You are on page 1of 110

LOCAL STATISTICS AND CONCENTRATION FOR NON-INTERSECTING

BROWNIAN BRIDGES WITH SMOOTH BOUNDARY DATA

AMOL AGGARWAL AND JIAOYANG HUANG


arXiv:2308.04318v1 [math.PR] 8 Aug 2023

Abstract. In this paper we consider non-intersecting Brownian bridges, under fairly general
upper and lower boundaries, and starting and ending data. Under the assumption that these
boundary data induce a smooth limit shape (without empty facets), we establish two results.
The first is a nearly optimal concentration bound for the Brownian bridges in this model. The
second is that the bulk local statistics of these bridges along any fixed time converge to the sine
process.

Contents
1. Introduction 1
2. Non-Intersecting Bridges and Dyson Brownian Motion 9
3. Limit Shapes for Non-intersecting Brownian Bridges 16
4. Proof of the Concentration Estimate 26
5. The Local Behaviors of Height Profiles 31
6. Proof of Concentration Bounds 39
7. Concentrating Functions With Given Local Profiles 47
8. Universality of Bulk Statistics 54
9. Regularity Properties of Solutions 65
Appendix A. Concentration Bounds for the Height Function 74
Appendix B. Proofs of Results From Section 2 76
Appendix C. Proofs of Results From Section 3 90
Appendix D. Proofs of Results From Section 8 93
Appendix E. Proofs of Results From Section 9 100
References 108

1. Introduction
1.1. Preface. Over the past several decades, a substantial amount of effort has been devoted to
the study of non-intersecting Brownian bridges. Since the works of Dyson [22] in 1962 and de
Gennes [32] in 1968, this model has been found to be ubiquitous across probability theory and
mathematical physics, with applications including to the spectra of random matrices [22] and their
universality (see the book [25] of Erdős–Yau); directed fibrous structures [32]; random surfaces (see
the works of Fisher [29] and Huse–Fisher [43]); stochastic growth models (see the surveys of Spohn
[60], Johansson [46], and Ferrari–Spohn [26]); and certain quantum many-body systems (see the
survey of Dean–Le Doussal–Majumdar–Schehr [20]).
1
2 AMOL AGGARWAL AND JIAOYANG HUANG

The general form of this model can be described as follows. Let n ≥ 1 be an integer, and fix two
n-tuples of starting data u = (u1 , u2 , . . . , un ) and ending data v = (v1 , v2 , . . . , vn ); a time interval
[a, b]; a lower boundary f : [a, b] → R ∪ {−∞}; and an upper boundary g : [a, b] → R ∪ {∞}. Then
sample a family x = (x1 , x2 , . . . , xn ) of n Brownian bridges on [a, b], starting at u and ending at
v (so that xj (a) = uj and xj (b) = vj for each j), that are conditioned to not intersect, stay above
f , and stay below g (so that f (t) < xn (t) < xn−1 (t) < · · · < x1 (t) < g(t) for each t ∈ (a, b)); see
Figure 1. One is then interested in the asymptotic behavior of this model as n becomes large.
This question is most extensively understood in the absence of an upper and lower boundary,
that is, when f = −∞ and g = ∞. The first choice of starting and ending data to have been
analyzed is the Brownian watermelon, corresponding to when all entries of u and v are equal to
0. In this case, the joint law of the Brownian paths at any fixed time coincides with that of the
spectrum of a Gaussian Unitary Ensemble (GUE) random matrix [22]. The latter has long been
known to be exactly solvable through the framework determinatal point processes and orthogonal
polynomials; its bulk local statistics converge to the sine processes, due to the works of Gaudin and
Mehta [56, 31]. The bulk local statistics, jointly over multiple times, of the Brownian watermelon
was shown by Prahöfer–Spohn [59] and Johansson [45] to converge to the extended sine process.
The Brownian bridge model has since been proven to be amenable to the framework of deter-
minantal point processes (though often requiring a considerably more involved analysis) for other
special choices of starting and ending data (u; v), when both consist of a bounded number (often
at most two) distinct points. In many of these situations, different limiting statistics were found
appear at the edge, such as Airy processes [59, 45] with wanderers by Adler–Ferrari–van Moerbeke
[1]; Pearcey processes by Brézin–Hikami [12], Tracy–Widom [62], and Bleher–Kuijlaars [9]; and
tacnode processes by Delvaux–Kuijlaars–Zhang [21], Ferrari–Vető [27], and Johansson [47].
Under general starting and ending data (but still with no upper and lower boundary), the physics
work of Matytsin [55] predicted a limit shape for non-intersecting Brownian bridges, in the form of a
complex Burgers equation. This was proven in the papers of Guionnet–Zeitouni [36] and Guionnet
[35] through a large deviations principle for spherical integrals, obtained by combining a stochastic
analysis of Dyson Brownian motion with ideas from free probability. More recent work [40] of
Huang proved (under additional, mild restrictions on the starting and ending data) that the edge
statistics of this model converge to the Airy point process.
The behavior of non-intersecting Brownian bridges is significantly less understood in the presence
of (finite) upper and lower boundaries, as in this situation the exact connection to random matrices,
Dyson Brownian motion, and free probability appears to be lost. The determinantal structure
remains, but is in general quite intricate. Only for special choices of the upper and lower boundary
(f ; g) (and starting and ending data (u; v)) has this structure proven to be tractable for asymptotics.
Three papers in this direction are those of Tracy–Widom [63], Ferrari–Vető [28], and Leichty–
Wang [53], which observed different statistics arising in the presence of upper and lower boundaries.
The first [63] considered non-intersecting Brownian bridges conditioned to be positive, correspond-
ing to the choice f = 0 and g = ∞, that start and end at 0; it then showed that the local limit of the
bottom curve converges to a Bessel process with parameter 21 . The second [28] considered the same
model, but where the paths started and ended at the same nonzero point; it then showed that the
local limit of the bottom curve is given by the hard edge tacnode process. The third [53] considered
non-intersecting Brownian bridges, conditioned to stay above f = 0 and below g = 1, that start
and end at 0; it then showed that the limiting statistics at the point where these bridges first meet
the upper boundary g = 1 are given by certain variants of the Pearcey and tacnode processes.
LOCAL STATISTICS AND CONCENTRATION FOR NON-INTERSECTING BROWNIAN BRIDGES 3

In this paper we study non-intersecting Brownian bridges under a fairly wide class of boundary
conditions (subject to a regularity constraint to be described below), with the intent of proving
various universality phenomena about their statistical behavior. Our results are two-fold; the first
is a concentration bound on the Brownian paths, and the second analyzes their bulk local statistics.
To be more specific, we fix as above a time interval, normalized to be [0, L−1 ]; upper and lower
boundaries (f ; g) on [0, L−1 ]; and starting and ending data (u; v). Let G : [0, L−1 ] × [0, 1] → R
denote the solution to the equation
(1.1) ∂t2 G(t, x) + π 2 (∂y G)−4 · ∂y2 G(t, x) = 0,
with boundary data along the south, north, west, and east sides of the rectangle [0, L−1 ] × [0, 1]
prescribed by f , g, u, and v, respectively. This function G will serve as the limit shape for the
non-intersecting Brownian bridges, with t and y serving as the time coordinate and normalized
index of the Brownian bridges (so that G(t, y) ≈ x⌊yn⌋ (t)). While we have not seen (1.1) previously
appear exactly as written for this model, we explain in Section 3 below that it is equivalent (after a
change of variables) to equations prescribing the limit shape for non-intersecting Brownian bridges
derived in the above-mentioned works [55, 36, 35], in the case of no upper and lower boundary.
Given an integer m ≥ 1, the regularity constraint (see Assumption 1.3 below) we will impose on
the boundary data (f ; g; u; v) stipulates that all of the first m + 1 derivatives of G are bounded
from above, and that |∂y G| is bounded from below. The latter ensures that the equation (1.1) is
uniformly elliptic; it also precludes the existence of macroscopic “empty” facets, through which no
Brownian paths likely pass (such regions are, in a sense, “compressed out” by the lower and upper
boundaries f and g). As such, these boundary data are qualitatively different from many of those
considered in the previous works.
Under this setup, our first result (see Theorem 1.4 below) states that xj (t) = G(t, jn−1 ) +
O(n2/m−1+o(1) ). In particular, if G is smooth (enabling m to be arbitrarily large), then this implies
that the Brownian paths xj concentrate around G with the nearly optimal error O(no(1)−1 ). The
only concentration bound for non-intersecting Brownian bridges, applicable to general upper and
lower boundary, we knew of prior to this work is one for its normalized height function, which has
error O(no(1)−1/2 ) (and does not provide any characterization for its expectation).1
−1
Our second result  (see Theorem 1.5 below)  states that, for a given (t0 , y20 ) ∈ (0, 1) × (0, L ),
the point process n · xj (t0 ) − G(t0 , y0 ) converges to the sine process. The condition that
(t0 , y0 ) ∈ (0, 1) × (0, L−1 ) is in the strict interior of the rectangle [0, 1] × [0, L−1 ] corresponds to
probing the “bulk” statistics of the Brownian bridges xj . For β = 2 Dyson Brownian motion (which
is equivalent after changing coordinates to the special case of the Brownian bridge model without
upper and lower boundary, with arbitrary starting data u and all entries of the ending data v
equal to 0), such bulk universality results have appeared in several forms over the past twenty-
five years; see the works of Brézin–Hikami [11], Johansson [44], Erdős–Péché–Ramı́rez–Schlein–Yau
[23], Landon–Yau [50], Erdős–Schnelli [24], Landon–Sosoe–Yau [49], and Claeys–Neuschel–Venker
[15]. However, we are unaware of any previous such universality statements for non-intersecting
Brownian bridges in the presence of upper and lower boundaries (and with general ending data).

1Starting with the work of Cohn–Elkies–Propp [16, Proposition 22], such concentration bounds are well-
documented in the context of discrete random surfaces, such as dimers. While we have not seen the counterpart
for Brownian bridges appear in the literature before, the proof is very similar. Although we will not need it in this
paper, we provide the short proof of this estimate in Appendix A below, for completeness.
2Our method should also apply to give convergence to the extended sine process for the joint correlation functions
of this point process across multiple times, but we will not address this here.
4 AMOL AGGARWAL AND JIAOYANG HUANG

Probing the “edge” statistics of the model instead corresponds to taking t ∈ {0, 1}, which
amounts to analyzing the Brownian bridges around the upper or lower boundaries. If f and g
are sufficiently regular (which is implied by our assumption), then locally they are approximately
linear, and so we would expect the xj to look (up to an affine shift) locally like non-intersecting
Brownian excursions. In view of the results in [63], we then predict that the edge scaling limit of
the extreme path is a Bessel process with parameter 21 . We will not pursue this question here.
Next let us briefly describe the proofs underlying our results. Given the concentration bound,
we establish universality for the local statistics of the non-intersecting Brownian bridges x by
sandwiching them between two Dyson Brownian motion models w− and w+ with slightly different
drifts (and almost the same initial data). The concentration bound not only enables the existence of
this sandwiching, but it can also be used to show that the initial data for w− and w+ are reasonably
regular. The latter point, together with the results of [50, 49] proving fast local equilibration of
Dyson Brownian motion under regular initial data, indicates that the local statistics for w− and w+
models quickly converge to (the same) sine process; by the sandwiching, this means that those for
x do as well. This comparison is analogous to (and, in a sense, simpler than) similar ones recently
used to establish universality results in the context of random tilings and non-intersecting Brownian
bridges [2, 40]. So, to us, the main novelty in this paper is in the proof of the concentration bound.
To do this, instead of directly studying the trajectories xj (t)  of the Brownian paths, it will be
convenient to analyze the associated height function Hx (t, x) = # 1 ≤ j ≤ n : xj (t) > x , counting
how many bridges are above x at time t; this enables us to interpret the Brownian bridge model x
as a random surface. The concentration bound on the xj (t) then becomes equivalent to one for this
height function, namely, n−1 · Hx (t, x) = H(t, x) + O(n2/m−1+o(1) ), where H(t, x) is a continuum
height function (which, for fixed t, is essentially the functional inverse of G(t, y)).
To that end, we will define two families of “barrier functions” φ− +
i and φi to bound n
−1
· Hx
3
from below and above. These functions will be chosen so that they satisfy several properties. First,
at i = 0, φ− + −
0 and φ0 will be small and large, respectively, guaranteeing φ0 ≤ H (t, x) ≤ φ0 to
x +
+ 2/m−1+o(1) −
hold deterministically. Second, the φi (t, x) will decrease to H(t, x) + n , and the φi will
increase to H(t, x) − n2/m−1+o(1) , as i increases from 0 to some integer I ≥ 1. We will then show
that n−1 ·Hx ≤ φ+ i−1 likely implies n
−1
·Hx ≤ φ+ i (and n
−1
·Hx ≥ φ− i−1 likely implies n
−1
·Hx ≥ φ− i ),
for each 1 ≤ i ≤ I. Inductively applying this estimate will then confirm the concentration bound,
as then H(t, x) + n2/m−1+o(1) = φ− I ≤n
−1
· Hx ≤ φ+ I = H(t, x) + n
2/m−1+o(1)
would likely hold.
The proof of the inductive implication stated above makes use of the following third fundamental

property satisfied by the φ+ +
i and φi . For any point (t0 , x0 ), the function φi is more concave at
(t0 , x0 ) than is any continuum height function H0 , whose gradient at (t0 , x0 ) matches that of φ+ i
(any such H0 will have a constrained Hessian, since it satisfies an elliptic partial differential equation
similar to the one (1.1) satisfied by G; see Lemma 3.16 below); an analogous statement holds for
φ−i . Now suppose that n
−1
· Hx is (at most) equal to φ+ i−1 . Then condition on the paths of x
outside of some neighborhood U of a fixed point (t0 , x0 ), and resample the paths of x inside U as
non-intersecting Brownian bridges with boundary data induced by the conditioning. Assuming a
sufficiently strong concentration bound for the Brownian bridge height function Hx on U around
its continuum limit, the above concavity property will imply that Hx (t0 , x0 ) likely “deflates” upon
this resampling, decreasing below φ+ i (t0 , x0 ), thus establishing the inductive step.

3In the actual proof, these barrier functions will instead be indexed by a pair of integers (j; k), but we will ignore
this point in the introductory exposition here.
LOCAL STATISTICS AND CONCENTRATION FOR NON-INTERSECTING BROWNIAN BRIDGES 5

Verifying this last assumption will amount to finding a “large” class of boundary data, for which
the associated non-intersecting Brownian bridges satisfy a nearly optimal O(no(1)−1 ) concentration
bound around their continuum height function; if this holds, we refer to the continuum height
function as a concentrating function. In practice, “large” means that any possible m-th order
Taylor expansion at a fixed point of some continuum height function, which we refer to as a local
profile, must be encompassed by at least one concentrating function in our family.
The above methodology was initially developed Laslier–Toninelli [52, 51], in the seemingly unre-
lated context of analyzing Glauber dynamics on random tilings. Their sequences {φ− +
i } and {φi }
were meant to serve as discrete-time, and more tractable, proxies for the mean curvature flow that
is widely predicted (dating back to the physics paper [54] of Lifshitz) to govern the macroscopic
behavior of random surfaces under reversible stochastic dynamics. These ideas were used to show
concentration estimates for random tilings in the paper [2] of Aggarwal, but the bounds there were
quite weak, involving an error of O(n−α ) for some small constant α > 0. The reason is that work
took m = 2, that is, it only produced concentrating functions (coming from limit shapes of random
tilings of hexagons, as in [52]) matching the first two derivatives of any continuum height function.
In this paper, we must take m to be arbitrary and introduce concentrating functions encom-
passing any local profile of order m. We find them as the continuum height functions for β = 2
Dyson Brownian motion with arbitrary initial data; nearly optimal concentration bounds for the
latter were proven by Huang–Landon in [41]. By the generality of the initial data, this provides an
infinite-dimensional family of concentrating functions. We show as Corollary 5.11 (see also Propo-
sition 7.1) below that they encompass any local profile of order m, thus enabling us to implement
the above framework to prove the O(n2/m−1+o(1) ) concentration bound for the non-intersecting
Brownian bridges x.
1.2. Non-Intersecting Brownian Bridges and the Sine Kernel. In this section we discuss
non-intersecting Brownian bridges and the sine kernel. In what follows, for any integer d ≥ 1
and subset I ⊆ Rd , we let C(I) denote the space of real-valued, continuous functions f : I → R.
Moreover, for any subset I0 ⊆ I and measurable function f : I → C, let f |I0 denote the restriction
of f to I0 . Let S ⊆ Z≥1 and I ⊆ R denote intervals. We may identify the set S × C(I) with families
of continuous functions x = (xj )j∈S , where xj : I → R is a continuous function for each j ∈ S; we
then also set x(t) = xj (t) j∈S for each t ∈ I. We will frequently use this notation for families of
(non-intersecting) Brownian bridges.
We next provide notation for the probability measure of n non-intersecting Brownian bridges
with given starting and ending points, and for given upper and lower boundaries. In what follows,
we set R = R ∪ {−∞, ∞}; we also let H = {z ∈ C : Im z > 0} denote the upper half-plane and H
denote its closure. For any real numbers a, b ∈ R with a < b, we set Ja, bK = [a, b] ∩ Z. For any
integer k ≥ 1, we denote the entries of any k-tuple y ∈ Rk by y = (y1 , y2 , . . . , yk ), unless stated
otherwise. We also let Wk = {y ∈ Rk : y1 > y2 > · · · > yk } and let Wk denote its closure. For any
integer k ≥ 1 and k-tuples x = (x1 , x2 , . . . , xk ) ∈ Rk and y = (y1 , y2 , . . . , yk ) ∈ Rk , we write x < y
(equivalently, y > x) if xj < yj for each j ∈ J1, kK; we similarly write x ≤ y (equivalently, y ≥ x)
if xj ≤ yj for each j ∈ J1, kK. For any subset I ⊆ R and measurable functions f, g : I → R we write
f < g (equivalently, g > f ) if f (t) < g(t) for each t ∈ I; we similarly write f ≤ g (equivalently,
g ≥ f ) if f (t) ≤ g(t) for each t ∈ I.
Definition 1.1. Fix an integer n ≥ 1; a real number σ > 0; two n-tuples u, v ∈ Wn ; an interval
[a, b] ⊂ R; and measurable functions f, g : [a, b] → R such that f < g, f < ∞, and g > −∞. Let
Qu;v
f ;g (σ) denote the law of x = (x1 , x2 , . . . , xn ) ∈ J1, nK×C [a, b] , given by n independent Brownian
6 AMOL AGGARWAL AND JIAOYANG HUANG

Figure 1. An example of non-intersecting Brownian bridges sampled from


Qu;v
f ;g (σ) is shown above.

motions of variances σ on the time interval t ∈ [a, b], conditioned on satisfying the following three
properties.
(1) The xj do not intersect, that is, x(t) ∈ Wn for each t ∈ (a, b).
(2) The xj start at uj and end at vj , that is, xj (a) = uj and xj (b) = vj for each j ∈ J1, nK.
(3) The xj are bounded below by f and above by g, that is, f < xj < g for each j ∈ J1, nK.
We assume f (a) ≤ un ≤ u1 ≤ g(a) and f (b) ≤ vn ≤ v1 ≤ g(b), even when not stated explicitly;
see Figure 1 for a depiction. We refer to u as starting data for x, and to v as its ending data.
We also refer to f as the lower boundary for x, and to g as its upper boundary. If g = ∞, then
we abbreviate Qu;v u;v
f (σ) = Qf ;∞ (σ); if additionally f = −∞, we further abbreviate Q
u;v
(σ) =
u;v u;v −1
Q−∞ (σ) = Q−∞;∞ (σ). If σ = n , then we omit the parameter σ from the notation, writing
Qu;v u;v
f ;g = Qf ;g (n
−1
), Qu;v
f = Qu;v
f (n
−1
), and Qu;v = Qu;v (n−1 ).
We will show in this paper that (under certain conditions on u, v, f , and g) the local statistics
of x(t) will converge to a point process described by the sine kernel, defined as follows.
Definition 1.2 ([6, Equation (3.6.1)]). We define the sine kernel Ksin : R2 → R by setting

sin π(x − y)
Ksin (x, y) = , for any x, y ∈ R.
π(x − y)
For any integer m ≥ 1 and m-tuple x = (x1 , x2 , . . . , xm ) ∈ Rm , define the m × m matrix Ksin (x)
(m)
whose (i, j) entry is Ksin (xi , xj ). Also define the m-point correlation function psin : Rm → R by
(m)
psin (x) = det Ksin (x), for any x ∈ Rm .
LOCAL STATISTICS AND CONCENTRATION FOR NON-INTERSECTING BROWNIAN BRIDGES 7

1.3. Results. In this section we state two results on the large n behavior of families of non-
intersecting Brownian bridges sampled from the measure Q of Definition 1.1; the first is a concen-
tration bound, and the second analyzes convergence of its local statistics. We begin by specifying
the assumptions to which we will subject our boundary data, namely, the starting and ending data
(u and v) and the lower and upper boundaries (f and g). To that end, we require some notation.
For any set S ∈ Cd and point z ∈ Cd , we set dist(z, S) = inf w∈S |z − w|. We also let ∂S
Qd
denote the boundary of S. We further denote the derivative ∂γ = i=1 (∂xi )γi for any d-tuple
Pd
(γ1 , γ2 , . . . , γd ) ∈ Zd≥0 ; we also set |γ| = j=1 γj . For any open subset R ⊆ Rd , let C k (R) denote
the set of f ∈ C(R) such that ∂γ f ∈ C(R), for each γ ∈ Zd≥0 with |γ| ≤ k. Further let C k (R) denote
the set of functions f ∈ C k (R) such that ∂γ f ∈ C(R) extends continuously to R, for each γ ∈ Zd≥0
with |γ| ≤ k. For any function f ∈ C(R) and integer k ∈ Z≥0 , we further define the (semi)norms
∥f ∥0 = ∥f ∥0;R , [f ] = [f ]k;0;R , and ∥f ∥C k (R) = ∥f ∥C k (R) = ∥f ∥k = ∥f ∥k;0;R on these spaces by
k
X

∥f ∥0 = sup f (z) ; [f ]k = max ∂γ f ∥0 ; ∥f ∥C k (R) = [f ]j .
z∈R γ∈Zd≥0 j=0
|γ|=k

For any bounded, open set R ⊂ R2 , let Adm(R) be the set of locally Lipschitz functions F :
R → R such that ∂y F (z) < 0 for almost all (with respect to Lebesgue measure) z ∈ R. For any
real number ε > 0, further let Admε (R) ⊆ Adm(R) denote the set of F ∈ Adm(R) such that
−ε−1 < ∂y F (z) < −ε, for almost all z ∈ R (with respect to Lebesgue measure).
We can now state the following assumption.
Assumption 1.3. Fix an integer m ≥ 4 and three real numbers ε ∈ 0, 21 , δ ∈ 0, 5m 1
 
2 , and

B0 > 1. Let n ≥ 1 be an integer and L ∈ (B0−1 , nδ ) be a real number; define the open rectangle
R = (0, L−1 ) × (0, 1) ⊂ R2 . Let G ∈ Admε (R) ∩ C m+1 (R) be such that G − G(0, 0) C m+1 (R) ≤ B0 ,
and such that it solves the equation
X 
(1.2) djk ∇G(t, y) ∂j ∂k G(t, y) = 0, for each (t, y) ∈ R,
j,k∈{t,y}

where the djk : R2 → R are defined by


(1.3) dtt (u, v) = 1; dty (u, v) = 0 = dyt (u, v); dyy (u, v) = π 2 v −4 .
Define f, g : [0, L−1 ] → R by setting f (s) = G(s, 1) and g(s) = G(s, 0), for each s ∈ [0, L−1 ].
Further let κ > 0 be a real number, and let u, v ∈ Wn be n-tuples with
max uj − G(0, jn−1 ) ≤ κ; max vj − G(L−1 , jn−1 ) ≤ κ.

(1.4)
j∈J1,nK j∈J1,nK

Sample non-intersecting Brownian bridges x = (x1 , x2 , . . . , xn ) ∈ J1, nK × C [0, L−1 ] from the


measure Qu;v
f ;g .

Let us briefly explain Assumption 1.3. The function G will eventually be the “limit shape” for
the family x of non-intersecting Brownian bridges, in the sense that we will have xj (t) ≈ G(t, jn−1 );
see Theorem 1.4 below. The conditions that f (s) = G(s, 1) and g(s) = G(s, 0) ensure that this holds
for the upper and lower boundaries of the model (formally, when j ∈ {0, n + 1}), and (1.4) ensures
that this holds (up to an error of κ) when t ∈ {0, L−1 }. Similarly to how limit shapes for random
tilings satisfy a nonlinear partial differential equation [17, Theorem 12.1], the function G here does
as well, given by (1.2). The constraint that G ∈ Admε (R) ensures that G has no “frozen facets”
8 AMOL AGGARWAL AND JIAOYANG HUANG


(macroscopic regions containing no curves), and the constraint that G − G(0, 0) C m+1 (R) ≤ B0
ensures that G has some regularity.
The below result is a concentration bound stating that, with high probability, xj (t) = G(t, jn−1 )+
O(nδ+2/m−1 + κ) (so that the error can be made smaller by increasing the parameter m that ac-
counts for the regularity of the boundary data); it will be established in Section 4.1 below.
Theorem 1.4. Adopt Assumption 1.3. There exists a constant c = c(ε, δ, B0 , m) > 0 such that
"   #
2
−1 −1 2/m+δ−1
≤ c−1 e−c(log n) .

P sup max xj (s) − G(s, jn ) > κ + c n

s∈[0,L−1 ] j∈J1,nK

The next result4 states that the local statistics of x(t0 ) at some t0 ∈ (0, L−1 ) are prescribed
by the sine kernel; it will be established in Section 8.1 below. In what follows, for any integers
n ≥ 1 and k ∈ J1, nK, and random n-tuple y ∈ Wn admitting a density with respect to Lebesgue
(k)
measure on Rn , define its k-point correlation function p(k) = py : Rk → R as follows. Letting
(n)
p(n) = py : Rn → R denote the symmetrized density of y, for any (x1 , x2 , . . . , xk ) ∈ Rk , set
Z n
Y
(1.5) p(k) (x1 , x2 , . . . , xk ) = p(n) (x1 , x2 , . . . , xn ) dxj .
Rn−k j=k+1

In what follows, for any integer k ≥ 1, complex numbers a, b ∈ R, vector x = (x1 , x2 , . . . , xk ) ∈


Ck , and set S ⊆ C, define a · x + b = (ax1 + b, ax2 + b, . . . , axn + b) and a · S + b = {as + b}s∈S .
For any additional subset S ′ ⊆ C, we denote S + S ′ = {s + s′ : s ∈ S, s′ ∈ S ′ }.
Theorem 1.5. Adopt Assumption 1.3; further suppose that m ≥ 225 and that κ ≤ nδ−1 . Fix an
integer k ≥ 1; a smooth  compactly supported function F : Rk → R; and a point (t0 , y0 ) ∈ R, such
−δ
that dist (t0 , y0 ), ∂R ≥ n  . Denote θ = −∂y G(t0 , y0 ), and set z = (z1 , z2 , . . . , zn ) ∈ Wn by
zj = θn · xj (t0 ) − G(t0 , y0 ) for each j ∈ J1, nK. Then,
Z Z
(k) (k)
lim F (a)pz (a)da = F (a)psin (a)da.
n→∞ Rk Rk

1.4. Outline. The remainder of this paper is organized as follows. In Section 2 we collect mis-
cellaneous (known) facts about non-intersecting Brownian bridges and Dyson Brownian motion,
including invariances, Hölder regularity bounds, and concentration estimates. In Section 3 we ex-
plain how to recover the equation (1.2), in the absence of upper and lower boundaries, from the
results in [36, 35]. In Section 4 we reduce the concentration bound Theorem 1.4 to the inductive
step, Proposition 4.4, explained in Section 1.1. In Section 5 we discuss local profiles and state
Corollary 5.11, providing a large class of concentrating functions. Using this, we establish Proposi-
tion 4.4 in Section 6. We next prove Corollary 5.11 in Section 7 and Theorem 1.5 in Section 8. We
conclude by discussing various regularity properties of the equation (1.2) in Section 9. While these
regularity results are not directly used in the proofs of Theorem 1.4 and Theorem 1.5, some (such as
Lemma 9.4 and Lemma 9.6) attest to the robustness of Assumption 1.3 under perturbations (and
others will also be used in the forthcoming work [4]). The Appendix proves various statements from
previous parts of the paper, and it largely constitutes of known (or mild variants of known) results.

4Let us clarify that, in Theorem 1.5, we view m, ε, δ, and B as fixed with n, while the other parameters L, G, u,
0
v, κ, t0 , and y0 can vary arbitrarily as n tends to ∞ (subject to the conditions of Assumption 1.3 and Theorem 1.5).
LOCAL STATISTICS AND CONCENTRATION FOR NON-INTERSECTING BROWNIAN BRIDGES 9

Acknowledgements. Amol Aggarwal was partially supported by a Packard Fellowship for Science
and Engineering, a Clay Research Fellowship, by NSF grant DMS-1926686, and by the IAS School
of Mathematics. The research of Jiaoyang Huang is supported by NSF grant DMS-2054835 and
DMS-2331096. The authors also thank Fabio Toninelli and Horng-Tzer Yau for helpful interactions
on non-intersecting Brownian bridges, and Camillo De Lellis, László Székelyhidi Jr., and Vladimı́r
Sverák for useful comments on elliptic partial differential equations. They wish to acknowledge the
NSF grant DMS-1928930, which supported their participation in the Fall 2021 semester program
at MSRI in Berkeley, California titled, “Universality and Integrability in Random Matrix Theory
and Interacting Particle Systems.

2. Non-Intersecting Bridges and Dyson Brownian Motion


In this section we collect several miscellaneous facts about non-intersecting Brownian bridges
and Dyson Brownian motion.
2.1. Height Monotonicity, Hölder Estimates, and Invariances. In this section we state cer-
tain invariances, monotone couplings, and Hölder estimates for non-intersecting Brownian bridges.
We first review two transformations that leave non-intersecting Brownian bridge measures invariant;
the first constitutes affine transformations, and the second constitutes diffusive scalings.
Remark 2.1. Nonintersecting Brownian bridges satisfy the following invariance property under affine
transformations. Adopt the notation of Definition 1.1, and fix real numbers α, β ∈ R. Define the
n-tuples u e ∈ Wn and functions fe, ge : [a, b] → R by setting
e, v
u
ej = uj + α, and vej = vj + (b − a)β + α, for each j ∈ J0, nK;
fe(t) = f (t) + (t − a)β + α, and ge(t) = g(t) + (t − a)β + α, for each t ∈ [a, b].
Sampling x e = (ex1 , x en ) under Qfũ;ṽ
e2 , . . . , x ˜,g̃ , there is a coupling between x
e and x such that x
ej (t) =
xj (t) + (t − a)β + α for each (j, t) ∈ J1, nK × [a, b].
Indeed, this follows from the analogous affine invariance of a single Brownian random bridge,
together with the fact that affine transformations do not affect the non-intersecting property. More
specifically, if x(t), for t ∈ [a, b], is a Brownian bridge from some u ∈ R to some v ∈ R then
x(t) + (t − a)β + α is a Brownian bridge from u + α to v + (b − a)β + α, and any y(t) ∈ Wn (namely,
is non-intersecting) if and only if y(t) + (t − a)β + α ∈ Wn .
Remark 2.2. Nonintersecting Brownian bridges also satisfy the following invariance property under
diffusive scaling. Again adopt the notation of Definition 1.1; assume that (a, b) = (0, T ), for some
real number T > 0. Further fix a real number σ > 0, and set Te = σT . Define the n-tuples
u e ∈ Wn and functions fe, ge : [0, Te] → R by setting
e, v
ej = σ 1/2 uj ,
u and vej = σ 1/2 vj , for each j ∈ J0, nK;
1/2 −1 1/2 −1
fe(t) = σ f (σ t), and ge(t) = σ g(σ t), for each t ∈ [0, Te].
Sampling x e = (e
x1 , x en ) under Qfũ;ṽ
e2 , . . . , x ˜,g̃ , there is a coupling between x
e and x such that x
ej (t) =
1/2 −1
σ xj (σ t) for each (j, t) ∈ J1, nK×[0, Te]. Similarly to Remark 2.1, this follows from the analogous
scaling invariance of a single Brownian bridge.
The next lemma recalls a monotone coupling for non-intersecting Brownian bridges that was
shown in [18]; we refer to it as height monotonicity.
10 AMOL AGGARWAL AND JIAOYANG HUANG

Lemma 2.3 ([18, Lemmas 2.6 and 2.7]). Fix an integer n ≥ 1; four n-tuples u, u e ∈ Wn ;
e , v, v
an interval [a, b] ∈ R; and measurable functions f, fe, g, ge : [a, b] → R with f ≤ g and fe ≤ ge.
Sample two families of non-intersecting Brownian bridges x = (x1 , x2 , . . . , xn ) ∈ J1, nK × C [a, b]
en ) ∈ J1, nK × C [a, b] from the measures Qu;v ũ;ṽ

and x
e = (e
x1 , x
e2 , . . . , x f ;g and Qf˜;g̃ , respectively. If

(2.1) f ≤ fe; g ≤ ge; u≤u


e; v≤v
e,
e so that xj (t) ≤ x
then there exists a coupling between x and x ej (t), for each (j, t) ∈ J1, nK × [a, b].
We next state the following variant of the above coupling, whose second part provides a linear
bound on the difference between two families of non-intersecting Brownian bridges, which have the
same starting data but different ending data. Its proof will be given in Appendix B.3 below.
Lemma 2.4. Fix an integer n ≥ 1; a real number B ≥ 0; a finite interval [a, b] ⊂ R; four n-tuples
u, u e ∈ Wn ; and four measurable functions f, fe, g, ge : [a, b] → R. Assume for each j ∈ J1, nK
e , v, v
and t ∈ [a, b] that
(2.2) uj ≤ u
ej ≤ uj + B; vj ≤ vej ≤ vj + B; f (t) ≤ fe(t) ≤ f (t) + B; g(t) ≤ ge(t) ≤ g(t) + B.

Sample two families of non-intersecting Brownian bridges x = (x1 , x2 , . . . , xn ) ∈ J1, nK × C [a, b]
en ) ∈ J1, nK × C [a, b] from the measures Qu;v ũ;ṽ

and x
e = (e
x1 , x
e2 , . . . , x f ;g and Qf˜;g̃ , respectively.

(1) There is a coupling between x and xe so that xj (t) ≤ x ej (t) ≤ xj (t) + B for each (j, t) ∈
J1, nK × [a, b].
(2) Further assume that u = ue and for each t ∈ [a, b] that
t−a t−a
(2.3) f (t) ≤ fe(t) ≤ f (t) + · B; g(t) ≤ ge(t) ≤ g(t) + · B.
b−a b−a
Then, it is possible to couple x and x
e such that
t−a
xj (t) ≤ x
ej (t) ≤ xj (t) + · B, for each (j, t) ∈ J1, nK × [a, b].
b−a
Remark 2.5. Let us mention two quick consequences of Lemma 2.4 that we will often use without
comment. First, suppose instead of (2.2) that for each j ∈ J1, nK and t ∈ [a, b] we have

(2.4) |uj − u ej | ≤ B; |vj − vej | ≤ B; f (t) − fe(t) ≤ B; g(t) − ge(t) ≤ B.

Then, it is possible to couple x and x e such that xj (t) − x ej (t) ≤ B for each (j, t) ∈ J1, nK × [a, b].
Indeed, define the n-tuples u b ∈ Wn by setting u
b, v bj = u ej +B and vbj = vej +B, for each j ∈ J1, nK;
also define the functions fb, gb : [a, b] → R by setting fb(t) = fe(t) + B and gb(t) = ge(t) + B,  for each
t ∈ [a, b]. Then, sample n Brownian bridges x b = (b x1 , x bn ) ∈ J1, nK × C [a, b] from the
b2 , . . . , x
measure Qû;v̂
fˆ;ĝ
. Observe that we may couple x b with x e such that x bj (t) = xej (t) + B, for each (j, t).
Applying the first part of Lemma 2.4 with the (x; x e; B) there given by (x; x b; 2B) here (using
(2.4) to verify (2.2)) yields a coupling between x and x b such that for each
(j, t) we have xj (t) ≤
bj (t) ≤ xj (t) + 2B. Since x
x bj (t) = xej (t) + B, it follows that xj (t) − x ej (t) ≤ B for each (j, t).
Second, assume that (2.4) holds; that we have u = u e ; and that, instead of (2.3), we have
t−a t−a
sup f (t) − fe(t) ≤ · B; sup f (t) − fe(t) ≤ · B.
t∈[a,b] b−a t∈[a,b] b−a
LOCAL STATISTICS AND CONCENTRATION FOR NON-INTERSECTING BROWNIAN BRIDGES 11

Then, by similar reasoning as used above, it is possible to couple x and x e such that
t−a
xj (t) − x
ej (t) ≤ · B, for each (j, t) ∈ J1, nK × [a, b].
b−a
We next recall the following Hölder estimate for non-intersecting Brownian bridges from [19]
(scaled by a factor of n−1/2 , as the variances of the bridges there were 1 while here they are n−1 ).
Lemma 2.6 ([19, Proposition 3.5]). There exist constants c > 0 and C > 1 such that the following
holds. Let n ≥ 1 be an integer; B ≥ 1 be a real number; [a, b] ⊂  R be an interval; and u, v ∈ Wn be
two n-tuples. Sampling x = (x1 , x2 , . . . , xn ) ∈ J1, nK × C [a, b] under Qu;v , we have
" n  #
[ [ vj − uj 1/2
−1 2
≤ CeCn−cB n .

xj (t + s) − xj (t) − s · ≥ Bs log 2s (b − a)

P
j=1
b−a
a≤t<t+s≤b

We will also require the following variant of Lemma 2.6 that allows for lower and upper boundaries
f and g. Its second part provides a Hölder bound assuming these two boundaries are differentiable.
Without at least some continuity constraints on f and g, such a bound cannot hold.5 However, when
g = ∞ (in the absence of a continuity condition on f ), the first part of the below lemma provides
an “interior” Hölder bound on these bridges, that is, one that holds away from the boundaries
t ∈ {a, b} of the time interval. The proof of the following result is similar to that of Lemma 2.6 and
will be established in Appendix B.1 below.
Lemma 2.7. Let n, B, a, b, u, and v be as in Lemma 2.6. Further let A ≥ 1 be a real number  and
f, g : [a, b] → R be measurable functions. Sampling x = (x1 , x2 , . . . , xn ) ∈ J1, nK × C [a, b] under
Qu;v
f ;g , the following two statements hold. In what follows, we set T = b − a.
(1) Assume that g = ∞; that un − f (r) ≥ −AT 1/2 ; and that vn − f (r) ≥ −AT 1/2 , for each
r ∈ [a, b]. For any real number 0 < κ < min T2 , 1 , we have
" n 
[ [
xj (t + s) − xj (t) − sT −1 (vj − uj )

P
j=1 a+κ≤t<t+s≤b−κ
#
1/2 −1 −1
2 2
≥s B log |2s T| + κ (A + B) ≤ CeCn−cB n .

(2) If instead f and g are differentiable and ∂t f (r) + ∂t g(r) ≤ A, for each r ∈ [a, b], then
" n 
[ [
xj (t + s) − xj (t) − sT −1 (vj − uj )

P
j=1 a≤t<t+s≤b
#
1/2 −1
2 2
≥s B log |9s T | + 2(A + B) ≤ CeCn−cB n .

2.2. Stieltjes Transforms and Free Convolutions. In this section we recall various results
concerning Stieltjes transforms and free convolution with the semicircle distribution. We first require
some notation on measures. In what follows, we let Pfin = Pfin (R) denote the set of nonnegative
measures µ on R with finite total mass, namely, µ(R) < ∞. Further let P = P(R) ⊂ Pfin denote the
set of probability measures on R, and let P0 = P0 (R) ⊂ P denote the set of probability measures
that are compactly supported; the support of any measure ν ∈ P is denoted by supp ν. We say that
5For example, if x (a) = f (a) and f (a+ ) > f (a), then x will be discontinuous at a.
n n
12 AMOL AGGARWAL AND JIAOYANG HUANG

a probability measure µ ∈ P has density ϱ (with respect to Lebesgue measure) if ϱ : R → R is a


measurable function satisfying µ(dx) = ϱ(x)dx. For any real number x ∈ R, we let δx ∈ P0 denote
the delta function at x.
Fix a measure µ ∈ Pfin . Its Stieltjes transform is the function m = mµ : H → H defined by, for
any complex number z ∈ H, setting
Z ∞
µ(dx)
(2.5) m(z) = .
−∞ x − z

If µ has a density ϱ ∈ L1 (R), then ϱ can be recovered from its Stieltjes transform by the identity
[58, Equation (8.14)],
(2.6) π −1 lim Im m(x + iy) = ϱ(x); π −1 lim Re m(x + iy) = Hϱ(x),
y→0 y→0

for any x ∈ R. In the latter, Hf denotes the Hilbert transform of any function f ∈ L1 (R), given by
Z ∞
f (w)dw
(2.7) Hf (x) = π −1 · PV ,
−∞ w − x
where PV denotes the Cauchy principal value (assuming the integral exists as a principal value).
The semicircle distribution is a measure µsemci ∈ P(R) whose density ϱsemci : R → R≥0 is
(4 − x2 )1/2
(2.8) ϱsemci (x) = · 1x∈[−2,2] , for all x ∈ R.

(t) (t)
For any real t > 0, we denote the rescaled semicircle density ϱsemci and distribution µsemci ∈ P by
(t) (t) (t)
(2.9) ϱsemci (x) = t−1/2 ϱsemci (t−1/2 x); µsemci (dx) = ϱsemci (x)dx.
We next discuss the free convolution of a probability measure µ ∈ P with the (rescaled) semicircle
(t)
distribution µsemci . First, for any real number t > 0, define the function M = Mt;µ : H → C and
the set Λt = Λt;µ ⊆ H by
( Z ∞ )
n  o µ(dx) 1
(2.10) M (z) = z − tm(z); Λt = z ∈ H : Im z − tm(z) > 0 = z ∈ H : 2
< .
−∞ |z − x| t

Lemma 2.8 ([8, Lemma 4]). The function M is a homeomorphism from Λt to H. Moreover, it is
a holomorphic map from Λt to H and a bijection from ∂Λt to R.
For any real number t ≥ 0, define mt = mt;µ : H → H as follows. First set m0 (z) = m(z); for
any real number t > 0, define mt so that

(2.11) mt z − tm0 (z) = m0 (z), for any z ∈ Λt .
Since by Lemma 2.8 the function M (z) = z − tm0 (z) is a bijection from Λt to H, (2.11) defines mt
on H. By [8, Proposition 2], mt is the Stieltjes transform of a probability measure µt ∈ P(R). This
(t) (t)
measure is called the free convolution between µ and µsemci , and we often write µt = µ ⊞ µsemci .
µ
By [8, Corollary 2], µt has a density ϱt = ϱt : R → R≥0 for any t > 0.
The following lemma provides an expression of the t-derivative of ϱt ; it is essentially known [64, 8]
but we provide the proof of its second part for convenience.
Lemma 2.9 ([64,
 8]). For any real number t > 0, both ϱt (y) and its Hilbert transform Hϱt (y) are
smooth in y ∈ y ′ ∈ R : ϱt (y ′ ) > 0 . Moreover, we have
for y ∈ y ′ ∈ R : ϱt (y ′ ) > 0 .
 
(2.12) ∂t ϱt (y) = π · ∂y ϱt (y)Hϱt (y) ,
LOCAL STATISTICS AND CONCENTRATION FOR NON-INTERSECTING BROWNIAN BRIDGES 13

Proof. The first part follows from [8, Corollary 3] (together with [8, Lemma 2]). For the second
part, by taking the derivative on both sides of (2.11) with respect to t, we get for any z ∈ Λt that
   
0 = ∂t m0 (z) = ∂t mt z − tm0 (z) = ∂t mt z − tm0 (z) − ∂z mt z − tm0 (z) · m0 (z)
  
= ∂t mt z − tm0 (z) − ∂z mt z − tm0 (z) · mt z − tm0 (z) .
Setting w = z − tm0 (z) and applying Lemma 2.8, it follows for any w = y + iη ∈ H that
1 2
∂t mt (y + iη) = · ∂z mt (y + iη) .
2
Sending η to zero and applying (2.6) (with the first part of the lemma), we find
 π2  2 
(2.13) π · ∂t Hϱt (y) + iϱt (y) = · ∂y Hϱt (y) + iϱt (y) .
2
Then (2.12) follows from comparing the imaginary parts on both sides of (2.13). □

The following lemmas indicate that ϱt is smooth on its support and bounds its derivatives there;
we provide their proofs (which follow similar ones in [8]) in Appendix B.2 below.
Lemma 2.10. For any real number t > 0, we have ϱµt (x) ≤ π −1 t−1/2 . Moreover, ϱt is smooth
on its support; in particular, for any integer k ∈ Z≥1 there exists a constant C = C(k) > 1
such
k that the following holds. For any real numbers t, δ ∈ (0, 1) and x0 ∈ R, if ϱt (x0 ) ≥ δ then
∂x ϱt (x0 ) ≤ C(t4 δ 6 )−k .

Lemma 2.11. For any integer k ≥ 2; real numbers 0 < δ ≤ c < 1 and B > 1, there exist constants
t0 = t0 (k, δ, c, B) > 0 and C = C(k, δ, c, B) > 1 such that the following holds. Let µ0 ∈ P0 denote
a compactly supported probability measure with a bounded density ϱ0 : R → R≥0 with respect to
Lebesgue measure. Assume that inf |x|≤c ϱ0 (x) ≥ δ and sup|x|≤c ∂xm ϱ0 (x) ≤ B m+1 δ −m for any
m ∈ J0, kK. Then,
δ ′ ′
sup sup ∂tk−k −2 ∂xk ϱt (x) ≤ C.

(2.14) inf inf ϱt (x) ≥ ; max
|x|≤c/4 t∈[0,t0 ] 2 ′
k ∈J0,k−2K |x|≤c/4 t∈[0,t0 ]

Remark 2.12. While free convolutions are typically defined between probability measures, the re-
lation (2.11) also defines the free convolution of any measure µ ∈ Pfin , satisfying A = µ(R) < ∞,
(t)
e ∈ P from
with the rescaled semicircle distribution µsemci . Indeed, define the probability measure µ
−1 1/2
e(I) = A · µ(A I), for any interval I ⊆ R. Furthermore, for any real number s ≥ 0,
µ by setting µ
(s)
define the probability measure µ es = µ
e ⊞ µsemci , and denote its Stieltjes transform by m
e s . Then,
(t)
define the free convolution µt = µ ⊞ µsemci and its Stieltjes transform mt = mt;µ by setting
et (A−1/2 I),
µt (I) = A · µ for any interval I ⊆ R, so that e t (A−1/2 z),
mt (z) = A1/2 · m
where the second equality follows from the first by (2.5). Then,
 
mt z − tm0 (z) = A1/2 · m e t A−1/2 z − tm0 (z) = A1/2 · m e t A−1/2 z − tm
e 0 (A−1/2 z)
 

e 0 (A−1/2 z) = m0 (z),
= A1/2 · m
so that (2.11) continues to hold for mt . In particular, Lemma 2.8 and Lemma 2.9 (the latter by its
proof) also hold for µ.
14 AMOL AGGARWAL AND JIAOYANG HUANG

Remark 2.13. Let us describe a scaling invariance for Stieltjes transforms. Fix a measure µ ∈ Pfin
(s)
with µ(R) < ∞, and let ms denote the Stieltjes transform of µ ⊞ µsemci , for any real number s ≥ 0.
Fix a real number β > 0, and define the measure µ e ∈ Pfin by setting µ
e(I) = µ(β 1/2 · I), for any
interval I ⊆ R. Denote the Stieltjes transform of µ e by me = mµ̃ , and let me s denote the Stieltjes
(s)
e ⊞ µsemci for any s > 0. Then, observe for any real number t ≥ 0 and complex number
transform of µ
z ∈ H that
(2.15) e t (z) = β 1/2 · mβt (β 1/2 z).
m
Indeed, this holds at t = 0 by (2.5); for t > 0, we have
e 0 (z) = β 1/2 · m0 (β 1/2 z) = β 1/2 · mβt β 1/2 z − βtm0 (β 1/2 z)
 
me t z − tm
e 0 (z) = m
 
= β 1/2 · mβt β 1/2 z − tme 0 (z) ,
which by Lemma 2.8 (and Remark 2.12, if µ is not a probability measure) implies (2.15).
2.3. Dyson Brownian Motion. In this section we recall properties about Dyson  Brownian mo-
tion. Fix an integer n ≥ 1 and a sequence λ(0) = λ1 (0), λ2 (0), . . . , λn (0) ∈ Wn . Define

λ(t) = λ1 (t), λ2 (t), . . . , λn (t) ∈ Wn , for t ≥ 0, to be the unique strong solution (see [6, Proposition
4.3.5]) to the stochastic differential equations
X dt
(2.16) dλi (t) = n−1 + n−1/2 dBi (t).
λi (t) − λj (t)
1≤j≤n
j̸=i

The system (2.16) is called Dyson Brownian motion, run for time t, with initial data λ(0); the λi
are sometimes referred to as particles.
Remark 2.14. As in Remark 2.2, Dyson Brownian motion admits the following invariance un-
der diffusive scaling for any real number σ > 0. If λ(t) = λ1 (t), λ2 (t), . . . , λn (t) ∈ Wn solves
(2.16) then, denoting λ e (t) = σ −1 λj (σ 2 t) for each (j, t) ∈ J1, nK × R≥0 , the process λ(t) =
 j
e
λ1 (t), λ2 (t), . . . , λn (t) ∈ Wn also solves (2.16). This again follows from the invariance of the
e e e
Brownian motions Bi under the same scaling.
We next describe the relation between Dyson Brownian motion, random matrices, and non-
intersecting Brownian bridges, to which end we require some additional terminology. A random
matrix is a matrix whose entries are random variables. The Gaussian Unitary Ensemble is an
n × n random Hermitian matrix G = Gn with random complex entries {wij } (for i, j ∈ J1, nK)
defined as follows. Its diagonal entries {wjj } are real Gaussian random variables of variance n−1 ;
its upper-triangular entries {wij }i<j are complex random variables, whose real and imaginary parts
are independent Gaussian random variables, each of variance (2n)−1 . These entries are mutually
independent, and the lower triangular entries {wij }i>j are determined from the upper triangular
ones by the Hermitian symmetry relation wij = wji .
The Hermitian Brownian motion  G(t) = Gn (t) is a stochastic process (over t ≥ 0) on
 n×n
random matrices, whose entries wij (t) are defined as follows. Its diagonal entries wjj (t)
are Brownian motions of variance n−1 , and its upper triangular entries wij (t) i<j are complex


Brownian motions whose real and imaginary parts are independent Brownian motions, each of
−1
variance
 (2n) . These entries are again mutually independent, and the lower triangular entries
wij (t) i>j are determined from its upper triangular ones by symmetry, wij (t) = wji (t). Observe
that G(1) has the same law as a GUE matrix G.
LOCAL STATISTICS AND CONCENTRATION FOR NON-INTERSECTING BROWNIAN BRIDGES 15

The following lemma from [22] (stated as below in [34]) interprets Dyson Brownian motion
in terms of sums of random matrices, and also in terms of non-intersecting Brownian motions
conditioned to never intersect (for the definition of the latter in terms of Doob h-transforms, see
[34, Section 6.2]).
Lemma 2.15 ([34, Theorems 3 and 4]). Fix an integer n ≥ 1 and a sequence λ(0) ∈ Wn . For any
real number s > 0, let λ(s) ∈ Wn denote Dyson Brownian motion, run for time s, with initial data
λ(0). Further let A denote an n × n diagonal matrix whose eigenvalues are given by λ(0), and let
G(s) = Gn (s) denote an n × n Hermitian Brownian motion.
(1) The law of the eigenvalues of A + G(s) coincides with that of λ(s), jointly over s ≥ 0.
(2) Consider n Brownian motions x = (x1 , x2 , . . . , xn ) ∈ J1, nK × C(R≥0), with variances n−1
and starting data λ(0), conditioned to never intersect. Then, x(s) s≥0 has the same law

as λ(s) s≥0 .
We next discuss concentration results from [41] for Dyson Brownian motion, to which end we
require the notion of a classical location with respect to a measure.
Definition 2.16. Let µ ∈ P denote a probabilty measure. For any integers n ≥ 1 and j ∈ Z, we
µ
define the classical location (with respect to µ) γj;n = γj;n ∈ R by setting
( Z ∞ )
2j − 1
γj;n = sup γ ∈ R : µ(dx) ≥ , if j ∈ J1, nK,
γ 2n
and also setting γj = ∞ if j ≤ 0 and γj = −∞ if j > n.
The following lemma due to6 [41] (together with the scale invariance Remark 2.14) provides a
concentration, or rigidity, estimate for the locations of bulk (namely, those sufficiently distant from
the first and last) particles under Dyson Brownian motion around the classical locations of a free
convolution measure (recall the latter from Section 2.2).
Lemma 2.17 ([41, Corollary 3.2]). For any real number A > 1, there exists a constant C =
C(A) > 1 such that the following holds. Fix an integer n ≥ 1 and sequence λ(0) ∈ Wn with −nA ≤
Pn (t)
min λ(0) ≤ max λ(0) ≤ nA . Denote the measure µ = n−1 j=1 δλj (0) ∈ P, and set µt = µ ⊞ µsemci ;
µt 
also denote the classical locations γj (t) = γj;n ∈ R. Letting λ(s) = λ1 (s), λ2 (s), . . . , λn (s) ∈ Wn
denote Dyson Brownian motion with initial data λ(0), we have
" n #
\ \  2
−A −A
≥ 1 − Ce−(log n) .

