Sunteți pe pagina 1din 18

KP GOVERNS RANDOM GROWTH OFF A ONE DIMENSIONAL SUBSTRATE

JEREMY QUASTEL AND DANIEL REMENIK

A BSTRACT. The logarithmic derivative of the marginal distributions of randomly fluctuating interfaces
in one dimension on a large scale evolve according to the Kadomtsev–Petviashvili (KP) equation. This is
derived algebraically from a Fredholm determinant obtained in [MQR17] for the KPZ fixed point as the
limit of the transition probabilities of TASEP, a special solvable model in the KPZ universality class. In
addition, it is noted that known exact solutions of the KPZ equation also solve KP.
arXiv:1908.10353v1 [math.PR] 27 Aug 2019

C ONTENTS

1. Matrix KP equation for multidimensional distributions 1


1.1. Initial data 3
1.2. Tracy-Widom distributions 4
1.3. PDEs for other initial data 5
1.4. Airy process 5
2. KP-II in special solutions of the KPZ equation 6
3. The determinantal formula and derivation of KP 8
3.1. KPZ fixed point formula 8
3.2. The logarithmic derivative 9
3.3. Formulas for the partial derivatives 10
Appendix A. Multipoint initial data 12
A.1. t → 0 limit of the Brownian scattering operator 12
A.2. Matrix KP initial data 14
Appendix B. Alternative derivation of KP-II for narrow wedge multipoint distributions 15
Appendix C. An identity for the flat KPZ Fredholm Pfaffian kernel 16
References 17

1. M ATRIX KP EQUATION FOR MULTIDIMENSIONAL DISTRIBUTIONS

The one dimensional KPZ universality class consists of random growth models, last passage percola-
tion and directed polymers, and random stirred fluids. All models in the class have an analogue of the
height function h(t, x) (free energy, integrated velocity) whose long time large scale evolution is the
principal object of study. The name of the class comes from the Kardar-Parisi-Zhang equation,
∂t h = λ(∂x h)2 + ν∂x2 h + σξ (1.1)
with ξ a space-time white noise, a canonical continuum equation for random growth introduced in
[KPZ86]. However, the key interest is on the universal features which are only found in large space-time
scales, under the 1:2:3 scaling corresponding to ε → 0 in
ε1/2 h(ε−3/2 t, ε−1 x) − Cε t. (1.2)

Date: August 28, 2019.


1
KP(Z) 2

The KPZ equation is not invariant under this scaling, which sends (λ, ν, σ) to (λ, ε1/2 ν, ε1/4 σ). A key
problem is to find the true, scaling invariant equation for random interface growth.
Since the early 2000’s [Joh03; Sas05; BFPS07] it was known, for a number of models in the class,
and special scaling invariant initial data narrow wedge and flat, that the distributional limits of (1.2) were
the Tracy-Widom distributions of random matrix theory. In an earlier article [MQR17] it was shown
that, at least for one model in the class, TASEP, (1.2) converges to a 1:2:3 invariant Markov process
h(t, x) with completely integrable transition probabilities given in terms of Fredholm determinants of
kernels K depending on the initial data (in [NQR19; MQR+] this is extended to other models related to
TASEP). It is widely believed that this KPZ fixed point governs the limiting fluctuation for all models in
the class.
The KPZ fixed point does not satisfy a stochastic differential equation. In place of that, it inherits
a variational formulation from TASEP; a Hopf-Lax type formula involving a non-trivial input noise
called the Airy sheet A(x, y),
dist
h(t, x) = sup t1/3 A(t−2/3 x, t−2/3 y) − 1t (x − y)2 + h0 (y) .

(1.3)
y∈R

The Airy sheet A(x, y) can be thought of as the height function at x at time 1, starting from a narrow
wedge at y at time 0, and therefore involves coupling different initial conditions. As far as we know at
the present time, the coupled initial condition problem is not integrable, and therefore the distribution of
the Airy sheet is unknown. This led to a problem in that it was unclear that (1.3) even involved a unique
object on the right hand side. An important advance is in [DOV19], who show that the Airy sheet is a
functional of the Airy line ensemble. This puts the variational formula (1.3) on a solid footing, as it
obviates the need for uniqueness of the Airy sheet. However, the functional is completely non-explicit.
In this sense, (1.3) is not satisfying as a universal scaling invariant equation.
Instead of a universal stochastic equation, one can study the n-space point distribution functions,
F (t, x1 , . . . , xn , r1 , . . . , rn ) = Ph0 (h(t, x1 ) ≤ r1 , . . . , h(t, xn ) ≤ rn ) (1.4)
where h(t, x) is the KPZ fixed point starting from h0 . In the cases of narrow wedge and flat initial
data, it was known [Joh00; Sas05; BFPS07] that the one-dimensional distributions F (1, x, r) were,
respectively, the Tracy-Widom GUE and GOE random matrix distributions (but, except in the particular
case of narrow wedge initial data, the connection between random growth and random matrices has
remained tangential and murky). The multidimensional distributions in these cases are given by
Fredholm determinants, and define the Airy2 and Airy1 processes. The one-dimensional distributions
can be written in terms of the Hastings-McLeod solution of Painlevé II; a longstanding open question
was whether the distributions satisfy an equation in the more general setting.
In (3.8) we will define a tau function Q = ((I − K)−1 K)(0, 0) which is an n × n matrix valued
function of t, x1 , . . . , xn , , r1 , . . . , rn , and the initial height profile h0 ; at this point its exact definition
is not important. Define
Xn Xn
Dr = ∂ri , Dx = ∂xi . (1.5)
i=1 i=1
Our main result is:
Theorem 1.1. Q and its derivative q = Dr Q solve the matrix Kadomtsev–Petviashvili (KP) equation
3
∂t q + qDr q + 1
12 Dr q + 41 Dx2 Q = 0, (1.6)
and the logarithmic derivative of the n point distribution (1.4) is given by
Dr log F = tr Q.
In particular, the one point marginals φ = ∂r2 log F satisfy the KP-II equation
∂t φ + φ∂r φ + 1 3
12 ∂r φ + 41 ∂r−1 ∂x2 φ = 0. (1.7)
KP(Z) 3

Remark 1.2.
1. The result is completely unexpected1. We do not have physical intuition why it is true; it follows
by, essentially, algebra from the form of the kernel in the Fredholm determinant for (1.4), and
we believe it is the first example of a physical law having been obtained in such a fashion.
In retrospect, there are not so many natural partial differential equations with the necessary
invariance under
φ(t, x, r) 7→ α−2 φ(α−3 t, α−2 x, α−1 r), h0 (x) 7→ α−1 h0 (α2 x).
The question is why they should satisfy a closed equation at all.
2. The KP equation (1.7) was originally derived from studies of long waves in shallow water
[AS79]. It has come to be accepted as the natural two dimensional extension of the Kortweg
deVries equation (KdV); when φ is independent of x, corresponding in our case to flat initial
data, it reduces to KdV. KP is completely integrable and plays an important role in the Sato
theory as the first equation in the KP hierarchy [MJD00]. None of the previous derivations
seem to be related to the problem at hand, and it could well be that our evolution is through a
class of functions where the equation is formally the same, because of the similarity of weakly
nonlinear asymptotics, but the physics is completely different.
3. The one dimensional distribution functions themselves therefore satisfy the equivalent Hirota
bilinear equation,
2 4
F ∂tr F − ∂t F ∂r F + 1
12 F ∂r F − 13 ∂r F ∂r3 F + 14 (∂r2 F )2 + 14 F ∂x2 F − 14 (∂x F )2 = 0,
which again has the necessary 1:2:3 invariance, now under
F (t, x, r) 7→ F (α−3 t, α−2 x, α−1 r), h0 (x) 7→ α−1 h0 (α2 x).
4. Unlike other limit points for fluctuation universality classes in probability, the Tracy-Widom
distributions themselves lack any invariance. Thm. 1.1 recovers the invariance of the scaling
limit (see Sec. 1.2).

1.1. Initial data. The natural class of initial data for our problem (the “one dimensional substrate”)
are upper semicontinuous functions h : R → [−∞, ∞) with a growth condition2 h(x) ≤ A|x| + B. A
function is upper semicontinuous if and only if its hypograph hypo(h) = {(x, y) : y ≤ h(x)} is closed
in [−∞, ∞) × R. We endow [−∞, ∞) with the distance d[−∞,∞) (y1 , y2 ) = |ey1 − ey2 |, and use the
topology of local Hausdorff convergence, which means Hausdorff convergence of the restrictions to
−L ≤ x ≤ L of hypo(hn ) to hypo(h) for each L > 0. This space is called UC.
Example 1.3. (Finite collection of narrow wedges) Let a1 < a2 < · · · < ak , b1 , b2 , . . . , bk ∈ R
~
and take h = d~ab , defined by
~ ~
d~ab (x) = bi if x = ai for some i, d~ab (x) = −∞ otherwise.

