Journal of Functional Analysis 270 (2016) 1269–1298

Contents lists available at ScienceDirect

Journal of Functional Analysis

www.elsevier.com/locate/jfa

Asymptotics for the heat kernel in multicone

domains
Pierre Collet a , Mauricio Duarte b,∗,1 , Servet Martínez c,2 ,
Arturo Prat-Waldron d , Jaime San Martín c,2
a
CNRS Physique Théorique, Ecole Polytechnique, 91128 Palaiseau cedex, France
b
Departamento de Matematica, Universidad Andres Bello, Republica 220, Piso 2,
Santiago, Chile
c
Departamento de Ingeniería Matemática, Facultad de Ciencias Físicas y
Matemáticas, Universidad de Chile, Beauchef 851, Torre norte, Piso 5, Santiago,
Chile
d
Max Planck Institute for Mathematics, Vivatsgasse 7, 53111 Bonn, Germany

a r t i c l e i n f o a b s t r a c t

Article history: A multicone domain Ω ⊆ Rn is an open, connected set

Received 11 May 2015 that resembles a finite collection of cones far away from the
Accepted 17 October 2015 origin. We study the rate of decay in time of the heat kernel
Available online 3 December 2015
p(t, x, y) of a Brownian motion killed upon exiting Ω, using
Communicated by L. Saloff-Coste
both probabilistic and analytical techniques. We find that the
MSC: decay is polynomial and we characterize limt→∞ t1+α p(t, x, y)
primary 60J65, 35K08, 35B40 in terms of the Martin boundary of Ω at infinity, where α > 0
secondary 60H30 depends on the geometry of Ω. We next derive an analogous
result for tκ/2 Px (T > t), with κ = 1 + α − n/2, where T is the
Keywords: exit time from Ω. Lastly, we deduce the renormalized Yaglom
Heat kernel limit for the process conditioned on survival.
Brownian motion © 2015 Elsevier Inc. All rights reserved.
Yaglom limit
Martin boundary

* Corresponding author.
E-mail addresses: [email protected] (P. Collet), [email protected]
(M. Duarte), [email protected] (S. Martínez), [email protected] (A. Prat-Waldron),
[email protected] (J. San Martín).
1
M.D. thanks the support from project FONDECYT 3130724, and the Programa Iniciativa Cientifica
Milenio grant number NC130062 through the Nucleus Millenium Stochastic Models of Complex and
Disordered Systems.

http://dx.doi.org/10.1016/j.jfa.2015.10.021
0022-1236/© 2015 Elsevier Inc. All rights reserved.
1270 P. Collet et al. / Journal of Functional Analysis 270 (2016) 1269–1298

1. Introduction

Let O be a domain (open and connected set) in Rn , regular for the Dirichlet problem.
Consider an n-dimensional Brownian motion Bt starting from the interior of O, with
exit time T O . The heat kernel pO (t, x, y) is the Radon–Nikodym derivative of the Borel
measure A → Px (Bt ∈ A, T O > t) with respect to the n-dimensional Lebesgue measure,
and it is characterized to be the fundamental solution of the heat equation with Dirichlet
boundary condition, that is: as a function of (t, y) it solves the heat equation ∂t u = 12 Δu,
it vanishes continuously on ∂O, and it satisfies the initial condition u(0, y) = δx (y).
It is well known that pO (t, x, y) tends to zero as time grows to infinity. A classical
problem is to find the exact asymptotic (in time) for the decay of the heat kernel and
the survival probability. This is well understood for bounded domains (see [10] and
[11]). For results in some planar domains we refer the reader to [2]. The large time
asymptotic problem is treated in [9] for a large class of (non-symmetric) diffusions under
some integrability conditions on the ground state. Exact asymptotic are computed for
Benedicks domains in [4], and for exterior domains in [5]. Our work focuses on finding
the exact asymptotic in time for pΩ (t, x, y) and Px (T Ω > t) for a multicone domain Ω,
which we define next.
Let Sn−1 = {x ∈ Rn : |x| = 1} be the unit sphere in Rn . Points in Rn will be regarded
as x = rθ, where r = |x| and θ ∈ Sn−1 . Given a Lipschitz, proper subdomain D of Sn−1 ,
and a vector a ∈ Rn , a truncated cone with opening D and vertex a is the set

C(a, D, R) = {a + x : x = rθ ∈ Rn : r > R, θ ∈ D} ,

where R ≥ 0. When R > 0, the set S = a + RD will be called the base of the truncated
cone. When R = 0, we will refer to the set in the previous display as cone with vertex a.
In the same context as above, given a base S = a +RD, let 0 < λ1 < λ2 ≤ λ3 ≤ · · · be
the eigenvalues of the Laplace–Beltrami operator on D, with corresponding orthonormal
basis {m1 , m2 , m3 , . . .} of L2 (D, σ), where σ is the surface measure on Sn−1 . Let αi =
 i 1/2
λ + ( n2 − 1)2 . We define the character of the base S as the number α = α(D) = α1 .
The character of the truncated cone C(a, D, R) is also defined as α.
A multicone domain Ω ⊆ Rn is a connected, open set such that there exist a bounded
domain Ω0 ⊆ Ω and finitely many truncated cones Ωj = C(aj , Dj , Rj ), with j = 1, . . . N ,
such that Ωj ∩ Ωi = ∅ for 1 ≤ j < i ≤ N , and


N
Ω \ Ω0 = Ωj .
j=1

Here Ω0 is the closure of the set Ω0 . The set Ω0 will be called the core, and for j ≥ 1,
the sets Ωj are called branches of the multi-cone set. Notice that by construction, the

2
S.M. and J.S.M thank the Center for Mathematical Modeling (CMM) Basal CONICYT Program PFB 03.
 P. Collet et al. / Journal of Functional Analysis 270 (2016) 1269–1298 1271

branches are disjoint from the core. Also, we will denote the base of the truncated cone Ωj
by Sj . Without loss of generality, we can assume that Rj = 1, which makes the exposition
that follows much easier. The character of the truncated cone Ωj will be denoted by αj .
We define the character of the multicone Ω as the number α = min {αj : j = 1, . . . , N }.
An index l such that αl = α will be called maximal. We denote by M the set of maximal
indices.
To state the main results of this article, we need to introduce the Martin boundary
at infinity for Ω.
It is well known that there is a unique minimal harmonic function v on a cone with
vertex C0 = C(a, D, 0) that vanishes continuously on ∂C0 . Actually, there is only one pos-
itive harmonic function in C0 that vanishes continuously on its boundary (Theorem 1.1
in [1]). For x = a + |x − a| θ ∈ C0 this function is given by:
 
α− n2 −1
v(x) = |x − a| m1 (θ), (1.1)

where α is the character of D and m1 is the first eigenfunction of the Laplace–Beltrami

operator on D. Notice how we have chosen to normalize v in terms of the normalization
of m1 in L2 (D, σ). In order to simplify our exposition, we set κ = 1 + α − n/2, so that
κ
v(x) = |x − α| m1 (θ). We call vj to the function (1.1) corresponding to the cone with
vertex C(aj , Dj , 0).
Similarly, if C = C(a, D, R) is a truncated cone, there is a unique (minimal) positive
harmonic function w in C that vanishes continuously on ∂C, which is defined as follows:
let T C be the exit time of a Brownian motion Bt from the cone C. Then

w(x) = v(x) − Ex (v(BT C )), x ∈ C. (1.2)

Let wj be the unique minimal harmonic function in Ωj . By a standard balayage

argument [7], one can extend wj to a minimal harmonic function in Ω. Such extension
is given by

1
uj (x) = wj (x)1Ωj (x) + G(x, y)∂n wj (y)σj (dy), x ∈ Ω, (1.3)
2
Sj

where ∂n denotes the (inward) normal derivative on Sj , and σj is the translation of σ

by aj , and G is the Green function of the domain Ω:
∞
G(x, y) = p(t, x, y)dt.
0

Reciprocally, we have that

wj (x) = uj (x) − Ex (u(BT j )) , x ∈ Ωj , (1.4)

where B is an n-dimensional Brownian motion, stopped at its exit time T j from Ωj .

1272 P. Collet et al. / Journal of Functional Analysis 270 (2016) 1269–1298

It is direct to verify from the last two equations that the function uj is bounded in
Ω \ Ωj , and satisfies that for x = aj + rθ,

uj (aj + rθ)
lim = 1, (1.5)
r→∞ wj (aj + rθ)

for fixed θ ∈ Dj .
We are ready to state the main results of this paper.

Theorem 1.1. Let Ω be a multicone domain with branches Ω1 , . . . , ΩN . Let α > 0 be the
character of Ω, and let M be the set of maximal indices. Then,

1 
lim t1+α p(t, x, y) = ul (x)ul (y). (1.6)
t→∞ 2α Γ(1 + α)
l∈M

The limit is in the topology of uniform convergence on compact sets.