(2.17) P γj+⌊(log n)5 ⌋ (t) − n ≤ λj (t) ≤ γj−⌊(log n)5 ⌋ (t) + n
j=1 t∈[0,nA ]

2.4. Brownian Watermelon Estimates. In this section we provide estimates for the locations
of paths in an ensemble of non-intersecting Brownian bridges conditioned to start and end at 0.
Such ensembles are sometimes referred to as Brownian watermelons; see Figure 2. In what follows,
for any integers n ≥ 1 and j ∈ Z we let γsemci;n (j) be the classical location (recall Definition 2.16)
with respect to the semicircle law, given by
Z 2
µsemci 1 2j − 1
(2.18) γsemci;n (j) = γj;n , which satisfies (4 − x2 )1/2 dx = ,
2π γsemci;n (j) n

6In [41], the probability on the right side of (2.17) was written to be 1 − Cn−D for any D > 1, but it can be seen
5 2
from the proof (see that of [41, Proposition 3.8], where δ there is 4
here) that it can be taken to be 1 − Ce−(log n)
instead.
16 AMOL AGGARWAL AND JIAOYANG HUANG

whenever j ∈ J1, nK.


The following lemma then provides a concentration bound for the paths in a Brownian water-
melon; its proof is given in Appendix B.5 below.
Lemma 2.18. For any real number D > 1, there exist constants c > 0 and C = C(D) > 1 such that
the following holds. Adopt the notation of Lemma 2.6; assume that b − a ≤ nD ; fix real numbers
u, v ∈ R; and assume that u = (u, u, . . . , u) ∈ Wn and v = (v, v, . . . , v) ∈ Wn (where u and v
appear with multiplicity n).
5
(1) With probability at least 1 − Ce−c(log n) , we have
 1/2
(b − t)(t − a) b−t t − a
max sup xj (t) − · γsemci;n (j) − ·u− · v

j∈J1,nK t∈[a,b] (b − a) b−a b−a
≤ (log n)9 · n−2/3 (b − a)1/2 · min{j, n − j + 1}−1/3 .
5
(2) With probability at least 1 − Ce−c(log n) , we have
 1/2 !
b−t t − a 1/2 (b − t)(t − a)
≤ n−D ;

max sup xj (t) − ·u− ·v −8
j∈J1,nK t∈[a,b] b−a b−a b−a
 1/2 !
b − t t − a (b − t)(t − a)
1/2
≥ −n−D .

min inf xj (t) − ·u− · v + 8
j∈J1,nK t∈[a,b] b−a b−a b−a

3. Limit Shapes for Non-intersecting Brownian Bridges


In this section we consider limit shapes for non-intesection Brownian bridges, in the case without
upper and lower boundary, and explain how they solve the partial differential equation (1.2).

3.1. Limit Shapes. In this section we recall results from [36, 35, 7] concerning the limiting macro-
scopic behavior of non-intersecting Brownian bridges under given starting and ending data, with
no upper and lower boundaries. Throughout, we use coordinates (t, x), or sometimes (t, y), for R2 .
For any sequence a = (a1 , a2 , . . . , an ) ∈ Wn , we denote its empirical measure emp(a) ∈ P by
n
1X
(3.1) emp(a) = δa .
n j=1 j

Fix an interval I ⊆ R. A measure-valued process (on the time interval I) is a family µ = (µt )t∈I of
measures µt ∈ Pfin for each t ∈ I. Given a real number A > 0, we say that µ has constant total mass
A if µt (R) = A, for each t ∈ I. If µ has constant total mass 1 (so each µt ∈ P), we call µ a probability
measure-valued process. Measure-valued processes can be interpreted as elements of I × Pfin and
probability measure-valued processes as ones of I × P. We let C(I; Pfin ) and C(I; P) denote the sets
of measure-valued processes and probability measure-valued processes that are continuous in t ∈ I,
under the topology of weak convergence on Pfin and P, respectively.
Given two measures µ, ν ∈ Pfin of finite total mass, the Lévy distance between them is
( Z y−a Z y Z y+a )
(3.2) dL (µ, ν) = inf a > 0 : µ(dx) − a ≤ ν(dx) ≤ µ(dx) + a, for all y ∈ R .
−∞ −∞ −∞
LOCAL STATISTICS AND CONCENTRATION FOR NON-INTERSECTING BROWNIAN BRIDGES 17

Figure 2. Shown above is a Brownian watermelon, consisting of non-intersecting


Brownian bridges starting and ending at 0.

Given an interval I ⊆ R and two measure-valued processes µ = (µt )t∈I ∈ I × Pfin and ν = (νt )t∈I ∈
I × Pfin on the time interval I, the Lévy distance between them is defined to be
dL (µ, ν) = sup dL (µt , νt ).
t∈I

The following lemma from [7] (based on results from [35, 36]) states that, as n tends to ∞,
the empirical measure for n non-intersecting Brownian bridges (whose starting and ending data
converge in a certain way) has a limit; see Figure 2 in the case of the Brownian watermelon, whose
limit shape is shown on the left side of Figure 4. The following lemma was stated in [7, 36] in the
case when [a, b] = [0, 1] and A = 1 but, by the scaling invariance (Remark 2.2) for non-intersecting
Brownian bridges, it also holds for any interval [a, b] and real number A > 0, as below.
Lemma 3.1 ([7, Claim 2.13]). Fix real numbers a < b and compactly supported measures µa , µb ∈
Pfin , both of total mass µa (R) = A = µb (R) for some real number A > 0. There is a measure-valued
process µ⋆ = (µ⋆t )t∈[a,b] ∈ C [a, b]; Pfin on [a, b] of constant total mass A, which is continuous in the
pair (µa , µb ) ∈ P2 under the Lévy metric, such that the following holds. For each integer n ≥ 1, let
u = un ∈ Wn and v = v n ∈ Wn denote sequences such that A·emp(un ) and A·emp(v n ) converge to
µa and µb under the Lévy metric as n tends to ∞, respectively. Sample n non-intersecting Brownian
bridges xn = (xn1 , xn2 , . . . , xnn ) ∈ J1, nK × C [a, b] from Qu;v (An−1 ); for

 any t ∈ [a, b],
 denote
νtn = A · emp xn (t) ∈ P. For any real number ε > 0, we have limn→∞ P dL (ν n , µ⋆ ) > ε = 0.
Terminology for the limit shape provided by Lemma 3.1 is given through the following definition.
Definition 3.2. Adopting the notation of Lemma 3.1, the measure-valued process µ⋆ = (µ⋆t )t∈[a,b]
is called the bridge-limiting measure process (on the interval [a, b]) with boundary data (µa ; µb ).
The following lemma indicates how bridge-limiting measure processes restrict to others. Its proof
is given in Appendix C.1 below.
Lemma 3.3. Adopt the notation and assumptions of Lemma 3.1, and let e a, eb ∈ R be real numbers
such that a ≤ ea < b ≤ b. Then, the bridge-limiting measure process on the interval [e
e a, eb] with
⋆ ⋆ ⋆
boundary data (µea ; µb̃ ) is given by (µt )t∈[ã,b̃] .
18 AMOL AGGARWAL AND JIAOYANG HUANG

We will in many cases make use of a height function and inverted height functions associated
with a measure-valued process, defined below.

Definition 3.4. Fix an interval I = [a, b] ⊆ R and a measure-valued process µ = (µt )t∈I of
constant total mass A > 0. The height function associated with µ is defined to be the function
H = H µ : I × R → [0, A] obtained by setting
Z ∞
(3.3) H(t, x) = µt (dw), for each (t, x) ∈ I × R.
x

The invertedheight function associated with µ is G = Gµ : I × [0, A] → R, defined by setting


G(t, 0) = inf x : H(t, x) = 0 and

(3.4) G(t, y) = x, where x = x(t, y) = sup x ∈ R : H(t, x) ≥ y , for y ∈ (0, A].

If µt = ϱt (x)dx has a density with respect to Lebesgue measure for each t ∈ (a, b), then we
sometimes associate H (or G) with ϱ = (ϱt ). Moreover, if µ is the bridge-limiting measure process
with boundary data (µa ; µb ) we say that H (or G) is associated with boundary data (µa ; µb ).

To simplify notation in what follows, it will often be useful to set (a, b) = (0, 1) and A = 1. The
following lemma indicates a transformation that implements this specialization. Its proof follows
from a quick consequence of Remark 2.2 and is provided in Appendix C.1 below.

Lemma 3.5. Adopt the notation and assumptions of Lemma 3.1; let H ⋆ : [a, b] × R → R and
G⋆ : [a, b] × [0, A] → R denote the height function and inverted height function associated with µ⋆ ,
respectively. Define the probability measures µ e1 by for any interval I ⊆ R setting
e0 , µ

e0 (I) = A−1 · µa A1/2 (b − a)1/2 · I ; e1 (I) = A−1 · µb A1/2 (b − a)1/2 · I


 
µ µ

µ⋆t )t∈[0,1] ∈ C [0, 1]; P and functions



Further define the probability measure-valued process µ e ⋆ = (e
He ⋆ : [0, 1] × R → R and G e ⋆ : [0, 1] × [0, 1] → R by setting

e⋆t (I) = A−1 · µ⋆a+t(b−a) A1/2 (b − a)1/2 · I ,



µ for any interval I ⊆ R;
⋆ −1 ⋆ 1/2 1/2

(3.5) He (t, x) = A · H a + t(b − a), A (b − a) x , for any (t, x) ∈ [0, 1] × R;
e ⋆ (t, y) = A−1/2 (b − a)−1/2 · G⋆ a + t(b − a), Ay ,

G for any (t, y) ∈ [0, 1] × [0, 1].

Then, µe ⋆ is the bridge-limiting measure process on [0, 1] with boundary data (e µ0 ; µ


e1 ). Moreover, its
associated height function and inverted height function are H e ⋆ and G
e ⋆ , respectively.

The following corollary reformulates Lemma 3.1 in terms of the associated (inverted) height
functions; in particular, it explains that the trajectory of the ⌊yn⌋-th path is approximated by the
inverted height function G⋆ associated with the bridge-limiting process. In what follows, given an
integer n ≥ 1, real
 numbers a < b, and a family of non-intersecting curves x = (x1 , x2 , . . . , xn ) ∈
J1, nK × C [a, b] , the associated (discrete) height function H = Hx : [a, b] × R → R is defined by

H(t, x) = Hx (t, x) = # j ∈ J1, nK : xj (t) > x .



(3.6)

Corollary 3.6. Adopt the notation and assumptions of Lemma 3.1, and fix a real number ε > 0.
Let H ⋆ : [a, b] × R → [0, A] and G⋆ : [a, b] × [0, A] → R denote the height and inverted height
LOCAL STATISTICS AND CONCENTRATION FOR NON-INTERSECTING BROWNIAN BRIDGES 19

functions associated with µ⋆ , respectively. We have


" #
\  −1 ⋆ −1 x −1 ⋆

lim P A H (t, w + ε) − ε ≤ n H (t, w) ≤ A H (t, w − ε) + ε = 1;
n→∞
(t,w)∈[a,b]×R
(3.7) " #
\
G⋆ (t, Ay + ε) − ε ≤ x⌊yn⌋ (t) ≤ G⋆ (t, Ay − ε) + ε

lim P = 1.
n→∞
(t,y)∈[a,b]×[0,1]

Proof. By Lemma 3.1, dL (ν n , µ⋆ ) tends to 0 in probability, as n tends to ∞. Combined with the


definition (3.2) of Lévy metric, this yields the first statement of (3.7). This, with the definition (3.4)
of G⋆ in terms of H ⋆ and that (3.6) of Hx in terms of x, gives the second statement of (3.7). □

The following lemma essentially due to [35] indicates that the measures µ⋆t have a density, and it
also discusses properties of this density. Its proof is given in Appendix C.1 below. In what follows,
(t)
we recall the free convolution and semicircle law µsemci ∈ P0 from Section 2.2. We also restrict to
the case (a, b) = (0, 1) and A = 1 to simplify the statement, though an analog of the below lemma
can be derived under arbitrary such parameters by scaling (see Remark 3.14 below).
Lemma 3.7. Adopt the notation and assumptions of Lemma 3.1, and assume (a, b) = (0, 1) and
A = 1. The following statements hold for each real number t ∈ (0, 1).
(1) There exists a measurable function ϱ⋆t : R → R≥0 such that µ⋆t (dx) = ϱ⋆t (x)dx.
(2) There exists some compactly supported probability measure νt ∈ P0 , dependent on µ0 and
(t−t2 )
ϱ1 with supp νt ⊆ supp µ0 + supp µ1 , such that ϱ⋆t = νt ⊞ µsemci .
(3) We have ϱ⋆t (x) ≤ (t − t2 )−1/2 , for any x ∈ R. Moreover, for any integer k ≥ 1, there exists
a constant C = C(k) > 1 such that, for any real number δ > 0 and point (t, x0 ) ∈ Ω with
ϱ⋆t (x0 ) ≥ δ, we have ∂xk ϱ⋆t (x0 ) ≤ Cδ −6k (t − t2 )−4k .
(4) The function ϱt (x) is continuous on (0, 1) × R.
Next, fix an interval (a, b) ⊆ R and a family of measures µ = (µt )t∈(a,b) ∈ (a, b) × Pfin of constant
total mass A > 0. Assume for each t ∈ (a, b) that each µt has a density ϱt (x), for some function
ϱt : R → R≥0 that is continuous in (t, x). We define the associated liquid region Ω ⊂ (a, b) × R and
inverted liquid region Ωinv ⊆ (a, b) × (0, 1) by
(3.8)
Ωinv = (t, y) ∈ (a, b) × [0, A] : y = H µ (t, x), (t, x) ∈ Ω .
 
Ω = (t, x) ∈ (a, b) × R : ϱt (x) > 0 ;

Observe that the map (t, x) 7→ t, H(t, x) is a bijection to from Ω to Ωinv . Moreover, the continuity
of ϱt (x) implies that the set Ω is open, which implies that Ωinv is also open. See Figure 3.
Corollary 3.8. Under the assumptions of Corollary 3.6, for any y ∈ (0, 1) such that (t, Ay) ∈ Ωinv
for each t ∈ (a, b), we have that G⋆ (t, Ay) is continuous in t ∈ [a, b], and
" #
\ 
⋆ n ⋆

(3.9) lim P G (t, Ay) − ε ≤ x⌊yn⌋ (t) ≤ G (t, Ay) + ε = 1.
n→∞
t∈(a,b)

Proof. Observe that G⋆ is continuous on Ωinv , by (3.4) and the fact that H(t, x) is continuous
and strictly increasing in x whenever (t, x) ∈ Ω. In particular, G⋆ is uniformly continuous on
compact subsets of Ωinv ; this, together with the second statement of (3.7), yields for any real
20 AMOL AGGARWAL AND JIAOYANG HUANG

Figure 3. The left panel is the liquid region Ω and the right panel is the inverted
liquid region Ωinv .

number δ ∈ 0, 12 that, with probabiliy 1 − o(1) (that is, tending to 1 as n tends to ∞), we have


sup x⌊yn⌋ (t) − G⋆ (t, Ay) < δ.



(3.10)
t∈[δ,1−δ]

Next observe that there is a constant C1 = C1 (y, µ0 , µ1 ) > 1 such that, with probability 1 − o(1),
we have
sup x⌊yn⌋ (t) − G⋆ (δ, Ay) ≤ C1 δ 1/3 ; sup x⌊yn⌋ (b) − G⋆ (1 − δ, Ay) ≤ C1 δ 1/3 ,

(3.11)
t∈[0,δ] t∈[1−δ,1]

where we used (3.10) and Lemma 2.6. Letting n tend to ∞ in (3.11) and applying (3.10), we obtain
sup G⋆ (t, Ay) − G⋆ (t′ , Ay) ≤ 2C1 δ 1/3 ; sup G⋆ (t, Ay) − G⋆ (t′ , Ay) ≤ 2C1 δ 1/3 ,

t,t′ ∈[0,δ] t,t′ ∈[1−δ,1]

This, with the continuity of G⋆ on compact subsets of Ωinv , implies that G⋆ (t, Ay) is continuous in
t ∈ [a, b]. Moreover, with (3.11) and (3.10), it also yields (3.9) by taking δ sufficiently small with
respect to ε. □

Before proceeding, let us provide two examples of the above notions. The first concerns the case
when µ0 and µ1 are delta measures, in which the associated non-intersecting Brownian bridges form
a Brownian watermelon (recall Section 2.4); see the left side of Figure 4.
Example 3.9. Fix real numbers a < b; u, v ∈ R; and A > 0. Assume that (µa , µb ) = (A · δu , A · δv ),
where δx ∈ P0 denotes the delta measure at x ∈ R. Then, it follows from Lemma 2.18 (multiplying
its results by A1/2 to account for the fact that the Brownian motions have variance An−1 here) and
the second statement of (3.7) (and also the continuity of γsemci (y) below in y) that the inverted
height function G⋆ : [0, 1] × [0, A] → R associated with boundary data (µ0 ; µ1 ) is given by
 1/2
A(b − t)(t − a) y b−t t−a
G⋆ (t, y) = · γsemci + ·u+ · v,
b−a A b−a b−a
LOCAL STATISTICS AND CONCENTRATION FOR NON-INTERSECTING BROWNIAN BRIDGES 21

1.0 0.5

0.8 0.4

0.6 0.3

0.4 0.2

0.2 0.1

0.0 0.0

Figure 4. The left panel is the density from Example 3.9 with (a, b, u, v, A) =
(0, 6, 0, 0, 2), and the whole shaded region is the liquid region Ω. The right panel
is the density from Corollary 3.10 with (a, b, d, A) = (0, 10, 2, 2), and the whole
shaded region is the liquid region Ω.

where γsemci (y) is the classical location of the semicircle law, defined to be the
Z 2
(3.12) unique γ ∈ [−2, 2] solving the equation (2π)−1 (4 − x2 )1/2 dx = y.
γ

Together with (3.3) and (3.4), it follows that the associated density process (ϱ⋆t ) and the height
function H ⋆ : [0, 1] × R are given by
 
⋆ (
A(b−t)(t−a)
(b−a)
) b−t t−a
ϱt (x) = A · ϱsemci x− ·u− ·v ,
b−a b−a
R∞ (t)
and H ⋆ (t, x) = x ϱ⋆t (y)dy, where we recall the rescaled semcircle density ϱsemci from (2.9).
The second example concerns the case when µ0 and µ1 are rescaled semicircle laws (recall (2.9)),
which is obtained by restricting a larger watermelon to a smaller interval. See the right side of
Figure 4.
(d)
Corollary 3.10. Fix real numbers a < b, d > 0, and A > 0; assume that µa = A · µsemci = µb .
Then, the inverted height function G⋆ : [a, b] × [0, A] → R and density process (ϱ⋆t ) associated with
boundary data (µa ; µb ) are given by
 1/2

A(b−t)(t−a)
(d+ b−a+2κ ) ⋆ A(b − t)(t − a) y
(3.13) ϱt (x) = A · ϱsemci (x); G (t, y) = d + · γsemci ,
b − a + 2κ A
where κ = κ(a, b, d) > 0 is defined by
 1/2
d a−b b − a 2  d 2
(3.14) κ= + + + .
A 2 2 A
Proof. This will follow from using Lemma 3.3 to restrict the limit shape from Example 3.9 to a
µ⋆t )t∈[a−κ,b+κ] denote the bridge-limiting measure process
smaller interval. To implement this, let (e
22 AMOL AGGARWAL AND JIAOYANG HUANG

on [a − κ, b + κ] with boundary data (δ0 , δ0 ). Then, for any t ∈ [a − κ, b + κ], we have


(
A(b+κ−t)(t−a+κ)
A(b + κ − t)(t − a + κ)
) A(b − t)(t − a)
(3.15) e⋆t = A · µsemci b−a+2κ
µ ; =d+ ,
b − a + 2κ b − a + 2κ
where the first statement follows from Example 3.9 and the second from the choice (3.14) of κ. It
(d)
follows that µe⋆a = µsemci = µ
e⋆b , and so from Lemma 3.3 we find µ⋆t = µe⋆t for each t ∈ [a, b]. Applied
with (3.15), this yields the first statement of (3.13), which (together with (3.3), (3.4), and (3.12))
implies its second statement. □
3.2. The Functions (u⋆ , ϱ⋆ ). In this section we recall results from [35, 36] stating that, under
conditions on µa and µb , the associated bridge-limiting density process µt gives rise to a pair of
functions (u⋆t , ϱ⋆t ) that satisfy certain equations (which will later be interpreted in terms of elliptic
partial differential equations for the associated height and inverted height functions in Section 3.3
below). For notational convenience, we at first restrict to the case of Lemma 3.1 when (a, b) = (0, 1)
and A = 1 (explaining how the more general case follows from it through scaling in Remark 3.14
below).
The following result appears as stated below7 in [7] and was originally established in [35, 36]
(see8 [35, Theorem 2.1]). Here, we recall from Lemma 3.7 that, at a given time t ∈ (0, 1), any
bridge-limiting measure process µ⋆ has a density ϱ⋆t : R → R≥0 ; we also recall from (3.8) the liquid
region Ω associated with any ϱ ∈ (0, 1) × R → R≥0 .
Lemma 3.11 ([7, Theorems 2.3 and 2.11]). Fix probability measures µ0 , µ1 ∈ P with densities
ϱ0 , ϱ1 , respectively, such that

(3.16) ϱ0 and ϱ1 are compactly supported; sup ϱ0 (x) < ∞; sup ϱ1 (x) < ∞.
x∈R x∈R

∈ C [0, 1]; P denote the bridge-limiting measure process on [0, 1] with boundary data

Let (µ⋆t )t∈[0,1]
(µ0 ; µ1 ). For each t ∈ [0, 1], denote the density of µ⋆t ∈ P by ϱ⋆t : R → R≥0 . There exists a
measurable function u⋆ : [0, 1] × R → R, denoted by u⋆t (x) = u⋆ (t, x), satisfying the following three
properties.

(1) We have ∂t ϱ⋆t (x) + ∂x ϱ⋆t (x)u⋆t (x) = 0 weakly (in the sense of distributions) on (0, 1) × R.
More specifically, for any smooth, compactly supported function φ : (0, 1) × R → R, we have
Z 1Z ∞ Z 1Z ∞
ϱ⋆t (x)∂t φ(t, x)dxdt + ϱ⋆t (x)u⋆t (x)∂x φ(t, x)dxdt = 0.
0 −∞ 0 −∞

(2) The functions ϱ and u are smooth in (t, x) on the liquid region Ω associated with ϱ⋆ .
⋆ ⋆

(3) For each (t, x) ∈ Ω, the pair (u⋆ , ϱ⋆ ) satisfies


 π2 ⋆ 3 
(3.17) ∂t ϱ⋆t (x) + ∂x u⋆t (x)ϱ⋆t (x) = 0; ∂t u⋆t (x)ϱ⋆t (x) + ∂x u⋆t (x)2 ϱ⋆t (x) −
 
ϱ (x) = 0.
3 t
7Let us mention that, as stated, those results apply to Dyson Brownian motion (2.16) started at some u ∈ W ,
n
conditioned to be v ∈ Wn at time 1. By Lemma 2.15, this is equivalent to non-intersecting Brownian bridges of
variance n−1 , starting at u and ending at v.
8
An alternative proof for the smoothness of u⋆ and ϱ⋆ (the second part of Lemma 3.11) would proceed by first
using Lemma 3.16 (see also Remark 3.17) below to show that the associated height function H ⋆ weakly solves
the divergence-form elliptic partial differential equation (3.27). The bounds of [35, Corollary 2.8(c)] verify that this
equation is elliptic (using Remark 3.19 below), uniformly on compact subsets of Ω. Together with regularity estimates
for weak solutions to such elliptic differential equations given by [33, Lemma 6.27, Theorem 8.20, Equation (13.41)]
and [5, Chapter 3.6], this yields the smoothness of H ⋆ , implying that of u⋆ and ϱ⋆ by Lemma 3.13 below.
LOCAL STATISTICS AND CONCENTRATION FOR NON-INTERSECTING BROWNIAN BRIDGES 23

Although we will not make substantial use of it, the next lemma from [7] (we recall its quick
proof in Appendix C.2 below) reformulates (3.17) as an equation for the associated complex slope
f = f (µ0 ;µ1 ) : (0, 1) × R → H, defined by
(3.18) f (t, x) = u⋆t (x) + πiϱ⋆t (x).
Lemma 3.12 ([7, Equation (2.16)]). Adopting the notation and assumptions of Lemma 3.11, the
associated complex slope f defined by (3.18) satisfies the complex Burgers equation,
(3.19) ∂t f (t, x) + f (t, x) · ∂x f (t, x) = 0, for all (t, x) ∈ Ω.
The following lemma expresses the derivatives of the height function (recall Definition 3.4) asso-
ciated with non-intersecting Brownian bridges in terms of the (u⋆ , ϱ⋆ ) above. The first statement
of (3.20) follows from the definition (3.3) of the height function. Together with the equation
∂t ϱ⋆ + ∂x (u⋆ ϱ⋆ ) = 0, this formally implies the second statement of (3.20); its detailed proof is given
in Appendix C.2 below.
Lemma 3.13. Adopt the notation and assumptions of Lemma 3.11, and let H ⋆ denote the height
function associated with (µ⋆t ). Then H ⋆ is locally Lipschitz on (0, 1) × R and, for almost all (t, x) ∈
(0, 1) × R (with respect to Lebesgue measure) we have
(3.20) ∂x H ⋆ (t, x) = −ϱt (x); ∂t H ⋆ (t, x) = ut (x)ϱt (x).
Remark 3.14. Using Lemma 3.5, let us provide analogs of the above results when the parameters
(a, b) and A in Lemma 3.1 are arbitrary. To that end, adopt the notation of Lemma 3.5, and assume
(analogously to in (3.16)) that µa and µb have densities ϱa : R → R≥0 and ϱb : R → R≥0 with
respect to Lebesgue measure, satisfying

(3.21) ϱa and ϱb are compactly supported; sup ϱa (x) < ∞; sup ϱb (x) < ∞.
x∈R x∈R

Letting (ϱ⋆t )t∈[a,b]


and ϱ⋆t )t∈[0,1]
(e denote the densities of and (eµ⋆t )t∈[0,1] with respect to
(µ⋆t )t∈[a,b]
Lebesgue measure, respectively (guaranteed to exist by the first part of Lemma 3.7), we have from
the first statement of (3.5) that
ϱe⋆t (x) = A−1/2 (b − a)1/2 · ϱ⋆a+(b−a)t A1/2 (b − a)1/2 x ,

(3.22) for any (t, x) ∈ [0, 1] × R.
By the second part of Lemma 3.7, there exists νet ∈ P with supp νet ⊆ supp µ e0 + supp µ e⋆t =
e1 and µ
(t−t2 )
νet ⊞µsemci , for each t ∈ (0, 1). Hence, by Remark 2.12, the β = b−a case of Remark 2.13, and (3.22),
for any t ∈ (a, b) there exists some νt ∈ Pfin with νt (R) = A, such that supp νt ⊆ supp µa + supp µb
((t−a)(b−t)/(b−a))
and µ⋆t = νt ⊞ µsemci . This verifies that (the rescaled version) of the second part of
Lemma 3.7 holds for µ⋆t .
Further setting (as in (3.20))
∂t H ⋆ (t, x) e t⋆ (x)
∂t H
(3.23) u⋆t (x) = , if ϱt (x) > 0; e⋆t (x) =
u , if ϱe⋆t (x) > 0,
ϱt (x) ϱe⋆t (x)
we have from (3.22) and the second statement of (3.5) that
e⋆t (x) = A−1/2 (b − a)1/2 · u⋆a+(b−a)t A1/2 (b − a)1/2 x .

(3.24) u
Defining (as in (3.18)) the functions f (t, x) = u⋆t (x) + πiϱ⋆t (x) and fe(t, x) = ue⋆t (x) + πie
ϱ⋆t (x), we
have from (3.22) and (3.24) that
fe(t, x) = A−1/2 (b − a)1/2 · f a + (b − a)t, A1/2 (b − a)1/2 x .

24 AMOL AGGARWAL AND JIAOYANG HUANG

Together with the fact that fe satisfies the complex Burgers equation (3.19) by Lemma 3.12, this
implies that f also satisfies (3.19) on the liquid region Ω associated with ϱ⋆ .
Remark 3.15. Adopt the notation and assumptions of Lemma 3.1. Let H ⋆ : [a, b] × R → R and G⋆ :
[a, b] × [0, A] → R denote the height and inverted height functions associated with µ⋆ , respectively,
and let ϱ⋆t and u⋆t be as in Lemma 3.11 (and Remark 3.14); denote the associated inverted liquid
region (recall (3.8)) by Ωinv ⊆ (a, b) × (0, A). By Definition 3.4, we have for (t, y) ∈ Ωinv that
H ⋆ t, G⋆ (t, y) = y. Differentiating this equation in y or t, we find that (3.20) can equivalently be
formulated in terms G⋆ as
1
ϱ⋆t G⋆ (t, y) = − , and u⋆t G⋆ (t, y) = ∂t G⋆ (t, y), if (t, y) ∈ Ωinv ,
 

∂y G (t, y)
which defines ϱ⋆t and u⋆t everywhere on Ω (since G⋆ bijectively maps Ωinv to Ω).
3.3. Elliptic Partial Differential Equations for the Height Function. In this section we
reformulate (3.17) as elliptic partial differential equations satisfied by the height and inverted height
functions associated with µ⋆t . The following lemma, which follows quickly from (3.17), implements
this for the height function.
Lemma 3.16. Adopt the notation and assumptions of Lemma 3.11, and let H ⋆ denote the height
function associated with (µ⋆t ). Recalling the liquid region Ω ⊂ (0, 1) × R from (3.8), we have
X
bij ∇H ⋆ (t, x) ∂i ∂j H ⋆ (t, x) = 0,

(3.25) for any (t, x) ∈ Ω,
i,j∈{t,x}

where, for any (u, v) ∈ R2 , we have denoted


(3.26) btt (u, v) = v 2 ; btx (u, v) = −uv = bxt (u, v); bxx (u, v) = u2 + π 2 v 4 .
Proof. Applying Lemma 3.13 and the second statement of (3.17) yields
 2
(∂t H)2

π 3
(3.27) ∂tt H + ∂x (∂x H) − = 0,
3 ∂x H
or equivalently,
2∂t H (∂t H)2
(3.28) ∂tt H − · ∂tx H + · ∂xx H + π 2 (∂x H)2 · ∂xx H = 0.
∂x H (∂x H)2
Multiplying this equation by (∂x H)2 then yields (3.25). □

Remark 3.17. The derivation of (3.27) does not require smoothness of u⋆ and ϱ⋆ . In particular, if one
were to only assume that the second equation in (3.17) holds weakly (in the sense of distributions)
on Ω (as was shown in [35, Property 2.9(4)]), the proof of Lemma 3.16 shows that H is a weak
solution to the divergence-form ellliptic partial differential equation (3.27).
The next lemma, which follows from changing variables in Lemma 3.16, shows that the inverted
height function G⋆ satisfies (1.2); in view of Corollary 3.8, this implies that the limiting trajectories
of paths in a family of non-intersecting Brownian motions (without north or south boundary) is
govered by solutions of (1.2); in the below, we recall the open set Ωinv from (3.8).
Lemma 3.18. Adopt the notation and assumptions of Lemma 3.11, and let G⋆ denote the inverted
height function associated with (µ⋆t ). Then G⋆ satisfies the equation (1.2) for (t, y) ∈ Ωinv .
LOCAL STATISTICS AND CONCENTRATION FOR NON-INTERSECTING BROWNIAN BRIDGES 25

Proof. Let (t, y) ∈ Ωinv


 , and set x ∈ R such that H(t, x) = y; then, (t, x) ∈ Ω. Differentiating the
equation G t, H(t, x) = x, it is quickly verified that
1 ∂t H ∂xx H ∂t H ∂tx H
∂y G = ; ∂t G = − ; ∂yy G = − ; ∂ty G = · ∂xx H − ;
∂x H ∂x H (∂x H)3 (∂x H)3 (∂x H)2
2∂t H (∂t H)2 ∂tt H
∂tt G = 2
· ∂tx H − · ∂xx H − .
(∂x H) (∂x H)3 ∂x H
where in the above the arguments of H and G (and their derivatives) are (t, x) and (t, y), respec-
tively. Dividing (3.28) by ∂x H and applying the above equalities yields
∂tt G + π 2 (∂y G)−4 ∂yy G = ∂tt G − π 2 (∂x H)∂xx H
(∂t H)2
 
2∂t H ∂tt H
= 2
· ∂tx H − − + π 2 ∂x H · ∂xx H = 0,
(∂x H) ∂x H (∂x H)3
which establishes the lemma. □
Remark 3.19. The regularity properties of solutions to (3.25) and (1.2) depend on the 2×2 matrices
   2 
btt (u, v) btx (u, v) v −uv
B = B(u, v) = = , and
bxt (u, v) btt (u, v) −uv u2 + π 2 v 4
   
dtt (u, v) dty (u, v) 1 0
D = D(u, v) = = ,
dyt (u, v) dyy (u, v) 0 π 2 v −4
respectively; both are nonnegative definite. The two eigenvalues of B are bounded above and below
whenever |u| is bounded above and v is bounded above and below; the two eigenvalues of D are
bounded above and below whenever v is. In particular, that the latter only requires bounds on
v = ∂y G⋆ = (∂x H ⋆ )−1 , while the former requires those on both ∂x H ⋆ and ∂t H ⋆ .
Further regularity properties of solutions to (1.2) are provided in Section 9 below.
Remark 3.20. Formally, one can interpret (3.25) and (1.2) as Euler–Lagrange equations for mini-
mizers of the functionals
Z 1Z ∞  2
π2

∂t F (t, x) 3
E(F ) = − + ∂x F (t, x) dxdt, and
0 −∞ ∂x F (t, x) 3
Z 1Z 1
2 π 2 −2 
Einv (F ) = ∂t F (t, x) + ∂x F (t, x) dxdt, respectively.
0 0 3
However, we will not make this fully precise, as we will not require this interpretation.
It will be useful to make use of invariances of the equations (3.25) and (1.2) under the below
(linear and multiplicative) transformations.
Lemma 3.21. Fix a bounded, open subset R ⊂ R2 and two functions G, H ∈ Adm(R) ∩ C 2 (R);
assume on R that H satisfies (3.25) and that G satisfies (1.2). Fix real numbers α, β > 0, and
denote the open sets R
e =Re α;β ⊂ R2 and R b =Rb α ⊂ R2 by
e = (t, x) ∈ R2 : (αt, βx) ∈ R ; R b = (t, x) ∈ R2 : (t, x − αt) ∈ R .
 
R
e ∈ C 2 (R)
(1) (a) Define H e x) = αβ −2 H(αt, βx). Then H
e by H(t, e satisfies (3.25) on R.
e
2 b
(b) Define H ∈ C (R) by H(t, x) = H(t, x − αt). Then, H satisfies (3.25) on R.
b b b b
e ∈ C 2 (R)
(2) (a) Define G e y) = (αβ)−1/2 G(αt, βy). Then G
e by G(t, e satisfies (1.2) on R.
e
26 AMOL AGGARWAL AND JIAOYANG HUANG

b ∈ C 2 (R)
(b) Define G b by G(t,
b y) = G(t, y) + αt. Then, G
b satisfies (1.2) on R.

Proof. We only show the first part of the first statement of the lemma, as the proofs of the remaining
ones are entirely analogous. To that end, fix (e e) ∈ R
t, x e and set (t, x) = (αe x) ∈ R. Then,
t, βe
∂t H(
e e e) = α2 β −2 ∂t H(t, x);
t, x ∂x H(
e e e) = αβ −1 ∂x H(t, x);
t, x
∂tt H(
e e e) = α3 β −2 ∂tt H(t, x);
t, x ∂tx H(
e e e) = α2 β −1 ∂tx H(t, x);
t, x ∂xx H(t, x
e e e) = α∂xx H(t, x).
It follows from (3.25) and (3.26) that
X X
e) = α5 β −4
 
bij ∇H( t, x
e e e) ∂i ∂j H(t, x
e e bij ∇H(t, x) ∂i ∂j H(t, x) = 0,
i,j∈{t,x} i,j∈{t,x}

verifying that H
e satisfies (3.25) on R.
e □

Let us mention that, while the scalings provided by the α = β cases of parts (1a) and (2a) of
Lemma 3.21 are satisfied by any equation of the general form (3.25) (only assuming that the bij
and dij are measurable), those provided by general (α, β) are special to the explicit forms of the
coefficients (3.26) and (1.3).
Remark 3.22. Let us extend the above statements to the case of general parameters a < b and
A > 0. Adopt the notation and assumptions of Remark 3.14. Since e⋆
−1 −1/2 −1/2
 H satisfies (3.25), the second
statement of (3.5) and the (α, β) = (b − a) , A (b − a) case of part (1a) of Lemma 3.21
e⋆
implies that H ⋆ also satisfies (3.25) on the liquid region Ω associated with (ϱ⋆t ). Similarly, since G
satisfies (1.2), the third statement of (3.5) and the (α, β) = (b − a)−1 , A−1 ) case of part (2a) of
Lemma 3.21 implies that G⋆ satisfies (1.2) on the inverted liquid region Ωinv associated with (ϱ⋆t ).
We conclude this section by mentioning that, as stated, the results of Section 3.2 and Section 3.3
largely imposed the assumption (3.21) (due to (3.16)). However, the following lemma indicates that
many of these results continue to hold if we only assume that the boundary measures µa and µb
are compactly supported (without imposing that they admit a bounded density, as in (3.21)). The
proof is given in Appendix C.2 below, by combining the results of Section 3.2 and Section 3.3 with
Lemma 3.7 and Lemma 3.3 (to verify (3.21) in the interior of the domain).
Lemma 3.23. Adopt the notation and assumptions of Lemma 3.1, and assume that µa and µb
are compactly supported. Denote the density process associated with µ by (ϱ⋆t )t∈(a,b) ; denote the
associated height and inverted height functions by H ⋆ : [a, b] × R → R and G⋆ : [a, b] × [0, A] → R,
respectively; and denote the associated liquid and inverted liquid regions by Ω ⊂ (a, b) × R and
Ωinv ⊆ (a, b) × (0, A), respectively. Then, the following statements hold.
(1) The function H ⋆ (t, x) is smooth for (t, x) ∈ Ω; moreover, G⋆ (t, y) is smooth for (t, y) ∈ Ωinv .
(2) Defining u⋆ : Ω → R as in Remark 3.15 and the complex slope f : Ω → H as in (3.18), f
satisfies (3.19) on Ω.
(3) The function H ⋆ satisfies (3.25) on Ω, and G⋆ satisfies (1.2) on Ωinv .

4. Proof of the Concentration Estimate


In this section we reduce Theorem 1.4 to Proposition 4.4 below, which is an inductive statement
that bounds the height function associated with the family x of non-intersecting Brownian motion
by a sequence of barrier functions.
LOCAL STATISTICS AND CONCENTRATION FOR NON-INTERSECTING BROWNIAN BRIDGES 27

4.1. Proof of Theorem 1.4. The proof of Theorem 1.4 will proceed by first establishing the
following variant of it for the associated height function instead of for the path trajectories xj (t).
We begin with the following definition for a certain class of open sets.
Definition 4.1. For any real numbers a < b and functions f, g : [a, b] → R with f < g, we define
the open set P = Pf ;g ⊂ R2 by

P = (t, x) ∈ (a, b) × R : f (t) < x < g(t) .
We call any open set of the form Pf ;g for some f, g : [a, b] → R with f < g a strip domain. The
width of Pf ;g is defined to be b − a.
Observe in particular that the non-intersecting Brownian motions from Definition 1.1 all lie
within the strip domain Pf ;g .
The following proposition, which quickly implies Theorem 1.4, is established in Section 4.2 below.
In what follows we recall that any function G : [a, b] × [0, 1] → R, that is nonincreasing in its second
coordinate, is associatedwith a unique height function
H : [a, b] × R → [0, 1] through (3.4), namely,
by setting H(t, x) = inf y ∈ [0, 1] : G(t, y) ≥ x . We also recall the height function Hx associated
with a family x of non-intersecting curves from (3.6).
Proposition 4.2. Adopt the notation and assumptions of Theorem 1.4; assume that κ = 0 and
that G(0, 0) = 0. There exist constants c = c(ε, δ, B0 , m) > 0 and C = C(ε, δ, B0 , m) > 1 such that
the following holds. Letting H : [0, L−1 ] × R → R denote the height function associated with G, we
have
 
2
P sup n−1 Hx (z) − H(z) > Cn2/m+δ−1 < Ce−c(log n) .

z∈Pf ;g

In the below, we will use the fact that under the notation of Proposition 4.2 we have
(4.1) −ε−1 < ∂x H(t, x) < −ε, for each (t, x) ∈ Pf ;g ,
which follows from (3.4) and the fact that G ∈ Admε (R) (recall Assumption 1.3). Also observe
that there exists a constant B = B(ε, B0 , m) > 1 such that
(4.2) ∥H∥C m+1 (Pf ;g ) < B,

which follows from (3.4) and the facts that G ∈ Admε (R) (so H t, G(t, y) = y for each (t, y) such
that G(t, y) ∈ Pf ;g ) and that ∥G∥C m+1 (R) ≤ B0 .

Proof of Theorem 1.4. We may assume by shifting in what follows that G(0, 0) = 0. Let us first
address the case when κ = 0. By Proposition 4.2, it suffices to show for any real number C > 1
the inclusions of events
( )
−1 −1 2/m+δ−1

max sup xj (t) − G(t, jn ) ≤ Cε n
j∈J1,nK t∈[0,1/L]
(4.3) ( )
−1 x 2/m+δ−1
⊆ sup n H (z) − H(z) ≤ Cn
.
z∈Pf ;g

To that end, we begin by restricting to the event on the right side of (4.3). Fix t0 ∈ (0, L−1 ), and
apply that event twice, first with z = (t0 , y1 ) and then with z = (t0 , y2 ), where y1 = G t0 , jn−1 +
28 AMOL AGGARWAL AND JIAOYANG HUANG

−1
 
Cn2/m+δ−1 and  y2 = G t0 , jn − Cn
2/m+δ−1
. Since (3.4) and the fact that G ∈ Adm(R) imply
H t0 , G(t0 , y) = y for each y ∈ (0, 1), this yields
Hx t0 , G(t0 , jn−1 − Cn2/m+δ−1 ) ≤ j ≤ Hx t0 , G(t0 , jn−1 + Cn2/m+δ−1 ) ,
 

and hence
G t0 , jn−1 + Cn2/m+δ−1 ≤ xj (t0 ) ≤ G t0 , jn−1 − Cn2/m+δ−1 .
 

Together with the fact that G ∈ Admε (R), this yields


G(t0 , jn−1 ) − Cε−1 n2/m+δ−1 ≤ xj (t0 ) ≤ G(t0 , jn−1 ) + Cε−1 n2/m+δ−1 ,
confirming (4.3). This establishes the theorem if κ = 0.
If κ ̸= 0, then define the n-tuples u− , u+ , v − , v + ∈ Wn and the functions f − , f + , g − , g + :
[0, L−1 ] → R by setting

j = G(0, jn
−1
) ± κ; vj± = G(L−1 ; jn−1 ) ± κ; f ± (s) = f (s) ± κ; g ± (s) = g(s) ± κ,
for each index ± ∈ {+, −} and pair (j, s) ∈ J1, nK × [0, L−1 ]. Sample non-intersecting Brown-
ian bridges x− = (x− − −
1 , x2 , . . . , xn ) ∈ J1, nK × C [0, L
−1
] and x+ = (x+ + +
1 , x2 , . . . , xn ) ∈ J1, nK ×
− − + +
C [0, L−1 ] under the measures Qu ;v u ;v −
 +
f − ;g − and Qf + ;g + , respectively. Observe that u ≤ u ≤ u and
− + − + − +
v ≤ v ≤ v by (1.4), and that f ≤ f ≤ f and g ≤ g ≤ g . Thus, by Lemma 2.3, we may
couple (x− , x, x+ ) such that
(4.4) x− +
j (s) ≤ xj (s) ≤ xj (s), for each (j, s) ∈ J1, nK × [0, L−1 ].
Then, Theorem 1.4 applies to both x− and x+ , with the G there given by G− = G − κ and
+
G = G + κ here, respectively. It yields a constant c = c(ε, δ, B0 , m) > 0 such that, on an event
2
of probability at least 1 − c−1 e−c(log n) , we have x− −
j (s) ≥ G (s, jn
−1
) − n2/m+δ−1 = G(s, jn−1 ) −
2/m+δ−1 + −1 2/m+δ−1
κ−n and similarly xj ≤ G(s, jn ) + κ + n . This, with (4.4) and a union bound,
yields the theorem. □

To establish Proposition 4.2 , we make use of the following initial estimate that provides a high-
probability 21 -Hölder bound, with associated constant nδ/2 , for the path trajectories xj (t) from
Proposition 4.2 .
Lemma 4.3. Adopt the notation and assumptions of Theorem 1.4; assume that κ = 0. There
exists a constant C = C(δ, B0 ) > 1 such that
 
xj (t + s) − xj (t) ≤ nδ/2 s1/2 + n−5 ≤ Ce−n .

P max sup
j∈J1,nK 0≤t<t+s≤1/L

Proof. First observe by Assumption 1.3 that f ′ (s) + g ′ (s) ≤ 2 G − G(0, 0) C m+1 (R) ≤ 2B0 , for
each s ∈ (0, L−1 ). Thus we may apply the second part of Lemma 2.7 with the (A, B, T ) there equal
to (2B0 , C1 , L−1 ) here, for some sufficiently large constant C1 > 1. Since L ≤ nδ , this yields some
constant C2 > 1 such that
(4.5)
" n  #
xj (t + s) − xj (t) − sL(vj − uj ) ≥ C2 s1/2 | log 9s−1 nδ | + B0 2
[ [
≤ C2 e−n .

P
j=1 a≤t<t+s≤b
LOCAL STATISTICS AND CONCENTRATION FOR NON-INTERSECTING BROWNIAN BRIDGES 29

1
Since δ < 80 , by Assumption 1.3, there is a constant C3 = C3 (B0 ) > 1 such that for n > C3 we
have
2  2 
C2 s1/2 log |9s−1 nδ | + B0 ≤ C2 B02 s1/2 (log n21 + 1)2 · 1s≥n−20 + log |sn| + 1 · 1s≤n−20
≤ nδ/4 s1/2 + n−5 .
−1

Inserting
this into (4.5), and using the facts that s ≤ L ≤ B 0 , that |v j − uj |L ≤ 2 G −
G(0, 0) C m+1 (R) ≤ 2B0 (by Assumption 1.3), and that n is sufficiently large, yields the lemma. □

4.2. Proof of Proposition 4.2. In this section we establish Proposition 4.2 assuming Proposi-
tion 4.4 below. Define the real numbers ρ, ω > 0; fix a real number ζ; and define the function
ψ : R2 → R so that
(4.6) ρ = nδ−1/m ; ω = n2/m+δ−1 ; ζ > 20000(B 5 + ε−5 ); ψ(t, x) = Υ − eζt − eζx ,
where Υ > 0 is chosen so that inf z∈Pf ;g ψ(z) = 2. Set J = ⌊3ω −1 ⌋ − 1, and let K be the
smallest integer so that (1 − Kρ2 n−δ ) supz∈Pf ;g ψ(z) ≤ 14 . Analogously to in [52, Equation (52)],
for any integers j ∈ J0, JK and real number (typically an integer) k ∈ [0, K], define the functions

φ+j;k , φj;k : Pf ;g → R by for z ∈ Pf ;g setting
2 −δ
φ+
j;k (z) = H(z) − (j + 1)ω + (1 − kρ n )ωψ(z) + 3;
(4.7)
φ− 2 −δ
j;k (z) = H(z) + (j + 1)ω − (1 − kρ n )ωψ(z) − 3,
and by for z ∈ ∂Pf ;g setting φ− +
j;k (z) = H(z) and φj;k (z) = H(z). Further define the events
−1 x
Ω− H (z) < φ−
 −1 x
Ω+ H (z) > φ+

j;k (z) = n j;k (z) ; j;k (z) = n j;k (z) .