The initial data for (1.7) is the “escarpment” φ(0, x, r) = 0 for r ≥ h0 (x) and φ(0, x, r) = −∞
for r < h0 (x). These are unusual and do not fit into any well-posedness schemes known for the KP
equation3. It appears the solutions to the equations with such initial data do not develop solitons and
that they are well posed, but we leave the proofs for future work. Since F is given by a Fredholm
determinant, these initial conditions represent an entirely new class of integrable initial data for KP.
The initial data for the matrix KP equation (1.6) is

≥h0 ,≤−d−~r
−Pxi ,xj ~x (ri , rj ) if i < j,


Qi,j = ∞ if i = j and ri < h0 (xi ),


0 otherwise.
1However, see the preprint [Pro19] from two days before this article was posted where it is shown that particular finite
volume solutions [BL19] can be written as superpositions of solitons.
2With some work this can be relaxed to h(x) ≤ A|x|2 + B up to a finite time t = t(A).
3[KPV97] consider as initial data an odd polynomial with positive leading coefficient which is similar in form.
KP(Z) 4

where
−~r
≥h0 ,≤−d~x
Pxi ,xj (ri , rj )drj

:= PB(xi )=ri B(y) ≥ h0 (y) ∀ y ∈ [xi , xj ], B(xn ) ≤ rn for each xn ∈ (xi , xj ), B(xj ) ∈ drj .
The probability is with respect to a Brownian motion B(·) with diffusivity 2 starting at ri at time xi .
This is derived in Appdx A for finite collections of narrow wedges.

1.2. Tracy-Widom distributions. A key observation is that the GUE and GOE Tracy-Widom distri-
butions are now seen to simply arise as special similarity solutions of the KP equation (1.7):
Example 1.4. (Tracy-Widom GUE distribution) If h0 = d0 , the narrow wedge initial condition
defined as d0 (0) = 0 and d0 (x) = −∞, x 6= 0, we look for a self-similar solution of (1.7) the form
φ(t, x, r) = t−2/3 ψ(t−1/3 r + t−4/3 x2 ).
This turns (1.7) into
ψ 000 + 12ψψ 0 − 4rψ 0 − 2ψ = 0.
The transformation ψ = −q 2 takes (1.7) into Painlevé II
q 00 = rq + 2q 3 . (1.8)
r x2
As r → −∞ the solution is approximately φ(t, x, r) ∼ −( 2t + ), picking out the Hastings-McLeod
2t2
solution q(r) ∼ − Ai(r) as r → ∞. The Tracy-Widom GUE distribution is usually written in the
equivalent form
 Z ∞ 
2
F2 (s) = exp − du (u − s)q (u) .
s

Example 1.5. (Tracy-Widom GOE distribution) If h0 (x) ≡ 0 (flat initial condition) there is no x
dependence and (1.7) reduces to KdV. We look for a self-similar solution of (1.7) the form
φ(t, r) = (t/4)−2/3 ψ((t/4)−1/3 r),
obtaining the ordinary differential equation
ψ 000 + 12ψ 0 ψ − rψ 0 − 2ψ = 0.
Miura’s transform
ψ = 21 (q 0 − q 2 )
brings this to Painlevé II (1.8). So we recover
F (t, x, r) = F1 (41/3 t−1/3 r)
where F1 is the GOE Tracy-Widom distribution, usually written in the equivalent form
1 ∞
 Z 
F1 (r) = exp − q(u)du F2 (r)1/2 .
2 r

These two examples also have the following interpretation. Let λmax,GUEN and λmax,GOE
N be the
largest eigenvalues of N × N matrices √ chosen from the √ Gaussian Unitary and Gaussian Orthogonal
Ensembles, multiplied, respectively, by N and 4−1/3 N , so that λmax,GUEN ∼ 2N + N 1/3 ζGUE and
max,GOE
λN ∼ 41/6 N + N 1/3 4−1/3 ζGOE (the 4−1/3 is just to coordinate conventions with random growth).
Let
FGOE (t, r) = limN →∞ P(N −1/3 (λmax,GOE
Nt − 41/6 N t) ≤ r),
FGUE (t, x, r) = limN →∞ P(N −1/3 (λmax,GUE
Nt − 2N t) ≤ r + x2 /t).
As we have seen, ∂r2 log FGOE and ∂r2 log FGUE satisfy the KP equation (1.7). In the former case, there
is no dependence on x and KP reduces to KdV.
KP(Z) 5

1.3. PDEs for other initial data. Another question is whether there are analogues of Painlevé II for
other self-similar solutions. It is natural to observe φ in the frame of the inviscid solution 14 (∂x h̄)2 −
∂t h̄ = 0 of Burgers’ equation,
φ(t, x, r) := φ̄(t, x, r − h̄(t, x));
one obtains
∂t φ̄ + φ̄∂r φ̄ + 1 3
12 ∂r φ̄ + 14 ∂r−1 ∂x2 φ̄ + V φ̄ = 0
with V = − 14 ∂x2 h̄0 and with initial data ψ(0, x, r) = 0 for r ≥ 0 and −∞ for r < 0. In order to get a
solution for the rescaled spatial process, let
φ(t, x, r) := t−2/3 ψ(t, t−2/3 x, t−1/3 (r − h̄(t, x)))
Then
t1/3
0 = − 31 r∂r ψ + 1 3
12 ∂r ψ − 23 ψ + ψ∂r ψ + 41 ∂r−1 ∂x2 ψ − 4t ∂x2 h̄ψ − ( 32 t−2/3 x + 2 ∂x h̄)∂x ψ + t∂t ψ
Example 1.6. (Half-flat initial data) Consider h0 (x) = 0, x ≤ 0 and h0 (x) = −∞, x > 0. Now
h̄(t, x) = −x2 /t1x≥0 . There is dependence on x, though not on t. This gives rise to a partial differential
equation for ψ(x, r);
0 = − 13 r∂r ψ + 1 3
12 ∂r ψ − ( 16 1x≥0 + 23 1x<0 )ψ + ψ∂r ψ + 41 ∂r−1 ∂x2 ψ − ( 32 1x<0 − 13 1x≥0 )x∂x ψ.
Remark 1.7. (Lower tail heuristics) Typically the equation is controlled on large scales by the
equation with the third derivative dropped, and the Burgers’ equation makes sense for such wedge type
initial data. Let
ψ̄ = ψ − η
c
where η = t r1r<0 . Then ψ̄(0, x, r) = 0 and
∂t ψ̄ + ψ̄∂r ψ̄ + 1 3
12 ∂r ψ̄ + 14 ∂r−1 ∂x2 ψ̄ + ∂r (η ψ̄) + µψ̄ + γ∂r ψ̄ + ν∂x ψ̄ + V = 0,
c 0
where V = 1t (µ − 1 + c)η + δµ ct 1r<0 − 12t δ0 (r). One hopes to set things up so that ψ̄ has good decay
at ±∞. Consider our two basic examples. In the flat case δ = 0, µ = 0 and we take c = 1 to make the
1 3
last term drop out, which will lead to the conclusion that φ(t, r) ∼ r/t, or F (t, r) ∼ exp{− 6t r } as
r → −∞, recovering the Tracy-Widom GOE lower tail. In the narrow wedge case µ = 1/2 and we
1 2
take c = 1/2 leading to F (t, x, r) ∼ exp{− 12t (r + xt )3 }, recovering the Tracy-Widom GUE lower
tail.
The conclusion is that the lower tail of the distributions can be seen directly from the “Burgers” part of
the KP equation, which dominates in that region.