Theorem 1.2. Let Ω be a multicone domain with branches Ω1 , . . . , ΩN . Let α > 0 be the
character of Ω, and let M be the set of maximal indices. Set κ = 1 + α − n/2. Then

 κ+n  ⎛ ⎞
Γ  
lim tκ/2 Px (T > t) = κ/2  2 n  ⎝ m1l (θ)σ(dθ)⎠ ul (x). (1.7)
t→∞ 2 Γ κ + 2 l∈M
Dl

The limit is in the topology of uniform convergence on compact sets.

Theorem 1.3. Let Ω be a multicone domain with character α > 0, and set β = 1 +α+n/2.
√
Fix x ∈ Ω, and 1 ≤ j ≤ N . For each y = |y| θ, with θ ∈ Dj , we have that aj + ty ∈ Ωj
for large enough values of t, and

√ uj (x)vj (y) −|y|2 /2

lim tβ/2 p(t, x, aj + ty) = 1M (j) e . (1.8)
t→∞ 2α Γ(1 + α)

The limit is in the sense of uniform convergence on compact sets on the variables x and y.

In (1.6), the variables x and y are interchangeable. This differs from (1.8), where
the variables x and y have very different roles: on the left hand side, x ∈ Ω is fixed, but
√
arbitrary, whereas y needs to be in the cone with vertex C(0, Dj , 0) so that aj + ty ∈ Ωj
makes sense for large values of t. This difference is reflected on the right hand side of (1.8),
2
where we find uj (x) and vj (y), plus an exponential decay in |y| .
The paper is organized as follows. Section 2 lists some key results that we take from the
literature on heat kernels for killed diffusions, in particular, subsection 2.1 includes our
main theorems for the case of a cone with vertex. Section 3 deals with the asymptotics
for truncated cones, and Section 4 includes some lemmas leading up to the proofs of
the main theorems, which are contained at the end of Section 4 for the decay of the
 P. Collet et al. / Journal of Functional Analysis 270 (2016) 1269–1298 1273

heat kernel, and in Section 5 for the decay of the survival probability. Finally, Section 6
includes the proof of Theorem 1.3 and discusses a renormalized Yaglom limit for the
killed Brownian motion.

2. Preliminary results

In what follows we make the following simplifications, in order to keep the exposition
clear. We set T = T Ω , T j = T Ωj and denote by p and pj the respective heat kernels. In
some of the formulas below, integrals over Sj are understood to be with respect to the
translated measure σ j , but we will omit the index since the dependence on j is clear from
the domain of integration. Also, we will abuse the notation by omitting the vector aj
form all the formulas involving functions in cones, since its inclusion affects all such
functions by a simple translation of coordinates. In particular, we will write pj (t, x, y)
for x = |x| θ, y = |y| η for θ, η ∈ Dj instead of pj (t, x + aj , y + aj ) in order to simplify
our exposition. In this spirit, we will often say that x → ∞ radially in Ωj to mean that
x = aj + rθ, and r → ∞.
We start by listing some general properties of heat kernels in unbounded domains.

Lemma 2.1. (Lemma 2.1 in [5]) Let O be a regular domain for the Dirichlet problem.
Let u(t, x) be a positive solution of the heat equation in R+ × O, and consider a function
a : R+ → R+ such that

a(t + s)
sup < ∞, (2.1)
t≥t0 ,|s|≤2 a(t)

for some t0 > 0. Further, assume that the family of functions {a(t)u(t, ·) : t ≥ t0 } is
bounded on compact sets. Then, the family {a(t)u(t, ·) : t ≥ t0 + 1} is equicontinuous on
compact sets of O.

The next lemma corresponds to Lemmas 2.1–2.4 in [4], which are proved for Benedicks
domains in Rn . Nonetheless, the proofs work in a much more general setting, as long as
the domain O is a regular domain for the Dirichlet problem, with infinite interior radius.

Lemma 2.2. (Lemmas 2.1–2.4 in [4]) In the same setting of Lemma 2.1, for x, y ∈ O
and s ∈ R we have
p(t + s, x, y)
lim = 1. (2.2)
t→∞ p(t, x, y)

The limit is uniform in compact sets of O. Also, the map t → p(t, x, x) is decreasing.

Lemma 2.3. In the same setting as in Lemma 2.2, further assume that for all s ∈ R,

a(t + s)
lim = 1. (2.3)
t→∞ a(t)
1274 P. Collet et al. / Journal of Functional Analysis 270 (2016) 1269–1298

1+|y|
If a(t)p(t, x, y) ≤ Cx for large enough t, then any limit point of a(t)p(t, · , ·) (in the
topology of uniform convergence on compact sets) has the following properties:

(i) is a symmetric, non-negative function;

(ii) is harmonic in each component;
(iii) and vanishes continuously on ∂O.

Proof. For the sake of simplicity we denote ht (x, y) = a(t)p(t, x, y). Let tk → ∞ be a
sequence such that htk converges uniformly on compact sets of O to a function h. It is
clear that h is symmetric and non-negative. Notice that for any s ∈ R, the sequence
htk +s also converges uniformly on compact sets of O. This is direct from Lemma 2.2 and
the hypothesis.
By the Chapman–Kolmogorov equation, for any s ∈ R and large enough k ∈ N,

a(tk + s)
htk +s (x, y) = htk (x, z)p(s, z, y)dz.
a(tk )
Ω

1+|z|
By assumption, htk (x, z) ≤ Cx , which is p(s, z, y)dz-integrable as it can be checked by
comparing p with the free Brownian motion’s kernel. Thus, we can apply the Dominated
Convergence Theorem to obtain

h(x, y) = h(x, z)p(s, z, y)dz = Ey (h(x, Xs )).
Ω

It is standard to show that h(x, Xs ) is a martingale, from where it’s standard to deduce
that y → h(x, y) is harmonic by means of the optional sampling theorem.
Consider a sequence yn ∈ O, with yn → y ∈ ∂O. By using once again the Gaussian up-
per bound on p, and applying the Dominated Convergence Theorem to h(x, z)p(1, z, yn ),
it is deduced that h(x, ·) vanishes continuously on ∂O. 2

Lemma 2.4. Let U and O be domains in Rn that are regular for the Dirichlet problem.
For ξ ∈ ∂U , x ∈ U

1
Px (BT U ∈ σ(dξ), T U ∈ ds) = ∂n pU (s, x, ξ)σ(dξ)ds. (2.4)
2

Here, ∂n represents the inward normal derivative at ξ ∈ ∂U .

Also, if U ⊆ O, then

t 
1
O U
p (t, z, y) = p (t, z, y) + ∂n pU (s, x, ξ)pO (t − s, ξ, y)σ(dξ)ds. (2.5)
2
0 ∂U
 P. Collet et al. / Journal of Functional Analysis 270 (2016) 1269–1298 1275

Proof. These results are well known so we only are going to comment their proofs. The
proof of (2.4) uses Green’s theorem and the heat equation, and it is very straightforward
to carry out. Equation (2.5) follows as an elementary application of the strong Markov
property at time T U . 2

The following lemma characterizes all positive, harmonic functions vanishing on ∂Ω.
In other words, we characterize the Martin boundary of Ω. We use the notation from
the Introduction.

Lemma 2.5. Let u1 , . . . , uN be the minimal harmonic functions given by (1.3). For every
nonnegative harmonic function u in Ω, vanishing continuously on ∂Ω, there are unique
nonnegative coefficients γ1 , · · · , γN such that


N
u(x) = γj uj (x), x ∈ Ω. (2.6)
j=1

Proof. For x ∈ Ωj , consider the harmonic function w̃j (x) = u(x) − Ex (u(BT j )). It is
standard to check that w̃j is harmonic in Ωj , and that it vanishes continuously on ∂Ωj .
For m > R, let Tm j
be the exit time from the set Ωj ∩ B(aj , m). By Itô’s formula, the
process u(Bt∧Tmj ) is a bounded martingale under Px , for x ∈ Ωj . Therefore,

   
   +E  )
u(x) = Ex u(BTm
j ) = Ex u(BTm
j )1
T j <T j x u(BT j )1 T j =T j
m m

 
≥ Ex (u(BT j )) − Ex u(BT j )1T j <T j ) .
m

j
Since Tm  T j , monotone convergence shows that u(x) ≥ Ex (u(BT j )), that is, w̃j is
nonnegative. Thus, w̃j (x) = γj wj (x) by uniqueness. For z ∈ Ω, set


N
ũ(z) = γj uj (z) − u(z),
j=1

which is harmonic in Ω, and vanishes continuously on ∂O. We will next show that ũ is
bounded, for this it is enough to show that it is bounded in each branch of Ω.
Fix i ∈ {1, . . . , N }, and consider x ∈ Ωi . We have


N
ũ(x) = −Ex (u(BT i )) + γi Ex (ui (BT i )) + γj uj (x)
j=1,j=i

The first term on the right hand side is bounded by supx∈Γi |u(x)|, and the second one
by γi supx∈Γi |ui (x)|. The summation is bounded as each term uj (x) is bounded in Ωi .
1276 P. Collet et al. / Journal of Functional Analysis 270 (2016) 1269–1298

We conclude that ũ is harmonic and bounded in Ω, and vanishes continuously on ∂Ω. It

follows that ũ (Bt∧T ) is a martingale, and so

ũ(z) = Ez (ũ (Bt∧T )) → 0, as t → ∞.