The following proposition indicates that if H lies below (or above) φ+j;k−1 (or φ;k−1 , respectively),

then with high probability it also lies below (or above) φ+ j;k (or φj;k , respectively); it will be
established in Section 6.2 below. Here, we denote the complement of any event E by E∁ .
Proposition 4.4. Adopt the notation and assumptions of Proposition 4.2. For any index ± ∈
{+, −}, integer pair (j, k) ∈ J0, JK × [0, K], and point z ∈ Pf ;g , we have
" #
\ 2
± ±
P Ωj;k (z) ∩ Ωj;k−1 (w) ≤ Ce−c(log n) .

w∈Pf ;g

Given Proposition 4.4, we can quickly establish Proposition 4.2.


Proof of Proposition 4.2. It suffices to show that
" #
[  2
−1 x 2/m+δ−1
< Ce−c(log n) ;

P n H (z) ≥ H(z) + 2n
z∈Pf ;g
(4.8) " #
[  2
n−1 Hx (z) ≤ H(z) − 2n 2/m+δ−1
< Ce−c(log n) .

P
z∈Pf ;g

We only establish the first bound in (4.8), for the proof of the latter is entirely analogous. To that
end, for any integer j ∈ J0, JK, define the function Φj : Pf ;g → R by
Φj (z) = H(z) − jω + 3.
30 AMOL AGGARWAL AND JIAOYANG HUANG

and for any z ∈ Pf ;g define the event


Ωj (z) = n−1 Hx (z) > Φj (z) .


Then, observe from (4.7) and (4.6) that


(4.9) φ+ +
j;K (z) ≤ Φj (z) ≤ Φj−1 (z) ≤ φj;0 (z),

where the first bound follows from the fact that (1 − Kρ2 n−δ ) supz∈Pf ;g ψ(z) ≤ 41 ; the second from
the definition of Φj ; and the third from the fact that inf z∈Pf ;g ψ(z) = 2.
Define the set S ⊂ Pf ;g consisting of all points (t, x) ∈ Pf ;g such that t ∈ n−15 ·Z and x = G(t, y),
for some y ∈ n−1 · Z; then, |S| ≤ n20 . Taking a union over z ∈ S of Proposition 4.4 yields constants
c1 = c1 (ε, δ, B0 , m) > 0 and C1 = C1 (ε, δ, B0 , m) > 1 satisfying
 
−c1 (log n)2
\
Ω+j;k−1 (w) \ Ej;k ≤ C1 e

(4.10) P  ,
w∈Pf ;g

where
\ \
Ej;k = Ω+ ′ ∁
j;k (z ) ∩ Ω+ ∁
j;k−1 (w) .
z ′ ∈S w∈Pf ;g

By Lemma 4.3, there also exists a constant C2 = C2 (δ, B0 ) > 1 such that
(4.11)
( )
−n
F= δ/2 1/2 −5

P[F] ≥ 1 − C2 e , where max sup xj (t + s) − xj (t) ≤ n s + n .
j∈J1,nK 0<t<t+s≤1/L

Next, for any j ∈ J0, JK, k ∈ [0, K], s ∈ [0, L−1 ], and r ≥ 0, let
+
(r; s) = sup γ ∈ R : φ+

γj;k j;k (s, γ) ≥ r .

Fix (i, s) ∈ J1, nK × [0, L−1 ], and let s′ ∈ [0, L−1 ] be such that s′ , G(s′ , in−1 ) ∈ S and |s − s′ | is


minimal; then, |s − s′ | ≤ n−15 . On the event Ej;k ∩ F, we have for sufficiently large n that
xi (s) ≤ xi (s′ ) + 2n−5 ≤ γj;k
+
(in−1 ; s′ ) + 2n−5 < γj;k
+
(in−1 ; s) + 4n−5 < γj;k
+
(in−1 + n−4 ; s).
Here, the first inequality follows from restricting to F; the second follows from restricting to Ej;k ;
and the third and fourth follow from the facts that φ+ +
j;k ∈ Admε/2 (Pf ;g ), that ∥φj;k ∥C 1 (Pf ;g ) ≤ C3
for some constant C3 = C3 (ε, δ, B, m) > 1, and that n sufficiently large (where these bounds follow
from (4.7), since H ∈ Admε (Pf ;g ) by (4.1), ∥H∥C 1 (Pf ;g ) ≤ B by (4.2), ∥ψ∥C 1 (Pf ;g ) ≤ 10ζeζ , and
ω = n2/m+δ−1 ). Since this holds for each (i, s) ∈ J1, nK × [0, L−1 ], it follows on Ej;k ∩ F that
n−1 Hx (z) ≤ φ+
j;k (z) + n
−4
< φ+
j;k−1/2 (z), for each z ∈ Pf ;g , on the event Ej;k ∩ F.

Applying this bound (decreasing k by 21 at each point) 2K + 2 ≤ 2n times, and using (4.10) and
(4.11), yields constants c2 = c2 (ε, δ, B0 , m) > 0 and C4 = C4 (ε, δ, B0 , m) > 1 such that
" #
[  \ 2
−1 x +
Ωj;0 (w) ≤ C4 e−c2 (log n) .
+ ∁

P n H (z) > φj;K (z) ∩
z∈Pf ;g w∈Pf ;g
LOCAL STATISTICS AND CONCENTRATION FOR NON-INTERSECTING BROWNIAN BRIDGES 31

By (4.9), we deduce
" #
[ \ 3
P Ωj (z) ∩ ∁
Ωj−1 (z) ≤ C4 e−c2 (log n) ,
z∈Pf ;g w∈Pf ;g

which upon applying J ≤ n times implies


" #
[ \ 2
P ΩJ (z) ∩ Ω0 (z)∁ ≤ C4 ne−c2 (log n) .
z∈Pf ;g w∈Pf ;g

With a union bound and the facts that ΦJ (z) ≤ H(z) + 2ω (since J = ⌊3ω −1 ⌋ − 1) and Φ0 (z) ≤
H(z) + 3, this yields
" #
[ 
−1 x 2/m+δ−1

P n H (z) ≥ H(z) + 2n
z∈Pf ;g
" #
[  2
<P n H (z) ≥ H(z) + 3} + C4 ne−c2 (log n) .
−1 x

z∈Pf ;g

Since n−1 Hx (z) ≤ 1 ≤ H(z) + 3 holds deterministically, this yields the first bound in (4.8) and thus
the proposition. □

5. The Local Behaviors of Height Profiles


In this section we explain how to approximate the first several derivatives (which we refer to
as a “local profile”) of a limit shape by another one, associated with a family of non-intersecting
Brownian bridges with known concentration bounds.

5.1. Local Profiles. In this section we introduce the notion of a local profile, which constitutes
information about the first several derivatives of a function. To that end, for any integer k ≥ 1;
open subset R ⊂ R2 ; and function F : C k (R), and point z ∈ R, define the k-th order local profile
(k)
qF (z) ∈ Rk+1 by
(k)
qF (z) = ∂xk F (z), ∂t ∂xk−1 F (z), . . . , ∂tk F (z) .

(5.1)
(m) (k) 
For any integer m ≥ 1, the full-m local profile is the sequence of vectors QF (z) = qF (z) k∈J1,mK .
(m)
We next describe two properties that a full-m local profile QF must satisfy, if some admissible
F solves the equation (3.25) on some open subset R ⊆ R2 . The first is that the second coordinate
(1)
of qF must be negative, due to the admissibility of F . The second is that F must satisfy (3.25),
(1) (2)
which imposes a consistency relation between qF and qF . More generally, applying the differential
(1) (2) (i+j+2) 
operator ∂ti ∂xj to (3.25) imposes a relation between qF , qF , . . . , qF , for each (i, j) ∈ Z2≥0
such that i + j ≤ m − 2. These admissibility and consistency properties are made precise through
the following definition. Throughout, unless mentioned otherwise, we will denote sequences of m
 (k) (k) (k) 
vectors Q = q (k) J1,mK ∈ R2 × R3 × · · · × Rm+1 by q (k) = q0 , q1 , . . . , qk , for each k ∈ J1, mK.
32 AMOL AGGARWAL AND JIAOYANG HUANG


Definition 5.1. A sequence of m vectors Q = q (k) ∈ R2 × R3 × · · · × Rm+1 is admissible if
(1)
q0 < 0. For any integers i, j ≥ 0, define the function vi,j : R2 × R3 × · · · × Ri+j+3 → R so that
!
(i+j+2)
X
i j
 
vi,j QF (z) = ∂t ∂x bi′ j ′ ∇F (z) ∂i′ ∂j ′ F (z) .
i′ ,j ′ ∈{t,x}

We say that the sequence Q is consistent if it is admissible and


vi,j q (1) , q (2) , . . . , q (i+j+2) = 0, for each (i, j) ∈ Z2≥0 with i + j ≤ m − 2.

(5.2)
Example 5.2. Since
(3)    
v1,0 QF (z) = btt ∇F (z) ∂ttt F (z) + 2btx ∇F (z) ∂ttx F (z) + bxx ∇F (z) ∂txx F (z)
  
+ ∂tt F (z)∂u btt ∇F (z) + ∂tx F (z)∂v btt ∇F (z) ∂tt F (z)
  
+ 2 ∂tt F (z)∂u btx ∇F (z) + ∂tx F (z)∂v btx ∇F (z) ∂tx F (z)
  
+ ∂tt F (z)∂u bxx ∇F (z) + ∂tx F (z)∂v bxx ∇F (z) ∂xx F (z),
 (k) (k) (k) 
we have, for any sequence Q(3) = q (k) k∈J1,3K with q (k) = q0 , q1 , . . . , qk for each k ∈ J1, 3K,
 (3)  (3)  (3)   (2)
(2) (2)
v1,0 Q(3) = btt q (1) q3 + 2btx q (1) q2 + bxx q (1) q1 + q2 ∂u btt q (1) + q1 ∂v btt q (1) q2
 
  (2)  (2)  (2)
(2) (2) (2)
+ 2 q2 ∂u btx q (1) + q1 ∂v btx q (1) q1 + q2 ∂u bxx q (1) + q1 ∂v bxx q (1) q0 .
 

(m)
Remark 5.3. For any function F ∈ C 2 (R) ∩ Adm(R) satisfying (3.25), the sequence QF (z) is
consistent, for each integer m ≥ 1 and point z ∈ R; this follows from applying ∂ti ∂xj to (3.25) (for
each (i, j) ∈ Z2≥0 with i + j ≤ m + 2).
Observe that consistency imposes m

2 relations (one for each (i, j) ∈ Z≥0 with i + j ≤ m − 2)
between the 2 − 1 coordinates of Q; this suggests that there are m+2
m+2
−1− m
  
2 2 = 2m “free
parameters” describing a consistent Q. The following lemma indicates this to indeed be the case,
(k) (k) 
taking these 2m parameters to be the first two coordinates q0 , q1 k∈J1,mK of each vector in Q.
(k) (k)  (1)
Lemma 5.4. Fix an integer m ≥ 1 and m pairs of real numbers q0 , q1 ∈ (R2 )m with q0 < 0.

(1) There is a unique consistent sequence of m vectors Q = q (k) ∈ R2 × R3 × · · · × Rm+1 so
(k) (k) 
that, for each k ∈ J1, mK, the first two coordinates of q (k) are q0 , q1 .
(2) For any real numbers ε ∈ (0, 1) and B > 1, there exists a constant C = C(ε, B, m) > 1 so
that, if
 (k) 
(1) (k)
(5.3) q0 ≤ −ε, and max q0 + q1 ≤ B,
1≤k≤m
(k)
then the consistent sequence Q sastisfies qj ≤ C for each k ∈ J1, mK and j ∈ J0, kK.
Proof. Observe for any (i, j) ∈ Z2>0 that
!
X  ′ ′  
∂ti ∂xj ∇F (z) ∂i′ ∂j ′ F (z) = btt ∇F (z) ∂ti+2 ∂xj F (z) + ui,j ∂ti ∂xj F (z) i′ ,j ′ ,
 
bi′ j ′
i′ ,j ′ ∈{t,x}
LOCAL STATISTICS AND CONCENTRATION FOR NON-INTERSECTING BROWNIAN BRIDGES 33

for some function ui,j that depends polynomially (as the bij (u, v) from (3.26) are polynomials in
′ ′
(u, v)) only on those derivatives ∂ti ∂xj F (z) such that either  i′ ≤ i + 1 or i′ + j ′ < i + j + 2. As such,
(1) (2) (m)
for any sequence of m vectors Q = q , q , . . . , q ∈ R2 × R3 × · · · Rm+1 , we have
(k′ ) 
 
(i+j+2)
vi,j q (1) , q (2) , . . . , q (i+j+2) = qi+2 btt q (1) + ui,j qi′ i′ ,k ,
 
(5.4)
(k′ )
where again ui,j is a polynomial in those qi′ with either i′ ≤ i + 1 or k ′ ≤ i + j + 1. Thus,
(k′ ) 
 
ui,j qi′ i′ ,k′
(i+j+2)
vi,j q (1) , q (2) , . . . , q (i+j+2) = 0,

(5.5) if and only if qi+2 =−  .
btt q (1)
(1) 
Hence since q0 < 0, meaning btt q (1) > 0, the consistency of Q provides a unique solution for
(i+j+2) (k′ ) 
qi+2 , in terms of all qi′ such that either i′ ≤ i + 1 or k ′ ≤ i + j + 1. In particular, given the
(k) (k) 
pairs q0 , q1 , this by induction on the lexicographic pair (i + j + 2, i + 2) determines a unique
consistent solution for the remaining entries of Q, establishing the first statement of the lemma.
 (1)
To establish the second, observe from (3.26) that btt q (1) ≥ ε2 , as q0 ≤ −ε. Thus, since the
ui,j are polynomials with coefficients and degrees only dependent on i, j ∈ J1, mK, it follows by (5.5)
(5.3), and induction on the lexicographic pair (i + j + 2, i + 2) ∈ J2, mK2 that for each (i, j) ∈ J0, mK2
(i+j+2)
there exists a constant Ci,j = Ci,j (ε, B) > 1 such that qi+2 ≤ Ci,j . This (together with the
(k) (k)
fact from (5.3) that q0 + q1
≤ B for each k ∈ J1, mK) yields the second statement of the
lemma, with C = maxi+j≤m−2 Ci,j + B. □

Definition 5.5. Adopting the notation of Lemma 5.4, we call Q = q (k) the consistent extension
(k) (k) 
of the sequence of m pairs q0 , q1 ∈ (R2 )m .
We conclude this section with the following two lemmas. The first is a continuity bound, saying
that if two sequences of real pairs are close, then their consistent extensions are as well. The second
implies that, if Q
e is a sequence that “almost satisfies” the consistency relations (5.2), then there is
a consistent sequence Q close to it.
Lemma 5.6. For any integer m ≥ 1 and real numbers ε ∈ (0, 1) and B > 1, there exists a constant
(k) (k)  (k) (k) 
C = C(ε, B, m) > 1 such that the following holds. Let q0 , q1 ∈ (R2 )m and qe0 , qe1 ∈ (R2 )m
be two sequences of m real number pairs such that
 (1) (1)  (k) (k) (k) 
(k)
max q0 , qe0 } ≤ −ε; max q0 + q1 + qe0 + qe1 ≤ B.
1≤k≤m

e = qe(k) denote the consistent extensions of q (k) , q (k) and qe(k) , qe(k) ,
(k)
   
Letting Q = q and Q 0 1 0 1
respectively, we have
k 
(k) X (k′ ) (k′ ) (k′ ) (k′ )

q − qe(k) ≤ C

q
0 − qe0 + q1 − qe1 , for each k ∈ J1, mK.
k′ =1

Proof. Throughout, we recall the notation from the proof of Lemma 5.4. First observe from the
second part of Lemma 5.4 that there exists a constant C1 = C1 (ε, B, m) > 1 such that
m 
X (k) (k) 
(5.6) q + qe ≤ C1 .
k=1
34 AMOL AGGARWAL AND JIAOYANG HUANG

 (1) (1)
Next, due to (5.6); the definition (3.26) of btt ; and the fact that max q0 , qe0 ≤ −ε, there exists
a constant C2 = C2 (ε, m) > 1 such that

1 1 ≤ C2 q (1) − qe(1) ;
1 1
(5.7)
b q (1)  −   +  ≤ C2 .
tt btt qe(1) btt q (1) btt qe(1)

Moreover, by (5.6); the fact that the ui,j (satisfying (5.5)) are polynomials that only depend on the
(k′ )
qi′ with either i′ ≤ i + 1 or k ′ ≤ i + j + 1; and the fact that the coefficients and degrees of ui,j
only depend on i, j ∈ J1, mK, there exists a constant C3 = C3 (ε, B, m) > 1 such that
(5.8)
i+1 i+j+2 i+2 i+j+1
 !
′ 
  ′ 
 X (k′ ) ′ X (k′ ) ′
ui,j q (k′ ) ′ ′ − ui,j qe(k′ ) ′ ′ ≤ C3 (k ) (k )
X X
q ′ − qe ′ + q ′ − qe ′ ;
i i ,k i i ,k i i i i
i′ =0 k′ =1 i′ =0 k′ =1
   
′  ′ 
ui,j q (k′ ) ′ ′ + ui,j qe(k′ ) ′ ′ ≤ C3 .

i i ,k i i ,k

Thus, (5.5), (5.7), and (5.8) together yield a constant C4 = C4 (ε, B, m) > 1 so that
i+1 i+j+2 i+2 i+j+1
!
(i+j+2) (i+j+2) X X (k′ ) (k ′
)
X X (k′ ) (k ′
)
q
i+2 − qe i+2
≤ C4 q ′ − qe ′ +
i i
q ′ − qe ′ .
i i
i′ =0 k′ =1 i′ =0 k′ =1

Induction on the lexicographic pair (i + j + 2, i + 2) ∈ J2, mK2 then gives a constant Ci,j =
Ci,j (ε, B, m) > 1 for any (i, j) ∈ J0, mK2 with i + j + 2 ≤ m such that
i+j+2
X  (k′ ) (k′ ) (k′ ) (k′ )
(i+j+2) 
(i+j+2)
q
i+2 − qe i+2 ≤ Ci,j q − qe + q
0 0 1− qe . 1
k′ =1

Summing over all such (i, j) with i + j = k − 2, for any k ∈ J1, mK, we deduce the lemma. □

Lemma 5.7. For any integer m ≥ 1 and real numbers ε ∈ (0, 1) and B > 1, there exists a constant
(k) (k) 
C = C(ε, B, m) > 1 such that the following holds. Let q0 , q1 ∈ (R2 )m denote a sequence of
(k)

m pairs of real numbers satisfying (5.3), and let Q = q denote its consistent extension. Let
(k)
 2 3 m+1 (k) (k) (k) (k)
Q = qe
e ∈ R × R × ··· × R be such that qe0 = q0 and qe1 = q1 for each k ∈ J1, mK. If
(k)
qe ≤ B for each k ∈ J1, mK and j ∈ J0, kK, then
j

k h−2
X
(k) X 
q − qe(k) ≤ C vi,h−i−2 qe(1) , qe(2) , . . . , qe(h) ,

for each k ∈ J1, mK.
h=2 i=0

Proof. Throughout, we recall the notation from the proof of Lemma 5.4. Following the beginning
of the proof of Lemma 5.6, we deduce the existence of a constant C1 = C1 (ε, B, m) > 1 such that
(5.6) holds (by the assumption of this lemma, as well as the second part of Lemma 5.4); a constant
C2 = C2 (ε, m) > 1 such that (5.7) holds (due to (5.6); the definition (3.26) of btt ; and the fact that
(1) (1)
q0 = qe0 ≤ −ε); and a constant C3 = C3 (ε, B, m) > 1 such that (5.8) holds (by (5.6); the fact
(k′ )
that the ui,j are polynomials that only depend on the qi′ with either i′ ≤ i + 1 or k ′ ≤ i + j + 1;
and the fact the coefficients and degrees of ui,j only depend on i, j ∈ J1, mK).
LOCAL STATISTICS AND CONCENTRATION FOR NON-INTERSECTING BROWNIAN BRIDGES 35

Next, by (5.4), we have


(k′ )  (k′ ) 
   
ui,j qi′ i′ ,k′ u q
(1) e(2) (i+j+2)

(i+j+2) (i+j+2) vi,j q ,q ,...,q
e e i,j e i ′ ′
i ,k ′
qi+2 =−  ; qei+2 =  −  .
btt q (1) btt qe(1) btt qe(1)
This, together with (5.7) and (5.8), yields a constant C4 = C4 (ε, m) > 1 so that
i+1 i+j+2
X (k′ ) i+2 i+j+1
(i+j+2) ′ X (k′ ) ′
(i+j+2) q ′ − qe(k′ ) + q ′ − qe(k′ )
X X
q
i+2 − qei+2 ≤ C4 i i i i
i′ =0 k′ =1 i′ =0 k′ =1
!

(1) (2) (i+j+2)
+ vi,j qe , qe , . . . , qe .

(k) (k) 
Induction on the lexicographic pair (i+j+2, i+2) ∈ J2, mK2 (together with the fact that q0 , q1 =
(k) (k) 
qe0 , qe1 , for each k ∈ J1, mK) then gives a constant Ci,j = Ci,j (ε, B, m) > 1 for any (i, j) ∈ J0, mK2
with i + j + 2 ≤ m such that
i+j+2 i+2
(i+j+2) (i+j+2)
X X 
q − q ≤ C vi′ ,h−i′ −2 qe(1) , qe(2) , . . . , qe(h) .

i+2 ei+2 i,j
h=2 i′ =0

Summing over all such (i, j) with i + j = k − 2, for any k ∈ J2, mK (recall q (1) = qe(1) , addressing
the case k = 1), we deduce the lemma. □
5.2. Concentrating Profiles. In this section we discuss families of (deterministic) height profiles
that are limit shapes for systems of non-intersecting Brownian bridges that admit a (nearly) optimal
concentration bound. They are given by the below definition.
Definition 5.8. Fix real numbers a < b and two functions f, g : [a, b] → R with f < g. Denote
the strip domain S = Pf ;g ⊂ R2 , and let H : S → R be a function. Given real numbers C > 1 and
1

c, δ ∈ 0, 2 , we say that H is (C; c; δ)-concentrating if it is satisfies the equation (3.25) for each
z ∈ S, and the following holds for any integer n ≥ 1. There exists an integer n′ ∈ [cn, n]; starting
data u ∈ Wn′ ; and ending data v ∈ Wn′ on S so that, sampling n′ non-intersecting Brownian
bridges y = (y1 , y2 , . . . , yn′ ) ∈ J1, n′ K × C [a, b] from the measure Qfu;v ;g (n
−1
) and denoting the
y
associated height function by H (recall (3.6)), we have
" #
−1 y 2
δ−1
≤ Ce−c(log n) .

P sup n H (z) − H(z) ≥ n

z∈S

We will be interested in height functions that are (C; c; δ)-concentrating with small δ > 0. While
it is plausible that essentially any solution of (3.25) on R has this property, it is not transparent to
us how to show this directly. We instead show the following result stating that, if Q ∈ R2 × R3 ×
· · · × Rm+1 is an admissible sequence of m vectors in the sense of Definition 5.1, then (for small
δ) there exists a (C; c; δ)-concentrating height function H whose full-m local profile at z = (0, 0) is
given by Q; see Figure 5 for a depiction. It will be established in Section 7.2 below.
Proposition 5.9. For any integer m ≥ 1 and real numbers ε ∈ (0, 1) and B > 1, there exist
constants ν = ν(ε, B, m) ∈ (0, 1), c = c(ε, B, m) > 0, and C = C(ε, B, m) > 1 such that the
(k) (k) 
following holds. Fix a sequence of m pairs q0 , q1 ∈ (R2 )m satisfying (5.3); denote its consistent
extension by Q. There is a strip domain S ⊂ R2 of width ν, such that [0, ν] × [−ν, ν] ⊆ S ⊆ [0, ν] ×
36 AMOL AGGARWAL AND JIAOYANG HUANG

Figure 5. Shown above is a depiction of Proposition 5.9 and Corollary 5.11.

[−Cν, Cν], and a function H : S → R, such that ∥H∥C m+1 (S) ≤ C and supz∈S ∂x H(z) ≤ −C −1 ,
which satisfies the following two properties.
(m)
(1) The full-m local profile of H satisfies limz→(0,0) QH (z) = Q.
(2) For any real number δ ∈ (0, 1), there exists a constant C1 = C1 (ε, δ, B, m) > 1 such that H
is (C1 ; c; δ)-concentrating.

Observe that Proposition 5.9 fixes the local profile of H at the point (0, 0) on the boundary of
∂R . However, it will be useful to (at least approximately) fix the local profile of H at some point
in the interior of R, to which end we have the following lemma and corollary; see Figure 5 for a
depiction of the latter.

Lemma 5.10. For any integer m ≥ 1 and real numbers ε ∈ (0, 1), ν ∈ (0, 1), and B > 1, there
exist constants c = c(ε, B, m) ∈ 0, ν2 and C = C(ε, B, m) > 1 such that the following holds. Fix a
 (k) (k) (k) 
consistent sequence Q = q (k) ∈ R2 ×R3 ×· · ·×Rm+1 ; denoting q (k) = q0 , q1 , . . . , qk ∈ Rk+1
for each k ∈ J1, mK, assume that (5.3) holds. Fix a real number ρ ∈ (0, c) and set

m−k m−k
(k)
X (−ρ)i (k+i) (k)
X (−ρ)i (k+i)
(5.9) r0 = q , and r1 = q , for each k ∈ J1, mK.
i=0
i! i i=0
i! i+1
LOCAL STATISTICS AND CONCENTRATION FOR NON-INTERSECTING BROWNIAN BRIDGES 37

(k) (k) 
Let R ∈ R2 ×R3 ×· · ·×Rm+1 denote the consistent extension of the sequence of m pairs r0 , r1 ∈
(R2 )m . If there exists a function H : [0, ν] × [−ν, ν] → R satisfying (3.25) on [0, ν] × [−ν, ν], with
(m)
(5.10) ∥H∥C m+1 ≤ B, and lim QH (z) = R,
z→(0,0)

(k) (k)
then qH (ρ, 0) − q (k) ≤ Cρm−k+1 for each k ∈ J1, mK, where qH is as defined in (5.1).
(0)
Proof. Throughout this proof, we set q0 = 0. First observe that, since (5.3) holds, Lemma 5.4
yields a constant C1 = C1 (ε, B, m) > 1 such that
(k)
(5.11) max max qj ≤ C1 .
k∈J0,mK j∈J0,kK

(1) (k)
This, together with the fact that q0 ≤ −ε (by (5.3)) implies (by the definition (5.9) of r0 and
(k)
r1 ) that there exist constants c1 = c1 (ε, B, m) > 0 and C2 = C2 (ε, B, m) > 1 such that for
(1) (k)
ρ ∈ (0, c1 ) we have r0 ≤ − 2ε and ri ≤ C2 , for each i ∈ {0, 1} and k ∈ J1, mK. Hence, the second
part of Lemma 5.4 yields a constant C3 = C3 (ε, B, m) > 1 such that
(1) ε (k)
(5.12) r0 ≤ − ; max max rj ≤ C3 .
2 k∈J0,mK j∈J0,kK


Next, define the (not necessarily consistent) sequence of m + 1 vectors R e = re(k) ∈
k∈J0,mK
(k) (k) (k) 
R × R2 × · · · × Rm+1 , where re(k) = re0 , re1 , . . . , rek ∈ Rk+1 for each k ∈ J0, mK, by
m−k ′
(k)
X (−ρ)i (k+i′ )
(5.13) rei = q ′ ,
i′ ! i+i
i′ =0

so in particular
(k) (k) (k) (k)
(5.14) re0 = r0 ; re1 = r1 .
(k) (k)
Let us express the qi through the rei . To that end, define the polynomials A0 , A1 : R2 → R by
k
m X (k) m X
k (k)
X qj X rej
(5.15) A0 (t, x) = tj xk−j ; A1 (t, x) = tj xk−j .
j!(k − j)! j!(k − j)!
k=1 j=0 k=0 j=0

so that
m X j
k X
X (−ρ)j−i (k)
A0 (t − ρ, x) = q ti xk−j
(k − j)!i!(j − i)! j
k=1 j=0 i=0
(5.16) ′
m m−i
X ti xj ′ m−i−j ′ m m−i (i+j ′ )
X X (−ρ)k (i+j ′ +k′ ) X X rei ′
= q ′ = ti xj = A1 (t, x),
i=0 ′
i!j ′ ! k ′ ! i+k i=0 ′
i!j ′!
j =0 k′ =0 j =0

where in the second equality we changed variables (j ′ , k ′ ) = (k − j, j − i) (and used the fact that
(0) (i+j)
q0 = 0), in the third we used the definition (5.13) of rei , and in the fourth we again changed
38 AMOL AGGARWAL AND JIAOYANG HUANG

(0)
variables (i, j ′ ) = (j, k − j). It follows (again since q0 = 0) that
m Xk m Xk X j
X tj xk−j (k) X ρj−i (k)
qj = A0 (t, x) = A1 (t + ρ, x) = rej ti xk−j
j=0
j!(k − j)! j=0 i=0
(k − j)!i!(j − i)!
k=0 k=0

m m−i
X t i xj ′ m−i−j ′
X X ρk (i+j ′ +k′ )
= re ′
i=0 ′
i!j ′ ! k ′ ! i+k
j =0 k′ =0
m k m−k ′
X X tj xk−j X ρi (k+i′ )
= re ′ ,
j!(k − j)! ′ i′ ! j+i
k=0 j ′ =0 i =0
′ ′
where in the fourth equality we again changed variables (j , k ) = (k − j, j − i), and in the fifth we
changed variables (i, j ′ , k ′ ) = (j, k − j, i′ ). Comparing coefficients of tj xk−j , we deduce
m−k ′
(k)
X ρi (k+i′ )
(5.17) qi = re ′ , for each k ∈ J0, mK and i ∈ J0, kK.
i′ ! i+i
i′ =0
 (k) (k) (k) 
Next, set R = r (k) , where r (k) = r0 , r1 , . . . , rk ∈ Rk+1 for each k ∈ J1, mK. Due to
(5.10) and a Taylor expansion, we have
m−i−j
m−i−j

X ρi′ (i+i′ +j) X ρi′ ′

i j i j i+i j
∂t ∂x H(ρ, 0) − r = ∂ ∂ H(ρ, 0) − ∂ ∂ H(0, 0) ≤ Bρm−i−j+1 .

i+i′ t x t x
i′!


i′!
′ i =0 ′ i =0

Hence, by (5.17), it suffices to show


(k)
r − re(k) ≤ Cρm−k+1 ,

(5.18) i i for any k ∈ J1, mK and i ∈ J0, kK.
To that end, define D : R2 → R by, for any (t, x) ∈ R2 , setting
X  X 
D(t, x) = bjk ∇A1 (t, x) ∂j ∂j A1 (t, x) = bj,k ∇A0 (t − ρ, x) ∂j ∂k A0 (t − ρ, x),
j,k∈{t,x} j,k∈{t,x}

were in the last equality we used (5.16). We will use a Taylor expansion to view D as an approxima-
tion of the left side of (3.25), where the H there is A here. For each integer k ∈ J0, mK, set Q(k) =
 
q (1) , q (1) , . . . , q (k) ∈ R2 × R3 × · · · × Rk+1 and R e (k) = re(1) , re(2) , . . . , re(k) ∈ R2 × R3 × · · · × Rk+1 .
(k)
Under this notation, we have from Definition 5.1 (and the fact from (5.15) that QA0 (0, 0) = Q(k)
(k) e (k) for each k ∈ J1, mK), that
and QA1 (0, 0) = R
∂ti ∂xj D(0, 0) = vi,j R
e (i+j+2) ; ∂ti ∂xj D(ρ, 0) = vi,j Q(i+j+2) ,
 

for each (i, j) ∈ R2≥0 with i + j ≤ m − 2. Since Q is consistent, we have from (5.2) and (5.11) (and
the fact that all derivatives of D of order at most 2m are uniformly bounded on compact sets) that
there exists a constant C4 = C4 (ε, B, m) > 1 with
∂ti ∂xj D(ρ, 0) = 0, for any (i, j) ∈ Z2≥0 with i + j ≤ m − 2; max sup ∂ti ∂xj D(z) ≤ C4 .

i,j∈J0,mK |z|≤1

Hence, it follows from a Taylor expansion that



e (i+j+2) = ∂ i ∂ j D(0, 0) ≤ C4 ρm−i−j−1 .

vi,j R

t x

This, with Lemma 5.7, (5.12), (5.11), (5.13), and (5.14), implies (5.18) and thus the lemma. □
LOCAL STATISTICS AND CONCENTRATION FOR NON-INTERSECTING BROWNIAN BRIDGES 39

Corollary 5.11. Adopting the notation and assumptions of Proposition 5.9, there exist constants
ν = ν(ε, B, m) ∈ (0, 1), c = c(ε, B, m) > 0, and C = C(ε, B, m) > 1; a strip domain S ⊂ R2 ; and a
function H : S → R, such that all statements but the first enumerated one in Proposition 5.9 hold.
Moreover, the following variant of that first statement holds, for any real number ρ ∈ (0, c).
(m) (k)  (k)
1. Denoting QH (ρ, 0) = qH (ρ, 0) k∈J1,mK , we have qH (ρ, 0) − q (k) ≤ Cρm−k+1 for all
k ∈ J1, mK.
(k) (k) 
Proof. Define the m real pairs r0 , r1 ∈ (R2 )m from Q through (5.9), and let R denote its
consistent extension. By Proposition 5.9 (applied with the Q there equal to R here) and (5.12),
there exist constants ν = ν(ε, B, m) ∈ (0, 1), c = c(ε, B, m) > 1, and C1 = C1 (ε, B, m) > 1; a strip
domain S ⊂ R2 of width ν, such that [0, ν] × [−ν, ν] ⊆ S ⊆ [0, ν] × [−C1 ν, C1 ν]; and a function H :
(m)
S → R with ∥H∥C m+1 (S) ≤ C1 and inf z∈S ∂x H(z) ≤ −C1−1 , satisfying limz→(0,0) QH (z) = R and
the following property. For any real numbers δ ∈ (0, 1), there exists a constant C2 = C2 (ε, δ, B, m) >
1 such that H is (C2 ; c; δ)-concentrating. This H satisfies the last condition of the corollary, as
(k)
Lemma 5.10 yields a constant C3 = C3 (ε, B, m) > 1 such that qH (ρ, 0) − q (k) ≤ C3 ρm−k+1 , for
each k ∈ J1, mK and sufficiently small ρ. □

6. Proof of Concentration Bounds


In this section we establish Proposition 4.4.

6.1. Local Approximations of H. In this section we will establish various statements needed to
show Proposition 4.4 in Section 6.2 below; we only establish the bound in that statement on Ω+ j;k (z),

as the proof of the bound on Ωj;k (z) is entirely analogous. Throughout, we recall the notation from
Section 4.2; fix (j, k) ∈ J0, JK × [0, K] and a point z ∈ Pf ;g ; and set (recall (4.7))
[
φ = φ+ j;k−1 ; Ω = Ω+j;k−1 ; κ = 1 − (k − 1)ρ2 n−δ ; P = Pf ;g .
w∈P

Throughout this section, we further restrict to the event Ω∁ .


Let us now briefly outline how we will establish Proposition 4.4. We will proceed by locally
comparing the Brownian bridges x in a neighborhood of z with a family y of non-intersecting
Brownian bridges, whose associated normalized height function n−1 Hy is known to concentrate
around its limit shape H. By Corollary 5.11, there are many choices for this limit shape H; we
will make its second derivatives slightly smaller than those of φ (while making its gradient match
that of φ), so as to ensure that φ+ j;k (z) > H(z) while φ|∂Bρ (z) < H|∂Bρ (z) . Since the normalized
−1
height function n H of x is bounded above by φ on the event Ω∁ , comparing x to y using
monotonicity (Lemma 2.3) and applying the concentration bound for y (Corollary 5.11) will yield
n−1 Hx (z) ≤ H(z) < φ+j;k (z) with high probability, establishing Proposition 4.4.
To implement this, we must fix the derivatives of the function H, to which end we require some
additional notation. Recalling the bij from (3.26), define for any (u, v) ∈ R × R<0 the vector
p(u, v) = bxx (u, v), 2btx (u, v), btt (u, v) ∈ R3 .


(2)
Observe under this notation, and recalling qF for any function F ∈ C 2 (P) from (5.1), the constraint
 (2)
(3.25) is equivalent to p ∇F (z) · qF (z) = 0.
Lemma 6.1. The following two statements hold for n sufficiently large.
40 AMOL AGGARWAL AND JIAOYANG HUANG

(1) For any z0 ∈ P, we have


ε
∂x φ(z0 ) ≤ − ; ∥φ∥C m+1 (P) ≤ 2B.
2
(2) For any z0 ∈ P, we have
 (2)
p ∇φ(z0 ) · qH

(z0 ) κω
≤ 2 .
 (2)
p ∇φ(z0 ) · qψ (z0 )

Proof. From the definition (4.7) of φ = φ+


j;k−1 , we have for any z0 = (t, x) ∈ P that

∇φ(z0 ) − ∇H(z0 ) ≤ κω ∇ψ(z0 ) ≤ 2κωζ max{eζt , eζx }.



(6.1)
where in the second inequality we used the definition (4.6) of ψ. Together with the fact that any
z0 = (t, x) ∈ P satisfies |x| + |t| ≤ 3B0 (as |x| + |t| ≤ ∥f ∥C m+1 + ∥g∥C m+1 + L−1 ≤ 2∥G∥C m+1 + B0 ≤
3B0 ) and the definition (4.6) of ω, this implies for sufficiently large n that
∂x φ(z0 ) − ∂x H(z0 ) ≤ 2κωζe3B0 ζ ≤ ε .

(6.2)
2
Since (4.1) gives ∂x H(z0 ) ≤ −ε, this yields the first part of the first statement of the lemma; the
proof of the second part of the first statement is entirely analogous.
To establish the second, first observe from (6.2) and (4.2) that ∇φ(z0 ) ≤ ∇H(z0 ) + ε ≤ 2B
for each z0 ∈ P. Moreover, by (3.26), each bij (u, v) is 100B 3 -Lipschitz on the region where
  (2)
max |u|, |v| ≤ 2B. Thus, since p ∇H(z0 ) · qH (z0 ) = 0 (by the fact that H solves (3.25)), we
have for any z0 ∈ P that
 (2)    (2)


p ∇φ(z0 ) · qH (z0 ) = p ∇φ(z0 ) − p ∇H(z0 ) · qH (z0 )

(6.3)
(2)
≤ 200B 3 ∇φ(z0 ) − ∇H(z0 ) qH (z0 ) ≤ 200B 4 ∇φ(z0 ) − ∇H(z0 ) ,

where in the last bound we used (4.2). Again by (3.26) and the first part of the lemma, we have
   ε4
inf min btt ∇φ(z0 ) , bxx ∇φ(z0 ) ≥ .
z0 ∈P 16
(2)
Since (4.6) gives for z0 = (t, x) ∈ P that qψ (z0 ) = (−ζ 2 eζx , 0, −ζ 2 eζt ), this implies
 (2) ζ 2 ε4
p ∇φ(z0 ) · qψ (z0 ) ≥ max{eζt , eζx }.

16
Together with (6.3) and (6.1), this yields
 (2)
p ∇φ(z0 ) · qH (z0 )

6400B 4
sup ≤ κω,
 (2)
z0 ∈P p ∇φ(z0 ) · q (z )
ε4 ζ
ψ 0

which by (4.6) yields the lemma. □

Now let us proceed to prescribe the derivatives of H at z, which will be a higher order variant of
what was done in [52, Equation (72)]. They will be set to be near an admissible, consistent profile
LOCAL STATISTICS AND CONCENTRATION FOR NON-INTERSECTING BROWNIAN BRIDGES 41

 (k) (k) (k) 


Q = q (k) ∈ R2 × R3 × · · · × Rm+1 , defined as follows. Denoting q (k) = q0 , q1 , . . . , qk ∈ Rk+1
for each k ∈ J3, mK, set
(6.4)
 (2)
(1) (2) (2) p ∇φ(z) · qH (z) (2) (k) (k)
q = ∇φ(z); q = qH (z) −  (2) qψ (z); q0 = ∂xk φ(z); q1 = ∂t ∂xk−1 φ(z).
p ∇φ(z) · qψ (z)
(1) (k) (k) 
Observe from the first part of Lemma 6.1 that q0 < 0. Thus, the sequence of m pairs q0 , q1 ∈
(R2 )m admits a consistent extension,9 which we denote by Q = q (k) .


By Corollary 5.11 (and the first statement of Lemma 6.1), there exist constants ν = ν(ε, B, m) >
0 and C1 = C1 (ε, B, m) > 1; a strip domain S = Pf;g ⊂ R2 with [0, ν] × [−ν, ν] ⊆ S ⊆ [0, ν] ×
[−C1 ν, C1 ν]; and a function H : S → R satisfying (recall (5.1))
(6.5)
1 (k)
q (ρ, 0) − q (k) ≤ C1 ρm−k+1 , for each k ∈ J1, mK,

∥H∥C m+1 (S) ≤ C1 ; inf ∂x H(z) ≤ − ; H
z∈S C1
and the below property.  There exist constants c0 = c0 (ε, δ, B, m) > 0 and C2 = C2 (ε, δ, B, m) > 1
so that H is C2 ; c0 ; δ -concentrating, meaning (recall Definition 5.8) that H satisfies (3.25) and the
following. There exist an integer n′ ∈ [c0 n, n], entrance data u ∈ Wn′ and ending data v ∈ Wn′ on
S such that, sampling non-intersecting Brownian bridges y = (y1 , y2 , . . . , yn′ ) from Qu;v
f;g (n
−1
) and
y
denoting the associated height function by H : S → Z, we have
" #
−1 y 2
δ−1
≤ C2 e−c0 (log n) .

P sup n H (z) − H(z) ≥ n

z∈S

The following lemma bounds the distance between Qφ (z) and QH (z) under this notation.
Lemma 6.2. There exists a constant C = C(ε, B, m) > 1 such that
qφ (z) − q (k) (ρ, 0) ≤ C(ρm−k+1 + ω),
(k)
(6.6) H for any k ∈ J1, mK.
Proof. The bound (6.6) holds at k = 1 by the first statement of (6.4) and the second statement of
(6.5). So, we may assume in what follows that k ∈ J2, mK. To establish (6.6) in that case, it suffices
to show that
(k) q − q (k) (ρ, 0) ≤ Cρm−k+1 ;
(k)
(6.7) qφ(k) (z) − qH (z) ≤ Cω;
(k)
q (z) − q (k) ≤ Cω.

H H

To verify the first, observe from (4.7) and (4.6) that


qφ (z) − q (k) (z) ≤ ω q (k) (z) ≤ (m + 1)ζ m eζ|z| ω ≤ C3 ω,
(k)
(6.8) H ψ

for some constant C3 = C3 (ε, B, m) > 1 (where in the last inequality we implicitly used the fact
that |z| ≤ 3B0 , which holds since |z| ≤ ∥f ∥C m+1 + ∥g∥C m+1 + L−1 ≤ 3B0 ). The second follows from
the last statement of (6.5).

9Recall in Definition 5.5 that consistent extensions were defined on sequences of pairs, while q (2) above is a triple.
Our choice of q (2) ensures that this introduces no conflict in notation. Indeed, (6.4) implies that p q (1) · q (2) = 0,


which by (3.25) and (5.2) gives v0,0 q (1) , q (2) = 0. In particular, the second vector of the admissible extension Q

(k) (k) 
of q0 , q1 must indeed be q (2) .
42 AMOL AGGARWAL AND JIAOYANG HUANG

(m)
To establish the third statement of (6.7), observe since QH (z) and Q are consistent that it
suffices by Lemma 5.6 (and the first part of Lemma 6.1) to show the existence of a constant
C4 = C4 (ε, B, m) > 1 such that
∂x H(z) − q (k) ≤ C4 ω; ∂t ∂x H(z) − q (k) ≤ C4 ω,
k k−1
(6.9) 0 1 for each k ∈ J1, mK.
For k ̸= 2, this follows from (6.4) and (6.8), and for k = 2 it follows from the fact that

p ∇φ(z) · q (2) (z)
(2)
q (z) − q (2) = H (2) (2) ωκ
≤ 2ζ 2 e3B0 ζ ωκ,

 (2) qψ (z) ≤ qψ (z)

H
p ∇φ(z) · q (z) 2
ψ

where the first statement holds by (6.4), the second by the second part of Lemma 6.1; and the third
from the explicit form (4.6) for ψ, with the fact that |z| ≤ 3B0 for z ∈ P. This confirms (6.9) and
thus the lemma. □

 In the below, for any point w ∈ Rd and real number r > 0, we denote the open disk Br (w) =
′ d ′
w ∈ R : |w − w| ≤ r . We next have the following lemma, which is similar to [52, Equation
(88)], indicating the relative height of φ(z) with respect to φ|∂Bρ (z) is larger than that of H(ρ, 0)
with respect to that of H|∂Bρ (ρ,0) .
Lemma 6.3. There exists a constant c = c(ε, B, m) > 0 such that the following holds. Fix w ∈ P
2
with |z − w| ≤ n1/2m ρ; set z0 = (ρ, 0) ∈ S; and set w0 = z0 − z + w. If w0 ∈ S, then
φ(z) − φ(w) ≥ H(z0 ) − H(w0 ) + cω|z − w|2 − c−1 ρm |z − w|.
Proof. For any continuously twice-differentiable function F and point z ′ in its domain, let the 2 × 2
matrix HF (z ′ ) denote its Hessian at z ′ , given by
∂t2 F (z ′ ) ∂t ∂x F (z ′ )
 

HF (z ) = .
∂t ∂x F (z ′ ) ∂x2 F (z ′ )
Then, by (6.4), (6.5), a Taylor expansion, the fact that w0 − z0 = w − z, the first part of Lemma 6.1,
and Lemma 6.2, there is a constant C3 = C3 (ε, B, m) > 1 such that
 
φ(w) − φ(z) − H(w0 ) − H(z0 ) − (w − z) · Hφ (z) − HH (z0 ) (w − z)

m
(1) X (k)
≤ |z − w| · qφ(1) (z) − qH (z0 ) + |z − w|k qφ(k) (z) − qH (z0 )

(6.10) k=3
+ |z − w|m+1 ∥φ∥C m+1 (S) + ∥H∥C m+1 (S)


m
X
m m+1
≤ C3 ρ |z − w| + C3 |z − w| + C3 (ρm−k+1 + ω)|z − w|k .
k=3

Next, observe that



(w − z) · Hφ (z) − HH (z0 ) (w − z)

= (w − z) · HH (z) − HH (z0 ) + κωHψ (z) (w − z)
(6.11)  (2) !
m−1 2 p ∇φ(z) · qH (z) 
≤ C1 ρ |z − w| + κω −  (2) (w − z) · Hψ (z) (w − z).
p ∇φ(z) · qψ (z)
LOCAL STATISTICS AND CONCENTRATION FOR NON-INTERSECTING BROWNIAN BRIDGES 43

where for the first statement we applied the definition (4.7) of φ and for the second we applied (6.4)
(2)
and (6.5). By the second part of Lemma 6.1, and the fact that qψ has all negative entries (by the
explicit form (4.6) for ψ), we have
 (2) !
p ∇φ(z) · qH (z) (2) κω (2)
κω − qψ (z) ≤ q (z), entrywise.
2 ψ
 (2)
p ∇φ(z) · q (z) ψ
(2)
Since qψ (t, x) = (−ζ 2 eζt , 0, −ζ 2 eζx ), since κ ≥ 14 , and since |z| ≤ 3B0 for any z ∈ P (as |z| ≤
∥f ∥C m+1 + ∥g∥C m+1 + L−1 ≤ 3B0 ), it follows that there is a constant c1 = c1 (ε, m) > 0 such that
 (2) !
p ∇φ(z) · qH (z)
(w − z) · Hψ (z) (w − z) ≤ −c1 ω|z − w|2 .

κω −  (2)
p ∇φ(z) · qψ (z)
Together with (6.10) and (6.11), this gives
m+1
X
ρm−k+1 |z − w|k + C3 mω|z − w|3 − c1 ω|z − w|2 .