1.4. Airy process. The Airy process A(x) is defined as


A(x) := h(1, x; d0 ) + x2
where h(t, x; d0 ) is the KPZ fixed point starting from a narrow wedge d0 at the origin. The Airy process
is stationary and the one point distribution is the Tracy-Widom GUE distribution.
K. Johansson famously asked whether there is an equation for the multipoint distribution. Equations
were given by [AM05; TW03]; starting with narrow wedge initial data the matrix KP equation (1.6)
gives us an equation which is similar to that in [TW03] (and possibly equivalent). The relation with
the equation derived in [AM05] is completely unclear. But by skew time reversal invariance we get
something better.
Let d~~rx denote a multiple narrow wedge (see Ex. 1.3). For y ∈ R write ~x + y = (x1 + y, . . . , xm + y)
and write similarly ~r + a. By skew time reversibility and translation and affine invariance of the KPZ
fixed point [MQR17, Thm. 4.5],
F (t, ~x + y, ~r + a) = P h(t, xi + y; d0 ) ≤ ri + a for all i = P h(t, ·; d0 ) ≤ −d~−~
r−a
 
x+y

= P h(t, ·; d~−~
r−a −~
r
 
x+y ) ≤ −d0 = P h(t, 0; d~ x+y ) ≤ a

= P h(t, −y; d~−~ r



x )≤a .
KP(Z) 6

Now the right hand side is just the one point distribution at −y with a given, fixed initial condition. So
if we let G(t, y, a) = F (t, ~x + y, ~r + a) we see that ∂r2 G satisfies (1.7) in (t, y, a).
But ∂y G = ∂y F = Dx F and similarly ∂a G = ∂a F = Dr F . So setting now a = y = 0 we deduce
that φ = Dr2 log F satisfies
∂t φ + φDr φ + 1 3
12 Dr φ + 14 Dr−1 Dx2 φ = 0.
1
From Thm. 1.1 we know that Dr2 log F = tr q and ∂t tr q + tr(qDr q) + 12 Dr3 tr q + 14 Dx2 tr Q = 0. But
the above argument implies then that tr q itself solves KP-II. As a consequence, we deduce in the narrow
wedge case that
tr(qDr q) = tr(q) tr(Dr q).
0 hypo(h )
This can also be proved directly using the fact that in this case (∂u + ∂v )Kt (u, v) is a rank one
kernel, which implies that Q is a rank one matrix.
An alternative derivation using the path integral formula for the KPZ fixed point can be found in
Appendix B.

2. KP-II IN SPECIAL SOLUTIONS OF THE KPZ EQUATION

The proof of Thm. 1.1 in Sec. 3 shows also that various explicit solutions for one dimensional
distributions of the Kardar-Parisi-Zhang equation (1.1) also satisfy the KP-II equation (1.7). At this
point we do not know if this is a general fact, or if KP-II only holds in these special cases because of
some symmetry. All we have is examples.
Example 2.1. (Narrow wedge solution of KPZ) Let hnw be the narrow wedge solution of (1.1) with
λ = ν = 14 and σ = 1. In other words, hnw = log Z where Z is the fundamental solution of the
stochastic heat equation with multiplicative noise
∂t Z = 41 ∂x2 Z + ξZ, Z(0, x) = δ0 (x).

The KPZ generating function is


  
t
h(t,x)+ 12 −r
Gnw (t, x, r) = E exp −e . (2.1)

The distribution of hnw (t, x) was computed in 2010 in [ACQ11; SS10; Dot10; CDR10], with the result
that
Gnw (t, x, r) = det(I − K)L2 [0,∞)
with
Z ∞
1
K(u, v) = dy t−2/3 y
Ai(t−1/3 (u + r − y) + t−4/3 x2 ) Ai(t−1/3 (v + r − y) + t−4/3 x2 ).
−∞ 1 + e
If we conjugate the operator by multiplying the kernel by e(v−u)x/t we get
2 2
K(u, v) = Ut e−x∂ Mex∂ U−1
t (u + r, v + r)
with M the multiplication operator Mf (u) = (1 + eu )−1 f (u) and Ut the Airy unitary operator defined
2
in (3.2) (see Sec. 3.1 also for the meaning of ex∂ Ut for general x). In particular, K satisfies the
(one-point version of the) same differential relations (3.7)/(3.9) as the KPZ fixed point kernel, whence
it follows
√ that φnw := ∂r2 log Gnw also satisfies KP-II. The initial condition is limt&0 Gnw (t, x, r +
log πt) = e−r (1x6=0 + e−1 1x=0 ).
Example 2.2. (Flat solution of KPZ) Consider next the solution hflat of the KPZ equation (1.1)
with λ = ν = 14 , σ = 1 and flat initial data hflat (0, x) ≡ 0, and define Gflat (t, r) as in (2.1) (there is
no dependence on x in this case because hflat (t, x) is stationary in x). A non-rigorous computation
[LDC12], supported by a rigorous version for the asymmetric exclusion process in [OQR17] which
leads to a formally similar result, indicates that G is given by a Fredholm Pfaffian
Gflat = Pf(J − K)L2 ([0,∞))
KP(Z) 7
 
1 K (u+r,v+r) f (u+r)
with K an explicit 2 × 2 matrix kernel of the form K(u, v) = −f (v+r) K2 (u,v)
where K1
and K2 are antisymmetric and K satisfies the same differential relations (3.7)/(3.9) as the multipoint
KPZ fixed point kernel (ignoring the one involving Dx ). Thus defining the 2 × 2 matrix Q =
−1 −1 −1 −1 0 I

(I − J K) J K(0, 0) (here J = −I 0 ), we get
1 3
∂t Q + 12 ∂r Q + 21 (∂r Q)2 = 0.
On the other hand, in this case we have Φflat := ∂r log Gflat = 21 tr(Q) (the important factor 12 comes
from the fact that now we are dealing with a Fredholm Pfaffian). We claim that Φflat = 21 tr(Q) satisfies
the (integrated) KdV equation,
1 3
∂t Φflat + 12 ∂r Φflat + 12 (∂r Φflat )2 = 0.
Comparing the two equations we see that this can only hold if tr(∂r Q)2 = 2 tr((∂r Q)2 . Now ∂r Q is
just a 2 × 2 (real-valued) matrix, so this is equivalent to ∂r Q having a single repeated eigenvalue. This
is proved directly from the stated properties of K in Prop. C.1. The initial condition is Φflat (0, x, r) =
−e−r .
Example 2.3. (Spiked/half-Brownian initial data) Consider now the solution hb of (1.1) with
λ = ν = 14 , σ = 1 and m-spiked initial data, where b = (b1 , · · · , bm ) ∈ Rm are the spike parameters.
When m = 1, this corresponds to half-Brownian initial data (more precisely at the level of the SHE one
sets Z(0, x) = eB(x)+b1 x 1x≥0 where B(x) is a Brownian motion with diffusivity 2); for the general
m ≥ 1 case we refer to [BCF14, Defn. 1.9] for the definition. Define Gb as in (2.1) again. Then from
[BCF14, Thm. 1.10] we get now that Gb (t, 0, r) = det(I − K0 )L2 [0,∞) with
3 −1/3 r)η m
t−1/3 π eη /3−(u+t Y Γ(t−1/3 ξ − bk )
Z Z
K0 (u, v) = dη dξ ,
Ct Ct0 sin(t−1/3 π(η − ξ)) eξ /3−(v+t
3 −1/3 r)ξ
k=1
Γ(t−1/3 η − bk )

where Ct goes from 14 t1/3 − i∞ to 14 t1/3 + i∞ crossing the real axis to the right of t1/3 b1 , . . . , t1/3 bm
and Ct0 = Ct + 12 t1/3 . We scale (η, ξ) 7→ (t1/3 η, t1/3 ξ) and (u, v) 7→ (t−1/3 u, t−1/3 v) (in the Fred-
3
π etη /3−(u+r)η Qm Γ(ξ−bk )
R R
holm determinant) so that K0 (u, v) is now given as C1 dη C 0 dξ sin(π(η−ξ)) tξ3 /3−(v+r)ξ k=1 Γ(η−bk ) .
1 e
Since hb+x/t (t, x) − x2 /t is stationary in x [BCF14, Rem. 1.14]), we may write
Gb (t, x, r) = det(I − K)L2 [0,∞)

with
3 2 m
etη /3−(u+r+x /t)η Y Γ(ξ − bk − x/t)
Z Z
π
K(u, v) = dη dξ
Ce1 Ce10 sin(π(η − ξ)) etξ3 /3−(v+r+x2 /t)ξ Γ(η − bk − x/t)
k=1
3 /3+xη 2 −(u+r)η−ux/t m
t2/3 π tη Γ(ξ − bk )
Z Z
e Y
= dη dξ
C1 C10 sin(π(η − ξ)) etξ3 /3+xξ2 −(v+r)ξ−vx/t k=1
Γ(η − bk )