Uniqueness follows from the boundedness of uj in Ω \ Ωj , and its unboundedness

in Ωj . 2

2.1. Asymptotics in a cone with vertex

In what follows we consider a cone V , with opening D and vertex a = 0, that is,
V = C(0, D, 0). Let pV be the heat kernel in V . Let 0 < λ1 < λ2 ≤ λ3 ≤ · · · be
the eigenvalues of the Laplace–Beltrami operator on D, with corresponding orthonormal
 1/2
basis {m1 , m2 , m3 , . . .} of L2 (D, σ). We also denote αi = λi + ( n2 − 1)2 .
The behaviour of the heat kernel with Dirichlet boundary conditions is well known
for a cone with vertex. The following results are taken from [3].

Theorem 2.6. For x = rθ, y = ρω ∈ V , with θ, ω ∈ D and r = |x|, ρ = |y|, the heat
kernel with Dirichlet boundary conditions in V is given by:
 2 2
exp − r 2t +ρ
∞  rρ 
V
p (t, x, y) = n
−1
Jα i mi (θ)mi (ω) (2.7)
t (rρ) 2 i=1
t

where Jν is the modified Bessel function of first kind of order ν, that is, the solution of

z 2 Jν (z) + zJν − (z 2 + ν 2 )Jν = 0,

satisfying the growing conditions:

zν zν
≤ Jν (z) ≤ ν ez , (2.8)
2ν Γ(1 + ν) 2 Γ(1 + ν)

for z > 0, and ν ≥ 0.

Recall that the

unique

minimal positive harmonic function in V is given by v(x) =
α1 − n2 −1
v(|x| θ) = |x| 1
m (θ).

Corollary 2.7. For each x, y ∈ V , we have

1 v(x)v(y)
lim t1+α pV (t, x, y) = (2.9)
t→∞ 2α1 Γ(1 + α1 )
pV (t, x, y) v(x)v(y)
lim V
= (2.10)
t→∞ p (t, w, z) v(w)v(z)

Both limits are uniform in compact sets.

 P. Collet et al. / Journal of Functional Analysis 270 (2016) 1269–1298 1277

Proof. Clearly, (2.10) follows from (2.9), so we only prove the latter. From Theorem 2.6,
we get the bound
   1 r2 +s2 
 1+α1 V v(x)v(y)   tα e− 2t  rs  (rs)α1 −( n2 −1) 
t  
 p (t, x, y) − α1 ≤C J 1 − α1 +
2 Γ(1 + α1 )   (rs) 2 −1 α 2 Γ(1 + α1 ) 
n
t
∞
 (rs)αk −( 2 −1)
n
−(α2 −α1 )
+t ,
2αk Γ(1 + αk )
k=2

where C = supθ∈D m1 (θ)2 . The uniform convergence on compact sets for the first term
is easily deduced from (2.8). The series on the right hand side converges uniformly in
compact sets, so the whole term converges to zero, as t → ∞, since α2 > α1 . 2

3. Asymptotics in a truncated cone

The main goal of this section is to extend Corollary 2.7 to a truncated cone C =
C(a, D, R). As before, we assume that R = 1 and a = 0.
We will often use the following version of the Harnack inequality up to the boundary.

Theorem 3.1. (From [12], see also [8]) Let O be a precompact, regular domain for the
Dirichlet problem, and let u ≥ 0 be a solution of the heat equation on O × [0, T ) with
Dirichlet boundary condition. Then, given x ∈ O, there is C1 > 0 such that u(t, z) ≤
C1 u(T, x), for all (t, z) ∈ [0, T ) × O where the constant C1 depends only on x and T − t.

Corollary 3.2. Let V be a cone with vertex. For any x ∈ V there is a constant Cx > 0,
only dependent on x, such that for all y ∈ V the following inequality holds for all t > 1:

pV (t, x, y) ≤ Cx1+|y| pV (t, x, x). (3.1)

Proof. Assume |x| = 1, otherwise the corollary follows by scaling. The inequality holds
for small |y|, by a direct application of the boundary Harnack inequality (Theorem 3.1),
so we assume that |y| > 2.
Let r be positive, but small enough so that B (x, r) ⊆ V . It follows by scaling that
B (νx, r) ⊆ V for all ν ≥ 1. Thus, applying the standard parabolic Harnack inequality
several times in the ball B(0, r) to the function u(s, z) = pV (t + s, νx + z, y) for fixed,
but arbitrary ν > 1, y ∈ V , we get

pV (t, νx, y) ≤ C21+rν pV (t + 1 + rν, x, y) ≤ C22+2rν pV (t + 2 + 2rν, νx, y),

for a positive constant C2 that only depends on x.

The heat kernel in V has the following scaling property:
 
t x y
pV (t, x, y) = λ−n pV , , , λ > 0.
λ2 λ λ
1278 P. Collet et al. / Journal of Functional Analysis 270 (2016) 1269–1298

From all the inequalities above, it follows that

1+|y| V
pV (t, x, y) ≤ C3 p (t + 1 + r |y| , x |y| , y)
 
1+|y| −n V t + 1 + r |y| y
= C3 |y| p , x,
|y|
2 |y|
 
1+|y| −n V t + 1 + r |y|
≤ C1 C3 |y| p 2 + 1, x, x
|y|
 
1+|y| −n V t
≤ C1 C3 |y| p 2 , x, x ,
|y|

where the second-to-last line comes from the boundary Harnack inequality, whereas the
last one comes form the fact that t → pV (t, x, x) is decreasing (see Lemma 2.2). Applying
scaling once again,

1+|y| V
pV (t, x, y) ≤ C1 C3 p (t, x |y| , x |y|)
3+3|y| V
≤ C1 C3 p (t, x, x) ,

as desired. 2

Lemma 3.3. Let C = C(a, D, 1), S = a + D, and V = C(a, D, 0). There is a universal
constant Q > 0 such that for any x ∈ C, and ξ ∈ S,

∂n pC (t, x, ξ) Q
lim sup ≤ , (3.2)
t→∞ pV (t, x, x) v(x)

where v is the unique minimal harmonic function in V , normalized as in (1.1).

Proof. By a translation of coordinates, we can assume a = 0. Set U = B(0, 1)c . By

monotonicity of domains, and since both pC (t, x, ·) and pU (t, x, ·) vanish on S, we have
that 0 ≤ ∂n pC (t, x, ξ) ≤ ∂n pU (t, x, ξ). Recall that there are constants A > 0, B > 0 such
that ∂n pU (1, x, ξ) ≤ A exp(−B |x| ), so 0 ≤ ∂n pC (1, x, ξ) ≤ A exp(−B |x| ) for ξ ∈ S.
2 2

These bounds allow us to compute the normal derivative from the Chapman–Kolmogorov
equation as follows

∂n pC (t + 1, x, ξ) = pC (t, x, z)∂n pC (1, z, ξ)dz
C

≤ pV (t, x, z)∂n pC (1, z, ξ)dz. (3.3)
C
 P. Collet et al. / Journal of Functional Analysis 270 (2016) 1269–1298 1279

Thus,

∂n pC (t + 1, x, ξ) pV (t, x, z)
≤ ∂n pC (1, z, ξ)dz.
pV (t, x, x) pV (t, x, x)
C

We intend to apply the Dominated Convergence Theorem to the integral on the right
hand side. Equation (2.10) shows pointwise convergence as t → ∞, and Corollary 3.2
together with the remarks at the beginning of this proof show that the integrand is
dominated. Therefore

∂n pC (t + 1, x, ξ) 1
lim sup ≤ v(z)∂n pC (1, z, ξ)dz.
t→∞ pV (t, x, x) v(x)
C

The integral can be estimated using the explicit formula for v(z), and the bound for
∂n pC (1, z, ξ) discussed at the beginning of this proof. Finally, we use Lemma 2.2 to
conclude. 2

Theorem 3.4. Let V be a cone with opening D and vertex a, and let its truncated version
be C = C(a, D, 1). Let w be the unique minimal positive harmonic function in C. Then,
for all x, y ∈ C,

w(x)w(y)
lim t1+α pC (t, x, y) =
1
, (3.4)
t→∞ 2α1 Γ(1 + α1 )

where α1 is the character of C. The limit is in the sense of uniform convergence on

compact sets.

