior between the
symmetric and asymmetric exclusion processes. 2010 Wiley Periodicals, Inc.
1 Introduction
1.1 KPZ/ Stochastic Heat Equation / Continuum Random Polymer
Despite its popularity as perhaps the default model of stochastic growth of a
one-dimensional interface, we are still far from a satisfactory theory of the Kardar-
Parisi-Zhang (KPZ) equation,
(1.1) d
T
h =
1
2
(d
A
h)
2
1
2
d
2
A
h

W
Communications on Pure and Applied Mathematics, Vol. LXIV, 04660537 (2011)
2010 Wiley Periodicals, Inc.
CONTINUUM RANDOM POLYMER 467
where

W (T. X)
1
is Gaussian space-time white noise,
1

W (T. X)

W (S. Y )| = (T S)(Y X).
The reason is that, even for nice initial data, the solution at time T > 0 will look
locally like a Brownian motion in X. Hence the nonlinear term is ill-dened. Nat-
urally, one expects that an appropriate Wick ordering of the nonlinearity can lead
to well-dened solutions. However, numerous attempts have led to nonphysical
answers [9]. By a physical answer one means that the solution should be close to
discrete growth models. In particular, for a large class of initial data, the solution
h(T. X) should look like
(1.2) h(T. X) ~ C(T ) T
1{3
(X)
where C(T ) is deterministic and where the statistics of ts into various universal-
ity classes depending on the regime of initial data one is looking at. More precisely,
one expects that the variance scales as
(1.3) Var(h(T. X)) ~ CT
2{3
.
The scaling exponent is the result of extensive Monte Carlo simulations and a few
theoretical arguments [13, 17, 18, 37].
The correct interpretation of (1.1) appears to be that of [3], where h(T. X) is
simply dened by the Hopf-Cole transform:
(1.4) h(T. X) = log Z(T. X)
where Z(T. X) is the well-dened [39] solution of the stochastic heat equation,
(1.5) d
T
Z =
1
2
d
2
A
Z Z

W .
Recently [1] proved upper and lower bounds of the type (1.3) for this Hopf-Cole
solution h of KPZ dened through (1.4), in the equilibrium regime, corresponding
to starting (1.1) with a two-sided Brownian motion. Strictly speaking, this is not
an equilibrium solution for KPZ, but for the stochastic Burgers equation
d
T
u =
1
2
d
A
u
2
1
2
d
2
A
u d
A

W .
formally satised by its derivative u(T. X) = d
A
h(T. X). See also [28] for a
similar bound for the free energy of a particular discrete polymer model.
In this article, we will be interested in a very different regime, far from equilib-
rium. It is most convenient to state in terms of the stochastic heat equation (1.5)
for which we will have as initial condition a delta function,
(1.6) Z(T = 0. X) =
A=0
.
1
We attempt to use capital letters for all variables (such as X and T ) on the macroscopic level of
the stochastic PDEs and polymers. Lowercase letters (such as . and t ) will denote WASEP variables,
the microscopic discretization of these SPDEs.
468 G. AMIR, I. CORWIN, AND J. QUASTEL
This initial condition is natural for the interpretation in terms of random polymers,
where it corresponds to the point-to-point free energy. The free energy of the con-
tinuum directed random polymer in 1 1 dimensions is
(1.7) F(T. X) = log 1
T,A
_
: exp:
_
_
T
0
W (t. b(t ))Jt

__
where 1
T,A
denotes expectation over the Brownian bridge b(t ), 0 _ t _ T , with
b(0) = 0 and b(T ) = X. The expectation of the Wick ordered exponential : exp:
is dened using the n-step probability densities
t
1
,...,t
n
(.
1
. . . . . .
n
) of the bridge
in terms of a series of multiple It integrals,
(1.8) 1
T,A
_
: exp:
_
_
T
0
W (t. b(t ))Jt

__
=
o
n=0
_
Z
n
(T)
_
R
n
(1)
n
t
1
,...,t
n
(.
1
. . . . . .
n
)W (Jt
1
J.
1
) W (Jt
n
J.
n
).
where ^
n
(T ) = {(t
1
. . . . . t
n
) : 0 _ t
1
_ _ t
n
_ T ]. Note that the series is
convergent in L
2
(W ), as one can check that
_
Z
n
(T)
_
R
n
2
t
1
,...,t
n
(.
1
. . . . . .
n
)Jt
1
J.
1
Jt
n
J.
n
_ C(n)
-1{2
and hence the square of the norm,
o
n=0
_
Z
n
(T)
_
R
n
2
t
1
,...,t
n
(.
1
. . . . . .
n
)Jt
1
J.
1
Jt
n
J.
n
.
is nite. Let
(1.9) (T. X) =
1
_
2T
e
-A
2
{2T
denote the heat kernel. Then we have
(1.10) Z(T. X) = (T. X) exp{F(T. X)].
as can be seen by writing the integral equation for Z(T. X),
(1.11) Z(T. X) = (T. X)
_
T
0
_
o
-o
(T S. X Y )Z(S. Y )W (JY. JS)
and then iterating. The factor (T. X) in (1.10) represents the difference between
conditioning on the bridge going to X, as in (1.8), and having a delta function
initial condition, as in (1.6). The initial condition corresponds to
F(0. X) = 0. X R.
In terms of KPZ (1.1), there is no precise mathematical statement of the initial
conditions; what one sees as T _ 0 is a narrowing parabola. In the physics
literature this is referred as the narrow wedge initial conditions.
We can now state our main result, which is an exact formula for the probabil-
ity distribution for the free energy of the continuum directed random polymer in
1+1 dimensions, or, equivalently, the one-point distribution for the stochastic heat
equation with delta initial condition, or the KPZ equation with narrow wedge initial
conditions.
For a function o(t ), dene the operator 1
c
through its kernel,
(1.12) 1
c
(.. ,) =
_
o
-o
o(t ) Ai(. t ) Ai(, t )Jt.
where Ai(.) =
1
t
_
o
0
cos(
1
3
t
3
.t )Jt is the Airy function.
THEOREM 1.1 The crossover distributions,
(1.13) J
T
(s)
def
= 1
_
F(T. X)
T
4
_ s
_
.
are given by the following equivalent formulas:
(1) The crossover Airy kernel formula,
(1.14) J
T
(s) =
_
C
J j
j
e
-
det(1 1
c
T, z J
)
1
2
(k
1
T
o,o)
.
where

C is dened in Denition 1.12 and 1
c
T, z J
is as above with
(1.15) o
T,
(t ) =
j
j e
-k
T
t
and
a = a(s) = s log
_
2T and k
T
= 2
-1{3
T
1{3
.
Alternatively, if
_
z is dened by taking the branch cut of the logarithm on the
negative real axis, then
J
T
(s) =
_
C
J j
j
e
-
det(1

1
c
T, z J
)
1
2
(-o,o)
. (1.16)
1
c
T, z J
(.. ,) =
o
T,
(. s)1
Ai
o
T,
(, s). (1.17)
where 1
Ai
(.. ,) is the Airy kernel, i.e., 1
Ai
= 1
c
with o(t ) = 1
0,o)
(t ).
(2) The Gumbel convolution formula,
J
T
(s) = 1
_
o
-o
G(r) (a r)Jr.
where G(r) is given by G(r) = e
-e
and where
(r) = k
-1
T
det(1 1
c
T
) tr((1 1
c
T
)
-1
P
Ai
).
where the operators 1
c
T
and P
Ai
act on 1
2
(k
-1
T
r. o) and are given by their ker-
nels with
P
Ai
(.. ,) = Ai(.) Ai(,). o
T
(t ) =
1
1 e
-k
T
t
.
For o
T
above, the integral in (1.12) should be intepreted as a principal value in-
tegral. The operator 1
c
T
contains a Hilbert transform of the product of Airy
functions which can be partially computed with the result that
1
c
T
(.. ,) =
_
o
-o
o
T
(t ) Ai(. t ) Ai(, t )Jt k
-1
T
G
(x-,){2
_
. ,
2
_
where
o
T
(t ) =
1
1 e
-k
T
t

1
k
T
t
.
G
o
(.) =
1
2
3{2
_
o
0
sin
_
.

3
12

o
2

t
4
_
_
J.
(1.18)
(3) The cosecant kernel formula,
J
T
(s) =
_
C
e
-
det(1 1
csc
o
)
1
2
(
I
)
)
J j
j
.
where the contour

C, the contour

I
)
, and the operator 1
csc
o
are dened in Deni-
tion 1.12.
The proof of the theorem relies on the explicit limit calculation for the weakly
asymmetric simple exclusion process (WASEP) contained in Theorem 1.10, as well
as the relationship between WASEP and the stochastic heat equation stated in The-
orem 1.14. Combining these two theorems proves the cosecant kernel formula.
The alternative versions of the formula are proved in Section 4.
We also have the following representation for the Fredholm determinant in-
volved in the above theorem. One should compare this result to the formula for
the GUE Tracy-Widom distribution given in terms of the Painlev II equation (see
[31, 32] or the discussion of Section 5.2).
PROPOSITION 1.2 Let o
T,
be as in (1.15). Then
J
2
Jr
2
log det(1 1
c
T, z J
)
1
2
(i,o)
=
_
o
-o
o
t
T,
(t )q
2
t
(r)Jt.
det(1 1
c
T, z J
)
1
2
(i,o)
= exp
_
_
o
i
(. r)
_
o
-o
o
t
T,
(t )q
2
t
(.)Jt J.
_
.
where
J
2
Jr
2
q
t
(r) =
_
r t 2
_
o
-o
o
t
T,
(t )q
2
t
(r)Jt
_
q
t
(r)
with q
t
(r) ~ Ai(t r) as r o and where o
t
T,
(t ) is the derivative of the
function in (1.15).
This proposition is proved in Section 5.2 and follows from a more general theory
developed in Section 5 about a class of generalized integrable integral operators.
It is not hard to show that lim
x-o
J
T
(s) = 1 from the formulas in Theo-
rem 1.1, but we do not at the present time know how to show directly from the
determinental formulas that lim
x--o
J
T
(s) = 0, or even that J
T
is nondecreas-
ing in s. However, for each T , F(T. X) is an almost surely nite random variable,
and hence we know from denition (1.13) that J
T
is indeed a nondegenerate dis-
tribution function.
The formulas in Theorem 1.1 suggest that in the limit as T goes to innity, under
T
1{3
scaling, we recover the celebrated J
GUE
distribution (sometimes written as
J
2
). This is the GUE Tracy-Widom distribution, i.e., the limiting distribution of
the scaled and centered largest eigenvalue in the Gaussian unitary ensemble.
COROLLARY 1.3 The following holds: lim
T,o
J
T
(T
1{3
s) = J
GUE
(2
1{3
s).
This is most easily seen from the cosecant kernel formula for J
T
(s). Formally,
as T goes to innity, the kernel 1
csc
o
behaves as 1
csc
T
1/3
x
, and making a change
of variables to remove the T from the exponential argument of the kernel, this
approaches the Airy kernel on a complex contour, as given in [35, eq. (33)]. The
full proof is given in Section 6.1.
An inspection of the formula for J
T
given in Theorem 1.1 immediately reveals
that there is no dependence on X. In fact, one can check directly from (1.8) that
the following holds:
PROPOSITION 1.4 For each T _ 0, F(T. X) is stationary in X.
This is simply because the Brownian bridge transition probabilities are afne trans-
formations of each other. Performing the change of variables, the white noise re-
mains invariant in distribution. The following conjecture is thus natural:
Conjecture 1.5. For each xed T > 0, as T ,o,
2
1{3
T
-1{3
_
F(T. T
2{3
X)
T
4
_
A
2
(X)
where A
2
(X) is the Airy
2
process (see [23]).
Unfortunately, the very recent extensions of the Tracy-Widom formula for ASEP
(1.31) to multipoint distributions [36] appear not to be conducive to the asymptotic
analysis necessary to obtain this conjecture following the program of this article.
Corollary 1.3 immediately implies the convergence of one-point distributions:
COROLLARY 1.6 The following holds:
lim
T,o
1
_
F(T. T
2{3
X)
T
4
T
1{3
_ s
_
= J
GUE
(2
1{3
s).
It is elementary to add a temperature
-1
into the model. Let
F
(T. X) = log 1
T,A
_
: exp:
_
_
T
0
W (t. b(t ))Jt

__
.
The corresponding function Z
(T. X) = (T. X) exp{F
(T. X)] is the solution

of d
T
Z
=
1
2
d
2
A
Z

W Z
with Z
(0. X) =
0
(X) and hence
Z
(T. X)
distr.
=
2
Z(
4
T.
2
X).
giving the relationship
~ T
1{4
.
between the time T and the temperature
-1
. Now, just as in (1.10) we dene
F
(T. X) in terms of Z
(T. X) and (T. X). From this we see that

F
(T. X)
distr.
= F(
4
T.
2
X).
Hence the following result about the low temperature limit is, just like Corol-
lary 1.3, a consequence of Theorem 1.1:
COROLLARY 1.7 For each xed X R and T > 0,
lim
-o
1
_
F
(T.
2{3
T
2{3
X)

4
T
4
4{3
T
1{3
_ s
_
= J
GUE
(2
1{3
s).
Now we turn to the behavior as T or _ 0.
PROPOSITION 1.8 As T
4
_ 0,
2
1{2
-1{4
-1
T
-1{4
F
(T. X)
converges in distribution to a standard Gaussian.
This proposition is proved in Section 6.2. As an application, with = 1 the
above theorem shows that
lim
T_0
J
T
(2
-1{2
1{4
T
1{4
s) =
_
x
-o
e
-x
2
{2
_
2
J..
Proposition 1.8 and Corollary 1.3 show that, under appropriate scalings, the family
of distributions J
T
cross over from the Gaussian distribution for small T to the
GUE Tracy-Widom distribution for large T .
The main physical prediction (1.3) is based on the exact computation [17],
(1.19) lim
T-o
T
-1
log 1Z
n
(T. 0)| =
1
4
n(n
2
1).
which can be performed rigorously [2] by expanding the Feynman-Kac formula
(1.7) for Z(T. 0) into an expectation over n independent copies (replicas) of the
Brownian bridge. In the physics literature, the computation is done by noting that
the correlation functions
(1.20) 1Z(T. X
1
) Z(T. X
n
)|
can be computed using the Bethe ansatz [21] for a system of particles on the line
interacting via an attractive delta function potential. Equation (1.19) suggests, but
does not imply, the scaling (1.3). The problem is that the moments in (1.19) grow
far too quickly to uniquely determine the underlying distribution. It is interesting
to note that the Tracy-Widom formula for ASEP (1.31), which is our main tool, is
also based on the same idea that hard core interacting systems in one dimension
can be rigorously solved via the Bethe ansatz. H. Spohn has pointed out (personal
communication, 2010), however, that the analogy is at best partial because the
interaction is attractive in the case of the -Bose gas.
The probability distribution for the free energy of the continuum directed ran-
dom polymer, as well as for the solution to the stochastic heat equation and the
KPZ equation, has been a subject of interest for many years, with a large physics
literature (see, for example, [8, 12, 19] and references therein.) The reason that
we can now compute the distribution is because of the exact formula of Tracy
and Widom for the asymmetric simple exclusion process (ASEP) with step initial
condition. Once we observe that the weakly asymmetric simple exlusion process
(WASEP) with these initial conditions converges to the solution of the stochastic
heat equation with delta initial conditions, the calculation of the probability dis-
tribution boils down to a careful asymptotic analysis of the Tracy-Widom ASEP
formula. This connection is made in Theorem 1.14, and the WASEP asymptotic
analysis is recorded by Theorem 1.10.
Remark 1.9. During the preparation of this article, we learned that T. Sasamoto and
H. Spohn [25, 26, 27] independently obtained a formula equivalent to (1.28) for the
distribution function J
T
. They also use a steepest-descent analysis on the Tracy-
Widom ASEP formula. Note that their argument is at the level of formal asymp-
totics of operator kernels and they have not attempted a mathematical proof. Very
recently two groups of physicists [8, 12] have successfully employed the Bethe
ansatz for the attractive -Bose gas and the replica trick to rederive the formulas for
J
T
. These methods are nonrigorous, employing divergent series. However, they
suggest a deeper relationship between the work of Tracy and Widom for ASEP and
the traditional methods of the Bethe ansatz for the -Bose gas.
Outline
There are three main results in this paper. The rst pertains to the KPZ/ stochas-
tic heat equation / continuum directed polymer and is contained in the theorems and
corollaries above in Section 1.1. The proof of the equivalence of the formulas of
Theorem 1.1 is given in Section 4. The Painlev-II-like formula of Proposition 1.2
is proved in Section 5.2 along with the formulation of a general theory about a class
of generalized integrable integral operators. The other results of the above section
are proved in Section 6. The second result is about the WASEP. In Section 1.2
we introduce the uctuation scaling theory of the ASEP and motivate the second
main result, which is contained in Section 1.3. The Tracy-Widom ASEP formula
is reviewed in Section 1.5, and then a formal explanation of the result is given in
Section 1.6. A full proof of this result is contained in Section 2 and its various
subsections. The third result is about the connection between the rst (stochastic
heat equation, KPZ equation, and directed random polymer) and second (WASEP)
and is stated in Section 1.4 and proved in Section 3.
1.2 ASEP Scaling Theory
The simple exclusion process with parameters . q _ 0 (such that q = 1)
is a continuous-time Markov process on the discrete lattice Z with state space
{0. 1]
Z
. The 1s are thought of as particles and the 0s as holes. The dynamics
for this process are given as follows: Each particle has an independent exponential
alarm clock that rings at rate 1. When the alarm goes off the particle ips a coin
and with probability attempts to jump one site to the right and with probability
q attempts to jump one site to the left. If there is a particle at the destination, the
jump is suppressed and the alarm is reset (see [22] for a rigorous construction of
the process). If q = 1 and = 0, this process is the totally asymmetric simple
exclusion process (TASEP); if q > it is the asymmetric simple exclusion process
(ASEP); and if q = , it is the symmetric simple exclusion process (SSEP). Fi-
nally, if we introduce a parameter into the model, we can let q go to zero with
that parameter, and then this class of processes is known as the weakly asymmetric
simple exclusion process (WASEP). It is the WASEP that is of central interest to
us. ASEP is often thought of as a discretization of KPZ (for the height function)
or stochastic Burgers (for the particle density). For WASEP the connection can be
made precise (see Sections 1.4 and 3).
There are many ways to initialize these exclusion processes (such as station-
ary, at, and two-sided Bernoulli) analogous to the various initial conditions for
KPZ/ stochastic Burgers. We consider a very simple initial condition known as the
step initial condition where every positive integer lattice site (i.e., {1. 2. . . .]) is ini-
tially occupied by a particle and every other site is empty. Associated to the ASEP
are occupation variables j(t. .) that equal 1 if there is a particle at position . at
time t and 0 otherwise. From these we dene j = 2j 1, which takes the values
1, and dene the height function for WASEP with asymmetry ; = q by
(1.21) h
;
(t. .) =
_
_
2N(t )
0~,_x
j(t. ,). . > 0.
2N(t ). . = 0.
2N(t )
x~,_0
j(t. ,). . < 0.
where N(t ) is equal to the net number of particles that crossed from the site 1
to the site 0 in time t . Since we are dealing with step initial conditions, h
;
is
initially given by (connecting the points with slope 1 lines) h
;
(0. .) = [.[. It is
easy to show that because of step initial conditions, the following three events are
equivalent:
{h
;
(t. .) _ 2m .] = {
J
;
(t. .) _ m] = {x
;
(t. m) _ .]
where x
;
(t. m) is the location at time t of the particle that started at m > 0 and
where

J
;
(t. .) is a random variable which records the number of particles that
started to the right of the origin at time 0 and ended to the left or at . at time t . For
this particular initial condition