(in the first line the contour Ce1 has to cross the real axis to the right of bi + x/t for all i, so after the
change of variables leading to the second equality the contours C1 and C10 have the same meaning
as above). As in Ex. 2.1, K satisfies the necessary differential relations, so it follows again that
φb := ∂r2 log Gb satisfies KP-II.
Example 2.4. (The BBP distribution for spiked random matrices) Consider Gb as in the previous
example. It is known (see [BCF14, Cor.1.15] for the case t = 1, x = 0, the general case follows in the
same way or by scaling and shift invariance) that
lim Gε1/2 b (ε−3/2 t, ε−1 x, ε−1/2 r) = FBBP,t1/3 b (t−1/3 r + t−4/3 x2 )
ε→0

where FBBP,b is the Baik-Ben Arous-Péché (BBP) distribution asrising from spiked (unitarily invariant)
random matrices [BBP05]. On the other hand, an easy computation shows that, for each fixed ε > 0,
KP(Z) 8

φεb (t, x, r) := ∂r2 log Gε1/2 b (ε−3/2 t, ε−1 x, ε−1/2 r) satisfies KP-II as well. As a consequence, one
expects that if
GBBP,b (t, x, r) = FBBP,t1/3 b (t−1/3 r + t−4/3 x2 )
then φBBP,b := ∂r2 log GBBP,b will also satisfy KP-II. This is indeed the case, as can be checked in a
similar way as above using the explicit Fredholm determinant formula for FBBP,b (see for instance
[BCF14, Eqn. (1.4)]).

3. T HE DETERMINANTAL FORMULA AND DERIVATION OF KP

3.1. KPZ fixed point formula. In [MQR17] a determinantal formula is given for (1.4). The first thing
to do is to rewrite the kernel in a natural way to obtain logarithmic derivatives in r, t and x. We recall
how the kernel is defined. For h ∈ UC let
hit h h No hit h
PNo
`1 ,`2 (u1 , u2 )du2 = PB(`1 )=u1 (B(y) > h(y) on [`1 , `2 ], B(`2 ) ∈ du2 ) , PHit
`1 ,`2 = I − P`1 ,`2 ,

where B is a Brownian motion with diffusion coefficient 2. The Brownian scattering transform of h is
the formal object
2
h −`2 ∂ 2 hit h −`2 ∂ 2 2
Khypo(h) = lim e`1 ∂ PHit
`1 ,`2 e = I − lim e`1 ∂ PNo
`1 ,`2 e , (3.1)
`1 →−∞ `1 →−∞
`2 →∞ `2 →∞

hit h
where PHit/No
`1 ,`2 are thought of as operators with the given integral kernels. This doesn’t make sense
since the backward heat operator is asked to act on non-analytic functions. In fact, Khypo(h) will never
actually be used by itself, but only after conjugation by the Airy unitary group,
1 3
Ut = e− 3 t∂ , (3.2)
with t 6= 0 and ∂3the third derivative operator. For a fixed vector a ∈ Rm and indices n1 < . . . < nm
we introduce the functions
χa (nj , x) = 1x>aj , χ̄a (nj , x) = 1x≤aj ,
which we also regard as multiplication operators acting on L2 ({x1 , . . . , xm } × R). For simplicity we
will write χa (x) when m = 1.
We take t > 0 in which case the Airy semigroup acts by convolution with Airy functions. These are
not themselves in L2 (R); however, for t > 0 and r > −∞, U−1 2
t χr maps L (R) into the domain of
2
ex∂ for any x ∈ R. So for t > 0 and r > −∞, we define on L2 ([r, ∞))
hypo(h) h −`2 ∂ 2 2
Kt = lim Ut e`1 ∂ PHit
`1 ,`2 e U−1
t . (3.3)
`1 →−∞
`2 →∞

For any t > 0 and r > −∞ the limit on the right hand side of (3.3) exists in trace class on L2 ([r, ∞)),
and defines the left hand side as a trace class operator in this space. It satisfies the semigroup property
hypo(h) hypo(h)
Us Kt U−1
s = Kt+s

Because it satisfies the semigroup property, we can write (at least informally)
hypo(h)
Kt = Ut Khypo(h) U−1
t . (3.4)
Note that we avoid the problem of domains by not defining the left hand side of (3.3) as a product
of three operators, but just as one operator with the semigroup property. In this sense the Brownian
scattering operator is the germ of the semigroup. Alternatively one can think of the Brownian scattering
operator as the entire semigroup (3.4). The fact that (3.1) is formal is important. We will see in (A.2)
that the limit of (3.4) as t & 0 is not Khypo(h) .
From Khypo(h) we build an extended Brownian scattering operator acting on L2 ({x1 , · · · , xm } × R),
hypo(h) 2 2 2
Kext (xi , ·; xj , ·) = −e(xj −xi )∂ 1xi <xj + e−xi ∂ Khypo(h) exj ∂ , (3.5)
KP(Z) 9

with the analogous caveat that in order to make sense each of the above (xi , xj ) entries should be
conjugated by Ut and the whole operator should be surrounded by χr which acts independently in each
coordinate by χri . In this language the KPZ fixed point formula reads
 
hypo(h )
F (t, x1 , . . . , xn , r1 , . . . , rn ) = det I − χr Ut Kext 0 U−1t χ r 2
. (3.6)
L ({x1 ,...,xm }×R)

hypo(h0 ) hypo(h0 )
Sometimes we write Kt,ext = Ut Kext U−1
t .

3.2. The logarithmic derivative. The next two sections contain the proof that the Fredholm determi-
nant (3.6) satisfies the matrix KP equation. After we performed the very complicated computation, we
discovered that a very similar argument was actually known [Po89] in the one dimensional case. It is
shown there that the Fredholm determinant of a kernel satisfying suitable differential relations solves
the Hirota equations. The differential relations turn out to be equivalent to the way the kernel depends
on t, x and r above. It seems to actually go back to [ZS74; ZS79] though it is not explicit there, and
rediscovered in the literature multiple times.
Call x = (x1 , . . . , xn ) ∈ Rn with x1 < x2 < · · · < xn , and r = (r1 , . . . , rn ) ∈ Rn . Let
hypo(h)
Φ(t, r, x) = Dr log(det(I − χr Kt,ext χr )L2 ({x1 ,...,xn }×R )
where Dr is defined in (1.5). Shifting variables in the kernel we get
Φ(t, r, x) = Dr log det(I − K)L2 (R≥0 )⊕···⊕L2 (R≥0 )
hypo(h)
where K(xa , ua ; xb , ub ) = Kt,ext (xa , ua + ra ; xb , ub + rb ). From now on we omit the subscript
on the Fredholm determinant and traces. We think think of K as an operator-valued matrix; to ease
notation we will write Kab = K(xa , ·; xb , ·).
Given an operator A acting on L2 (R) with kernel A(u, v) we will write d1 A and d2 A for the
operators with kernels given by
(d1 A)(u, v) = ∂u A(u, v) and (d2 A)(u, v) = ∂v A(u, v),
while in the matrix case we let

Di K = di Kab a,b=1,...,n
, i = 1, 2.

Note that this is just a notational device; di and Di are not meant to denote operators. By definition of
K we have
Dr K = (D1 + D2 )K. (3.7)
Then
X
Φ(t, r, x) = − tr((I − K)−1 ∂ra K) = − tr((I − K)−1 Dr K) = − tr((I − K)−1 (D1 + D2 )K)
a
= − tr((I − K)−1 D1 K + D2 ((I − K)−1 K))
= − tr(D1 K(I − K)−1 + D2 ((I − K)−1 K)) = − tr((D1 + D2 )((I − K)−1 K))
XZ ∞ X
dξ ∂ξ (I − K)−1 K aa (ξ, ξ) = (I − K)−1 K aa (0, 0),
 
=−
a 0 a

where we used the cyclicity of the trace. Introducing the notation



[A] = Aa,b (0, 0) a,b=1,...,n ,

this tells us that Φ can be expressed as an n-dimensional trace,


Φ(t, r, x) = tr[RK] with R = (I − K)−1 .
Note here that I − K is invertible because the determinant is non-zero.
KP(Z) 10

3.3. Formulas for the partial derivatives. Let now Q denote the matrix
Q = [RK] (3.8)
and write K0 = Dr K. The above argument shows that Dr log F = tr Q. The goal is now to prove that
Q satisfies the matrix KP equation (1.6).
Using the general formula ∂a (I − A(a))−1 = (I − A(a))−1 ∂a A(a)(I − A(a))−1 for an operator
A(a) depending smoothly on a parameter a we have
Dr Q = [Dr (RK)] = [ a (R∂ra KRK + R∂ra K)] = [RK0 RK + RK0 ] = [RK0 R],
P