Proof. The proof relies on equation (2.5) and the Dominated Convergence Theorem. Let

t/2
1
I1 (t) = ∂n pC (s, x, ξ)pV (t − s, ξ, y)σ(dξ)ds
2
0 S

t 
1
I2 (t) = ∂n pC (s, x, ξ)pV (t − s, ξ, y)σ(dξ)ds.
2
t/2 S

Then, equation (2.5) and Theorem 2.6 yield,

v(x)v(y)
≤ lim sup t1+α pC (t, x, y) + lim sup t1+α (I1 (t) + I2 (t)) .
1 1
(3.5)
2α1 Γ(1 + α1 ) t→∞ t→∞

1
We start by studying I1 (t). For 0 ≤ s ≤ t/2, Theorem 2.6 shows that t1+α pV (t −
s, ξ, y) converges to 2αv(ξ)v(y)
1
Γ(1+α1 )
. Besides, by using Corollary 3.2 we get the following
bound:
1280 P. Collet et al. / Journal of Functional Analysis 270 (2016) 1269–1298

1 1 1
t1+α pV (t − s, ξ, y) = 21+α (t/2)1+α pV (t − s, ξ, y)
1 1+|y| 1
≤ 21+α C1 C2 (t/2)1+α pV (t − s, x, x)
1 1+|y| 1+α1
≤ 21+α C1 C2 (t/2) pV (t/2, x, x).

The right hand side is uniformly bounded for t > 1, so the Dominated Convergence
Theorem applies:

∞ 
1 v(ξ)v(y) v(y)Ex (v(BT C ))
lim t 1+α1
I1 (t) = ∂n pC (s, x, ξ) σ(dξ)ds = .
t→∞ 2 2α1 Γ(1+α )1 2α1 Γ(1 + α1 )
0 S

Next, we study the asymptotics of I2 (t). Since we don’t have sharp asymptotics for
∂n pC yet, we are not able to use the Dominated Convergence Theorem. Instead, we will
resort to Fatou’s lemma. For t/2 ≤ s ≤ t, Lemma 3.3 and Theorem 2.6 imply that

Q1 v(x)
lim sup t1+α ∂n pC (s, x, ξ) ≤
1
1 ,
t→∞ 2 Γ(1 + α1 )
α

where Q1 only depends on x. To show domination, we combine equations (3.1) and (3.3)
to get the bound

t1+α ∂n pC (s, x, ξ) ≤ t1+α pV (s − 1, x, z)∂n pC (1, z, ξ)dz
1 1

C

∂n pC (1, z, ξ)dz.
1 1+|z|
≤ t1+α pV (t/2 − 1, x, x) C2
C

The right hand side is uniformly bounded in t > 2 by a constant Q2 that only depends
on x. It follows that
∞ 
1 Q1 v(x)
lim sup t1+α I2 (t) ≤ pV (s, ξ, y)σ(dξ)ds
t→∞ 2α1 Γ(1 + α1 )
0 S

Q1 v(x)
= GV (S, y).
2α1 Γ(1 + α1 )

Using these two estimates in equation (3.5) we obtain,

v(x)v(y) v(y)Ex (v(BT C )) + Q1 v(x)GV (S, y)

≤ lim sup t1+α pC (t, x, y) +
1
.
2α1 Γ(1+α )1
t→∞ 2α1 Γ(1 + α1 )

Recall that w(x) = v(x) − Ex (v(BT C )) for x ∈ C. Thus,

w(x)v(y) Q1 v(x)GV (S, y)

≤ lim sup t1+α pC (t, x, y) +
1
. (3.6)
2α1 Γ(1 +α ) 1
t→∞ 2α1 Γ(1 + α1 )
 P. Collet et al. / Journal of Functional Analysis 270 (2016) 1269–1298 1281

We will deduce the theorem from this last estimate. By Corollary 3.2, it is possible
to apply Lemma 2.3 to t1+α pC (t, x, y). Therefore, any limit point of this family has the
1

form ηw(x)w(y) for some η ≥ 0. Different limit points will correspond to different values
of η. We will show that this is not the case: by monotonicity of domains, ηw(x)w(y) ≤
v(x)v(y)
2α1 Γ(1+α1 )
, and since w(z) and v(z) have the same asymptotic behaviour for radially
convergent z → ∞, we deduce that η ≤ 2α1 Γ(1+α1
1)
. Let η ∗ = sup η, where the supremum
is taken over all possible limit points. Equation (3.6) then yields

w(x)v(y) ∗ Q1 v(x)GV (S, y)

1 ≤ η w(x)w(y) + .
2α Γ(1 + α1 ) 2α1 Γ(1 + α1 )

Dividing this equation by v(y) and taking y radially to infinity, the second term on the
right hand side vanishes in the limit as GV (S, y) is bounded for y away form S. We
obtain that η ∗ ≥ 2α1 Γ(1+α
1
1)
, which shows that the only possible limit point is the one
given by (3.4). Uniform convergence on compact sets follows from Lemma 2.1. 2

4. Asymptotics in multicone domains

We start by fixing x0 ∈ Ω, and a sequence (tk ) such that

p(tk , x, y)  k
F (x, y) = lim = γij ui (x)uj (y),
k→∞ p(tk , x0 , x0 )
i,j=1

where the convergence is uniform in compact sets of Ω × Ω, and uj are the minimal
harmonic functions in Ω. This is obtained by a double application of Lemma 2.5. The
coefficients γij ≥ 0 might depend on the sequence (tk ). Notice that F (x0 , x0 ) = 1.
By passing to subsequences of (tk ), we can assume that for all j = 1, . . . , k we have
that

pj (tk , x, y)
Fj (x, y) = lim = μj wj (x)wj (y)
k→∞ p(tk , x0 , x0 )

is well defined. The convergence is uniform in compact sets of Ωj × Ωj . The coefficient

μj ≥ 0 may also depend on the sequence (tk ).
Our goal is to compute explicitly the coefficients γij . In order to do this, we will use
equation (2.5) with O = Ω, and U = Ωj , and estimate the integral involved in (2.5).
It will be convenient to fix points ξj ∈ Sj , and zj ∈ Ωj . For x ∈ Ωj , y ∈ Ω, j =
1, . . . , N , we define the following object

b 
Ijx,y (a, b; t) = Px (BT j ∈ dξ, T j ∈ du)p(t − u, ξ, y). (4.1)
a Sj
1282 P. Collet et al. / Journal of Functional Analysis 270 (2016) 1269–1298

Most of the technical work of this section will be devoted to find convenient estimates
for Ijx,y .
We start with a lemma about the function Fj . Recall that M denotes the set of
maximal indices.

Lemma 4.1. We have that μj = μ for j ∈ M, and μj = 0 for j ∈ / M. Also, there is a

constant C, independent of the sequence (tk ), such that μ ≤ C.

Proof. Recall that αj denotes the character of the branch Ωj . For j, l = 1, . . . , N , and
points x ∈ Ωj , y ∈ Ωl we have

pj (t, x, x) t1+αj pj (t, x, x) pl (t, y, y) 1

= 1+α l .
p(t, x0 , x0 ) t l p (t, y, y) p(t, x , x ) tαj −αl
0 0

It follows that

pj (tk , x, x) 2αl Γ(1 + αl )wj (x)2 1

μj wj (x)2 = lim = αj μl wl (y)2 lim αj −αl .
k→∞ p(tk , x0 , x0 ) 2 Γ(1 + αj )wl (y)2 t→∞ t

If j ∈
/ M and l ∈ M, we have αl < αj , and so μj = 0. If both j, l ∈ M, we have αj = αl ,
and so μj = μl = μ, only depending on (tk ).
Pick any j ∈ M. By Harnack’s inequality we have

pj (tk , zj , zj ) j
2 p (tk , zj , zj )
≤ CH ≤ CH
2
,
p(tk + 2, x0 , x0 ) p(tk , zj , zj )

by monotonicity of domains. Using Lemma 2.2, we see that the left hand side above
converges to μwj (zj )2 , thus,

2 2
CH CH
μ≤ ≤
wj (zj )2 inf j wj (zj )2

as desired. 2

Lemma 4.2. There is a constant C > 1 such that, for every M > 2, every index 1 ≤ j ≤
N , m ∈ M, and points x ∈ Ωj , y ∈ Ω we have for tk > 2M + 1

Ijx,y (0, M ; tk )
lim sup ≤ CPx (BT j ∈ Sj ) F (ξj , y), (4.2)
k p(tk , x0 , x0 )
Ijx,y (tk − M, tk ; tk )
lim sup ≤ C1M (j)G(Sj , y)wj (x)wj (zj ), (4.3)
k p(tk , x0 , x0 )
Ijx,y (L, tk − L; tk ) wj (x)
lim sup lim sup ≤ C1M (j) F (z, y), (4.4)
L k p(tk , x0 , x0 ) wm (z)
 P. Collet et al. / Journal of Functional Analysis 270 (2016) 1269–1298 1283

where the last equation holds for any z ∈ Ωm . The constant C depends only on the
domain Ω and our choices of zj .