J
;
(t. .) = J
;
(t. .) . 0 where J
;
(t. .) is the
usual time-integrated current which measures the signed number of particles that
cross the bond (.. . 1) up to time t (positive sign for jumps from . 1 to . and
negative for jumps from . to . 1). The ; throughout emphasizes the strength of
the asymmetry.
In the case of the ASEP (q > , ; (0. 1)) and the TASEP (q = 1, = 0,
; = 1) there is a well-developed uctuation theory for the height function. We
briey review this, since it motivates the time / space / uctuation scale we will use
throughout the paper, and also since we are ultimately interested in understanding
the transition in behavior from WASEP to ASEP.
The following result was proved for ; = 1 (TASEP) by Johansson [15] and for
0 < ; < 1 (ASEP) by Tracy and Widom [35]:
lim
t -o
1
_
h
;
(
t
;
. 0)
1
2
t
t
1{3
_ s
_
= J
GUE
(2
1{3
s).
In the case of TASEP, the one-point distribution limit has been extended to a
process-level limit. Consider a time t , a space scale of order t
2{3
, and a uctu-
ation scale of order t
1{3
. Then, as t goes to innity, the spatial uctuation process,
scaled by t
1{3
, converges to the Airy
2
process (see [6, 10] for this result for TASEP,
[16] for DTASEP, and [23] for the closely related PNG model). Specically, for
m _ 1 and real numbers .
1
. . . . . .
n
and s
1
. . . . . s
n
,
lim
t -o
1
_
h
;
(t. .
k
t
2{3
) _
1
2
t
_
.
2
k
2
s
k
_
t
1{3
. k {1. . . . . m]
_
=
1(A
2
(.
k
) _ 2
1{3
s
k
. k {1. . . . . m])
where A
2
is the Airy
2
process (see, for example, [6, 10]) and has one-point mar-
ginals J
GUE
. In [16], it is proved that this process has a continuous version and
that (for DTASEP) the above convergence can be strengthened due to tightness.
Notice that in order to get this process limit, we needed to deal with the parabolic
curvature of the height function above the origin by including (.
2
k
,2 s
k
) rather
than just s
k
. In fact, if one were to replace t by t T for some xed T , then the
parabola would become .
2
k
,(2T ). We shall see that this parabola comes up again
soon.
An important takeaway from the result above is the relationship between the
exponents for time, space, and uctuationstheir 3 : 2 : 1 ratio. It is only with
this ratio that we encounter a nontrivial limiting spatial process. For the purposes
of this paper, it is more convenient for us to introduce a parameter c that goes to
zero, instead of the parameter t , which goes to innity.
Keeping in mind the 3 : 2 : 1 ratio of time, space, and uctuations, we dene
scaling variables
t = c
-3{2
T. . = c
-1
X.
where T > 0 and X R. With these variables the height function uctuations
around the origin are written as
c
1{2
_
h
;
_
t
;
. .
_
1
2
t
_
.
Motivated by the relationship we will establish in Section 1.4, we are interested in
studying the Hopf-Cole transformation of the height function uctuations given by
exp
c
1{2
_
h
;
_
t
;
. .
_
1
2
t
__
.
When T = 0 we would like this transformed object to become, in some sense,
a delta function at X = 0. Plugging in T = 0 we see that the height function is
given by [c
-1
X[ and so the exponential becomes exp{c
-1{2
[X[]. If we introduce
a factor of c
-1{2
,2 in front of this, then the total integral in X is 1, and this does
approach a delta function as c goes to zero. Thus we consider
(1.22)
c
-1{2
2
exp
c
1{2
_
h
;
_
t
;
. .
_
1
2
t
__
.
As we shall explain in Section 1.3, the correct scaling of ; to see the crossover
behavior between ASEP and SSEP is ; = bc
1{2
. We can set b = 1, as other
values of b can be recovered by scaling. This corresponds with setting
; = c
1{2
. =
1
2

1
2
c
1{2
. q =
1
2

1
2
c
1{2
.
Under this scaling, the WASEP is related to the KPZ equation and stochastic heat
equation. To help facilitate this connection, dene
v
t
= q 2
_
q =
1
2
c
1
S
c
2
O(c
3
). (1.23)
z
t
=
1
2
log
_
q
]
_
= c
1{2
1
3
c
3{2
O(c
5{2
). (1.24)
and the discrete Hopf-Cole transformed height function
(1.25) 7
t
(T. X) =
1
2
c
-1{2
exp
z
t
h
;
_
t
;
. .
_
v
t
c
-1{2
t
_
.
Observe that this differs from the expression in (1.22) only to second order in c.
This second-order difference, however, introduces a shift of T,4, which we will
see now. Note that the same factor appears in [3]. With the connection to the
polymer-free energy in mind, write
7
t
(T. X) = (T. X) exp{J
t
(T. X)].
where (T. X) is the heat kernel dened in (1.9). This implies that the eld should
be dened by
J
t
(T. X) = log
_
c
-1{2
2
_
z
t
h
;
_
t
;
. .
_
v
t
c
-1{2
t
X
2
2T
log
_
2T .
We are interested in understanding the behavior of 1(J
t
(T. X) _ s) as c goes to
zero. This probability can be translated into a probability for the height function,
the current, and nally the position of a tagged particle:
(1.26)
1
_
J
t
(T. X)
T
4
_ s
_
= 1
_
log
_
c
-1{2
2
_
z
t
h
;
_
t
;
. .
_
v
t
c
-1{2
t
X
2
2T
log
_
2T
T
4
_ s
_
= 1
_
h
;
_
t
;
. .
_
_
z
-1
t
_
s log
_
2T log
_
c
-1{2
2
_
X
2
2T
v
t
c
-1{2
t
T
4
__
= 1
_
h
;
_
t
;
. .
_
_ c
-1{2
_
a log
_
c
-1{2
2
_
X
2
2T
_
t
2
_
= 1
_
J
;
_
t
;
. .
_
_ m
_
= 1
_
x
;
_
t
;
. m
_
_ .
_
.
where m is dened as
m =
1
2
_
c
-1{2
_
a log
_
c
-1{2
2
_
X
2
2T
_
1
2
t .
_
.
a = s log
_
2T .
(1.27)
1.3 WASEP Crossover Regime
We now turn to the question of how ; should vary with c. The simplest heuristic
argument is to use the KPZ equation
d
T
h
;
=
;
2
(d
A
h
;
)
2
1
2
d
2
A
h
;

W
as a proxy for its discretization ASEP and rescale
h
t,;
(t. .) = c
1{2
h
;
(t ,;. .)
to obtain
d
t
h
t,;
=
1
2
(d
x
h
t,;
)
2
c
1{2
;
-1
2
d
2
x
h
t,;
c
1{4
;
-1{2

W .
from which we conclude that we want ; = bc
1{2
for some b (0. o). We expect
Gaussian behavior as b _ 0 and J
GUE
behavior as b , o. On the other hand, a
simple rescaling reduces everything to the case b = 1. Thus it sufces to consider
; := c
1{2
.
From now on we will assume that ; = c
1{2
unless we state explicitly otherwise.
In particular, J
t
(T. X) should be considered with respect to ; as dened above.
The following theorem is proved in Section 2 though an informative but nonrig-
orous derivation is given in Section 1.6.
THEOREM 1.10 For all s R, T > 0, and X R we have the following conver-
gence:
J
T
(s) : = lim
t-0
1
_
J
t
(T. X)
T
4
_ s
_
=
_
C
e
-
det(1 1
csc
o
)
1
2
(
I
)
)
J j
j
.
(1.28)
where a = a(s) is given as in the statement of Theorem 1.1 and where the contour
C, the contour

I
)
, and the operator 1
csc
o
is dened below in Denition 1.12.
Remark 1.11. The limiting distribution function J
T
(s) above is, a priori, unrelated
to the crossover distribution function (notated suggestively as J
T
(s) too) dened
in Theorem 1.1, which pertains to KPZ, the stochastic heat equation, etc., and not
to WASEP. Theorem 1.14 below, however, establishes that these two distribution
function denitions are, in fact, equivalent.
DEFINITION 1.12 The contour

C is dened as
C = {e
i0
]
2
_0_
3
2
L {. i ]
x>0
.
The contours

I
)
and

I
(
are dened as
I
)
=
_
c
3
2
i r : r (o. o)
_
.
I
(
=
_
c
3
2
i r : r (o. o)
_
.
where the constant c
3
is dened henceforth as
c
3
= 2
-4{3
.
The kernel 1
csc
o
acts on the function space 1
2
(
I
)
) through its kernel,
1
csc
o
( j. j
t
) =
_
I
(
e
-
T
3
(
(
3
- )
03
)2
1/3
o(
(- )
0
)
_
2
1{3
_
o
-o
je
-2
1/3
t (
(- )
0
)
e
t
j
Jt
_
J
j
.
(1.29)
Remark 1.13. It is very important to observe that our choice of contours for

and
j
t
ensure that Re(2
1{3
(
j
t
)) =
1
2
. This ensures that the integral in t above
converges for all

and j
t
. In fact, the convergence holds as long as we keep
Re(2
1{3
(
j
t
)) in a closed subset of (0. 1). The inner integral in (1.29) can be
evaluated and we nd the following equivalent expression:
1
csc
o
( j. j
t
) =
_
I
(
e
-
T
3
(
(
3
- )
03
)2
1/3
o(
(- )
0
)
2
1{3
( j)
-2
1/3
(
(- )
0
)
sin(2
1{3
(
j
t
))
J
j
.
This serves as an analytic extension of the rst kernel to a larger domain of j, j
t
,
and

. We do not, however, make use of this analytic extension and simply record
it as a matter of interest.
1.4 Connection between WASEP and the Stochastic Heat Equation
We now state a result about the convergence of the 7
t
(T. X) from (1.25) to
the solution Z(T. X) of the stochastic heat equation (1.5) with delta initial data
(1.6). First we take the opportunity to state (1.5) precisely: W (T ), T _ 0, is the
cylindrical Wiener process, i.e., the continuous Gaussian process taking values in
H
-1{2-
loc
(R) =
_
~-1{2
H
loc
(R) with
1(. W (T ))([. W (S))| = min(T. S)(. [)
for any . [ C
o
c
(R), the smooth functions with compact support in R. Here
H
loc
(R), < 0, consists of distributions such that for any C
o
c
(R), is in
the standard Sobolev space H
-
(R), i.e., the dual of H
(R) under the 1

2
pairing.
H
-
(R) is the closure of C
o
c
(R) under the norm
_
o
-o
(1 [t [
-2
)[

(t )[
2
Jt
where

denotes the Fourier transform. The distributional time derivative

W (T. X)
is the space-time white noise,
1

W (T. X)

W (S. Y )| = (T S)(Y X).
Note the mild abuse of notation for the sake of clarity; we write

W (T. X) even
though it is a distribution on (T. X) 0. o)Ras opposed to a classical function
of T and X. Let F(T ), T _ 0, be the natural ltration, i.e., the smallest o-eld
with respect to which W (S) is measurable for all 0 _ S _ T .
The stochastic heat equation is then shorthand for its integrated version (1.11)
where the stochastic integral is interpreted in the It sense [39], so that, in particu-
lar, if (T. X) is any nonanticipating integrand,
1
___
T
0
_
o
-o
(S. Y )W (JY. JS)
_
2
_
= 1
___
T
0
_
o
-o
2
(S. Y
_
JY JS
_
.
The awkward notation is inherited from stochastic partial differential equations:
W is for the (cylindrical) Wiener process,

W is for white noise, and stochastic
integrals are taken with respect to white noise W (JY. JS).
Note that the solution can be written explicitly as a series of multiple Wiener
integrals. With X
0
= 0 and X
n1
= X,
(1.30) Z(T. X) =
o
n=0
(1)
n
_
Z
0
n
(T)
_
R
n
n
i=0
(T
i1
T
i
. X
i1
X
i
)
n
i=1
W (JT
i
JX
i
)
where ^
t
n
(T ) = {(T
1
. . . . . T
n
) : 0 = T
0
_ T
1
_ _ T
n
_ T
n1
= T ].
Returning now to the WASEP, the random functions 7
t
(T. X) from (1.25) have
discontinuities both in space and in time. If desired, one can linearly interpolate in
space so that they become a jump process taking values in the space of continuous
functions. But it does not really make things easier. The key point is that the
jumps are small, so we use instead the space D
u
(0. o): D
u
(R)) where D
u
refers
to right-continuous paths with left limits with the topology of uniform convergence
on compact sets.
Let P
t
denote the probability measure on D
u
(0. o): D
u
(R)) corresponding
to the process 7
t
(T. X).
THEOREM 1.14 P
t
, c (0.
1
4
), are a tight family of measures and the unique
limit point is supported on C(0. o): C(R)) and corresponds to solution (1.30) of
the stochastic heat equation (1.5) with delta function initial conditions (1.6).
In particular, for each xed X, T , and s,
lim
t_0
1(J
t
(T. X) _ s) = 1(F(T. X) _ s).
The result is motivated by, but does not follow directly from, the results of [3].
This is because of the delta function initial conditions and the consequent differ-
ence in the scaling. It requires a certain amount of work to show that their basic
computations are applicable to the present case. This is done in Section 3.
1.5 Tracy-Widom Step Initial Condition ASEP Formula
Due to the process-level convergence of WASEP to the stochastic heat equa-
tion, exact information about WASEP can be, with care, translated into informa-
tion about the stochastic heat equation. Until recently very little exact information
was known about ASEP or WASEP. The work of Tracy and Widom in the past few
years, however, has changed the situation signicantly. The key tool in determin-
ing the limit as c goes to zero of 1(J
t
(T. X)
T
4
_ s) is their exact formula
for the transition probability for a tagged particle in ASEP started from step initial
conditions. This formula was stated in [35] in the form below and was developed
in the three papers [33, 34, 35]. We will apply it to the last line of (1.26) to give us
an exact formula for 1(J
t
(T. X)
T
4
_ s).
Recall that x
;
(t. m) is the location at time t of the particle that started at m > 0.
Consider q > such that q = 1 and let ; = q and t = ,q. For m > 0,
t _ 0, and . Z, it is shown in [35] that
(1.31) 1(x(;
-1
t. m) _ .) =
_
S
:
C
Jj
j
o
k=0
(1 jt
k
) det(1 jJ
t,n,x,
)
1
2
(I
)
)
where S
r
C is a circle centered at zero of radius strictly between t and 1, and where
the kernel of the Fredholm determinant (see Section 2.3) is given by
(1.32) J
t,n,x,
(j. j
t
) =
_
I
(
exp{
t,n,x
()
t,n,x
(j
t
)]
(j. ,j
t
)
j
t
( j)
J
where j and j
t
are on I
)
, a circle centered at zero of radius j strictly between t
and 1, and the integral is on I
(
, a circle centered at zero of radius strictly between
1 and jt
-1
(so as to ensure that [,j[ (1. t
-1
)), and where, for xed ,
(j. z) =
o
k=-o
t
k
1 t
k
j
z
k
.
t,n,x
() =
t,n,x
()
t,n,x
().
t,n,x
() = . log(1 )
t
1
mlog .
Remark 1.15. Throughout the rest of the paper we will only include the subscripts
on J, , and when we want to emphasize their dependence on a given variable.
1.6 Weakly Asymmetric Limit of the Tracy-Widom ASEP Formula
The Tracy and Widom ASEP formula (1.31) provides an exact expression for
the probability 1(J
t
(T. X)
T
4
_ s) by interpreting it in terms of a probability
of the location of a tagged particle (1.26). It is of great interest to understand this
limit (J
T
(s)) since, as we have seen, it describes a number of interesting limiting
objects.
We will now present a formal computation of the expressions given in Theo-
rem 1.10 (see Section 1.3) for J
T
(s). After presenting the formal argument, we
will stress that there are a number of very important technical points that arise dur-
ing this argument, many of which require serious work to resolve. In Section 2
we will provide a rigorous proof of Theorem 1.10 in which we deal with all of the
possible pitfalls.
DEFINITION 1.16 Recall the denitions for the relevant quantities in this limit:
=
1
2
1
2
c
1{2
. q =
1
2
1
2
c
1{2
.
; = c
1{2
. t =
1 c
1{2
1 c
1{2
.
. = c
-1
X. t = c
-3{2
T.
m =
1
2
_
c
-1{2
_
a log
_
c
-1{2
2
_
X
2
2T
_
1
2
t .
_
.
J
t
(T. X)
T
4
_ s
_
=
x
_
t
;
. m
_
_ .
_
.
where a = a(s) is dened in the statement of Theorem 1.1. We also dene the
contours I
)
and I
(
to be
I
)
=
z : [z[ = 1
1
2
c
1{2
_
and I
(
=
z : [z[ = 1
1
2
c
1{2
_
.
The rst term in the integrand of (1.31) is the innite product
o
k=0
(1 jt
k
).
Observe that t ~ 1 2c
1{2
and that S
r
C, the contour on which j lies, is a circle
centered at zero of radius between t and 1. The innite product is not well-behaved
along most of this contour, so we will deform the contour to one along which the
product is not highly oscillatory. Care must be taken, however, since the Fredholm
determinant has poles at every j = t
k
. The deformation must avoid passing
through them. Observe now that
o
k=0
(1 jt
k
) = exp
_
o
k=0
log(1 jt
k
)
_
.
and that for small [j[
o
k=0
log(1 j(1 2c
1{2
)
k
) ~ c
-1{2
_
o
0
log(1 je
-2i
)Jr
~ c
-1{2
j
_
o
0
e
-2i
Jr =
c
-1{2
j
2
.
(1.33)
With this in mind dene
j = c
-1{2
j.
from which we see that if the Riemann sum approximation is reasonable then the
innite product converges to e
- {2
. We make the j c
-1{2
jchange of variables
and nd that the above approximations are reasonable if we consider a j-contour
C
t
= {e
i0
]
2
_0_
3
2
L {. i ]
0~x~t
1/2
-1
.
Thus the innite product goes to e
- {2
.
Now we turn to the Fredholm determinant. We determine a candidate for the
pointwise limit of the kernel. That the combination of these two pointwise limits
gives the actual limiting formula as c goes to zero is, of course, completely un-
justied at this point. Also, the pointwise limits here disregard the existence of a
number of singularities encountered during the argument.
The kernel J(j. j
t
) is given by an integral, and the integrand has three main
components: an exponential term
exp{() (j
t
)].
a rational function term (we include the differential with this term for scaling pur-
poses)
J
j
t
( j)
.
and the term
j (j. ,j
t
).
We will proceed by the method of steepest descent; in order to determine the
region along the - and j-contours that affects the asymptotics, we consider the
exponential term rst. The argument of the exponential is given by () (j
t
)
where
() = . log(1 )
t
1
mlog().
and where, for the moment, we take
m =
1
2
_
c
-1{2
_
a
X
2
2T
_
1
2
t .
_
.
The real expression for m has a log(c
-1{2
,2) term which we absorb into a for the
moment (recall that a is dened in the statement of Theorem 1.1). Recall that ., t ,
and m all depend on c. For small c, () has a critical point in an c
1{2
neighbor-
hood of 1. For purposes of having a nice ultimate answer, we choose to center
on
= 1 2c
1{2
X
T
.
We can rewrite the argument of the exponential as (()())((j
t
)()) =
() (j
t
). The idea in [35] for extracting asymptotics of this term is to deform
the - and j-contours to lie along curves such that outside the scale c
1{2
around ,
Re () is large and negative, and Re (j
t
) is large and positive. Hence we can
ignore those parts of the contours. Then, rescaling around to blow up this c
1{2
scale, we obtain the asymptotic exponential term. This nal change of variables
then sets the scale at which we should analyze the other two terms in the integrand
for the J kernel.
Returning to (), we make a Taylor expansion around and nd that in a
neighborhood of ,
() ~
T
48
c
-3{2
( )
3
a
2
c
-1{2
( ).
This suggests the change of variables
(1.34)

= 2
-4{3
c
-1{2
( ). j
t
= 2
-4{3
c
-1{2
(j
t
).
and likewise for j. After this our Taylor expansion takes the form
(1.35) (
) ~
T
3
3
2
1{3
a
.
In the spirit of steepest-descent analysis, we would like the -contour to leave
in a direction where this Taylor expansion is decreasing rapidly. This is accom-
plished by leaving at an angle
2t
3
. Likewise, since (j) should increase rapidly,
j should leave at angle
t
3
. The -contour was originally centered at zero and
of radius 1 c
1{2
,2 and the j-contour of radius 1 c
1{2
,2. In order to deform
these contours without changing the value of the determinant, care must be taken
since there are poles of whenever ,j
t
= t
k
, k Z. We ignore this issue
for the formal calculation and deal with it carefully in Section 2 by using different
contours.
Let us now assume that we can deform our contours to curves along which
rapidly decays in and increases in j, as we move along them away from . If we
apply the change of variables in (1.34), the straight part of our contours become
innite at angles
2t
3
and
t
3
, which we call

I
(
and

I
)
. Note that these are not
the actual denitions of these contours that we use in the statement and proof of
Theorem 1.1 because of the singularity problem mentioned above.
Applying this change of variables to the kernel of the Fredholm determinant
changes the 1
2
-space and hence we must multiply the kernel by the Jacobian term
2
4{3
c
1{2
. We will include this term with the j (j. z)-term and take the c 0
limit of that product.
As noted before, the term 2
1{3
a
should have been 2

