TEST FUNCTION APPROACH TO FULLY NONLINEAR

EQUATIONS IN THIN DOMAINS

arXiv:2404.19577v1 [math.AP] 30 Apr 2024

ISABEAU BIRINDELLI, ARIELA BRIANI, AND HITOSHI ISHII

Abstract. In this note we extend to fully nonlinear operators the well known
result on thin domains of Hale and Raugel [7]. The result is more general even
in the case of the Laplacian.

Contents
1. Introduction 1
2. Preliminaries 4
3. Convergence results 8
3.1. Relaxed limits 8
3.2. Uniform convergence 11
References 14

1. Introduction
The classical result of Hale and Raugel [7] in thin domains states that if uε are
solutions of

 −∆uε + uε = f (x, y) in Ωε
∂uε
=0 on ∂Ωε
∂νε


2020 Mathematics Subject Classification. 35B40, 35D40, 35J25 49L25 .

Key words and phrases. asymptotic behavior of solutions, thin domains.
A. Briani was partially supported by l’Agence Nationale de la Recherche (ANR), project ANR-
22-CE40-0010 COSS; I. Birindelli was partially supported by project Leoni 2023 GNAMPA-
INDAM and project ”At the Edge of Reaction-diffusion equations” Sapienza Università di Roma.
H. Ishii was partially supported by the JSPS KAKENHI Grant Nos. JP20K03688, JP20H01817
and JP21H00717. The project was very much advanced while I. Birindelli was visiting prof. Ishii
in Tsuda University, she wished to thank the Institution for the invitation.
1
2 I. BIRINDELLI, A. BRIANI, AND H. ISHII

where Ωε = {(x, y) ∈ RN × R : x ∈ Ω, 0 < y < εg(x)}, for some g ∈ C 3 (Ω) such

that 0 < inf g ≤ sup g < ∞ then uε converges to uo solution of
Ω Ω

 −(∆uo + Dg·Du g
o
) + uo = f (x, 0) in Ω
∂u0
 =0 on ∂Ω.
∂ν
This result has been extended in a wide variety of related problems see e.g. the
works of Arrieta, Pereira, Raugel [1, 2, 10]. But all the above results concern
variational problems, where the appearance of the first order seems to come from
a typical integration by part, related to the variational nature of the problem.
In this paper, instead, we treat fully nonlinear equation in thin domains i.e.
where the equation is given by
F (D 2u, Du, u, (x, y)) = 0 in Ωε
where F : S(N + 1) × RN +1 × R × Ωε → R is a proper functional in the sense of
the User’s guide [4]. Of course the solutions are viscosity solutions and the proof
follows the test function approach of Evans [5] which is somehow more direct and
completely different from the papers mentioned above. Furthermore the technique
does not require the operator to be uniformly elliptic as it will be evident from the
hypotheses below. An example of thin domains for degenerate elliptic operator
will be given explicitly below.
Even though the results will be proved for a large class of operators, in this
introduction, we will illustrate the special case where the fully non linear operator
is one of the extremal Pucci operators e.g. for 0 < λ ≤ Λ
X X
M+ λ,Λ (D 2
u) := sup (tr A(D 2
u)) := λ ei + Λ ei ,
λI≤A≤ΛI
ei ≤0 ei ≥0

where ei = ei (D 2 u) denotes the i-th eigenvalue of the Hessian matrix D 2 u.

Under the hypothesis
(H1) g ∈ C 1 (Ω) and 0 < inf g ≤ sup g < ∞,
Ω Ω
we will prove that uε , the solutions of

 −M+ 2
λ,Λ (D uε ) + uε = f (x, y) in Ωε
(1) ∂uε
=0 on ∂Ωε
∂νε


converge uniformly to uo solution of

(2)
  +  −
 −M+ 2 Dg·Duo Dg·Duo
λ,Λ (D u o (x)) − Λ g
+ λ g
+ uo (x) = f (x, 0) in Ω
 ∂uo = 0 on ∂Ω.
∂ν
 TEST FUNCTION APPROACH TO FULLY NONLINEAR EQUATIONS IN THIN DOMAINS 3