Proof. Take k large enough such that tk > 2M +1. By the boundary Harnack inequality,
there exists a positive C1 such that for all u ∈ [0, tk − 1], all y ∈ Ω, and all i = 1, . . . , N

p(tk − u, ξ, y) ≤ C1 p(tk − u + 1, ξi , y). (4.5)

By Fatou’s lemma and Lemma 2.2 we get

M 
Ijx,y (0, M ; tk )
lim ≤ C1 Px (BT j ∈ dξ, T j ∈ du)F (ξj , y),
k p(tk , x0 , x0 )
0 Sj

from which (4.2) follows easily.

For u ∈ [0, M ], Lemma 3.3, Theorem 3.4, equation (3.3) and the Dominated Conver-
gence Theorem yield
∂n pj (tk − u, x, ξ) ∂n wj (ξ)
lim = , (4.6)
k→∞ pj (tk , x, zj ) wj (zj )

where the convergent sequence is bounded by a constant that only depends on M , x

and zj (see Lemma 3.3). It follows by the Dominated Convergence Theorem, and Har-
nack’s inequality, that
M 
Ijx,y (tk − M, tk ; tk ) pj (tk , x, zj ) 1 ∂n wj (ξ)
lim ≤ lim p(u, ξ, y)σ(dξ)du
k p(tk , x0 , x0 ) k p(tk , x0 , x0 ) 2 wj (zj )
0 Sj

pj (tk , x, zj )
≤ C2 G(Sj , y) lim ,
k p(tk , x0 , x0 )
1

where C2 = max supξ∈Sj ∂n wj (ξ), and G(Sj , y) = G(ξ, y)σ(dξ). This
j=1....,N 2wj (zj ) Sj
inequality and Lemma 4.1 prove (4.3).
Recall that there is r0 > 0 such that for all i = 1, . . . , N , we have B2r0 (ξi ) ∩{|x| = 1} ⊆
Si . For x ∈ Ωj , z ∈ Ωm , set

Sj n
∂ pj (t, x, ξ)σ(dξ)
Cjm (x, z) = sup 
L
m
,
t>L Sm ∩Br (ξm ) ∂n p (t, z, ξ )σ(dξ )
0

which is finite since m ∈ M. From Lemma 3.3, Theorem 3.4, and the Dominated Con-
vergence Theorem, we obtain

wj (x) ∂ w (ξ)σ(dξ)
Sj n j wj (x)
L
lim Cjm (x, z) = 1M (j)  ≤ C3 1M (j) ,
L→∞ wm (z) Sm ∩Br (ξm ) ∂n wm (ξ )σ(dξ ) wm (z)
0

for a constant C3 > 0.

1284 P. Collet et al. / Journal of Functional Analysis 270 (2016) 1269–1298

From the standard Harnack inequality we have p(t −u +1, ξm , y) ≤ C4 p(t −u +2, ξ , y)
for all ξ ∈ Br0 (ξm ). The previous discussion yields the following series of inequalities

t
k −L
1
Ijx,y (L, tk − L; tk ) ≤ C1 ∂n pj (u, x, ξ)p(tk − u + 1, ξm , y)σ(dξ)du
2
L Sj

t
k −L 
1
≤ C1 Cjm
L
(x, z) ∂n pm (u, z, ξ )p(tk − u + 1, ξm , y)σ(dξ )du
2
L Sm ∩Br0 (ξm )

t
k −L 
1
≤ L
C1 C4 Cjm (x, z) ∂n pm (u, z, ξ )p(tk − u + 2, ξ , y)σ(dξ )du
2
L Sm ∩Br0 (ξm )

t
k −L 
1
≤ L
C1 C4 Cjm (x, z) ∂n pm (u, z, ξ )p(tk − u + 2, ξ , y)σ(dξ )du
2
L Sm

≤ C1 C4 Cjm
L
(x, z)p(tk + 2, z, y).

Equation (4.4) now follows from Lemma 2.2. 2

Lemma 4.3. For the coefficients of the function F defined at the beginning of this section:

(i) γij = 0 if i ∈
/ M or j ∈
/ M.
(ii) There is a universal constant C depending only on the domain Ω such that γij ≤
Cγjm for i, j, m ∈ M, with i = j.

Proof. Let x ∈ Ωi and y ∈ Ωj . By (2.5),

p(t, x, y) = pi (t, x, y)δij + Iix,y (0, t; t).

From Lemmas 4.2 and 4.1 we obtain for m ∈ M, z ∈ Ωm ,


F (x, y) ≤ δij μi wi (x)wi (y) + C F (ξi , y) + G(Si , y)1M (i)wi (x)wi (zi )

wi (x)
+ 1M (i) F (z, y)
wm (z)
 
F (z, y)
≤ C1M (i)wi (x) δij wj (y) + G(Si , y)wi (zi ) + + CF (ξi , y).
wm (z)

The use of this inequality is twofold. First, if i ∈/ M, by taking x radially to infinity

in Ωi we find that γij ui (x)uj (y) ≤ CF (ξi , y) is only possible if γij = 0. By symmetry of
the kernel we conclude (i).
 P. Collet et al. / Journal of Functional Analysis 270 (2016) 1269–1298 1285

Secondly, consider i, j, m ∈ M, with i = j and z ∈ Ωm . Dividing the inequality by

wi (x)wj (y), and taking x, y, z → ∞ radially in their respective branches, we obtain that
γij ≤ Cγmj , as desired. 2

Remark 4.1. Set γ ∗ = maxi∈M γii and fix m ∈ M such that γ ∗ = γmm . The previous
lemma states that γij ≤ Cγ ∗ for all i, j ∈ M. Also, notice that

γmm um (x)um (y) ≤ F (x, y) ≤ (1 + C)γmm ui (x)uj (y).
i,j∈M

Since F (x0 , x0 ) = 1, we obtain that γ ∗ is bounded below by a constant that is indepen-

dent of the sequence (tk ).
It follows that

p(tk , x, y) F (x, y)  ui (x)uj (y)

lim = ≤ (1 + C)
k p(tk , z, z ) F (z, z ) um (z)um (z )
i,j∈M

where the constant C comes from Lemma 4.2. In particular, if x ∈ Ωm , the inequality
⎛ ⎞
p(t, x, ξm ) 1+C  1 
lim sup ≤ uj (ξm ) ⎝1 + sup ui (z)⎠ (4.7)
t→∞ p(t, x, x) um (x) um (x) z∈Ωm
j∈M i=m

 ∈ Dm , and let r > 0 be sufficiently large, for x = am +r

holds. If we fix x x this inequality
implies that

p(t, x, x0 ) C5
lim sup ≤ , (4.8)
t→∞ p(t, x, x) um (x)

where C5 > 0 is independent of r.

Lemma 4.4. The following inequalities hold

0 < lim inf t1+α p(t, x, y), lim sup t1+α p(t, x, y) < ∞.

Proof. The first inequality is direct from monotonicity of domains and Theorem 3.4
applied to Ωm ⊆ Ω.
For the second one, notice that by Harnack’s inequality, it suffices to prove the theorem
for x = y ∈ Ωm . We start by setting some constants that will be relevant to our estimates:
fix x ∈ Dm , and consider x = am + r x. Then, the Harnack constant

p(s, ξ, r
x)
CH = sup ,
s>1,ξ∈Sm p(s + 1, ξm , r
x)

is independent of r > 1.
1286 P. Collet et al. / Journal of Functional Analysis 270 (2016) 1269–1298

Fix 0 < θ < 1. In view of (4.8), we can find x ∈ Ωm such that

p(t, ξm , x) C5 θ
CH 21+α lim sup ≤ CH 21+α = .
t→∞ p(t, x, x) um (x) 2

We fix such an x = am +r
x from now on. It follows that for large enough t0 , the inequality

p(t, ξm , x) θ
≤ 1+α , ∀ t > t0 (4.9)
p(t, x, x) 2 CH

holds.
Next we get estimates using techniques somewhat similar to the ones we have used in
Lemma 4.2. Let t > 2t0 . By Harnack’s inequality, and since t → p(t, x, x) is decreasing

t/2
x,x
Im (0, t/2; t) ≤ CH Px (BT m ∈ Sm , T m ∈ du)p(t − u + 1, ξm , x)du
0

θ
≤ Px (BT m ∈ Sm , T m < t/2)p(t/2, x, x)
21+α
θ
≤ 1+α p(t/2, x, x).
2

It follows that for all t > 2t0 ,

 1+α
t
t1+α Im
x,x
(0, t/2; t) ≤θ p (t/2, x, x) .
2

On the other hand,

t/2 
t1+α
t1+α Im
x,x
(t/2, t; t) ≤ ∂n pm (t − u, x, ξ)p(u, ξ, x)σ(dξ)du.
2
0 Sm

The right hand side converges by (4.6), Theorem 3.4, and the Dominated Convergence
Theorem, to