1{3
(a log(c
-1{2
,2))
in
the Taylor expansion above, giving
(
) ~
T
3
3
2
1{3
_
a log
_
c
-1{2
2
__
.
which would appear to blow up as c goes to zero. We now show how the extra log c
in the exponent can be absorbed into the 2
4{3
c
1{2
j (j. ,j
t
)-term. Recall
j (j. z) =
o
k=-o
jt
k
1 t
k
j
z
k
.
If we let n
0
= ]log(c
-1{2
), log(t), then observe that for 1 < [z[ < t
-1
,
j (j. z) =
o
k=-o
jt
kn
0
1 t
kn
0
j
z
kn
0
= z
n
0
t
n
0
j
o
k=-o
t
k
1 t
k
t
n
0
j
z
k
.
By the choice of n
0
, t
n
0
~ c
-1{2
so
j (j. z) ~ z
n
0
j ( j. z).
The discussion on the exponential term indicates that it sufces to understand the
behavior of this function when and j
t
are within c
1{2
of . Equivalently, letting
z = ,j
t
, it sufces to understand j (j. z) ~ z
n
0
j ( j. z) for
z =

j
t
=
2
4{3
c
1{2
2
4{3
c
1{2
j
t
~ 1 c
1{2
z. z = 2
4{3
(
j
t
).
Let us now consider z
n
0
using the fact that log(t) ~ 2c
1{2
:
z
n
0
~ (1 c
1{2
z)
t
1/2
(
1
4
log t)
~ e
-
1
4
;log(t)
.
Plugging back in the value of z in terms of

and j
t
, we see that this prefactor of
z
n
0
exactly cancels the log c term that accompanies a in the exponential.
What remains is to determine the limit of 2
4{3
c
1{2
j ( j. z) as c goes to zero
for z ~ 1c
1{2
z. This can be found by interpreting the innite sum as a Riemann
sum approximation for a certain integral. Dene t = kc
1{2
and observe that
(1.36) c
1{2
j ( j. z) =
o
k=-o
jt
t t
1/2
z
t t
1/2
1 jt
t t
1/2
c
1{2
_
o
-o
je
-2t
e
- ;t
1 je
-2t
Jt.
This used the fact that t
t t
1/2
e
-2t
and that z
t t
1/2
e
- ;t
, which hold at
least pointwise in t . For (1.36) to hold, we must have Re z bounded inside (0. 2),
but we disregard this difculty for the heuristic proof. If we change variables of t
to
t
2
and multiply the top and bottom by e
-t
, then we nd that
2
4{3
c
1{2
j (j. ,j
t
) 2
1{3
_
o
-o
je
- ;t {2
e
t
j
Jt.
The nal term to consider is the rational expression. Under the change of variables
and zooming in on , the factor of 1,j
t
goes to 1 and the J,( j
t
) goes to
J
,(
j
t
).
Thereby we formally obtain from jJ the kernel 1
csc
o
0
( j. j
t
) acting on 1
2
(
I
)
),
where
1
csc
o
0 ( j. j
t
) =
_
I
(
e
-
T
3
(
(
3
- )
03
)2
1/3
o
0
(
(- )
0
)
_
2
1{3
_
o
-o
je
-2
1/3
t (
(- )
0
)
e
t
j
Jt
_
J
j
with a
t
= a log 2. Recall that the log 2 came from the log(c
-1{2
,2) term.
We have the identity
(1.37)
_
o
-o
je
- ;t {2
e
t
j
Jt = ( j)
- ;{2
csc
_
z
2
_
.
where the branch cut in j is taken along the positive real axis; hence ( j)
- ;{2
=
e
-log(- ) ;{2
where the log is taken with the standard branch cut along the negative
real axis. We may use the identity to rewrite the kernel as
1
csc
o
0 ( j. j
t
) =
_
I
(
e
-
T
3
(
(
3
- )
03
)2
1/3
o
0
(
(- )
0
)
2
1{3
( j)
-2
1/3
(
(- )
0
)
sin(2
1{3
(
j
t
))
J
j
.
Therefore we have shown formally that
lim
t-0
1
_
J
t
(T. X)
T
4
_ s
_
:= J
T
(s) =
_
C
e
- {2
J j
j
det(1 1
csc
o
0 )
1
2
(
I
)
)
.
where a
t
= a log 2. To make it cleaner we replace

2
with j. This only affects
the j-term above, given now by
(2 j)
- ;{2
= ( j)
-2
1/3
(
(- )
0
)
e
-2
1/3
log 2(
(- )
0
)
.
This can be absorbed and cancels the log 2 in a
t
and thus we obtain
J
T
(s) =
_
C
e
-
J j
j
det(1 1
csc
o
)
1
2
(
I
)
)
.
which, up to the denitions of the contours

I
)
and

I
(
, is the desired limiting
formula.
We now briey note some of the problems and pitfalls of the preceding formal
argument, all of which will be addressed in the real proof of Section 2.
First, the pointwise convergence of both the prefactor innite product and the
Fredholm determinant is certainly not enough to prove convergence of the j-
integral. Estimates must be made to control this convergence or to show that we
can cut off the tails of the j-contour at negligible cost and then show uniform
convergence on the trimmed contour.
Second, the deformations of the j- and -contours to the steepest-descent curves
is entirely illegal, as it involves passing through many poles of the kernel (coming
from the -term). In the case of [35] this problem could be dealt with rather simply
by just slightly modifying the descent curves. However, in our case, since t tends
to 1 like c
1{2
, such a patch is much harder and involves very ne estimates to show
that there exists suitable contours that stay close enough together, yet along which
displays the necessary descent and ascent required to make the argument work.
This issue also comes up in the convergence of (1.36). In order to make sense of
this we must ensure that 1 < [,j
t
[ < t
-1
or else the convergence and the resulting
expression make no sense.
Finally, one must make precise tail estimates to show that the kernel convergence
is in the sense of the trace class norm. The Riemann sum approximation argument
can in fact be made rigorous (following the proof of Proposition 2.5). We choose,
however, to give an alternative proof of the validity of that limit in which we iden-
tify and prove the limit of via analysis of singularities and residues.
S
1
C
Cauchys Theorem Change of vars

C
1/2
1
~
~
FIGURE 2.1. The S
t
-contour is deformed to the C
t
-contour via
Cauchys theorem and then a change of variables leads to

C
t
, with its
innite extension

C.
2 Proof of the Weakly Asymmetric Limit
of the Tracy-Widom ASEP Formula
2.1 Proof of Theorem 1.10
In this section we give a proof of Theorem 1.10, for which a formal derivation
was presented in Section 1.6. The heart of the argument is Proposition 2.4, which
is proved in Section 2.2 and also relies on a number of technical lemmas. These
lemmas as well as all of the other propositions are proved in Section 2.3.
The expression given in (1.31) for 1(J
t
(T. X)
T
4
_ s) contains an integral
over a j-contour of a product of a prefactor innite product and a Fredholm deter-
minant. The rst step towards taking the limit of this as c goes to zero is to control
the prefactor,
o
k=0
(1 jt
k
). Initially j lies on a contour S
r
C that is centered at
zero and of radius between t and 1. Along this contour the partial products (i.e.,
product up to N) form a highly oscillatory sequence and hence it is hard to control
the convergence of the sequence.
The rst step in our proof is to deform the j-contour S
r
C to
C
t
= {c
1{2
e
i0
] L {. i c
1{2
]
0~x_1-t
1/2 L {1 c
1{2
c
1{2
i,]
-1~,~1
.
a long, skinny, cigar-shaped contour (see Figure 2.1). We orient C
t
counterclock-
wise. Notice that this new contour still includes all of the poles at j = t
k
associ-
ated with the -function in the J-kernel.
In order to justify replacing S
r
C by C
t
we need the following (for the proof see
Section 2.3):
LEMMA 2.1 In (1.31) we can replace the contour S
t
with C
t
as the contour of
integration for j without affecting the value of the integral.
Having made this deformation of the j-contour, we now observe that the natural
scale for j is on order c
1{2
. With this in mind we make the change of variables
j = c
1{2
j.
Remark 2.2. Throughout the proof of this theorem and its lemmas and proposi-
tions, we will use the tilde to denote variables that are c
1{2
rescaled versions of the
original, untilded variables.
The j-variable now lives on the contour
C
t
= {e
i0
] L {. i ]
0~x_t
1/2
-1
L {c
-1{2
1 i,]
-1~,~1
.
which grow and ultimately approach
C = {e
i0
] L {. i ]
x>0
.
In order to show convergence of the integral as c goes to zero, we must consider two
things, the convergence of the integrand for j in some compact region around the
origin on

C, and the controlled decay of the integrand on

C
t
outside of that compact
region. This second consideration will allow us to approximate the integral by a
nite integral in j, while the rst consideration will tell us what the limit of that
integral is. When all is said and done, we will paste back in the remaining part
of the j-integral and have our answer. With this in mind we give the following
bound, which is proved in Section 2.3.
LEMMA 2.3 Dene two regions, depending on a xed parameter r _ 1,
1
1
=
_
j : [ j[ _
r
sin(,10)
_
.
1
2
=
_
j : Re( j)
_
r
tan(,10)
. c
-1{2
_
and Im( j) 2. 2|
_
.
1
1
is compact and 1
1
L 1
2
contains all of the contour

C
t
. Furthermore, dene
the function (the innite product after the change of variables)
g
t
( j) =
o
k=0
(1 c
1{2
jt
k
).
Then uniformly in j 1
1
,
(2.1) g
t
(j) e
- {2
.
Also, for all c < c
0
(some positive constant) there exists a constant c such that for
all j 1
2
we have the following tail bound:
(2.2) [g
t
( j)[ _ [e
- {2
[ [e
-ct
1/2

2
[.
(By the choice of 1
2
, for all j 1
2
, Re( j
2
) > > 0 for some xed . The
constant c can be taken to be
1
S
).
We now turn our attention to the Fredholm determinant term in the integrand.
Just as we did for the prefactor innite product in Lemma 2.3, we must establish
uniform convergence of the determinant for j in a xed compact region around
the origin and a suitable tail estimate valid outside that compact region. The tail
estimate must be such that for each nite c, we can combine the two tail estimates
(from the prefactor and from the determinant) and show that their integral over the
tail part of

C
t
is small and goes to zero as we enlarge the original compact region.
For this we have the following two propositions (the rst is the most substantial
and is proved in Section 2.2, while the second is proved in Section 2.3).
PROPOSITION 2.4 Fix s R, T > 0, and X R. Then for any compact subset of
C we have that for all > 0 there exists an c

0
> 0 such that for all c < c
0
and all
j in the compact subset,
det(1 c
1{2
jJ
t
1/2

)
1
2
(I
)
)
det(1 1
csc
o
0 )
1
2
(
I
)
)
< .
Here a
t
= a log 2 and 1
csc
o
0
is dened in Denition 1.29 and depends implicitly
on j.
PROPOSITION 2.5 There exist c. c
t
> 0 and c
0
0
and
all j

C
t
,
g
t
( j) det(1 c
1{2
jJ
t
1/2

)
1
2
(I
)
)
_ c
t
e
-c[ [
.
This exponential decay bound on the integrand shows that, by choosing a suit-
ably large (xed) compact region around zero along the contour

C
t
, it is possi-
ble to make the j-integral outside of this region arbitrarily small, uniformly in
c (0. c
0
). This means that we may assume henceforth that j lies in a compact
subset of

C.
Now that we are on a xed compact set of j, the rst part of Lemma 2.3 and
Proposition 2.4 combine to show that the integrand converges uniformly to
e
- {2
j
det(1 1
csc
o
0 )
1
2
(
I
)
)
and hence the integral converges to the integral with this integrand.
To nish the proof of the limit in Theorem 1.10, it is necessary that for any we
can nd a suitably small c
0
such that the difference between the two sides of the
limit differ by less than for all c < c
0
. Technically we are in the position of a

3
argument. One portion of

3
goes to the cost of cutting off the j-contour outside
of some compact set. Another

3
goes to the uniform convergence of the integrand.
The nal portion goes to repairing the j-contour. As gets smaller, the cut for the
j-contour must occur further out. Therefore the limiting integral will be over the
limit of the j-contours, which we called

C. The nal

3
is spent on the following
proposition, whose proof is given in Section 2.3.
PROPOSITION 2.6 There exists c. c
t
> 0 such that for all j

C with [ j[ _ 1,
e
- {2
j
det(1 1
csc
o
)
1
2
(
I
)
)
_ [c
t
e
-c
[.
Recall that the kernel 1
csc
o
is a function of j. The argument used to prove this
proposition immediately shows that 1
csc
o
is a trace class operator on 1
2
(
I
)
).
It is an immediate corollary of this exponential tail bound that for sufciently
large compact sets of j, the cost to include the rest of the j-contour is less than

3
.
This, along with the change of variables in j described at the end of Section 1.6,
nishes the proof of Theorem 1.10.
2.2 Proof of Proposition 2.4
In this section we provide all of the steps necessary to prove Proposition 2.4. To
ease understanding of the argument, we relegate more technical points to lemmas
whose proof we delay to Section 2.3.
During the proof of this proposition, it is important to keep in mind that we are
assuming that j lies in a xed compact subset of

C. Recall that j = c
-1{2
j. We
proceed via the following strategy to nd the limit of the Fredholm determinant as
c goes to zero. The rst step is to deform the contours I
)
and I
(
to suitable curves
along which there exists a small region outside of which the kernel of our operator
is exponentially small. This justies cutting the contours off outside of this small
region. We may then rescale everything so that this small region becomes order 1 in
size. Then we show uniform convergence of the kernel to the limiting kernel on the
compact subset. Finally, we need to show that we can complete the nite contour
on which this limiting object is dened to an innite contour without signicantly
changing the value of the determinant.
Recall now that I
(
is dened to be a circle centered at zero of radius 1c
1{2
,2,
that I
)
is a circle centered at zero of radius 1 c
1{2
,2, and that
= 1 2c
1{2
X
T
.
The function (j. ,j
t
) that shows up in the denition of the kernel for J has
poles as every point ,j
t
= z = t
k
for k Z.
As long as we simultaneously deform the I
(
-contour as we deform I
)
so as to
keep ,j
t
away from these poles, we may use Proposition 2.19 [35, proposition 1]
to justify the fact that the determinant does not change under this deformation. In
this way we may deform our contours to the following modied contours I
),I
and
I
(,I
:
DEFINITION 2.7 Let I
),I
and I
(,I
be two families (indexed by l > 0) of simple
closed contours in C dened as follows: Let
(2.3) k(0) =
2X
T
tan
2
_
0
2
_
log
_
2
1 cos 0
_
.
Both I
),I
and I
(,I
will be symmetric across the real axis, so we need only dene
them on the top half. I
),I
begins at c
1{2
,2 and moves along a straight vertical
line for a distance lc
1{2
and then joins the curve
(2.4) 1 c
1{2
(k(0) )|e
i0
parametrized by 0 from lc
1{2
O(c) to zero, and where =
1
2
O(c
1{2
)
(see Figure 2.2 for an illustration of these contours). The small errors are necessary
to make sure that the curves join up at the end of the vertical section of the curve.
FIGURE 2.2. I
(,I
(the outermost curve) is composed of a small, vertical
section near labeled I
vert
(,I
and a large, almost circular (small modica-
tion due to the function k(0)) section labeled I
circ
(,I
. Likewise I
),I
is the
middle curve, and the inner curve is the unit circle. These curves depend
on c in such a way that [,j[ is bounded between 1 and t
-1
~ 12c
1{2
.
We extend this to a closed contour by reection through the real axis and orient it
clockwise. We denote the rst, vertical part of the contour by I
vert
),I
and the second,
roughly circular part by I
circ
),I
. This means that I
),I
= I
vert
),I
L I
circ
),I
, and along this
contour we can think of parametrizing j by 0 0. |.
We dene I
(,I
similarly except that it starts out at c
1{2
,2 and joins the curve
given by (2.4) where the value of 0 ranges from 0 = lc
1{2
O(c) to 0 = 0
and where =
1
2
O(c
1{2
). We similarly denote this contour by the union of
I
vert
(,I
and I
circ
(,I
.
By virtue of these denitions, it is clear that c
-1{2
[,j
t
t
k
[ stays bounded
away from zero for all k and that [,j
t
[ is bounded in a closed set contained in
(1. t
-1
) for all I
(,I
and j I
),I
. Therefore, for any l > 0 we may, by
deforming both the j- and -contours simultaneously, assume that our operator
acts on 1
2
(I
),I
) and that its kernel is dened via an integral along I
(,I
. It is
critical that we now show that, due to our choice of contours, we are able to forget
about everything except for the vertical part of the contours. To formulate this we
have the following:
DEFINITION 2.8 Let
vert
I
and
circ
I
be projection operators acting on 1
2
(I
),I
) that
project onto 1
2
(I
vert
),I
) and 1
2
(I
circ
),I
), respectively. Also, dene two operators J
vert
I
and J
circ
I
that act on 1
2
(I
),I
) and have kernels identical to J (see (1.32)) except
the -integral is over I
vert
(,I
and I
circ
(,I
, respectively. Thus we have a family (indexed
by l > 0) of decompositions of our operator J as follows:
J = J
vert
I

vert
I
J
vert
I

circ
I
J
circ
I

vert
I
J
circ
I

circ
I
.
We now show that it sufces to just consider the rst part of this decomposition
(J
vert
I

vert
I
) for sufciently large l.
PROPOSITION 2.9 Assume that j is restricted to a bounded subset of the contour
C. For all > 0 there exist c

0
> 0 and l
0
0
and all
l > l
0
,
[det(1 jJ)
1
2
(I
),l
)
det(1 J
vert
I
)
1
2
(I
vert
),l
)
[ < .
PROOF. As was explained in the introduction, if we let
(2.5) n
0
=
log(c
-1{2
)
log(t)
.
then it follows from the invariance of the doubly innite sum for (j. z) that
j (j. z) = z
n
0
( j ( j. z) O(c
1{2
)).
Note that because the O(c
1{2
) does not play a signicant role in what follows, we
drop it.
Using the above argument and the following two lemmas (which are proved in
Section 2.3) we will be able to complete the proof of Proposition 2.9.
LEMMA 2.10 For all c > 0 there exist l
0
> 0 and c
0
> 0 such that for all l > l
0
,
c < c
0
, and j I
circ
),I
,
Re((j) n
0
log(j)) _ c[ j[c
-1{2
.
where n
0
is dened in (2.5). Likewise, for all c < c
0
and I
circ
(,I
,
Re(() n
0
log()) _ c[ [c
-1{2
.
LEMMA 2.11 For all l > 0 there exist c
0
> 0 and c > 0 such that for all c < c
0
,
j
t
I
),I
, and I
(,I
,
[ j ( j. ,j
t
)[ _
c
[ j
t
[
.
It now follows that for any > 0, we can nd l
0
large enough that [J
vert
I

circ
I
[
1
,
[J
circ
I

vert
I
[
1
, and [J
circ
I

circ
I
[
1
are all bounded by ,3. This is because we may
factor these various operators into a product of Hilbert-Schmidt operators and then
use the exponential decay of Lemma 2.10 along with the polynomial control of
Lemma 2.11 and the remaining term 1,( j) to prove that each of the Hilbert-
Schmidt norms goes to zero (for a similar argument, see [35, bottom of p. 27]).
This completes the proof of Proposition 2.9.
We now return to the proof of Proposition 2.4. We have successfully restricted
ourselves to considering J
vert
I
acting on 1
2
(I
vert
),I
). Having focused on the region
of asymptotically nontrivial behavior, we can now rescale and show that the kernel
converges to its limit, uniformly on the compact contour.
DEFINITION 2.12 Recall c
3
= 2
-4{3
and let
j = c
-1
3
c
1{2
j. j
t
= c
-1
3
c
1{2
j
t
. = c
-1
3
c
1{2
.
Under these changes of variables the contours I
vert
),I
and I
vert
(,I
become
I
),I
=
_
c
3
2
i r : r (c
3
l. c
3
l)
_
.

I
(,I
=
_
c
3
2
i r : r (c
3
l. c
3
l)
_
.
As l increases to innity, these contours approach their innite versions,
I
)
=
_
c
3
2
i r : r (o. o)
_
.