φ(w) − φ(z) − H(w0 ) − H(z0 ) ≤ C3
k=1
2
This, since ρ |z − w| + ω|z − w| ≤ 2n−1/m ω|z − w|2 for each k ∈ J2, m + 1K (which holds
m−k+1 k 3
2
1
by (4.6), the fact that |z − w| ≤ n1/2m ρ, and the fact that δ ≤ 5m 2 ), implies the lemma. □

6.2. Proof of Proposition 4.4. In this section we establish Proposition 4.4. As indicated in
Section 6.1, we only show the bound in that proposition on Ω+ j;k (z), as the proof of the bound on
Ω−j;k (z) is entirely analogous; recall we assume that Ω ∁
holds. We proceed by comparing the bridges
x(t) with those y(t), concentrating around H, in a local neighborhood (of diameter about ρ) around
z. To implement this, we must reindex the bridges in x and y to form x′ and y ′ , defined as follows.
Let z = (t, x) ∈ P and z0 = (ρ, 0) ∈ S; see Figure 6 for a depiction. Letting c1 = c1 (ε, B, m) > 0
denote the constant c from Lemma 6.3 (and recalling C1 from (6.5)), set
c1 ρ2 ωn
 
; B = nH(z0 ) + nδ ; ς1 = min{ρ, t}; ς2 = min{ρ, L−1 − t};
 
A = nφ(z) −
(6.12) 2
d1 = min ⌈5C1 ρn⌉, A − 1 ; d2 = min ⌈5C13 ρn⌉, n − A ; ς = ς1 + ς2 ; d = d1 + d2 .
3
 

In what follows, we omit the floors and ceilings, as they do not affect the proofs; we may also assume
that A ≤ n−5nδ , for otherwise φ(z) ≥ 1 +n−δ ρ2 ω and so n−1 Hx (z) ≤ 1 ≤ φ(z)−n−δ ρ2 ω = φj;k (z)
holds deterministically. For each integer j ∈ J0, d + 2K and real number s ∈ [0, ς], set
(6.13) x′j (s) = xA+j−d1 −1 (s + t − ς1 ) − x; yj′ (s) = yB+j−d1 −1 (s + ρ − ς1 ),
where x0 = g and xn+1 = f . Further denote x′ = (x′1 , x′2 , . . . , x′d+1 ) ∈ J1, d + 1K × C [0, ς] and


y ′ = (y1′ , y2′ , . . . , yd+1



) ∈ J1, d + 1K × C [0, ς] . Also define the functions x− , x+ , y− , y+ : [0, ς] → R


by for each s ∈ [0, ς] setting


x− (s) = x′d+2 (s); x+ (s) = x′0 (s); y− (s) = yd+2

(s); y+ (s) = y0′ (s).
Further define the entrance data u′ , u′′ ∈ Wd+1 and ending data v ′ , v ′′ ∈ Wd+1 by setting u′ =
x′ (0), v ′ = x′ (ς), u′′ = y ′ (0), and v ′′ = y ′ (ς).
Then, the laws of x′ and y ′ are given by non-intersecting Brownian bridges sampled from the

;v ′ ′′
;v ′′
measures Qu x− ;x+ (n
−1
) and Qu y− ;y+ (n
−1
), respectively. Denote their height functions (recall (3.6))
44 AMOL AGGARWAL AND JIAOYANG HUANG

Figure 6. For the proof of Proposition 4.4, we compare the non-intersecting Brow-
nian bridges x(t) in a neighborhood of z with bridges y(t) in one of z0 .

′ ′
by Hx : Px− ;x+ → R and Hy : Py− ;y+ → R, respectively. Observe by (6.13) that for any
(t0 , x0 ) ∈ Px− ,x+ ∩ Py− ,y+ we have for any (t0 , x0 ) ∈ [0, ς] × R that

Hx (t0 , x0 ) = Hx (t0 + t − ς1 , x0 + x) − A + d1 + 1;
(6.14) ′
Hy (t0 , x0 ) = Hy (t0 + ρ − ς1 , x0 ) − B + d1 + 1.

Lemma 6.4. There exists a constant c = c(ε, B, m) > 1 and a coupling between x′ and y ′ such
2
that the following holds. On an event F with P[F] ≥ 1 − c−1 e−c(log n) , we have x′j (s) ≤ yj′ (s) for
each j ∈ J1, d + 1K and s ∈ [0, ς].

Proof. By monotonicity (Lemma 2.3), we must show on an event of probability at least 1 −


2
c−1 e−c(log n) that
x′j (0) ≤ yj′ (0), and x′j (ς) ≤ yj′ (ς), for each j ∈ J1, d + 1K;
(6.15)
x′0 (s) ≤ y0′ (s), and x′d+2 (s) ≤ ′
yd+2 (s), for each s ∈ [0, ς].

Defining the event


( )
2
E=
−1 y
sup n H (z) − H(z) ≤ nδ−1 , P[E] ≥ 1 − C2 e−c0 (log n) ,

(6.16) we have
z∈S
LOCAL STATISTICS AND CONCENTRATION FOR NON-INTERSECTING BROWNIAN BRIDGES 45

since H is (C2 ; c0 ; δ)-concentrating. On E ∩ Ω∁ , we have


 j  j
φ s, xj (s) ≥ ; H s, yj−nδ (s) ≤ , for any j ∈ J1, nK,
n n
where the first statement holds from Ω∁ and the second holds from E. Replacing (j, s) with (A +
j − d1 − 1, s + t − ς1 ) in the former; replacing (j, s) with (B + j + nδ − d1 − 1, s) in latter; and using
(6.13) and (6.12), we find for j ∈ J1, nK and s ∈ [0, ς] that
c1 ρ2 ω j − d1 − 1
φ s + t − ς1 , x′j (s) + x ≥ φ(z) −

+ ;
2 n
j − d1 − 1
H s + ρ − ς1 , yj′ (s) ≤ H(z0 ) + 2nδ−1 +

.
n
It follows that
 c1 ρ2 ω d1 − j + 1
φ(z) − φ s + t − ς1 , x′j (s) + x ≤ +
(6.17) 2 n
 c1 ρ2 ω
≤ H(z0 ) − H s + ρ − ς1 , yj′ (s) + + 2nδ−1 .
2
We next apply Lemma 6.3 with that w0 = s + ρ − ς1 , yj′ (s) for either (j, s) ∈ {0, d + 2} × [0, ς]

2
or (j, s) ∈ J1, d + 2K × {0, ς}. To that end, we must first verify that |z0 − w0 | ≤ n1/2m ρ. In both
cases, we have from the event E from (6.16), the first two bounds in (6.5), and (6.12) that
|z0 − w0 | ≤ yj′ (s) + ρ = yB+j−d1 −1 (s + ρ − ς1 ) + ρ


B + j − d1 − 1
− H(s + ρ − ς1 , 0) + C1 nδ−1 + ρ

≤ C1
n

B + j − d1 − 1 2
− H(z0 ) + C1 nδ−1 + 2C12 ρ ≤ 20C14 ρ ≤ n1/2m ρ,

≤ C1

n
2
which establishes the bound |w0 − z0 | ≤ n1/2m ρ. We may thus apply Lemma 6.3 with w0 =
s + ρ − ς1 , yj′ (s) . By (6.17), this yields
c1 ρ2 ω
φ(z) − φ s + t − ς1 , x′j (s) + x ≤ H(z0 ) − H(w0 ) + + 2nδ−1

2
(6.18)  c1 ρ2 ω
≤ φ(z) − φ s + t − ς1 , yj′ (s) + x + + 2nδ−1
2
− c1 ω|w0 − z0 |2 + c−1 m
1 ρ |w0 − z0 |.
2
If |w0 − z0 | ≥ ρ, then since c1 ρ2 ω = c1 n3δ−1 ≥ 4nδ−1 and c−1 m −1 1/2m m+1
1 ρ |w0 − z0 | ≤ c1 n ρ ≤
−1 (m+1)δ−1−1/m+1/2m 2
δ−1 1
c1 n ≤n (by (4.6) and the fact that δ ≤ 5m2 ) it follows that
φ s + t − ς1 , yj′ (s) + x ≤ φ s + t − ς1 , x′j (s) + x .
 

Since ∂x φ < 0 (by the first part of Lemma 6.1), this yields yj′ (s) ≥ x′j (s), confirming (6.15).
If instead |w0 − z0 | < ρ, then we can improve (6.18) using the fact that φ|∂P = H|∂P . In
particular, assume first in (6.15) that s ∈ {0, ς}; since the two cases are entirely analogous, we
let s = 0. Since ς1 ≤ |w0 − z0 | < ρ, it follows from (6.12) that t < ρ and so ς1 = t. Then,
φ s + t − ς1 , x′j (s) + x = φ 0, x′j (s) + x . We next claim that
1
φ 0, yj′ (s) + x = H 0, yj′ (s) + x ≤ φ 0+ , yj′ (s) + x − c2 ω,
  
(6.19) where c2 = 3B0 ζ .
8e
46 AMOL AGGARWAL AND JIAOYANG HUANG

Indeed, the first statement follows from the definition of φ (see below (4.7)) with the fact that
0, yj′ (s) + x ∈ ∂P. To see the second, first observe that
(6.20) sup ψ(z) = sup (Υ − eζt − eζx ) ≤ 2 + sup (eζt + eζx ) ≤ 2e3B0 ζ + 2 ≤ 4e3B0 ζ ,
z∈P z∈P z∈P

where in the first statement we used the definition (4.6) of ψ; in the second we used the fact that

inf z∈P ψ(z) = 2; in the third we used the fact that |t| + |x| ≤ supt∈[0,1/L] f (t) + supt∈[0,1/L] g(t) +
L−1 ≤ 3B0 ; and in the fourth we used the fact that B0 , ζ ≥ 0. Next, observe for any z ∈ P that
φ(z) − H(z) = 3 − (j + 1)ω + κωψ(z) ≥ κωψ(z)
ω inf z∈P ψ(z) ω 2 ω
≥ sup ψ(z) · ≥ · 3B0 ζ ≥ 3B0 ζ = c2 ω,
4 z∈P supz∈P ψ(z) 4 4e 8e
where the first statement is from (4.7); the second is from the fact that j ≤ J = 3ω −1 − 1; the third
is from the fact κ ≥ 41 ; the fourth is from the fact that inf z∈P ψ(z) ≥ 2 (from below equation (4.6))
and (6.20); and the fifth is from the definition of c2 as in (6.19). This verifies (6.19).
The estimate (6.19), together with the first bound in (6.18), (the w0 = s+ − ς1 + ρ, yj′ (s) case


of) Lemma 6.3, and (4.6), yields (recalling that s = 0 and t = ς1 )


φ(z) − φ 0, x′j (s) + x


 c1 ρ 2 ω
≤ H(z0 ) − H s+ + ρ − ς1 , yj′ (s) + + 2nδ−1
2
 c1 ρ2 ω
≤ φ(z) − φ 0+ , yj′ (s) + x + + 2nδ−1 − c1 ω|w0 − z0 |2 + c−1 m
1 ρ |w0 − z0 |
2
 c1 ρ2 ω − 2c2 ω
≤ φ(z) − φ 0, yj′ (s) + x + + 2nδ−1 + c−1 m+1
≤ φ(z) − φ 0, yj′ (s) + x ,

1 ρ
2
from which it again follows that yj′ (s) ≥ x′j (s) since ∂x φ < 0.
Now assume that j ∈ {0, d + 2}; the two cases are again entirely analogous, so let j = 0. Then,
d1
ρ > y0′ (s) = yB−d1 −1 (s + ρ − ς1 ) ≥ − C1 n−1 B − H(s + ρ − ς1 , 0) − C1 nδ−1

C1 n
d1
− C1 H(ρ, 0) − H(s + ρ − ς1 , 0) − 2C1 nδ−1


C1 n
d1
≥ − C12 ρ − 2C1 nδ−1 .
C1 n
Here, the first statement follows from the fact that y0′ (s) ≤ |w0 − z0 | < ρ; the second follows from
(6.13); the third follows from the facts that we are restricting to the event E from (6.16), and also
from the first two bounds in (6.5); the fourth follows from the definition (6.12) of B; and the fifth
follows from the first bound in (6.5). Hence, d1 < 5C13 ρn, implying by (6.12) that d1 = A − 1.
We must show that y0′ (s) ≥ x′0 (s). Since x′0 (s) = xA−d−1 (s + t − ς1 ) − x = x0 (s + t − ς) − x =
g(s + t − ς1 ) − x for each s ∈ [0, ς], this is equivalent to yB−d1 −1 (s + ρ) ≥ g(s + t) − x for each
s ∈ [−ς1 , ς2 ]. To that end, observe that
H(ρ, 0) − φ(t, x)
1/2m2 m+1
≤ H s + ρ − ς1 , g(t + s − ς1 ) − x+ − φ(t + s − ς1 , g(t + s − ς1 ) − 0+ + c−1
 
(6.21) 1 n ρ
 c2 ω
≤ H s + ρ − ς1 , g(s + t − ς1 ) − x − .
2
LOCAL STATISTICS AND CONCENTRATION FOR NON-INTERSECTING BROWNIAN BRIDGES 47

Here, in the first inequality, we applied Lemma 6.3; in the second, we used the fact that φ t−ς1 , g(t+
1/2m2 m+1
s − ς1 ) − 0+ ≥ c2 ω ≥ 2c−1
 
1 n ρ by (4.6) and (6.19) (since H t + s − ς1 , g(t + s − ς1 ) = 0).
Moreover,
H s + ρ − ς1 , y0′ (s) = H s + ρ − ς1 , yB−d1 −1 (s + ρ − ς1 )
 

≤ n−1 (B − d1 − 1) + nδ−1 = n−1 (B − A) + nδ−1 ,


where in the first statement we used the definition of yj′ from (6.13); in the second we used the fact
that we are restricting to E (from (6.16)); and in the third we used the equality d1 = A − 1. Thus,
H s + ρ − ς1 , y0′ (s) ≤ n−1 (B − A) + nδ−1


c1 ρ2 ω
≤ H(ρ, 0) − φ(t, x) + + 2nδ−1
2
 c1 ρ2 ω − c2 ω
≤ H s + ρ − ς1 , g(s + t − ς1 ) − x + + 2nδ−1
 2
≤ H s − ρ − ς1 , g(s + t − ς1 ) − x .
where in the second bound we used the definitions (6.12) of A and B; in the third we applied (6.21);
and in the fourth we used the fact that c2 ω − c1 ρ2 ω ≥ c22ω > 4nδ−1 for sufficiently large n. Since
∂x H < 0, it follows that y0′ (s) ≥ g(s + t − ς1 ) − x. In particular, x′0 (s) ≤ y0′ (s), which again verifies
(6.15) and thus the lemma. □

We can now establish Proposition 4.4.

Proof of Proposition 4.4. By Lemma 6.4, there exists a constant c2 = c2 (ε, B, m) > 0 and an event
−c2 (log n)2
F with P[F] ≥ 1 − c−1
2 e such that x′j (s) ≥ yj′ (s) for each (j, s) ∈ J1, d + 1K × [0, ς] on F. By
(6.16) (and recalling the event E there), we may assume that F ⊆ E, for otherwise we can replace
F with E ∩ F; we restrict to F in what follows. Observe on E (and thus on F) that (ς1 , 0) ∈ Py− ;y+ ,
since by (6.14) and (6.12) we have (as A ≤ n − 5nδ )

Hy (ς1 , 0) = Hy (z0 ) − n H(z0 ) + nδ−1 ) + d1 + 1 ∈ [d1 − 2nδ + 1, d1 + 2nδ + 1] ⊆ [1, d + 1].
′ ′
The above coupling thus implies that Hx (ς1 , 0) ≤ Hy (ς1 , 0), so by (6.14) we deduce
′ ′
Hx (z) = Hx (ς1 , 0) + A − d1 − 1 ≤ Hy (ς1 , 0) + A − d1 − 1
c1 ρ 2 ω
 
= Hy (ρ, 0) + A − B ≤ n φ(z) − + Hy (ρ, 0) − B.
2
Since on the event F ⊆ E, we have Hy (ρ, 0) ≤ B, so
c1 ρ 2 ω
n−1 Hx (z) ≤ φ(z) − ≤ φ − n−δ ρ2 ωψ(z) ≤ φ+
j;k (z),
2
where in the second bound we used the definition (4.7) of φ+
j;k (and the uniform boundedness of ψ,
from (4.6)); this establishes the proposition. □

7. Concentrating Functions With Given Local Profiles


In this section we establish Proposition 5.9.
48 AMOL AGGARWAL AND JIAOYANG HUANG

7.1. Taylor Approximations Through Dyson Brownian Motion. The following proposition
produces measures whose Stieltjes transforms have specified derivatives. In what follows, we recall
the Stieltjes transform mµ of any measure µ ∈ P from Section 2.2.
Proposition 7.1. For any integer m ≥ 1 and real numbers ϖ > 0, ε ∈ (0, 1), and A, B > 1, there
exists a constant δ = δ(ε, ϖ, A, B) > 0 such that the following holds. Let r = (r0 , r1 , r2 , . . . , rm ) ∈
Rm+1 be an (m + 1)-tuple of real numbers with |rj | ≤ A for each j ∈ J0, mK. Also let µ ∈ P be a
compactly supported probability measure with density ϱ : R → R≥0 that is C m in a neighborhood
of
0, satisfying ε ≤ ϱ(x) ≤ B for each x ∈ [−ϖ, ϖ]; ϱ(x) = 0 for |x| > ϖ; and Re mµ (0) ≤ A. There
exists a measure µe ∈ P0 with density ϱe : R → R≥0 compactly supported on [−ϖ, ϖ], satisfying the
following two properties.
(1) We have 2ε ≤ ϱe(x) ≤ 2B for each x ∈ [−ϖ, ϖ], and ϱe(x) = ϱ(x) for each x ∈ [−δ, δ].
(2) For each integer k ∈ J0, mK we have
(7.1) ∂zk mµ̃ (0) = rk + i Im ∂zk mµ (0).
Proof of Proposition 7.1. If Re mµ (0) = r0 , then we will establish the lemma with the stronger
bound on ϱe given by 3ε 4 ≤ ϱe(x) ≤ B + 4ε . Let us first reduce to this statement. If Re mµ (0) ̸= r0 ,
then denote ϑ = r0 − Re mµ (0); assume that ϑ > 0, as the case when ϑ < 0 is entirely analogous.
By our assumption, ϑ ≤ |r0 | + Re mµ (0) ≤ A + B. Then replace ϱ(x) with ϱ0 : R → R, defined
by for any x ∈ R setting
ε
ϱ0 (x) = ϱ(x) + · 1(d ≤ x ≤ e8ϑ/ε d) − 1(e−4ϑ/ε d ≤ x ≤ d) ,

(7.2)
4
for any real number d < e−10(A+B)/ε ϖ. Then, ϱ0 continues to be a probability measure supported
in [−ϖ, ϖ] and satisfies 3ε ε
4 ≤ ϱ0 (x) ≤ B + 4 for x ∈ [−ϖ, ϖ]. Moreover, ϱ0 = ϱ in a neighborhood
of 0, and thus Im m(0) = Im mµ0 (0) by the first statement of (2.6), where mµ0 denotes the Stieltjes
transform of the measure µ0 = ϱ0 (x)dx ∈ P0 . Additionally,
ε
Re mµ0 (0) = Re m(0) + · (8ε−1 ϑ − 4ε−1 ϑ) = Re m(0) + ϑ = r0 .
4
Thus, µ0 and ϱ 0 continue to satisfy the hypotheses of the lemma (after replacing (ε, B) with
3ε ε

4 , B + 4 and altering the r j for j ∈ J1, mK to account for the difference in (7.2)) but now with
Re mµ0 (0) = r0 . It follows that there exists a measure µ e with density ϱe satisfying the statement of
the lemma (and in fact the bounds 2ε ≤ 16 9ε 7ε
≤ ϱe(x) ≤ B + 16 ≤ 2B).
Thus, we will assume in what follows that Re mµ (0) = r0 . We will take ϱe(x) to be of the form
m+2
X 
(7.3) ϱe(x) = ϱ(x) + cℓ · 1 a(yℓ + y) ≤ x ≤ a(yℓ + y + b) ,
ℓ=1
where the coefficients cℓ ∈ R and positive real numbers a, b, y, yℓ > 0 will be chosen later, in a way
continuous in the parameters (ε, A, B), so that their uniform boundedness (fixing (ε, A, B)) follows
from compactness. In this way, for any complex number z ∈ C in a neighborhood of 0, we have
m+2  
X a(yℓ + y + b) − z
(7.4) mµ̃ (z) = mµ (z) + cℓ log .
a(yℓ + y) − z
ℓ=1
Thus, for any integer k ≥ 1, we have
m+2
!
X 1 1
(7.5) ∂zk mµ̃ (z) = ∂zk mµ (z) + cℓ (−1)k−1 (k − 1)! k − k .
ℓ=1 z − a(yℓ + y + b) z − a(yℓ + y)
LOCAL STATISTICS AND CONCENTRATION FOR NON-INTERSECTING BROWNIAN BRIDGES 49

Since the parameters a, b, y, and yℓ are all positive, we have ϱe(x) = ϱ(x) for x ∈ R in a
sufficiently small neighborhood of 0; this verifies the second part of the first statement of the
proposition. Moreover, since by (2.6) (as mµ and mµ̃ are well-defined near 0, by the above) we have
Im mµ̃ (x) = π ϱe(x) = πϱ(x) = Im mµ (x) for x in a neighborhood of 0. Differentiating this equation
in x, it follows that ∂zk Im mµ̃ (0) = ∂zk Im mµ (0) for any integer k ≥ 1, from which the imaginary
part of (7.1) follows.
By (7.5), the real part of the second part of (7.1) holds for any k ∈ J1, mK if and only if
 m+2
 
X 1 1
(7.6) ak ∂zk Re mµ (0) − rk = cℓ (k − 1)! −
(yℓ + y + b)k (yℓ + y)k
ℓ=1
R∞
Moreover, ϱe has total mass 1 (that is, −∞ ϱe(x)dx = 1) and the second part of (7.1) holds at k = 0
if and only if
m+2 m+2  
X X yℓ + y + b
(7.7) cℓ = 0, and cℓ log = 0,
j=1
yℓ + y
ℓ=1

respectively, by (7.3) (since ϱ has total mass 1) and (7.4) (since Re mµ = r0 ). We denote the
(m + 2)-dimensional vectors c = (c1 , c2 , · · · , cm+2 ) and p = (p1 , p2 , · · · , pm+2 ), where p1 = 0 = p2
and
pk = ∂zk−2 Re mµ (0) − rk−2 , for each k ∈ J3, m + 2K.
2 m
We the (m + 2) × (m + 2) matrices A = A(a) = diag(a, a, a, a , . . . , a ) and W (y, b) =
 also define
Wk,ℓ (y, b) ℓ,k∈J1,m+2K , with entries given by
 
1 1
Wk,ℓ (y, b) = (k − 3)! − , for k ∈ J3, m + 2K;
(yℓ + y + b)k−2 (yℓ + y)k−2
 
yℓ + y + b
W2,ℓ (y, b) = 1, W1,ℓ (y, b) = log ,
yℓ + y
for each ℓ ∈ J1, m + 2K. Then we can rewrite the relations (7.6) and (7.7) as A · p = W (y, b) · c or
equivalently as c = W (y, b)−1 A · p.
Next we show there exists some choice of y, yℓ , b > 0 such that W (y, b) is invertible. Observe
that, for generic b, the first row of W (y, b) is linearly independent from the bottom m + 1 rows
(since the logarithmic function is not rational). It therefore suffices to show that the left-bottom
(m + 1) × (m + 1) block W0 (y, b) of W (y, b) (constituting its (k, ℓ) ∈ J2, m + 2K × J1, m + 1K entries)
is invertible. Let us write
 this block as W0 (y, b) = V (y) − E(y, b), where the (m + 1) × (m + 1)
matrices V = V (y) = Vk,ℓ (y) and E = E(y, b) = Ek,ℓ (y, b) are defined by
V1,ℓ (y) = 1, and Vk,ℓ (y) = −(k − 2)!(yℓ + y)1−k , for k ∈ J2, mK;
E1,ℓ (y, b) = 0, and Ek,ℓ (y, b) = −(k − 2)!(yℓ + y + b)1−k , for k ∈ J2, mK,
for each ℓ ∈ J1, m + 1K. Then, V (y) is a Vandermonde matrix (up to multiplication by a diagonal
matrix), and is therefore invertible if we take y1 , y2 , . . . , ym+1 > 0 mutually distinct. Moreover, by
taking b sufficiently large (and generic) the norm of E(y, b) can be taken to be sufficiently small so
as to guarantee that W (y) = V (y) − E(y, b) remains invertible.
There thus exists a choice of the parameters y, b, y1 , y2 , . . . , ym+1 such that W0 (y, b) and thus
W (y, b) is invertible; then, c = W (y)−1 A · p. By taking a sufficiently small, we can ensure that
each entry of c is smaller than (4m + 8)−1 ε and that a(yℓ + y + b) ≤ ϖ for each ℓ ∈ J1, m + 2K.
50 AMOL AGGARWAL AND JIAOYANG HUANG

Since ε ≤ ϱ(x) ≤ B for x ∈ [−ϖ, ϖ] and ϱ(x) = 0 for |x| > −ϖ, it follows from (7.3) that
3ε e(x) ≤ B + ε for x ∈ [−ϖ, ϖ] and that ϱe(x) = 0 for |x| > ϖ. In particular, ϱe is a
4 ≤ ϱ 4
nonnegative measure satisfying the (previously mentioned improvement of the) first statement of
the proposition. That it is a probability measure satisfying (7.1) follows from our choices of y, b,
and the (yℓ ). □

The next lemma, which is a quick consequence of Lemma 2.17, indicates that bridge-limiting
measures (recall Definition 3.2) can be obtained through free convolution (recall Section 2.2). In
(t)
what follows, we recall the rescaled semicircle law µsemci from (2.9).
Lemma 7.2. Fix a compactly supported measure ν0 ∈ Pfin and a real number t > 0. For any real
(s)
number s > 0, denote the measure νs = ν ⊞µsemci ∈ Pfin . Then, ν = (νs )s∈[0,t] is the bridge-limiting
measure process on [0, t] with boundary data (ν0 ; νt ).
Proof. By Remark 2.12 and Lemma 3.5, we may assume that ν0 (and thus νs for each s ≥ 0)
µ
is a probability measure. Recalling (from Definition 2.16) the classical locations γj;n ∈ R with
(ν )
respect to any probability measure µ, abbreviate γj;n (s) = γj;ns , for each real number s ≥ 0 and
integers n ≥ 1 and j ∈ J1, nK. Further define the n-tuple un = (un1 , un2 , . . . , unn ) ∈ Wn by setting
unj = γj;n (0), and let un (s) denote the process obtained by running Dyson Brownian motion for

time s on u; set v n = un (t). Denote µns = emp un (s) for each s ∈ [0, t] (recall (3.1)).
By (the A = 2 case of) Lemma 2.17, we have for sufficiently large n that, with probability at
least 1 − n−1 ,
Z y−2(log n)5 /n Z y Z y+2(log n)5 /n
(7.8) νs (dx) ≤ µns (dx) ≤ νs (dx), for all (s, y) ∈ [0, t] × R.
−∞ −∞ −∞

Moreover, the second part of Lemma 2.15 indicates that, conditional on v n = un (t), the law of
n n
un (s) is given by Qu ;v . This, together with (7.8), Lemma 3.1, and Definition 3.2, indicates that
ν = (νs )s∈[0,t] is the bridge-limiting measure process on [0, t] with boundary data (ν0 ; νt ). □

7.2. Proof of Proposition 5.9. In this section we establish Proposition 5.9. In what follows, we
recall the notation from that proposition; on free convolutions and Stieltjes and Hilbert transforms
from Section 2.2; on bridge-limiting measure processes from Definition 3.2; and on the associated
(inverted) height functions from Definition 3.4.
The function H of the proposition will eventually be the height function associated with a bridge-
limiting measure process arising from a free convolution (as in Lemma 7.2), to which end we must
define the initial data for the free convolution. This will proceed using Proposition 7.1, and so we
must first define the real numbers r and measure ϱ appearing there; we begin with the former.
Define real numbers r0 , r1 , . . . , rm−1 ∈ R through the equations
k−1
X k − 1 (k−j) (k)
(7.9) q0 rj = q1 , for each k ∈ J1, mK.
j=0
j
(1)
Since q0 ≤ −ε by (5.3), these equations uniquely define the rj by induction on j. Furthermore,
by (5.3) and induction on j, there exists a constant C2 = C2 (ε, B, m) > 1 such that |rj | < C2 for
each j ∈ J0, m − 1K.
Next, observe that there exists a constant C3 = C3 (ε, B, m) > 1 and a probability measure
µ ∈ P0 with density ϱ ∈ L2 (R) such that supp ϱ = [−1, 1], such that ∥ϱ∥C m+3 ([−1,1]) ≤ C2 and such
LOCAL STATISTICS AND CONCENTRATION FOR NON-INTERSECTING BROWNIAN BRIDGES 51

that
(7.10)
(k+1)
∂xk ϱ(0) = −q0 , for each k ∈ J0, m − 1K, and C3−1 < ϱ(x) ≤ C3 , for each x ∈ [−1, 1].
Moreover, after increasing C3 if necessary, it follows from (2.6), the fact that supp ϱ ⊆ [−1, 1], and
the bound ∥ϱ∥C m+3 ([−1,1]) ≤ C2 , that
1 ϱ(x)dx 1 ϱ(x) − ϱ(0) dx
Z Z 

Re mµ (0) = π Hϱ(0) = = ≤ 2∥ϱ∥C ([−1,1]) ≤ 2C2 ≤ C3 .
1

−1 x −1 x
This verifies the assumptions in Proposition 7.1. Thus, by Proposition 7.1, there is a constant
c1 = c1 (ε, B, m) > 1 and a probability measure µ0 ∈ P0 with density ϱ0 ∈ L1 (R), such that
supp ϱ0 = [−1, 1] and
(2C3 )−1 ≤ ϱ0 (x) ≤ 2C3 , for each x ∈ [−1, 1]; ϱ0 (x) = ϱ(x), for each x ∈ [−c1 , c1 ];
(7.11)
∂zk mµ0 (0) = rk + i Im ∂zk mµ (0), for each k ∈ J0, mK.
(t)
For each real number t > 0, define the free convolution measure µt = µ0 ⊞ µsemci ∈ P0 and
c1
the associated density ϱt ∈ L1 (R) satisfying µt (dx) = ϱt (x)dx. Define the constant c2 = 3C 3
, and
denote the probabilility measure-valued process µ = (µt )[0,c2 ] ; observe that it is the bridge-limiting
measure process on [0, c2 ] with boundary data (µ0 ; µc2 ), by Lemma 7.2. Denote its height function
by H ⋆ : [0, c2 ] × R → [0, 1]. We will later define H as the restriction of H ⋆ to a certain strip domain.
Lemma 7.3. There exist constants c3 = c3 (ε, B, m) > 0 and C4 = C4 (ε, B, m) > 1 such that the
following statements hold. In what follows, we denote U0 = (0, c3 ) × (−c3 , c3 ).
(1) For each z ∈ U0 , we have ∂x H ⋆ ≤ −c3 and ∥H ⋆ ∥C m+2 (U0 ) ≤ C4 .
(2) The function H ⋆ satisfies the equation (3.25) on U0 .
(m)
(3) The full-m local profile of H ⋆ satisfies limz→(0,0) QH ⋆ (z) = Q.
Proof. The first statement of the lemma follows from Lemma 2.11, whose conditions are verified by
(7.10) and the fact that ∥ϱ∥C m+1 ([−1,1]) ≤ C3 . This, together with Lemma 3.16, implies the second
statement of the lemma. To verify the third statement observe, since H ⋆ satisfies (3.25), its local
profile at any point is consistent in the sense of Definition 5.1. Hence, by Lemma 5.4, to show
(m)
limz→(0,0) QH ⋆ (z) = Q, it suffices to show for each k ∈ J1, mK that
(k) (k)
(7.12) lim ∂xk H ⋆ (z) = q0 , and lim ∂xk−1 ∂t H ⋆ (z) = q1 .
z→(0,0) z→(0,0)
(k)
The first statement in (7.12) holds since limz→(0,0) ∂xk H ⋆ (z) = −∂xk−1 ϱ0 (0) = −∂xk−1 ϱ(0) = q0 ,
where the first follows from (3.3); the second from (7.11); and the third from (7.10).
To verify the second statement in (7.12), observe that
 Z ∞ 
k−1 ⋆ k−1
lim ∂x ∂t H (z) = ∂y ∂s ϱs (x)dx

z→(0,0) y
s=0,y=0
 Z ∞ 
k−1

= π∂y ∂x ϱ0 (x)Hϱ0 (x) dx

y
y=0

  k−1
X k − 1
= −π ∂yk−1 ϱ0 (y)Hϱ0 (y) ∂xj Hϱ0 (0) · ∂xk−j−1 ϱ0 (0),

= −π

j=0
j
y=0
52 AMOL AGGARWAL AND JIAOYANG HUANG

where first statement follows from (3.3); the second from (2.12); the third from performing the
integration (and using the fact that supp ϱ0 = [−1, 1]); and the fourth from taking the (k − 1)-fold
derivative. Together with (7.10), the second statement of (2.6), the last statement of (7.11), and
(7.9), this yields
k−1
X k − 1 (k−j) k−1
X k − 1 (k−j) (k)
lim ∂xk−1 ∂t H ⋆ (z) = q0 ∂xj Re mµ0 (0) = q0 rj = q1 ,
z→(0,0)
j=0
j j=0
j

confirming the second statement of (7.12) and thus the lemma. □


Now we can establish Proposition 5.9.
Proof of Proposition 5.9. Let G⋆ denote the inverted height function associated with H ⋆ and, re-
calling the notation of Lemma 7.3, fix
c33 c23
(7.13) c0 = H ⋆ (0, 0); ν= ; c4 = .
4C43 2C4
For each real number s ∈ [0, ν], let f (s) = G⋆ (s, c0 + c4 ) and let g(s) = G⋆ (s, c0 − c4 ). Observe that
Z c3

c0 = H (0, 0) ≥ −c3 ∂x H ⋆ (0, y)dy ≥ c23 ,
0

by the first part of Lemma 7.3, so c0 ≥ ≥ 4c4 > c4 . Then define the strip domain S = Pf ;g ⊂ R2
c23
and the function H : S → R by setting H(z) = H ⋆ (z) − c0 + c4 for each z ∈ S. See the left side of
Figure 7. Let us verify that the conditions of the proposition hold for this choice of (S, H).
To see that [0, ν] × [−ν, ν] ⊆ Pf ;g ⊆ [0, ν] × [−Cν, Cν], observe again by the first part of
c3
Lemma 7.3, and using H ⋆ (0, 0) = c0 and H ⋆ (0, f (0)) = c0 + c4 , we have − 2C 4
= −c−1
3 c4 ≤ f (0) ≤
2 −1
−1 c3 ′


−1
−C c4 = − 2 . Since f (s) = (∂t H )(s, f (s)) · (∂x H )(s, f (s)) ≤ c C4 for s ∈ [0, ν] by
4 2C4 3
the first part of Lemma 7.3, it follows from (7.13) that, for each s ∈ [0, ν],
4C42 ν c3 C4 ν C4 ν c2 C4 ν
− 2 ≤− − ≤ f (s) ≤ − 32 = − ≤ −ν.
c3 2C4 c3 c3 2C4 c3
where we also used ν = c33 (4C43 )−1 from (7.13). Similar reasoning implies that ν ≤ g(s) ≤ 4c−2 2
3 C4 ν,
2 −2
which indicates that [0, ν] × [−ν, ν] ⊆ Pf ;g ⊆ [0, ν] × [−Cν, Cν] for C = 4C4 c3 .
To establish the other parts of the proposition, observe that the bounds on −∂x H and on ∥H∥C m+1
follow from the first part of Lemma 7.3. Moreover, the fact that the full-m local profile of H around
(0, 0) is given by Q follows from the third part of Lemma 7.3.
It therefore remains to show that H is (C1 ; c; δ) concentrating in the sense of Definition 5.8.
That it satisfies (3.25) follows from the second part of Lemma 7.3. Now let n ≥ 1 be an integer;
set n′ = ⌊2c4 n⌋ (in what follows, we again omit the floors and ceilings, as they will not affect the
proofs), and define the n′ -tuples u, v ∈ Wn′ by setting
(7.14) uj = G⋆ (0, c0 − c4 + jn−1 ), and vj = G⋆ (ν, c0 − c4 + jn−1 ), for each j ∈ J1, n′ K;
and sample n′ non-intersecting Brownian bridges y = (y1 , y2 , . . . , yn′ ) ∈ J1, nK × C [0, ν] from the


measure Qu;v
f ;g (n
−1
). We must show that
" #
2
P sup n H (z) − H ⋆ (z) + c0 − c4 ≤ nδ−1 ≥ 1 − C1 e−c(log n) .
−1 y
(7.15)
z∈S
LOCAL STATISTICS AND CONCENTRATION FOR NON-INTERSECTING BROWNIAN BRIDGES 53

Figure 7. Depicted in the left panel is the strip domain S = Pf ;g . Depicted in


the right panel is the coupling such that yej− ≤ yj ≤ yej+ .

To that end, define the n-tuples u− , u+ ∈ Wn by for each j ∈ J1, nK setting


(7.16) u− ⋆
j = G (0, jn
−1
+ nδ/2−1 ); u+ ⋆
j = G (0, jn
−1
− nδ/2−1 ).
Let y − (t) = y1− (t), y2− (t), . . . , yn− (t) and y + (t) = y1+ (t), y2+ (t), . . . , yn+ (t) denote Dyson Brownian
 

motion run with initial data u− and u+ , respectively. Since H ⋆ is the height function associated
(t) 
with the measure-valued process µ0 ⊞µsemci t∈[0,c2 ] , Lemma 2.17 gives constants c = c(ε, δ, B, m) >
2
0 and C1 = C1 (ε, δ, B, m) > 1 such that, with probability at least 1 − C1 e−c(log n) , we have

H ⋆ (t, x) − nδ/2−1 − nδ/4−1 ≤ n−1 Hy (t, x) ≤ H ⋆ (t, x) − nδ/2−1 + nδ/4−1 ;
(7.17) +
H ⋆ (t, x) + nδ/2−1 − nδ/4−1 ≤ n−1 Hy (t, x) ≤ H ⋆ (t, x) + nδ/2−1 + nδ/4−1 ,
for any (t, x) ∈ S. In the below, we restrict to this event.
Since f (t) = G⋆ (t, c0 − c4 ) and g(t) = G⋆ (t, c0 + c4 ), applying (7.17) with x equal to each
− + − +
of y(c 0 +c4 )n+1
(t), y(c0 +c4 )n+1
(t), y(c0 −c4 )n
(t), y(c0 −c4 )n
(t), f (t), and g(t) (and using the fact that
− +
H ⋆ (t, x), Hy (t, x), and Hy (t, x) are decreasing in x) yields
− + − +
(7.18) y(c 0 +c4 )n+1
(t) ≤ f (t) ≤ y(c0 +c4 )n+1
(t); y(c0 −c4 )n
(t) ≤ g(t) ≤ y(c0 −c4 )n
(t).
Now, denote c = c0 − c4 , and condition on the yj− (s) and yj+ (s) with j ∈
/ Jcn + 1, cn + n′ K. Define
− + − +
the functions fe , fe , ge , ge : [0, ν] → R by setting
±
fe± (s) = ycn+n ′ +1 (s); ge± (s) = ycn
±
(s),
54 AMOL AGGARWAL AND JIAOYANG HUANG

for each index ±  ∈ {+,+−} and real number s ∈ [0, ν]. Further −
 define ye = (ey1− , ye2− , . . . , yen−′ ) ∈
′ + + + ′ −
y1 , ye2 , . . . , yen′ ) ∈ J1, n K × C [0, ν] by reindexing y and y + , respec-
J1, n K × C [0, ν] and ye = (e
tively, namely, by setting
(7.19) yej± (s) = yj+cn
±
(s), for each ± ∈ {+, −} and (j, s) ∈ J1, n′ K × [0, ν],
and define the n′ -tuples u
e− , u e− , v
e+ , v e+ ∈ Wn′ by setting
(7.20) e±
u ±
ej± (0),
j = uj+cn = y and vej± = yej± (ν), for each ± ∈ {+, −} and j ∈ J1, n′ K.
− −
Then, by the second part of Lemma 2.15, the laws of ye− and ye+ are given by Qfũ˜− ;g̃
;ṽ

(n−1 ) and
+ +
Qfũ˜+ ;g̃
;ṽ
+
(n−1 ), respectively; see the right side of Figure 7. Observe from (7.14), (7.16), and (7.19)
that u e− ≤ u ≤ u e + , since ∂y G⋆ ≤ 0. Moreover, (7.18) and (7.19) imply that fe− = yen−′ +1 ≤ f ≤
ye+′ = fe+ and ge− = ye− ≤ g ≤ ye+ = ge+ . Additionally, by (7.14) and by the t = ν (or s = ν)
n +1 0 0
case of (7.17), (7.19), (7.20), and the first statement of Lemma 7.3, we find that v e− ≤ v ≤ v e+ .
Together with the monotone coupling Lemma 2.3, these indicate that it is possible to couple ye− ,
y, and ye+ in such a way that yej− (s) ≤ yj (s) ≤ yej+ (s) for each (j, s) ∈ J0, n′ K × [0, ν]. This, with
(7.17) and (7.19), yields (7.15) and thus the proposition. □

8. Universality of Bulk Statistics


In this section we establish Theorem 1.5.
8.1. Proof of Theorem 1.5. In this section we establish Theorem 1.5. Given Theorem 1.4,
its proof will be similar to what was done in [2, Sections 3 and 4], by comparing non-intersecting
Brownian bridges to Dyson Brownian motion (though the affine invariance of our model will simplify
the arguments here). Throughout, we adopt the notation from Theorem 1.5, and we set
(8.1) ω = 90000(δ + m−1 ); n′ = ⌊n1/2 ⌋; n′′ = ⌊n1/4 ⌋, s0 = nω/2−1 , s1 = n3ω−1 ,
1 1
so that ω ≤ 20 (as m ≥ 225 and δ ≤ 5m 2 ). In what follows, we omit all floors and ceilings, as they

do not affect the proofs. We also assume (as we may by  shifting) that G(t0 − s0 , y0 ) = 0.
Define x̆ = (x̆1 , x̆2 , . . . , x̆n′ ) ∈ J1, n′ K × C [0, s0 + s1 ] by reindexing x, specifically, by setting
x̆j (s) = xy0 n−n′ /2+j (t0 − s0 + s), for each (j, s) ∈ J1, n′ K × [0, s0 + s1 ].
Also define the n′ -tuples ŭ = (ŭ1 , ŭ2 , . . . , ŭn′ ) ∈ Wn′ and v̆ = (v̆1 , v̆2 , . . . , v̆n ) ∈ Wn′ , and functions
f˘, ğ : [0, s0 + s1 ] → R by for each j ∈ J1, n′ K and s ∈ [0, s0 + s1 ] setting
ŭj = x̆j (0); v̆j = x̆j (s0 + s1 ); f˘(s) = x̆n′ /2+n′′ +1 (s); ğ(s) = x̆n′ /2−n′′ −1 (s).
q n′ ′′ n′
′′
Next, we condition on x̆j (s) for (j, s) ∈
/ 2 − n , 2 + n K × (0, s0 + s1 ). Defining the event
( )
′ 
\  n j
(8.2) A0 = ω/4−1

x̆j (s) − G t0 − s0 + s, y0 − + ≤n ,
2n n
/ ′ /2−n′′ ,n′ /2+n′′ K×(0,s0 +s1 )
(j,s)∈Jn

we have by Theorem 1.4 and the fact that nω/4−1 ≥ κ + n2/m+3δ/2−1 (by (8.1) and the fact that
κ ≤ nδ−1 ) that there exists a constant c0 = c0 (ε, δ, B0 , m) > 0 such that
2
(8.3) P[A0 ] ≥ 1 − c−1
0 e
−c0 (log n)
.
We restrict to A0 in what follows.
LOCAL STATISTICS AND CONCENTRATION FOR NON-INTERSECTING BROWNIAN BRIDGES 55

Next, for each index ± ∈ {+, −}, let G± : [0, 1] → R be a function satisfying the following six
properties.
(1) For all y ∈ [0, 1], we have G(t0 − s0 , y) − nω/2−1 ≤ G− (y) ≤ G(t0 − s0 , y) and that
G(t0 − s0 , y) ≤ G+ (y) ≤ G(t0 − s0 , y) + nω/2−1 .
′′
(2) For all y ∈ [0, 1] with |y−y0 | ≤ 5n6n , we have G(s0 −t0 , y)−n
3ω/10−1
≤ G− (y) ≤ G(t0 −s0 , y)
+ 3δ/10−1
and that G(t0 − s0 , y) ≤ G (y) ≤ G(t0 − s0 , y) + n .
′′
(3) For y ∈ [0, 1] with |y − y0 | ≤ n2n , we have G± (y) = G(t0 − s0 , y).
 ′′ 5n′′  ±
(4) For all y ∈ [0, 1] with |y −y0 | ∈ 2n3n , 6n , we have G (t0 −s0 , y) = G(t0 −s0 , y)±n
3ω/10−1
.
n′′ ± ω/2−1
(5) For all y ∈ [0, 1] with |y − y0 | ≥ n , we have G (y) = G(t0 − s0 , y) ± n .
(6) For all y ∈ [0, 1], we have − 2ε ≤ ∂y G± (y) ≤ − 2ε , and ∥G± ∥C m+1 ≤ 2B0 .
′′
The existence of such a function follows from the facts that nω−1 ≤ n−3/4 = nn , that G ∈
Admε (R), and that ∥G∥C m+1 (R) ≤ B0 (by Assumption 1.3). Then, define the n′ -tuples ŭ− , ŭ+ ∈
Wn′ by, for each index ± ∈ {+, −} and integer j ∈ J1, n′ K, setting
r n′ 2n′′ n′ 2n′′ z  n′ j
(8.4) ŭ±
j = ŭj , if j ∈ − , + ; ŭ±
j = G ±
y0 − + , otherwise,
2 3 2 3 2n n
observing that u± ∈ Wn′ on A0 due to (8.2).
Let w̆(s) = w̆1 (s), w̆2 (s), . . . , w̆n′ (s) denote Dyson Brownian motion with initial data ŭ, as in
(2.16), but with the dw̆i there multiplied by a factor of n−1 n′ and the dBi multiplied by a factor
of (n−1 n)1/2 ; more specifically, they solve (instead of (2.16)) the stochastic differential equation
X 1
dw̆i (s) = n−1 ± ± + n−1/2 dBi (s).
′ w̆i (s) − w̆j (s)
1≤j≤n
j̸=i

Similarly, for each ± ∈ {+, −}, let w̆± (s) = w̆1± (s), w̆2± (s), . . . , w̆n±′ (s) denote Dyson Brownian


motion with initial data ŭ± , but with the same rescaling of the dw̆i± (s) as above; see Figure 8.
Moreover define the functions f˘− , f˘+ , ğ − , ğ + : [0, s0 + s1 ] → R by, for each index ± ∈ {+, −} and
real number s ∈ [0, s0 + s1 ], setting
f˘± (s) = w̆n±′ /2+n′′ +1 (s); ğ ± (s) = w̆n±′ /2−n′′ −1 (s).
We then have the following proposition, to be established in Section 8.2 below; it indicates that
the middle paths in x̆ are bounded between those in w̆− and w̆+ (after adding a suitable drift).
Proposition 8.1. Set a = ∂t G(t0 , y0 ). There exist constants c = c(ε, ω, m, B0 ) > 0 and C =
C(ε, ω, m, B0 ) > 1; a coupling between x̆, w̆− , and w̆+ ; and an event A1 , with P[A1 ] ≥ 1 −
2
Ce−c(log n) , such that on A1 we have
w̆j− (s) + (a − n−ω )s ≤ x̆j (s) ≤ w̆j+ (s) + (a + n−ω )s,
q n′ ′ y
for each (j, s) ∈ 2 − n′′ , n2 + n′′ × [0, s0 + s1 ].
We next have the following three lemmas, whose proofs are both consequences of results in [50, 49]
and are therefore deferred to Appendix D below. The first, to be established in Appendix D.1
below, indicates that the local statistics of w̆ are prescribed by the sine process. The second, to
be established in Appendix D.2 below, indicates that the middle paths of w̆− and w̆+ are close to
those of w̆. The third, to be established in Appendix D.3 below, states that w̆− and w̆+ are likely
close in expectation, under any coupling between them.
56 AMOL AGGARWAL AND JIAOYANG HUANG

Figure 8. Shown above are the paths of w̆+ and w̆ in the left panel, and of w̆−
and w̆ in the right panel.

Lemma 8.2. Let a > 0 be a (bounded, uniformly in n) constant, and denote the n-tuple w =
θn · w̆(s0 ) + as0 ∈ Wn . We have
Z Z
(k) (k)
lim F (a)pw (a)da = F (a)psin (a)da.
n→∞ Rk Rk

Lemma 8.3. There exist constants c = c(ε, δ, B0 , m) > 0 and C = C(ε, δ, B0 , m) > 1; a coupling
between w̆− , w̆; and w̆+ ; and an event A2 , with P[A2 ] ≥ 1 − Cn−c , such that on A2 we have
w̆j (s0 ) − w̆− (s0 ) ≤ n−1−ω/5 ;

max j
j∈Jn′ /2−nω/15000 ,n′ /2+nω/15000 K
(8.5)
w̆j (s0 ) − w̆+ (s0 ) ≤ n−1−ω/5 .

max j
j∈Jn′ /2−nω/15000 ,n′ /2+nω/15000 K

Lemma 8.4. For any real number χ > 0, there exist constants c = c(ε, δ, B0 , m) > 0 and C =
C(χ, ε, δ, B0 , m) > 1 such that, under any coupling between w̆− and w̆+ , we have
h i 2
E w̆j+ (s0 ) − w̆j− (s0 ) · 1|w̆+ (s0 )−w̆− (s0 )|>nχ−1 < Ce−c(log n) .

max
j∈Jn′ /2−nω/15000 ,n′ /2+nω/15000 K j j

Given the above four results, we can establish Theorem 1.5.


Proof of Theorem 1.5. By a Taylor expansion (and the fact that G(t0 − s0 , y0 ) = 0), we have
G(t0 , y0 ) − as0 = G(t0 , y0 ) − G(t0 − s0 , y0 ) − as0 ≤ B0 s20 ≤ n−1−ω .