Of course in the first equation M+ λ,Λ acts on matrices in S(N + 1) while, in the
second equation, it acts on matrices in S(N).
In the special case λ = Λ = 1, when M+ λ,Λ = ∆, we recover Hale and Raugel
result, but we improve the condition on g that is only required to be the natural
condition C 1 and not C 3 .
We wish to explain the heuristic behind the formal proof which will be given
in this paper, for a much larger class of operators. Let uε be a solution of (1) and
let
vε (x, y) := uε (x, εg(x)y)
so that we have ”flattened” the top boundary. In similarity with the linear varia-
tional case we can suppose that there exists a constant C such that
|∂yy vε | ≤ Cε2 .
This in turn implies that for ε → 0, ∂yy vε → 0 and then, using the boundary
condition, we get
vε (x, y) → vo (x).
On the other hand, the above estimates implies also that, for some function k(x),
we get that
∂yy vε (x, y)
→ k(x) := g 2(x)h(x).
ε2
So we may use the following ansatz :
y2
vε (x, y) = w(x) + ε2 k(x) + o(ε2 ).
2
Substituting the ansatz in the equation, we let, formally ε go to zero, after a
tedious but simple computation it is easy to see that we obtain
 2 
+ D w(x) 0
−Mλ,Λ ( ) + w(x) = f (x, 0)
0 h(x)
We use the condition on the ”top” boundary, in order to determine h(x):
1
Dg(x) · [Dw(x) + ε2 D(g 2h)(x) ] = g(x)h(x)[1 + ε2 Dg(x)].
2
Passing to the limit we find
Dg(x) · Dw(x)
h(x) =
g
i.e. the limit equation becomes:
 2
D w(x) 0

+
−Mλ,Λ ( Dg(x)·Dw(x) ) + w(x) = f (x, 0).
0 g(x)
In the rest of the paper we will treat the general case and make rigorous the above
idea.
4 I. BIRINDELLI, A. BRIANI, AND H. ISHII

We will first give some a priori bounds, which allow to prove that the upper
and lower relaxed limits u+ and u− of {uε }ε∈(0,ε0 ] are respectively sub and super
solutions of the limit equation:
  2  
D w 0
F , (Dw, 0), w, (x, 0) = 0 in Ω,


0 Dg · Dw/g
(3)
 ∂w = 0 on ∂Ω.


∂ν
Under the further condition that the comparison principle holds, we will prove
that the convergence of uε to the solution of (3) is uniform.
The scope of this paper is to open the results in thin domains from a perspective
that is, to our knowledge completely different from the previous results, see e.g.
[1, 2, 7, 10]. So, in order to make the exposition the clearer possible, we have
decided to concentrate on the more classical case i.e. domains like Ωε that in
one direction have a one flat boundary and with a Neumann boundary condition.
We plan to investigate in a further research, more generale domains, for examples
with jumps or with non flat sides, or thin domains that also have an oscillatory
boundary. We hope the reader will appreciate this choice.

2. Preliminaries
Let F : S(N + 1) × RN +1 × R × Ωε → R be a proper functional in the sense of
the User’s guide [4], i.e.
(
F ∈ C(S(N + 1) × RN +1 × R × Ωε , R),
(H2)
F (X, p, r, (x, y)) ≤ F (Y, p, s, (x, y)) whenever r ≤ s, and Y ≤ X.
Furthermore, for simplicity of the presentation, we strengthen the monotonicity
condition on F in the above as follows.
(H3) There exists α > 0 such that
α(r − s) ≤ F (X, p, r, (x, y)) − F (X, p, s, (x, y))
for r ≥ s and (X, p, (x, y)) ∈ S(N + 1) × RN +1 × Ωε .
Our PDE problem is:
∂uε
(4) F (D 2 uε , Duε , uε , (x, y)) = 0 in Ωε and = 0 on ∂Ωε ,
∂νε
where νε denotes the outward (unit) normal to Ωε .
Since our concern is the asymptotic behavior of solutions uε to (4), we will
restrict ourself to the parameter ε in the range (0, ε0], where ε0 > 0 is a number
fixed throughout (see the comment after Proposition 1). Let us note that we shall
use ν to indicate the normal to Ω and νε for Ωε . Obviously, the assumptions above
are not enough to ensure the existence of viscosity solutions in the sense of the
User’s guide [4] to (4). To keep the generality of the assumptions made above, we
 TEST FUNCTION APPROACH TO FULLY NONLINEAR EQUATIONS IN THIN DOMAINS 5