I
(
=
_
c
3
2
i r : r (o. o)
_
.
With respect to the change of variables, dene an operator

J acting on 1
2
(
I
)
) via
the kernel:
j
J
I
( j. j
t
) =
c
-1
3
c
1{2
_
I
(,l
e
(c
1
3
t
1/2
()-(c
1
3
t
1/2
)
0
)
j
_
j.
c
1
3
t
1/2
(
c
1
3
t
1/2
)
0
_
( c
-1
3
c
1{2
j
t
)(
j)
J
.
Lastly, dene the operator
I
that projects 1
2
(
I
)
) onto 1
2
(
I
),I
)
It is clear that by applying the change of variables, the Fredholm determinant
det(1 J
vert
I
)
1
2
(I
vert
),l
)
becomes det(1
I
j
J
I

I
)
1
2
(
I
),l
)
.
We now state a proposition that gives, with respect to these xed contours

I
),I
and

I
(,I
, the limit of the determinant in terms of the uniform limit of the kernel.
Since all contours in question are nite, uniform convergence of the kernel sufces
to show trace class convergence of the operators and hence convergence of the
determinant.
Recall the denition of the operator 1
csc
o
given in Denition 1.12. For the pur-
poses of this proposition, modify the kernel so that the integration in occurs now
only over

I
(,I
and not all of

I
(
. Call this modied operator 1
csc
o
0
,I
.
PROPOSITION 2.13 For all > 0 there exist c
0
> 0 and l
0
> 0 such that for all
c < c
0
, l > l
0
, and j in our xed compact subset of

C,
det(1
I
j
J
I

I
)
1
2
(
I
),l
)
det(1
I
1
csc
o
0
,I

I
)
1
2
(
I
),l
)
< .
where a
t
= a log 2.
PROOF. The proof of this proposition relies on showing the uniform conver-
gence of the kernel of j
J to the kernel of 1
csc
o
0
,I
, which sufces because of the
compact contour. Furthermore, since the -integration is itself over a compact set,
it sufces to show uniform convergence of this integrand. The two lemmas stated
below will imply such uniform convergence and hence complete this proof.
First, however, recall that j (j. z) = z
n
0
( j ( j. z) O(c
1{2
)) where n
0
is
dened in equation (2.5). We are interested in having z = ,j
t
, which, under the
change of variables, can be written as
z = 1 c
1{2
z O(c). z = c
-1
3
(
j
t
) = 2
4{3
(
j
t
).
Therefore, since n
0
=
1
2
log(c
-1{2
)c
-1{2
O(1) it follows that
z
n
0
= exp{2
1{3
(
j
t
) log(c
-1{2
)](1 o(1)).
This expansion still contains an c and hence the argument blows up as c goes to
zero. However, this exactly counteracts the log(c
-1{2
) term in the denition of m
that goes into the argument of the exponential of the integrand. We make use of
this cancellation in the proof of this rst lemma and hence include the n
0
log(,j
t
)
term in the exponential argument.
The following two lemmas are proved in Section 2.3.
LEMMA 2.14 For all l > 0 and all > 0 there exists c
0
j
t

I
),I
and

I
(,I
we have for 0 < c _ c
0
,
((
) ( j
t
) n
0
log(,j
t
))
_
T
3
(
3
j
t3
) 2
1{3
a
t
(
j)
_
< .
where a = a
t
log 2. Similarly, we have
e
(
()-( )
0
)n
0
log(({)
0
)
e
-
T
3
(
(
3
- )
03
)2
1/3
o
0
(
(- )
0
)
< .
LEMMA 2.15 For all l > 0 and all > 0 there exists c
0
j
t

I
),I
and

I
(,I
we have for 0 < c _ c
0
,
c
1{2
j
_
j.
c
-1
3
c
1{2
c
-1
3
c
1{2
j
t
_
_
o
-o
je
-2
1/3
t (
(- )
0
)
e
t
j
Jt
< .
As explained in Denition 1.12, the nal integral converges since our choices of
and j
t
ensure that Re(2
1{3
(
j
t
)) =
1
2
. Note that the above integral also has
a representation (1.37) in terms of the cosecant function. This gives the analytic
extension of the integral to all z 2Z where z = 2
4{3
(
j
t
).
Finally, the sign change in front of the kernel of the Fredholm determinant comes
from the 1,j
t
term, which, under the change of variables, converges uniformly
to 1.
Having successfully taken the c to zero limit, all that now remains is to paste the
rest of the contours

I
)
and

I
(
to their abbreviated versions

I
),I
and

I
(,I
. To justify
this, we must show that the inclusion of the rest of these contours does not signi-
cantly affect the Fredholm determinant. Just as in the proof of Proposition 2.9 we
have three operators that we must reinclude at provably small cost. Each of these
operators, however, can be factored into the product of Hilbert-Schmidt operators
and then an analysis similar to that done following Lemma 2.11 (see in particular
[35, pp. 2728]) shows that because Re(
3
) grows like [
[
2
along

I
(
(and like-
wise but opposite for j
t
) there is sufciently strong exponential decay to show that
the trace norms of these three additional kernels can be made arbitrarily small by
taking l large enough.
This last estimate completes the proof of Proposition 2.4.
2.3 Technical Lemmas, Propositions, and Proofs
Properties of Fredholm Determinants
Before beginning the proofs of the propositions and lemmas, we give the deni-
tions and some important properties for Fredholm determinants, trace class opera-
tors, and Hilbert-Schmidt operators. For a more complete treatment of this theory
see, for example, [29].
Consider a (separable) Hilbert space H with bounded linear operators L(H).
If L(H), let [[ =
_
+
be the unique positive square root. We say
that B
1
(H), the trace class operators, if the trace norm [[
1
< o. Re-
call that this norm is dened relative to an orthonormal basis of H as [[
1
:=
o
n=1
(e
n
. [[e
n
). This norm is well-dened as it does not depend on the choice
of orthonormal basis {e
n
]
n_1
. For B
1
(H), one can then dene the trace
tr :=
o
n=1
(e
n
. e
n
). We say that B
2
(H), the Hilbert-Schmidt operators,
if the Hilbert-Schmidt norm [[
2
:=
_
tr([[
2
) < o.
LEMMA 2.16 [7, p. 40] from [29, theorem 2.20] The following conditions are
equivalent:
v [1
n
1[
1
0,
v tr 1
n
tr 1 and 1
n
1 in the weak operator topology.
For B
1
(H) we can also dene a Fredholm determinant det(1 )
H
.
Consider u
i
H and dene the tensor product u
1
u
n
by its action on
1
. . . . .
n
H as
u
1
u
n
(
1
. . . . .
n
) =
n
i=1
(u
i
.
i
).
Then
n
i=1
H is the span of all such tensor products. There is a vector subspace
of this space that is known as the alternating product,
n
_
(H) =
_
h
n
i=1
H : Vo S
n
. oh = h
_
.
where ou
1
u
n
= u
c(1)
u
c(n)
. If e
1
. . . . . e
n
is a basis for H, then
e
i
1
. .e
i
k
for 1 _ i
1
< . . . < i
k
_ n form a basis of
_
n
(H). Given an operator
L(H), dene
I
n
()(u
1
u
n
) := u
1
u
n
.
Note that any element in
_
n
(H) can be written as an antisymmetrization of tensor
products. It then follows that I
n
() can be restricted to act as an operator from
_
n
(H) into
_
n
(H). If B
1
(H), then tr I
(n)
() _ [[
n
1
,n, and we can
dene
det(1 ) = 1
o
k=1
tr(I
(k)
()).
As one expects, det(1 ) =
}
(1 z
}
) where z
}
are the eigenvalues of
counted with algebraic multiplicity [24, theorem XIII.106].
LEMMA 2.17 [29, chap. 3] det(1 ) is a continuous function on B
1
(H).
Explicitly,
[det(1 ) det(1 T)[ _ [ T[
1
exp([[
1
[T[
1
1).
If B
1
(H) and = TC with T. C B
2
(H), then
[[
1
_ [T[
2
[C[
2
.
For B
1
(H),
[det(1 )[ _ e
||
1
.
If B
2
(H) with kernel (.. ,), then
[[
2
=
__
[(.. ,)[
2
J. J,
_
1{2
.
LEMMA 2.18 If 1 is an operator acting on a contour and is a projection
operator unto a subinterval of , then
det(1 1)
1
2
(,)
= det(1 1)
1
2
(,)
.
In performing steepest-descent analysis on Fredholm determinants, the follow-
ing proposition allows one to deform contours to descent curves.
LEMMA 2.19 [35, prop. 1] Suppose s I
x
is a deformation of closed curves
and a kernel 1(j. j
t
) is analytic in a neighborhood of I
x
I
x
C
2
for each s.
Then the Fredholm determinant of 1 acting on I
x
is independent of s.
The following lemma, provided to us by Percy Deift, with proof provided in
the Appendix, allows us to use Cauchys theorem when manipulating integrals that
involve Fredholm determinants in the integrand.
LEMMA 2.20 Suppose (z) is an analytic map from a region D C into the
trace class operators on a (separable) Hilbert space H. Then z det(1 (z))
is analytic on D.
Proofs from Section 2.1
We now turn to the proofs of previously stated lemmas and propositions.
PROOF OF LEMMA 2.1. This lemma follows from Cauchys theorem once we
show that for xed c, the integrand j
-1
o
k=0
(1 jt
k
) det(1 jJ
) is analytic
in j between S
t
and C
t
(note that we now include a subscript j on J to emphasize
the dependence of the kernel on j). It is clear that the innite product and the j
-1
are analytic in this region. In order to show that det(1 jJ
) is analytic in the
desired region, we may appeal to Lemma 2.20. Therefore it sufces to show that
the map J(j) dened by j J
is an analytic map from this region of j between

S
t
and C
t
into the trace class operators (this sufces since the multiplication by j
is clearly analytic). The rest of this proof is devoted to the proof of this fact.
In order to prove this, we need to show that J
h
= (J
h
J
),h converges to
some trace class operator as h C goes to zero. By the criteria of Lemma 2.16 it
sufces to prove that the kernel associated to J
h
converges uniformly in j. j
t
I
)
to the kernel of J
t
. This will prove both the convergence of traces as well as

the weak convergence of operators necessary to prove trace norm convergence and
complete this proof. The operator J
t
acts on I
)
, the circle centered at zero and of
radius 1
1
2
c
1{2
, as
J
t
(j. j
t
) =
_
I
(
exp{() (j
t
)]
t
(j. ,j
t
)
j
t
( j)
J
where
t
(j. z) =
o
k=-o
t
2k
(1 t
k
j)
2
z
k
.
Our desired convergence will follow if we can show that
[h
-1
( (j h. ,j
t
) (j. ,j
t
))
t
(j. ,j
t
)[
tends to zero uniformly in I
(
and j
t
I
)
as [h[ tends to zero. Expanding this
and taking the absolute value inside of the innite sum, we have
(2.6)
o
k=-o
h
-1
_
t
k
1 t
k
(j h)
t
k
1 t
k
(j)
_
t
2k
(1 t
k
(j))
2
z
k
where z = [,j
t
[ (1. t
-1
). For c and j xed there is a k
+
such that for k _ k
+
,
t
k
h
1 t
k
j
< 1.
Furthermore, by choosing [h[ small enough we can make k
+
negative. As a
result we also have that for small enough [h[, for all k < k
+
,
h
t
-1
j
< 1.
Splitting our sum into k < k
+
and k _ k
+
and using the fact that 1,(1 n) =
1 n O(n
2
) for [n[ < 1, we can Taylor-expand as follows: For k _ k
+
t
k
1 t
k
(j h)
=
t
k
1 t
k
j
1
1
r
k
h
1-r
k
=
t
k
_
1
r
k
h
1-r
k

_
r
k
1-r
k
_
2
O(h
2
)
_
1 t
k
j
.
Similarly, after expanding the second term inside the absolute value in (2.6) and
canceling with the third term, we are left with
o
k=k
t
3k
(1 t
k
j)
3
O(h)z
k
.
The sum converges since t
3
z < 1 and thus behaves like O(h) as desired. Likewise
for k < k
+
, by multiplying the numerator and denominator by t
-k
, the same type
of expansion works and we nd that the error is given by the same summand as
above but over k from o to k
+
1. Again, however, the sum converges since
the numerator and denominator cancel each other for k large negative, and z
k
is
a convergent series for k going to negative innity. Thus this error series also
behaves like O(h) as desired. This shows the needed uniform convergence and
completes the proof.
PROOF OF LEMMA 2.3. We prove this with the scaling parameter r = 1 as the
general case follows in a similar way. Consider
log(g
t
( j)) =
o
k=0
log
_
1 c
1{2
jt
k
t
_
.
We have
o
k=0
c
1{2
t
k
=
1
2
(1 c
1{2
c
t
) where c
t
= O(1). So for j 1
1
we
have
log(g
t
( j))
j
2
(1 c
1{2
c
t
)
k=0
log(1 c
1{2
jt
k
) c
1{2
jt
k
_
o
k=0
[log(1 c
1{2
jt
k
) c
1{2
jt
k
[
_
o
k=0
[c
1{2
jt
k
[
2
=
c[ j[
2
1 t
2
=
c
1{2
[ j[
2
4 4c
1{2
_ cc
1{2
[ j[
2
_ c
t
c
1{2
.
The second inequality uses the fact that for [z[ _
1
2
, [log(1 z) z[ _ [z[
2
. Since
j 1
1
it follows that [z[ = c
1{2
[ j[ is bounded by
1
2
for small enough c. The
constants here are nite and do not depend on any of the parameters. This proves
(2.1) and shows that the convergence is uniform in j on 1
1
.
We now turn to the second inequality, (2.2). Consider the region
D = {z : arg(z)
t
10
.
t
10
|] {z : Im(z) (
1
10
.
1
10
)] {z : Re(z) _ 1].
For all z D,
(2.7) Re(log(1 z)) _ Re(z z
2
,2).
For j 1
2
, it is clear that c
1{2
j D. Therefore, using (2.7),
Re(log(g
t
( j))) =
o
k=0
Relog(1 c
1{2
jt
k
)|
_
o
k=0
_
Rec
1{2
jt
k
| Re
_
(c
1{2
jt
k
)
2
2
__
_ Re( j,2)
1
8
c
1{2
Re( j
2
).
This proves (2.2). Note that from the denition of 1
2
we can calculate the argu-
ment of j and we see that [arg j[ _ arctan(2 tan(
t
10
)) <
t
4
and [ j[ _ r _ 1.
Therefore Re( j
2
) is positive and bounded away from zero for all j 1
2
.
PROOF OF PROPOSITION 2.5. This proof proceeds in a similar manner to the
proof of Proposition 2.6. However, since in this case we have to deal with c going
to zero and changing contours, it is, by necessity, a little more complicated. For
this reason we encourage readers to rst study the simpler proof of Proposition 2.6.
In that proof we factor our operator into two pieces. Then, using the decay of
the exponential term and the control over the size of the cosecant term, we are able
to show that the Hilbert-Schmidt norm of the rst factor is nite and that for the
second factor it is bounded by [ j[
for < 1 (we show it for =

1
2
though any
> 0 works, just with the constant getting large as _ 0). This gives an estimate
on the trace norm of the operator, which, by exponentiating, gives an upper bound
e
c[ [
on the size of the determinant. Due to this upper bound and the exponential
decay of the prefactor term g
t
, the product of these two terms is bounded by a
quantity that is exponentially decaying.
For the proof of Proposition 2.5, we do the same sort of factorization of our
operator into T, where here
(. j) =
e
c(()n
0
log(()j
j
with n
0
as explained prior to the statement of Lemma 2.10, 0 < c < 1 xed, and
T(j. ) = e
-c(()n
0
log(()j
e
(()-())
j (j. ,j)
1
j
.
We must be careful in keeping track of the contours on which these operators
act. As we have seen, we may assume that the j-variables are on I
),I
and the
-variables on I
(,I
for any xed choice of l _ 0. Now using the estimates of Lem-
mas 2.10 and 2.14, we compute that [[
2
< o (uniformly in c < c
0
and, triv-
ially, also in j). Here we calculate the Hilbert-Schmidt norm using Lemma 2.17.
Intuitively this norm is uniformly bounded as c goes to zero, because, while the
denominator blows up as badly as c
-1{2
, the numerator is roughly supported only
on a region of measure c
1{2
(owing to the exponential decay of the exponential
when differs from by more than order c
1{2
).
We wish to control [T[
2
now. Using the discussion before Lemma 2.10 we may
rewrite T as
T(j. ) = e
-c(()n
0
log(()j
e
((()n
0
log(())-(())-n
0
log()))
j ( j. ,j)
1
j
.
Lemmas 2.10 and 2.14 apply and tell us that the exponential terms decay at least as
fast as exp{c
-1{2
c
t
[ j[]. So the nal ingredient in proving our proposition is
control of [ j ( j. z)[ for z = ,j
t
. We break it up into two regions of j
t
and : the
rst (case 1) when [j
t
[ _ c for a very small constant c and the second (case 2)
when [j
t
[ > c. We will compute [T[
2
as the square root of
(2.8)
_
),(case 1
[T(j. )[
2
Jj J
_
),(case 2
[T(j. )[
2
Jj J.
We will show that the rst term can be bounded by C[ j[
2
for any < 1, while
the second term can be bounded by a large constant. As a result, [T[
2
_ C[ j[
,
which is exactly as desired since then [T[
1
_ e
c[ [
.
Consider case 1, where [j
t
[ _ c for a constant c that is positive but small
(depending on T ). One may easily check from the denition of the contours that
c
-1{2
([,j[ 1) is contained in a compact subset of (0. 2). In fact, ,j
t
lies
almost exactly along the curve [z[ = 1 c
1{2
, and in particular (by taking c
0
and c small enough) we can assume that ,j never leaves the region bounded by
[z[ = 1 (1 r)c
1{2
for any xed r < 1. Let us call this region 1
t,i
. Then we
have the following:
LEMMA 2.21 Fix c
0
and r (0. 1). Then for all c < c
0
, j

C
t
, and z 1
t,i
,
[ j ( j. z)[ _ c
[ j[
[1 z[
for some (0. 1) with c = c() independent of z, j, and c.
Remark 2.22. By changing the value of in the denition of k(0) (which then
goes into the denitions of I
),I
and I
(,I
) and also focusing the region 1
t,i
around
[z[ = 1 2c
1{2
, we can take arbitrarily small in the above lemma at a cost of
increasing the constant c = c() (the same also applies for Proposition 2.6). The
[ j[
comes from the fact that

(1 2c
1{2
)
1
2
t
1/2
log [ [
~ [ j[
.
Another remark is that the proof below can be used to provide an alternative proof
of Lemma 2.15 by studying the convergence of the Riemann sum directly rather
than by using functional equation properties of and the analytic continuations.
We complete the ongoing proof of Proposition 2.5 and then return to the proof
of the above lemma.
Case 1 is now done since we can estimate the rst integral in (2.8) using
Lemma 2.21 and the exponential decay of the exponential term outside of [j
t
[ =
O(c
1{2
). Therefore, just as with the operator, the c
-1{2
blowup of [ j ( j. ,j
t
)[
is countered by the decay of the exponential, and we are just left with a large con-
stant times [ j[
.
Turning to case 2, we need to show that the second integral in (2.8) is bounded
uniformly in c and j

C
t
. This case corresponds to [j
t
[ > c for some
xed but small constant c. Since c
-1{2
([,j[ 1) stays bounded in a compact
set, using an argument almost identical to the proof of Lemma 2.11, we can show
that [ j ( j. ,j)[ can be bounded by C[ j[
C
0
for positive yet nite constants C
and C
t
. The important point here is that there is only a nite power of [ j[. Since
[ j[ < c
-1{2
, this means that this term can blow up at most polynomially in c
-1{2
.
On the other hand, we know that the exponential term decays exponentially fast
like e
-t
1/2
c
and hence the second integral in (2.8) goes to zero.
We now return to the proof of Lemma 2.21, which will complete the proof of
Proposition 2.5.
PROOF OF LEMMA 2.21. We will prove the desired estimate for z : [z[ = 1
c
1{2
. The proof for general z 1
t,i
follows similarly. Recall that
j ( j. z) =
o
k=-o
jt
k
1 jt
k
z
k
.
Since j has imaginary part 1, the denominator is smallest when t
k
= 1,[ j[,
corresponding to
k = k
+
= ]
1
2
c
-1{2
log [j[.
We start, therefore, by centering our doubly innite sum at around this value,
j ( j. z) =
o
k=-o
jt
k
t
k
1 jt
k
t
k
z
k
z
k
.
By the denition of k
+
,
[z[
k
= [ j[
1{2
(1 O(c
1{2
)):
thus we nd that
[ j ( j. z)[ = [ j[
1{2
k=-o
ot
k
1 ot
k
z
k

where
o = jt
k
and is roughly on the unit circle except for a small dimple near 1. To be more
precise, due to the rounding in the denition of k
+
, the o is not exactly on the unit
circle. However, we do have the following two properties:
[1 o[ > c
1{2
. [o[ 1 = O(c
1{2
).
The section of