∞ 
wm (x)
C2 = 1+α ∂n wm (ξ)p(u, ξ, x)σ(dξ)du < ∞.
2 Γ(1 + α)
0 Sm

Putting together these two estimates, we have that, for the continuous function ϕ(t) =
t1+α p(t, x, x), t ≥ 2t0 , it holds that

ϕ(t) ≤ C3 + θϕ(t/2), t > 2t0 ,

 P. Collet et al. / Journal of Functional Analysis 270 (2016) 1269–1298 1287

2
where C3 = C2 + 2αwΓ(1+α)
m (x)
. By iteration of the inequality above, it is easy to deduce
that, if t/2 ∈ [2t0 , 4t0 ] then
N


N −1
C3
ϕ(t) ≤ C3 θk + θN ϕ(t/2N ) ≤ + sup ϕ(s),
1 − θ s∈[2t0 ,4t0 ]
k=0

which finishes the proof. 2

4.1. Proof of Theorem 1.1

The proof is reminiscent of the one we gave for Theorem 3.4. For fixed x, y ∈ Ω,
Lemma 4.4 ensures that t1+α p(t, x, y) is bounded. Harnack’s inequality then ensures
that the same holds for x, y in any compact set of Ω. This shows that Lemmas 2.1
and 2.3 apply. Let (tk ) be a sequence such that t1+α p(tk , · , ·) converges uniformly on
k N
compact sets. The limit then has the form H(x, y) = i,j=1 ηij ui (x)uj (y), where ηij
might depend on the sequence (tk ).
Let x ∈ Ωm , and y ∈ Ω. Set

t/2 
1
I1 (t) = ∂n pm (s, x, ξ)p(t − s, ξ, y)σ(dξ)ds,
2
0 Sm

t 
1
I2 (t) = ∂n pm (s, x, ξ)p(t − s, ξ, y)σ(dξ)ds.
2
t/2 Sm

An application of the Dominated Convergence Theorem, as in the proof of Lemma 4.4,

yields that

∞ 
1
lim t1+α
k I1 (tk ) = ∂n pm (s, x, ξ)H(ξ, y)σ(dξ)ds
k→∞ 2
0 Sm

= Ex (H (BT m , y)) .

On the other hand, by using equations (1.3) and (1.4)

∞ 
1
lim t1+α
k I2 (tk ) = 1M (m) wm (x)∂n wm (ξ)p(t − s, ξ, y)σ(dξ)ds
k→∞ 2
0 Sm

wm (x) (um (y) − wm (y))

= 1M (m) .
2α Γ(1 + α)
1288 P. Collet et al. / Journal of Functional Analysis 270 (2016) 1269–1298

Using these two estimates, and (2.5), we arrive at

wm (x)um (y)
H(x, y) = Ex (H (BT m , y)) + 1M (m) , x ∈ Ωm , y ∈ Ω.
2α Γ(1 + α)

Recall that Ex (H (BT m , y)) is bounded as function of x. By taking the limit of

H(x, y)/um (x), with x going radially to infinity, we find that


N
um (y)
ηmj uj (y) = 1M (m) .
j=1
2α Γ(1 + α)

By uniqueness of the decomposition (2.6), we find that the only nonzero coefficients are
1
γmm = 2α Γ(1+α) for m ∈ M. This shows (1.6). Uniform convergence on compact sets is
direct from Lemma 2.1. 2
The following corollary is a direct consequence of the previous theorem.

Corollary 4.5. Let Ω be a multicone domain, with maximal index set M. Then,

p(t, x, y) j∈M uj (x)uj (y)
lim = . (4.10)
t→∞ p(t, w, z)
j∈M uj (w)uj (z)

The convergence is uniform in compact sets.

5. Asymptotics for the exit time

The following result is taken form [3].

Theorem 5.1. Let V ⊆ Rn be a cone with vertex 0 and opening D. Assume that D is
regular for the Laplace–Beltrami operator on Sn−1 , and let α be the character of D. Set
κ = 1 + α − n/2, and let T V be the Brownian exit time from V . Then, for each x ∈ V ,

lim tκ/2 Px (T V > t) = γV v(x). (5.1)

t→∞

κ
Here v(x) = |x| m1 (x/ |x|) is the harmonic function defined in (1.1), where m1 is the
only non-negative eigenfunction of the Laplace–Beltrami operator on D with Dirichlet
boundary conditions. Also,
  
Γ κ+n
γV =  2 n  m1 (θ)σ(dθ).
2κ/2 Γ κ + 2
D

Remark 5.1. From this theorem, the scaling property of the heat kernel in V , and Har-
nack’s inequality up to the boundary, we get the following bound:
 P. Collet et al. / Journal of Functional Analysis 270 (2016) 1269–1298 1289

   
t t
t κ/2
Px (T V
> t) = tκ/2
Px/|x| T V
> 2 ≤ CH t κ/2
Pξ T V
> 2
|x| |x|
 κ/2  
κ t t
≤ CH |x| 2 Pξ T V
> 2 ,
|x| |x|

2
where ξ ∈ D is fixed. Using (5.1), we can pick tξ such that whenever t/ |x| > tξ , the
κ
right hand side of the last display is bounded by C |x| , where C depends on our choice
2 κ/2 κ
of ξ. For t/ |x| ≤ tξ , we have that tκ/2 ≤ tξ |x| . We deduce that there is a universal
constant C > 0 such that
κ
tκ/2 Px (T V > t) ≤ C |x| , x ∈ V, t > 0. (5.2)

By monotonicity of domains, the same inequality holds for T C , and x ∈ C, where C is

any truncated cone with opening D.

The following lemma will be the key tool when extending the previous result to mul-
ticones.

Lemma 5.2. Let T be the exit time from a multicone set Ω, let T j be the exit time from Ωj ,
and pick x ∈ Ωi for some i = 1, . . . , N . We have

Px (T > t) = Px (T i > t) + Px (Bt ∈ Ω0 , T > t) +

t 
1
N
+ ∂n Pz (T j > t − s)p(s, x, z)σ(dz)ds. (5.3)
2 j=1
0 Sj

Proof. For j = 1, . . . , k, and 0 ≤ s ≤ t define the functions


fj (s) = Pz (T j > t − s)p(s, x, z)dz.
Ωj

For s < t, since u(s, z) = Pz (T j > s), and v(s, z) = p(s, x, z) are solutions of the heat
equation with Dirichlet boundary condition in Ωj and Ω respectively, we have by Green’s
formula

dfj (s) 1
= −p(s, x, z)Δz Pz (T j > t − s) + Pz (T j > t − s)Δz p(s, x, z)dz
ds 2
Ωj

1
= p(s, x, z)∂n Pz (T j > t − s) − Pz (T j > t − s)∂n p(s, x, z)σ(dz)
2
Sj

1
= p(s, x, z)∂n Pz (T j > t − s)σ(dz),
2
Sj
1290 P. Collet et al. / Journal of Functional Analysis 270 (2016) 1269–1298

where, as usual, ∂n represents the (inward) normal derivative. Then, for every ε > 0,

t−ε
1
fj (t − ε) − fj (0) = p(s, x, z)∂n Pz (T j > t − s)σ(dz). (5.4)
2
0 Sj

In order to extend this equation to ε = 0, we need an estimate for ∂n Pz (T j > u) for u

near zero. The process B leaves Ωj before the norm of B hits level 1. Since ρt = |Bt | is
a Bessel process, if we let τ be the hitting time of 1 for an n-dimensional Bessel process,
then, for t < τ ,

t
n−1 n−1
ρt = |z| + βt + ρ−1
s ds ≤ |z| + βt + t,
2 2
0

for some one dimensional Brownian motion βt . Let τ μ be the hitting time of zero for the
Brownian motion with drift βt + μt, with μ = n−1
2 . It follows that

Pz (T j > u) ≤ P|z| (τ > u) ≤ Pr (τ μ > u), r = |z| − 1.