consider the notion of viscosity solutions to (4) which eliminates the continuity
requirement. That is, we call a bounded function u on Ωε a (viscosity) solution
of (4) if its upper and lower semicontinuous envelopes are viscosity sub and super
solutions, in the sense of [4], to (4), respectively.
We assume throughout that
(H4) Ω is a bounded C 1 domain of RN .
Accordingly, we may choose a function ρ ∈ C 1 (RN ) so that
(5) ρ(x) < 0 for x ∈ Ω, Dρ(x) 6= 0, and ρ(x) > 0 for x ∈ RN \ Ω.
Note that the outward unit normal ν to Ω at x ∈ ∂Ω is given by ν = |Dρ(x)|−1 Dρ(x).
The domain Ωε has corners, where the N–dimensional hypersurface ∂Ω × R inter-
sects either the hypersurfaces y = g(x) or y = 0, respectively.
We denote by ∂L Ωε , ∂B Ωε , and ∂T Ωε the lateral, bottom, and top portions of the
boundary ∂Ωε , which are described respectively as
{(x, y) ∈ ∂Ωε : x ∈ ∂Ω}, {(x, y) ∈ ∂Ωε : y = 0}, and
{(x, y) ∈ ∂Ωε : y = εg(x)}.
Furthermore, the outward unit normal to Ωε at the lateral boundary, at the
bottom, y = 0, and at the top boundary, y = εg(x), is given, respectively, by
νL = (|Dρ(x)|−1 Dρ(x), 0), νB = −eN +1 = −(0, . . . , 0, 1), and
(−εDg(x), 1)
νT = p .
1 + ε2 |Dg(x)|2
The appearance of corners of the domain Ωε requires a little care in the definition
of sub and super solutions to (4). For instance, when u is a bounded upper
semicontinuous function on Ωε , we call u a viscosity subsolution of (4) if the
following condition holds: whenever φ ∈ C 2 (Ωε ), ẑ = (x̂, ŷ) ∈ Ωε and maxΩε (u −
φ) = (u − φ)(ẑ), we have
(6) F (D 2 φ(ẑ), Dφ(ẑ), u(ẑ), ẑ) ≤ 0
if ẑ ∈ Ωε ; we have either (6) or
(7) νL · Dφ(ẑ) ≤ 0
if ẑ ∈ ∂L Ωε \ (∂B Ωε ∪ ∂T Ωε ); we have either (6) or
(8) νB · Dφ(ẑ) ≤ 0
if ẑ ∈ ∂B Ωε \ ∂L Ωε ; we have either (6) or
(9) νT · Dφ(ẑ) ≤ 0
if ẑ ∈ ∂T Ωε \ ∂L Ωε ; we have either (6), (7), or (8) if ẑ ∈ ∂L Ωε ∩ ∂B Ωε . Finally we
have either (6), (7), or (9) if ẑ ∈ ∂L Ωε ∩ ∂T Ωε . Replacing “max” and “≤” with
“min” and “≥”, respectively, in the above condition yields the right definition of
viscosity supersolution.
6 I. BIRINDELLI, A. BRIANI, AND H. ISHII

Thanks to (H2), we can fix a constant C0 > 0 so that

|F (0, 0, 0, (x, y))| ≤ C0 for (x, y) ∈ Ωε0 .
Proposition 1. Assume that (H1)–(H4) hold.
(1) The constant functions α−1 C0 and −α−1 C0 are classical super and sub
solutions to (4), respectively.
(2) There is a viscosity solution to (4).
(3) There is a constant ε0 > 0 such that if 0 < ε ≤ ε0 , then any viscosity
solution u to (4) satisfies supΩε |u| ≤ α−1 C0 .
Henceforth, we will assume that ε0 is small enough so that if 0 < ε ≤ ε0 , then
supΩε |u| ≤ α−1 C0 for any solution u to (4).

Proof. (1) Set u(z) = α−1 C0 for z ∈ Ωε . It is clear that νL ·Du(z) = 0 for z ∈ ∂L Ωε ,
νB · Du(z) = 0 for z ∈ ∂B Ωε , and νT · Dφ(z) = 0 for z ∈ ∂T Ωε . It follows that
F (D 2 u(z), Du(z), u(z), z) = F (0, 0, u(z), z) ≥ F (0, 0, 0, z) + αu(z)
≥ −C0 + C0 = 0 for z ∈ Ωε .

Thus, the constant function α−1 C0 is a classical supersolution of (4). Similarly,

the constant function −α−1 C0 is a classical subsolution of (4).
(2) Thanks to (H2), the constant functions α−1 C0 and −α−1 C0 are viscosity sub
and super solutions of (4), respectively. Hence, the Perron method readily yields
a solution to (4). Indeed, if we set
u(z) = sup{v(z) : v is a viscosity subsolution of (4),
− α−1 C0 ≤ v ≤ α−1 C0 on Ωε } for z ∈ Ωε ,
then the function u is a viscosity solution to (4).
(3) We claim that if ε0 is sufficiently small, then there is a function ψ ∈ C 2 (Ωε )
such that