Dr2 Q = 2[RK0 RK0 R] + [RK00 R],


Dr3 Q = 6[RK0 RK0 RK0 R] + 3[RK00 RK0 R] + 3[RK0 RK00 R] + [RK000 R],
∂t Q = [R∂t KR].
On the other hand, from the definition of the Brownian scattering transform we have
∂t K = − 13 (D31 + D32 )K and Dx K = (D22 − D21 )K (3.9)
2
(we are using here also the fact that (d31 + d32 )e`∂ = 0 for any ` > 0).
Next we want to compute (Dr Q)2 . Note that, in general,
XZ ∞
([A][B])a,b = − dη ∂η (Aac (0, η)Bcb (η, 0))
c 0
XZ ∞ 
=− dη d2 Aac (0, η)Bcb (η, 0) + Aac (0, η)d1 Bcb (η, 0)
c 0

= −(D2 AB)ab (0, 0) − (AD1 B)ab (0, 0)


so that the following integration by parts formula holds:
[A][B] = −[AD1 B + D2 AB]. (3.10)
We will use this in the formula
(Dr Q)2 = ([RK0 RK] + [RK0 ])2 = [RK0 RK]2 + [RK0 RK][RK0 ] + [RK0 ][RK0 RK] + [RK0 ]2 .
The first term equals (using D2 (Ka Kb ) = Ka D2 Kb )
−[RK0 R(D2 KR+KD1 R)K0 RK] = −[RK0 RK0 RK0 RK]+[RK0 R(D1 KR−KD1 R)K0 RK].
Similarly, the fourth term equals
−[RK00 RK0 ] + [R(d1 K0 R − K0 d1 R)K0 ]
and the two middle ones equal
−[RK0 RK0 RK0 ]−[RK00 RK0 RK]+[RK0 R(d1 KR−Kd1 R)K0 ]+[R(d1 K0 R−K0 d1 R)K0 RK].
Using this together with our formulas for ∂t Q and Dr3 Q yields
12∂t Q(t, r, x) + Dr3 Q(t, r, x) + 6(Dr Q(t, r, x))2
= −4[R(D31 + D32 )KR] + 6[RK0 RK0 RK0 R] + 3[RK00 RK0 R] −(1)+(2)+(3)
0 00
+ 3[RK RK R] + [R(D31 + D32 + 3D21 D2 + 3D1 D22 )KR] +(4)+(5)
− 6[RK00 RK0 ] + 6[R(D1 K0 R − K0 D1 R)K0 ] −(6)+(7)
0 0 0 00 0
− 6[RK RK RK ] − 6[RK RK RK] −(8)−(9)
0 0 0 0 0
+ 6[RK R(D1 KR − KD1 R)K ] + 6[R(D1 K R − K D1 R)K RK] +(10)+(11)
− 6[RK0 RK0 RK0 RK] + 6[RK0 R(D1 KR − KD1 R)K0 RK]. −(12)+(13)
KP(Z) 11

We will use the identity KR = RK = R − I repeatedly. (12) equals 6[RK0 RK0 RK0 R] −
6[RK0 RK0 RK0 ], so
SI := −(1) + (2) + (5) − (8) − (12) = −3[R(D31 + D32 − D21 D2 − D1 D22 )KR]
= −3[R(D1 − D2 )2 (D1 + D2 )KR] = −3[R(D1 − D2 )2 K0 R].
Similarly
SII := +(3) + (4) − (6) − (9) = −3[RK00 RK0 R] + 3[RK0 RK00 R]
and
SIII := +(7)+(10)+(11)+(13) = +6[R(D1 K0 R−K0 D1 R)K0 R]+6[RK0 R(D1 KR−KD1 R)K0 R].

So
4∂t Q(t, r, x) + 31 Dr3 Q(t, r, x) + 2(Dr Q(t, r, x))2 = 31 (SI + SII + SIII )
= −[RK00 RK0 R] + [RK0 RK00 R] − [R(D31 + D32 − D21 D2 − D1 D22 )KR]
+ 2[R(D1 K0 R − K0 D1 R)K0 R] + 2[RK0 R(D1 KR − KD1 R)K0 R]
= −[R(D2 − D1 )K0 RK0 R] + [RK0 RK00 R] − [R(D31 + D32 − D21 D2 − D1 D22 )KR]
− 2[RK0 RD1 RK0 R] + 2[RK0 RD1 KRK0 R].
Using now −[RK0 RD1 RK0 R] + [RK0 RD1 KRK0 R] = −[RK0 RD1 (I − K)RK0 R], which equals
−[RK0 RD1 K0 R], yields

4∂t Q(t, r, x) + 13 Dr3 Q(t, r, x) + 2(Dr Q(t, r, x))2


= −[R(D2 − D1 )K0 RK0 R] − [RK0 R(D1 − D2 )K0 R] − [R(D1 − D2 )2 (D1 + D2 )KR]. (3.11)

Remark 3.1. At this stage we can already see that the one point distribution in the flat case h0 ≡ 0
satisfies the (integrated) KdV equation. In fact, the arguments in [MQR17, Sec. 4.4] lead in this case to
hypo(h)
Kt = I − Ut (I − %)χ̄0 (I − %)U−1 −1
t = Ut %Ut with %f (x) = f (−x), which means that K is a
Hankel kernel, so that the right hand side in (3.11) vanishes.

Next we add the derivatives in the xi variables. As for Dr2 Q, we have


Dx2 Q = 2[RDx KRDx KR] + [RDx2 KR]. (3.12)
On the other hand, if we apply Dr to (3.11) we get
− [RK0 RDx KRK0 R] − [RDx K0 RK0 R] − 2[RDx KRK0 RK0 R] − [RDx KRK00 R]
+ 2[RK0 RK0 RDx KR] + [RK00 RDx KR] + [RK0 RDx K0 R] + [RK0 RDx KRK0 R]
+ [RK0 R(d1 − d2 )Dx KR] + [R(d1 − d2 )Dx K0 R] + [R(d1 − d2 )Dx KRK0 R].

Note that the first and eight terms cancel. We want to add Dx2 Q(t, r, x). Note that (D1 − D2 )Dx K0 =
Dx (D21 − D22 )K = −Dx2 K, so the next-to-last term in cancels the second bracket on the right hand side
of (3.12). Using additionally Dx K0 + (D1 − D2 )Dx K = 2D1 Dx K and −Dx K0 + (D1 − D2 )Dx K =
−2D2 Dx K and writing
q = Dr Q,
we deduce that
4∂t q + 31 ∂r3 q + 4qDr q + Dx2 Q(t, r, x)
= −2[RD2 Dx KRK0 R] − 2[RDx KRK0 RK0 R] − [RDx KRK00 R]
(3.13)
+ 2[RK0 RK0 RDx KR] + [RK00 RDx KR] + 2[RK0 RD1 Dx KR]
+ 2[RDx KRDx KR].
KP(Z) 12

We claim that the right hand side equals two times


− [RDx KRD1 KRK0 R] + [RDx D2 KRK0 R] + [RK0 RD2 KRDx KR] + [RK0 RD1 Dx KR]
 

− [RDx KRD2 KRK0 R] + [RDx KRD1 K0 R] + [RK0 RD1 KRDx KR] + [RD2 K0 RDx KR] .
 