As Pz (T j > u) vanishes on Sj , we get ∂n Pz (T j > u) ≤ ∂r Pr (τ μ > u)|r=0 . The distribu-

tion of τ μ is well known (see equation (5.12), p. 197, [6]):
 
r (r + μt)2
Pr (τ μ ∈ dt) = √ exp − dt, t > 0.
2πt3 2t

A direct computation shows that

∞  
r (r + μt)2
Pr (τ μ > u) = √ exp − dt,
2πt3 2t
u
∞  2 
1 μ t
∂r Pr (τ > u)|r=0 =
μ
√ exp − dt
2πt 3 2
u

1  2 ∞
t−3/2
1 μ t
≤ √ dt +
√ exp − dt
2π2π 2
u 1
   
2 1 2 −2 −μ2 /2
= √ −1 + μ e , u < 1,
π u π

which is integrable (in u) near zero. Therefore, we can apply the Dominated Convergence
Theorem in (5.4) and use the continuity of fj to deduce that this equation also holds for
ε = 0. Adding all the equations for j = 1, . . . , N , using that fj (0) = 1Ωj (x)Px (T j > t)
and fj (t) = Px (Bt ∈ Ωj , T > t), and adding the contribution from Ω0 , we obtain
(5.3). 2
 P. Collet et al. / Journal of Functional Analysis 270 (2016) 1269–1298 1291

5.1. Proof of Theorem 1.2 for truncated cones

Theorem 5.1 is also valid if we change the cone V = C(a, D, 0) for its truncated
version C = C(a, D, 1), but the limit turns out to be γV w(x), where w(·) is the
unique positive harmonic function in C that vanishes on ∂C, normalized at infinity,
such that limr→∞ r−κ w(a + rθ) = 1 for any θ ∈ D. Recall that S = a + D is the base
of C.
Indeed, notice that (t, x) → Px (T C > t) solves the heat equation in C. By Har-
nack’s inequality, the family of functions h(t, x) = tκ/2 Px (T C > t), indexed by t > 0,
is bounded on compact sets of C. Since Px (T C > t) ≤ Px (T V > t), Lemma 2.3 applies
and we conclude that any limit point has the form μw(x). Of course, the constant μ
may depend on the sequence (tk ) that makes h(tk , ·) converge. Nevertheless, we have
μ ≤ γV by monotonicity of domains, where γV is the same constant as in Theo-
rem 5.1.
On the other hand, from Lemma 5.2, we have that for all x ∈ C

Px (T V > t) = Px (T C > t) + Px (|Bt | < 1, T V > t) +

t 
1
+ ∂n Pz (T C > t − s)pV (s, x, z)σ(dz)ds.
2
0 S

Also, by Harnack’s inequality tκ/2 Px (|Bt | < 1, T V > t) ≤ C1 tκ/2 p(t + 1, x, x0 ), which
converges to zero, as t → ∞ by Theorem 1.1. Thus,

t 
tκ/2
γV v(x) ≤ μw(x) + lim ∂n Pz (T C > t − s)pV (s, x, z)σ(dz)ds. (5.5)
t→∞ 2
0 Sj

We will next show how to control the integral on the right hand side of (5.5).
By applying Fubini’s theorem to the Chapman–Kolmogorov equation for the heat
kernel, we get for t, s > 0,

Px (T C > t + s) = pC (s, x, z)Pz (T C > t)dz.
C

Using the heat kernel of the exterior of a ball we get the upper bound ∂x pC (s, x, z) ≤
Ae−B|z| for s > 1 and all x ∈ C. We can apply the Dominated Convergence Theorem
2

to get for x ∈ S

C
∂n Px (T > t + s) = ∂n pC (s, x, z)Pz (T C > t)dz. (5.6)
C
1292 P. Collet et al. / Journal of Functional Analysis 270 (2016) 1269–1298

Moreover, using (5.2), it is easy to obtain the following limit by using again the Domi-
nated Convergence Theorem

lim tk ∂n Px (T > tk ) = μ ∂n pC (s, x, z)w(z)dz = μ∂n w(x).
C
κ/2
k
C

The last equality holds because w is harmonic.

As usual, we split the integral from (5.5) into:

t/2
I1 (t) = ∂n Pz (T C > t − s)pV (s, x, z)σ(dz)ds
0 S

 t/2
C
≤ CH ∂n Pz (T > t/2)σ(dz) pV (s + 1, x, ξ)ds.
S 0

This shows that limt→∞ tα/2 I1 (t) ≤ C1 GV (x, ξ), where C1 > 0 is universal.
Also, by the boundary Harnack inequality

t 
I2 (t) = ∂n Pz (T C > t − s)pV (s, x, z)σ(dz)ds
t/2 S

t 
≤ CH ∂n Pz (T C > t − s)σ(dz)ds pV (t/2, x, x)
t/2 S

t/2
−1−α
≤ Cx t ∂n Pz (T C > s)σ(dz)ds
0 S

where Cx only depends on x. Using bounds for the exit time for a Bessel process from
1
[1, ∞) as in Lemma 5.2, we get that 0 ∂n Pz (T C > s)ds ≤ Q, independently of z ∈ S.
Then
⎛ ⎞
 
t
tκ/2 I2 (t) ≤ Cx t−1−α+κ/2 ⎝Q |S| + −1 ∂n Pz (T C > 1)σ(dz)⎠ .
2
S

It follows that tκ/2 I2 (t) → 0 as t → ∞. Equation (5.5) now reads

γV v(x) ≤ μw(x) + C1 GV (ξ, x), x ∈ C.

Since GV (ξ, x) remains bounded as x → ∞ radially in C, we deduce that γV = μ, which

proves the asymptotic for the survival probability. 2
 P. Collet et al. / Journal of Functional Analysis 270 (2016) 1269–1298 1293

5.2. Proof of Theorem 1.2

In formula (5.3), the first term is controlled by our result from the previous section.
The second term goes to zero by using Harnack’s inequality up to the boundary, that is,
for some x0 ∈ Ω0 ,

Px (Bt ∈ Ω0 , T > t) = p(t, x, z)dz ≤ CH |Ω0 | p(t + 1, x, x0 ),
Ω0

where |Ω0 | stands for the Lebesgue measure of the core Ω0 . It follows that tκ/2 Px (Bt ∈
Ω0 , T > t) converges to zero as t → ∞ for each x ∈ Ω.
Next, we deal with the summation terms. In order to do this, we will find limits for
the following two objects:

t/2
t κ/2 κ/2
I1 (t) = t ∂n Pz (T j > t − s)p(s, x, z)σ(dz)ds,
0 Sj

t/2
t κ/2 κ/2
I2 (t) = t ∂n Pz (T j > s)p(t − s, x, z)σ(dz)ds.
0 Sj

An analogous proof as the one in the last part of the previous section, shows that tκ/2 I2 (t)
converges to zero.
As for tκ/2 I1 (t), if 0 ≤ s ≤ t/2, our computations in the previous section show that
t ∂n Pz (T j > t − s) converges to γj ∂n wj (z) for z ∈ Sj and j ∈ M, otherwise, it
κ/2

converges to zero. Here,

  
Γ κ+n
γj =  2 n  m1j (θ)σ(dθ).
2κ/2 Γ κ + 2
Sj

To show domination, we use monotonicity of t → ∂n Px (T j > t), equation (5.6), the

bound ∂n pC (1, x, z) ≤ Ae−B|z| , and equation (5.2). We find that
2


Ae−B|z| |z| dz < ∞.
2 κ
tκ/2 ∂n Px (T j > t − s) ≤ C3
C

Thus, by the Dominated Convergence Theorem, we deduce that

∞ 
κ/2
lim t I1 (t) = γj ∂n wj (z)p(s, x, z)σ(dz)ds
t→∞
0 Sj
1294 P. Collet et al. / Journal of Functional Analysis 270 (2016) 1269–1298


= γj ∂n wj (z)G(x, z)σ(dz) = 2γj (uj (x) − wj (x)) ,
Sj

by Fubini’s theorem, and equation (1.3).

Putting all together, for x ∈ Ω,

lim tκ/2 Px (T > t) = γk uk (x),
t→∞
k∈M

which is (1.7). 2

6. Renormalized Yaglom limit for multicones

In what follows, we set β = 1 + α + n/2. Notice that β/2 + κ/2 = 1 + α, that will be
conveniently used later.
From Theorem 2.6, it is straightforward to get that for x, y ∈ V = C(0, D, 0)

√ v(x)v(y) −|y|2 /2
lim tβ/2 pV (t, x, ty) = α e . (6.1)
t→∞ 2 Γ(1 + α)

The limit above holds uniformly in compact sets of V .

In order to extend this result to multicones, we start with the case of a truncated cone
C = C(0, D, 1).

Lemma 6.1. Let x ∈ C and y ∈ V . Then,

√ w(x)v(y) −|y|2 /2
lim tβ/2 pC (t, x, ty) = α e . (6.2)
t→∞ 2 Γ(1 + α)

Proof. As in the proof of Lemma 3.3, we have that t1+α ∂n pC (t, x, ξ) ≤ Qx for all t > t0 ,
ξ ∈ S. Thus, using the boundary Harnack inequality, there exist Cx > 0 only dependent
on x, such that

t/2
tβ/2 √ √
∂n pC (t − s, x, ξ)pV (s, ξ, ty)dξds ≤ Cx GV (x, ty)tβ/2−(1+α) . (6.3)
2
0 S

√
For large t, the quantity GV (x, ty) is bounded, and as β/2 − (1 + α) = −κ/2 < 0, we
deduce that

t/2
tβ/2 √
lim ∂n pC (t − s, x, ξ)pV (s, x, ty)dξds = 0. (6.4)
t→∞ 2
0 S
 P. Collet et al. / Journal of Functional Analysis 270 (2016) 1269–1298 1295