ν = νL and z ∈ ∂L Ωε ,


(10) ν · Dψ(z) > 0 for ν = νB and z ∈ ∂B Ωε ,


ν = ν and z ∈ ∂ Ω .
T T ε

We will come back to the proof of (10) later and now assume the existence of such
a ψ, and complete the proof of (3). Let u be any viscosity solution of (4). Let v and
w be the upper and lower semicontinuous envelopes of u on Ωε , respectively. Fix
any δ > 0. We prove by contradiction that v ≤ α−1 C0 on Ωε . Thus, we suppose
that maxΩε v > α−1 C0 . Choosing positive constants δ and γ small enough, we
have maxΩε (v − γψ) > δ + α−1 C0 . Set φ = γψ + δ + α−1 C0 on Ωε . Let ẑ ∈ Ωε be
 TEST FUNCTION APPROACH TO FULLY NONLINEAR EQUATIONS IN THIN DOMAINS 7

a maximum point of the function v − φ. Noting that (10) holds with φ in place of
ψ, we find by the subsolution property of v that
0 ≥ F (D 2 φ(ẑ), Dφ(ẑ), v(ẑ), ẑ)
≥ F (γD 2 ψ(ẑ), γDψ(ẑ), γψ(ẑ) + α−1 C0 , ẑ) + αδ.
Sending γ → 0+ , we obtain F (0, 0, α−1C0 , z̃) + αδ ≤ 0 for some z̃ ∈ Ωε , which
contradicts that α−1 C0 is a classical supersolution of (4). Hence, we conclude that
u ≤ v ≤ α−1 C0 on Ωε . A parallel argument ensures that u ≥ w ≥ −α−1 C0 on Ωε .
It remains to prove that there is ε0 > 0 such that for each ε ∈ (0, ε0 ], there
exists ψ = ψε ∈ C 2 (Ωε ) for which (10) holds. Fix any ε > 0. Let ρ ∈ C 1 (RN ) be
a function satisfying (5). Choose a function η ∈ C 1 (R) such that
η(r) = 0 for r ≤ −g0 and 0 < η ′ (r) ≤ 1 for r > −g0 ,
where g0 := inf Ω g > 0. For a positive constant δ, we define ψ = ψε on RN +1 by
setting
  y  y − εg(x) 
ψ(x, y) = δρ(x) + ε η − + η for (x, y) ∈ RN × R.
ε ε
If z = (x, y) ∈ ∂L Ωε , then
  y
νL · Dψ(z) = (|Dρ|−1 Dρ, 0) · δ(Dρ, 0) − η ′ − eN +1
ε
′ y − εg(x)
  
+η (−εDg(x), 1) ≥ δ|Dρ(x)| − ε0 |Dg(x)|.
ε
Similarly, if z = (x, 0) ∈ ∂B Ωε , then
νB · Dψ(z) = −eN +1 · Dψ(z) = η ′ (0) − η ′ (−g(x)) = η ′ (0),
and if z = (x, y) ∈ ∂T Ωε , then
1 
′ ′ 2 2

νT · Dψ(z) ≥ p − δε|Dρ||Dg| − η (−g(x)) + η (0)(ε |Dg(x)| + 1))
ε2 |Dg|2 + 1
1  
≥p − δε|Dρ||Dg| + η ′ (0)(ε2 |Dg(x)|2 + 1))
ε2 |Dg|2 + 1
≥ −δ|Dρ(x)| + η ′ (0).
We select δ > 0 and ε0 > 0 sufficiently small so that
η ′ (0) > δ max |Dρ| and δ min |Dρ| > ε0 max |Dg|,
Ω ∂Ω ∂Ω

which assures that (10) holds if ε ∈ (0, ε0]. At this point, we only have the C 1 –
regularity of ψ, but the standard mollification procedure provides a C 2 –function
ψ which satisfies (10) as far as ε ∈ (0, ε0 ]. 
8 I. BIRINDELLI, A. BRIANI, AND H. ISHII

3. Convergence results
3.1. Relaxed limits. Let uε be a solution of (4). By the choice of ε0 (see also
Proposition 1), we have
C0
kuε k∞ ≤ .
α
This allows us to define the upper and lower relaxed limits u+ and u− of {uε }ε∈(0,ε0 ] :
 +
 u (x) = lim+ sup{uε (ξ, η) : (ξ, η) ∈ Ωε , |ξ − x| < r, 0 < ε < r},
r→0
(11)
 u (x) = lim inf{uε (ξ, η) : (ξ, η) ∈ Ωε , |ξ − x| < r, 0 < ε < r}.
−
r→0+