(8.6)
LOCAL STATISTICS AND CONCENTRATION FOR NON-INTERSECTING BROWNIAN BRIDGES 57

So, it suffices by the continuity of F to show that


Z Z
(k) (k)
(8.7) lim F (a)pθn·(x̆(s0 )−as0 ) (a)da = F (a)psin (a)da.
n→∞ Rk Rk
ω

We next claim that there exist constants c = c(ε, δ, B0 , m) ∈ 0, 30000 and C = C(ε, δ, B0 , m) > 1,
and a coupling between x, w̆− , and w̆+ , such that
(8.8) P[A] ≥ 1 − Cn−c , where A = A′0 ∩ A1 ∩ A3 ,
where we have recalled A1 from Proposition 8.1 and defined the events
\ n o
A′0 = xj (s0 ) − G(s0 , jn−1 ) ≤ nω/30000−1 ;

|j−y0 n|≤nω/15000
\ n o
A3 = w̆− (s0 ) − w̆+ (s0 ) ≤ n−1−c .

j j
|j−y0 n|≤nω/15000

Let us establish the theorem assuming (8.8); throughout this proof, we set ω ′ = 30000
ω
. Since
limn→∞ P[A] = 1, we may restrict to the event A in what follows. Setting a = ∂t G(t0 , y0 ) and
applying the definition of A1 from Proposition 8.1 (at s = s0 = nω/2−1 ) gives
(8.9) w̆j− (s0 ) + as0 − n−1−ω/2 ≤ x̆j (s0 ) ≤ w̆j+ (s0 ) + as0 + n−1−ω/2 ,
q ′ ′ ′ ′
for each integer j ∈ n2 − n′′ , n2 + n′′ . Further observe on A′0 that for j > n2 + n2ω we have
y
′ ′ ′ ′ ′
x̆j (s0 ) − as0 ≤ G(t0 , y0 + n2ω −1 ) − as0 + nω −1 ≤ n−1−ω − εn2ω −1 + nω −1 < −nω −1 (where we
have used (8.6) and the fact from Assumption 1.3 that ∂y G ≤ −ε). By similar reasoning, we have
′ ′ ′
on A0 that for j < n2 − n2ω that x̆j (s0 ) > nω −1 . By these two bounds, together with (8.9) and
− ′ ′ ′ ′y
the fact that w̆j (s0 ) − w̆j+ (s0 ) ≤ n−1−c for j ∈ J n2 − n2ω , n2 + n2ω on A, it follows from the

continuity of F that, to verify (8.7), it suffices to show


Z Z
(k) (k)
lim F (a)pθn·w̆− (s0 ) (a)da = F (a)psin (a)da.
n→∞ Rk Rk

This follows from the first statement of (8.5) and Lemma 8.2 (with the continuity of F ).
It remains to confirm (8.8). To that end, it suffices  by a−c union bound to show that P[A′0 ] ≥
1 − Cn , that P[A1 ] ≥ 1 − Cn , and that P A1 ∩ A3 ≤ Cn . The first follows by Theorem 1.4,
−c −c ∁

since ω ′ = 3(m−1 + δ) > 2m−1 + 3δ, and the second follows from Proposition 8.1.
The third will follow from (8.5), Lemma 8.4, and a Markov estimate. To implement this, observe
that (8.5) and Lemma 8.4 together imply the existence of constants c1 = c1 (ε, δ, B0 , m) > 0 and
C1 = C1 (ε, δ, B0 , m) > 1, and a coupling between w̆− and w̆+ (possibly different from the one
above), such that
X h i
E w̆j+ (s0 ) − w̆j− (s0 ) ≤ C1 n−1−c1 ,
|j−n′ /2|≤nω/15000

from which it follows that


X   
E w̆j+ (s0 ) − E w̆j− (s0 ) ≤ C1 n−1−c1 .
 
(8.10)
|j−n′ /2|≤nω/15000

Since this bound is independent of the coupling between w̆+ and w̆− , it also holds for our original
one given by Proposition 8.1. Next, restricting to the event A1 , we have w̆j− (s0 ) ≤ w̆j+ (s0 )+n−ω s0 =
58 AMOL AGGARWAL AND JIAOYANG HUANG

w̆j+ (s0 ) + n−1−ω/2 . Together with (8.10), this implies under our original coupling that
X h i
E 1A1 · w̆+ (s0 ) − w̆− (s0 )

j j
|j−n′ /2|≤nω/15000
X   
E w̆j+ (s0 ) − E w̆j− (s0 ) + 2n−1−ω/2 ≤ 3C1 n−1−c2 ,
  

|j−n′ /2|≤nω/15000

for some constant c2 = c2 (ε, δ, B0 , m) > 0. By a Markov estimate, we thus have for c = c22 that
X h i
P A1 ∩ A∁3 ≤ n1+c2 /2 · E 1A1 · w̆j+ (s0 ) − w̆j− (s0 ) ≤ 3C1 n−1−c2 /2 .
 

|j−n′ /2|≤nω/15000

This gives the third bound above, P A1 ∩ A∁3 ≤ Cn−c , and thus the theorem.
 

8.2. Proof of Proposition 8.1. In this section we establish Proposition 8.1; throughout, we recall
the notation of that proposition. We first rescale s0 , s1 , x̆(s), w̆− (s), w̆(s), w̆+ (s), G, G− , and
G+ to form se0 , se1 , x
e(s), w e − (s), w(s),
e w e + (s), G,
e Ge − , and G e + , respectively, defined by setting
nsi n  n′ s   ′ ′
e y) = n · G n s + t0 − s0 , n y + y0 − n ,
′
sei = ′ ; zej (s) = ′ · z̆j ; G(s,
(8.11) n n n n′ n n 2n
′ ′
± n ±
 n y n
Ge (y) = ·G + y0 − ,
n′ n 2n
for any integers i ∈ {0, 1} and j ∈ J1, n′ K; indices z ∈{x, w− , w, w+ } and ± ∈ {+, −}; and real
numbers s ∈ [0, se0 + se1 ] and y ∈ 1 − y0 n1/2 , (1 − y0 )n1/2 . In this way, w(s) e is obtained by applying
Dyson Brownian motion (2.16) to the initial data w(0) e for time s (without the additional rescaling
described in Section 8.1); similar statements are true for w− (s) and w+ (s). By (8.2), and the fact
that (n′ )2 = n, we have
( )
\  j 
A0 = ′ ω/2−1

(8.12) ej (s) − G s, n′ ≤ (n )
x e .
(j,s)∈Jn
/ ′ ′′ ′ ′′
/2−n ,n /2+n K×(0,s̃0 +s̃1 )

Further observe since G ∈ Admε/2 (R); since − 2ε ≤ ∂y G± ≤ 2ε ; since G(t0 − s0 , y0 ) = 0; since


∥G∥C m+1 (R) ≤ B0 ; and since ∥G± ∥C m+1 ≤ 2B0 that, for each ± ∈ {+, −} and k ∈ J1, m + 1K,

e ∈ Admε/2 n · R ; 2 e ± (y) ≤ − ε , for y ∈ 1 − y0 n1/2 , (1 − y0 )n1/2 ;


   
G − ≤ ∂y G
n′ ε 2
(8.13)   ′ k−1
e 0, 1 = G e ± 0, 1 = 0; e ± ]k ≤ 2B0 n
  
G [G]
e k , [G ,
2 2 n
The proof of Proposition 8.1 will make use of the following lemma, to be established in Section 8.3
below, that approximates the locations of the middle paths in w e − (s) and w
e + (s).
Lemma 8.5. There exists a coupling between w e − , w,
e and we + , and a constant c = c(ε, δ, B0 , m) > 0
−1 −c(log n)2
such that P[A1 ] ≥ 1 − c e , where we have defined the event
(   )
− −
A1 = ′ 2ω/3−1

sup max e (s) − w
w e (0) ≤ (n )
j j
s∈[0,s̃0 +s̃1 ] j∈Jn′ /2−2n′′ ,n′ /2+2n′′ K
(8.14) (   )
e+ (0) ≤ (n′ )2ω/3−1 ∩ A0 .
+
∩ sup max e (s) − w
w
j j
s∈[0,s̃0 +s̃1 ] j∈Jn′ /2−2n′′ ,n′ /2+2n′′ K
LOCAL STATISTICS AND CONCENTRATION FOR NON-INTERSECTING BROWNIAN BRIDGES 59

Using Lemma 8.5, we can establish Proposition 8.1.

Proof of Proposition 8.1. Fix a = ∂t G(t0 , y0 ), which is uniformly bounded in [−B0 , B0 ] by Assump-
tion 1.3. We claim on the event A1 from (8.14) that
w̆j− (0) ≤ ŭj ≤ w̆j+ (0);
w̆j− (s0 + s1 ) + (a − n−ω/2 )(s0 + s1 ) ≤ v̆j ≤ w̆j+ (s0 + s1 ) + (a + n−ω/2 )(s0 + s1 );
(8.15)
f˘− (s) + (a − n−ω/2 )s ≤ f˘(s) ≤ f˘+ (s) + (a + n−ω/2 )s,
ğ − (s) + (a − n−ω/2 )s ≤ ğ(s) ≤ ğ + (s) + (a + n−ω/2 )s,
q ′ ′ y
for each integer j ∈ n2 − n′′ , n2 + n′′ and real number s ∈ [0, s0 + s1 ].
Let us quickly establish the proposition assuming (8.15). To that end, denote the (2n′′ +1)-tuples
ů = (ŭn′ /2−n′′ , . . . , ŭn′ /2+n′′ ) ∈ W2n′′ +1 and v̊ = (v̆n′ /2−n′′ , . . . , v̆n′ /2+n′′ ) ∈ W2n′′ +1 . Similarly, for
each index ± ∈ {+, −} and real number s ∈ [0, s0 + s1 ], denote
 
ẘ± (s) = w̆n±′ /2−n′′ (s) + (a ± n−ω/2 )s, . . . , w̆n±′ /2+n′′ (s) + (a ± n−ω/2 )s ∈ W2n′′ +1 ;

f˚± (s) = f˘± (s) + (a ± n−ω/2 )s; g̊ ± (s) = ğ ± (s) + (a ± n−ω/2 )s.

Then, by the second part of Lemma 2.15 and Remark 2.1, the laws of the processes x̆j (s) and
′ ′
ẘj± (s) for each ± ∈ {+, −} and (j, s) ∈ n2 −n′′ , n2 +n′′ ×[0, s0 +s1 ] are given by Qfů;v̊
 q y −1
˘;ğ (n ) and
ẘ± (0);ẘ± (s0 +s1 )
Qf˚± ;g̊± (n−1 ), respectively. This, together with (8.15) and Lemma 2.3 yields a coupling
on A1 that w̆− (s) + (a − n−ω/2 )s = ẘ− (s) ≤ x̆(s) ≤ ẘ+ (s) = w̆+ (s) + (a + n−ω/2 )s for each
s ∈ [s0 , s1 ], verifying the proposition.
The second bound in the first statement of (8.15) follows the fact that w̆± (0) = ŭ± ; (8.4); the
fact that A1 ⊆ A0 ; and the fact that on A0 we have (by (8.2)) that uj ≤ G(t0 −s0 , jn−1 )+nω/4−1 ≤
q ′ ′′ ′ ′′ y
G+ (jn−1 ) = ŭ+ j for j ∈/ n2 − 2n3 , n2 + 2n3 . The proof of the first bound there is entirely analogous.
For the second statement of (8.15), we only show that w̆j− (s0 + s1 ) + (a − n−ω/2 )(s0 + s1 ) ≤ v̆j ,
as the proof that v̆j ≤ w̆j+ (s0 + s1 ) + (a + n−ω/2 )(s0 + s1 ) is entirely analogous. Since we have
restricted to the event A0 (and since n = (n′ )2 ), we have that
n′
  
n n j ω/4−1 e se0 + se1 , j − (n′ )ω/2−1 ,
 
· v̆ j ≥ · G t 0 + s1 , y0 − + − n = G
n′ n′ 2n n n′
and so it suffices to show that
e se0 + se1 , j − (a − n−ω/2 )(e
 
ej− (e
w s0 + se1 ) ≤ G s0 + se1 ) − (n′ )ω−1 .
n′
q ′ ′
Thus, since we have restricted to the event A1 from (8.14); since for j ∈ n2 − n′′ , n2 + n′′ we have
y

n  j
wej− (0) = ′ · ŭ− j ≤ G 0, ′ + (n )
e ′ 2ω/3−1
,
n n
by (8.4), (8.12), and the definition of G− ; and since se0 = (n′ )ω−1 (by (8.1) and (8.11)), it suffices
to show that
e 0, j ≤ G e se0 + se1 , j − a(e
   
G s0 + se1 ) + (n′ )−ω se1 − 2(n′ )ω−1 .
n′ n′
60 AMOL AGGARWAL AND JIAOYANG HUANG

This follows from the fact that


′′ 2

e 0, j − G e se0 + se1 , j − a(e ≤ B0 se0 + se1 + n
   
≤ 9B0 (n′ )12ω−2 + (n′ )−1

G s0 + s1 )
′ ′ ′
e
n n n
≤ (n′ )−ω se1 − 2(n′ )ω−1 ,
for sufficiently large n. Here, in the first inequality we applied a Taylor expansion, the fact that

a = ∂t G(t0 , y0 ), the fact that j − n2 ≤ n′′ , and the fact that ∥G∥C 2 (R) ≤ B0 ; in the second
s1 ≤ 2(n′ )6ω−1 and that n′′ = (n′ )1/2 ; and in the third
inequality, we used the facts that se0 + se1 ≤ 2e
inequality we used the facts that ω ≤ 20 (and again that se1 = (n′ )6ω−1 ).
1

It therefore remains to establish the last two statements of (8.15); among these, we only verify
the bound f˘− ≤ f˘ there, as the proofs of the remaining three are entirely analogous. To that end,
observe on A1 that for each s ∈ [0, s0 + s1 ] we have, denoting se = ns
n′ ,
(8.16)
n ˚−
· f (s) = w en−′ /2+n′′ +1 (e
s) + (a − n−ω )es≤w en−′ /2+n′′ +1 (0) + ae s + (n′ )2ω/3−1
n′
′′
e − 1 + n + 1 + ae
 
=G s + (n′ )2ω/3−1
2 n′
′′
e 0, 1 + n + 1 + ae
 
=G ′
s − (n′ )ω−1 + (n′ )2ω/3−1 ,
2 n
where in the first statement we applied the definitions of f˚ and f˘− (as well as (8.11)); in the
second we used the definition of A1 from (8.14); in the third we used the fact that w̆n−′ /2+n′′ +1 (0) =
′′
G− y0 + n n+1 (by (8.4)) and the rescaling (8.11); and in the fourth we used the definition of G e−


in terms of G. e We further have that


 1 n′′ + 1 
Ge 0, + s − (n′ )ω−1 + (n′ )2ω/3−1
+ ae
2 n′
′′ ′ ′′ 2
e se, 1 + n + 1 − (n′ )3ω/4−1 + B0 n se + n + 1
 
≤G
2 n′ n n′
2 ′
2B0 se n n n
≤x en′ /2+n′′ +1 (es) − (n′ )3ω/4−1 + + 2(n′ )ω/2−1 ≤ ′ · x̆n′ /2+n′′ +1 (s) = ′ · f˘(s),
n n n
where in the first statement we applied  a Taylor −expansion for G,
e using the facts that a =

∂t G(t0 , y0 ) = ∂t G (t0 , y0 ) = ∂t G −
e 0, 1
and that [G e ]2 = ·[G ]2 ≤ 2B0 n′ ; in the second we used
n ′

2 n n
the fact that we are restricting to the event A0 , (8.12), and the fact that n′ n−1 (n′′ n′−1 )2 = n−1 ;
in the third we used (8.11) and the fact that se2 ≤ 4es21 = 4(n′ )12ω−2 ≤ (n′ )−1 (as ω ≤ 201
); and in
˘
the fourth we used the definition of f . Combining this with (8.16) yields the first inequality in the
third statement of (8.15). As mentioned previously, the proof of the second inequality is entirely
analogous, as is the proof of the fourth statement of (8.15); this establishes the proposition. □
8.3. Proof of Lemma 8.5. To establish Lemma 8.5, we set some additional notation, which will
eventually more precisely prescribe the limiting behaviors of the paths in w(s). e To that end,
let He : [0, se0 + se1 ] × R → [0, 1] denote the height function associated with the inverted height

function G e (recall (3.4)), given by H(t, e x) = sup y ∈ R : G(t,
e y) ≥ x . In particular, for any
(t, x) ∈ [0, se0 + se1 ] × [−nω , nω ], we have
e x) is the unique value of y ∈ − (n′ )3ω/2 , (n′ )3ω/2 such that G(t,
 
(8.17) H(t, e y) = x,
LOCAL STATISTICS AND CONCENTRATION FOR NON-INTERSECTING BROWNIAN BRIDGES 61

where this uniqueness follows from the fact that G(s, e y) is decreasing in y (by (8.13)), and we
implicitly used the fact that any (s, y) ∈ [0, se0 + se1 ] × − (n′ )3ω/2 , (n′ )3ω/2 is in the domain of G
 
e
′ 1+2ω ′ ′ 1+2ω ′
since 0 ≤ y0 − (n )n n
≤ y0 + (n )n n
≤ 1 (as min{y0 , 1 − y0 } ≥ dist (t0 , y0 ), ∂R ≥ n−δ ,

− 2n − 2n
as stipulated in Theorem 1.5). Next, define the density ϱe0 ∈ L1 (R) by setting
ϱe0 (x) = −∂x H(0,
e x), e 1) ≤ x ≤ G(0,
if G(0, e 0);
(8.18)
ϱe0 (x) = 0, if x < G(0,
e 1) or if x > G(0, e 0).
R∞  
Observe in particular that −∞ ϱe0 (x)dx = H e 1) − H
e 0, G(0, e 0, G(0,
e 0) = 1. Denoting the
measure µ e0 = ϱe0 (x)dx, we then find that µ e0 ∈ P0 , that is, it is a probability measure. Denote
(t)
its free convolution with the rescaled semicircle law (recall Section 2.2) by µ et = µe0 ⊞ µsemci for
1
any t > 0; also denote the associated density by ϱet ∈ L (R), which satisfies µ et (dx) = ϱet (x)dx.
Denote the classical locations (recall Definition 2.16 with n replaced by n′ ) with respect to µ et by
µ̃t ′
ej (t) = γj;n′ for each j ∈ J1, n K.
γ
We define the analogous quantities associated with G e − and G e + very similarly. In particular, for
e ± (x) = sup y ∈ R : G e ± (y) ≥ x , so that for any x ∈ [−nω , nω ]

each index ± ∈ {+, −}, we define H
e ± (x) is the unique value of y ∈ − (n′ )3ω/2 , (n′ )3ω/2 such that G e ± (y) = x. We
 
we have that H
± ± e ± (x) if G
e ± (1) ≤ x ≤ G e ± (0), and
further define the density ϱe0 ∈ L1 (R) by setting ϱe0 (x) = −∂x H
± ± ±
ϱe0 (x) = 0 otherwise. Then, µ e0 = ϱe0 (x)dx is a probability measure. Denote its free convolution
(t)
with the semicircle law by µ e±
t =µ e±
0 ⊞ µsemci and the associated density by ϱ e± 1
t ∈ L (R). Also denote
µ̃±
the classical locations with respect to µ e±t by γej± (t) = γj;n
t
′.

The next lemma provides bounds on this density ϱet , as well as on the classical locations γ
ej (t)
(and their analogs for ϱe±
t and γ
e ±
j (t)); it is established in Section 8.4 below.
Lemma 8.6. There exists a constant C = C(ε, δ, B0 , m) > 1 such that the following hold for each
integer n ≥ C and index ± ∈ {+, −}.
(1) If (t, x) ∈ [0, n−ω ] × [−C −1 , C −1 ], then −4ε−1 ≤ ϱet (x) ≤ − 4ε and −4ε−1 ≤ ϱe± t (x) ≤ − 4 .
ε

en′ /2+i (0) ≤ Ct nn + n|i|′ + t
q y
(2) If i ∈ − (n′ )1−ω , (n′ )1−ω and t ∈ [0, n−ω ], then γ

en′ /2+i (t) − γ

en±′ /2+i (0) ≤ Ct nn + n|i|′ + t .
± 
and γ en′ /2+i (t) − γ
We also require the below lemma indicating a monotone coupling for Dyson Brownian motion,
established in Section 8.4 below.
Lemma 8.7. Let M ≥ 1 and k ≥ 2 be integers, and fix k M -tuples λ(1) , λ(2) , . . . , λ(k) ∈ WM with
(j) (j) (j) 
λ(1) ≤ λ(2) ≤ · · · ≤ λ(k) . For each integer j ∈ J1, kK, let λ(j) (s) = λ1 (s), λ2 (s), . . . , λM (s)
denote Dyson
 (j) Brownian motion run for time s with initial data λ(j) . Then it is possible to couple
the λ (s) such that λ (s) ≤ λ(2) (s) ≤ · · · ≤ λ(k) (s), for each real number s ≥ 0.
(1)

Now we can establish Lemma 8.5.


Proof of Lemma 8.5. By (8.3), we may restrict to the event A0 in what follows. Define the two
n′ -tuples λ− = (λ− − − + + + +
1 , λ2 , . . . , λn′ ) ∈ Wn′ and λ = (λ1 , λ2 , . . . , λn′ ) ∈ Wn′ by setting

e ± 2j − 1 ,
 
λ±
j =G for each index ± ∈ {+, −} and integer j ∈ J1, n′ K.
n′
For each index ± ∈ {+, −}, let λ± (s) = λ± ± ±

1 (s), λ2 (s), . . . , λn′ (s) denote Dyson Brownian motion,
run for time s, with initial data λ± . Since we have restricted to the event A0 , we have by (8.12);
62 AMOL AGGARWAL AND JIAOYANG HUANG

(8.4) (with the definitions of G− and G+ ); and the fact from (8.13) that ∂y G e ± ≤ − ε , that λ− −
2
′ 3ω/5−1 − − ′ 3ω/5−1 + ′ 3ω/5−1 e (0) ≤ λ +2(n′ )3ω/5−1 . Thus,
+ +
2(n ) ≤w e (0) ≤ λ +2(n ) and λ −2(n ) ≤w
by Lemma 8.7, it is possible to couple the processes λ− (s), w e − (s) such that λ− (t)−2(n′ )3ω/5−1 ≤


e − (t) ≤ λ+ (t) + 2(n′ )3ω/5−1 for each t ≥ 0; similarly, it is possible to couple λ+ (s), w

w e + (s) such
that λ+ (t) − 2(n′ )3ω/5−1 ≤ w e + (t) ≤ λ− (t) + 2(n′ )3ω/5−1 for each t ≥ 0. Hence, it suffices to show
" #
\ n o 2
− −
P A0 ∩ ′ ω/2−1
≥ 1 − c−1 e−c(log n) ;

λ (s) − λ (0) ≤ (n )
j j
s∈[0,s̃0 +s̃1 ]
j∈Jn′ /2−2n′′ ,n′ /2+2n′′ K
(8.19) " #
\ n o 2
P A0 ∩ λ+ (s) − λ+ (0) ≤ (n′ )ω/2−1 ≥ 1 − c−1 e−c(log n) .

j j
s∈[0,s̃0 +s̃1 ]
j∈Jn′ /2−2n′′ ,n′ /2+2n′′ K

We only show the first bound in (8.19), as the proof of the second is entirely analogous. Under
this notation, Lemma 2.17 applies with the (n, λ) there given by (n′ , λ− ) here and the µ there given
by µe−
0 here. It therefore yields a constant c1 > 0 such that
2
P[E− −1 −c1 (log n)
1 ] ≥ 1 − c1 e ,
where we have defined the event
\n′ \  
E−
1 = γ
e −
j+⌊(log n)5 ⌋ (t) − n −1
≤ λ −
j (t) ≤ γ
e −
j−⌊(log n)5 ⌋ (t) + n −1
.
j=1 t∈[0,1]

Observe on E−
1 that

λ (s) − λ− (0) ≤ γ − − −1

j j ej−⌊(log n)5 ⌋ (s) − γ
ej+⌊(log n)5 ⌋ (0) + 2n .
q ′
This yields on E− 1 a constant C1 = C1 (ε, δ, B0 , m) > 1 such that, for each pair (j, s) ∈ n2 −
′ y
2n′′ , n2 + 2n′′ × [0, n−ω ], we have
 ′ ′′
λ (s) − λ− (0) ≤ C1 s n + n + s + γ 2
− 
− −

j j ej−⌊(log n) 5 ⌋ (0) − γ
ej+⌊(log n) 5 ⌋ (0) +
n n ′ n
 1  4(log n)5 2
≤ 2C1 s ′′ + s + + ,
n εn′ n
where in the first inequality we used the second part of Lemma 8.6; in the second inequality we
− − ε 1 1

used (8.1) and the fact that ϱe0 = −∂x G ≥ 2 . Together with the fact that s n′′ + s ≤ n′ for
e
sufficiently large n (since n′′ = (n′ )1/2 , s ≤ se0 + se1 ≤ 2(n′ )6ω−1 , and ω ≤ 20 1
), this implies the first
statement of (8.19) and thus the lemma. □
8.4. Proofs of Lemma 8.6 and Lemma 8.7. In this section we establish first Lemma 8.6 and
then Lemma 8.7.
Proof of Lemma 8.6. We only establish the estimates in the lemma for ϱet and γ ej (t), as the proofs
of the analogous bounds for ϱe± ej± (t) are entirely analogous. We begin with the first statement
t and γ
e ∈ Admε/2 n′ · R (by (8.13)), we have

of the lemma. To that end observe that, since G n
2 ε  
(8.20) − ≤ ϱe0 (x) ≤ − , for each x ∈ G(0,
e 1), G(0,
e 0) .
ε 2
LOCAL STATISTICS AND CONCENTRATION FOR NON-INTERSECTING BROWNIAN BRIDGES 63

Again by (8.13), we also have that [G]e k ≤ 2B0 n′ for each integer k ∈ J2, mK, from which it quickly
n
follows from (8.17), (8.18), and (8.20) that there exists a constant C1 = C1 (ε, δ, B0 , m) > 1 with
C 1 n′
(8.21) ϱ0 ]k ≤
[e , on supp ϱe0 , for each k ∈ J1, m − 1K.
n
As such, Lemma 2.11 applies to ϱe0 around any point in the interior of supp ϱe0 and yields (since
e ≤ − ε , by (8.13)) a constant C2 = C2 (ε, δ, B0 , m) > 1 such that
e 0, 1 = 0 and ∂ G

G 2 2

max ∂t ∂xj ϱe0 (x) ≤ C2 .



(8.22) sup sup ∂t ϱet (x) ≤ C2 ; sup sup
t∈[0,1/C2 ] |x|≤1/C2 t∈[0,1/C2 ] |x|≤1/C2 j∈J1,m−3K

The first bound in (8.22), with the first two statements of (8.13), yields the first statement of the
lemma for sufficiently large n (so that 4C2 n−ω ≤ ε). q y
To establish the second part of the lemma, first observe for i ∈ − (n′ )1−ω , (n′ )1−ω that
   
 1 i  2i 2i 2 2
(8.23) en′ /2+i (0) = G 0, + ′ ∈ − ′ , ′ ⊆ −
γ e , ,
2 n εn εn ε(n′ )ω ε(n′ )ω
due to the first and third statements of (8.13). We then claim that there exists a constant C3 =
C3 (ε, δ, B0 , m) > 1 such that (recalling the Hilbert transform from (2.7))
 ′
H ϱes (y) ≤ C3 n + |y| + s ,

for each (s, y) ∈ [0, n−ω ] × − (n′ )−ω/2 , (n′ )−ω/2 .
 
(8.24)
n
Before verifying (8.24) let us confirm the second part of the lemma assuming it. To that end,
observe for any j ∈ J1, n′ K that
Z ∞ Z ∞ Z ∞ Z ∞ Z t
2j − 1
ϱet (x)dx − = ϱ t (x)dx − ϱ 0 (x)dx = ∂s ϱes (x)dsdx
n′
e e
γ̃j (0) γ̃j (0) γ̃j (0) γ̃j (0) 0
Z tZ ∞

=π ∂x ϱes (x)H ϱes (x) dxds
0 γ̃j (0)
Z t  
=π ej (0) · H ϱes γ
ϱes γ ej (0) ds,
0

where the first statement follows from the definition of the classical location γ ej (0); the second from
performing the integration in s; the third from integrating (2.12); and the fourth from performing

the integration
q in x (using ythe fact that ϱes (x) = 0 for x sufficiently large). This with j = n2 + i
for i ∈ − (n′ )1−ω , (n′ )1−ω , together with (8.24), (8.23), and the fact (from the first part of the
lemma) that ϱes (x) ≤ 4ε−1 for each (s, x) ∈ [0, n−ω ] × [−n−ω/8 , n−ω/8 ], yields (since |γj | ≤ 3|i|

εn , by
the second and third statements in (8.13))
1 i 8πC3 t  n′ 3|i|  Z ∞ 1 i 8πC3 t  n′ 3|i| 
+ ′− + ′ +t ≤ ϱet (x)dx ≤ + ′ + + ′ +t .
2 n ε n εn γ̃n′ /2+i (0) 2 n ε n εn
R∞ ′
Together with the facts that γ̃ ′ ϱe (x)dx = 21 + 2i−1
(t) t n′ for each j ∈ J1, n K and that ϱ et (x) ≥ 2ε
n /2+i

for |x| ≤ n−ω/8 , it follows that


30πC3 t  n′ |i| 
γen′ /2+i (t) − γ
en′ /2+i (0) ≤ + + t ,
ε3 n n′
which verifies the second part of the lemma.
64 AMOL AGGARWAL AND JIAOYANG HUANG


It then remains to establish (8.24), to which end we will bound H ϱe0 (y) and ∂t H ϱes (y) . First
 
observe that supp µ
e0 = G(0,
e 1), G(0,
e 0) and that


ε ε e 1) ≤ 4B0 n ,
− 2ε−1 ≤ G(0, e 0) ≤ 2ε−1 ;

(8.25) e 1) ≤ − ≤ ≤ G(0, G(0,
e 0) + G(0,
4 4 n
where the first statement follows from the first and third statements of (8.13), and the second
follows from the first and last statement of (8.13). We next have that
Z Z
∞ ϱe (w)dw ∞ ϱe (w) − ϱe (y)dw
−1 0 −1 0 0
(8.26) H ϱe0 (y) = π =π .

−∞ w − y −∞ w−y

Thus, for y ∈ [−n−ω/8 , n−ω/8 ], we have

(8.27)
Z Z Z ∞
G̃(0,0) ϱe (w) − ϱe (y)dw G̃(0,1) dw dw
−1
H ϱe0 (y) ≤ π 0 0 −1
+ π ϱe0 (y) +

w−y w−y G̃(0,0) w − y

G̃(0,1) −∞
C1 n′ e 
e 1) + ε−1 G(0,

|w − y|−2

≤ G(0, 0) − G(0, e 0) + G(0, e 1) + 2|y| · max
n w∈[
/ G̃(0,1),G̃(0,0)]

4C1 n′ 256B0  n′ 
≤ + + |y| ,
εn ε3 n
 
where the first bound follows from (8.26) and the fact that supp ϱe0 = G(0, e 1), G(0,e 0) ; the second
follows from (8.21) and (8.20); and the third follows from (8.25) (using the fact that |w − y| ≥ 8ε
for |y| ≤ n−ω and w ∈
 
/ G(0,
e 1), G(0,
e 0) , due to the first statement of (8.25)).
Next, we claim that there exists a constant C4 = C4 (ε, δ, B0 , m) > 1 such that

(8.28) sup sup ∂t H ϱes (y) ≤ C4 .
s∈[0,n−ω ] |y|≤(n′ )−ω/2

Together with (8.27), this would yield (8.24) and thus the lemma. To that end, observe that
Z
∞ ∂ ϱe (w)dw Z ∞ ∂ ϱe (w)H ϱe (w)dw
−1 t s x s s
(8.29) ∂t H ϱes (y) = π = ,

−∞ w − y −∞ w−y

where in the second equality we used (2.12). Further observe for s ∈ [0, n−ω ] that
(s) 
e 0) + 2t1/2 ⊆ [−4ε−1 , 4ε−1 ],
e 1) − 2t1/2 , G(0,

(8.30) supp ϱes ⊆ supp µ esemci ⊆ G(0,
e0 + supp µ

(s)
where the first holds since µe =µ e0 ⊞ µsemci , and the second holds by (2.9), (2.8), and the fact that
 s
supp µ
e0 = G(0, 1), G(0, 0) ; and the third holds by (8.25). Thus, due to (8.29), (8.30), (8.20), and
e e
(8.21), to verify (8.28) it suffices to show the existence of a constant C5 = C5 (ε, δ, B0 , m) > 1 such
that

sup ∂yk H ϱes (y) ≤ C5 ;



(8.31) sup for each k ∈ {0, 1, 2}.
s∈[0,n−ω ] |y|≤1/C5
LOCAL STATISTICS AND CONCENTRATION FOR NON-INTERSECTING BROWNIAN BRIDGES 65

To confirm (8.31), fix k ∈ {0, 1, 2}, and observe that


Z Z
∞ 1/C2 ∂ k ϱe (y + w)dw Z
k −1

k ϱ s (w)dw y s ϱ s (w)dw

∂y H ϱes (y) = π ∂y e e
+

w − y w (w − y)k+1
−∞ −1/C2 |w−y|>1/C2
Z
1/C2 ∂ k ϱe (y + w) − ϱe (y)dw
y s s
≤ + (2C2 )3 ≤ 10C23 ,

−1/C2 w
where in the first statement we used (2.7); in the second we changed variables in the first integral
by replacing w with y + w; in the third we used the facts that k ≤ 2 and that ϱet is a probability
measure; and in the fourth we used (8.21) and the second bound in (8.22). This verifies (8.31),
which as mentioned above establishes the lemma. □
Proof of Lemma 8.7. Let G(s) = GM (s) denote an M × M Hermitian Brownian motion as de-
scribed in Section 2.3, and for every j ∈ J1, kK let Λ(j) denote the M × M diagonal matrix whose
(j)
(i, i) diagonal entry is given by λi (0), for each i ∈ J1, M K. Since λ(1) ≤ λ(2) ≤ · · · ≤ λ(k) , we have
Λ(1) ≤ Λ(2) ≤ · · · ≤ Λ(k) , where we write Λ ≤ Λ′ if Λ′ − Λ is nonnegative definite; in particular,
Λ(1) + G(s) ≤ Λ(2) + G(s) ≤ · · · ≤ Λ(k) + G(s). Together with the fact that (by the first part of
Lemma 2.15) λ(j) (s) has the same law as the eigenvalues of Λ(j) + G(s) for each j, jointly in s ≥ 0,
this yields λ(1) (s) ≤ λ(2) ≤ · · · ≤ λ(k) (s) for each s ≥ 0. □

9. Regularity Properties of Solutions


In this section we state and establish various regularity properties of solutions to the equation
(1.2) governing the limit shape for non-intersecting Brownian bridges. These include comparison
principles (Lemma 9.1 below); derivative bounds (Lemma 9.2 and Lemma 9.4 below); and exponen-
tial decay of discrepancy estimates (Proposition 9.5 below). We also provide a result that explains
how perturbing regular solutions to (1.2) preserves their regularity (Lemma 9.6 below), thereby
showing that Assumption 1.3 is in a sense stable under smooth perturbations of boundary data.
9.1. Regularity Estimates. In this section we state estimates for solutions to (1.2). In what
follows, we recall the sets of functions Adm(R) and Adm  ε (R) from Section 1.3. For any real
number r > 0 and point z ∈ R2 , we recall that Br (z) = z ′ ∈ R : |z ′ − z| < r denotes the open
disk of radius r centered at z; if z = (0, 0), we abbreviate Br (z) = Br . We begin with the below
maximum and comparison principles for solutions of (1.2), which are quick consequences of the
analogous principles for solutions to nonlinear elliptic partial differential equations; it is established
in Appendix E.1 below.
Lemma 9.1. Fix an open set R ⊂ R, and let F1 , F2 , F ∈ Adm(R) ∩ C 2 (R) be solutions to (1.2) on
R.
(1) If F1 (z) ≤ F2 (z) for each z ∈ ∂R, ≤ F2 (z) for each z ∈ R.
then F1 (z)
(2) We have supz∈R F1 (z) − F2 (z) ≤ supz∈∂R F1 (z) − F2 (z) . In particular, supz∈R F (z) =

supz∈∂R F (z) .
We next have the following lemma indicating boundedness of the derivatives for a solution to
(1.2); it is established in Appendix E.1 below, again as a consequence of its analog in the context
of elliptic partial differential equations. In what follows, R′ ⊆ R is a subset that accounts for
boundary regularity, in that F is regular in the interior of R but up to the boundary of R′ ; in
particular, if R′ is empty, the below lemma is an interior regularity estimate and, if R′ = R, then
66 AMOL AGGARWAL AND JIAOYANG HUANG

Figure 9. Shown on the left is depiction for Lemma 9.2. Shown in the middle
is a depiction for Proposition 9.5 stating, if F1 (z) = F2 (z) for each z on the blue
sides of R, then F1 − F2 is exponentially small in the shaded region. Shown on the
right is a depiction for Lemma 9.6.

it is a global one. Here, for any sets S and S ′ , we let S \ S ′ = S \ (S ∩ S ′ ). See the left side of
Figure 9.
Lemma 9.2. For any integer m ≥ 1; real numbers r > 0, ε ∈ (0, 1), and B > 1; and bounded
open subsets R′ ⊆ R ⊂ R2 such that R′ has a smooth boundary, there exists a constant C =
C(ε, r, B, m, R′ ) > 1 such that the following holds. Let f ∈ C(∂R) be a function satisfying ∥f ∥0 ≤ B,
and let F ∈ Admε (R) ∩ C 2 (R) be a solution to (1.2) on R with boundary data F |∂R = f . Assume

that there exists a function φ ∈ C m+1 (R ) with ∥φ∥C m+1 (R′ ) ≤ B, such that f (z) = φ(z) for each

z ∈ ∂R ∩ ∂R. Letting
Dr = z ∈ R : dist(z, ∂R \ ∂R′ ) > r ,

(9.1)
we have ∥F ∥C m (Dr ) ≤ C.
The following lemma states that the solutions of (1.2) are real analytic (that is, they have an
absolutely convergent power series expansion at each point in the interior of their domains); it is
shown in Appendix E.1 below, as a quick consequence of classical results stating real analyticity
for solutions of elliptic partial differential equations.
Lemma 9.3. Fix a real number ε ∈ (0, 1) and some open set R ⊂ R2 ; let F ∈ Admε (R) ∩ C 2 (R)
denote a solution to (1.2) on R. Then, F is real analytic R.
LOCAL STATISTICS AND CONCENTRATION FOR NON-INTERSECTING BROWNIAN BRIDGES 67

We next have the following lemma bounding the effect of a C 0 boundary perturbation on the
derivatives of a solution to (1.2); its proof is similar to [2, Proposition 2.13] and is given in Appen-
dix E.2 below.
Lemma 9.4. For any integer m ≥ 1; real numbers r > 0, ε ∈ (0, 1), and B > 1; and bounded
open subsets R′ ⊆ R ⊂ R2 such that R′ has a smooth boundary, there exists a constant C =
C(ε, r, B, m, R′ ) > 1 such that the following holds. For each index i ∈ {1, 2}, let Fi ∈ Admε (R) ∩
C 2 (R) denote a solution to (1.2) such that supz∈∂R Fi (z) ≤ B. Assume that there exist functions
φi ∈ C m+5 (R) such that ∥φi ∥C m+5 (R) < B and Fi (z) = φi (z) for each z ∈ ∂R′ ∩ ∂R. Then, letting
Dr ⊆ R be as in (9.1), we have

∥F1 − F2 ∥C m (Dr ) ≤ C ∥F1 − F2 ∥C 0 (R) + ∥φ1 − φ2 ∥C m+2 (R) .
The following result is established in Section 9.2 below. It states that, given two solutions F1 , F2
to (1.2) on a tall rectangle of height 1 and width L−1 (with L bounded below), whose boundary
data match on its west and east boundaries, |F1 − F2 | decays exponentially in L in the middle of
the rectangle. As such, it can be viewed as indicating that the “discrepancy” between F1 and F2
(originating from their different boundary data along the north and south sides of the rectangle)
decay exponentially around the middle of the rectangle. See the middle of Figure 9.
Proposition 9.5. For any real numbers ε, r ∈ 0, 41 and B > 1, there exists a constant c =


c(ε, r, B) > 0 such that the following holds. Let L > 0 be a real number, and define the open
rectangle R = (0, L−1 ) × (−1, 1) ⊂ R2 . Let F1 , F2 ∈ Admε (R) ∩ C 5 (R) be two solutions to (1.2)
such that ∥Fi ∥C 5 (R) ≤ B for each i ∈ {1, 2}. If F1 (t, x) = F2 (t, x) for any (t, x) ∈ ∂R with
t ∈ {0, L−1 }, then
F1 (t, x) − F2 (t, x) ≤ c−1 e−cL1/8 , for any (t, x) ∈ [0, L−1 ] × [r − 1, 1 − r].

We conclude this section with the following lemma, which is a consequence of the above bounds
and is established in Appendix E.3 below; it essentially states the following. Fix a solution F to
(1.2), bounded in C m for some integer m on a rectangle R, as well some boundary data g0 and
g1 on the two vertical sides on the rectangle; assume that g0 and g1 are small perturbations of F
(restricted to those boundaries). Then, it is possible to find a solution G to (1.2) on a slightly
shorter rectangle S, whose boundary data on the vertical sides of the rectangle are given by g0 and
g1 (the first condition in the below lemma), and that is close to F (quantified through the second
and third conditions of the below lemma). The second part of the lemma states that F and G are
close in any C k norm in the interior of S, and the third part states that F and G are close in C m−5
(that is, fewer derivatives than the original assumed bound on F ) up to the boundary of S. See
the right side of Figure 9.
Lemma 9.6. For any integers m, k ≥ 7, and real numbers ε > 0; r ∈ 0, 14 ; and B > 1, there


exist constants δ = δ(ε, B) > 0, C1 = C1 (ε, r, B, k) > 1, and C2 = C2 (ε, B, m) > 1 such that the
following holds. Fix a real number L > 2, and define the open rectangles
 1  r 1 − r r + 1 r + 1
R = 0, × (−1, 1); Sr = , × − 1, 1 − ; S = S0 .
L L L L L
Let F ∈ Admε (R) ∩ C m (R) denote a solution to (1.2) such that ∥F ∥C m (R) ≤ B, and define the
functions f0 , f1 : [−1, 1] → R by setting fi (x) = F (iL−1 , x) for each (i, x) ∈ {0,
1} × [−1, 1]. Further
let g0 , g1 : [−1, 1] → R denote two functions such that ∥gi ∥C m (−1,1) ≤ B and gi (x) − fi (x) ≤ δL−1
for each (i, x) ∈ {0, 1} × [−1, 1]. Then, there exists a solution G ∈ Admε/2 (S) ∩ C m−5 (S) to (1.2)
satisfying the following three properties.
68 AMOL AGGARWAL AND JIAOYANG HUANG

(1) For each i ∈ {0, 1} and x ∈ L−1 − 1, 1 − L−1 , we have G(iL−1 , x) = gi (x).
 

(2) We have ∥F − G∥C k (Sr ) ≤ C1 Lk · ∥f0 − g0 ∥C 0 + ∥f1 − g1 ∥C 0 .



(3) We have ∥F −G∥C m−5 (S) ≤ C2 Lm−5 · ∥f0 −g0 ∥C m−3 +∥f1 −g1 ∥C m−3 and ∥F −G∥C m−5 (S) ≤
3/m 3/m 
C2 Lm−5 · ∥f0 − g0 ∥C 0 + ∥f1 − g1 ∥C 0 .
9.2. Proof of Proposition 9.5. In this section  L we establish
 Proposition 9.5. In what follows, we
fix real numbers ε > 0, L > 1, and L0 ∈ 2B , 2BL . We also denote the rectangular open set
R0 = (0, 1) × (−L0 , L0 ) and let F1 , F2 ∈ Admε (R0 ) ∩ C 2 (R0 ) denote two solutions to the equation
(1.2).
The next lemma bounds F1 (z) − F2 (z) , for a point z ∈ R0 such that F1 = F2 on the part of
∂R0 near z; we establish it in Section 9.3 below.
Lemma 9.7. For any real number B > 1, there exist constants c = c(ε, B) > 0 and C = C(ε, B) > 1
such that the following holds. Fix a point z = (t0 , x0 ) ∈ R0 , such that [0, 1] × [x0 − 2L1/2 , x0 +
2L1/2 ] ⊆ R0 . Let A ≤ B be a real number, and assume the below three statements hold.

(1) For
any indices
i ∈ {1, 2} and j, k ∈ {t, x}, and point w ∈ R 0 , we have ∇Fi (w) ≤ B and
∂j ∂k Fi (w) ≤ BL−1 ; we also have ∥Fi ∥C 5 (R ) − ∥Fi ∥C 0 (R ) ≤ B.
0 0

(2) For any point w ∈ ∂R0 with |w − z| ≤ 2L1/2 , we have F1 (w) = F2 (w).
(3) For any point w ∈ ∂R0 , we have F1 (w) − F2 (w) ≤ A.
1/8
Then, F1 (z) − F2 (z) ≤ C(L−1/8 A + e−cL ).
We can quickly prove Proposition 9.5 by repeated application of Lemma 9.7.

Proof of Proposition 9.5. We may assume throughout this proof that L is sufficiently large (in a way
only dependent on ε and B), for otherwise the lemma holds by altering the values of c and C (using
the fact that F1 (t, x) − F2 (t, x) ≤ 2B). Define the rescaled domain Rb = L · R = (0, 1) × (−L, L)
2
and, for each index i ∈ {1, 2}, set Fbi ∈ C (R)
b by for each w ∈ [0, 1] × [−L, L] letting

Fbi (w) = L · Fi (L−1 w), so ∇Fbi (w) = ∇Fbi (L−1 w), and ∂j ∂k Fbi (w) = L−1 · ∂i ∂j Fi (L−1 w),

for any indices j, k ∈ {t, x}. In particular, the Fbi continue to solve (1.2).
1/8
It therefore suffices to show that Fbi (t, x) − Fb2 (t, x) ≤ Ce−cL

for each (t, x) ∈ [0, 1] × (r −

1)L, (1 − r)L . Denoting the open rectangle R b m = (0, 1) × (mL2/3 − L, L − mL2/3 ) for any integer
m ≥ 1, we will more generaly show by induction on m ∈ J0, L1/8 K that there exist constants
c = c(ε, B) > 0 and C = C(ε, B) > 1 such that
Fb1 (t, x) − Fb2 (t, x) ≤ 2BL−m/9 + Cme−cL1/8 , for any (t, x) ∈ R

(9.2) b m,

from which the lemma would follow upon taking m = ⌊L1/8 ⌋; see the left side of Figure 10.
Let us first show that the Fbi satisfy the first and third assumptions
of Lemma
9.7 (with
the −1
A
there equal to 2B here). Since ∥Fi ∥C 5 (R) ≤ B, we have that ∇Fi (w) ≤ B and ∂j ∂k Fi (w) ≤ BL
b b
for each w ∈ R b and that ∥Fbi ∥ 5
C (R̂m ) − ∥Fi ∥C 0 (R̂m
b ≤ B; this verifies the first assumption. Moreover,
)
since F1 (t, x) = F2 (t, x) for t ∈ {0, L }, since ∇Fi (w) ≤ B, and since R = (0, L−1 ) × (−1, 1) we
−1

have F1 (w) − F2 (w) ≤ 2BL−1 for each w ∈ R. Hence, Fb1 (w) − Fb2 (w) ≤ 2B for each w ∈ R, b
verifying the third assumption of Lemma 9.7. This in particular confirms (9.2) at m = 0.
LOCAL STATISTICS AND CONCENTRATION FOR NON-INTERSECTING BROWNIAN BRIDGES 69

We therefore assume in what follows that (9.2) holds at some value of m ∈ J0, L1/8 − 1K and
verify that it also holds with m replaced by m + 1. For any integer k ∈ J0, L1/8 K, set

ςk = sup Fb1 (w) − Fb2 (w) .
w∈Rk

Then the second assumption of Lemma 9.7 applies, with the (R0 , F1 , F2 , z) there given by the
b m+1 . Since L − (m + 1)L2/3 ≥ L + 2L1/2 for m ∈ J0, L1/8 K,
b m , Fb1 , Fb2 , z) with any z ∈ R
(R 2B
Lemma 9.7 therefore yields constants c1 = c1 (ε, B) > 0 and C1 = C1 (ε, B) > 1 such that
1/8
ςm+1 = sup Fb1 (w) − Fb2 (w) ≤ C1 L−1/8 · sup Fb1 (w) − Fb2 (w) + C1 e−c1 L

w∈R
b m+1 w∈R
bm
1/8 1/8
= C1 L−1/8 ςm + C1 e−c1 L ≤ L−1/9 ςm + C1 e−c1 L ,
where in the last inequality we used the fact that L is sufficiently large. Applying this with (9.2),
with the (c, C) there equal to (c1 , C1 ) here yields
1/8 1/8 1/8
ςm+1 ≤ 2BL−(m+1)/9 + C1 mL−1/9 e−c1 L + C1 e−c1 L ≤ 2BL−(m+1)/9 + C1 (m + 1)e−c1 L .
This verifies (9.2) (with the m there replaced by m + 1 here), establishing the proposition. □
To establish Lemma 9.7, we use a probabilistic interpretation of the Fi . To that end, for any
real numbers u ∈ R and v ∈ [−ε−1 , ε], we recall from Remark 3.19 the 2 × 2 matrix D = D(u, v).
Let W : R≥0 → R2 denote a standard two-dimensional Brownian motion. For any point z ∈ R0
and index i ∈ {1, 2}, define the process Yi = Yiz : [0, τi ] → R2 through the stochastic differential
equation
 1/2
(9.3) dYi (s) = D ∇Fi Yi (s) · dW (s), with initial data Yi (0) = z,
so that Y1 and Y2 are coupled through the same Brownian motion W (s); see the right side of
Figure 10 for a depiction. Here, the hitting time τi = τi;z is given by

τi = inf t ≥ 0 : Yi (s) ∈ ∂R0 .
That the stochastic differential equation (9.3) is well-defined follows from the fact that the coeffi-
1/2
cients D ∇Fi (t, x) of the equation are uniformly Lipschitz (as Fi ∈ C 2 (R0 )).

Lemma 9.8. For each index i ∈ {1, 2}, the process Fi Yi (s) is a martingale. In particular,
   1/2
(9.4) dFi Yi (s) = ∇Fi Yi (s) · D ∇Fi Yi (s) · dW (s).

Proof. Since τi = 0 if z ∈ ∂R0 , we may assume that z ∈ R0 . Then, denoting the Hessian of Fi by
 
∂tt Fi (z) ∂tx Fi (z)
H(z) = ,
∂xt Fi (z) ∂xx Fi (z)
we have by Itô’s lemma that
  
   1/2 1   
dFi Yi (s) = ∇Fi Yi (s) · D ∇Fi Yi (s) · dW (s) + Tr D ∇Fi Yi (s) H Fi Yi (s)
2
  1/2
= ∇Fi Yi (s) · D ∇Fi Yi (s) · dW (s),
where in the last equality we used the fact thatFi satisfies (1.2) (with the (t, y) there equal to Yi (s)
here). This yields (9.4) and thus that Fi Yi (s) is a martingale. □
70 AMOL AGGARWAL AND JIAOYANG HUANG

Figure 10. Shown to the left is the rectangle R b m used in the proof of Proposi-
tion 9.5. Shown to the right are the coupled stochastic processes Y1 (s) and Y2 (s)
as in (9.3).

Hence, we must analyze properties of the process Yi , to which end we have the following lemma.
Lemma 9.9. There exists a constant c = c(ε) > 0 such that, for any real number A > 1, index
i ∈ {1, 2}, and point z = (t0 , x0 ) ∈ R0 , we have
 i
P[τi;z ≥ A] ≤ c−1 e−cA ; Yi (s) − z ≥ A ≤ c−1 e−cA .