(3.14)
To see this, express the right hand side of (3.13) as 2(r1 +r2 +. . .+r7 ), express (3.14) as q1 +q2 +. . .+q8 ,
and note first that r1 = q2 , r6 = q4 , r2 = q1 + q5 and r4 = q3 + q7 . On the other hand we have
r3 = − 12 [RDx KR(D1 + D2 )K0 R] = q6 + 12 [RDx KR(D1 − D2 )K0 R] = q6 − 21 r7 and similarly
r5 = 12 [R(D1 + D2 )K0 RDx KR] = q8 + 12 [R(D1 − D2 )K0 RDx KR] = q8 − 21 r7 . This gives
r3 + r5 + r7 = q6 + q8 , and finishes proving the claim.
Integrating by parts (i.e. using (3.10)) within each parenthesis in (3.14) we get
1
2 (3.13) = −[R∂x K(RD1 KR − D1 R)K0 R] + [RDx K][RK0 R]
+ [RK0 (RD2 KR − D2 R)Dx KR] − [RK0 R][Dx KR]
− [RDx K(RD2 KR − D2 R)K0 R] + [RDx KR][K0 R]
+ [RK0 (RD1 KR − D1 R)Dx KR] − [RK0 ][RDx KR].
Write this as s1 + · · · + s8 . Notice that s1 + s5 yields a term involving
(RD1 KR−D1 R)+(RD2 KR−D2 R) = RK0 R−(D1 +D2 )R = R(KD1 K+D2 KK)R−(D1 +D2 )I,
where we have used RK = KR = R − I again, and thus integrating by parts one more time we get
s1 + s5 = −[RDx K(R(KD1 K + D2 KK)R − (D1 + D2 )I)K0 R]
= [RDx KRK][KRK0 R] + [RDx K((D1 + D2 )I)K0 R]
= ([RDx KR] − [RDx K])([RK0 R] − [K0 R]) + [RDx K((D1 + D2 )I)K0 R]
= [RDx KR][RK0 R] − s6 − s2 + [RDx K][K0 R] + [RDx K((D1 + D2 )I)K0 R].
In a similar fashion we get
s3 + s7 = −[RK0 R][RDx KR] − s4 − s8 − [RK0 ][Dx KR] − [RK0 ((D1 + D2 )I)Dx KR].
Therefore
1
2 (3.13) = [RDx K][K0 R]+[RDx K((D1 +D2 )I)K0 R]−[RK0 ][Dx KR]−[RK0 ((D1 +D2 )I)Dx KR].
In order to complete the proof we note that if A and B are nice kernels then integrating by parts we get
[A((D1 + D2 )I)B] = [AD1 B] + [AD2 IB] = −[A][B],

which immediately yields (3.13) = 0, as desired.

A PPENDIX A. M ULTIPOINT INITIAL DATA

A.1. t → 0 limit of the Brownian scattering operator. Let the initial data for the KPZ fixed point
be a finite collection of narrow wedges as in Ex. 1.3. Fix x1 < . . . < xm . We want to compute
~
2 hypo(d~ba ) xj ∂ 2
lim e−xi ∂ Kt e .
t→0

Throughout this section we will use the notation


1 3 2 2
St,x = e− 3 t∂ +x∂ = ex∂ Ut .
R∞
The St,x act by convolution St,x f (z) = −∞ dy St,x (z − y)f (y) where, for t > 0,
2x3 ux
St,x (u) = t−1/3 e 3t2 − t Ai(−t−1/3 u + t−4/3 x2 ),
while for t < 0, St,x (u) = S−t,x (−u) (see [MQR17, Eqn. (3.10)]).
KP(Z) 13

We consider first the single narrow wedge case, k = 1, writing a = a1 , b = b1 . In this case, by
hypo(h)
definition of Kt we have
hypo(dba ) Hit db 2 2 2
Kt = lim S−t,−` P−`,`a e−`∂ St,−` = lim S−t,−` e(a+`)∂ χ̄b e(`−a)∂ St,−` = S−t,a χ̄b St,−a .
`→∞ `→∞

In particular,
2 hypo(dba ) a∂ 2
lim e−a∂ Kt e = lim S−t,0 χ̄b St,−0 = χ̄b ,
t→0 t→0

so in the case xi ≤ a ≤ xj we get


2 hypo(dba ) xj ∂ 2 2 2 hypo(dba ) a∂ 2  (xj −a)∂ 2
lim e−xi ∂ Kt e = lim e(a−xi )∂ e−a∂ Kt e e
t→0 t→0
2 2 b
= e(a−xi )∂ χ̄b e(xj −a)∂ = PHit da
xi ,xj .

Next consider the case when the inequality xi ≤ a ≤ xj does not hold. We have

−x ∂ 2 hypo(db ) x ∂ 2
Z b
e i K
t
a
e j (u, v) ≤ dη |St,a−xi (u − η)||St,xj −a (v − η)|
−∞
Z b
−2/3 3 /3t2 +2(x 3 2
j −a) /3t −(u−η)(a−xi )/t−(v−η)(xj −a)/t
=t dη e2(a−xi )
−∞
× |Ai(t−1/3 (u − η) + t−4/3 (a − xi )2 )Ai(t−1/3 (v − η) + t−4/3 (xj − a)2 )|.

Let κ ∈ (0, 1) and note that both t−1/3 (u − η) + (1 − κ2/3 )t−4/3 (a − xi )2 > 0 and t−1/3 (v − η) +
(1 − κ2/3 )t−4/3 (xj − a)2 > 0 for all η ≤ b if t is small enough. Therefore, using the classical bound
2 3/2
on the Airy function |Ai(u)| ≤ C e− 3 (u∨0) , the last integral is bounded by
Z b
−2/3 2[(a−xi )3 −κ|a−xi |3 +(xj −a)3 −κ|xj −a|3 ]/3t2
Ct e dη e−(u−η)(a−xi )/t−(v−η)(xj −a)/t .
−∞
2
It is easy to check that whenever xi ≤ a ≤ xj does not hold, the prefactor can be bounded by e−c/t
for some c > 0 by choosing κ close enough to 1, and hence computing the η integral shows that the
whole expression goes to 0 as t → 0.
The conclusion of all this is that, in the case of narrow wedge initial data dba ,
2 hypo(dba ) xj ∂ 2
lim e−xi ∂ Kt e = lim S−t,a−xi χ̄b St,xj −a
t→0 t→0 (A.1)
(a−xi )∂ 2 (xj −a)∂ 2 dba
=e χ̄b e 1xi ≤a≤xj = PHit
xi ,xj 1xi ≤xj .

~
Now we turn to the general case h = d~ab . For ` > |a1 | ∨ |a2 | we have, by inclusion-exclusion,

~ k
Hit db X X 2 2 2 2
P−`,`~a = (−1)n+1 e(ap1 −`)∂ χ̄bp1 e(ap2 −ap1 )∂ χ̄bp2 · · · e(apn −apn−1 )∂ χ̄bpn e(`−apn )∂ ,
n=1 1≤p1 <···<pn ≤k
~
∂2 hypo(d~ba ) xj ∂ 2
so e−xi Kt e equals
k
X X 2 2
(−1)n+1 S−t,ap1 −xi χ̄bp1 e(ap2 −ap1 )∂ χ̄bp2 · · · e(apn −apn−1 )∂ χ̄bpn St,xj −apn .
n=1 1≤p1 <···<pn ≤k

Each summand can be factored as


  
S−t,ap1 −xi χ̄bp1 St,0 S−t,ap2 −ap1 χ̄bp2 St,0 · · · S−t,apn −apn−1 χ̄bpn St,xj −apn .
KP(Z) 14

bp b
Hit dap1 Hit dapn
By (A.1), as t → 0 the first factor goes to Pxi ,ap1 1 1xi ≤ap1 , the last factor goes to Papn−1p,x
n
j 1apn ≤xj ,
b
Hit daps
and each of the inner factors goes to Paps−1p,as ps , 2 ≤ s ≤ n − 1. Therefore

k bp bp b
2
~
hypo(d~ba ) xj ∂ 2 X X Hit dap1 Hit dap2 Hit dapn
lim e−xi ∂ Kt e = (−1)n+1 Pxi ,ap1 1 Pap1 ,ap22 · · · Papn−1p,x
n
j 1xi ≤ap1 , xj ≥apn
t→0
n=1 1≤p1 <···<pn ≤k

and then, using inclusion-exclusion again, we deduce finally that


~ ~
2 hypo(d~ba ) xj ∂ 2 Hit db
lim e−xi ∂ Kt e = Pxi ,xj~a 1xi ≤xj .
t→0

A.2. Matrix KP initial data. Consider now compactly supported initial data h ∈ UC, meaning that
~
h(y) = −∞ for y outside some compact interval. Approximating h by initial data of the form d~ab we
obtain
2 hypo(h) xj ∂ 2
lim e−xi ∂ Kt e = PHit h
xi ,xj 1xi ≤xj .
t→0

In terms of the extended Brownian scattering operator (3.5), this gives


hypo(h) 2 hypo(h) 2
K0,ext (xi , ·; xj , ·) := lim e−xi ∂ Kt,ext (xi , ·; xj , ·)exj ∂
t→0
No hit h

−Pxi ,xj
 if i < j, (A.2)
= χ̄h(xi ) if i = j,

0 if i > j.

hypo(h)
Remark A.1. This formula recovers correctly the KPZ fixed point initial data: since K0,ext (xi , ·; xj , ·)
is upper triangular, we have
m m
hypo(h) 
Y  Y
det I − χr K0,ext χr = det I − χri χ̄h(xi ) χri = 1ri ≥f h(xi )
i=1 i=1

as desired.