C
t
in which j = c
-1{2
1 i, for , (1. 1) corresponds to o
lying along a small dimple around 1 (and still respects [1o[ > c
1{2
). We call the
curve on which o lies C.
We can bring the [ j[
1{2
factor to the left and split the summation into three
parts, so that [ j[
-1{2
[ j ( j. z)[ equals
(2.9)
-t
1/2
k=-o
ot
k
1 ot
k
z
k
t
1/2
k=-t
1/2
ot
k
1 ot
k
z
k
k=t
1/2
ot
k
1 ot
k
z
k
.
We will control each of these term separately. The rst and the third are easiest.
Consider
(z 1)
-t
1/2
k=-o
ot
k
1 ot
k
z
k
.
We wish to show this is bounded by a constant that is independent of j and c.
Summing by parts, we can write the argument of the absolute value as
(2.10)
ot
-t
1/2
1
1 ot
-t
1/2
1
z
-t
1/2
1
(1 t)
-t
1/2
k=-o
ot
k
(1 ot
k
)(1 ot
k1
)
z
k
.
We have t
-t
1/2
1
~ e
2
and [z
-t
1/2
1
[ ~ e
-1
(where e ~ 2.718). The
denominator of the rst term is therefore bounded from zero. Thus the absolute
value of this term is bounded by a constant. For the second term of (2.10) we can
bring the absolute value inside of the summation to get
(1 t)
-t
1/2
k=-o
ot
k
(1 ot
k
)(1 ot
k1
)
[z[
k
.
The rst absolute value term stays bounded above by a constant times the value
at k = c
-1{2
. Therefore, replacing this by a constant, we can sum in [z[ and
get [z[
-t
1/2
,(1 1,[z[). The numerator, as noted before, is like e
-1
but the
denominator is like c
1{2
,2. This is cancelled by the term 1 t = O(c
1{2
) in
front. Thus the absolute value is bounded.
The argument for the third term of (2.9) works in the same way except that
rather than multiplying by [1 z[ and showing the result is constant, we multiply
by [1 tz[. This is, however, sufcient since [1 tz[ and [1 z[ are effectively
the same for z near 1, which is where our desired bound must be shown carefully.
We now turn to the middle term in (2.9), which is more difcult. We will show
that
(1 z)
t
1/2
k=-t
1/2
ot
k
1 ot
k
z
k
= O(log [ j[).
This is of smaller order than [ j[ raised to any positive real power and thus nishes
the proof. For the sake of simplicity we will rst show this with z = 1c
1{2
. The
general argument for points z of the same radius and nonzero angle is very similar
as we will observe at the end of the proof. For the special choice of z, the prefactor
(1 z) = c
1{2
.
The method of proof is to show that this sum is well approximated by a Riemann
sum. This idea was mentioned in the formal proof of the c goes to zero limit. In
fact, the argument below can be used to make that formal observation rigorous and
thus provides an alternative method to the complex analytic approach we take in
the proof of Lemma 2.15. The sum we have is given by
(2.11) c
1{2
t
1/2
k=-t
1/2
ot
k
1 ot
k
z
k
= c
1{2
t
1/2
k=-t
1/2
o(1 c
1{2
O(c))
k
1 o(1 2c
1{2
O(c))
k
where we have used the fact that tz = 1 c
1{2
O(c).
Observe that if k = t c
-1{2
, then this sum is close to a Riemann sum for
(2.12)
_
1
-1
oe
-t
1 oe
-2t
Jt.
We use this formal relationship to prove that the sum in (2.11) is O(log [ j[). We do
this in a few steps. The rst step is to consider the difference between each term in
our sum and the analogous term in a Riemann sum for the integral. After estimating
the difference we show that this can be summed over k and gives us a nite error.
The second step is to estimate the error of this Riemann sum approximation to the
actual integral. The nal step is to note that
_
1
-1
oe
-t
1 oe
-2t
Jt ~ [log(1 o)[ ~ log [ j[
for o C (in particular, where [1 o[ > c
1{2
). Hence it is easy to check that it is
smaller than any power of [ j[.
A single term in the Riemann sum for the integral looks like
c
1{2
oe
-kt
1/2
1 oe
-2kt
1/2
.
Thus we are interested in estimating
(2.13) c
1{2
o(1 c
1{2
O(c))
k
1 o(1 2c
1{2
O(c))
k

oe
-kt
1/2
1 oe
-2kt
1/2
.
We claim that there exists C < o, independent of c and k, satisfying kc
1{2
_ 1
such that the previous line is bounded above by
(2.14)
Ck
2
c
3{2
(1 o o2kc
1{2
)
Ck
3
c
2
(1 o o2kc
1{2
)
2
.
To prove that (2.13) _ (2.14), we expand the powers of k and the exponentials.
For the numerator and denominator of the rst term inside of the absolute value in
(2.13) we have o(1 c
1{2
O(c))
k
= o okc
1{2
O(k
2
c) and
1 o(1 2c
1{2
O(c))
k
= 1 o o2kc
1{2
o2k
2
c O(kc) O(k
3
c
3{2
)
= (1 o o2kc
1{2
)
_
1
o2k
2
c O(kc) O(k
3
c
3{2
)
1 o o2kc
1{2
_
.
Using 1,(1 z) = 1 z O(z
2
) for [z[ < 1 we see that
o(1 c
1{2
O(c))
k
1 o(1 2c
1{2
O(c))
k
=
o okc
1{2
O(k
2
c)
1 o o2kc
1{2
_
1
o2k
2
c O(kc) O(k
3
c
3{2
)
1 o o2kc
1{2
_
=
(o okc
1{2
O(k
2
c))(1 o o2kc
1{2
o2k
2
c O(kc) O(k
3
c
3{2
))
(1 o o2kc
1{2
)
2
.
Likewise, the second term from (2.13) can be estimated and shown to be
oe
-kt
1/2
1 oe
-2kt
1/2
=
(o okc
1{2
O(k
2
c))(1 o o2kc
1{2
o2k
2
c O(k
3
c
3{2
))
(1 o o2kc
1{2
)
2
.
Taking the difference of these two terms and noting the cancellation of a number
of the terms in the numerator gives (2.14).
To see that the error in (2.14) is bounded after the summation over k in the range
{c
-1{2
. . . . . c
-1{2
], note that this gives
c
1{2
t
1/2
-t
1/2
(2kc
1{2
)
2
1 o o(2kc
1{2
)
(2kc
1{2
)
3
(1 o o(2kc
1{2
))
2
~
_
1
-1
(2t )
2
1 o o2t
(2t )
3
(1 o o2t )
2
Jt.
The Riemann sums and integrals are easily shown to be convergent for our o,
which lies on C, which is roughly the unit circle, and avoids the point 1 by dis-
tance c
1{2
.
Having completed this rst step, we now must show that the Riemann sum for
the integral in (2.12) converges to the integral. This uses the following estimate,
(2.15)
t
1/2
k=-t
1/2
c
1{2
max
(k-1{2)t
1/2
_t _(k1{2)t
1/2
oe
-kt
1/2
1 oe
-2kt
1/2

oe
-t
1 oe
-2t
_ C.
To show this, observe that for t c
1{2
k
1
2
. k
1
2
| we can expand the second
fraction as
(2.16)
oe
-kt
1/2
(1 O(c
1{2
))
1 oe
-2kt
1/2
(1 2lc
1{2
O(c))
where l
1
2
.
1
2
|. Factoring the denominator as
(2.17) (1 oe
-2kt
1/2
)
_
1
oe
-2kt
1/2
(2lc
1{2
O(c))
1 oe
-2kt
1/2
_
.
we can use 1,(1 z) = 1 z O(z
2
) (valid since [1 oe
-2kt
1/2
[ > c
1{2
and
[l[ _ 1) to rewrite (2.16) as
oe
-kt
1/2
(1 O(c
1{2
))
_
1
oe
2ks
1/2
(2It
1/2
O(t))
1-oe
2ks
1/2
_
1 oe
-2kt
1/2
.
Canceling terms in this expression with the terms in the rst part of (2.15), we nd
that we are left with terms bounded by
O(c
1{2
)
1 oe
-2kt
1/2

O(c
1{2
)
(1 oe
-2kt
1/2
)
2
.
These must be summed over k and multiplied by the prefactor c
1{2
. Summing
over k, we nd that these are approximated by the integrals
c
1{2
_
1
-1
1
1 o o2t
Jt. c
1{2
_
1
-1
1
(1 o o2t )
2
Jt.
where [1 o[ > c
1{2
. The rst integral has a logarithmic singularity at t = 0,
which gives [log(1 o)[. This is clearly bounded by a constant time [log c
1{2
[ for
o C. When multiplied by c
1{2
, this term is clearly bounded in c. Likewise, the
second integral diverges like [1,(1 o)[, which is bounded by c
-1{2
, and again
multiplying by the c
1{2
factor in front shows that this term is bounded. This proves
the Riemann sum approximation.
This estimate completes the proof of the desired bound when z = 1c
1{2
. The
general case of [z[ = 1c
1{2
is proved along a similar line by letting z = 1jc
1{2
for j on a suitably dened contour such that z lies on the circle of radius 1 c
1{2
.
The prefactor is no longer c
1{2
but rather jc
1{2
, and all estimates must take into
account j. However, going through this carefully one nds that the same sort of
estimates as above hold and hence the theorem is proved in general.
This lemma completes the proof of Proposition 2.5.
PROOF OF PROPOSITION 2.6. We will focus on the growth of the absolute
value of the determinant. Recall that if 1 is trace class, then [det(1 1)[ _
e
|1|
1
. Furthermore, if 1 can be factored into the product 1 = T where
and T are Hilbert-Schmidt, then [1[
1
_ [[
2
[T[
2
. We will demonstrate such a
factorization and follow this approach to control the size of the determinant.
Dene : 1
2
(
I
(
) 1
2
(
I
)
) and T : 1
2
(
I
)
) 1
2
(
I
(
) via the kernels
(
. j) =
e
-[Im(
()[
j
.
T( j.
) = e
[Im(
()[
e
-
T
3
(
(
3
- )
3
)o ;
2
1{3
( j)
;
sin( z)
.
where we let z = 2
1{3
(
j). Notice that we have put the factor e

-[Im(
()[
into the
kernel and removed it from the T-contour. The point of this is to help control
the kernel without signicantly impacting the norm of the T kernel.
Consider rst [[
2
, which is given by
[[
2
2
=
_
I
(
_
I
)
J
J j
e
-2[Im(
()[
[
j[
2
.
The integral in j converges and is independent of

(recall that [
j[ is bounded
away from zero), while the remaining integral in

is clearly convergent (it is ex-
ponentially small as

goes away from zero along

I
(
). Thus [[
2
< c with no
dependence on j at all.
We now turn to computing [T[
2
. First consider the cubic term

3
. The contour
I
(
is parametrized by c
3
,2 c
3
i r for r (o. o), that is, a straight up-and-
down line just to the left of the ,-axis. By plugging this parametrization in and
cubing it, we see that Re(
3
) behaves like [Im(
)[
2
. This is crucial; even though
our contours are parallel and only differ horizontally by a small distance, their
relative locations lead to very different behavior for the real part of their cube. For
j on the right of the ,-axis, the real part still grows quadratically, although with a
negative sign. This is important because this implies that [e
-(T{3)(
(
3
- )
3
)
[ behaves
like the exponential of the real part of the argument, which is to say, like
e
-
T
3
([Im(
()[
2
[Im( ))[
2
)
.
Turning to the j-term, observe that
[( j)
- ;
[ = e
Re(log [ [i arg(- ))(-Re( ;)-i Im( ;))j
= e
-log [ [ Re( ;)arg(- ) Im( ;)
.
The cosecant term behaves, for large Im( z), like e
-t[Im( ;)[
, and putting all these
estimates together gives that for

and j far from the origin on their respective
contours, [T( j.
)[ behaves like the following product of exponentials:

e
[Im(
()[
e
-
T
3
([Im(
()[
2
[Im( ))[
2
)
e
-log [ [ Re( ;)arg(- ) Im( ;)-t[Im( ;)[
.
Now observe that, due to the location of the contours, Re( z) is constant and less
than 1 (in fact, equal to
1
2
by our choice of contours). Therefore we may factor out
the term e
-log [ [ Re( ;)
= [ j[
for =
1
2
< 1.
The Hilbert-Schmidt norm of what remains is clearly nite and independent
of j. This is just due to the strong exponential decay from the quadratic terms
Im()
2
and Im(j)
2
in the exponential. Therefore we nd that [T[
2
_ c[ j[
for some constant c.

This shows that [1
csc
o
[
1
behaves like [ j[
for < 1. Using the bound that

[det(1 1
csc
o
)[ _ e
|1
csc
u
|
, we nd that [det(1 1
csc
o
)[ _ e
[ [
. Comparing this
to e
-
we have our desired result. Note that the proof also shows that 1
csc
o
is trace
class.
Proofs from Section 2.2
PROOF OF LEMMA 2.10. Before starting this proof, we remark that the choice
(2.3) of k(0) was specically to make the calculations in this proof more tractable.
Certainly other choices of contours would do; however, the estimates could be
harder. As it is, we used Mathematica as a preliminary tool to assist us in comput-
ing the series expansions and simplifying the resulting expressions.
Dene g(j) = (j)n
0
log(j). We wish to control the real part of this function
for both the j-contour and the -contour. Combining these estimates proves the
lemma.
We may expand g(j) into powers of c with the expression for j in terms of k(0)
from (2.3) with =
1
2
(similarly
1
2
for the -expansion). Doing this we see that
the n
0
log(j) term plays an important role in canceling the log(c) term in the ,
and we are left with
(2.18) Re(g(j)) =
c
-1{2
_
1
4
c
-1{2
T cot
2
(
0
2
)
1
S
T k(0)|
2
cot
2
(
0
2
)
_
O(1).
We must show that everything in the parentheses above is bounded below by a pos-
itive constant times [j [ for all j that start at roughly angle lc
1{2
. Equivalently,
we can show that the terms in the parentheses behave bounded below by a positive
constant times [ 0[, where 0 is the polar angle of j.
The second part of this expression is clearly positive regardless of the value
of . What this suggests is that we must show (in order to also be able to deal with
=
1
2
corresponding to the -estimate) that for j starting at angle lc
1{2
and going
to zero, the rst term dominates (if l is large enough).
To see this, we rst note that since =
1
2
, the rst term is clearly positive and
dominates for 0 bounded away from . This proves the inequality for any range
of j with 0 bounded from . Now observe the following asymptotic behavior of
the following three functions of 0 as 0 goes to :
cot
2
_
0
2
_
~
1
4
( 0)
2
. tan
2
_
0
2
_
~ 4( 0)
-2
.
log
2
_
2
1 cos(0)
_
~
1
16
( 0)
4
.
The behavior expressed above is dominant for 0 close to . We may expand the
square in the second term in (2.18) and use the above expressions to nd that for
some suitable constant C > 0 (which depends on X and T only), we have
Re(g(j)) = c
-1{2
_
1
16
c
-1{2
T( 0)
2
C( 0)
2
_
O(1).
Now use the fact that 0 _ lc
1{2
to give
(2.19) Re(g(j)) = c
-1{2
_
1
16
lT( 0)
X
2
8T
( 0)
2
_
O(1).
Since 0 is bounded by , we see that taking l large enough, the rst term always
dominates for the entire range of 0 0. lc
1{2
|. Therefore since =
1
2
, we
nd that we have have the desired lower bound in c
-1{2
and [ 0[.
Turn now to the bound for Re(g()). In the case of the j-contour we took
=
1
2
; however, since we now are dealing with the -contour, we must take
=
1
2
. This change in the sign of and the argument above shows that (2.19)
implies the desired bound for Re(g()) for l large enough.
Before proving Lemma 2.11, we record the following key lemma on the mero-
morphic extension of j (j. z). Recall that j (j. z) has poles at j = t
}
, Z.
LEMMA 2.23 For j = t
}
, Z, j (j. z) is analytic in z for 1 < [z[ < t
-1
and extends analytically to all z = 0 or t
k
for k Z. This extension is given by
rst writing j (j. z) = g
(z) g
-
(z) where
g
(z) =
o
k=0
jt
k
z
k
1 t
k
j
. g
-
(z) =
o
k=1
jt
-k
z
-k
1 t
-k
j
.
and where g
is now dened for [z[ < t

-1
and g
-
is dened for [z[ > 1. These
functions satisfy the following two functional equations, which imply the analytic
continuation:
g
(z) =
j
1 tz
jg
(tz). g
-
(z) =
1
1 z
1
j
g
-
(z,t).
By repeating this functional equation we nd that
g
(z) =
1
k=1
j
k
1 t
k
z
j
1
g
(t
1
z).
g
-
(z) =
1-1
k=0
j
-k
1 t
-k
z
j
-1
g
-
(zt
-1
).
PROOF. We prove the g
functional equation, since the g

-
one follows simi-
larly. Observe that
g
(z) =
o
k=0
j(tz)
k
_
1
1
1 jt
k
1
_
=
j
1 tz
k=0
j
2
t
k
1 jt
k
(tz)
k
=
j
1 tz
jg
(tz).
which is the desired relation.
PROOF OF LEMMA 2.11. Recall that j lies on a compact subset of

C and hence
that [1 jt
k
[ stays bounded from below as k varies. Also, observe that due to our
choices of contours for j
t
and , [,j
t
[ stays bounded in (1. t
-1
). Write z = ,j
t
.
Split j ( j. z) as g
(z) g
-
(z) (see Lemma 2.23 above); we see that g
(z) is
bounded by a constant time 1,(1tz) and likewise g
-
(z) is bounded by a constant
time 1,(1 z). Writing this in terms of and j
t
again, we have our desired upper
bound.
PROOF OF LEMMA 2.14. By the discussion preceding the statement of this
lemma it sufces to consider the expansion without n
0
log(,j
t
) and without the
log c term in m since, as we will see, they exactly cancel out. Therefore, for the
sake of this proof we modify the denition of m given in (1.27) to be
m =
1
2
_
c
-1{2
_
a
t
X
2
2T
_
1
2
t .
_
where a
t
= a log 2.
The argument now amounts to a Taylor series expansion with control over the
remainder term. Let us start by recording the rst four derivatives of ():
() = . log(1 )
t
1
mlog .
t
() =
.
1
t
(1 )
2

m
tt
() =
.
(1 )
2

2t
(1 )
3

m
2
.
ttt
() =
2.
(1 )
3

6t
(1 )
4

2m
3
.
tttt
() =
6.
(1 )
4

24t
(1 )
5

6m
4
.
We Taylor-expand () = () () around and then expand in c as c goes
to zero and nd that
t
() =
a
t
2
c
-1{2
O(1).
tt
() = O(c
-1{2
).
ttt
() =
T
8
c
-3{2
O(c
-1
).
tttt
() = O(c
-3{2
).
A Taylor series remainder estimate then shows that
()
_
t
()( )
1
2
tt
()( )
2
1
3
ttt
()( )
3
_
_
sup
t B(,[(-[)
1
4
[
tttt
(t )[[ [
4
.
where T(. [ [) denotes the ball around of radius [ [. Now considering
the scaling we have that = c
-1
3
c
1{2
, so that when we plug this in along with

the estimates on derivatives of at , we nd that the equation above becomes
()
_
2
1{3
a
t

T
3
3
_
= O(c
1{2
).
From this we see that if we included the log c term in with m it would, as claimed,
exactly cancel the n
0
log(,j
t
) term. The above estimate therefore proves the de-
sired rst claimed result.
The second result follows readily from [e
;
e
u
[ _ [zn[ max{[e
;
[. [e
u
[] and
the rst result, as well as the boundedness of the limiting integrand.
PROOF OF LEMMA 2.15. Expanding in c we have that
z =
c
-1
3
c
1{2
c
-1
3
c
1{2
j
t
= 1 c
1{2
z O(c)
where the error is uniform for our range of j
t
and

and where
z = c
-1
3
(
j
t
).
We now appeal to the functional equation for , explained in Lemma 2.23. There-
fore we wish to study c
1{2
g
(z) and c
1{2
g
-
(z) as c goes to 0 and show that they
converge uniformly to suitable integrals.
First consider the g
case. Let us, for the moment, assume that [ j[ < 1. We

know that [tz[ < 1; thus for any N _ 0, we have
c
1{2
g
(z) = c
1{2
1
k=1
j
k
1 t
k
z
c
1{2
j
1
g
(t
1
z).
Since, by assumption, [ j[ < 1, the rst sum is the partial sum of a convergent
series. Each term may be expanded in c. Noting that
1 t
k
z = 1 (1 2c
1{2
O(c))(1 c
1{2
z O(c))
= (2k z)c
1{2
kO(c).
we nd that
c
1{2
j
k
1 t
k
z
=
j
k
2k z
kO(c
1{2
).
The last part of the expression for g
is bounded in c; thus we end up with the

following asymptotics:
c
1{2
g
(z) =
1
k=1
j
k
2k z
N
2
O(c
1{2
) j
1
O(1).
It is possible to choose N(c), which goes to innity, such that N
2
O(c
1{2
) = o(1).
Then for any xed compact set contained in C \ {2. 4. 6. . . .] we have uni-
form convergence of this sequence of analytic functions to some function, which
is necessarily analytic and equals
o
k=1
j
k
2k z
.
This expansion is valid for [ j[ < 1 and for all z C \ {2. 4. 6. . . .].
Likewise for c
1{2
g
-
(z), for [ j[ > 1, and for z C \ {2. 4. 6. . . .], we
have uniform convergence to the analytic function
0
k=-o
j
k
2k z
.
We now introduce the Hurwitz-Lerch transcendental function and relate some
basic properties of it, which can be found in [30]:
(a. s. n) =
o
k=0
a
k
(n k)
x
for n > 0 real and either [a[ < 1 and s C or [a[ = 1 and Re(s) > 1. For
Re(s) > 0 it is possible to analytically extend this function using the integral
formula
(a. s. n) =
1
I(s)
_
o
0
e
-(u-1)t
e
t
a
t
x-1
Jt.
where additionally a C \ 1. o) and Re(n) > 0.
Observe that we can express our series in terms of this function as
o
k=1
j
k
2k z
=
1
2
j( j. 1. 1 z,2).
0
k=-o
j
k
2k z
=
1
2
( j
-1
. 1. z,2).
These two functions can be analytically continued using the integral formula onto
the same region where Re(1 z,2) > 0 and Re( z,2) > 0, i.e., where Re( z,2)
(1. 0). Additionally, the analytic continuation is valid for all j not along R
.
We now wish to use Vitalis convergence theorem to conclude that j ( j. z)
converges uniformly for general j to the sum of these two analytic continuations.
In order to do that, we need a priori boundedness of c
1{2
g
and c
1{2
g
-
for com-
pact regions of j away from R
. This, however, can be shown directly as follows.