(9.5) P sup
s∈[0,τi;z ]

Proof. Denote W (s) = W1 (s), W2 (s) ∈ R, so that W1 , W2 : R≥0 → R are two independent
standard Brownian motions; also let Yi (s) = Yi;1 (s), Yi;2 (s) ∈ R2 . Then, by (9.3) and the fact
that D(u, v) = diag(1, π 2 v −4 ), we have Yi;1 (s) − t0 = W1 (s), for each s ∈ [0, τi;z ]. In particular we
have for any s ≤ min{τ1;z , τ2;z } that
(9.6) Y1;1 (s) = Y2;1 (s) = W1 (s) + t0 .
Since R0 ⊂ [0, 1] × R, it follows that τi;z ≥ A can only hold if Wi (s) = Yi;1 (s) ∈ [−1, 1], for each
s ∈ [0, A]. Thus,
" # " ⌊A⌋ #
\  \ n o
(9.7) P[τi;z ≥ A] ≤ P W1 (s) ∈ [−1, 1] = P W1 (k) − W1 (k − 1) ≤ 2 .
s∈[0,A] k=1
 
Now observe that there exists a constant c1 > 1 such that P W1 (k) − W1 (k − 1) ≤ 2 < 1 − c1 , for
any k ≥ 1. Hence,
" ⌊A⌋ # ⌊A⌋
\ n o Y h i
P W1 (k) − W1 (k − 1) ≤ 2 ≤ (1 − c1 )⌊A⌋ ,

P W1 (k) − W1 (k − 1) ≤ 2 =
k=1 k=1
which together with (9.7) yields the first bound in (9.5).
LOCAL STATISTICS AND CONCENTRATION FOR NON-INTERSECTING BROWNIAN BRIDGES 71

Next, since Fi ∈ Admε (R0 ), we have ∂x Fi (t, x) ∈ [−ε−1 , −ε] for each (t, x) ∈ R0 ; it follows that
D ∇Fi (t, x) 1/2 ≤ 5ε−2 for each (t, x) ∈ R0 . By the definition (9.3) of Yi (and the fact that Yi is


a time-changed Brownian motion by [48, Theorem 3.4.6]), this yields


   
−2

P sup Yi (s) − z ≥ A ≤ P sup W (s) ≥ 10ε A + P[τi ≥ A]

s∈[0,τi ] s∈[0,A]
 
≤ P sup W (s) ≥ 10ε−2 A1/2 + c−1 e−cA ,

s∈[0,1]

where in the last inequality we used the first bound in (9.5). This, together with the fact (see [48,

Chapter 4, Equation (3.40)]) that there is a constant c2 = c2 (ε) > 0 such that sups∈[0,1] W (s) ≤
10ε−2 A1/2 holds with probability at least 1 − c−1
2 e
−c2 A
, implies the second bound in (9.5). □

9.3. Proof Lemma 9.7. In this section we establish Lemma 9.7.

Proof of Lemma 9.7. We may assume throughout this proof that L > 10 (for otherwise
taking the
constant C = 2 in the lemma gives F1 (z) − F2 (z) ≤ supz′ ∈∂R0 F1 (z ′ ) − F2 (z ′ ) ≤ A ≤ CL−1/8 A,
by Lemma 9.1). Fix M = L1/4 ; abbreviate τi = τi;z for each i ∈ {1, 2}; set τ = min{τ1 , τ2 }; and
define the event
 
A = {τ ≤ M } ∩ so that P A∁ ≤ c−1 −c1 M
 
(9.8) max sup Yi (s) − z ≤ M ,

1 e ,
i∈{1,2} s∈[0,τi ]


for some constant  c1 = c1 (ε) > 0, by Lemma 9.9. Denote W (s) = W1 (s), W2 (s) and Yi (s) =
Yi;1 (s), Yi;2 (s) , for each index i ∈ {1, 2} and real number s ≥ 0. Recall from (9.6) that we have
Y1;1 (s) = W1 (s) + t0 = Y2;1 (s) for each
s ∈ [0, τ ]. Since on the event A we have Yi0 (τ ) ∈ ∂R0 for
some i0 ∈ {1, 2} while Yi0 ;2 (τ ) − x0 ≤ L1/4 ≤ L1/2 , and since (0, 1) × [x0 − 2L1/2 , x0 + 2L1/2 ] ⊆ R0 ,
it follows that Yi0 ;1 (τ ) ∈ {0, 1} (that is, the t-coordinate of Yi0 (τ ) causes it to be on ∂R0 ). Hence,

(9.9) Y1,1 (τ ) = Y2,1 (τ ) ∈ {0, 1}, and thus τ1 = τ2 , on the event A.

By (9.3), we have
  
 1/2  1/2
(9.10) d Y1 (t) − Y2 (t) = D ∇F1 Y1 (t) − D ∇F2 Y2 (t) · dW (t).

Thus, for any real numbers t > 0 and p > 1, the Burkholder–Davis–Gundy inequality [13, Theorem
3.2] (see also [14, Theorem 11.2.1]) yields

(9.11)
" Z 2p #

2p
 t 1/2 1/2
2p

E sup Y1 (s) − Y2 (s) ≤ (72p) · E D ∇F1 (Y1 (s) − D ∇F2 (Y2 (s)) .

ds
s∈[0,t] 0

To bound the integral on the right side of (9.11), first observe that since ∂x F1 (w), ∂x F2 (w) ∈
[−ε−1 , −ε] that the D(∇Fi ) are uniformly smooth (in a way dependent on ε) in ∇Fi . Thus, there
72 AMOL AGGARWAL AND JIAOYANG HUANG

exists a constant C1 = C1 (ε) > 0 such that



1/2  1/2
D ∇F1 Y1 (t) − D ∇F 2 Y2 (t)

 
≤ C1 ∇F1 Y1 (t) − ∇F2 Y2 (t)

(9.12)  
   
≤ C1 ∇F1 Y1 (t) − ∇F1 Y2 (t) + ∇F1 Y2 (t) − ∇F2 Y2 (t)


≤ C1 BL−1 Y1 (t) − Y2 (t) + C1 ∇F1 Y2 (t) − ∇F2 Y2 (t) ,



where in the last bound we used the first assumption of the lemma ∂j ∂k Fi (w) ≤ BL−1 . Since
we deterministically have Y1 (t) − Y2 (t) ≤ 4L + 1 ≤ 5L (as R0 ⊆ [0, 1] × [−2L, 2L]) and ∇Fi (w) ∈
[−ε−1 , −ε], we therefore have that
" Z t 2p #

≤ P A∁ · (C1 t)2p (5B + ε−1 )2p
  
E 1A∁ · D ∇F1 (Y1 (s)) − D ∇F2 (Y2 (s)) ds

(9.13) 0

≤ (C2 t)2p e−c1 M ,

for some constant C2 = C2 (ε, B) > 1, where in the last inequality we applied (9.8).

To bound the second term on the right  R0 ⊆ R0 of Y2 (t)
 side of (9.12), fix some ′neighborhood 
with smooth boundary such that [0, 1] × Y2 (t) − 1, Y2 (t) + 1 ⊆ R0 ⊆ [0, 1] × Y2 (t) − 2, Y2 (t) + 2 .
On the event A, we have (since M = L1/2 ) that Y2 (t) ∈ [0, 1] × [x0 − 2L1/2 , x0 + 2L1/2 ] ⊆ R0 , and
so we may take R′0 to be of some fixed shape (for example, a rectangle of uniformly bounded height
and width, with smoothened corners) up to translation, independently of ε, A, B, and L. Then,
(9.14)
 
1A · ∇F1 Y2 (t) − ∇F2 Y2 (t) ≤ sup ∇F1 (w) − ∇F2 (w)

w∈R′0

≤ C3 sup F1 (w) − F2 (w) ≤ C3 sup F1 (w) − F2 (w) ≤ C3 A,
w∈∂R′0 w∈∂R0

for some some constant C3 = C3 (ε, B) > 1, where in the second inequality we applied (the m = 1
case of) Lemma 9.4 (applied with both φi = F1 |R′0 , which satisfy φ1 − φ2 = 0 and ∥φi ∥C 5 (R′0 ) −
∥φi ∥C 0 (R′0 ) ≤ ∥F1 ∥C 5 (R0 ) − ∥F1 ∥C 0 (R0 ) ≤ B), in the third we applied Lemma 9.1, and in the fourth
we applied the third assumption of the lemma. Inserting this, (9.12), and (9.13) into (9.11) gives
 
2p
E sup Y1 (s) − Y2 (s)

s∈[0,t]
Z t 2p !
2p 2p −c1 M 2p −1 2p

≤ (72p) (C2 t) e + (2C1 C3 At) + (2C1 BL ) Y1 (s) − Y2 (s) ds .
0

By Markov’s inequality, we deduce the existence of a constant C4 = C4 (ε, B) > 1 such that
" Z t #
P sup Y1 (s) − Y2 (s) ≥ C4 pt(A + e−c1 M/2p ) + C4 pL−1 Y1 (s) − Y2 (s) ds ≤ 4−p .

s∈[0,t] 0
LOCAL STATISTICS AND CONCENTRATION FOR NON-INTERSECTING BROWNIAN BRIDGES 73

1/8
c1
Taking p = L1/8 gives (after replacing c1 by with probability at least 1 − 4−L that
2 )
Z t
3/8 −c1 L1/8 −7/8

(9.15) sup Y1 (s) − Y2 (s) ≤ C4 L (A + e
) + C4 L Y1 (s) − Y2 (s) ds,
s∈[0,t] 0

for any t ∈ [0, L1/4 ]. Together with Gronwall’s inequality,10 this yields a constant C5 = C5 (ε, B) > 1
such that
 
1/8 1/8
sup Y1 (s) − Y2 (s) ≥ C5 L3/8 (A + e−c1 L ) ≤ 2−L .

(9.16) P
s∈[0,L1/2 ]

Since by Lemma 9.8 the process Fi Yi (s) is a martingale, we have
h  i
(9.17) F1 (z) − F2 (z) = E F1 Y1 (τ ) − F2 Y2 (τ ) .

Setting
z0 = z − (t0 , 0) = (0, x0 ) ∈ ∂R0 , Taylor expanding F1 and F2 , and using the fact that
∂j ∂k Fi (w) ≤ BL−1 , we find that
h
 i h  i h i
E F1 Y1 (τ ) − F2 Y2 (τ ) − E F1 Y1 (τ ) − F2 Y1 (τ ) − E ∇F2 (z0 ) · Y1 (τ ) − Y2 (τ )

 
   B h 2 i
− E ∇F2 Y2 (τ ) − ∇F2 (z0 ) · Y1 (τ ) − Y2 (τ ) ≤ · E Y1 (τ ) − Y2 (τ ) .
2L
   
Together with (9.17), the fact that E Y 1 (τ ) = z = E Y2 (τ ) (as Y1 and Y2 are martingales by
(9.3)), and again the bound ∂j ∂k Fi (w) ≤ BL−1 , this yields

h  i
F1 (z) − F2 (z) = E F1 Y1 (τ ) − F2 Y2 (τ )

h
 i B h i
≤ E F1 Y1 (τ ) − F2 Y1 (τ ) + · E Y2 (τ ) − z0 · Y2 (τ ) − Y1 (τ )
L
B h 2 i
+ · E Y1 (τ ) − Y2 (τ ) ,
2L
and hence

h  i
F1 (z) − F2 (z) ≤ E F1 Y1 (τ ) − F2 Y1 (τ )

(9.18)  
2B 
+ · E z0 − Y1 (τ ) + Y2 (τ ) − z0 · Y1 (τ ) − Y2 (τ ) .
L
Since F1 Y1 (τ ) = F2 Y 1 (τ ) on the event A (by (9.9)), and since F1 Y1 (τ ) − F2 Y2 (τ ) ≤
   

10BL (as ∇F1 (w) ≤ B, ∇F2 (w) ≤ B, and the diameter of R0 is at most 5L), it follows from
(9.8) that
h  
i 
= E 1 ∁ · F1 Y1 (τ ) − F2 Y1 (τ ) ≤ 10c−1 BLe−c1 L1/8
 
(9.19)
E F 1 Y 1 (τ ) − F 2 Y1 (τ ) A 1

10Here, we implicitly take a union bound so that (9.15) holds for all t in a 2−L1/8 -mesh of [0, L1/8 ], and apply
the discrete Gronwall inequality [39, Equation (8)].
74 AMOL AGGARWAL AND JIAOYANG HUANG


Moreover, from (9.16), (9.8), and the fact that Yi (τ ) − z0 ≤ M + 1 ≤ L1/2 for each i ∈ {1, 2} on
A (and again the fact that the diameter of R0 is at most 5L), we deduce the existence of constants
c2 = c2 (ε, B) > 0 and C6 = C6 (ε, B) > 1 such that
(9.20)
 

E Y1 (τ ) − z0 + Y2 (τ ) − z0 · Y1 (τ ) − Y2 (τ )

1/8 1/8 1/8 1/8
≤ 2L1/2 · C5 L3/8 (A + e−c1 L ) + c−1
1 e
−c1 L
+ 2−L
) · 2(5L)2 ≤ C6 L7/8 A + C6 e−c2 L ,
where we have used the fact that, on A, we have |Y1 (τ ) −z0 + Y2 (τ ) − z0 ≤ 2M = 2L1/2 . Inserting

(9.19) and (9.20) into (9.18) yields


F1 (z) − F2 (z) ≤ 2C6 BL−1/8 A + 2C6 BL−1 e−c2 L1/8 + 10c−1 BLe−c1 L1/8 ,

1

which implies the lemma. □

Appendix A. Concentration Bounds for the Height Function


In this section we establish a concentration bound for non-intersecting Brownian bridges; it (and
its proof) is analogous to the one that appears in the context of random tilings (see [16, Theorem 
21], for example). For any family of continuous curves x = (x1 , x2 , . . . , xn ) ∈ J1, nK × C [a, b] ,
x
we recall from height function H = H : [a, b] × R → R is defined by
 (3.6) that the associated
H(t, w) = # j ∈ J1, nK : xj (t) > w . In what follows, we also recall the measure Qfu;v ;g from
Definition 1.1.
Lemma A.1. Let n ≥ 1 be an integer; T > 0 and r, B ≥ 1 be real numbers; u, v ∈ Wn be
n-tuples; and f, g : [0, T ] → R be measurable functions with f ≤ g. Sample non-intersecting
Brownian bridges x ∈ J1, nK × C [0, T ] from the measure Qfu;v

. Fix real numbers t ∈ [0, T ] and
;g
w ∈ f (t), g(t) . Denoting the event E = H(t, w) ≤ B , there exists a deterministic number
  

Y = Y(u; v; f ; g; T ; t; w; B) ≥ 0 such that


h i 2
P H(t, w) − Y ≥ rB 1/2 ≤ 2e−r /4 + 2 · P E∁ .
 
(A.1)
2
In particular, setting B = n, we have P H(t, w) − Y ≥ rn1/2 ≤ 2e−r /4 .
 

To establish Lemma A.1, we proceed through a discretization. For any integer S ≥ 0, a discrete
path is a function p : J0, SK → Z such that p(t) − p(t − 1) ∈ {0, 1}, for each t ∈ J1, SK. A
discrete path ensemble on J0, SK is an ordered sequence of discrete paths p = (p1 , p2 , . . . , pk ), with
pj : J0, SK → Z for each j ∈ J1, kK. Given integer sequences u = (u1 , u2 , . . . , uk ) ∈ Zk and
v = (v1 , v2 , . . . , vk ) ∈ Zk , we say that p starts at u and ends at v if pj (0) = uj and pj (S) = vj for
each j ∈ J1, kK, respectively; observe that this is only possible if uj ≤ vj ≤ uj + S for each j. Given
two functions f, g : J0, SK → Z, we further say that the ensemble p remains above f and remains
below g if f (s) < pj (s) and pj (s) < g(s) for each (j, s) ∈ J1, nK × J0, SK, respectively. See the left
side of Figure 11 for a depiction. 
Denoting p(t) = p1 (t), p2 (t), . . . , pk (t) for any t, we additionally say that the ensemble p
is non-intersecting if p(t) ∈ Wk , for each t ∈ J0, SK. Given a non-intersecting path ensemble
p = (p1 , p2 , . . . , pk ) on J0, SK, we define its height function Hp : J0, SK × Z → Z≥0 by, for any
(t, w) ∈ J0, SK × Z, setting
Hp (t, w) = # j ∈ J1, kK : pj (t) > w .

LOCAL STATISTICS AND CONCENTRATION FOR NON-INTERSECTING BROWNIAN BRIDGES 75

Figure 11. Shown to the left is a discrete path ensemble p = (p1 , p2 , p3 ). Shown
to the right is a level set decomposition D = S0 ∪ S1 ∪ · · · ∪ S3 ; there, (3, 2, 0) is
an exit sequence of S3 .

Given the above notation, we can state the discrete variant of Lemma A.1 that quickly implies
it. Observe that, in the below concentration bound, the error estimates and probabilities do not
depend on the size S of the domain; this is in contrast to analogous statements in the context of
tilings (for example, in [16, Theorem 21], the error in the height function grows at least as S1/2 ),
though the proof is similar.
Lemma A.2. Let n, S ≥ 1 be integers; r, B ≥ 1 be real numbers; u, v ∈ Wn ∩ Zn be two n-
tuples of integers; and f, g : J0, SK → Z be functions with f ≤ g. Let p = (p1 , p2 , . . . , pn ) denote
a uniformly random discrete path ensemble conditioned to start at u; end at v; remain above
f and below g; and to be non-intersecting (assuming at least one
such ensemble exists). Fixing
(t, w) ∈ J0, SK × Z, and denoting the event E(p) = Hp (t, w) ≤ B , there exists a deterministic real


number Y = Y(u; v; f ; g; t; w; B) ∈ J0, nK such that


h i 2
P Hp (t, w) − Y ≥ rB 1/2 ≤ 2e−r /4 + 2 · P E(p)∁ .
 

Proof. This proof follows the discussion in [3, Section 3.2]. Let p′ = (p′1 , p′2 , . . . , p′n ) denote a
non-intersecting discrete path ensemble with the same law as, but independent from, p. Denote
D = D(S) = J0, SK × Z ⊂ Z2 ; for any subset S ⊆ D, the boundary ∂S of S denotes the set of
vertices (s, y) ∈ S that are adjacent11 to a vertex in Z2 \ S. Define the function F : D → Z by

setting F (s, y) = Hp (s, y) − Hp (s, y), for each (s, y) ∈ D.
Sk
A family S = (S0 , S2 , . . . , Sk ) of mutually disjoint subsets of D satisfying j=0 Sj = D is a level
set decomposition of D with respect to F if the following two properties hold. First, for any two
vertices (s, y), (s′ , y ′ ) ∈ Sj of the same subset, F (s, y) = F (s′ , y ′ ). Second, for any two adjacent
′ ′
vertices (s, y) ∈ Sj and (s , y ) ∈ Sj ′ with j ̸= j ′ (namely, in different subsets), F (s, y) ̸= F (s′ , y ′ );
observe in this case that F (s, y) − F (s , y ) = 1, as we either have that both Hp (s, y) − Hp (s′ , y ′ ) ∈
′ ′
′ ′
{0, 1} and Hp (s, y) − Hp (s′ , y ′ ) ∈ {0, 1} hold or that both Hp (s, y) − Hp (s′ , y ′ ) ∈ {0, −1} and
11Here, two vertices (s, y) and (s′ , y ′ ) are adjacent if (s−s′ , y−y ′ ) ∈ (0, 1), (1, 0), (1, 1), (0, −1), (−1, 0), (−1, −1)
(in this way, we view vertices as on a triangular lattice, instead of on a square one).
76 AMOL AGGARWAL AND JIAOYANG HUANG

′ ′
Hp (s, y) − Hp (s′ , y ′ ) ∈ {0, −1} hold. We set F (Sj ) = F (s, y), for any (s, y) ∈ Sj . In the below, we
condition on the level set decomposition S of D with respect to F .
Further observe that F (s, y) = 0 for any (s, y) ∈ J0, SK × Z that either satisfies (s, y) ∈ ∂D (since
p and p′ both start and end at the same n-tuples u and v, respectively), or that satisfies y < f (s)
′ ′
(as then Hp (s, y) = n = Hp (s, y)) or y ≥ g(s) (as then Hp (s, y) = 0 = Hp (s, y)). Since this family
of vertices is connected, there is one infinite set in S, which we let be S0 . It has the property that
D \ S0 is finite, so the number of elements k + 1 in S is finite.
Next, given i, j ∈ J1, nK with i ̸= j, we say that Si is adjacent to Sj if there exists a vertex in
Si that is adjacent to a vertex in Sj . In this case, we say that Si is exterior adjacent to Sj if Si
is contained in the infinite connected component of D \ Sj ; we otherwise call Si interior adjacent
to Sj . In particular, S0 is exterior adjacent to any Sj adjacent to it. It is quickly verified (see [3,
Lemma 12]) that, for every j ∈ J1, kK, there is at most one element of S that is exterior adjacent
to Sj . See the right side of Figure 11 for a depiction.
An exit sequence of any Sj ∈ S is an ordered sequence (i0 , i1 , . . . , im ) of integers in J0, kK so that
i0 = j, im = 0, and Sia is exterior adjacent to Sia−1 , for each a ∈ J1, mK; here, m = m(j) = m(j; S)
depends on j P and S. Then, it is quickly verified (see the proof of [3, Lemma 14]) that F (Sj ) has the
m
same law as a=1 ξa , where the ξa are mutually independent {−1, 1}-Bernoulli random variables
with P[ξa = 1] = 21 = P[ξa = −1]. Hence, by Chernoff’s inequality,
 
 2
P F (Sj ) − E F (Sj ) ≥ rm S ≤ 2e−r /2 .
1/2

(A.2)

Now fix j ∈ J0, mK to be such that (t, w) ∈ Sj . Observe on the event E = E(p) ∩ E(p′ ) =
  ′
Hp (t, w) ≤ B ∩ Hp (t, w) ≤ B that m(j; S) ≤ 2B. Indeed, define a sequence (w0 , w1 , . . . , wm )
by first setting w0 = w and then defining wi+1 for i ∈ J0, m − 1K inductively as follows. Let
ji ∈ J0, kK be such that (t, wi ) ∈ Sji (so j0 = j), and let wi+1 > wi be the maximum integer such
that (t, wi+1 ) ∈/ Sji but (t, wi+1 − 1) ∈ Sji (assuming ji ̸= 0; if instead ji = 0 then the sequence
stops). Then, (j0 , j1 , . . . , jm ) is an exit sequence for Sj . Moreover, for each i ∈ J0, m − 1K, we either
′ ′
have Hp (t, wi+1 ) = Hp (t, wi ) − 1 or Hp (t, wi+1 ) = Hp (t, wi ) − 1, since F (t, wi−1 ) ̸= F (t, wi ) and

since Hp (t, y) and Hp (t, y) are nonincreasing in y. On the event E, at most B such decrements can

occur for Hp and for Hp , meaning that m ≤ 2B.
Combining this estimate 1E · m(j; S) ≤ 2B with (A.2) and a union bound, we obtain
   
 
1/2
≤ P 1E · F (Sj ) − E F (Sj ) ≥ r(2B) S + P E∁
1/2
   
P F (Sj ) − E F (Sj ) ≥ r(2B)

2
≤ 2e−r /2 + 2 · P E(p)∁ ,
 

which gives the lemma. □


Proof of Lemma A.1. The second statement follows from the first, since H(t, w) ≤ n holds determin-
istically. The first statement follows from Lemma A.2 and the fact that non-intersecting random dis-
crete paths converge to non-intersecting Brownian motions (namely, the  T 1/2 S−1/2 · pj (⌊sT −1 S⌋)
from Lemma A.2 over (j, s) ∈ J1, nK×[0, T ] converges in law to the xj (s) over (j, s) ∈ J1, nK×[0, T ]
from Lemma A.1). □

Appendix B. Proofs of Results From Section 2


B.1. Proof of Lemma 2.7. In this section we establish Lemma 2.7. We begin with the follow-
ing lemma that bounds non-intersecting Brownian bridges constrained to lie below a linear lower
LOCAL STATISTICS AND CONCENTRATION FOR NON-INTERSECTING BROWNIAN BRIDGES 77

Figure 12. The left panel depicts Lemma B.1. The right panel depicts its proof
by coupling x with the non-intersecting Brownian bridges z.

boundary; see the left side of Figure 12 for a depiction. In the below, we let 0n = (0, 0, . . . , 0) ∈ Wn
(where 0 appears with multiplicity n).
Lemma B.1. There exist constants c > 0 and C > 1 such that the following holds. Fix an integer
n ≥ 1; real numbers A ≥ 0, T > 0, and v ≥ AT ; and define g : [0, T ] → R by setting g(x) = −Ax,
for each x ∈ [0, T ]. Let v = (−v, −v, . . . , −v) ∈ Wn (where −v appears with multiplicity  n),
and sample non-intersecting Brownian bridges x = (x1 , x2 , . . . , xn ) ∈ J1, nK × C [0, T ] under the
measure Q0−∞;g
n ;v
. Then, for any real numbers t ∈ [0, T ] and B ≥ 1, we have
h i 2
P xn (t) ≤ −vtT −1 − Bt1/2 log |9t−1 T | ≤ CeCn−cB n .

Proof. In view of the invariance of non-intersecting Brownian bridges under affine transformations
and diffusive scaling (recall Remark 2.1 and Remark 2.2), we may assume that T = 1. Then,
set θ = B8 · t1/2 log |9t−1 | and v ′ = (t−1 + 1)θ + v; denote u = (−θ, −θ, . . . , −θ) ∈ Wn and
v ′ = (−v ′ , −v ′ , . . . , −v ′ ) ∈ Wn . Sample non-intersecting Brownian bridges y = (y1 , y2 , . . . , yn ) ∈

u;v ′
J1, nK × C [0, T ] and z = (z1 , z2 , . . . , zn ) ∈ J1, nK × C [0, T ] from the measures Qu;v
 
−∞;g and Q ,
respectively; see the right side of Figure 12. By Lemma 2.4, we may couple x and y so that
yj (s) − xj (s) ≤ (t−1 s + 1)θ,

(B.1) for each (j, s) ∈ J1, nK × [0, 1].
Now, let us compare y and z. To that end, first observe from Lemma 2.6 that there exist
constants c ∈ (0, 1) and C1 > 1 such that
" n #
\ \ n B 1/2 −1 −1
o 2
P zj (s) ≥ − · s log |9s | − (t s + 1)θ − vs ≥ 1 − C1 eC1 n−cB n ;
j=1 s∈[0,1]
8
(B.2) " n #
\ \ n B 1/2 −1 −1
o 2
P zj (s) ≤ · s log |9s | − (t s + 1)θ − vs ≥ 1 − C1 eC1 n−cB n .
j=1
8
s∈[0,1]
78 AMOL AGGARWAL AND JIAOYANG HUANG

Observe that B8 · s1/2 log |9s−1 | − (t−1 s + 1)θ − vs ≤ −As, for each s ∈ [0, 1]. Indeed, if s ≤ t, this
follows from the facts that v ≥ A and t1/2 log |9t−1 | ≥ s1/2 log |9s−1 | when 0 < s ≤ t ≤ 1; for s > t,
this follows from the facts that v ≥ A and t−1/2 log |9t−1 | ≥ s−1/2 log |9s−1 | when 0 < t ≤ s ≤ 1.
Hence, the second statement of (B.2) indicates that zj (s) remains below g with probability at least
2
1 − C1 eC1 n−cB . It follows that we may couple z and y to coincide on an event of this probability
2
1 − C1 eC1 n−cB . Together with (B.1), the first part of (B.2), and a union bound, this gives
h B i 2
P xn (t) ≥ −4θ − vt − · t1/2 log |9t−1 | ≥ 1 − 2C1 eC1 n−cB n ,
8
which yields the lemma, since θ = B
8 · t1/2 log |9t−1 |. □

We then deduce the following variant of Lemma 2.7 that proves high probability Hölder bounds
for a single choice of (j, s, t), which quickly implies Lemma 2.7. We detail the proof of the first part
and outline that of the second, which is similar.

Lemma B.2. Adopt the notation of Lemma 2.7.


(1) Under the assumption of part 1 of Lemma 2.7, we have
h i 2 2
P xj (t + s) − xj (t) − sT −1 (vj − uj ) ≥ s1/2 B log |2s−1 T | + κ−1 (A + B) ≤ CeCn −cB n .

(2) Under the assumption of part 2 of Lemma 2.7, we have


h i 2 2
P xj (t + s) − xj (t) − sT −1 (vj − uj ) ≥ s1/2 B log |9s−1 T | + 2 A + B) ≤ CeCn −cB n .

Proof of Lemma B.2(1). By Remark 2.2 and Remark 2.1, we may assume that (a, b) = (0, 1) and
that (uj , vj ) = (0, 0). Since (sT −1 )1/2 ≤ 1 (as s ≤ T ), it suffices to show that
2 2
P xj (t + s) − xj (t) ≥ Bs1/2 log |2s−1 | + κ−1 s(A + B) ≤ CeCn −cB n ;
 
(B.3) 2 2
P xj (t + s) − xj (t) ≤ −Bs1/2 log |2s−1 | − κ−1 s(A + B) ≤ CeCn −cB n .
 

We only establish the first bound in (B.3), as then


 that of the second would follow by symmetry
(namely, by reflecting the bridges into the line t = 12 ). To that end, we first claim that there
exist constants c1 ∈ (0, 1) and C1 > 1 such that
" #
2
(B.4) P sup xj (r) ≤ A + B ≥ 1 − C1 eC1 n−c1 B n .
r∈[0,1]

To see this, define u′ , v ′ ∈ Wn by setting


B B
u′k = uk + + A; vk′ = vk + + A,
2 2
for each k ∈ J1, nK. Then, sample non-intersecting Brownian bridges y = (y1 , y2 , . . . , yn ) ∈ J1, nK ×
′ ′ ′ ′
C [0, 1] and y ′ = (y1′ , y2′ , . . . , yn′ ) ∈ J1, nK × C [0, 1] according to the measures Qfu ;v and Qu ;v ,
 

respectively. Then, by Lemma 2.3, we may couple x and y so that

(B.5) xk (r) ≤ yk (r), for each (k, r) ∈ J1, nK × [0, 1].


LOCAL STATISTICS AND CONCENTRATION FOR NON-INTERSECTING BROWNIAN BRIDGES 79

Figure 13. The left panel depicts the proof of Lemma B.2(1). The right panel
depicts that of Lemma B.2(2).

Moreover, by Lemma 2.6, there exist constants c1 ∈ (0, 1) and C2 > 1 such that
" #
B 2
P inf yn′ (r) ≥ sup f (r) + ≥ 1 − C2 eC2 n−c1 B n ;
r∈[0,1] r∈[0,1] 4
(B.6) " #
2
P sup yj′ (r) ≤ A + B ≥ 1 − C2 eC2 n−c1 B n ,
r∈[0,1]

where in the first inequality we used that min{u′n , vn′ } = min{un , vn } + A + B2 ≥ supr∈[0,1] f (r) + B2 ,
and in the second inequality we used the fact that u′j = A + B2 = vj′ (recalling that uj = vj = 0).
2
The first bound in (B.6) indicates that y ′ lies above f with probability at least 1−C2 eC2 n−c1 B n .
2
Hence, we may couple y and y ′ so that they coincide with probability at least 1 − C2 eC2 n−c1 B n .
This, together with the second bound in (B.6), (B.5), and a union bound then yields (B.4).
Now, to show the first bound in (B.3), we condition on xk (r) for (k, r) ∈ Jj + 1, nK × [0, 1] and for
(k, r) ∈ J1, nK × [t + s, 1], and set u′′ , v ′′ ∈ Wn−j+1 by letting u′′k = xj (0) = 0 and vk′′ = xj (t + s), for
each k ∈ J1, jK. Sample non-intersecting Brownian bridges z = (z1 , z2 , . . . , zj ) ∈ J1, jK × C [0, t + s]
′′ ′′
from the measure Qu ;v (n−1 ). Then, Lemma 2.3 yields a coupling between x and z such that
xk (r) ≥ zk (r), for each (k, r) ∈ J1, jK × [0, 1]; see the left side of Figure 13. Thus,
 
1/2 −1 s(A + B)
P xj (t) ≤ xj (t + s) − Bs log |2s | −
κ
 
s(A + B) 2
≤ P zj (t) ≤ zj (t + s) − Bs1/2 log |2s−1 | − ≤ 2C1 eC1 n−c1 B n ,
κ
80 AMOL AGGARWAL AND JIAOYANG HUANG

where in the last equality we used Lemma 2.6 and the facts that s + t ≥ t ≥ κ and that
vj′′ − u′′j A+B
= (s + t)−1 · zj (t + s) ≤ (s + t)−1 · max xj (r) ≤ ,
s+t r∈[0,1] κ

on the event maxr∈[0,1] xj (r) ≤ A + B . This, with (B.4), establishes the first bound in (B.3) and
thus the first statement of the lemma. □

Proof of Lemma B.2(2). Since the proof of the second part of the lemma is similar to that of the
first part, we only outline it. We may again assume that (a, b) = (0, 1) and (uj , vj ) = (0, 0). We
may further assume by symmetry that t ≤ 21 . It suffices (again since (sT −1 )1/2 ≤ 1) to show that
2 2
P xj (t + s) − xj (t) ≥ Bs1/2 log |9s−1 | + 2s(A + B) ≤ CeCn −cB n ;
 
(B.7) 2 2
P xj (t + s) − xj (t) ≤ −Bs1/2 log |9s−1 | − 2s(A + B) ≤ CeCn −cB n ,
 

and by symmetry it suffices to show the latter. Following the derivation of (B.4), we again find
that there exist constants c1 ∈ (0, 1) and C1 > 1 such that
" #
2
(B.8) P sup xj (r) ≤ A + B ≥ 1 − C1 eC1 n−c1 B n .
r∈[0,1]

Now condition on xk (r) for (k, r) ∈ Jj +1, nK×[0, 1] and for (k, r) ∈ J1, nK×[0, t]; set ge : [t, 1] → R
by letting ge(r) = xj (t) + (t − r)A, for each r ∈ [t, 1]; and set u′ , v ′ ∈ Wj by letting u′k = xj (t)
and vk′ = min{xj (1), ge(1)} = min{0, ge(1)} (where we used that xj (1) = vj = 0) for eachk ∈ J1, jK.
Then sample non-intersecting Brownian bridges y = (y1 , y2 , . . . , yj ) ∈ J1, jK × C [t, 1] from the
′ ′
measure Qu ;v (n−1 ); see the right side of Figure 13. Since ∂t g(r) ≤ A for each r ∈ [t, 1], and

−∞;g̃
since g(t) ≥ xj (t) = u′j , we have g ≥ ge. Thus, Lemma 2.3 yields a coupling between x and y such
that

(B.9) xk (r) ≥ yk (r), for each (k, r) ∈ J1, jK × [t, 1].

Applying Lemma B.1 to y yields constants c1 ∈ (0, 1) and C2 > 1 such that
(u′j − vj′ )s
 
2
(B.10) P yj (t + s) ≥ yj (t) − − Bs1/2 log |9s−1 | ≥ 1 − C2 eC2 n−c1 B n .
1−t
If vj′ = ge(1) = u′j − A(1 − t), then u′j − vj′ = A(1 − t) ≤ A. If instead vj′ = 0, and (B.8) implies
2
P[u′j ≤ A + B] ≥ 1 − C1 eC1 n−c1 B n ; since t ≤ 12 , we have in either case that
 ′
uj − vj′

2
P ≤ 2(A + B) ≥ 1 − C1 eC1 n−cB n .
1−t
Together with (B.10), (B.9), and a union bound, this gives the second bound in (B.7), which implies
the second statement of the lemma. □

Proof of Lemma 2.7 (Outline). The proof follows quickly from Lemma B.2 and (the proof of) [19,
Lemma 3.3], very similarly to the proof of [19, Proposition 3.5] given [19, Lemmas 3.2 and 3.3], so
further details are omitted. □
LOCAL STATISTICS AND CONCENTRATION FOR NON-INTERSECTING BROWNIAN BRIDGES 81

B.2. Proofs of Lemma 2.10, and Lemma 2.11. In this section we establish first Lemma 2.10,
and next Lemma 2.11.

Proof of Lemma 2.10. By [8, Lemma 3], we have mµt (z) ≤ t−1/2 for any z ∈ Λt;µ . By (2.6), this

yields ϱt (x) ≤ π −1 t−1/2 , for any x ∈ R, confirming the first statement of the lemma.
To verify the latter, fix x ∈ R and set z ∈ ∂Λt such that M (z) = z − tm0 (z) = x ∈ R (where
such a z exists by Lemma 2.8). Then, (2.6) and (2.11) give
ϱt (x) = π −1 Im mt z − tm0 (z) = π −1 Im m0 (z).

(B.11)
−1
Let z0 ∈ ∂Λt be such that M (z0 ) = x0 . Repeatedly applying the relation ∂x = ∂∂xz ∂z =
′ −1
(1 − tm0 ) ∂z , it follows that there exists a constant C = C(k) > 1 such that
′ ′′
∂x ϱt (x0 ) ≤ C · max ∂zk′ m0 (z0 ) 2k/k · max 1 − tm′0 (z0 ) −k .
k
(B.12) ′ ′′
k ∈J1,kK k ∈J0,2kK

To bound the first term on the right side, observe since |z0 − y| ≥ Im z0 for any y ∈ R that
(B.13)
Z ∞ Z ∞
k
∂z m0 (z0 ) ≤ k! ϱ0 (y)dy k! ϱ0 (y)dy Im m0 (z0 ) k!
k+1
≤ = k! · ≤ ,
−∞ |z0 − y| (Im z0 )k−1 −∞ |z0 − y| 2 (Im z0 ) k t(tδπ)k−1
where in the last two statements we used the facts that
Im m0 (z0 ) = πϱt (x0 ) ≥ πδ; Im z0 = t Im m0 (z0 );
Z ∞ Z ∞
(B.14) ϱ0 (y)dy 1 ϱ0 (y)dy Im m0 (z0 )
2
= · Im = = t−1 ,
−∞ |z0 − y| Im z0 −∞ z0 − y Im z0

Here, the first statement follows from (B.11) and the assumption ϱt (x0 ) ≥ δ; and the second follows
from the fact that M (z0 ) = z0 − tm0 (z0 ) = x0 ∈ R; and the third follows from the definition of m0 .
To estimate the last term on the right side of (B.12), observe that
1 − tm′0 (z0 ) ≥ 1 − t Re m′0 (z0 ),

(B.15)
and that
Z ∞ Z ∞
ϱ0 (y)dy  ϱ0 (y)dy
Re m′0 (z0 ) = Re = |y − z0 |2 − 2(Im z0 )2 ·
−∞ (y − z0 )2 −∞ |y − z0 |4
Z ∞
1 ϱ0 (y)dy
= − 2(Im z0 )2 4
,
t −∞ |y − z0 |

where in the last equality we used the third statement of (B.14). Inserting this into (B.15), we find
(B.16)
∞ ∞ 2
2(Im z0 )2
Z Z
1 − tm′0 (z0 ) ≥ 2t(Im z0 )2 ϱ0 (y)dy ϱ0 (y)dy
≥ 2t(Im z0 )2 ≥ 2t(πδ)2 ,


−∞ |y − z0 |4 −∞ |y − z0 |2 t
where in the second statement we used the fact that ϱ0 is a probability density; in the third we
used the third statement of (B.14); and in the fourth we used the first two statements of (B.14).
The lemma then follows from inserting (B.13) and (B.16) into (B.12). □

Proof of Lemma 2.11. Denote the Stieltjes transform of µ0 by m0 = mµ0 . We will use (B.12) to
estimate the x-derivatives of ϱt , to which end we must upper bound derivatives of m0 and lower
82 AMOL AGGARWAL AND JIAOYANG HUANG

bound |1 − tm′0 |. To implement this, we claim that there exists a constant C1 = C1 (k) > 1 such
that, for any integer j ∈ J0, k − 1K,
∂z m0 (z) ≤ C1 B j+2 (δc)−j−1 ,
j c
for any z ∈ H such that | Re z| ≤ ;
2
(B.17)
1 − tm′0 (z) ≥ tδc , c

for any z ∈ H such that | Re z| ≤ and Im z ∈ [0, c].
2 Im z 2
To see this, fix j ∈ J0, k − 1K and observe that (2.5) implies for | Re z| < 2c that
(B.18)
Z Z
∞ ϱ (x)dx c ϱ (x)dx
0 0
+ 2j+1 j!c−j−1
j
∂z m0 (z) = j! ≤ j!

−∞ (x − z)j+1 −c (x − z)j+1
Z c j
∂x ϱ0 (x)dx −j−1
j+1
+ 2j+1 (j + 1)!B j+1 (δc)−j .

≤ + 2 j!c
−c x − z
Here, to obtain the second statement we used the facts that | Re z| ≤ 2c and that ϱ0 is the density
of a probability measure. To obtain the third, we integrated by parts j times, using the estimate
on the boundary terms,
j j−i j
∂x ϱ0 (c) ∂xj−i ϱ0 (−c) 2j X j+1 j+1

∂xj−i ϱ0 (c) + ∂xj−i ϱ0 (−c) ≤ 2 (j + 1)!B
X  
i!

i
− i

j
i! j
,
i=0
(z − c) (z + c) c i=0 (δc)

where the first bound holds again since | Re z| ≤ 2c and the second holds since ∂xi ϱ0 (x) ≤ B i+1 δ −i
for any i ∈ J0, kK. Then,
Z c j Z
∂x ϱ0 (x) − ∂xj ϱ0 (Re z) c dx

+ O B j+1 (δc)−j
j j 
∂z m0 (z) ≤ dx + ∂x ϱ0 (Re z) ·
−c
z − x −c x − z
Z c j j

∂x ϱ0 (x) − ∂x ϱ0 (Re z)
dx + O B j+1 (δc)−j


−c
x − z
≤ 2c · sup ∂xj+1 ϱ0 (x) + O B j+1 cδ)−j−1 ≤ O B j+2 (cδ)−j−1 ,
 
|x|≤c

where in the first and second estimates we used (B.18), the bound ∂xj ϱ0 (x) ≤ B j δ −j for |x| ≤ c
Rc
(as | Re z| ≤ c), and the fact that PV −c (x − z)−1 dx ≤ 2 (as | Re z| ≤ 2c ); in the fourth, we used
the same bound on derivatives of ϱ0 , but with j replaced by j + 1. Here, the implicit constant only
depends on j, verifying the first statement of (B.17).
To verify the second, observe for | Re z| ≤ 2c and Im z ∈ [0, c] that
Z ∞ Z c
1 − tm′0 (z) ≥ 2t(Im z)2
ϱ0 (x)dx 2 dx
4
≥ 2tδ(Im z)
−∞ |x − z| −c |x − z|4
Z c/2
dx
≥ 2tδ(Im z)2 4
−c/2 |x| + Im z
 −3 
4 2 −3
c tδc
= · tδ(Im z) (Im z) − + Im z ≥ ,
3 2 2 Im z
where in the first statement we used the first bound of in (B.16); in the second we used the fact
that ϱ0 (x) ≥ δ for |x| ≤ c; in the third we changed variables from x to x − Re z and used the
LOCAL STATISTICS AND CONCENTRATION FOR NON-INTERSECTING BROWNIAN BRIDGES 83

bound | Re z| ≤ 2c ; in the fourth we performed the integration; and in the fifth we used the fact that
Im z ≤ c. This verifies the second statement of (B.17).
To apply (B.17), let C2 = C2 (k) > 10 denote a sufficiently large constant (to be determined
later) and fix real numbers

c h (δc)20 i
(B.19) x0 , t ∈ R, such that |x0 | ≤ , and t ∈ 0, .
4 C2 B 20

By Lemma 2.8, there exists a complex number z0 = z0 (x0 , t) ∈ ∂Λt such that z0 − tm0 (z0 ) = x0 .
We claim that | Re z0 | ≤ 2c . To see this, first observe that z0 = z0 (x0 , t) is continuous in t.
Indeed, since the map M (z) = z − tm0 (z) is a homeomorphism from Λt (recall (2.10)) to H
by Lemma 2.8 that is uniformly continuous in t, its inverse M −1 is also continuous in t. Thus,
z0 (x0 , t) = M −1 (x0 ) is continuous
 c c  in t. Now, assume to the contrary that
c
Re z0 (x, t) > c2 . Then

since z0 (x0 , 0) = x0 ∈ − 4 , 4 , there exists some s ∈ [0, t] such that Re z0 (x0 , s) = 2 . Since

|x0 | ≤ 4c , it follows that tm0 (z0 ) = |z0 − x0 | ≥ 4c at z0 = z0 (x0 , s). This contradicts the fact (from
the j = 0 case of (B.17) and (B.19)) that

20
tm0 (z0 ) ≤ (δc) · C1 B 2 (δc)−1 ≤ c ,

C2 B 20 8

for C2 > C1 , and so | Re z0 | ≤ 2c . Further observe that Im z0 ≤ c


2, since otherwise the third
statement of (B.14) would yield (as C2 > 4)

∞ ∞
C2 B 20
Z Z
1 ϱ0 (x)dx
= < 4c−2 ϱ0 (x)dx = 4c−2 < ,
t −∞ |x − z0 |2 −∞ (δc)20

which contradicts (B.19).


Therefore,
Z ∞ Z c
1 ϱ0 (x)dx dx Bπ
= ≤B + 4c−2 ≤ + 4c−2 ,
t −∞ |x − z0 |2 −c |x − z0 |2 Im z

where in the first statement we used the third part of (B.14); in the second we used the facts that
sup|x|≤c ϱ(x) ≤ B, that ϱ is a probability density, and that | Re z0 | ≤ 2c ; and in the third we used the
R∞ 2
equality −∞ |x − z|−2 dx = π(Im z)−1 . Since t ≤ c8 , it follows that t(Im z)−1 ≥ (2Bπ)−1 . Inserting
this into the second bound of (B.17) (which applies as | Re z0 | ≤ 2c and Im z0 ≤ 2c ) gives

1 − tm′0 (z0 ) ≥ δc .

(B.20)
4πB
c
Inserting the first bound of (B.17) (at z = z0 , which applies since | Re z0 | ≤ 2) and (B.20) into
(B.12), it follows that there exists a constant C3 = C3 (k) > 1 such that

∂x ϱt (x0 ) ≤ C3 (B 8 δ −6 c−6 )j+1 ,


j
(B.21) for any j ∈ J0, k − 1K.
84 AMOL AGGARWAL AND JIAOYANG HUANG

Before proceeding to bound the t-derivative of ϱt , let us quickly establish the first statement of
(2.14). To that end, observe that
ϱt (x0 ) − ϱ0 (x0 ) = π −1 Im mt (x0 ) − Im m0 (x0 )


= π −1 Im mt z0 − tm0 (z0 ) − Im m0 (x0 )


= π −1 Im m0 (z0 ) − Im m0 (x0 )

δ
≤ C3 B 16 δ −12 c−12 · |z0 − x0 | ≤ C3 B 16 δ −12 c−12 · t m0 (z0 ) ≤ ,

4
where in the first statement, we applied (2.6); in the second, the definition of z0 ; in the third
statement, (2.11); in the fourth, the first bound in (B.17) at j = 1; in the fifth, the fact that
z0 − x0 = tm0 (z0 ); and in the sixth, the bound on t from (B.19) (and making C2 > 4C3 sufficiently
large) with first bound in (B.17) at j = 0. Since ϱ0 (x0 ) ≥ δ (as |x| ≤ 4c ≤ c), this implies the first
statement of (2.14).
To prove the second statement of (2.14), observe by (B.21) that  it suffices to bound the t-
derivatives of ϱt . Recall from (2.12) that ∂t ϱt (x) = π∂x ϱt (x)Hϱ t (x) , where Hϱt (x) is the Hilbert
transform of ϱt (recall (2.7)). For any real number x ∈ − 4c , 4c and integer j ∈ J0, k − 2K, we can


then bound the derivatives of Hilbert transform by


 Z ∞ 

∂x Hϱt (x) = π −1 ∂xj PV
j ϱ t (x − y)dy
y

−∞
 Z c/2 
ϱt (x − y)dy ϱt (x − y)dy
Z
−1 j j
= π ∂x PV + ∂x
y y

(B.22)
−c/2 |y|>c/2

c/2

ϱt (x − y) − ϱt (x) dy
Z  Z

−1 j j ϱt (y)dy
= π ∂x PV + ∂x
y |y−x|>c/2 x − y

−c/2

≤ c · sup ∂xj+1 ϱt (y) + 2j j!c−j = O (B 8 δ −6 c−6 )j+2 ,



|y|≤c

where in the first and second statements we used (2.7); in the third we used the fact that PV(a) = 0
for any constant a in the first integral,
and we changed
variables from y to x − y in the second;
in the fourth we used the fact that ∂xj (x − y)−1 ≤ j!2j+1 c−j−1 for |x − y| ≥ 2c and that ϱt is a
probability measure; and in the fifth we applied (B.21). Combining this with (B.21) and (2.12)
then yields the second bound in (2.14). □
B.3. Proof of Lemma 2.4. In this section we establish Lemma 2.4, to which end we discretize in
time (leaving space continuous). For any integer T ≥ 1, a (T -step)  Gaussian walk starting at u ∈ R
is a probability measure on (T + 1)-tuples x(0), x(1), . . . , x(T ) ∈ RT with x(0) = u such that, for
each j ∈ J1, T K, the jump x(j) − x(j − 1) is a centered Gaussian random variable of variance 1. A
(T -step) Gaussian bridge from u to v is a Gaussian walk starting at u, conditioned to end at v (so
x(T ) = v). The following definition, similar to Definition 1.1, provides notation for non-intersecting
Gaussian bridges.
Definition B.3. Fix integers T, n ≥ 1; two n-tuples u, v ∈ Wn ; and two  functions f, g : J0, T K → R.
Let Gu;v
f ;g denote the law on sequences x(t) = x 1 (t), x2 (t), . . . , x n (t) , with t ∈ J0, T K, given by n
independent T -step Gaussian walks, conditioned on satisfying the following three properties.
(1) The xj do not intersect, that is, x(t) ∈ Wn for each t ∈ J1, T − 1K.
LOCAL STATISTICS AND CONCENTRATION FOR NON-INTERSECTING BROWNIAN BRIDGES 85

(2) The xj start at uj and end at vj , that is, xj (0) = uj and xj (T ) = vj for each j ∈ J1, nK.
(3) The xj are bounded below by f and above by g, that is, f < xj < g for each j ∈ J1, nK.
We assume f (0) ≤ un ≤ u1 ≤ g(0) and f (T ) ≤ vn ≤ v1 ≤ g(T ), even when not stated explicitly.
Analogous to Remark 2.1, the following states that non-intersecting Gaussian walks are invariant
under affine transformations.
Remark B.4. Nonintersecting Gaussian walks satisfy the following invariance property under affine
transformations. Adopt the notation of Definition B.3, and fix real numbers α, β ∈ R. Define the
n-tuples u e ∈ Wn and functions fe, ge : J0, nK → R by setting
e, v
u
ej = uj + α, and vej = vj + T β + α, for each j ∈ J0, nK;
ge(t) = f (t) + tβ + α, and ge(t) = g(t) + tβ + α, for each t ∈ J0, T K.

en (t) under Gfũ;ṽ



Sampling x e= x e1 (t), x
e2 (t), . . . , x ˜,g̃ , there is a coupling between x e and x such that
ej (t) = xj (t) + βt + α for each t ∈ J0, T K and j ∈ J1, nK.
x
As in Remark 2.1, this follows from the analogous affine invariance of a single Gaussian random
bridge, together with the fact that affine transformations do not affect the non-intersecting property.
The following lemma states height monotonicity (the analog of Lemma 2.3) for non-intersecting
Gaussian bridges. Its proof is very similar to that of Lemma 2.3 (from [18]) and is thus omitted.