Now we compute [RK] with K as in Section 3.3. Recalling that in that section we shifted the entries
of the kernel by ri , this corresponds in the current setting to evaluating the (i, j) entry of (I − K)−1 K at
hypo(h)
(ri , rj ). By a simple argument using the fact that K0,ext is upper triangular we can expand (formally)
the entries of (I − K)−1 K as

X |π|−1
Y
−1

(I − K) K i,j = 1i≤j χrπ(n) Kπ(n),π(n+1) χrπ(n+1) , (A.3)
π:i→j n=1
π incr.

where the sum is over non-decreasing paths π going from i to j along integers and |π| denotes the
length of the path. Fix i ≤ j and assume first that r` ≥ h(x` ) for each i ≤ ` ≤ j. Consider a fixed path
π from i to j. If π(n) = π(n + 1) for some n then the corresponding factor in the product inside the
sum will be χrπ(n) χ̄h(xπ(n) ) χrπ(n) = 0, so only strictly increasing paths contribute to the sum and we
get (note that this sum is now finite)
|π|−1
X Y
(I − K)−1 K (−1)|π|−1 χrπ(n) PNo hit h

i,j
= 1i<j xπ(n) ,xπ(n+1) χrπ(n+1) .
π:i→j n=1
π str. incr.
KP(Z) 15

Evaluating at (ri , rj ) and applying inclusion-exclusion again, we deduce that


−~r
≥h,≤−d~x
[RK]i,j = −1i<j Pxi ,xj (ri , rj )
:= −1i<j PB(xi )=ri B(y) ≥ h(y) ∀ y ∈ [xi , xj ], (A.4)

B(xn ) ≤ rn for each xn ∈ (xi , xj ), B(xj ) ∈ drj /drj .

Suppose next that r` < h(x` ) for some i ≤ ` ≤ j, and for simplicity assume that this is the
only such index satisfying the condition (the argument can be generalized easily). Assume also that
i < j. From the argument in the previous case we know that if π : i → j has a constant piece which
stays at any index other than `, then π does not contribute to the sum in (A.3). Hence any path π
from i to j which does contribute to the sum can be decomposed as π1 ◦ υ ◦ π2 with π1 : i → `
and π2 : ` → j strictly increasing (we allow for π1 or π2 to be empty if ` = i or ` = j), and υ
staying at ` for a given number of steps (which could be 0). The product inside the sum in (A.3) splits
between factors coming from the three pieces of the path, and from the middle part we get a factor
|υ|
χr` χ̄h(x` ) χr` = I · 1|υ|=0 + χr` χ̄h(x` ) 1|υ|>0 . In other words, and repeating the previous argument,
 
|π1 |−1
X X X Y
(I − K)−1 K (−1)|π1 |−1 PNo hit h

i,j
=− χr x ,x
π1 (n)
χr
π1 (n)

π1 (n+1) π1 (n+1)
π1 :i→` π2 :`→j ν≥0 n=1
π1 str. incr. π2 str. incr.
 
|π2 |−1
Y
× I · 1ν=0 + χr` χ̄h(x` ) 1ν>0 (−1)|π2 |−1 χrπ2 (n) PNo hit h

xπ (n) ,xπ χrπ2 (n+1) 
2 2 (n+1)
n=1
X ≥h,≤d~~rx  ≥h,≤d~r
=− χri P xi ,x` I · 1ν=0 + χr` χ̄h(x` ) 1ν>0 Px` ,xj ~x χrj .
ν≥0

≥h,≤d~~rx
But P xi ,x` (u, v) = 0, because at the endpoint v it requires h(x` ) ≤ v ≤ r` (the analogous statement
holds for the other factor). Hence we conclude that, in this case, [RK]i,j = 0, which by the same
reason means that (A.4) still holds.
Suppose finally that i = j and ri < h(xi ). Now the only possible paths in (A.3) are constant paths π
of arbitrary length |π| ≥ 1. Each such |π| contributes a term of the form χri χ̄h(xi ) , which evaluated at
(ri , ri ) is taken to be 1, and hence [RK]i,i diverges to ∞ in this case (which coincides with the physical
meaning of this quantity, namely ∂ri log F (t, x1 , . . . , xn , r1 , . . . , rn )).
The conclusion is then that

≥h,≤d~~x
r
−Pxi ,xj (ri , rj ) if i < j,


[RK]i,j = ∞ if i = j and ri < h(xi ),


0 otherwise.

A PPENDIX B. A LTERNATIVE DERIVATION OF KP-II FOR NARROW WEDGE MULTIPOINT


DISTRIBUTIONS

In this section we will derive the KP-II equation (1.7) for the Airy2 process directly using the
path-integral formula for the KPZ fixed point [MQR17, Prop. 4.3]. Define F (t, ~x + y, ~r + a) as in Sec.
hypo(h0 )
1.4. Then letting Kt,x = Kt (x, ·; x, ·) we have
2 2
F (t, ~x + y, ~r + a) = det(I − Kt,x1 +y + χ̄r1 +a e(x2 −x1 )∂ χ̄r2 +a · · · χ̄rm +a e(x1 −xm )∂ Kt,x1 +y )
2 2
= det(I − K + χ̄r1 e(x2 −x1 )∂ χ̄r2 · · · χ̄rm e(x1 −xm )∂ K)

with K = Kt,x1 +y (a + ·, a + ·) = ea∂ Kt,x1 +y e−a∂ . Note that the product of operators preceding K in
the last term does not depend on t, y or a; call it I−P so that F = det(I−PK). Up to here this is general,
KP(Z) 16

but now we specialize to the narrow wedge case, for which K = ea∂ (St,−x1 −y )∗ χ̄0 St,x1 +y e−a∂ . Using
the cyclic property of the determinant we get
F = det(I − χ̄0 St,x1 +y e−a∂ Pea∂ (St,−x1 −y )∗ χ̄0 ) = det(I − χ̄0 e−a∂ St,x1 +y P(St,−x1 −y )∗ ea∂ χ̄0 )
= det(I − χ0 %e−a∂ St,x1 +y P(St,−x1 −y )∗ ea∂ %χ0 ).
So letting
L = %e−a∂ St,x1 +y P(St,−x1 −y )∗ ea∂ % = ea∂ (St,x1 +y )∗ (%P%)St,−x1 −y e−a∂
(the second equality is a simple computation) we get F = det(I − L). Now ∂a L = (d1 + d2 )L,
∂t L = − 13 (d31 + d32 ), and ∂y L = (d21 − d22 )L, and this is the same equation that the kernel satisfies
in the one point computation (except for the change y 7→ −y, which again is irrelevant), so the same
computations yields that φ = ∂a2 log(F ) solves KP-II, and translating back to the Dr , Dx derivatives
yields the same result.

A PPENDIX C. A N IDENTITY FOR THE FLAT KPZ F REDHOLM P FAFFIAN KERNEL

Proposition C.1. Consider two 2 × 2 matrix kernels


   
M M1,2 E F1
M = M1,1 ∗
2,1 M1,1
and N= F2 E∗

where M1,2 , M2,1 , F1 and F2 are antisymmetric. Assume that I − M is invertible and let R =
(I − M)−1 , and assume that I − M1,1 is also invertible. Then the 2 × 2 matrix RNR(0, 0) is a multiple
of the identity.

Proof. We claim first that 


R= A B
C A∗
with B and C antisymmetric. To see this we use the formula
 
A AM1,2 (I−M∗ )−1
R = (I−M∗ )−1 M A (I−M∗ )−1 +(I−M∗ )−1 M1,1 AM ∗ −1
1,1 2,1 1,1 1,1 2,1 1,2 (I−M1,1 )

with A = (I − M1,1 − M1,2 (I − M∗1,1 )−1 M2,1 )−1 . We need to show that A−1 R∗2,2 = I. We have
A−1 = I − M1,2 (I − M∗1,1 )−1 M2,1 (I − M1,1 )−1 (I − M1,1 ),


R∗2,2 = (I − M1,1 )−1 I + M1,2 A∗ M2,1 (I − M1,1 )−1 ,




so
A−1 R∗2,2 = I − M1,2 (I − M∗1,1 )−1 M2,1 (I − M1,1 )−1 I + M1,2 A∗ M2,1 (I − M1,1 )−1
 

= I + M1,2 (A∗ − (I − M∗1,1 )−1 )M2,1 (I − M1,1 )−1


− M1,2 (I − M∗1,1 )−1 M2,1 (I − M1,1 )−1 M1,2 A∗ M2,1 (I − M1,1 )−1 .
Thus it is enough to show that A∗ − (I − M∗1,1 )−1 = (I − M∗1,1 )−1 M2,1 (I − M1,1 )−1 M1,2 A∗ , which
is equivalent to A(I − M1,1 ) − I = AM1,2 (I − M∗1,1 )−1 M2,1 , and this is true by definition of A.
In order to finish proving our claim we need to check that R1,2 and R2,1 are antisymmetric. For the
first one we need to show that AM1,2 (I − M∗1,1 )−1 = (I − M1,1 )−1 M1,2 A∗ . Premultiplying by
(I − M1,1 ) and postmultiplying by its adjoint we see that this is equivalent to
−1 −1
I−M1,2 (I−M∗1,1 )−1 M2,1 (I−M1,1 ) M1,2 = M1,2 (I−(I−M∗1,1 )−1 M2,1 (I−M1,1 )−1 M1,2 ,

which is straightforwardly true. The second one is similar.