By assumption on j, we have that [1t
k
j[ > c
-1
for some positive constant c.
Consider c
1{2
g
rst.
[c
1{2
g
(z)[ _ c
1{2
j
o
k=0
[tz[
k
[1 t
k
j[
_ cc
1{2
1
1 [tz[
.
We know that [tz[ is bounded to order c
1{2
away from 1 and therefore this shows
that [c
1{2
g
(z)[ has an upper bound uniform in j. Likewise, we can do a similar

computation for c
1{2
g
-
(z) and nd the same result, this time using that [z[ is
bounded to order c
1{2
away from 1.
As a result of this a priori boundedness, uniform in j, we have that for com-
pact sets of j away from R
, uniformly in c, c
1{2
g
and c
1{2
g
-
are uniformly
bounded as c goes to zero. Therefore Vitalis convergence theorem implies that
they converge uniformly to their analytic continuation.
Now observe that
1
2
j( j. 1. 1 z,2) =
1
2
_
o
0
je
- ;t {2
e
t
j
Jt
and
1
2
( j
-1
. 1. z,2) =
1
2
_
o
0
e
-(- ;{2-1)t
e
t
1, j
Jt =
1
2
_
0
-o
je
- ;t {2
e
t
j
Jt.
Therefore, by a simple change of variables in the second integral, we can combine
these as a single integral
1
2
_
o
-o
je
- ;t {2
e
t
j
Jt =
1
2
_
o
0
js
- ;{2
s j
Js
s
.
The rst of the above equations proves the lemma, and for an alternative expression
we use the second of the integrals (which followed from the change of variables
e
t
= s) and thus, on the region Re( z,2) (1. 0), this integral converges and
equals
1
2
( j)
- ;
csc( z,2).
This function is, in fact, analytic for j C \ 0. o) and for all z C \ 2Z.
Therefore it is the analytic continuation of our asymptotic series.
3 Weakly Asymmetric Limit of the Corner Growth Model
Recall the denitions in Section 1.2 of WASEP, its height function (1.21), and,
for X cZ and T _ 0,
(3.1) 7
t
(T. X) =
1
2
c
-1{2
exp{z
t
h
t
1/2(c
-2
T. c
-1
X|) v
t
c
-2
T ]
where, for c (0.
1
4
), let =
1
2

1
2
c
1{2
and q =
1
2

1
2
c
1{2
; v
t
and z
t
are as in
(1.23) and (1.24), and the closest integer .| is given by
.| = ].
1
2
.
Let us describe in simple terms the dynamics in T of 7
t
(T. X) dened in (3.1).
It grows continuously exponentially at rate c
-2
v
t
and jumps at rates
r
-
(X) = c
-2
q(1 j(.))j(. 1) =
1
4
c
-2
q(1 j(.))(1 j(. 1))
to e
-22
s
7
t
and
r
(X) = c
-2
j(.)(1 j(. 1)) =
1
4
c
-2
(1 j(.))(1 j(. 1))
to e
22
s
7
t
, independently at each site X = c. cZ (recall that j = 2j 1). We
write this as follows:
J7
t
(X) = {c
-2
v
t
(e
-22
s
1)r
-
(X) (e
22
s
1)r
(X)]7
t
(X)JT
(e
-22
s
1)7
t
(X)JM
-
(X) (e
22
s
1)7
t
(X)JM
(X)
where JM
(X) = J1
(X) r
(X)JT where 1
-
(X). 1
(X), X cZ are
independent Poisson processes running at rates r
-
(X). r
(X), and J always refers

to change in macroscopic time T .
Let
D
t
= 2
_
q = 1
1
2
c O(c
2
)
and ^
t
be the cZ Laplacian, ^ (.) = c
-2
( (. c) 2 (.) (. c)). We
also have
1
2
D
t
^
t
7
t
(X) =
1
2
c
-2
D
t
(e
-2
s
)(x1)
2 e
2
s
)(x)
)7
t
(X).
The parameters have been carefully chosen so that
1
2
c
-2
D
t
(e
-2
s
)(x1)
2 e
2
s
)(x)
) =
c
-2
v
t
(e
-22
s
1)r
-
(X) (e
22
s
1)r
(X).
Hence [3, 14]
(3.2) J7
t
=
1
2
D
t
^
t
7
t
7
t
JM
t
where
JM
t
(X) = (e
-22
s
1)JM
-
(X) (e
22
s
1)JM
(X)
are martingales in T with
J(M
t
(X). M
t
(Y )) = c
-1
1(X = Y )b
t
(t
-t
1
Aj
j)JT.
Here t
x
j(,) = j(, .) and
b
t
(j) = 1 j(1) j(0)

b
t
(j)
where
b
t
(j) = c
-1
((e
-22
s
1)
2
4c) q((e
22
s
1)
2
4c)|
q(e
-22
s
1)
2
(e
22
s
1)
2
|( j(1) j(0))
q(e
-22
s
1)
2
(e
22
s
1)
2
c| j(1) j(0)
_
.
Clearly b
t
.

b
t
_ 0. It is easy to check that there is a C < osuch that
b
t
_ Cc
1{2
and, for sufciently small c > 0,
(3.3) b
t
_ 3.
Note that (3.2) is equivalent to the integral equation
7
t
(T. X) = c
Y tZ
t
(T. X Y )7
t
(0. Y )
_
T
0
c
Y tZ
t
(T S. X Y )7
t
(S. Y )JM
t
(S. Y )
(3.4)
where
t
(T. X) are the (normalized) transition probabilities for the continuous-
time random walk with generator
1
2
D
t
^
t
. The normalization is multiplication of
the actual transition probabilities by c
-1
so that
t
(T. X) (T. X) =
e
-A
2
{2T
_
2T
.
We need some a priori bounds.
LEMMA 3.1 For 0 < T _ T
0
, and for each q = 1. 2. . . . . there is a C
q
=
C
q
(T
0
) < osuch that
(i) 17
2
t
(T. X)| _ C
2
2
t
(T. X);
(ii) 1(7
t
(T. X) c
Y tZ
t
(T. X Y )7
t
(0. Y ))
2
| _ C
2
2
t
(T. X);
(iii) 17
2q
t
(T. X)| _ C
q
2q
t
(T. X).
PROOF. Within the proof, C will denote a nite number that does not depend
on any other parameters except T and q but may change from line to line. Also, for
ease of notation, we identify functions on cZ with those on R by (.) = (.|).
First, note that
7
t
(0. Y ) =
1
2
c
-1{2
exp{c
-1
z
t
[Y [] =
1
2
c
-1{2
exp{c
-1{2
[Y [ O(c
1{2
)]
is an approximate delta function, from which we check that
(3.5) c
Y tZ
t
(T. X Y )7
t
(0. Y ) _ C
t
(T. X).
Let
t
(T. X) = 17
2
t
(T. X)|.
From (3.4) and (3.5) we get
(3.6)
t
(T. X) _ C
2
t
(T. X) C
_
T
0
_
o
-o
2
t
(T S. X Y )
t
(S. Y )JS JY.
Iterating, we obtain
(3.7)
t
(T. X) _
o
n=0
C
n
1
n,t
(T. X)
where, for ^
n
= ^
n
(T ) = {0 = t
0
_ T
1
< < T
n
< T ], X
0
= 0,
1
n,t
(T. X) =
_
Z
n
_
R
n
n
i=1
2
t
(T
i
T
i-1
. X
i
X
i-1
)
2
t
(T T
n
. X .
n
)
n
i=1
JX
i
JT
i
.
One readily checks that
1
n,t
(T. X) _ C
n
T
n{2
(n)
-1{2
2
t
(T. X).
from which we obtain (i),
t
(T. X) _ C
o
n=0
(CT )
n{2
(n)
-1{2
2
t
(T. X) _ C
t
2
t
(T. X).
Now we turn to (ii). From (3.4), the term on the right-hand side is bounded by a
constant multiple of
_
T
0
_
o
-o
2
t
(T S. X Y )17
2
t
(S. Y )|JY JS.
Using (i), this is in turn bounded by C
_
T
2
t
(T. X), which proves (ii).
Finally, we prove (iii). Fix a q _ 2. By standard methods of martingale analysis
and (3.3), we have
1
___
T
0
c
Y tZ
t
(T S. X Y )7
t
(S. Y )JM
t
(S. Y )
_
2q
_
_
C1
___
T
0
c
Y tZ
2
t
(T S. X Y )7
2
t
(S. Y )JS
_
q
_
.
Let
g
t
(T. X) =
17
2q
t
(T. X)|
2q
t
(T. X)
.
From the last inequality and Schwarzs inequality, we have
g
t
(T. X) _
C(1
_
Z
0
q
(T)
_
R
q
q
i =1
2
t
(S
i
S
i -1
. X
i
X
i -1
)
2
t
(S
i
. Y
i
)g
1{q
t
(S
i
. Y
i
)JY
i
JS
i
).
Now use the fact that
q
i=1
g
1{q
t
(S
i
. Y
i
) _ C
q
i=1
}yi

2{(q-1)
t
(S
}
. Y
}
)
2
t
(S
i
. Y
i
)
g
t
(S
i
. Y
i
)
and iterate the inequality to obtain (iii).
We now turn to the tightness. In fact, although we are in a different regime,
the arguments of [3] actually extend to our case. For each > 0, let P
t
be
the distributions of the processes {7
t
(T. X)]
_T
on D
u
(. o): D
u
(R)) where
D
u
refers to right-continuous paths with left limits with the topology of uniform
convergence on compact sets. Because the discontinuities of 7
t
(T. ) are restricted
to c(
1
2
Z), it is measurable as a D
u
(R)-valued random function (see [4, sec. 18]).
Since the jumps of 7
t
(T. ) are uniformly small, local uniform convergence works
for us just as well as the standard Skhorohod topology.
The following summarizes results that are contained in [3] but not explicitly
stated there in the form we need.
THEOREM 3.2 [3] There is an explicit < o such that if there exist C. c < o
for which
(3.8)
_
o
-o
7
]
t
(. X)JP
t
_ Ce
c[A[
. X cZ.
then {P
t
]
0_t_1{4
is tight. Any limit point P
is supported on C(. o): C(R))

and solves the martingale problem for the stochastic heat equation (1.5) after
time .
It appears that = 10 works in [3], though it almost certainly can be improved
to = 4. Note that the process level convergence is more than we need for the one-
point function. However, it could be useful in the future. Although not explicitly
stated there, the theorem is proved in [3]. The key point is that all computations in
[3] after the initial time are done using (3.2) for 7
t
, which scales linearly in 7
t
.
So the only input is a bound like (3.8) on the initial data. In [3], this is made as an
assumption, which can easily be checked for initial data close to equilibrium. In
the present case, it follows from (iii) of Lemma 3.1.
The measures P
1
and P
2
for
1
<
2
can be chosen to be consistent on
C(
2
. o). C(R)); consequently, there is a limit measure P on C((0. o). C(R))
that is consistent with any P
when restricted to C(. o). C(R)). From the

uniqueness of the martingale problem for t _ > 0 and the corresponding martin-
gale representation theorem [20], there is a space-time white noise

W , on a possi-
bly enlarged probability space (
C.

F
T
.

P), such that under

P, for any > 0,
7(T. X) =
_
o
-o
(T . X Y )7(. Y )JY
_
T
_
o
-o
(T S. X Y )7(S. Y )

W(JY. JS).
Finally, (ii) of Lemma 3.1 shows that under

P,
_
o
-o
(T . X Y )7(. Y )JY (T. X)
as _ 0, which completes the proof.
4 Alternative Forms of the Crossover Distribution Function
We now demonstrate how the various alternative formulas for J
T
(s) given in
Theorem 1.1 are derived from the cosecant kernel formula of Theorem 1.10.
4.1 Proof of the Crossover Airy Kernel Formula
We prove this by showing that
det(1 1
csc
o
)
1
2
(
I
)
)
= det(1 1
c
T, z J
)
1
2
(k
1
T
o,o)
where 1
c
T, z J
and o
T,
are given in the statement of Theorem 1.1 and k
T
=
2
-1{3
T
1{3
.
The kernel 1
csc
o
( j. j
t
) is given by (1.29) as
_
I
(
e
-
T
3
(
(
3
- )
03
)2
1/3
o(
(- )
0
)
_
2
1{3
_
o
-o
je
-2
1/3
t (
(- )
0
)
e
t
j
Jt
_
J
j
.
where we recall that the inner integral converges since Re(2
1{3
(
j
t
)) =
1
2
(see the discussion in Denition 1.12). For Re(z) < 0 we have the following nice
identity:
_
o
o
e
x;
J. =
e
o;
z
.
which, noting that Re(
j
t
) < 0, we may apply to the above kernel to get
2
2{3
_
I
(
_
o
-o
_
o
o
e
-
T
3
(
(
3
- )
03
)-2
1/3
o )
0 je
-2
1/3
t (
(- )
0
)
e
t
j
e
2
1/3
(o-x) )
e
2
1/3
x
(
J. Jt J
.
This kernel can be factored as a product TC where
: 1
2
(a. o) 1
2
(
I
)
). T : 1
2
(
I
(
) 1
2
(a. o). C : 1
2
(
I
)
) 1
2
(
I
(
).
and the operators are given by their kernels
( j. .) = e
2
1/3
(o-x) )
. T(..
) = e
2
1/3
x
(
.
C(
. j) = 2
2{3
_
o
-o
exp
_
T
3
(
3
j
3
) 2
1{3
a j
_
je
-2
1/3
t (
(- ))
e
t
j
Jt.
Since det(1 TC) = det(1 TC), we consider TC acting on 1
2
(a. o)
with kernel
2
2{3
_
o
-o
_
I
z
(
_
I
z )
e
-
T
3
(
(
3
- )
3
)2
1/3
(x-t )
(-2
1/3
(,-t ) )
j
e
t
j
J j J
Jt.
Using the formula for the Airy function given by
Ai(r) =
_
I
(
exp
_
1
3
z
3
rz
_
Jz
and replacing t with t , we nd that our kernel equals
2
2{3
T
-2{3
_
o
-o
j
j e
-t
Ai(T
-1{3
2
1{3
(. t )) Ai(T
-1{3
2
1{3
(, t ))Jt.
We may now change variables in t as well as in . and , to absorb the factor of
T
-1{3
2
1{3
. To rescale . and , use
det(1 1(.. ,))
1
2
(io,o)
= det(1 r1(r.. r,))
1
2
(o,o)
.
This completes the proof.
4.2 Proof of the Gumbel Convolution Formula
Before starting we remark that throughout this proof we will dispense with the
tilde with respect to j and

C. We choose to prove this formula directly from the
form of the Fredholm determinant given in the crossover Airy kernel formula of
Theorem 1.1. However, we make note that it is possible, and in some ways simpler
(though a little messier), to prove this directly from the cosecant form of the kernel.
Our starting point is the formula for J
T
(s) given in (1.14). The integration in
j occurs along a complex contour, and even though we havent been writing it
explicitly, the integral is divided by 2i . We now demonstrate how to squish this
contour to the positive real line (at which point we will start to write the 2i ). The
pole in the term o
T,
(t ) for j along R
means that the integral along the positive

real axis from above will not exactly cancel the integral from below.
Dene a family of contours C
1
,
2
parametrized by
1
.
2
> 0 (small). The
contours are dened in terms of three sections
C
1
,
2
= C
-
1
,
2
L C
circ
1
,
2
L C
1
,
2
traversed counterclockwise, where
C
circ
1
,
2
= {
2
e
i0
:
1
_ 0 _ 2
1
]
and where C
1
,
2
are horizontal lines extending from
1
e
i
2
to o.
We can deform the original j-contour to any of these contours without chang-
ing the value of the integral (and hence of J
T
(s)). To justify this we use Cauchys
theorem. However, this requires the knowledge that the determinant is an ana-
lytic function of j away from R
. This may be proved similarly to the proof of