Lemma B.5. Fix integers T, n ≥ 1; four n-tuples u, u e ∈ Wn ; and functions f, fe, g, ge :


e , v, v
J0, T K → R with f ≤ g and f ≤ ge. Sample non-intersecting Gaussian bridges x(t) and x
e e(t) from
the measures Gu;v
f ;g and Gũ;ṽ
f˜;g̃
, respectively. If

f ≤ fe; g ≤ ge; u≤u


e; v≤v
e,
e such that x(t) ≤ x
then there exists a coupling between x and x e(t), for each t ∈ J0, T K.
The following lemma is a discrete (in time) variant of Lemma 2.4 and quickly implies it.
Lemma B.6. Adopt the notation and assumptions of Lemma B.5, and fix some real number B ≥ 0.
If uj ≤ u
ej ≤ uj + B and vj ≤ vej ≤ vj + B for all j ∈ J1, nK, then the following two statements hold.
(1) If f (t) ≤ fe(t) ≤ f (t) + B and g(t) ≤ ge(t) ≤ g(t) + B for each t ∈ J0, T K, then there is a
coupling between x and x e so that xj (t) ≤ x
ej (t) ≤ xj (t) + B for each (t, j) ∈ J0, T K × J1, nK.
(2) If u = u e , and f (t) ≤ fe(t) ≤ f (t) + tB
T and g(t) ≤ ge(t) ≤ g(t) + tBT for each t ∈ J0, T K,
then there is a coupling between x and x e so that xj (t) ≤ x ej (t) ≤ xj (t) + tBT for each
(t, j) ∈ J0, T K × J1, nK.
Proof of Lemma 2.4. This follows from Lemma B.6 with the convergence of Gaussian walks to
Brownian bridges; we omit further details. □
Similarly to the proof [18, Lemma 2.6 and Lemma 2.7], that of Lemma B.6 makes use of a local
Glauber dynamic, which in our setting is defined as follows.
Definition B.7. Fix integers  T, n ≥ 1 and two functions f, g : J0, T K → R. For t ∈ J0, T K, let
y(t) = y1 (t), y2 (t), . . . , yn (t) ∈ Wn be a family of n non-intersecting T -step walks. The Glauber
dynamics is the discrete-time Markov chain whose state P k y(t) = P k y1 (t), P k y2 (t), . . . , P k yn (t)
at time k ≥ 0 is determined as follows. Below, for each t ∈ J0, T K, let y0 (t) = g(t) and yn+1 (t) = f (t).
86 AMOL AGGARWAL AND JIAOYANG HUANG

Set P 0 y = y. For k ≥ 1, let (a, b, d) ∈ J0, T −2K×J1, nK×Z≥0 be such that k = dn(T −1)+an+b.
For (t, j) ̸= (a + 1, b), set P k yj (t) = P k−1 yj (t). For (t, j) = (a + 1, b), set P k yb (a + 1) to be the
middle (t = 1) value of a 2-step Gaussian random bridge, conditioned to start at P k−1 yb (a); end
at P k−1 yb (a + 2); be above P k−1 yb+1 (a + 1); and be below P k−1 yb−1 (a + 1).
Remark B.8. It follows from the Gibbs property (for non-intersecting Gaussian bridges) that
y(0);y(T )
Gf ;g is a stationary measure for the Glauber dynamics P.
The following lemma indicates that the Glauber dynamics converge to the stationary measure
y(0);y(T )
Gf ;g in the long-time limit; its proof is given in Appendix B.4 below.

Lemma B.9. Adopt the notation of Definition B.7, and assume that maxt∈J0,T K |f (t)| + |g(t)| <
y(0);y(T )
∞. Then, the law of P 2n(T −1)k y converges as k tends to ∞ to Gf ;g .
Now let us briefly outline the proof of Lemma B.6.
Proof of Lemma B.6 (Outline). We only discuss the second statement of the lemma,
as the proof
of the first is entirely analogous. We will assume throughout that f (t) + g(t) < ∞ for each
t ∈ J1, T − 1K, as the cases when f (s) = −∞ or g(s) = ∞ for some s ∈ J1, T − 1K can then be
obtained from a limiting procedure (in particular, one can consider a sequence fk of lower boundaries
such that fk (s) always remains finite but tends to −∞ as k tends to ∞, and similarly for g).
Let us first confirm the lemma when (T, n) = (2, 1), in which case we will verify the following
more general result. Fix a real number A ≥ 0 and assume that
x1 (0) ≤ xe1 (0) ≤ x1 (0) + A; x1 (0) ≤ x e1 (2) ≤ x1 (2) + A + B;
B B
f (1) ≤ fe(1) ≤ f (1) + A + ; g(1) ≤ ge(1) ≤ g(1) + A + .
2 2
Then, there exists a coupling between x1 and x e1 such that
B
(B.23) x1 (1) ≤ x e1 (1) ≤ x1 (1) + A + .
2
Indeed, this is a consequence of the fact (which follows quickly from Lemma B.5 and Remark B.4,
e1 (1) ≥ r ≤ P x1 (1) + A + B2 ≥ r , for
  
or can also be verified directly) that P x1 (1) ≥ r ≤ P x
any real number r ∈ R. This shows (B.23), whose A = 0 special case yields the second part of
Lemma B.6 at (T, n) = (2, 1).
Thus, we assume in  what follows that either T ≥ 3 or n ≥ 2. Fix two ensembles y(t) =
y1 (t), y2 (t), . . . , yn (t) ∈ Wn and ye(t) = ye1 (t), ye2 (t), . . . , yen (t) ∈ Wn of n non-intersecting T -step
walks, with starting points y(0) = u = ye(0), and ending points y(T ) = v and ye(T ) = v e, such that,
for each t ∈ J0, T K,
tB
y(t) ≤ ye(t) ≤ y(t) + ; f (t) ≤ yn (t) ≤ y1 (t) ≤ g(t); fe(t) ≤ ye1 (t) ≤ yen (t) ≤ ge(t).
T
Such initial data exist due to the assumptions of the second part of the lemma.
We next run the Glauber dynamics on y (with boundaries (f, g)) and ye (with  boundaries (f , ge)),
e
k−1 k−1
coupled in the following way. If we update the pair P yj (t), P yej (t) , then using (B.23)
(t−1)B 2B 
(with the (A; B) there given by T ; T here), we couple P yj (t) ≤ P k yej (t) ≤ P k yj (t) + tB
k
T .
By induction on k, this ensures that P k yj (t) ≤ P k yej (t) ≤ P k yj (t) + tB T holds for each (t, j) ∈
J0, T K × J1, nK and k ≥ 0. Letting k tend to ∞ and taking any limit point of this coupling, the
lemma then follows from Lemma B.9. □
LOCAL STATISTICS AND CONCENTRATION FOR NON-INTERSECTING BROWNIAN BRIDGES 87

B.4. Proof of Lemma B.9. In this section we establish Lemma B.9. We begin by recalling a
general statement on convergence to stationarity of infinite-dimensional Markov chains.
Let Ω be a measurable space with σ-algebra F; let P(Ω) denote the space of probability measures
on (Ω, F). Given any two probability measures ν1 , ν2 ∈ P(Ω), recall that the total variation distance
between them is defined by

(B.24) dTV (ν1 , ν2 ) = sup ν1 (A) − ν2 (A) .
A∈F

Assumption B.10. Adopting the above notation, let K : Ω × F → R≥0 be a Markov transition
kernel. For any function φ : Ω → R≥0 and measure µ on Ω, define the function Kφ : Ω → R≥0 and
measure Kµ on Ω by setting
Z Z
Kφ(x) = φ(y)K(x, dy); Kµ(A) = K(x, A)µ(dx),
Ω Ω
for any x ∈ Ω and measurable set A ∈ F. Assume that there exist constants α ∈ (0, 1), γ ∈ (0, 1),
2B
B ≥ 0, and R > 1−γ ; a potential function V : Ω → R≥0 ; and a probability measure ν on Ω, such
that the following two conditions hold.
(1) For each x ∈ Ω, we have KV (x) ≤ γV (x) + B, for each x ∈ Ω.
(2) For each x ∈ Ω with V (x) ≤ R, and any measurable set A ∈ F, we have K(x, A) ≥ αν(A).
The following result provides a convergence theorem for such Harris chains [38]. It appears in
[57], though it is stated as written below in [37].
Lemma B.11 ([37, Theorem 1.2]). Adopt Assumption B.10, and fix a measure µ on Ω. Then, the
Markov process defined by K has a unique stationary measure µ0 , and limm→∞ ∥Km µ − µ0 ∥TV = 0.
We next apply Lemma B.11 to the Glauber dynamics P of Definition B.7. Throughout the
remainder of this section, we adopt the notation of that definition. These include the functions 
f, g : J0, T K → R; the family y of n non-intersecting T -step walks y(t) = y1 (t), y2 (t), . . . , yn (t) ∈
Wn (over t ∈ J0, T K), and the associated Markov operator P (which we also interpret as a kernel)
for the Glauber dynamics. We also set u = y(0) and v = y(T ), which are fixed throughout the
dynamics. Then the state space for the Glauber dynamics P can be viewed as
n o
Ω0 = y(t) t∈J1,T −1K ∈ (Wn )T −1 : min
  
yn (t) − f (t) > 0, min g(t) − y1 (t) > 0 .
t∈J1,T −1K t∈J1,T −1K

The associated potential function V0 is defined by


 
V0 (y) = max max yj (t) + 1 .
j∈J1,nK t∈J1,T −1K

We then have the following two lemmas verifying Assumption B.10 for the Glauber dynamics.
Lemma B.12. There exist constants γ ∈ (0, 1) and B = B(u, v, f, g) ≥ 0 such that, for any family
y of n non-intersecting T -step walks with y(0) = u and y(T ) = v, we have PV0 (y) ≤ γV0 (y) + B.
Proof. This follows from taking γ = 21 and B = maxt∈J0,T K |f (t)| + |g(t)| + 1, which applies since


f (t) ≤ yj (t) ≤ g(t) holds for any (t, j) ∈ J0, T K × J1, nK. In particular, PV0 (y) ≤ B ≤ γV0 (y) + B
holds deterministically. □
Lemma B.13. There exists a constant α = α(u, v, f, g) > 0 such that the following holds. Letting
ν0 denote the Lebesgue measure on the set Ω0 , we have P 2n(T −1) (y, A) ≥ αν0 (A), for each y ∈ Ω0
and any measurable subset A ⊆ (Wn )T −1 .
88 AMOL AGGARWAL AND JIAOYANG HUANG

Proof (Outline). For any integer T ′ ≥ 2; two n-tuples u′ , v ′ ∈ Wn ; and two functions f ′ , g ′ :

;v ′
J0, T ′ K → R, the density of the measure Gu

f ′ ;g ′ on sequences x(t) = x1 (t), x2 (t), . . . , xn (t) is

T′  ′ −1
n
!
2 TY

Y Y 1
(B.25) C · 1x∈Ω0 · 1xj (0)=u′j 1xj (T ′ )=vj′ exp − xj (t) − xj (t − 1) dxj (t) ,
j=1 t=1
2 t=1

for some normalization constant C = C(f ′ , u′ , v ′ ) > 0. Observe that there exist some constant
c1 = c1 (f ′ , g ′ , u′ , v ′ ) > 0 such that C > c1 (that is, the normalization constant is bounded below),
since the property x∈ Ω 0 is an
open condition. Further observe that the density (B.25) is uniformly
bounded (since f (t) + g(t) is). Thus, there exists a constant c1 = c1 (f ′ ; u′ ; v ′ ) > 0 such that the
′ ′
measure Gu f′
;v
is absolutely continuous with respect to the Lebesgue measure on this set, and its
Radon–Nikodym derivative is bounded above by c−1 1 and below by c1 .
Adopting the notation of Definition B.7, we fix an integer k ≥ 1 and apply this in the case
(n′ , T ′ ) = (1, 2); (u′ , v ′ ) = (u′1 , v1′ ) = P k−1 yb (a), P k−1 yb (a + 2) ; and (f ′ , g ′ ) = f ′ (1), g ′ (1) =


P k−1 yb+1 (a + 1), P k−1 yb−1 (a + 1) . We then deduce (using the fact that the density (B.25) is
bounded below by a constant multiple of the Lebesgue measure on Ω0 ) that there is a constant
c2 = c2 (u, v, f, g, R) > 0 such that
Z
 k 
(B.26) P P yb (a + 1) ∈ Sk ≥ c2 dy,
Sk

for any measurable set Sk ⊆ x ∈ R : P k−1 yb+1 (a + 1) ≤ x ≤ P k−1 yb−1 (a + 1) . Applying this
n(T − 1) times yields a constant c3 = c3 (u, v, f, g, R) > 0 such that
P P k+n(T −1) y ∈ S ≥ c3 · ν0 (S),
 
(B.27)
for any measurable set
n o
S ⊆ x = x(t) t∈J1,T −1K ∈ (Wn )T −1 : P k yj+1 (t) ≤ xj (t) ≤ g(t),

for all (t, j) .

Applying (B.27) twice, first at k = 0 and then at k = n(T − 1), yields the lemma. □

Given these lemmas, we can quicky establish Lemma B.9.

Proof of Lemma B.9. That Gu;v


f is stationary for P follows from Remark B.8. Thus, the lemma
follows from Lemma B.11, using Lemma B.12 and Lemma B.13 to verify Assumption B.10. □

B.5. Proof of Lemma 2.18. In this section we establish Lemma 2.18. We begin with the following
concentration bound, which follows quickly from combining results of [61, 10], for the eigenvalues
of a GUE random matrix.
Lemma B.14 ([61, Corollary 5],[10, Corollary 1.6]). There exists a constant C > 1 such that the
following holds. Letting n ≥ 1 be an integer and (λ1 , λ2 , . . . , λn ) ∈ Wn denote the eigenvalues of an
n × n GUE matrix G = Gn , we have
" n #
[ n o 5
−2/3 −1/3
8
≤ Ce−(log n) .

P λj − γsemci;n (j) ≥ (log n) · n · min{j, n − j + 1}
j=1

Now we can establish Lemma 2.18.


LOCAL STATISTICS AND CONCENTRATION FOR NON-INTERSECTING BROWNIAN BRIDGES 89

Proof of Lemma 2.18. By Remark 2.1, we may assume throughout that u = v = 0. We may also
assume by shifting that a = −b = T , so that (a, b) = (−T, T ) for some T ≤ nD .
We next equate the law of x(t) with  the eigenvalues of a suitably scaled GUE matrix. To that
end, let λ(t) = λ1 (t), λ2 (t), . . . , λn (t) ∈ Wn denote Dyson Brownian motion with starting data
0n . From the second part of Lemma 2.15, λ(t) prescribes the law of n non-intersecting Brownian
motions (of variance n−1 ) run for time t, conditoned to never intersect; from the first part of
Lemma 2.15, it also prescribes the eigenvalues of an n × n GUE matrix, scaled by t1/2 .
Next, recall that a Brownian bridge B(t) on the interval [−T, T ], conditioned to start and end
T −t T +t

at 0 (that is, B(−T ) = B(T ) = 0), has the same law as (2T )1/2 · W T −t , where W : [0, ∞) → R is
a Brownian motion. This, implies (see also [19, Equation (2)]) that
(B.28) xj (t) has the same law as (2T )−1/2 (T 2 − t2 )1/2 · λj ;
Given (B.28), the proofs of both parts of the lemma are very similar, obtained by applying
Lemma B.14 on a suitable mesh of t; we therefore only detail the proof of the first statement of the
lemma. To that end, let C0 > 1 be the constant from Lemma B.14. For any integer j ∈ J1, nK and
real number t ∈ [−T, T ], define the events
[n  
G(t) = xj (t) − (2T )−1/2 2 2 1/2 8 1/2 −2/3
(T − t ) · γ(j) ≥ (log n) · T n · min{j, n − j + 1} −1/3
;

j=1

n ⌊T n20(D+1) ⌋
[ [ n o [  j 
G0 = xj (t + s) − xj (t) ≥ s1/2 n5 log |4s−1 T | ; G1 = G

,
j=1 a≤t<t+s≤b
n20(D+1)
j=−⌊T n20(D+1) ⌋

where we have abbreviated γ(j) = γsemci;n (j). By (B.28), Lemma B.14, (the B = n5 case of)
Lemma 2.6, a union bound, and the fact that T ≤ nD , there exist constants c > 0 and c1 =
c1 (D) > 1 such that
5 10 5
(B.29) P[G0 ∩ G1 ] ≤ 3C0 n20(D+1) T e−c1 (log n) + c−1 en/c−cn ≤ 2c−1
1 e
−c(log n)
,

for sufficiently large n. Moreover observe, since the function (T − t2 T −1 )1/2 is 12 -Hölder on [−T, T ]
with constant 21/2 , on the event G∁0 ∩ G∁1 we have for each j ∈ J1, nK that
t2 1/2
T

sup xj (t) − − · γ(j)
t∈[a,b] 2 2T

−1/2 2 2 1/2

≤ max xj (t) − (2T ) (T − t ) · γ(j)
n 20(D+1) t∈J−T n 20(D+1) ,T n20(D+1) K

!

(B.30) + sup sup xj (t + s) − xj (t)
t∈[a,b] s∈[0,n−20(D+1) ]
!

2 2 1/2 2 2 1/2

+ sup sup T − (t + s) − (T − t )
t∈[a,b] s∈[0,n−20(D+1) ]

≤ (log n)8 · T 1/2 n−2/3 · min{j, n − j + 1}−1/3 + 2n−10D−5 · log 10(T + 1)T n20(D+1)


≤ 2(log n)8 · T 1/2 n−2/3 · min{j, n − j + 1}−1/3 + n−D ,


for sufficiently large n. The lemma then follows from (B.29) and (B.30). □
90 AMOL AGGARWAL AND JIAOYANG HUANG

Appendix C. Proofs of Results From Section 3


C.1. Proofs of Lemma 3.3, Lemma 3.5, and Lemma 3.7. In this section we establish first
Lemma 3.3, then Lemma 3.5, and next Lemma 3.7.

Proof of Lemma 3.3. Denote the bridge-limiting measure process on [e a, eb] with boundary data
(µã , µb̃ ) by (eµ⋆t )t∈[ã,b̃] . Recalling the ensemble xn ∈ J1, nK × C [a, b] , condition on x(t) for

t∈/ [e a, eb]. Then, the law of the restriction y n = xn |[ã,b̃] ∈ J1, nK × C [e
a, eb] of x to [e a, eb] is given by
n n
Qx (ã);x (b̃) (An−1 ). Then Lemma 3.1 implies that, with probability 1 − o(1) (that is, tending to 1
as n tends to ∞),
(C.1) the measure A · emp y n (t) converges under the Lévy metric to µ⋆t , uniformly in t ∈ [e

a, eb].

In particular, (C.1) holds for t ∈ {e a, eb}, and so (with probability 1 − o(1)), the ensemble y
satisfies the conditions of Lemma 3.1, with (a, b, µa , µb ) there equal to (e a, eb, µã , µb̃ ) here. Hence,
 ⋆
with probability 1 − o(1), the measure A · emp y(t) converges to µ et , uniformly in t ∈ [e a, eb].
Together with (C.1), this yields the lemma. □

Proof of Lemma 3.5. If xn = (xn1 , xn2 , . . . , xnn ) ∈ J1, nK × C [a, b] is as in Lemma 3.1, then by
Remark 2.2 the ensemble x en = (e
xn1 , x
en2 , . . . , x
enn ) ∈ J1, nK × C [0, 1] defined by setting
ej (t) = A−1/2 (b − a)−1/2 · xj a + (b − a)t ,

x for each (j, t) ∈ J1, nK × [0, 1],
is as in Lemma 3.1,
 with the (a, b; A; µa , µb ) there equal to (0, 1; 1; µ
e0 , µ
e1 ) here. Then, recalling that
νtn = emp xn (t) and denoting νet = emp x en (t) for each t ∈ [0, 1], we have
νetn (I) = νa+t(b−a)
n
A1/2 (b − a)1/2 · I ,

for any interval I ⊆ R.
Letting n tend to ∞ and using Lemma 3.1, this verifies the first statement of (3.5). Together with
(3.3) and (3.4), this also quickly implies the second and third statements of (3.5). □

Proof of Lemma 3.7. The first statement of the lemma is due to [35, Corollary 2.8(b)], and the
second to [35, Theorem 2.7] (see also [7, Proposition 2.6(6)]). For the third, the estimates on ϱ⋆t
and ∂x ϱ⋆t follow from the second part of the lemma, together with Lemma 2.10.
For the fourth, let us first show that ϱt (x) is 71 -Hölder in x. In particular, we claim that there
exists a constant C1 > 1 such that, for any t ∈ (0, 1) and x, x′ ∈ R with |x − x′ | sufficiently small
(in a way only dependent on t), we have
ϱt (x ) − ϱ⋆t (x) ≤ C1 |x − x′ |1/7 t(1 − t) −1 .
⋆ ′ 
(C.2)
Assume to the contrary that this is false, so there exist t ∈ (0, 1) and x, x′ ∈ R such that (C.2)
does not hold; we may suppose by symmetry that x′ > x. Then, since ϱ⋆t (x) ≥ 0 and ϱ⋆t (x′ ) ≥ 0,
we either have ϱ⋆t (x) > C1 (x′ − x)1/7 (t − t2 )−1 or ϱ⋆t (x′ ) ≥ C1 (x′ − x)1/7 (t − t2 )−1 ; let us assume
the former holds. By the third part of the lemma, we have for some constant C2 > 1 that
−1 ′
∂x ϱ⋆t (y) ≤ C2 (x′ − x)−6/7 t(1 − t) 2 , whenever ϱ⋆t (y) ≥ t(1 − t) (x − x)1/7 .


Integrating this bound over a maximal subset of [x, x′ ] on which ϱ⋆t (y) ≥ (t − t2 )−1 |x − x′ |1/7 , we
find for any such y0 that
ϱt (y0 ) − ϱ⋆t (x) ≤ C2 |y0 − x| · (x′ − x)−6/7 t(1 − t) 2 ≤ C2 (x′ − x)1/7 t(1 − t) 2 .
⋆  
(C.3)
LOCAL STATISTICS AND CONCENTRATION FOR NON-INTERSECTING BROWNIAN BRIDGES 91

This implies for C1 > C2 + 2 that ϱ⋆t (y) ≥ |x − x′ |1/7 (t − t2 )−1 for any y ∈ [x, x′ ], so (C.3) applies
at y0 = x′ . This yields (C.2), which is a contradiction, so (C.2) holds for any (t, x) ∈ (0, 1) × R.
Now fix any point (t0 , x0 ) ∈ (0, 1) × R and sequence (t1 , x1 ), (t2 , x2 ), . . . ∈ (0, 1) × R with
limj→∞ (tj , xj ) = (t0 , x0 ). For any real number δ > 0, we have by the continuity of the mea-
sures µ⋆t in t that
Z x0 +δ Z x0 +δ
lim ϱtj (x)dx = ϱt0 (x)dx.
j→∞ x0 −δ x0 −δ

By (C.2), this implies for sufficiently large j (dependent on δ) that


Z x0 +δ  !
−1

ϱt0 (x0 ) − ϱtj (xj ) ≤ (2δ) ϱt0 (x) − ϱt0 (x0 ) + ϱtj (xj ) − ϱtj (x0 ) dx
x0 −δ
Z
x0 +δ Z x0 +δ −1 1/7
+ (2δ)−1 ϱt0 (x)dx − ϱtj (x)dx ≤ 4C1 t0 (1 − t0 ) δ + δ,

x0 −δ x0 −δ

which upon letting δ tend to 0 yields the continuity of ϱt (x) in (t, x). □

C.2. Proofs of Lemma 3.12, Lemma 3.13, and Lemma 3.23. In this section we establish
first Lemma 3.12, then Lemma 3.13, and next Lemma 3.23.

Proof of Lemma 3.12. Expanding the second and first equations in (3.17) yields
(C.4) ∂t ϱ + u∂x ϱ + ϱ∂x u = 0; u∂t ϱ + ϱ∂t u + 2uϱ∂x u + u2 ∂x ϱ − π 2 ϱ2 ∂x ϱ = 0.
The first equation of (C.4) is equivalent to the imaginary part of (3.19). Moreover, subtracting the
second equation of (C.4) from the first one multiplied by u then yields
ϱ(∂t u + u∂x u − π 2 ϱ∂x ϱ) = 0.
Since ϱt (x) > 0 for (t, x) ∈ Ω, this implies ∂t u + u∂x u − π 2 ϱ∂x ϱ = 0 on Ω, which is equivalent to
the real part of (3.19); this verifies (3.19). □

Proof of Lemma 3.13. Let us first show that H ⋆ is locally Lispchitz; that this holds in directions
parallel to the x-axis follows from the definition (3.3) and the bound on ϱ⋆t from the third part of
Lemma 3.7. It remains to show that H ⋆ is locally Lipschitz in directions parallel to the t-axis.
To that end, fix points (t1 , x0 ), (t2 , x0 ) ∈ (0, 1)×R with t1 < t2 ; by the second part of Lemma 3.7,
there exists some real number x1 > x0 such that ϱt (x) = 0 for each x > x1 − 1 and t ∈ (0, 1). For
any real number ε > 0, let φε : (0, 1) → R denote a smooth approximation of the indicator function
1t∈[t1 ,t2 ] · 1x∈[x0 ,x1 ] . Explicitly, fix a smooth function ψ : R2 → R≥0 with supp ψ ⊆ [−1, 1] × [−1, 1]
R∞ R∞
and with total integral −∞ −∞ ψ(t, x)dtdx = 1, and define ψε : R2 → R by setting ψε (t, x) =
ε−2 ψ(ε−1 t, ε−1 x) for each (t, x) ∈ R2 . Then, set
Z t2 Z x 1
φε (t, x) = ψε (t − s, x − w)dwds.
t1 x0

Since (u , ϱ ) weakly solves the equation ∂t ϱ⋆ = ∂x (u⋆ ϱ⋆ ), we have


⋆ ⋆

Z 1Z ∞ Z 1Z ∞
(C.5) ϱ⋆t (x)∂t φε (t, x)dxdt + u⋆t (x)ϱ⋆t (x)∂x φε (t, x)dxdt = 0.
0 −∞ 0 −∞
92 AMOL AGGARWAL AND JIAOYANG HUANG

The continuity of ϱ (due to the fourth part of Lemma 3.7), and the facts that ψε has total integral
equal to 1 and that supp ψε ⊆ [−ε, ε] × [−ε, ε], imply
Z 1Z ∞
lim ϱ⋆t (x)∂t φε (t, x)dxdt
ε→0 0 −∞
Z 1 Z ∞ Z x1
ϱ⋆t (x) ψε (t − t1 , x − w) − ψε (t − t2 , x − w) dwdxdt

(C.6) = lim
ε→0 0 −∞ x0
Z x1
ϱ⋆t1 (x) − ϱ⋆t2 (x) dx = Ht⋆1 (x0 ) − Ht⋆2 (x0 ),

=
x0

where we used the fact that Ht⋆j (x1 ) = 0 for each j ∈ {1, 2}, since ϱ⋆tj (x) = 0 for x > x1 .
Next, [35, Corollary 2.8(c)] states that, for any constant ω ∈ 0, 21 , there exists a constant


C = C(ω, µ0 , µ1 ) > 1 satisfying

sup sup u⋆t (x)ϱ⋆t (x) < C.



(C.7)
x∈R t∈[ω,1−ω]

We further have
Z 1 Z ∞
lim u⋆t (x)ϱ⋆t (x)∂x φε (t, x)dxdt
ε→0 0 −∞
Z 1 Z ∞ Z t2 Z x1
= − lim u⋆t (x)ϱ⋆t (x)∂w ψε (t − s, x − w)dwdsdxdt
ε→0 0 −∞ t1 x0
(C.8) Z 1 Z ∞ Z t2
u⋆t (x)ϱ⋆t (x) ψε (t − s, x − x0 ) − ψε (t − s, x − x1 ) dsdxdt

= lim
ε→0 0 −∞ t1
Z 1 Z ∞ Z t2
= lim u⋆t+s (x + x0 )ϱ⋆t+s (x + x0 )ψε (t, x)dsdxdt,
ε→0 0 −∞ t1

where in first equality we used the fact that ∂x ψε (t − s, x − w) = −∂w ψε (t − s, x − w); in the second
we performed the integration; and in the third we used the fact that u⋆t (x)ϱ⋆t (x)ψε (t − s, w − x1 ) = 0
for sufficiently small ε (as ϱ⋆t (x) = 0 for x ≥ x1 − 1 and supp ψε ⊆ [−ε, ε] × [−ε, ε]) and changed
variables from (t, x) to (t + s, x + x0 ). Now, since ψε is nonnegative with total integral
equal to
1, inserting (C.8) and (C.6) into (C.5), and using (C.7) yields Ht⋆1 (x0 ) − Ht⋆2 (x0 ) ≤ C1 for some
constant C1 = C1 (t1 , t2 , µ1 , µ2 ) > 1. In particular, Ht⋆ is locally Lipschitz.
It remains to verify (3.20); its first statement follows from (3.3), so it remains to confirm its
second. To that end, the Lebesgue differentiation theorem yields, for almost all (s, x0 ) ∈ (0, 1) × R,
Z 1Z ∞
lim u⋆t+s (x + x0 )ϱ⋆t+s (x + x0 )ψε (t, x)dxdt = u⋆s (x0 )ϱ⋆s (x0 ).
ε→0 0 −∞

With (C.7), this implies for almost every x0 ∈ R that


Z 1 Z ∞ Z t2 Z t2
lim u⋆t+s (x + x0 )ϱ⋆t+s (x + x0 )ψε (t, x)dsdxdt = u⋆s (x0 )ϱ⋆s (x0 )ds.
ε→0 0 −∞ t1 t1

Together this, (C.5), (C.6), and (C.8) gives for almost all x0 ∈ R that Ht⋆2 (x0 ) − Ht⋆1 (x0 ) =
R t2 ⋆
u (x )ϱ⋆ (x )ds, which establishes the second part of (3.20).
t1 s 0 s 0

LOCAL STATISTICS AND CONCENTRATION FOR NON-INTERSECTING BROWNIAN BRIDGES 93

Proof of Lemma 3.23. We only establish the first part of the lemma, as the proofs of the second
and third are entirely analogous. To do this, it suffices to show that H ⋆ is smooth on Ω, as from
(3.4) this quickly implies that G⋆ is smooth on Ωinv .
To that end, fix some point (t0 , x0 ) ∈ Ω; we must show that H ⋆ is smooth in a neighborhood
of (t0 , x0 ). Let a′ , b′ ∈ (a, b) be real numbers such that a′ < t0 < b′ . Then, Lemma 3.3 implies
that (µt )t∈[a′ ,b′ ] is the bridge-limiting measure process on [a′ , b′ ] with boundary data (µa′ ; µb′ ).
Furthermore, Lemma 3.7 implies that the densities ϱa′ of µa′ and ϱb′ of µb′ are compactly supported
and uniformly bounded. Hence, Lemma 3.3, the second part of Lemma 3.11, and Remark 3.14 (with
the (a, b; µa , µb ) there equal to (a′ , b′ ; µa , µb ) here) imply that ϱ⋆t (x) and u⋆t (x) are smooth for (t, x)
in a neighborhood of (t0 , x0 ). Together with (3.20) (again with Remark 3.14), this implies that H ⋆
is smooth in a neighborhood of (t0 , x0 ), verifying the lemma. □

Appendix D. Proofs of Results From Section 8


D.1. Proof of Lemma 8.2. In this section we establish Lemma 8.2. We first require the following
definition from [50], which defines a class of initial data for which it is possible to show quick
convergence of Dyson Brownian motion. In what follows, we recall the Stieltjes transform mν
associated with a measure ν from (2.5), and the empirical measure emp(a) associated with a
sequence a from (3.1).
Definition D.1 ([50, Definition 2.1]). Fix real numbers E0 ∈ R and ζ > 0. For each positive
real number n, let r = rn and R = Rn be two parameters satisfying nζ−1 ≤ r ≤ n−ζ and
nζ r ≤ R ≤ n−ζ . We call an n-tuple d = (d1 , d2 , . . . , dn ) ∈ Rn (r, R; ζ)-regular with respect to E0 if
the following holds. Denoting ν = ν n = emp(d), there exist constants c ∈ (0, 1) and C > 1 (both
independent of n) such that maxj∈J1,nK |dj | ≤ nC and
c ≤ Im mν n (E + iη) ≤ C, for all n ≥ C and (E, η) ∈ (E − R, E + R) × (r, 10).

Lemma D.2. On the event A0 , the n′ -tuples w


e − (0), w(0), e + (0) are (n′ )ω/2−1 ; (n′ )−ω/2 ; ω2 -

e and w
regular with respect to 0.
Proof. We only show the lemma for w(0),
e as the proof is entirely analogous
 for we − (0) and w
e + (0);
throughout, we restrict to the event A0 and denote νe = νe = emp w(0)
n e . Then, for any z ∈ H,

n
1 X 1
mν (z) = ′ .
ej (0) − z
n j=1 w

Let z = E + iη for some real numbers E and η with |E| ≤ (n′ )−ω/2 and (n′ )ω/2−1 ≤ η ≤ 10. Then,
n′ n′

1 X 1 1 X e jn′−1 ) − w
G(0, ej (0)
mν (z) ≤ + ′

n′

j=1 G(0, jn
e ′−1 ) − z n j=1 w e jn′−1 ) − z
ej (0) − z · G(0,
n′

1 X 1 1 j
≤ ′ + ′ 2 max′ G 0, ′ − w ej (0)
e
n ′−1 ) − z n η j∈J1,n K n
j=1 G(0, jn
e

n′
n

1 X 1 1 X 1
−2 ′ ω/2−2
≤ ′ + η (n ) ≤ ′ + (n′ )−ω/2 ,

n e jn′−1 ) − z
G(0, n e jn′−1 ) − z
G(0,
j=1 j=1
94 AMOL AGGARWAL AND JIAOYANG HUANG

j
2 2
where in the second bound we used the fact that w ej (0) − z · G(0,
e
n′ ) − z ≥ (Im z) = η ; for

the third we used the fact that we are restricting to A0 , together with (8.12); and in the fourth we
used the fact that η ≥ (n′ )ω/2−2 .
Now let j0 ∈ J1, n′ K be the integer (if it exists) such that G e 0, j0′ ≤ E < G e 0, j0 −1
 
n n′ . Observe
ε 1 ′ −ω/2
from (8.13) that ∂y G ≤ − 2 and G(0, 2 ) = 0. Therefore, since |E| ≤ (n )
e e , it follows for
sufficiently large n that j0 exists and

n′
j0 − ≤ 2ε−1 (n′ )−ω/2 .

2

n′ n′
We will assume in what follows that j0 ≥ 2 , as the proof in the alternative case when j0 < 2
is entirely analogous. Then,

n′

1 X 1


n

e jn′−1 ) − z
G(0,
j=1
n′ −n′ −1
  2j0X
1 1 1 1 1
X
≤ ′ + + ′

2n n

G(0, jn′−1 ) − z G(0, 2j n ′−1 − jn′−1 ) − z G(0, jn′−1 ) − z
0
e e e
j=2j0 −n′ j=1
n′

1 X e jn′−1 ) + G(0,
G(0, e 2j0 n′−1 − jn′−1 ) − 2z
≤ ′  + (n′ )−ω/8 ,

n


e jn′−1 ) − z G(0,
G(0, e 2j0 n′−1 − jn′−1 ) − z
j=2j0 −n

where in the second bound we used the fact that for j < 2j0 − n′ ≤ 4ε−1 (n′ )1−ω/2 we have

e 0, j − z ≥ G
e 0, j − E
  
G
n′ n′
′ −ω/2
e 0, j0 − 1 ≥ ε(j0 − j − 1) ≥ ε − 10(n )
 j   ε
≥Ge 0,

− G ′ ′
≥ ≥ (n′ )−ω/8 ,
n n 2n 4 ε 8

by (8.13). We additionally have for sufficiently large n that


  
e 0, j + Ge 0, 2j0 − j − 2z ≤ G
e 0, j + G e 0, 2j0 − j − 2G e 0, j0 + 2η + 4
      
G
n′ n′ n′ n′ n′ εn′
(D.1)
2B0 n′  j0 − j 2 2B0 (j − j0 )2
≤ · + 3η = + 3η,
n n′ nn′

e 0, j′ ≤ E < G e 0, j−1
 
where in the first inequality we used the facts that z = E + iη with G n n′ and

−2ε−1 ≤ ∂y G e ≤ − ε (by (8.13)), and in the second we used the facts that [G]e 2 ≤ 2B0 n (again by
2 n
(8.13)) and that η ≥ (n′ )ω−1 . We also have that
 
e 0, j − z ≥ ε|j − j0 | + η ; G 0, 2j0 − j − z ≥ ε|j − j0 | + η ,
 
(D.2) G

n 4n′ 4 n ′ 4n′ 4
LOCAL STATISTICS AND CONCENTRATION FOR NON-INTERSECTING BROWNIAN BRIDGES 95

e ≤ − ε . Together, (D.1) and (D.2) yield


where in both inequalities we again used the facts that ∂y G 2

n e jn′−1 ) + G(0,
e 2j0 n′−1 − jn′−1 ) − 2z
1 X G(0,
n′
 

e jn′−1 ) − z G(0,
G(0, e 2j0 n′−1 − jn′−1 ) − z
j=2j0 −n

n 2B0 |j−j0 |2
1 X
nn′ + 3η
≤ ′ ε2 |j−j0 |2 η2
n
j=2j0 −n′ 16(n′ )2 + 16
′ ′ ′
ηn 
n n
!
−2
X n−1 j 2 + ηn′ −2
X j2 1  X 1 ηn′ 
≤ 48B0 ε ≤ 48B0 ε + + + 2 ≤ C,
j=0
j 2 + (ηn′ )2 j=0

n(ηn ) 2 ηn′

n j
j=ηn

for some constant C = C(ε, δ, B0 ) > 1; this yields the lemma. □

The following result of [49]12 (see also earlier results from [50]) states that, if one runs Dyson
Brownian motion on initial data that is (r; R; ζ)-regular with respect to a point E0 ∈ R, then its
local statistics after run for some time t between r and R2 are given by the sine kernel. In what
follows, we recall notation on free convolutions from Section 2.2.

Lemma D.3 ([49, Theorem 2.2]). For any integer k ≥ 1, real numbers ζ, ς > 0, and com-
pactly supported, smooth function F ∈ C ∞ (Rk ), there exist constants c = c(k, ζ, ς) > 0 and
C = C(k, ζ, ς, F ) > 1 such that the following holds. Adopt the notation and assumptions of
(t)
Definition D.1; fix some real number t ∈ [nς r, n−ς R2 ]; define the measure νt = ν ⊞ ϱsemci ; and
denote the the density of νt by ϱt ∈ L1 (R). Fix some real number E ∈ E0 − R2 , E0 + R2 ; let
d(t) = d1 (t), d2 (t), . . . , dn (t) denote Dyson Brownian motion, run for time t, with initial data d;
(k)
and denote its k-point correlation function by pd(t) (recall (1.5)). Then, we have
Z   Z

(k) a (k)

(D.3) F (a)pd(t) E + da − F (a)psin (a)da < Cn−c .

nϱt (E)

Rk Rk

Now we can outline the proof of Lemma 8.2.

Proof of Lemma 8.2 (Outline). We only address the case when a = 0, as the proof is entirely
analogous if a ̸= 0. Recalling the notation from Section 8.2 (and, in particular, (8.11)), denote the
Pn′
measure ν0 = (n′ )−1 j=1 δw̃j (0) ∈ P0 . For any real number s ≥ 0, denote its free convolution with
(s)
the semicircle law by νs = ν0 ⊞ µsemci , and denote the associated density by ϱs ∈ L1 (R), which
satisfies νs (dx) = ϱs (x)dx.
Observe that se0 = (n′ )ω−1 (by (8.1) and (8.11)) and that G e se0 , 1 = 0 (by (8.13)). This,

2
together with Lemma D.2 and LemmaD.3 (with the parameters (n; E0 , E; t; r, R; ζ, ς) there equal
to n′ ; 0, 0, se0 ; (n′ )ω/2−1 , (n′ )−ω/2 ; ω2 , ω4 here) then imply that
Z   Z
(k) a (k)
lim F (a)pw̃(s̃0 ) da = F (a)psin (a)da.
n→∞ Rk n′ ϱs̃0 (0) Rk

12It was stated there for Dyson Brownian motion associated with the Gaussian Orthogonal Ensemble (GOE),
but it applies equally well to that associated with the GUE, as stated here.
96 AMOL AGGARWAL AND JIAOYANG HUANG


It follows from applying the change of variables replacing a with nn · a, and using (8.11), and that
Z   Z
(k) a (k)
lim F (a)pw̆(s0 ) da = F (a)psin (a)da.
n→∞ Rk nϱs̃0 (0) Rk

e se0 , 1 and since limn→∞ ∂x G e se0 , 1 = ∂x G e 0, 1 (as limn→∞ se0 = 0


  
Since θ = ∂y G(t0 , y0 ) = ∂y G 2 2 2
e 0, 1 .

and [G]e 1 = [G]1 ≤ B0 ), it therefore suffices to show that limn→∞ ϱs̃ (0) = ∂x G
0 2
Stated alternatively, since the approximate density of w ej (0) around 0 is ∂x G e 0, 1 , this amounts
2
s̃0
to showing that the effect of applying the free convolution with µsemci for se0 small has a negligible
effect on this density. This is a standard statement about free convolutions that has been proven
and applied numerous times in the random matrix literature; we therefore do not repeat its proof
here and refer, for example, to [42, Lemmas 3.3 and 3.4] for very similar arguments. □

D.2. Proof Outline of Lemma 8.3. In this section we provide a brief outline for the proof of
Lemma 8.3, using results of [50].

Proof of Lemma 8.3 (Outline). We only outline the proof of the first statement of (8.5), as that of
the second is entirely analogous. Following [50, Equation (4.10)], we define the processes ξ(s) =
ξ1 (s), ξ2 (s), . . . , ξn′ (s) and ξ − (s) = ξ1− (s), ξ2− (s), . . . , ξn−′ (s) by for each pair (j, s) ∈ J1, n′ K ×
 

[0, ns0 ] setting


s s
(D.4) ξj (s) = n · w̆j ; ξj− = n · w̆j− ,
n n
which satisfy the stochastic differential equations
X dt X dt
(D.5) ∂t ξi (t) = + dBi (t); ∂t ξi− (t) = − − + dBi (t).

ξi (t) − ξj (t) ′ ξi (t) − ξj (t)
1≤j≤n 1≤j≤n
j̸=i j̸=i

We further define for any distinct integers j, k ∈ J1, n′ K the quantity


∆jk = (2 · 1j<k − 1) · (n′ )−500 .

Then following [50, Equation (4.16)] we define the processes ξb = ξb1 (s), ξb2 (s), . . . , ξbn′ (s) and
ξb− (s) = ξb1− (s), ξb2− (s), . . . , ξbn−′ (s) obtained by “regularizing” the denominators on the right sides


of (D.5), namely, as the (strong) solutions to the stochastic differential equations


(D.6)
X dt X dt
∂t ξbi (t) = + dBi (t); ∂t ξbi− (t) = + dBi (t),
ξi (t) − ξj (t) + ∆ij ξi− (t) − ξj− (t) + ∆ij
1≤j≤n′ 1≤j≤n′
j̸=i j̸=i

b = ξ(0) and ξb− (0) = ξ − (0). It then follows from13 [50, Lemma 4.4] (together
with initial data ξ(0)
with Lemma D.2) that ξb and ξb− likely are close to ξ and ξ − , respectively, namely, with probability

13Let us mention that there are two differences between the setup of [50, Lemma 4.4] and that here. First, [50]
considers the β = 1 (GOE) case of Dyson Brownian motion, while here we consider the β = 2 (GUE) case. Second,
1
the differential equations (D.5) and (D.6) in [50] have an additional − 2n ′ · ξj (t) term on their right sides. However,
it is quickly verified that neither point affects the proofs or statements of the results from [50].
LOCAL STATISTICS AND CONCENTRATION FOR NON-INTERSECTING BROWNIAN BRIDGES 97

at least 1 − (n′ )90 we have


 
ξj (t) − ξbj (t) + ξ − (t) − ξb− (t) ≤ (n′ )−15 .

(D.7) sup j j
t∈[0,(n′ )1−3ω ]
j∈Jn′ /2−(n′ )1−ω ,n′ /2+(n′ )1−ω K

Now set14 K = (n′ )1/6 . Following [50, Equation (4.24)], we define the processes of (2K +1)-tuples
ξ(s) = ξen′ /2−K (s), . . . , ξen′ /2+K (s) and ξe− (s) = ξen−′ /2−K (s), . . . , ξen−′ /2+K (s) by only retaining the
 
e
middle 2K + 1 of the paths in (D.6), namely, as the (strong) solutions to the stochastic differential
equations
X dt
∂t ξei (t) = + dBi (t);
′ ′
ξi (t) − ξj (t) + ∆ij
n /2−K≤j≤n /2+K
j̸=i
(D.8) X dt
∂t ξei− (t) = + dBi (t),
ξ − (t) − ξj− (t) + ∆ij
n′ /2−K≤j≤n′ /2+K i
j̸=i
q ′ ′
with initial data ξei (0) = ξbi (0) and ξei− (0) = ξbi− (0), for each i ∈ n2 − K, n2 + K . Observe from
y
′ ′ ′ ′
+ nj = G− y0 − 2n + nj for j ∈ n2 − K, n2 + K that
n
 n
 q y
(8.4) and the fact that G t0 − s0 , y0 − 2n
(D.9) e = ξe− (0).
ξ(0)

Now define the processes of (2K + 1)-tuples ϑ(s) b = ϑbn′ /2−K (s), . . . , ϑbn′ /2+K (s) and ϑ(s)
e =

ϑn /2−K (s), . . . , ϑn /2+K (s) by setting
e ′ e ′

r n′ n′ z
ϑbi (s) = ξbi (s) − ξbi− (s), and ϑei (s) = ξei (s) − ξei− (s); for each j ∈ − K, +K .
2 2


To compare ϑ(s) and ϑ(s) first observe,
b e by (8.4) and the fact that G(t0 − s0 , y) − G (y) ≤ nω/2−1

(and that n′ = n1/2 ), that we have ξj (0) − ξj− (0) ≤ (n′ )ω for each j ∈ J1, n′ K. Together with

Lemma 8.7, it follows that ξj (s) − ξj− (s) ≤ (n′ )ω for each s > 0. Hence, subtracting (D.8) from

(D.6) and integrating over t ∈ [0, ns0 ], we obtain


(D.10)

ϑei (ns0 ) − ϑbi (ns0 ) = ξei (ns0 ) − ξe− (ns0 ) − ξbi (ns0 ) − ξb− (ns0 )

i i
 
X 1 1
≤ ns0 · sup − −


ξ (t) − ξ (t) + ∆ −

t∈[0,ns0 ]
j:|j−n′ /2|>K
i j ij ξi (t) ξ j (t) + ∆ ij
ξi (t) − ξ − (t) + ξj (t) − ξ − (t)

i j
X
≤ ns0 · sup ,
ξi (t) − ξj (t) + ∆ij ξ − (t) − ξ − (t) + ∆ij

t∈[0,ns0 ] i j
j:|j−n′ /2|>K
X 1
≤ 2(n′ )2ω sup ,
ξi (t) − ξj (t) ξ − (t) − ξ − (t)

t∈[0,ns0 ] i j
j:|j−n′ /2|>K

14The choice of K in [50] was slightly different, but it can quickly be verified that the below statements
will valid for K = (n′ )1/6 as well. Indeed, fixing K = (n′ )1/6 correponds in [50] to imposing that w e be
n−2/3−o(1) , n−1/3−o(1) ; 31 + o(1) -regular around 0; this is guaranteed Lemma D.2 (which implies that w

e is even
more regular than this).
98 AMOL AGGARWAL AND JIAOYANG HUANG

q ′ n′ +K
for any i ∈ n −K
y
2 , 2 , where in the second bound we used the fact that ns0 = (n′ )ω . By
Lemma 2.17, and the first statement of Lemma 8.6, there exists a constant c > 0 such that with
2
probability 1 − c−1 e−c(log n) we have ξi (t) − ξj (t) ≥ 8ε |i − j| ≥ (n′ )−ω K and ξi (t) − ξj (t) ≥
′ ′
n′ +K
K whenever j − n2 > K and i ∈ n −K
ε ′ −ω
q y
8 |i − j| ≥ (n ) 2 , 2 . Inserting this into (D.10) yields
r n′ − K n′ + K z
ϑei (ns0 ) − ϑbi (ns0 ) ≤ 2(n′ )4ω K −2 = 2(n′ )4ω−1/3 ,

(D.11) for each i ∈ , ,
2 2
where we have recalled that K = (n′ )1/6 . 
Following [50, Equation (4.38)], we next define the process ϑ̊(s) = ϑ̊n′ /2−K (s), . . . , ϑ̊n′ /2+K (s)
as the solution to the differential equation

X ϑ̊i (s) − ϑ̊j (s) dt
(D.12) ∂t ϑ̊i (t) = ,
ξi (t) − ξj (t) + ∆ij ξi− (t) − ξj− (t) + ∆ij

′ ′
n /2−K≤j≤n /2+K
j̸=i

with initial data ϑ̊(0) = ϑ(0)


e = (0, 0, . . . , 0), where the last equality follows from (D.9). Combining
this initial condition with the explicit form of (D.12), it follows
r n′ n′ z
(D.13) ϑ̊i (s) = 0, for each i ∈ − K, +K .
2 2
Next, [50, Lemma 4.8]15 implies (with a union bound over 2nω/15000 + 1 = 2(n′ )ω/7500 + 1 ≤
K indices) that, with probability at least 1 − K −ω/1000 we have
ω/1000

ϑ̊j (ns0 ) − ϑej (ns0 ) ≤ K 4ω−1/2 .