Having established the claim, we now have
AEA + BF2 A + AF1 C + BE∗ C AEB + BF2 B + AF1 A∗ + BE∗ A∗
 
RNR = .
CEA + A∗ F2 A + CF1 C + A∗ E∗ C CEB + A∗ F2 B + CF1 A∗ + A∗ E∗ A∗
But A∗ E∗ A∗ (0, 0) = AEA(0, 0), A∗ F2 B(0, 0) = B∗ F∗2 A(0, 0) = BF∗2 A(0, 0) by antisymmetry
of B and F2 , and similarly CF1 A∗ (0, 0) = AF1 C(0, 0), CEB(0, 0) = BE∗ C(0, 0). This shows
KP(Z) 17

that the diagonal entries of RNR(0, 0) coincide. On the other hand, AEB(0, 0) = B∗ E∗ A∗ (0, 0) =
−BE∗ A∗ (0, 0) while BF2 B and AF1 A∗ are antisymmetric so they vanish when evaluated at the
diagonal. This shows that the (1, 2) entry of RNR(0, 0) vanishes. The same argument shows that the
(2, 1) entry of the matrix vanishes, and this finishes the proof. 

Acknowledgements. JQ was supported by the Natural Sciences and Engineering Research Council of
Canada and by a Killam research fellowship. DR was supported by Conicyt Basal-CMM Proyecto/Grant
PAI AFB-170001, by Programa Iniciativa Científica Milenio grant number NC120062 through Nucleus
Millenium Stochastic Models of Complex and Disordered Systems, and by Fondecyt Grant 1160174.

R EFERENCES
[AS79] M. J. Ablowitz and H. Segur. On the evolution of packets of water waves. Journal of Fluid
Mechanics 92.4 (1979), 691âĂŞ715.
[AM05] M. Adler and P. van Moerbeke. PDEs for the joint distributions of the Dyson, Airy and sine
processes. Ann. Probab. 33.4 (2005), pp. 1326–1361.
[ACQ11] G. Amir, I. Corwin, and J. Quastel. Probability distribution of the free energy of the
continuum directed random polymer in 1 + 1 dimensions. Comm. Pure Appl. Math. 64.4
(2011), pp. 466–537.
[BBP05] J. Baik, G. Ben Arous, and S. Péché. Phase transition of the largest eigenvalue for nonnull
complex sample covariance matrices. Ann. Probab. 33.5 (2005), pp. 1643–1697.
[BL19] J. Baik and Z. Liu. Multipoint distribution of periodic TASEP. J. Amer. Math. Soc. 32.3
(2019), pp. 609–674.
[BCF14] A. Borodin, I. Corwin, and P. Ferrari. Free energy fluctuations for directed polymers in
random media in 1 + 1 dimension. Comm. Pure Appl. Math. 67.7 (2014), pp. 1129–1214.
[BFPS07] A. Borodin, P. L. Ferrari, M. Prähofer, and T. Sasamoto. Fluctuation properties of the
TASEP with periodic initial configuration. J. Stat. Phys. 129.5-6 (2007), pp. 1055–1080.
[CDR10] P. Calabrese, P. L. Doussal, and A. Rosso. Free-energy distribution of the directed polymer
at high temperature. EPL (Europhysics Letters) 90.2 (2010), p. 20002.
[DOV19] D. Dauvergne, J. Ortmann, and B. Virág. The directed landscape. 2019. arXiv: 1812 .
00309[math.PR].
[Dot10] V. Dotsenko. Bethe ansatz derivation of the Tracy-Widom distribution for one-dimensional
directed polymers. EPL (Europhysics Letters) 90.2 (2010), p. 20003.
[Joh00] K. Johansson. Shape fluctuations and random matrices. Comm. Math. Phys. 209.2 (2000),
pp. 437–476.
[Joh03] K. Johansson. Discrete polynuclear growth and determinantal processes. Comm. Math. Phys.
242.1-2 (2003), pp. 277–329.
[KPZ86] M. Kardar, G. Parisi, and Y.-C. Zhang. Dynamical scaling of growing interfaces. Phys. Rev.
Lett. 56.9 (1986), pp. 889–892.
[KPV97] C. E. Kenig, G. Ponce, and L. Vega. Global solutions for the KdV equation with unbounded
data. J. Differential Equations 139.2 (1997), pp. 339–364.
[LDC12] P. Le Doussal and P. Calabrese. The KPZ equation with flat initial condition and the directed
polymer with one free end. J. Stat. Mech. 2012.06 (2012), P06001.
[MQR+] K. Matetski, J. Quastel, and D. Remenik. In preparation.
[MQR17] K. Matetski, J. Quastel, and D. Remenik. The KPZ fixed point. 2017. arXiv: 1701 .
00018v1[math.PR].
[MJD00] T. Miwa, M. Jimbo, and E. Date. Solitons. Vol. 135. Cambridge Tracts in Mathematics.
Differential equations, symmetries and infinite-dimensional algebras, Translated from the
1993 Japanese original by Miles Reid. Cambridge University Press, Cambridge, 2000,
pp. x+108.
[NQR19] M. Nica, J. Quastel, and D. Remenik. Solution of the Kolmogorov equation for TASEP.
2019. arXiv: 1606.09228[math.PR].
KP(Z) 18

[OQR17] J. Ortmann, J. Quastel, and D. Remenik. A Pfaffian representation for flat ASEP. Comm.
Pure Appl. Math. 70.1 (2017), pp. 3–89.
[Po89] C. Pöppe. General determinants and the τ function for the Kadomtsev-Petviashvili hierarchy.
Inverse Problems 5.4 (1989), pp. 613–630.
[Pro19] S. Prolhac. Riemann surfaces for KPZ with periodic boundaries. 2019. arXiv: 1908 .
08907[cond-mat.stat-mech].
[Sas05] T. Sasamoto. Spatial correlations of the 1D KPZ surface on a flat substrate. Journal of
Physics A: Mathematical and General 38.33 (2005), p. L549.
[SS10] T. Sasamoto and H. Spohn. Exact height distributions for the KPZ equation with narrow
wedge initial condition. Nuclear Phys. B 834.3 (2010), pp. 523–542.
[TW03] C. A. Tracy and H. Widom. A system of differential equations for the Airy process. Electron.
Comm. Probab. 8 (2003), pp. 93–98.
[ZS74] V. E. Zaharov and A. B. Šabat. A plan for integrating the nonlinear equations of mathematical
physics by the method of the inverse scattering problem. I. Funkcional. Anal. i Priložen. 8.3
(1974), pp. 43–53.
[ZS79] V. E. Zaharov and A. B. Šabat. Integration of the nonlinear equations of mathematical
physics by the method of the inverse scattering problem. II. Funktsional. Anal. i Prilozhen.
13.3 (1979), pp. 13–22.

(J. Quastel) D EPARTMENT OF M ATHEMATICS , U NIVERSITY OF T ORONTO , 40 S T. G EORGE S TREET , T ORONTO ,


O NTARIO , C ANADA M5S 2E4
E-mail address: quastel@math.toronto.edu

(D. Remenik) D EPARTAMENTO DE I NGENIERÍA M ATEMÁTICA AND C ENTRO DE M ODELAMIENTO M ATEMÁTICO


(UMI-CNRS 2807), U NIVERSIDAD DE C HILE , AV. B EAUCHEF 851, T ORRE N ORTE , P ISO 5, S ANTIAGO , C HILE
E-mail address: dremenik@dim.uchile.cl

S-ar putea să vă placă și