On the other hand, from Theorem 2.6, it is direct to find the bound
 
∞
√ (|x| |y|)α − 2 −1
i n

t p (t, x, ty) ≤ C
β/2 V
< ∞, (6.5)
i=1
2αi Γ(1 + αi )

for t > 2 and some universal constant C > 0. It follows by (6.1) and the Dominated
Convergence Theorem that

t/2
tβ/2 √
lim ∂n pC (s, x, ξ)pV (t − s, ξ, ty)dξds =
t→∞ 2
0 S
∞ 
v(y)e−|y| /2
2
1
= α ∂n pC (s, x, ξ)v(ξ)dξds
2 Γ(1 + α) 2
0 S

/2 v(y)Ex (v(BT C ))
= e−|y|
2
(6.6)
2α Γ(1 + α)

Plugging the last two equations into (2.5), with O = V and U = C, we obtain

√ √ v(y)Ex (v(BT C ))
lim tβ/2 pC (t, x, ty) = lim tβ/2 pV (t, x, ty) − e−|y| /2
2
,
t→∞ t→∞ 2α Γ(1 + α)

from where (6.2) is direct to deduce by using (1.2). 2

Lemma 6.2. We have for each y ∈ V

√
sup sup tβ/2 ∂n pC (t, ty, ξ) < ∞ (6.7)
t>2 ξ∈S

Also, for y ∈ C and ξ ∈ S,

√ ∂n w(ξ)v(y) −|y|2 /2
lim tβ/2 ∂n pC (t, ty, ξ) = α e . (6.8)
t→∞ 2 Γ(1 + α)

The limit holds in the sense of uniform convergence in compact sets.

Proof. Recall that, from the bound for the heat kernel of the exterior of a ball, and
monotonicity of domains

√  √
C
t β/2
∂n p (t + 1, ty, ξ) = t β/2
∂n pC (1, z, ξ)pC (t, ty, z)dz (6.9)
C
 √
Ae−B|z| pV (t, ty, z)dz
2
≤t β/2
(6.10)
C
1296 P. Collet et al. / Journal of Functional Analysis 270 (2016) 1269–1298

 ∞

n 
(|z| |y|)α − −1
i
2
−B|z|2
≤ AC e (6.11)
i=1
2αi Γ(1 + αi )
C

where the last inequality follows from (6.5). From here, using bounds for the moment of
gaussian random variables, we arrive at the bound:
n   
∞ α − 2 −1 i i

C
√ α +n/2
C7 |y| 1 + αi n
t β/2
∂n p (t + 1, ty, ξ) ≤ C6 Γ + , (6.12)
i=1
2αi /2 Γ(1 + αi ) 2 4

which is finite.
The same steps as above show that it is possible to apply the Dominated Convergence
Theorem in (6.9). Equation (6.8) then follows from Lemma 6.1. 2

6.1. Proof of Theorem 1.3

As before, our starting point is equation (2.5). We will study the rate of decay of the
integral involved in such equation by splitting into two terms, as before. First, let us
study

t/2
1 √
t β/2
I1 (t) = t β/2
∂n pj (t − s, aj + ty, ξ)p(s, x, ξ)dξds. (6.13)
2
0 Sj

Using Lemma 6.2, we see that we can apply the Dominated Convergence Theorem to
this integral to obtain

∞ 
1 ∂n wj (ξ)e−|y| /2 vj (y)
2
β/2
lim t I1 (t) = 1M (j) p(s, x, ξ)dξds
t→∞ 2 2α Γ(1 + α)
0 Sj

e−|y| /2 vj (y)
2
1
= 1M (j) α ∂n wj (ξ)G(x, ξ)dξds
2 Γ(1 + α) 2
Sj

−|y|2 /2
e vj (y)
= 1M (j) α
(uj (x) − wj (x)) . (6.14)
2 Γ(1 + α)

Second, we look at

t/2 √
1
t β/2
I2 (t) = t β/2
∂n pj (s, aj + ty, ξ)p(t − s, x, ξ)dξds
2
0 Sj

t/2
−κ/2 1 √
=t ∂n pj (s, aj + ty, ξ)t1+α p(t − s, x, ξ)dξds (6.15)
2
0 Sj
 P. Collet et al. / Journal of Functional Analysis 270 (2016) 1269–1298 1297

From Theorem 1.1, we have that t1+α p(t − s, x, ξ) ≤ Cx for all s ∈ [0, t/2] as long as
t > 3. The constant Cx depends only on |x|. Then

t/2 √
−κ/2 1
t I2 (t) ≤ Cx t
β
∂n pj (s, aj + ty, ξ)dξds
2
0 Sj

= Cx t−κ/2 P√ty (BT j ∈ Sj , T j < t/2) ≤ Cx t−κ/2 ,

which converges to zero.

Putting together equation (2.5), Lemma 6.1, and the last estimates, we obtain

√ uj (x)vj (y) −|y|2 /2

lim tβ/2 p(t, x, aj + ty) = 1M (j) e , (6.16)
t→∞ 2α Γ(1 + α)

as desired. 2

6.2. Distributional convergence of the renormalized process

Theorem 1.3 suggests that, when conditioned on survival, most of the trajectories of
√
Brownian motion at time t stay within order t from the origin. Thus, it is natural to
√
study the convergence of the rescaled process Bt / t conditioned on survival.
Let A ⊆ Vj be a precompact, Borel set. Notice that β − κ = n. Then, by a simple
change of variable
 √  
p(t, x, aj + z)
Px (Bt − aj )/ t ∈ A|T > t = dz
√
Px (T > t)
tA
 √
p(t, x, aj + ty) n/2
= t dy
Px (T > t)
A
 √
tβ/2 p(t, x, aj + ty)
= dy.
tκ/2 Px (T > t)
A

By Theorems 1.2 and 1.3, the integrand on the right hand side converges to the function

vj (y)e−|y| /2
2
γj uj (x)
px (j, y) = · 1Vj (y). (6.17)
γj 2α Γ(1 + α) k∈M γk uk (x)

Equation (6.17) defines a probability distribution function on M × ∪j∈M Vj for a family

of random variables X x = (X1x , X2x ), with x ∈ Ω, which is simple to interpret. Fix x ∈ Ω
and let X1x be a discrete random variable with distribution given by

γj uj (x)
P(X1x = j) =  , j ∈ M. (6.18)
k∈M γk uk (x)
1298 P. Collet et al. / Journal of Functional Analysis 270 (2016) 1269–1298

This is a sample of one of the maximal branches of the multicone. As t → ∞ the

multicone Ωj is rescaled into the cone with vertex Vj . Correspondingly, we define X2x as
a continuous random variable on ∪j∈M Vj satisfying

vj (y)e−|y| /2
2

P(X2x ∈ dy|X1x = j) = 1V (y). (6.19)

γj 2α Γ(1 + α) j

Our computation at the beginning of the section, and the uniform convergence on
√
compact sets show that, under Px , the renormalized process Bt / t conditioned on sur-
vival converges weakly to X x .

References

[1] Alano Ancona, On positive harmonic functions in cones and cylinders, Rev. Mat. Iberoam. 28 (1)
(2012) 201–230.
[2] Rodrigo Bañuelos, Burgess Davis, Heat kernel, eigenfunctions, and conditioned brownian motion in
planar domains, J. Funct. Anal. 84 (1) (1989) 188–200.
[3] Rodrigo Bañuelos, Robert G. Smits, Brownian motion in cones, Probab. Theory Related Fields
108 (3) (1997) 299–319.
[4] Pierre Collet, Servet Martínez, Jaime San Martín, Ratio limit theorems for a Brownian motion
killed at the boundary of a Benedicks domain, Ann. Probab. 27 (3) (1999) 1160–1182.
[5] Pierre Collet, Servet Martínez, Jaime San Martín, Asymptotic behaviour of a Brownian motion on
exterior domains, Probab. Theory Related Fields 116 (3) (2000) 303–316.
[6] Ioannis Karatzas, Steven E. Shreve, Brownian Motion and Stochastic Calculus, second edition,
Grad. Texts in Math., vol. 113, Springer-Verlag, New York, 1991.
[7] Shin-ichi Matsushita, Generalized laplacian and balayage theory, J. Inst. Polytech. Osaka City Univ.
Ser. A 8 (1) (1957) 57–90.
[8] Jürgen Moser, A harnack inequality for parabolic differential equations, Comm. Pure Appl. Math.
17 (1) (1964) 101–134.
[9] Yehuda Pinchover, Large time behavior of the heat kernel and the behavior of the green function
near criticality for nonsymmetric elliptic operators, J. Funct. Anal. 104 (1) (1992) 54–70.
[10] Ross G. Pinsky, The lifetimes of conditioned diffusion processes, Ann. Inst. Henri Poincaré Probab.
Stat. 26 (1) (1990) 87–99.
[11] Ross G. Pinsky, Positive Harmonic Functions and Diffusion, Cambridge Stud. Adv. Math., Cam-
bridge University Press, Cambridge, UK, New York, 1995.
[12] Sandro Salsa, Some properties of nonnegative solutions of parabolic differential operators, Ann.
Mat. Pura Appl. 128 (1) (1981) 193–206.