It follows that u ≥ u− on Ω and u+ , −u− ∈ USC(Ω). The limit equation will be

+
  2  
D w 0
F , (Dw, 0), w, (x, 0) = 0 in Ω,


0 Dg · Dw/g
(12)
 ∂w = 0 on ∂Ω.


∂ν
Theorem 2. Suppose that (H1)–(H4) hold. The functions u+ and u− are, respec-
tively, sub and super solutions of (12).
Proof. We treat only the subsolution property. By replacing uε by its upper semi-
continuous envelope, we may assume that uε is upper semicontinuous on Ωε . Let
φ ∈ C 2 (Ω) and assume that for some x̂ ∈ Ω,
(u+ − φ)(x) < (u+ − φ)(x̂) if x 6= x̂.
In the following computation, we fix δ > 0 arbitrarily. We choose hδ ∈ C 2 (Ω)
so that  
Dg · Dφ
(x) − hδ (x) < δ for x ∈ Ω.
g
We set
1
ψδ± (x, y) = y 2 (±2δ + hδ (x)) ,
2
and consider the function
Φ(x, y) = φ(x) + ψδ+ (x, y) + γε2 ζ(y/ε),
where ζ ∈ C 2 (R) is a bounded function on R having the properties
−1 < ζ ′(0) < 0 < ζ ′ (y) < 1 for y ≥ min g and |ζ ′′ (y)| < 1 for y ∈ [0, max g],
and γ > 0.
We choose a maximum point (x̄, ȳ) = (x̄(ε, γ), ȳ(ε, γ)) of the function uε − Φ on
Ωε . We are to take the limit ε → 0+ . In our limit process as ε → 0+ , the choice
of γ depends on ε in such a way that lim γ/ε = 0. A possible choice is γ = ε2 . It
is a standard observation (see Remark 3 below) that as ε → 0+ ,
(13) (x̄, ȳ) → (x̂, 0) and uε (x̄, ȳ) → u+ (x̂).
 TEST FUNCTION APPROACH TO FULLY NONLINEAR EQUATIONS IN THIN DOMAINS 9

Since uε is a subsolution of (1), if

(i) (x̄, ȳ) ∈ Ωε ,
then we have
(14) F (D 2Φ(x̄, ȳ), DΦ(x̄, ȳ), uε (x̄, ȳ), (x̄, ȳ)) ≤ 0;
if
(ii) (x̄, ȳ) ∈ ∂T Ωε \ ∂L Ωε ,
then we have either (14) or
(15) −εDg(x̄) · Dx Φ(x̄, εg(x̄)) + Φy (x̄, εg(x̄)) ≤ 0;
if
(iii) (x̄, ȳ) ∈ ∂B Ωε \ ∂L Ωε ,
then we have either (14) or
(16) −Φy (x̄, 0) ≤ 0;
if
(iv) (x̄, ȳ) ∈ ∂L Ωε \ (∂T Ωε ∪ ∂B Ωε ),
then we have either (14) or
∂Φ(x̄, ȳ)
(17) = νL · DΦ(x̄, ȳ) ≤ 0;
∂νε
if
(v) (x̄, ȳ) ∈ ∂T Ωε ∩ ∂L Ωε ,
then we have either (14), (15), or (17); if
(vi) (x̄, ȳ) ∈ ∂B Ωε ∩ ∂L Ωε ,
then we have either (14), (16), or (17).
Observe that
y2 y
Dx Φ = Dφ(x) + Dhδ (x), Φy = y (2δ + hδ (x)) + εγζ ′ ,
2 ε
2 2 y2 2 ′′ y
 
Dx Φ = D φ(x) + D hδ (x), Φyy = 2δ + hδ (x) + γζ ,
2 ε
Φxi y = Φyxi = y (hδ )xi (x).
Inequalities (14), (15), (16), and (17), can be written, respectively, as
(18) F (X̄, p̄, uε (x̄, ȳ), (x̄, ȳ)) ≤ 0,
10 I. BIRINDELLI, A. BRIANI, AND H. ISHII

where  
2 ȳ 2 2
D φ(x̄) + 2 D hδ (x̄) ȳDhδ (x̄)
X̄ =   ȳ  ,

ȳDhδ (x̄)T 2δ + hδ (x̄) + γζ ′′
ε
2
ȳ  ȳ 
and p̄ = (Dφ(x̄) + Dhδ (x̄), ȳ (2δ + hδ (x̄)) + εγζ ′ ),
2 ε
ε2 g(x̄)2
 