Lemma 2.1 and relies on Lemma 2.20. As such we do not include this computation
here.
Fixing
2
for the moment we wish to consider the limit of the integrals over these
contours as
1
goes to zero. The resulting integral may be written as 1
circ
2
1
line
2
where
1
circ
2
=
_
[[=
2
Jj
j
e
-
det(1 1
T,
)
1
2
(k
1
T
o,o)
.
1
line
2
= lim
1
-0
_
o
2
Jj
j
e
-
det(1 1
T,i
i
) det(1 1
T,-i
i
)|.
Claim 4.1. 1
circ
2
exists and lim
2
-0
1
circ
2
= 1.
PROOF. It is easiest, in fact, to prove this claim by replacing the determinant by
the cosecant determinant, (1.29). From that perspective the j at 0 and at 2 are on
opposite sides of the branch cut for log(j) but are still dened (hence the 1
circ
2
is clearly dened). As far as computing the limit, one can do the usual Hilbert-
Schmidt estimate and show that, uniformly over the circle [j[ =
2
, the trace norm
goes to 0 as
2
goes to 0. Thus the determinant goes uniformly to 1 and the claim
follows.
Turning now to 1
line
2
, we see that this limit exists by going to the equivalent
cosecant kernel (where this limit is trivially just the kernel on different levels of
the log(j) branch cut). Notice now that we can write the operator 1
T,i
1
=
1
sym
1
1
asym
1
and likewise 1
T,-i
1
= 1
sym
1
1
asym
1
where 1
sym
1
and 1
asym
1
also act on 1
2
(k
-1
T
a. o) and are given by their kernels
1
sym
1
(.. ,) =
_
o
-o
j(j b)
2
1
(j b)
2
2
1
Ai(. t ) Ai(, t )Jt.
1
asym
1
(.. ,) =
_
o
-o
i
1
b
(j b)
2
2
1
where b = b(t ) = e
-k
T
t
.
From this it follows that
1
sym
(.. ,) := lim
1
-0
1
sym
1
(.. ,) = P. V.
_
j
j e
-k
T
t
As far as 1
asym
1
, since jb has a unique root at t
0
= k
-1
T
log j, it follows from
the Plemelj formula [11] that
lim
1
-0
1
asym
1
(.. ,) =
i
k
T
Ai(. t
0
) Ai(, t
0
).
With this in mind we dene
1
asym
(.. ,) =
2i
k
T
Ai(. t
0
) Ai(, t
0
).
We see that 1
asym
is a multiple of the projection operator onto the shifted Airy
functions.
We may now collect the calculations from above and we nd that
1
line
2
=
1
2i
_
o
2
Jj
j
e
-
_
det
_
1 1
sym
1
2
1
asym
_
det
_
1 1
sym
1
2
1
asym
__
=
1
2i
_
o
2
Jj
j
e
-
det(1 1
sym
)tr((1 1
sym
)
-1
1
asym
)
where both 1
sym
and 1
asym
act on 1
2
(k
-1
T
a. o). Here we have used the fact that
1
asym
is rank 1, and if you have and T, where T is rank 1, then
det(1 T) = det(1 ) det(1 (1 )
-1
T)
= det(1 )1 tr(1 )
-1
T)|.
As stated above, weve only shown the pointwise convergence of the kernels to
1
sym
and 1
asym
. However, using the decay properties of the Airy function and the
exponential decay of o, this can be strengthened to trace class convergence.
We may now take
2
to zero and nd that
J
T
(s) = lim
2
-0
(1
circ
2
1
line
2
)
= 1
1
2i
_
o
0
Jj
j
e
-
det(1 1
sym
) tr((1 1
sym
)
-1
1
asym
)
with 1
sym
and 1
asym
as above acting on 1
2
(k
-1
T
a. o) and where the integral is
improper at zero.
We can simplify our operators by changing variables and replacing . by . t
0
and , by ,t
0
. We can also change variables from jto e
-i
. With this in mind, we
redene the operators 1
sym
and 1
asym
to act on 1
2
(k
-1
T
(a r). o) with kernels
1
sym
(.. ,) = P. V.
_
o(t ) Ai(. t ) Ai(, t )Jt.
1
asym
(.. ,) = Ai(.) Ai(,).
where o(t ) =
1
1-e
k
T
I
. In terms of these operators we have
J
T
(s) = 1
_
o
-o
e
-e
(a r)Jr
where
(r) = k
-1
T
det(1 1
sym
)
1
2
(k
1
T
i,o)
tr((1 1
sym
)
-1
1
asym
)
1
2
(k
1
T
i,o)
.
Calling G(r) = e
-e
and observing that 1

sym
= 1
c
T
and 1
asym
= P
Ai
com-
pletes the proof of the rst part of the Gumbel convolution formula.
Turning now to the Hilbert transform formula, we may isolate the singularity of
o
T
(t ) from the above kernel 1
sym
(or 1
c
T
) as follows: Observe that we may write
o
T
(t ) as
o
T
(t ) = o
T
(t )
1
k
T
t
where o
T
(t ) (given in (1.18)) is a smooth function, nondecreasing on the real line,
with o
T
(o) = 0 and o
T
(o) = 1. Moreover, o
t
T
is an approximate delta
function with width k
-1
T
= 2
1{3
T
-1{3
. The principal value integral of the o
T
(t )
term can be replaced by a simple integral. The new term gives
P. V.
_
1
k
T
t
Ai(. t ) Ai(, t ).
This is k
-1
T
times the Hilbert transform of the product of Airy functions, which is
explicitly computable [38] with the result
P. V.
_
1
k
T
t
Ai(. t ) Ai(, t ) = k
-1
T
G
(x-,){2
_
. ,
2
_
where G
o
(.) is given in (1.18).
5 Formulas for a Class of Generalized Integrable Integral Operators
Presently we will consider a certain class of Fredholm determinants and make
two computations involving these determinants. The second of these computations
closely follows the work of Tracy and Widom and is based on a similar calculation
done in [32]. In that case the operator in question is the Airy operator. We deal
with the family of operators that arise in considering J
T
(s).
Consider the class of Fredholm determinants det(1 1)
1
2
(x,o)
with operator 1
acting on 1
2
(s. o) with kernel
(5.1) 1(.. ,) =
_
o
-o
o(t ) Ai(. t ) Ai(, t )Jt.
where o(t ) is a function that is smooth except at a nite number of points at which
it has bounded jumps and that approaches 0 at oand 1 at oexponentially fast.
These operators are, in a certain sense, generalizations of the class of integrable
integral operators (see [5]).
The kernel can be expressed alternatively as
(5.2) 1(.. ,) =
_
o
-o
o
t
(t )
(. t )[(, t ) [(. t )(, t )
. ,
Jt.
with (.) = Ai(.) and [(.) = Ai
t
(.), Ai(.) being the Airy function. This and
the entire generalization we will now develop is analogous to what is known for
the Airy operator, which is dened by its kernel 1
Ai
(.. ,) on 1
2
(o. o),
1
Ai
(.. ,) =
_
o
-o
(t ) Ai(. t ) Ai(, t )Jt =
Ai(.) Ai
t
(.) Ai(,) Ai
t
(.)
. ,
.
where presently (t ) = 1
t _0
.
Note that the o(t ) in our main result is not exactly of this type. However, one
can smooth out the o and apply the results of this section to obtain formulas, which
then can be shown to converge to the desired formulas as the smoothing is removed.
It is straightforward to control the convergence in terms of trace norms, so we will
not provide further details here.
5.1 Symmetrized Determinant Expression
It is well-known that
det(1 1
Ai
)
1
2
(x,o)
= det(1
_
x
1
Ai
_
x
)
1
2
(-o,o)
where
x
is the multiplication operator by 1
-_x
(i.e., (
x
)(.) = 1(. _ s) (.)).
The following proposition shows that for our class of determinants the same
relation holds and provides the proof of (1.16) in Theorem 1.1.
PROPOSITION 5.1 For 1 in the class of operators with kernel as in (5.1),
det(1 1)
1
2
(x,o)
= det(1

1
x
)
1
2
(-o,o)
.
where the kernel for

1
x
is given by
1
x
(.. ,) =
_
o(. s) 1(.. ,)
_
o(, s).
PROOF. Dene 1
x
: 1
2
(s. o) 1
2
(o. o) by
(1
x
)(.) =
_
o
x
Ai(. ,) (,)J,.
Also, dene o : 1
2
(o. o) 1
2
(o. o) by (o )(.) = o(.) (.), and
similarly
x
: 1
2
(o. o) 1
2
(s. o) by (
x
)(.) = 1(. _ s) (.). Then
1 =
x
1
-o
o1
x
.
We have
det(1 1)
1
2
(x,o)
= det(1

1
x
)
1
2
(-o,o)
where
1
x
=
_
o 1
x
x
1
-o
_
o.
The key point is that
1
x
x
1
-o
(.. ,) = 1
Ai
(. s. , s)
where 1
Ai
is the Airy kernel. One can also see now that this operator is self-adjoint
on the real line.
5.2 Painlev II Type Integrodifferential Equation
We now express det(1 1)
1
2
(x,o)
as an integrodifferential equation. This
provides the proof of Proposition 1.2.
Recall that J
GUE
(s) = det(1 1
Ai
)
1
2
(x,o)
can be expressed in terms of a
nonlinear version of the Airy function, known as Painlev II, as follows. Let q be
the unique (Hastings-McLeod) solution to Painlev II:
J
2
Js
2
q(s) = (s 2q
2
(s))q(s)
subject to q(s) ~ Ai(s) as s o. Then
J
2
Js
2
log det(1 1
Ai
)
1
2
(x,o)
= q
2
(s).
From this one shows that
J
GUE
(s) = exp
_
_
o
x
(. s)q
2
(.)J.
_
.
See [32] for details. We now show that an analogous expression exists for the class
of operators described in (5.1).
PROPOSITION 5.2 For 1 in the class of operators with kernel as in (5.1), let q(t. s)
be the solution to
(5.3)
J
2
Js
2
q
t
(s) =
_
s t 2
_
o
-o
o
t
(r)q
2
i
(s)Jr
_
q
t
(s)
subject to q
t
(s) ~ Ai(t s) as s o. Then we have
(5.4)
J
2
Js
2
log det(1 1)
1
2
(x,o)
=
_
o
-o
o
t
(t )q
2
t
(s)Jt.
det(1 1)
1
2
(x,o)
= exp
_
_
o
x
(. s)
_
o
-o
o
t
(t )q
2
t
(.)Jt J.
_
.
PROOF. As mentioned, we follow the work of Tracy and Widom [32] very
closely and make the necessary modications to our present setting. Consider an
operator 1 of the type described in (5.1). We use the notation 1
.
= 1(.. ,) to
indicate that operator 1 has kernel 1(.. ,). It will be convenient to think of our
operator 1 as acting, not on (s. o), but on (o. o) and to have kernel
1(.. ,)
x
(,)
where is the characteristic function of (s. o). Since the integral operator 1
is trace class and depends smoothly on the parameter s, we have the well-known
formula
(5.5)
J
Js
log det(1 1) = tr
_
(1 1)
-1
d1
ds
_
.
By calculus
(5.6)
d1
ds
.
= 1(.. s)(, s).
Substituting this into the above expression gives
J
Js
log det(1 1) = 1(s. s)
where 1(.. ,) is the resolvent kernel of 1, i.e., 1 = (1 1)
-1
1
.
= 1(.. ,).
The resolvent kernel 1(.. ,) is smooth in . but discontinuous in , at , = s. The
quantity 1(s. s) is interpreted to mean the limit of 1(s. ,) as , goes to s from
above:
lim
,-x
C
1(s. ,).
Representation for R.x; y/
If M denotes the multiplication operator, (M )(.) = . (.), then
M. 1|
.
= .1(.. ,) 1(.. ,), = (. ,)1(.. ,) =
_
o
-o
o
t
(t ){(. t )[(, t ) [(. t )(, t )]Jt
where (.) = Ai(.) and [(.) = Ai
t
(.). As an operator equation this is
M. 1| =
_
o
-o
o
t
(t ){t
t
t
t
[ t
t
[ t
t
]Jt.
where a b
.
= a(.)b(,) and . | denotes the commutator. The operator t
t
acts
as (t
t
)(.) = (. t ). Thus
M. (1 1)
-1
| = (1 1)
-1
M. 1|(1 1)
-1
=
_
o
t
(t ){(1 1)
-1
(t
t
t
t
[ t
t
[ t
t
)(1 1)
-1
]Jt
=
_
o
t
(t ){Q
t
1
t
1
t
Q
t
]Jt. (5.7)
where we have introduced
Q
t
(.: s) = Q
t
(.) = (1 1)
-1
t
t
.
1
t
(.: s) = 1
t
(.) = (1 1)
-1
t
t
[.
Note an important point here that as 1 is self-adjoint we can use the transformation
t
t
t
t
[(1 1)
-1
= t
t
(1 1)
-1
t
t
[. On the other hand, since (1 1)
-1
.
=
j(.. ,) = (. ,) 1(.. ,),
(5.8) M. (1 1)
-1
|
.
= (. ,)j(.. ,) = (. ,)1(.. ,).
Comparing (5.7) and (5.8), we see that
1(.. ,) =
_
o
-o
o
t
(t )
_
Q
t
(.)1
t
(,) 1
t
(.)Q
t
(,)
. ,
_
Jt. .. , (s. o).
Taking , . gives
1(.. .) =
_
o
-o
o
t
(t ){Q
t
t
(.)1
t
(.) 1
t
t
(.)Q
t
(.)]Jt
where the
t
denotes differentiation with respect to ..
Introducing
q
t
(s) = Q
t
(s: s) and
t
(s) = 1
t
(s: s).
we have
(5.9) 1(s. s) =
_
o
-o
o
t
(t ){Q
t
t
(s: s)
t
(s) 1
t
t
(s: s)q
t
(s)]Jt. s < .. , < o.
Formulas for Q
0
t
.x/ and P
0
t
.x/
As we just saw, we need expressions for Q
t
t
(.) and 1
t
t
(.). If D denotes the
differentiation operator
d
dx
, then
Q
t
t
(.: s) = D(1 1)
-1
t
t
= (1 1)
-1
Dt
t
D. (1 1)
-1
|t
t
= (1 1)
-1
t
t
[ D. (1 1)
-1
|t
t
= 1
t
(.) D. (1 1)
-1
|t
t
. (5.10)
We need the commutator
D. (1 1)
-1
| = (1 1)
-1
D. 1|(1 1)
-1
.
Integration by parts shows
D. 1|
.
=
_
d1
d.
d1
d,
_
1(.. s)(, s).
The -function comes from differentiating the characteristic function . Using the
specic form for and [ (
t
= [, [
t
= .),
_
d1
d.
d1
d,
_
=
_
o
-o
o
t
(t )t
t
(.)t
t
(,)Jt.
Thus
(5.11) D. (1 1)
-1
|
.
=
_
o
-o
o
t
(t )Q
t
(.)Q
t
(,)Jt 1(.. s)j(s. ,).
(Recall (1 1)
-1
.
= j(.. ,).) We now use this in (5.10):
Q
t
t
(.: s) = 1
t
(.)
_
o
-o
o
t
(
t )Q
t
(.)(Q
t
. t
t
)J
t 1(.. s)q
t
(s)
= 1
t
(.)
_
o
-o
o
t
(
t )Q
t
(.)u
t,t
(s) 1(.. s)q
t
(s)
where the inner product (Q
t
. t
t
) is denoted by u
t,t
(s) and u
t,t
(s) = u
t,t
(s).
Evaluating at . = s gives
Q
t
t
(s: s) =
t
(s)
_
o
-o
o
t
(
t )q
t
(s)u
t,t
(s) 1(s. s)q
t
(s).
We now apply the same procedure to compute 1
t
encountering the one new
feature: since [
t
(.) = .(.), we need to introduce an additional commutator
term:
1
t
t
(.: s) = D(1 1)
-1
t
t
[
= (1 1)
-1
Dt
t
[ D. (1 1)
-1
|t
t
[
= (M t )(1 1)
-1
t
t
(1 1)
-1
. M|t
t
D. (1 1)
-1
|t
t
[.
Writing it explicitly, we get (. t )Q
t
(.) 1(.. s)
t
(s) where
=
_
o
-o
o
t
(
t )(1
t
Q
t
Q
t
1
t
)t
t
J
t
_
o
-o
o(
t )(Q
t
Q
t
)t
t
[ J
t
=
_
o
-o
o
t
(
t ){1
t
(.)(Q
t
. t
t
) Q
t
(.)(1
t
. t
t
) Q
t
(.)(Q
t
. t
t
[)]J
t
=
_
o
-o
o
t
(
t ){1
t
(.)u
t,t
(s) Q
t
(.)
t,t
(s) Q
t
(.)
t,t
(s)]J
t .
with the notation
t,t
(s) = (1
t
. t
t
) = (t
t
[. Q
t
). Evaluating at . = s gives
1
t
(s: s) = (s t )q
t
(s)
_
o
-o
o
t
(
t ){
t
(s)u
t,t
(s) q
t
(s)
t,t
(s) q
t
(s)
t ,t
(s)]J
t
1(s. s)
t
(s).
Using this and the expression for Q
t
(s: s) in (5.9) gives
1(s. s) =
_
o
-o
o
t
(t )
_
2
t
sq
2
t

_
o
-o
o
t
(
t )
q
t
t

t
q
t
|u
t,t
q
t
q
t
t,t

t,t
|
_
_
J
t Jt.
First-Order Equations for q, p, u, and v
By the chain rule
(5.12)
Jq
t
Js
=
_
d
d.
d
ds
_
Q
t
(.: s)
x=x
.
We have already computed the partial of Q(.: s) with respect to .. The partial
with respect to s is, from (5.6),
d
ds
Q
t
(.: s) = (1 1)
-1
d1
ds
(1 1)
-1
t
t
= 1(.. s)q
t
(s).
Adding the two partial derivatives and evaluating at . = s gives
(5.13)
Jq
t
Js
=
t

_
o
-o
o
t
(
t )q
t
u
t,t
J
t .
A similar calculation gives
J
Js
= (s t )q
t

_
o
-o
o
t
(
t ){
t
u
t,t
q
t
t,t

t ,t
|]J
t .
We derive rst-order differential equations for u and by differentiating the inner
products: u
t,t
(s) =
_
o
x
t
t
(.)Q
t
(.: s) J.,
Ju
t,t
Js
= t
t
(s)q
t
(s)
_
o
x
t
t
(.)
dQ
t
(.: s)
ds
J.
=
_
t
t
(s)
_
o
x
1(s. .)t
t
(.) J.
_
q
t
(s)
= (1 1)
-1
t
t
(s)q
t
(s) = q
t
q
t
.
Similarly, J
t,t
,Js = q
t
t
.
Integrodifferential Equation for q
t
From the rst-order differential equations for q
t
, u
t
, and
t,t
, it follows that the
derivative in s of
_
o
-o
o
t
(t
t
)u
t,t
0 u
t
0
,t
Jt
t

t,t

t,t
| q
t
q
t
is zero. Examining the behavior near s = oto check that the constant of integra-
tion is zero then gives
_
o
-o
o
t
(t
t
)u
t,t
0 u
t
0
,t
Jt
t

t,t

t,t
| = q
t
q
t
.
a rst integral. Differentiate (5.13) with respect to s to get
q
tt
t
= (s t )q
t

_
o
-o
o
t
(
t )
_ _
o
-o
o
t
(t
t
)q
t
0 u
t ,t
0 Jt
t
u
t,t
q
t
t,t

t ,t
| q
t
q
2
t
_
J
t
and then use the rst integral to deduce that q satises (5.3).
Since the kernel of D. (1 1)
-1
| is (d,d. d,d,)1(.. ,), (5.11) says
_
d
d.
d
d,
_
1(.. ,) =
_
o
-o
o
t
(t )Q
t
(.)Q
t
(,)Jt 1(.. s)j(s. ,).
In computing dQ(.: s),ds we showed that
d
ds
(1 1)
-1
.
=
d
ds
1(.. ,) = 1(.. s)j(s. ,).
Adding these two expressions yields
_
d
d.
d
d,
d
ds
_
1(.. ,) =
_
o
-o
o
t
(t )Q
t
(.)Q
t
(,)Jt.
and then evaluating at . = , = s gives
J
Js
1(s. s) =
_
o
-o
o
t
(t )q
2
t
(s)Jt.
Hence q
tt
t
= {s t 21
t
]q
t
. Integrating and recalling (5.5) gives
J
Js
log det(1 1) =
_
o
x
_
o
-o
o
t
(t )q
2
t
(.)Jt J.
and hence
log det(1 1) =
_
o
x
__
o
,
_
o
-o
o
t
(t )q
2
t
(.)Jt J.
_
J,.
Rearranging gives (5.4). This completes the proof of Proposition 5.2.
6 Proofs of Corollaries to Theorem 1.1
6.1 Large-Time F
GUE
Asymptotics (Proof of Corollary 1.3)
We describe how to turn the idea described after Corollary 1.3 into a rigorous
proof. The rst step is to cut the j-contour off outside of a compact region around
the origin. Proposition 2.6 shows that for a xed T , the tail of the j-integrand is
exponentially decaying in j. A quick inspection of the proof shows that increas-
ing T only further speeds up the decay. This justies our ability to cut the contour
at minimal cost. Of course, the larger the compact region, the smaller the cost
(which goes to zero).
We may now assume that j is on a compact region. We will show the following
critical point: that det(1 1
csc
o
)
1
2
(
I
)
)
converges (uniformly in j) to the Fredholm
determinant with kernel
(6.1)
_
I
z
(
e
-
1
3
(
(
3
- )
03
)2
1/3
x(
(- )
0
)
J
( j
t
)( j)
.
This claim shows that we approach, uniformly, a limit that is independent of j.
Therefore, for large enough T we may make the integral arbitrarily close to the
integral of e
-
, j times the above determinant (which is independent of j) over
the cutoff j-contour. The j integral approaches 1 as the contour cutoff moves
towards innity, and the determinant is equal to J
GUE
(2
1{3
s), which proves the
corollary. A remark worth making is that the complex contours on which we are
dealing are not the same as those of [35]; however, owing to the decay of the kernel
and the integrand (in the kernel denition), changing the contours to those of [35]
has no effect on the determinant.
All that remains, then, is to prove the uniform convergence of the Fredholm
determinant claimed above.
The proof of the claim follows in a rather standard manner. We start by taking a
change of variables in the equation for 1
csc
o
in which we replace

by T
-1{3
and
likewise for j and j
t
. The resulting kernel is then given by
T
-1{3
_
I
(
e
-
1
3
(
(
3
- )
03
)2
1/3
(xo
0
)(
(- )
0
)
2
1{3
( j)
-2
1/3
T
1/3
(
(- )
0
)
sin(2
1{3
T
-1{3
(
j
t
))
J
j
.
Notice that the 1
2
-space as well as the contour of