(D.14) max
j∈Jn′ /2−nω/15000 ,n′ /2+nω/15000 K

Combining (D.14), (D.13), (D.11), and (D.7) yields with probability at least 1 − n−ω/15000 that
ξj (ns0 ) − ξj− (ns0 ) ≤ K 4ω−1/2 + 2(n′ )4ω−1/3 + (n′ )−15 ≤ (n′ )ω−1/12 . Together with (D.4) (and the
1
fact that ω ≤ 20 ), this yields the lemma. □

D.3. Proof Outline of Lemma 8.4. In this section we provide a brief outline for the proof of
Lemma 8.4.
q ′ ′
Proof of Lemma 8.4 (Outline). Throughout this proof, we fix an integer j ∈ n2 − nω/15000 , n2 +
y
nω/15000 ; we may also assume that χ < ω. It suffices to show that, for any real number A ≥ 1,
h i h i
P w̆j− (s0 ) ≥ An5 < Ce−cAn ; P w̆j+ (s0 ) ≥ An5 < Ce−cAn ,

(D.15)

and that, for any coupling between w̆− and w̆+ ,


h i 2
P w̆j+ (s0 ) − w̆j− (s0 ) > nχ−1 ≤ Ce−c(log n) .

(D.16)

15It was stated there when the ξ− paths come from applying Dyson Brownian motion to the eigenvalues of a
GOE random matrix, but it is quickly verified that the proof and statement of that result applies equally well for
the ξ− here (satisfying the regularity hypothesis shown in Lemma D.2).
LOCAL STATISTICS AND CONCENTRATION FOR NON-INTERSECTING BROWNIAN BRIDGES 99

Indeed, given these two bounds, it would then follow that


h i
E w̆j+ (s0 ) − w̆j+ (s0 ) · 1 +


− χ−1
|w̆j (s0 )−w̆j (s0 ) >n

h i h i
≤ 2n · P w̆j+ (s0 ) − w̆j− (s0 ) > nχ−1 + E w̆j+ (s0 ) · 1|w̆+ (s0 )|>n10
10

j
h i


+ E w̆j (s0 ) · 1|w̆− (s0 )|>n10

j
Z ∞ h i h i
2
≤ 2Cn10 e−c(log n) + n5 P w̆j+ (s0 ) ≥ n5 A + P w̆j− (s0 ) ≥ n5 A dA

n5
10 −c(log n)2 −1
≤ 2Cn e + 2c Cn5 e−cn ,
verifying the lemma. The bound (D.15) is a quick consequence of Lemma 2.6 with the fact that
Dyson Brownian motion λ(s0 ), run for time s0 ≥ 0 with initial data given by some n-tuple a, has
e s0 , where λ

the same law as (1 + s0 ) · λ 1+s0
e is a family of n non-intersecting Brownian bridges on
a;0n
[0, 1] sampled from the measure Q (where 0n = (0, 0, . . . , 0) consists of n zeroes).
To establish (D.16), it suffices to exhibit (deterministic) real numbers θj− , θj+ ∈ R such that
(D.17)
h i 2
h i 2
P w̆+ (s0 ) − θ+ > nχ/2−1 ≤ Ce−c(log n) ; P w̆j− (s0 ) − θj− > nχ/2−1 ≤ Ce−c(log n) .

j j

Indeed this, Lemma 8.3, a union bound, would yield a constant c1 = c1 (χ, ε, δ, B0 , m) > 1 such that
P |θj+ − θj− | > n3χ/4−1 ≥ P |θj+ − θj− | > 2nχ/2−1 + n−c1 −1 ≥ 1 − c−1 −c1
   
1 n > 0.
Since θj+ and θj− are deterministic, it follows that |θj+ − θj− | ≤ n3χ/4−1 . Together with (D.17), a
union bound, and the fact that nχ−1 > 2nχ/2−1 + n3χ/4−1 for sufficiently large n, this implies
h i h i h i
P w̆j+ (s0 ) − w̆j+ (s0 ) > nχ−1 ≤ P w̆j+ (s0 ) − θj+ > nχ/2−1 + P w̆j− (s0 ) − θj− > nχ/2−1

2
≤ 2Ce−c(log n) ,
thereby confirming (D.16).
It remains to establish (D.17); let us only explain the first bound there, since the proof of
the second is entirely analogous. This will follow from the concentration bound Lemma 2.17. In
Pn′
particular, denote the measure ν0+ = (n′ )−1 j=1 1nŭ+ /n′ . For any real number s ≥ 0, also define
j
(s)
its free convolution νs+ = ν0+ ⊞ µsemci with the semicircle distribution; let ϱ+
s denote its density with
ν+
respect to Lebesgue measure, and for each k ∈ J1, n′ K let γk+ (s) = γk;n s
′ denote its classical location

(recall Definition 2.16). By Lemma 2.17 (and Remark 2.13), denoting se0 = ns n′ = n
0 ω−1
(by (8.1)),
there exist constants c2 > 0 and C1 > 1 such that
 ′
n −2 n′ + −2
i 2
P +
· γj+⌊(log n)5 ⌋ (e
s0 ) − n ≤ w̆ +
j (s0 ) ≤ · γ j−⌊(log n)5 ⌋ (e
s0 ) + n ≥ 1 − C1 e−(log n) .
n n
n′
Taking θj+ = n · γj (e
s0 ), to prove (D.17) it therefore suffices to show for sufficiently large n that

+ + (log n)10
γj−⌊(log s0 ) − γj+⌊(log
n)5 ⌋ (e s0 ) ≤
n)5 ⌋ (e .
n′
 +
ε
s0 ) − (n′ )2ω , γj+ (e
s0 ) + (n′ )2ω , which can be

This follows from the fact that ϱ+s0 (x) ≥ 2 for x ∈ γj (e
deduced from the fact that se0 = n1−ω is small (by (8.1)); that nn′ · w̆+ (0) = nn′ · ŭ+ approximates G to
100 AMOL AGGARWAL AND JIAOYANG HUANG

q ′ ′′ n′ +n′′ y
within (n′ )ω/2−1 ≪ se0 , for indices k ∈ n −n
2 , 2 (by (8.4) and (8.2)); and that G ∈ Admε (R).
This last statement is a standard one about free convolutions that has been proven and applied
numerous times in the random matrix literature; we therefore do not repeat its proof here and refer,
for example, to [42, Lemmas 3.3 and 3.4] for very similar arguments. □

Appendix E. Proofs of Results From Section 9


E.1. Proofs of Lemma 9.1, Lemma 9.2, and Lemma 9.3. In this section we first recall several
results from [33] about regularity properties of uniformly elliptic partial differential equations and
then use them to establish Lemma 9.1, Lemma 9.2, and Lemma 9.3. To that end, we fix an integer
d ≥ 1 and an open subset R ⊆ Rd , and we recall the notation from Section 1.3. For any function
f ∈ C(R); real number α ∈ (0, 1); and integer k ∈ Z≥0 , define [f ]α = [f ]α;R , [f ]k,α = [f ]k,α;R , and
∥f ∥C k,α (R) = ∥f ∥C k,α (R) = ∥f ∥k,α = ∥f ∥k,α;R by

f (y) − f (z)
[f ]α = sup ; [f ]k,α = max [∂γ f ]α ; ∥f ∥C k,α (R) = ∥f ∥k + [f ]k,α .
y,z∈R |y − z|α γ∈Zd
≥0
y̸=z |γ|=k

Define the Hölder space C k,α (R) to be the set of f ∈ C k (R) such that [f ]
k,α;R0 < ∞, for any
compact subset R0 ⊂ R. Also define C k,α (R) = f ∈ C k (R) : [f ]k,α;R < ∞ . Then ∥f ∥k,α is a
norm on C k,α (R). By restriction, we have that C k,α (R) ⊆ C k,α (U) for any subset U ⊆ R. Observe
that, for any f, g ∈ C 0,α (R), these norms satisfy
(E.1) ∥f g∥C 0,α (R) ≤ ∥f ∥C 0,α (R) ∥g∥0 .

We further say that the boundary ∂R of some open set R ⊂ Rd is C k,α , if for any point z0 ∈ ∂R,
there exists a real number r > 0 and an injective function ψ = ψz0 ;r : Br (z0 ) → Rd such that
Dz0 ;r;ψ;R = ψ Br (z0 ) ∩ R ⊂ R>0 × Rd−1 ; ψ Br (z0 ) ∩ ∂R ⊂ {0} × Rd−1 ;
 

where ψ ∈ C k,α Br (z0 ) and ψ −1 ∈ C k,α (Dz0 ;r;ψ;R ). We call this boundary C k if it is C k,0 , and we


call it smooth if it is C k for each integer k ≥ 1.


The following lemma provides the Schauder estimates for solutions to uniformly elliptic partial
differential equations with Hölder continuous coefficients; as in Lemma 9.2 and Lemma 9.4, we state
a form of it that addresses interior (R′ empty) and boundary regularity (R′ = R) simultaneously.
Although these bounds hold in all dimensions d ≥ 2, we only state them for d = 2.
Lemma E.1 ([33, Theorem 6.6]). Fix real numbers r > 0, B > 1, and α ∈ (0, 1), and bounded open
sets R′ ⊆ R ⊂ R2 such that R′ has C 2,α boundary. There exists a constant C = C(r, B, α, R′ ) > 1
such that the following holds. For each j, k ∈ {1, 2}, let ajk , bj ∈ C 0,α (R) denote functions such
that axy = ayx and
X
max ∥ajk ∥0,α < B; max ∥bj ∥0,α < B; inf ajk (z)ξi ξj ≥ B −1 |ξ|2 ,
j,k∈{1,2} j∈{1,2} z∈R
j,k∈{1,2}

all hold for any ξ = (ξ1 , ξ2 ) ∈ R . Let g ∈ C (R) denote a function, and suppose that F ∈ C 2,α (R)
2 0,α

satisfies the elliptic partial differential equation


X X
(E.2) ajk (z)∂j ∂k F (z) + bj (z)∂j F (z) = g, for each z ∈ R.
j,k∈{1,2} j∈{1,2}
LOCAL STATISTICS AND CONCENTRATION FOR NON-INTERSECTING BROWNIAN BRIDGES 101

Assume that there exists some function φ ∈ C 2,α (R) such that F (z) = φ(z), for each z ∈ ∂R′ ∩ ∂R.
Then, recalling the set Dr ⊆ R from (9.1), we have

∥F ∥C 2,α (Dr ) ≤ C ∥F ∥C 0 (R) + |φ∥C 2,α (R) + ∥g∥C 0,α (R) .
We will be interested in properties of solutions to certain families of non-linear, uniformly elliptic
partial differential equations. The following result provides a comparison principle for such solutions.
Lemma E.2 ([33, Theorem 10.1]). Fix a constant B > 0 and a bounded open subset R ⊂ R2 . For
each j, k ∈ {1, 2}, let ajk ∈ C 1 (R2 ) denote functions such that axy = ayx and
X
(E.3) B −1 |ξ|2 ≤ ajk (w)ξj ξk ≤ B|ξ|2 , for any ξ = (ξ1 , ξ2 ) ∈ R2 and w ∈ R2 .
j,k∈{1,2}

Moreover, for each i ∈ {1, 2}, let fi : ∂R → R denote a continuous function, and suppose that
Fi : R → R with Fi ∈ C 2 (R) ∩ C(R) is a solution to the partial differential equation
X 
(E.4) ajk ∇F (z) ∂j ∂k F (z) = 0, for each z ∈ R,
j,k∈{1,2}

with boundary data Fi |∂R = fi . If f1 (z) ≥ f2 (z) for each z ∈ ∂R, then F1 (z) ≥ F2 (z) for each
z ∈ R.
The following lemma provides interior and boundary regularity estimates (which are stated
simultaneously, as in Lemma 9.2) for solutions F to equations of the form (E.4); it is known from
[33, Corollary 6.7], [33, Theorem 8.29], [33, Equation (13.41)], together with [5, Chapter 3.6].
Lemma E.3 ([5, 33]). For a fixed integer m ∈ Z>1 ; real numbers r, B ∈ R>0 ; and bounded open
subsets R′ ⊆ R ⊂ R2 with C m+1 boundaries, there exists a constant C = C(r, B, m, R′ ) > 1 such
that the following holds. For each j, k ∈ {1, 2}, let ajk ∈ C m−1 (R2 ) denote functions with axy = ayx
and
X
(E.5) max ∥ajk ∥C m−1 (R2 ) ≤ B; inf2 ajk (z)ξj ξk ≥ B −1 |ξ|2 ,
j,k∈{1,2} z∈R
j,k∈{1,2}

for any ξ = (ξ1 , ξ2 ) ∈ R2 . Then for any continuous function f ∈ C(∂R) such that ∥f ∥0 ≤ B, the
following two statements hold.
(1) There exists a unique F ∈ C 2 (R) ∩ C(R) satisfying the partial differential equation (E.4)
with boundary data F |∂R = f .

(2) Suppose there exists a function φ ∈ C m+1 (R ) such that ∥φ|∥C m+1 (R′ ) ≤ B and f (z) = φ(z)
for each z ∈ ∂R′ ∩∂R. Letting Dr = z ∈ R : d(z, ∂R\∂R′ ) > r , we have ∥F ∥C m (Dr ) ≤ C.


From the above statements, we can quickly establish Lemma 9.1, Lemma 9.2, and Lemma 9.3.
Proof of Lemma 9.1. Observe that the second statement
of the lemma follows from the first. In-

deed, set G− (z) = F2 (z) − supw∈∂R F1 (z) − F2 (z) and G+ (z) = F2 (z) + supz∈∂R F1 (z) − F2 (z) ,
for each z ∈ R. Then, G− and G+ solve (1.2) on R (since F2 does), and G− ≤ F1 ≤ G+ on ∂R;
hence, G− ≤ F1 ≤ G+ on R, implying sup z∈R
F1 (z) − F2 (z) ≤ supz∈∂R F1 (z) − F2 (z) . Taking

(F1 , F2 ) = (F, 0) yields supz∈R F (z) ≤ supz∈∂R F (z) .
Now we establish the first statement of the lemma, which would follow from Lemma E.2, except
that (E.3) might not be valid for a uniform constant B > 1. To circumvent this, we restrict to a
subset of R. More precisely, assume to the contrary that F1 (z0 ) > F2 (z0 ) for some z0 ∈ R. Then,
102 AMOL AGGARWAL AND JIAOYANG HUANG

since F1 , F2 ∈ C(R) and F1 |∂R ≤ F2 |∂R , there exists some real number ε > 0 and open subset R′
with R′ ⊂ R, such that F1 (z) − ε ≤ F2 (z) for each z ∈ ∂R′ , and F1 (z) − ε > F2 (z) for each z ∈ R′ .
Observe that F1 − ε solves (1.2). Furthermore, since F1 , F2 ∈ C 2 (R) ∩ Adm(R), we have by
−1
compactness that ∂y Fj (z) + ∂y Fj (z) ≤ B is uniformly bounded by some constant B > 1, over
z ∈ R′ . It follows from the explicit form (1.3) of the coefficients dij that (E.3) holds (for a larger
value of B, given by π 2 B 4 ). Applying Lemma E.2 then indicates that F1 (z) − ε ≤ F2 (z) for all
z ∈ R′ . This is a contradiction for z = z0 , confirming the lemma. □

Proof of Lemma 9.2. For each i, j ∈ {t, y}, let e dij : R2 → R be smooth functions such that
−1
dij (u, v) = dij (u, v) whenever v ∈ [−ε , −ε], and such that (E.5) holds for some constant B
e e =

B(ε) > 1. Then, Lemma E.3 yields a constant C = C(ε, r, B, m, R ) > 1 such that there is a unique
e
solution Fe to the equation
X 
(E.6) dij ∇Fe(z) ∂i ∂j Fe(z) = 0,
e for z ∈ R, with Fe|∂R = f ,
i,j∈{t,y}

and that this Fe satisfies ∥Fe∥C m (Dr ) ≤ C. Observe since ∂x F (z) ∈ [−ε−1 , ε] for each z ∈ R (as
F ∈ Admε (R)); since F satisfies (1.2); and since dij (u, v) = e dij (u, v) for v ∈ [−ε−1 , ε], that F also
satisfies (E.6). Hence, F = Fe, and so ∥F ∥C m (Dr ) = ∥Fe∥C m (Dr ) ≤ C. □

We can also establish Lemma 9.3 as a quick consequence of results from [30].

Proof of Lemma 9.3. This is a special case of [30, Theorem 1]; let us explain how to match our
setup to the one there. First some point z0 ∈ R; it suffices to show that F is real analytic in some
neighborhood of z0 . We may assume by translation in what follows that z0 = 0; let r > 0 be a real
number such that B4r ⊆ R. By Lemma 9.2, there exists a constant C > 1 such that ∥F ∥C 2 (B2r ) ≤ C
and F ∈ C 3 (B2r ).
Now, define the function Φ : R3 → R by setting Φ(v1 , v2 , v3 ) = v2 + π 2 v1−4 v3 ; also define the
set E = (v1 , v2 , v3 ) ∈ R3 : ε ≤ −v1 ≤ ε−1 , |v2 | ≤ C, |v3 | ≤ C . Then we can reformulate the


equation (1.2) as Φ(∂y F, ∂tt F, ∂yy F ) = 0; since F ∈ Admε (R) and ∥F ∥C 2 (B2r ) ≤ C, we also have
∂y F (z), ∂tt F (z), ∂yy F (z) ∈ E for any z ∈ B2r . On E, Φ is real analytic; in particular, for any
point (v1 , v2 , v3 ) ∈ E and integers k1 , k2 , k3 ≥ 0, we have
∂v 1 ∂v 2 ∂v 3 Φ(v1 , v2 , v3 ) ≤ π 2 Cε−k−4 (k + 3)!,
k k k
1 2 3

where we have denoted k = k1 + k2 + k3 ≥ 0. Together with the fact that F ∈ C 3 (B2r ), this verifies
the assumptions in [30, Theorem 1] (in particular, see [30, Equation (5)]), with the Mk there equal
to k! here. It follows that there exists some constant A > 1 such that, for any point (t, y) ∈ Br and
integers k1 , k2 ≥ 0, we have
k k
∂t 1 ∂y 2 F (t, y) ≤ Ak1 +k+2 (k1 + k2 )! ≤ A2 · (2A)k1 +k2 k1 !k2 !,

(E.7)
and hence F is real analytic on Br . □

E.2. Proof of Lemma 9.4. In this section we establish Lemma 9.4. To that end, we show
the following proposition (which quickly implies Lemma 9.4) bounding the derivatives of F1 − F2
assuming F1 and F2 satisfy the same nonlinear, elliptic partial differential equation with close
boundary data. As in Lemma 9.4, we state its interior and boundary forms simultaneously; its
proof will be similar to that of [2, Proposition A.5], which provides its interior variant (at m = 1).
LOCAL STATISTICS AND CONCENTRATION FOR NON-INTERSECTING BROWNIAN BRIDGES 103

Lemma E.4. For any integer m ≥ 1; real numbers r > 0 and B > 1; and bounded open subsets
R′ ⊆ R ⊂ R2 such that R′ has C m+5 boundary, there exists a constants C = (B, m, r, R′ ) > 1
such that the following holds. For each j, k ∈ {1, 2}, let ajk ∈ C m+4 (R2 ) denote functions satisfying
axy = ayx and
X
(E.8) max ∥ajk ∥C m+4 (R2 ) ≤ B; inf2 ajk (z)ξj ξk ≥ B −1 |ξ|2 ,
j,k∈{1,2} z∈R
j,k∈{1,2}
2
x , ξy ) ∈ R . For each i ∈ {1, 2}, let fi ∈ C(∂R) denote a continuous function such
for any ξ = (ξ
that supz∈∂R fi (z) ≤ B, and let Fi ∈ C m+4 (R) ∩ C(R) denote the solution to
X 
(E.9) ajk ∇Fi (x) ∂j ∂k Fi (z) = 0, for each z ∈ R,
j,k∈{1,2}

with boundary data Fi |∂R = fi . Assume that there exist functions φi ∈ C m+5 (R) such that
∥φi ∥C m+5 (R) < B and Fi (z) = φi (z) for each z ∈ ∂R′ ∩ ∂R. Then, letting Dr ⊆ R be as in
(9.1), we have

(E.10) ∥F1 − F2 ∥C m (Dr ) ≤ C ∥F1 − F2 ∥C 0 (R) + ∥φ1 − φ2 ∥C m+2 (R) .
Proof of Lemma 9.4 (Outline). This follows from Lemma E.4, very similarly to how Lemma 9.2
followed from Lemma E.3; we omit further details. □
To establish Lemma E.4, we will use the following lemma, which is an interpolation estimate that
bounds derivatives of a function in terms of its higher (and lower) derivatives; its proof is similar
to that of [33, Lemma 6.35].
Lemma E.5. For any integers k > j > 0 and bounded open subset R ⊂ R2 with C k+1 boundary,
there exists a constant C = C(k, R) > 1 such that the following holds. Let F ∈ C k (R) denote a
function, and suppose that 0 < A ≤ B are real numbers such that A ≥ ∥F ∥0 and B ≥ [F ]k;0;R .
Then,
(E.11) [F ]j,0;R ≤ CA1−j/k B j/k .
Proof. Since the proof of this lemma is similar to that of [33, Lemma 6.35], we only establish it in
the case (j, k) = (1, 2); the proof for general k > j ≥ 1 is entirely analogous.
Fix some constant C0 = C0 (R) > 1 to be determined later,
and assume to the contrary that
there exists some i ∈ {1, 2} and z0 ∈ R such that ∂i F (z0 ) > C0 (AB)1/2 . Assume that i = x
and ∂i F (z0 ) > C0 (AB)1/2 , as the proofs in the alternative cases are entirely analogous. Further
set w = (1, 0) ∈ R2 and w′ = (2−1/2 , 2−1/2 ) ∈ R2 . Since the boundary of R is C k+1 , there exists
a constant t0 = t0 (R) > 0 such that either z0 + tw ∈ R for each t ∈ [0, t0 ]; z0 − tw ∈ R for each
t ∈ [0, t0 ]; z0 + tw′ ∈ R for each t ∈ [0, t0 ]; or z0 − tw′ ∈ R for each t ∈ [0, t0 ]. Let us assume that
the remaining ones are very similar. Then, since [F ]2,0;R ≤ B,
the third case holds, as the proofs in
we have ∂x F (z0 + tw′ ) − ∂x F (z0 ) ≤ Bt for each t ∈ [0, t0 ].
Thus, ∂x F (z0 + tw′ ) > 2−1/2 C0 (AB)1/2 − Bt whenever t ∈ [0, t0 ], and so integration over t yields
F (z0 )−F (z0 +tw′ ) > 2−1/2 C0 (AB)1/2 t− Bt2 , for any t ∈ [0, t0 ]. Selecting t = A 1/2 t0 ≤ t0 yields

2 B
t2 
∥F (z0 ) − F (z0 + t0 w′ ) > 2−1/2 C0 t0 − 0 A. In particular, selecting C0 sufficiently large so that
2
2−1/2 C0 t0 − t20 > 3, this implies that ∥F ∥0;R > A, which is a contradiction, verifying (E.11). □
Now we can establish Lemma E.4.
104 AMOL AGGARWAL AND JIAOYANG HUANG

Proof of Lemma E.4. The proof of this proposition is similar to that of [2, Proposition A.5], ob-
tained by replacing the interior regularity, Schauder, and interpolation estimates (given by [2,
Lemma A.3], [2, Lemma A.1], and [2, Lemma A.6], respectively), by the boundary ones (given by
Lemma E.3, Lemma E.1, and Lemma E.5, respectively). We only address the case m = 1 here
(so that it suffices to bound |∇F1 − ∇F2 |), as bounding the difference of the higher derivatives
follows similarly by differentiating the equations below. Throughout, we assume that R′ = R for
notational simplicity; the proof in the case of more general R′ is entirely analogous.
Denoting
n o
(E.12) ς = max ∥F1 − F2 ∥C 0 (R) , ∥φ1 − φ2 ∥C 3 (R)

we have

(E.13) ∥F1 − F2 ∥C 0 (R) ≤ ς; max sup Fi (z) ≤ B,
i∈{1,2} z∈R

where the second bound follows from Lemma E.2 and the fact that fi (z) ≤ B, for each i ∈ {1, 2}
and z ∈ ∂R. Moreover, the m = 5 case of Lemma E.3 yields a constant C1 = C1 (B, R) > 1 such
that
(E.14) max ∥Fi ∥C 5 (R) < C1 .
i∈{1,2}

Thus (E.11) (with the F there equal to F1 − F2 here; the A there equal to ς here; the B there equal
to C1 here; and the (j, k) there equal to either (1, 5) or (2, 5)), together with (E.13) and (E.14),
yields a constant C2 = C2 (B, R) > 1 such that
sup ∇F1 (z) − ∇F2 (z) ≤ C2 ς 4/5 ; ∥F1 − F2 ∥C 2 (R) ≤ C2 ς 3/5 .

(E.15)
z∈R

Thus, we may assume in what follows that ς < δ, for some sufficiently small constant δ =
δ(ε, B) > 0 (to be specified later), for otherwise the result follows from the fact that supz∈R ∇F1 (z)−
∇F2 (z) ≤ C2 ς 4/5 from (E.15). To improve the first bound in (E.15) to (E.10), define g : R → R
by setting
g(z) = ς −1/2 F2 (z) − F1 (z) ,

for each z ∈ R.
Then the first bound in (E.13) and (E.15) together yield
(E.16) ∥g∥C 0 (R) ≤ ς 1/2 ; ∥g∥C 1 (R) ≤ 2C2 ς 3/10 ; ∥g∥C 2 (R) ≤ C2 ς 1/10 . .

Next, setting i = 2 and using the fact that F2 = F1 + ς 1/2 g in (E.9), we obtain
X
ajk ∇F1 (z) + ς 1/2 ∇g(z) ∂j ∂k F1 (z) + ς 1/2 ∂j ∂k g(z) = 0.
 
(E.17)
1≤j,k≤2

Then, subtracting the i = 1 case of (E.9) from (E.17), we deduce that


X  
µjk (z)∂j ∂k g(z) + νjk (z)κjk (z) · ∇g(z)
j,k∈{1,2}
(E.18) X  
= −ς 1/2 ∇g(z) · κjk (z) ∂j ∂k g(z) + νjk (z)hjk (z) + ς 1/2 hjk (z)∂j ∂k g(z) ,


j,k∈{1,2}
LOCAL STATISTICS AND CONCENTRATION FOR NON-INTERSECTING BROWNIAN BRIDGES 105

where for each j, k ∈ {1, 2}, we have defined µjk , νjk , hjk ∈ C(R) and κjk : R → R2 by setting
 
µjk (z) = ajk ∇F1 (z) ; κjk (z) = ∇ajk ∇F1 (z) ; νjk (z) = ∂j ∂k F1 (z);
(E.19) −1
 
1/2
ajk ∇F1 (z) + ς ∇g(z) − ajk ∇F1 (z) − ς 1/2 ∇g(z) · ∇ajk ∇F1 (z) ,
 
hjk (z) = ς

for each z ∈ R. Now, observe that (E.1), (E.14), and the fact that ∥ajk ∥C 4 (R2 ) ≤ B together imply
the existence of a constant C3 = C3 (B, R) > 1 such that
(E.20) ∥µjk ∥C 0,1/2 (R) ≤ C3 ; ∥κjk (z)∥C 0,1/2 (R) ≤ C3 ; ∥νjk (z)∥C 0,1/2 (R) ≤ C3 .

Since the ajk satisfy (E.8), it follows from (E.18), (E.12), and the α = 21 case of the Schauder
estimate Lemma E.1 that there exists a constant C4 = C4 (B, R) > 1 such that

∥g∥C 2,1/2 (R) ≤ C4 ∥g∥C 0 (R) + C4 ς 1/2 max ∇g(z) · κjk (z) ∂j ∂k g(z) 0,1/2

j,k∈{1,2} C (R)
1/2 1/2

+ C4 ς max νjk (z)hjk (z) C 0,1/2 (R) + C4 ς
j,k∈{1,2}

+ C4 ς max hjk (z)∂j ∂k g(z) C 0,1/2 (R) .
j,k∈{1,2}

Thus, by (E.1) and (E.20), there exists a constant C5 = C5 (B, R) > 1 such that
 
∥g∥C 2,1/2 (R) ≤ C5 ς 1/2 ∥g∥C 2,1/2 (R) ∥g∥C 1 (R) + ∥g∥1,1/2;R ∥g∥C 2 (R) + ∥h∥C 0,1/2 (R) + 1
(E.21)  
+ C5 ς ∥h∥C 0 (R) ∥g∥C 2,1/2 (R) + ∥h∥C 0,1/2 (R) ∥g∥C 2 (R) + C5 ∥g∥C 0 (R) .

Now, (E.16) yields a constant C6 = C6 (B, R) > 1 such that



max ∥g∥C 1 (R) , ∥g∥1,1/2;R , ∥g∥C 2 (R) , ∥h∥C 0 (R) , ∥h∥C 0,1/2 (R)
(E.22) 
≤ max ∥g∥C 2 (R) , ∥h∥C 1 (R) < C6 ,
where the bound on ∥h∥C 1 (R) follows from differentiating the definition (E.19) of h; (E.14); (E.16);
the fact that ∥ajk ∥C 4 (R2 ) ≤ B for each j, k ∈ {1, 2}; and a Taylor expansion.
Inserting the first bound of (E.16); (E.22); and the fact that ∥g∥C 2,1/2 (R) ≥ ∥g∥C 2 (R) into (E.21)
then yields a constant C7 = C7 (B, R) > 1 such that

(1 − C7 ς 1/2 )∥g∥C 2,1/2 (R) ≤ C7 ς 1/2 .


1

Thus, if δ < min 4C72
,1 and ς < δ, we obtain that

(E.23) ∥F1 − F2 ∥C 1 (R) ≤ ∥F1 − F2 ∥C 2,1/2 (R) = ς 1/2 ∥g∥C 2,1/2 (R) ≤ 2C7 ς,
from which the proposition follows. □

E.3. Proof of Lemma 9.6. In this section we establish Lemma 9.6.

Proof of Lemma 9.6. We first perform a rescaling, to which end we define the open rectangles
b = L · R = (0, 1) × ( − L, L);
R b r = L · Sr = (r, 1 − r) × (r + 1 − L, L − r − 1);
S
b = L · S = (0, 1) × (1 − L, L − 1),
S
106 AMOL AGGARWAL AND JIAOYANG HUANG

as well as the functions Fb ∈ C m (R)


b and fbi , gbi : [−L, L] → R for each i ∈ {0, 1} by, for any
w ∈ [0, 1] × [−L, L] and x ∈ [−L, L], setting
(E.24)
Fb(w) = L · F (L−1 w); fbi (x) = L · fi (L−1 x); gbi (x) = L · gi (L−1 x), so ∇Fb(w) = ∇F (L−1 w),

which in particular implies that Fb ∈ Admε (R). b Further define the functions Hb ∈ C m (R)
b and
hi : [−L, L] → R for each i ∈ {0, 1} by, for any (t, x) ∈ [0, 1] × [−L, L], setting
b

(E.25) hi (x) = gbi (x) − Fb(i, x) = gbi (x) − fbi (x);


b b x) = Fb(t, x) + (1 − t)b
H(t, h0 (x) + tb
h1 (x),
b x) = gbi (x) for each i ∈ {0, 1} and x ∈ [−L, L]. The scalings (E.24) (together with the
so that H(i,
assumptions of the lemma) then imply for any index i ∈ {0, 1} and integer j ∈ J0, mK that
(E.26)
[Fb]j ≤ BL1−j ; gi ]j ≤ BL1−j ;
[b b j ≤ 3BL1−j ;
[H] ∥Fb − H∥
b 0 b ≤ ∥b
C (R) h0 ∥C 0 + ∥b
h1 ∥C 0 ≤ 2δ.
b ∈ Admε/2 (S)
It suffices to show the existence of a solution G b to the equation (1.2) such that

G(i,
b x) = gbi (x), for each i ∈ {1, 2} and x ∈ [1 − L, L − 1];

(E.27) ∥Fb − G∥
b k b ≤ C1 · ∥b
C (Sr ) h0 ∥C 0 + ∥b
h1 ∥C 0 ;

∥Fb − G∥
b m−5 b ≤ C2 · ∥b
C (S) h0 ∥C m−3 + ∥b h1 ∥C m−3 ,
for some constants C1 = C1 (ε, r, B, k) > 1 and C2 = C2 (ε, B, m) > 1. Then, the function G ∈
Admε/2 (S) defined by G(w) = L−1 ·G(Lw)b for each w ∈ [0, L−1 ]×[L−1 −1, 1−L−1 ] would satisfy the
conditions of the lemma (where here, we again used the fact that ∇G(w) = ∇G(Lw), b as in (E.24); to
deduce the last bound in the third part of the lemma, we also used ∥b hi ∥C m−3 ≤ L∥hi ∥C m−3 for each
i ∈ {0, 1}, ∥F − G∥C m−5 ≤ Lm−6 ∥Fb − G∥ b C m−5 , and the interpolation estimate Lemma E.5 to deduce
3/m
that ∥fi − gi ∥C m−3 ≤ C2 ∥fi − gi ∥C 0 for each i ∈ {0, 1} and some constant C2′ = C2′ (ε, B, m) > 1,

which holds since ∥fi − gi ∥C m ≤ ∥fi ∥C m + ∥gi ∥C m ≤ 2B).


 end, let us fix an open set U ⊂ (0, 1) × (−1, 1) with smooth boundary, such that
To that
(0, 1) × − 21 , 12 ⊂ U; see the left side of Figure 14. For any real number a ∈ R, we also denote its
vertical translate Ua = U+(0, a), and we define the domain [0, 1]×[1−L, L−1] ⊂ V ⊂ [0, 1]×[−L, L]
by

(E.28) V = [0, 1] × [1 − L, L − 1] ∪ U1−L ∪ UL−1 .
See the right side of Figure 14. As in the proof of Lemma 9.2, for any indices i, j∈ {t, x} we also
dij (u, v) = dij (u, v) whenever v ∈ − 2ε , − 2ε , and such that

dij ∈ C m+3 (R2 ) so that e
fix fuctions e
(E.5) holds for some constant B e = B(ε)
e > 1.
Then, Lemma E.3 yields a solution G b ∈ C m−1 (V) to the equation
X 
(E.29) dij ∇G(z)
e b ∂i ∂j G(z)
b = 0, for z ∈ V, with G|
b ∂V = H|
b ∂V .
i,j∈{t,x}

We first verify that Gb satisfies the three properties in (E.27), and then we confirm that Gb ∈
Admε/2 (V) ⊆ Admε/2 (S) b (where the last inclusion holds by restricting G,
b as S b ⊂ V) and that
G solves the original equation (1.2). To confirm the first condition in (E.27), observe from the
b
LOCAL STATISTICS AND CONCENTRATION FOR NON-INTERSECTING BROWNIAN BRIDGES 107

Figure 14. Shown on the left is the open set (0, 1)× − 12 , 21 ⊂ U ⊂ (0, 1)×(−1, 1).
 

Shown on the right is the open set V as in (E.28).

definition (E.28) of V that {0, 1} × [1 − L, L − 1] ⊂ ∂V, so for any (i, x) ∈ {0, 1} × [1 − L, L − 1] we


have G(i,
b x) = H(i,
b x) = gbi (x).
The second and third statements will follow from a suitable application of Lemma E.4. To
implement this, first observe that
(E.30)

sup Fb(w) − G(w)
b ≤ sup Fb(w) − G(w)
b = sup Fb(w) − H(w)
b ≤ ∥b
h0 ∥C 0 + ∥b
h1 ∥C 0 ≤ 2δ.
w∈V w∈∂V w∈∂V

Here, the first inequality follows from the comparison principle Lemma E.2 (applied twice, as in the
proof of Lemma 9.1, first with (F1 , F2 ) there equal to (Fb, G b − supw∈∂V Fb(w) − G(w)
b ) and next

with (F1 , F2 ) there equal to (F , G + supw∈∂V F (w) − G(w) ) here); the second follows from the
b b b b
boundary data (E.29) for G; b and the third and fourth follow from the last bounds in (E.26).
Now, fix some w0 = (t0 , x0 ) ∈ Sb r so that Br (t0 , x0 ) ⊆ V, and apply Lemma E.4 with the (F1 , F2 )
e k; r ; Ux here;
 
there equal to F − F (0, x0 ); G − Fb(0, x0 ) here; (B, m, r; R) there equal to 8B + B,
b b b
2 0
and R′ there empty here (so we may take the (φ1 , φ2 ) there to be 0 here). This yields a constant
C1 = C1 (ε, r, B, k) > 1 (recalling that B e only depends on ε) such that
 
Fb − G∥
b C k (B (w )) ≤ C1 · sup Fb(w) − G(w) b ≤ C1 ∥b
h0 ∥C 0 + ∥b
h1 ∥C 0 ,
r/2 0
w∈∂Ux0

where the last bound follows from (E.30). Ranging over w0 ∈ S b r gives the second part of (E.27).
To establish the third part of (E.27), fix some w0 ∈ (t0 , x0 ) ∈ V, and let s ∈ [1 − L, L − 1] be
such that Br/4 (w0 ) ⊂ Us (which is guaranteed to exist by (E.28)). Then, apply Lemma E.4 with
b − Fb(0, x0 ) here; (B, m, r; R, R′ ) there equal to 8B + B,

the (F1 , F2 ) equal to Fb − Fb(0, x0 ), G e m−
108 AMOL AGGARWAL AND JIAOYANG HUANG

5, 8r ; Us , Us here; and φ1 (t, x) and φ2 (t, x) there equal to Fb(t, x) − Fb(0, x0 ) and H(t,

b x) − Fb(0, x0 )
here, respectively. This yields a constant C2 = C2 (ε, r, B, m) > 1 (which also depends on the open
set U, but we omit this dependence since U is fixed) such that
   

Fb − G b m−5 ≤ C 2 sup F
b (w) − G(w)
b + ∥Fb − H∥
b C m−3 ≤ 2C2 ∥h
b 0 ∥ C m−3 + ∥h1 ∥C m−3 ,
b
C (U )
s
w∈∂Us

where in the last bound we used (E.25) and (E.30). Ranging over all w ∈ V yields
 
(E.31) Fb − G
b m−5 ≤ 2C2 ∥h
b ∥
0 C m−3 + ∥h
b ∥
1 C m−3 ,
C (V)

which since S ⊂ V implies the third part of (E.27).


It remains to verify that G b ∈ Admε/2 (V) ⊆ Admε/2 (S) b and that G b satisfies (1.2). Observe
that it suffices to prove the former, as it implies the latter since G b satisfies (E.29), and edij (∇G)
b =
dij (∇G) on V if G ∈ Admε/2 (V). To show the former, observe from (E.31), (E.30), and the
b b
((F ; j, k; R) = (Fb − G;
b 1, 2; Us ) case of the) interpolation estimate Lemma E.5 that there exists a
constant C3 = C3 (ε, B, m) > 1 (also implicitly dependent on U) such that the following holds. For
sufficiently small δ = δ(ε, B, m) > 0 we have
≤ C3 δ 1/2 .

sup ∇Fb(w) − ∇G(w)
b
w∈V

Together with the fact that Fb ∈ Admε (R),


b it follows (after decreasing δ again if necessary) that
b ∈ Admε/2 (V), thereby establishing the lemma.
G □

References
[1] M. Adler, P. L. Ferrari, and P. van Moerbeke. Airy processes with wanderers and new universality classes. Ann.
Probab., 38(2):714–769, 2010.
[2] A. Aggarwal. Universality for lozenge tiling local statistics. To appear in Ann. Math., preprint, arXiv:1907.09991.
[3] A. Aggarwal and V. Gorin. Gaussian unitary ensemble in random lozenge tilings. Probab. Theory Related Fields,
184(3-4):1139–1166, 2022.
[4] A. Aggarwal and J. Huang. In preparation.
[5] L. Ambrosio, A. Carlotto, and A. Massaccesi. Lectures on elliptic partial differential equations, volume 18 of
Appunti. Scuola Normale Superiore di Pisa (Nuova Serie) [Lecture Notes. Scuola Normale Superiore di Pisa
(New Series)]. Edizioni della Normale, Pisa, 2018.
[6] G. W. Anderson, A. Guionnet, and O. Zeitouni. An introduction to random matrices, volume 118 of Cambridge
Studies in Advanced Mathematics. Cambridge University Press, Cambridge, 2010.
[7] S. Belinschi, A. Guionnet, and J. Huang. Large deviation principles via spherical integrals. Probability and
Mathematical Physics, 3(3):543–625, 2022.
[8] P. Biane. On the free convolution with a semi-circular distribution. Indiana University Mathematics Journal,
pages 705–718, 1997.
[9] P. M. Bleher and A. B. J. Kuijlaars. Large n limit of Gaussian random matrices with external source. III. Double
scaling limit. Comm. Math. Phys., 270(2):481–517, 2007.
[10] P. Bourgade, K. Mody, and M. Pain. Optimal local law and central limit theorem for β-ensembles. Comm. Math.
Phys., 390(3):1017–1079, 2022.
[11] E. Brézin and S. Hikami. Level spacing of random matrices in an external source. Phys. Rev. E (3), 58(6):7176–
7185, 1998.
[12] E. Brézin and S. Hikami. Universal singularity at the closure of a gap in a random matrix theory. Phys. Rev. E
(3), 57(4):4140–4149, 1998.
[13] D. L. Burkholder. Distribution function inequalities for martingales. Ann. Probability, 1:19–42, 1973.
[14] Y. S. Chow and H. Teicher. Probability theory. Springer Texts in Statistics. Springer-Verlag, New York, second
edition, 1988. Independence, interchangeability, martingales.
LOCAL STATISTICS AND CONCENTRATION FOR NON-INTERSECTING BROWNIAN BRIDGES 109

[15] T. Claeys, T. Neuschel, and M. Venker. Boundaries of sine kernel universality for Gaussian perturbations of
Hermitian matrices. Random Matrices Theory Appl., 8(3):1950011, 50, 2019.
[16] H. Cohn, N. Elkies, and J. Propp. Local statistics for random domino tilings of the Aztec diamond. Duke Math.
J., 85(1):117–166, 1996.
[17] H. Cohn, R. Kenyon, and J. Propp. A variational principle for domino tilings. J. Amer. Math. Soc., 14(2):297–
346, 2001.
[18] I. Corwin and A. Hammond. Brownian Gibbs property for Airy line ensembles. Invent. Math., 195(2):441–508,
2014.
[19] D. Dauvergne and B. Virág. Bulk properties of the Airy line ensemble. Ann. Probab., 49(4):1738–1777, 2021.
[20] D. S. Dean, P. Le Doussal, S. N. Majumdar, and G. Schehr. Noninteracting fermions in a trap and random
matrix theory. J. Phys. A, 52(14):144006, 32, 2019.
[21] S. Delvaux, A. B. J. Kuijlaars, and L. Zhang. Critical behavior of nonintersecting Brownian motions at a tacnode.
Comm. Pure Appl. Math., 64(10):1305–1383, 2011.
[22] F. J. Dyson. A Brownian-motion model for the eigenvalues of a random matrix. J. Mathematical Phys., 3:1191–
1198, 1962.
[23] L. Erdős, S. Péché, J. A. Ramı́rez, B. Schlein, and H.-T. Yau. Bulk universality for Wigner matrices. Comm.
Pure Appl. Math., 63(7):895–925, 2010.
[24] L. Erdős and K. Schnelli. Universality for random matrix flows with time-dependent density. Ann. Inst. Henri
Poincaré Probab. Stat., 53(4):1606–1656, 2017.
[25] L. Erdős and H.-T. Yau. A dynamical approach to random matrix theory, volume 28. American Mathematical
Soc., 2017.
[26] P. L. Ferrari and H. Spohn. Random growth models. In The Oxford Handbook of Random Matrix Theory. Oxford
University Press, 09 2015.
[27] P. L. Ferrari and B. Vető. Non-colliding Brownian bridges and the asymmetric tacnode process. Electron. J.
Probab., 17:no. 44, 17, 2012.
[28] P. L. Ferrari and B. Vető. The hard-edge tacnode process for Brownian motion. Electron. J. Probab., 22:Paper
No. 79, 32, 2017.
[29] M. E. Fisher. Walks, walls, wetting, and melting. J. Statist. Phys., 34(5-6):667–729, 1984.
[30] A. Friedman. On the regularity of the solutions of nonlinear elliptic and parabolic systems of partial differential
equations. J. Math. Mech., pages 43–59, 1958.
[31] M. Gaudin. Sur la loi limite de l’espacement des valeurs propres d’une matrice aléatoire. Nuclear Physics,
25:447–458, 1961.
[32] P.-G. d. Gennes. Soluble model for fibrous structures with steric constraints. The Journal of Chemical Physics,
48(5):2257–2259, 1968.
[33] D. Gilbarg and N. S. Trudinger. Elliptic partial differential equations of second order. Classics in Mathematics.
Springer-Verlag, Berlin, 2001. Reprint of the 1998 edition.
[34] D. J. Grabiner. Brownian motion in a Weyl chamber, non-colliding particles, and random matrices. Ann. Inst.
H. Poincaré Probab. Statist., 35(2):177–204, 1999.
[35] A. Guionnet. First order asymptotics of matrix integrals; a rigorous approach towards the understanding of
matrix models. Comm. Math. Phys., 244(3):527–569, 2004.
[36] A. Guionnet and O. Zeitouni. Large deviations asymptotics for spherical integrals. J. Funct. Anal., 188(2):461–
515, 2002.
[37] M. Hairer and J. C. Mattingly. Yet another look at Harris’ ergodic theorem for Markov chains. In Seminar
on Stochastic Analysis, Random Fields and Applications VI, volume 63 of Progr. Probab., pages 109–117.
Birkhäuser/Springer Basel AG, Basel, 2011.
[38] T. E. Harris. The existence of stationary measures for certain Markov processes. In Proceedings of the Third
Berkeley Symposium on Mathematical Statistics and Probability, 1954–1955, vol. II, pages 113–124. University
of California Press, Berkeley-Los Angeles, Calif., 1956.
[39] J. M. Holte. Discrete Gronwall lemma and its applications. MAA-NCS Meeting at the University of North
Dakota, Grand Forks, 2009.
[40] J. Huang. Edge universality for nonintersecting Brownian bridges. Preprint, arXiv:2011.01752.
[41] J. Huang and B. Landon. Rigidity and a mesoscopic central limit theorem for Dyson Brownian motion for general
β and potentials. Probab. Theory Related Fields, 175(1-2):209–253, 2019.
[42] J. Huang, B. Landon, and H.-T. Yau. Bulk universality of sparse random matrices. J. Math. Phys., 56(12):123301,
19, 2015.
110 AMOL AGGARWAL AND JIAOYANG HUANG

[43] D. A. Huse and M. E. Fisher. Commensurate melting, domain walls, and dislocations. Physical Review B,
29(1):239, 1984.
[44] K. Johansson. Universality of the local spacing distribution in certain ensembles of Hermitian Wigner matrices.
Comm. Math. Phys., 215(3):683–705, 2001.
[45] K. Johansson. Discrete polynuclear growth and determinantal processes. Comm. Math. Phys., 242(1-2):277–329,
2003.
[46] K. Johansson. Random matrices and determinantal processes. In Mathematical statistical physics, pages 1–55.
Elsevier B. V., Amsterdam, 2006.
[47] K. Johansson. Non-colliding Brownian motions and the extended tacnode process. Comm. Math. Phys.,
319(1):231–267, 2013.
[48] I. Karatzas and S. E. Shreve. Brownian motion and stochastic calculus, volume 113 of Graduate Texts in
Mathematics. Springer-Verlag, New York, second edition, 1991.
[49] B. Landon, P. Sosoe, and H.-T. Yau. Fixed energy universality of Dyson Brownian motion. Adv. Math., 346:1137–
1332, 2019.
[50] B. Landon and H.-T. Yau. Convergence of local statistics of Dyson Brownian motion. Comm. Math. Phys.,
355(3):949–1000, 2017.
[51] B. Laslier and F. L. Toninelli. The mixing time of lozenge tiling Glauber dynamics. Preprint, arXiv:2207.01444.
[52] B. Laslier and F. L. Toninelli. Lozenge tilings, Glauber dynamics and macroscopic shape. Comm. Math. Phys.,
338(3):1287–1326, 2015.
[53] K. Liechty and D. Wang. Nonintersecting Brownian bridges between reflecting or absorbing walls. Adv. Math.,
309:155–208, 2017.
[54] I. M. Lifshitz. Kinetics of ordering during second-order phase transitions. Sov. Phys. JETP, 15:939, 1962.
[55] A. Matytsin. On the large-N limit of the Itzykson-Zuber integral. Nuclear Phys. B, 411(2-3):805–820, 1994.
[56] M. L. Mehta and M. Gaudin. On the density of eigenvalues of a random matrix. Nuclear Phys., pages 420–427,
1960.
[57] S. Meyn and R. L. Tweedie. Markov chains and stochastic stability. Cambridge University Press, Cambridge,
second edition, 2009. With a prologue by Peter W. Glynn.
[58] J. A. Mingo and R. Speicher. Free probability and random matrices, volume 35 of Fields Institute Monographs.
Springer, New York; Fields Institute for Research in Mathematical Sciences, Toronto, ON, 2017.
[59] M. Prähofer and H. Spohn. Scale invariance of the PNG droplet and the Airy process. volume 108, pages
1071–1106. 2002. Dedicated to David Ruelle and Yasha Sinai on the occasion of their 65th birthdays.
[60] H. Spohn. Kardar–Parisi–Zhang equation in one dimension and line ensembles. Pramana, 64:847–857, 2005.
[61] T. Tao and V. Vu. Random matrices: sharp concentration of eigenvalues. Random Matrices Theory Appl.,
2(3):1350007, 31, 2013.
[62] C. A. Tracy and H. Widom. The Pearcey process. Comm. Math. Phys., 263(2):381–400, 2006.
[63] C. A. Tracy and H. Widom. Nonintersecting Brownian excursions. Ann. Appl. Probab., 17(3):953–979, 2007.
[64] D. Voiculescu. Addition of certain noncommuting random variables. J. Funct. Anal., 66(3):323–346, 1986.

You might also like