−Dg(x̄)· Dφ(x̄) + Dhδ (x̄)
(19) 2
+ g(x̄) (2δ + hδ (x̄)) + γζ ′ (g(x̄)) ≤ 0,

(20) ζ ′ (0) ≥ 0,

∂Φ(x̄, ȳ)
(21) = (|Dρ|−1 Dρ(x̄), 0) · DΦ(x̄, ȳ) ≤ 0.
∂νL
Choosing ε > 0 small enough, we may assume that
ε2 g(x̄)
δ≥ Dg(x̄) · Dhδ (x̄).
2
If (19) holds, then we have
 ε2 g(x̄)2 
0 ≥ −Dg(x̄) · Dφ(x̄) + Dhδ (x̄)
  2 
Dg · Dφ
+ g(x̄) δ + (x̄) + γζ ′(g(x̄))
g
′
≥ γζ (g(x̄)).
This contradicts our choice of ζ, and also (20) is a contradiction.
Thus, we have (18) in the case when either (i), (ii), or (iii) is valid, and we have
either (18) or (21) in the cases when either (iv), (v), or (vi) holds.
Sending ε → 0+ , we have
 2 
 2  D φ(x̂) 0
D φ(x̂) 0
X̄ → ≤ Dg(x̂) · Dφ(x̂) 
0 2δ + hδ (x̂) 0 3δ +
g(x̂)
and
p̄ → (Dφ(x̂), 0).
Therefore, we see that if x̂ ∈ Ω, then we have
 2  
D φ(x̂) 0
(22) F  Dg(x̂)Dφ(x̂)  , (Dφ(x̂), 0), u+(x̂), (x̂, 0) ≤ 0,
0 3δ +
g(x̂)
 TEST FUNCTION APPROACH TO FULLY NONLINEAR EQUATIONS IN THIN DOMAINS11

if x̂ ∈ ∂Ω, then we have either (22) or

∂φ(x̂)
νL · (Dφ(x̂), 0) = ≤ 0.
∂ν
This guarantees that u+ is a subsolution of (12).
A remark on the proof of the supersolution property of u− is that, in this case,
one should use the perturbed test function
y 
Φ(x, y) = φ(x) + ψδ− (x, y) − γε2 ζ . 
ε
Remark 3. For a general approach to the proof of (13), we may refer to the User’s
guide [4]. Here, for the reader’s convenience, we give a straightforward proof of
(13). By the definition of u+ (x̂), we may choose {(εj , xj , yj )}j∈N so that
1 1
εj → 0+ , (xj , yj ) ∈ Ωεj , |xj − x̂| < , u+ (x̂) < + uεj (xj , yj ).
j j
Since Φ depends on ε, we write Φε for Φ. Also, we write (x̄j , ȳj ) for (x̄, ȳ) with
ε = εj . Thus, (x̄j , ȳj ) is a maximum point of uεj − Φεj , and we have
1
(uεj − Φεj )(x̄j , ȳj ) ≥ (uεj − Φεj )(xj , yj ) > − + u+ (x̂) − Φεj (xj , yj ).
j
We may assume after passing to a subsequence that for some x̃ ∈ Ω and ũ ∈ R,
lim(x̄j , ȳj ) = (x̃, 0) and lim uεj (x̄j , ȳj ) = ũ.
Since (εj , x̄j , ȳj ) → (0, x̃, 0), we see, by the definition of u+ (x̃), that
u+ (x̃) ≥ lim uεj (x̄j , ȳj ) = ũ.
All the above together, we see in the limit as j → ∞ that
u+ (x̃) − Φ0 (x̃, 0) ≥ ũ − Φ0 (x̃, 0) ≥ u+ (x̂) − Φ0 (x̂, 0),
where Φ0 (x, y) := limε→0+ Φε (x, y) = φ(x) + ψδ− (x, y), that is,
(u+ − φ)(x̃) ≥ ũ − φ(x̃) ≥ (u+ − φ)(x̂),
which shows that x̃ = x̂ and ũ = u+ (x̂). 
3.2. Uniform convergence. Let F , Ω, and g be as in the previous section.
Define the function G ∈ C(S(N) × RN × R × Ω, R) by
 
X 0

(23) G(X, p, r, x) = F , (p, 0), r, (x, 0) .
0 Dg(x) · p/g(x)
Recall that the limit equation (12) for u is stated as
∂u
(24) G(D 2 u, Du, u, x) = 0 in Ω and = 0 on ∂Ω.
∂ν
A convenient assumption for Theorem 4 to draw a uniform convergence result
is the validity of the comparison principle for (24):
12 I. BIRINDELLI, A. BRIANI, AND H. ISHII

(H5) If v and w are viscosity sub and super solutions to (24), respectively, then
v ≤ w on Ω.
Indeed, we have
Theorem 4. Assume (H1)–(H5). Let uε be a viscosity solution to (4) for ε ∈
(0, ε0 ]. Then, for the unique continuous viscosity solution u0 of (24), we have
(25) lim+ max |uε (x, y) − u0 (x)| = 0.
ε→0 (x,y)∈Ωε