-integration should have been
dilated by a factor of T
1{3
. However, it is possible (using Lemma 2.19) to show
that we may deform these contours back to their original positions without chang-
ing the value of the determinant. We have also used the fact that a = T
1{3
s
log
_
2T and hence T
-1{3
a = s a
t
where a
t
= T
-1{3
log
_
2T .
We may now factor this, just as in Proposition 2.6, as T, and likewise we may
factor our limiting kernel (6.1) as 1
t
=
t
T
t
where
(
. j) =
e
-[Im(
()[
j
.
T( j.
) = e
[Im(
()[
e
-
1
3
(
(
3
- )
3
)2
1/3
(xo
0
)(
(- ))
2
1{3
T
-1{3
( j)
-2
1/3
T
1/3
(
(- ))
sin(2
1{3
T
-1{3
(
j))
.
and similarly
t
(
. j) =
e
-[Im(
()[
j
.
T
t
( j.
) = e
[Im(
()[
e
-
1
3
(
(
3
- )
03
)2
1/3
x)(
(- )
0
)
1
j
.
Notice that =
t
. Now we use the estimate
[det(1 1
csc
o
) det(1 1
t
)[ _ [1
csc
o
1
t
[
1
exp{1 [1
csc
o
[
1
[1
t
[
1
].
Observe that [1
csc
o
1
t
[
1
_ [T T
t
[
1
_ [[
2
[T T
t
[
2
. Therefore it
sufces to show that [T T
t
[
2
goes to zero (the boundedness of the trace norms
in the exponential also follows from this). This is an explicit calculation and is
easily made by taking into account the decay of the exponential terms and the fact
that a
t
goes to zero. The uniformness of this estimate for compact sets of j follows
as well. This completes the proof of Corollary 1.3.
6.2 Small-Time Gaussian Asymptotics
PROPOSITION 6.1 As T
4
_ 0, 2
1{2
-1{4
-1
T
-1{4
F
(T. X) converges in dis-

tribution to a standard Gaussian.
PROOF. We have from (1.8),
F
(T. X) = log(1 T
1{4
G(T. X)
2
T
1{2
C(. T. X))
where
G(T. X) = T
-1{4
_
T
0
_
o
-o
(T S. X Y )(S. Y )
(T. X)
W (JY. JS)
and
C(. T. X) =
T
-1{2
o
n=2
_
Z
n
(T)
_
R
n
()
n-2
t
1
,...,t
n
(.
1
. . . . . .
n
)W (Jt
1
J.
1
) W (Jt
n
J.
n
).
It is elementary to show that for each T
0
< o there is a C = C(T
0
) < o such
that, for T < T
0
,
1C
2
(. T. X)| _ C.
G(T. X) is Gaussian and
1G
2
(T. X)| = T
-1{2
_
T
0
_
o
-o
2
(T S. X Y )
2
(S. Y )
2
(T. X)
JY JS =
1
2
_
.
Hence by Chebyshevs inequality,
J
T
(2
-1{2
1{4
T
1{4
s) = 1
_
T
1{4
G(T. X)
2
T
1{2
C(. T. X)
_ e
2
1/2
t
1/4
T
1/4
x
1
_
=
_
x
-o
e
-x
2
{2
_
2
J. O(T
1{4
).
Appendix: Analytic Properties of Fredholm Determinants

The following appendix addresses the question of analytic properties of Fred-
holm determinants and is based on communications from Percy Deift. For a gen-
eral discussion of Fredholm determinants, see [24, 29].
Suppose (z) is an analytic map from the region D C into the trace class
operators on a (separable) Hilbert space H. Then we have the following result:
THEOREM A.1 With : D B
1
(H) as above, the map z det(1 (z)) is
analytic on D and
J
Jz
det(1 (z)) = tr
t
tr(
t

t
)
tr(
t

t

t
) .
We rst prove the following useful lemma:
LEMMA A.2 Suppose
1
. . . . .
k
B
1
(H). Then
I(
1
. . . . .
k
) =
tS
k
t(1)

t(k)
maps
_
k
(H) to
_
k
(H) and I(
1
. . . . .
k
) B
1
(
_
k
(H)) with norm
(A.1) [I(
1
. . . . .
k
)[
1
_ [
1
[
1
[
2
[
1
[
k
[
1
.
PROOF. Since
}
are trace class, they are also compact. Compact operators
have singular value decompositions, which is to say that for each 1. . . . . k
there exists a decomposition of
}
as
}
=
i_1
a
}i
(
}i
. v)
t
}i
.
where a
}i
_ 0,
o
i=1
a
}i
< o, and {
}i
] as well as {
t
}i
] are orthonormal. For
u
1
. . . . . u
k
H, we write
u
i
. . u
k
=
1
_
k
cS
k
sgn(o)u
c(1)
u
c(k)

k
_
(H).
We will show in a moment that
(A.2) I(
1
. . . . .
k
)u
1
. u
2
. . u
k
=
i
1
,...,i
k
_1
k
I=1
a
I,i
l
__
k
_
I=1
I,i
l
_
.
_
k
_
I=1
u
I
__
_
t
I,i
l
.
Hence, as linear combinations of u
1
. . u
k
are dense in
_
k
(H), we have
I(
1
. . . . .
k
) =
i
1
,...,i
k
_1
a
1,i
1
a
k,i
k
(
1,i
1
. .
k,i
k
. v)
t
1,i
1
. .
t
k,i
k
.
which is the generalization of the singular value decomposition to the alternating
product of operators. As [(u. v)[
B
1
= [(u. )[ _ [u[[[ for any rank 1 operator
in a Hilbert space, it follows that
[(
1,i
1
. .
k,i
k
. v)
t
1,i
1
. .
t
k,i
k
[
B
1
(
_
k
(H))
_
[
1,i
1
. .
k,i
k
[ [
t
1,i
1
. .
t
k,i
k
[ _ 1.
and hence it follows that
[I(
1
. . . . .
k
)[
B
1
(
_
k
(H))
_
i
1
,...,i
k
_1
a
1,i
1
a
k,i
k
= [
1
[
B
1
[
k
[
B
1
.
This proves (A.1).
It remains to proves (A.2). Note that the left-hand side can be written as
1
_
k
cS
k
tS
k
sgn(o)(
t(1)

t(k)
)u
c(1)
u
c(k)
=
i
1
,...,i
k
_1
1
_
k
c,tS
k
sgn(o)
k
I=1
a
t(I),i
l
k
I=1
((
t(I),i
l
. v)
t
]i(I),i
l
)
k
I=1
u
c(I)
We recognize that
cS
k
sgn(o)
k
I=1
(
t(I),i
l
. u
c(I)
) = det(
t(I),i
l
. u
n
)|
k
I,n=1
so that, after a permutation, the whole thing becomes
i
1
,...,i
k
_1
1
_
k
tS
k
sgn()
k
I=1
a
I,i
l
det(
I,i
l
. u
n
)|
k
I,n=1
k
I=1
t
t(I),i
(l)
=
i
1
,...,i
k
_1
k
I=1
a
I,i
l
__
k
_
I=1
I,i
l
_
.
_
k
_
I=1
u
I
__
1
_
k
tS
k
sgn()
k
I=1
t
t(I),i
(l)
=
i
1
,...,i
k
_1
k
I=1
a
I,i
l
__
k
_
I=1
I,i
l
_
.
_
k
_
I=1
u
I
__
_
t
I,i
l
.
which gives (A.2).
Now let . T B
1
(H). For l. m _ 0, k = l m, dene
I
(I,n)
(. T) =
1
l m
I(. . . . . . T. . . . T).
where there are l s and m Ts. Clearly I
(I,n)
(. T) =
c
1
c
k
where
the sum is over all
_
nI
n
_
ways of designating l of the c
i
s as and the other m as
T. As an example, I
(1,2)
(. T) = T T T T T T .
COROLLARY A.3 (Corollary to Lemma A.2)
[I
(I,n)
(. T)[
B
1
(
_
k
(H))
_
[[
I
1
l
[T[
n
1
m
.
We can now proceed with the proof of Theorem A.1:
PROOF OF THEOREM A.1. Fix z D and let (z h) = (z) = .
For k _ 1,
(z h) (z h)
=
= I
(1,k-1)
(. )
I
(2,k-2)
(. )
(I,k-I)
(. ) .
Thus
h
-1
{(z h) (z h) (z) (z)] =
(1,k-1)
(,h. ) ^(h).
where, by the corollary,
[^(h)[
B
1
(
_
k
(H))
_
1
[h[
[[
2
1
2
[[
k-2
1
(k 2)

1
[h[
[[
k
1
k
.
Observe that [[
1
= [(z h) (z)[
1
= O(h). Write
(1,k-1)
_
h
.
_
= I
(1,k-1)
(
t
. ) I
(1,k-1)
_
(z h) (z)
h

t
(z).
_
.
and then observe that by the corollary
(A.3)
I
(1,k-1)
_
(z h) (z)
h

t
(z). (z)
_
B
1
(
_
k
(H))
_
(z h) (z)
h

t
(z)
B
1
1
(k 1)
[(z)[
k-1
B
1
= O(h).
Combining these observations shows that
h
-1
{(z h) (z h) (z) (z)] =
I
(1,k-1)
(
t
. ) O(h).
and hence the function z (z) (z) = I
(k)
((z)) is an analytic map
from D to B
1
(
_
k
(H)) for all k _ 1 and
J
Jz
(z) (z) = I
(1,k-1)
(
t
. )
=
t

t
.
It then follows that z tr I
(k)
((z)) is analytic for k _ 1 from D to C. Hence
for any n _ 1, 1
n
k=1
tr I
(k)
((z)) is analytic in D and
1
n
k=1
tr I
(k)
((z))
_ 1
n
k=1
[I
(k)
((z))[
B
1
(
_
k
(H))
_ 1
n
k=1
[(z)[
k
B
1
(
_
k
(H))
k
.
which is bounded by e
|(;)|
, and so for z in a compact subset of D, the functions
1
n
k=1
tr I
(k)
((z)) are uniformly bounded in n. It follows that
z det(1 (z)) = lim
n-o
n
k=0
tr I
(k)
((z))
is analytic in D and
J
Jz
det(1 (z))
= lim
n-o
n
k=0
J
Jz
tr I
(k)
((z)) =
o
k=1
tr
_
I
(1,k-1)
(
t
(z). (z))
_
=
o
k=1
tr
_
t
(z) (z) (z) (z)
t
(z)
_
.
Acknowledgment. We would like to thank Percy Deift for his ongoing support
and assistance with this project, as well as travel funding he provided to IC. We
thank Craig Tracy and Harold Widom for discussing this matter during a visit in
the summer of 2009, and further thank Tracy for ongoing interest and support.
JQ and GA wish to thank Kostya Khanin and Balint Virg for many interesting
discussions and encouragement, and Tom Bloom and Ian Graham for their helpful
suggestions on function theory. IC wishes to thank Grard Ben Arous and Antonio
Aufnger for helpful discussions and comments as well as Sunder Setheraman for
an early discussion about the WASEP crossover. IC also wishes to thank all of the
participants of the ASEP seminar that occurred in 200809 year. This collaboration
was initially encouraged by Ron Peled, and we thank him graciously for playing
matchmaker. We also thank Alex Bloemendal and David and Nora Ihilchik for
hosting IC during his visits to Toronto. GA and JQ are supported by the Natural
Science and Engineering Research Council of Canada. IC is funded by a National
Science Foundation graduate research fellowship and has also received support
from Partnerships for International Research and Education (PIRE) Grant OISE-
07-30136.
Bibliography
[1] Balzs, M.; Quastel, J.; Sepplinen, T. Scaling exponent for the Hopf-Cole solution of
KPZ/stochastic Burgers. Preprint, 2009. arXiv:0909.4816v1
[2] Bertini, L.; Cancrini, N. The stochastic heat equation: Feynman-Kac formula and intermittence.
J. Statist. Phys. 78 (1995), no. 5-6, 13771401.
[3] Bertini, L.; Giacomin, G. Stochastic Burgers and KPZ equations from particle systems. Comm.
Math. Phys. 183 (1997), no. 3, 571607.
[4] Billingsley, P. Convergence of probability measures. 2nd ed. Wiley Series in Probability and
Statistics: Probability and Statistics. Wiley-Interscience, New York, 1999.
[5] Borodin, A.; Deift, P. Fredholm determinants, Jimbo-Miwa-Ueno t-functions, and representa-
tion theory. Comm. Pure Appl. Math. 55 (2002), no. 9, 11601230.
[6] Borodin, A.; Ferrari, P. L. Large time asymptotics of growth models on space-like paths.
I. PushASEP. Electron. J. Probab. 13 (2008), no. 50, 13801418.
[7] Borodin, A.; Okounkov, A.; Olshanski, G. Asymptotics of Plancherel measures for symmetric
groups. J. Amer. Math. Soc. 13 (2000), no. 3, 481515 (electronic).
[8] Calabrese, P.; Le Doussal, P.; Rosso, A. Free-energy distribution of the directed polymer at high
temperature. Euro. Phys. Lett. 90 (2010), no. 2, 20002.
[9] Chan, T. Scaling limits of Wick ordered KPZ equation. Comm. Math. Phys. 209 (2000), no. 3,
671690.
[10] Corwin, I.; Ferrari, P. L.; Pch, S. Limit processes for TASEP with shocks and rarefaction fans.
J. Stat. Phys. 140 (2010), 232267.
[11] Deift, P. A. Orthogonal polynomials and random matrices: a Riemann-Hilbert approach.
Courant Lecture Notes in Mathematics, 3. New York University, Courant Institute of Math-
ematical Sciences, New York; American Mathematical Society, Providence, R.I., 1999.
[12] Dotsenko, V. Bethe ansatz derivation of the Tracy-Widom distribution for one-dimensional di-
rected polymers. Euro. Phys. Lett. 90 (2010), no. 2, 20003.
[13] Forster, D.; Nelson, D. R.; Stephen, M. J. Large-distance and long-time properties of a randomly
stirred uid. Phys. Rev. A (3) 16 (1977), no. 2, 732749.
[14] Grtner, J. Convergence towards Burgers equation and propagation of chaos for weakly asym-
metric exclusion processes. Stochastic Process. Appl. 27 (1988), no. 2, 233260.
[15] Johansson, K. Shape uctuations and random matrices. Comm. Math. Phys. 209 (2000), no. 2,
437476.
[16] Johansson, K. Discrete polynuclear growth and determinantal processes. Comm. Math. Phys.
242 (2003), no. 1-2, 277329.
[17] Kardar, M. Replica Bethe Ansatz studies of two-dimensional interfaces with quenched random
impurities. Nuclear Phys. B 290 (1987), no. 4, 582602.
[18] Kardar, M.; Parisi, G.; Zhang, Y.-C. Dynamic scaling of growing interfaces. Phys. Rev. Lett. 56
(1986), 889892.
[19] Kolokolov, I. V.; Korshunov, S. E. Universal and nonuniversal tails of distribution functions in
the directed polymer and Kardar-Parisi-Zhang problems. Phys. Rev. B 78 (2008), no. 2, 024206.
[20] Konno, N.; Shiga, T. Stochastic partial differential equations for some measure-valued diffu-
sions. Probab. Theory Related Fields 79 (1988), no. 2, 201225.
[21] Lieb, E. H.; Liniger, W. Exact analysis of an interacting Bose gas. I. The general solution and
the ground state. Phys. Rev. 130 (1963), no. 4, 16051616.
[22] Liggett, T. M. Interacting particle systems. Reprint of the 1985 original. Classics in Mathemat-
ics. Springer, Berlin, 2005.
[23] Prhofer, M.; Spohn, H. Scale invariance of the PNG droplet and the Airy process. J. Statist.
Phys. 108 (2002), no. 5-6, 10711106.
[24] Reed, M.; Simon, B. Methods of modern mathematical physics. IV. Analysis of operators. Aca-
demic [Harcourt Brace Jovanovich], New YorkLondon, 1978.
[25] Sasamoto, T.; Spohn, H. The crossover regime for the weakly asymmetric simple exclusion
process. J. Stat. Phys. 140 (2010), no. 2, 209231.
[26] Sasamoto, T.; Spohn, H. Exact height distributions for the KPZ equation with narrow wedge
initial condition. Nuclear Phys. B 834 (2010), no. 3, 523542.
[27] Sasamoto, T.; Spohn, H. One-dimensional Kardar-Parisi-Zhang equation: an exact solution and
its universality. Phys. Rev. Lett. 104 (2010), 230602.
[28] Sepplinen, T. Scaling for a one-dimensional directed polymer with boundary conditions.
Preprint, 2009. arXiv:0911.2446v2
[29] Simon, B. Trace ideals and their applications. 2nd ed. Mathematical Surveys and Mono-
graphs, 120. American Mathematical Society, Providence, R.I., 2005.
[30] Srivastava, H. M.; Choi, J. Series associated with the zeta and related functions. Kluwer, Dor-
drecht, 2001.
[31] Tracy, C. A.; Widom, H. Level-spacing distributions and the Airy kernel. Comm. Math. Phys.
159 (1994), no. 1, 151174.
[32] Tracy, C. A.; Widom, H. Airy kernel and Painlev II. Isomonodromic deformations and ap-
plications in physics (Montral, QC, 2000), 8596. CRM Proceedings and Lecture Notes, 31.
Amererican Mathematical Society, Providence, R.I., 2002.
[33] Tracy, C. A.; Widom, H. A Fredholm determinant representation in ASEP. J. Stat. Phys. 132
(2008), no. 2, 291300.
[34] Tracy, C. A.; Widom, H. Integral formulas for the asymmetric simple exclusion process. Comm.
Math. Phys. 279 (2008), no. 3, 815844.
[35] Tracy, C. A.; Widom, H. Asymptotics in ASEP with step initial condition. Comm. Math. Phys.
290 (2009), no. 1, 129154.
[36] Tracy, C. A.; Widom, H. Formulas for joint probabilities for the asymmetric simple exclusion
process. J. Math. Phys. 51 (2010), 063302.
[37] van Beijeren, H.; Kutner, R.; Spohn, H. Excess noise for driven diffusive systems. Phys. Rev.
Lett. 54 (1985), no. 18. 20262029.
[38] Varlamov, V. Fractional derivatives of products of Airy functions. J. Math. Anal. Appl. 337
(2008), no. 1, 667685.
[39] Walsh, J. B. An introduction to stochastic partial differential equations. cole dt de prob-
abilits de Saint Flour, XIV1984, 265439. Lecture Notes in Mathematics, 1180. Springer,
Berlin, 1986.
GIDEON AMIR
Department of Mathematics
40 St. George St.
Toronto, Ontario
CANADA M5S 2E4
E-mail: gidi@math.toronto.edu
JEREMY QUASTEL
Department of Mathematics
40 St. George St.
Toronto, Ontario
CANADA M5S 2E4
E-mail: quastel@
math.toronto.edu
IVAN CORWIN
Courant Institute
251 Mercer St., Room 604
New York, NY 10012
E-mail: corwin@cims.nyu.edu
Received April 2010.

Probability Distribution of The Free Energy of The Continuum Directed Random Polymer in 1 + 1 Dimensions - Amir - 2010 - Communications On Pure and Applied Mathematics - Wiley Online Library PDF

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Probability Distribution of The Free Energy of The Continuum Directed Random Polymer in 1 + 1 Dimensions - Amir - 2010 - Communications On Pure and Applied Mathematics - Wiley Online Library PDF

Uploaded by

Copyright:

Available Formats

Probability Distribution of the Free Energy

of the Continuum Directed Random Polymer

W (t. b(t ))Jt

W (t. b(t ))Jt

W (t. b(t ))Jt

(T. X) = (T. X) exp{F

(T. X)] is the solution

(T. X) and (T. X). From this we see that

(R) under the 1

should have been 2

Cauchys Theorem Change of vars

C we have that for all > 0 there exists an c

C. For all > 0 there exist c

is an analytic map from this region of j between

. This will prove both the convergence of traces as well as

for < 1 (we show it for =

comes from the fact that

502 G. AMIR, I. CORWIN, AND J. QUASTEL

j). Notice that we have put the factor e

)[ behaves like the following product of exponentials:

for some constant c.

for < 1. Using the bound that

is now dened for [z[ < t

functional equation, since the g

, so that when we plug this in along with

case. Let us, for the moment, assume that [ j[ < 1. We

is bounded in c; thus we end up with the

. This, however, can be shown directly as follows.

(z)[ has an upper bound uniform in j. Likewise, we can do a similar

(X), and J always refers

is supported on C(. o): C(R))

when restricted to C(. o). C(R)). From the

means that the integral along the positive

. This may be proved similarly to the proof of

and observing that 1

(T. X) converges in dis-

Appendix: Analytic Properties of Fredholm Determinants

You might also like