Proof. The following argument is standard in the asymptotic analysis based on

the half-relaxed limits, but we here present it for the reader’s convenience. Let u+
and u− be the functions defined by (11). By the definition, we have u− ≤ u+ on
Ω and u+ , −u− ∈ USC(Ω). Theorem 2 ensures that u+ and u− are viscosity sub
and super solutions to (24), respectively. Furthermore, (H5) assures that u+ ≤ u−
on Ω. Hence, we see that u+ = u− on Ω, which readily shows that u+ = u− is
continuous on Ω. Writing u0 for u+ = u− , we find that u0 is a continuous viscosity
solution to (24).
To check (25), fix any δ > 0. By the definition of u+ , for any x ∈ Ω, we select
r = r(δ, x) > 0 so that
uε (ξ, η) < u0 (x) + δ if 0 < ε < r, (ξ, η) ∈ Ωε , and |ξ − x| < r.
Reselecting r > 0 sufficiently smaller, we may assume that u0 (x) < u0 (ξ) + δ if
ξ ∈ Ω and |ξ − x| < r. Now, the above inequality can be stated as
(26) uε (ξ, η) < u0 (ξ) + 2δ if 0 < ε < r, (ξ, η) ∈ Ωε , and |ξ − x| < r.
Since Ω is compact, we can choose a finite number of balls, B1 , . . . , Bm , which
cover Ω, such that for every j ∈ {1, . . . , m}, if xj and rj denote, respectively, the
center and radius of Bj , then (26), with (xj , rj ) in place of (x, r), holds. Setting
r0 = min{rj : j = 1, . . . , m}, we find that
uε (ξ, η) < u0 (ξ) + 2δ for (ξ, η) ∈ Ωε and 0 < ε < r0 .
An argument parallel to the above yields, after replacing r0 > 0 by a smaller one
if necessary,
uε (ξ, η) > u0 (ξ) − 2δ for (ξ, η) ∈ Ωε and 0 < ε < r0 ,
which completes the proof of (25). 
Let us recall that there are a number of contests where the comparison principle
(H5) holds. In particular, when dealing with Neumann boundary conditions, one
can refer to the results of Hitoshi Ishii [8], Guy Barles [3] and Stefania Patrizi [9].
We consider here the general comparison principle given in [4, Theorem 7.5].
This leads us to assume, further hypotheses on the domain Ω and the operator G.
 TEST FUNCTION APPROACH TO FULLY NONLINEAR EQUATIONS IN THIN DOMAINS13

On Ω, in addition to (H4), we need the uniform exterior sphere condition, i.e.

that there is a constant r0 > 0 such that
(27) Br0 (x + r0 ν(x)) ∩ Ω = ∅ for x ∈ ∂Ω,
where Br (x) denotes the open ball {y ∈ RN : |y − x| < r}. On the function G a
crucial and typical hypothesis is the following:


 There is a function ω : [0, ∞) → [0, ∞) that satisfies ω(0+ ) = 0 such



 that

G(Y, p, r, y) − G(X, p, r, x) ≤ ω(γ|x − y|2 + |x − y|(|p| + 1))


(28)

 whenever γ > 0, p ∈ RN , x, y ∈ Ω, r ∈ R, and X, Y ∈ S(N) satisfy

    

 X 0 I N −I N

 −3γI2N ≤ ≤ 3γ .
 0 −Y −IN IN
Here Im denotes the identity matrix of order m. We impose another continuity
condition on G, which states:

There is a neighborhood V of ∂Ω, relative to Ω, such that

(29) G(X, p, r, x) − G(Y, q, r, x) ≤ ω(kX − Y k + |p − q|)

for X, Y ∈ S(N), p, q ∈ RN , r ∈ R, and x ∈ V .


Note that if (H3) holds, then

(30) α(r − s) ≤ G(X, p, r, x) − G(X, p, s, x)
for r ≥ s and (X, p, x) ∈ S(N) × RN × Ω.
The next proposition is a direct consequence of [4, Theorem 7.5] and Theorem
4.
Proposition 5. Assume (H1)–(H4) and (27)–(29). Then (H5) is satisfied and
the uniform convergence (25) as in Theorem 4 is valid.
Before concluding our discussion, we present two important examples of equa-
tions to which Theorem 4 applies, one is fully nonlinear and the other is linear
but degenerate elliptic.
Example 6. We apply Proposition 5 to show the uniform convergence result for
the solution of equation (1), involving the extremal Pucci operator as presented in
the Introduction. The extremal Pucci operator −M+ λ,Λ (X) has the property (28).
Indeed, the matrix inequality on the right-hand side of (28) implies that X ≤ Y
and hence, −M+ +
λ,Λ (Y ) + Mλ,Λ (X) ≤ 0. If the regularity of g is strengthened so
that g ∈ C 1,1 (Ω), then both the functions
 ±
Dg(x) · p
H(p, x) =
g(x)
14 I. BIRINDELLI, A. BRIANI, AND H. ISHII

satisfy
|H(p, y) − H(p, x)| ≤ C|x − y||p|
for all p ∈ RN , x, y ∈ Ω and some constant C > 0. It is then obvious to see that
the operator
 +  −
+ Dg(x) · p Dg(x) · p
G(X, p, r, x) = −Mλ,Λ (X) − Λ +λ + αr − f (x, 0),
g(x) g(x)
where f ∈ C(Ωε0 ), satisfies (28). Thus, thanks to Proposition 5, we find that
the uniform convergence (25) for the solution uε to (1), as in Theorem 4, holds,
provided that α > 0, (H1), g ∈ C 1,1 (Ω), f ∈ C(Ωε0 ), (H4), and (27) are satisfied.
Of course the case of the Laplacian is recovered just by considering λ = Λ = 1.
Example 7. In these examples we concentrate on simple degenerate elliptic equa-
tions in order to emphasize how the nature of the limit equation depends on the
direction of the diffusion. Let uε be the solution of
∂ 2 uε ∂uε
− + uε = f (x, y) in Ωε , = 0 on ∂Ωε .
∂y 2 ∂νε
If g ∈ C 1,1 (Ω) and f ∈ C(Ωε0 ), then we are under the hypothesis of Proposition
5, therefore uε converges uniformly to uo solution of a first order equation precisely:
Dg · Duo ∂uo
− + uo = f (x, 0) in Ω, = 0 on ∂Ω.
g ∂ν
Instead, if uε is the solution of
∂ 2 uε ∂uε
− + uε = f (x, y) in Ωε , on ∂Ωε
∂x21 ∂νε
it will converge to uo solution of the second order equation
∂ 2 u0 ∂uo
− + uo = f (x, 0) in Ω, = 0 on ∂Ω.
∂x21 ∂ν

References
[1] Arrieta, José M.; Pereira, Marcone C. Homogenization in a thin domain with an oscillatory
boundary J. Math. Pures Appl. (9) 96 (2011), no. 1, 29–57.
[2] Arrieta, José M.; Nogueira, Ariadne; Pereira, Marcone C. Semilinear elliptic equations in
thin regions with terms concentrating on oscillatory boundaries Comput. Math. Appl. 77
(2019), no. 2, 536–554.
[3] Barles, Guy Fully nonlinear Neumann type boundary conditions for second-order elliptic
and parabolic equations J. Differential Equations 106 (1993), no. 1, 90–106.
[4] Crandall, Michael G.; Ishii, Hitoshi; Lions, Pierre-Louis User’s guide to viscosity solutions
of second order partial differential equations, Bull. Amer. Math. Soc. (N.S.) 27 (1992), no.
1, 1–67.
 TEST FUNCTION APPROACH TO FULLY NONLINEAR EQUATIONS IN THIN DOMAINS15

[5] Evans, Lawerence C. The perturbed test function method for viscosity solutions of nonlinear
PDE. Proc. of the Royal Society of Edinburgh: Section A Mathematics. (1989);111(3-4),
359-375.
[6] Gilbarg, David; Trudinger, Neil S. , Elliptic partial differential equations of second order
Classics Math. Springer-Verlag, Berlin, 2001, xiv+517 pp.
[7] Hale, Jack K.; Raugel, Geneviève Reaction-diffusion equation on thin domains, J. Math.
Pures Appl. (9) 71 (1992), no. 1, 33–95.
[8] Ishii, Hitoshi Fully nonlinear oblique derivative problems for nonlinear second-order elliptic
pde’s, Duke Mathematical Journal, 62 (1991), no. 3, 633-661.
[9] Patrizi, Stefania The Neumann problem for singular fully nonlinear operator, Jour. de
Math. Pure et Appliquées, 90 (2008), no. 3, 286-311.
[10] Raugel, Geneviève Dynamics of partial differential equations on thin domains, in: Dy-
namical Systems, Montecatini Terme, 1994, in: Lecture Notes in Mathematics, vol. 1609,
Springer, Berlin, 1995, 208–315.

(I. Birindelli) Dipartimento di Matematica Guido Castelnuovo, Sapienza Univer-

sità di Roma, Piazzale Aldo Moro 5, Roma, Italy.
Email address: [email protected]

(A. Briani) Institut Denis Poisson, Université de Tours, France.

Email address: [email protected]

(H. Ishii) Institute for Mathematics and Computer Science, Tsuda University,
2-1-1 Tsuda, Kodaira, Tokyo 187-8577 Japan.
Email address: [email protected]