Example Sheets of Elliptic PDEs

Davide Parise
April 26, 2024

Example Sheet 1
• Definition of locally uniformly convex in question 7.
• The proof of Question 7 only works in a neighborhood of a local maxima and not a local one?
• Issue for the expression of the Laplacian in polar coordinates not being valid in Q4. Thus, we
have to check that the function u is harmonic as well at the origin.
• The solution of exercise1 3 can be simplified.
• Typed solution of Q7.
• Typed solution of Q9.

Exercise 1. Let Ω ⊂ Rn be a domain and let u ∈ C 1 (Ω) ∩ C 2 (Ω \ {x; u(x) = 0}) be a non-negative
function such that ∆u = 0 on the set Ω \ {x; u(x) = 0}. Show that either u ≡ 0 on Ω or u > 0 on Ω.
Solution 1. Denote the set {x; u(x) = 0} by Σ and notice that by continuity the set is relatively closed
in Ω. It is either empty or non-empty. In the first case we obtain a harmonic function on the whole
domain Ω. By minimum principle we know that it cannot attain the minimum inside, thus we must
have u > 0, as it is non-negative, and we are done in this case. Thus, assume Σ 6= ∅. If Σ were open
we would be done as by connectedness we would have Σ = Ω, i.e. u ≡ 0. Thus, assume that Σ is not
open. There exists z ∈ Σ such that for every radius ρ > 0 we have Bρ (z) ∩ {x; u(x) > 0} = 6 ∅. Choose
ρ0 > 0 such that Bρ0 (z) ⊂ Ω, any y ∈ Bρ0 /2 (z) satisfies dist(y, Σ) < dist(y, ∂Ω). Consider now

R := sup{r > 0; Br (y) ⊂ Ω \ Σ}.

Then, by definition of R, we have that B R (y) ∩ Σ 6= ∅. In particular, there exists t ∈ B R (y) ∩ Σ, at

which the function u attains a minimum by non-negativity. Hopf boundary point lemma applied to
Ω \ Σ implies then ∂u
∂η (z) < 0 which is absurd because Du(z) = 0. Thus, Σ is open and we are done.

Exercise 2. Let (uj ) be a sequence of non-negative harmonic functions on B1 (0) such that the sequence
of numbers (uj (0)) converges. Show that there is a non-negative harmonic function u on the open unit
ball and a subsequence (ujl ) such that ujl → u in C k (Bρ (0)) for every k ∈ N and every ρ ∈ (0, 1).
Must the whole sequence (uj ) converge to u? What if we drop the non-negativity assumption on the
uj ’s?
Solution 2. Before proceeding with the exercise we recall Harnack’s inequality. Let u ∈ C 2 (B1 ) ∩
C 0 (B 1 ) be a non-negative harmonic function, then for every r ∈ (0, 1) there exists Cr = C(n, r) such
that
sup u ≤ Cr inf u.
Br Br

Thus, going back to our problem, we obtain

sup uj ≤ Cr inf uj ≤ Cr uj (0) ≤ Cr M,

Br Br

1
where the last inequality follows by convergence of the sequence of numbers (uj (0)). Indeed, converging
sequences are bounded. Interior Schauder estimates entail for all j ≥ 0 the following bound

|uj |2,α;Br/2 ≤ C|uj |0;Br ,

where we dropped the dependences of the constant C for the sake of clarity. Arzela-Ascoli implies
then convergence of some subsequence of (uj ) in C 2 (Br/2 ) to a limit u ∈ C 2,α (Br/2 ). Notice that
the convergence is not a priori in C 2,α . To obtain C k convergence higher interior regularity Schauder
estimates or bootstrap the previous estimates recalling that ∂ α u are harmonic for all multi-indices
α. A more detailed explanation of the latter can be found in the next paragraph. Fix a multi-index
α ∈ Nn of length |α| = 1 and recall that ∂ α uj is harmonic again for all j ≥ 0. Thus,

sup |∂ α uj | ≤ C inf |∂ α uj | ≤ C|(∂ α uj )(0)| ≤ C|uj |2,α;Br/2 ≤ C|uj |0;Br ≤ C̃.

Br Br

As before, interior Schauder estimates entail

|∂ α uj |2,α;Br/2 ≤ C|∂ α uj |0;Br ,

for all j ≥ 0. Considering the subsequence of before and applying Arzela-Ascoli again we obtain a
further subsequence, not relabelled, along which ∂ α uj converges in C 2 (Br/2 ) to a limit g ∈ C 2,α (Br/2 ).
We then have that g = ∂u, so that u ∈ C 3,α (Br/2 ). We can conclude by induction.

To answer the first question negatively consider B1 (0) ⊂ R and the functions f1 (x) = 1 + x and
f2 (x) = 1 − x. Defined ui to be f1 when i is odd and f2 otherwise. We then have ui (0) = 1 for
all i, whereas the sequence has two distinct limits. To answer negatively the second one consider the
sequence of functions uk (x) = kx. As k → ∞ the sequence diverges.
Exercise 3. Let Ω ⊂ Rn be open. If u ∈ C 0 (Ω) satisfies the mean value property
Z
n −1
u(x) = (ωn ρ ) u(y) dy,
Bρ (x)

for every x ∈ Ω and every ρ ∈ (0, dist(x, ∂Ω)). Show that u is smooth and harmonic in Ω. If (un ) is a
sequence of harmonic functions converging to a function u locally uniformly on Ω, deduce that u must
be harmonic.
Solution 3. Recall that the mean value property for spheres can take the following form
Z
1
u(x) = u(x + rω) dSω , for Br (x) ⊂ Ω
ωn |ω|=1

and Z
n
u(x) = u(x + rω) dz, for Br (x) ⊂ Ω.
ωn |ω|≤1

Here ωn denotes the surface area of the unit sphere R and dSω is the surface measure of the sphere.
Consider then a function φ ∈ Cc∞ (B1 (0)) satisfying B1 (0) φ = 1 and such that φ(x) = ψ(|x|), i.e.
Z 1
ωn rn−1 ψ(r) dr = 1.
0

n
Define then the family φ (z) = φ(z/)/ for  > 0. Now, for any x ∈ Ω consider  < dist(x, ∂Ω).

2
Then, the following holds
Z Z
u(y)φ (y − x) dy = u(x + y)φ (y) dy
Ω
Z
1
= n u(x + y)φ(y/) dy
 |y|<
Z
= u(x + y)φ(y) dy
|y|<1
Z1 Z
n−1
= r dr u(x + rω)φ(rω) dSω
0 ∂B1 (0)
Z 1 Z
= ψ(r)rn−1 dr u(x + rω) dSω
0 |ω|=1
Z 1
= u(x)ωn ψ(r)rn−1 dr = u(x),
0

where in the last step we used the mean value property. Thus, u(x) = (φ ∗ u)(x) for all x ∈ Ω = {y ∈
Ω; dist(y, ∂Ω) > }, so that u is smooth on the whole domain Ω. Moreover,
Z Z
n−1 ∂ ∂
∆u dx = r u(x + rω) dSω = rn−1 (ωn u(x)) = 0,
Br (x) ∂r |ω|=1 ∂r

for any Br (x) ⊂ Ω, where the first equality follows from the divergence theorem. Indeed,
Z Z Z Z
∂u n−1 ∂u n−1 ∂
∆u(y) dy = (z) dS(z) = ρ (x + ρω) dSω = ρ u(x + ρω) dSω .
Bρ (x) ∂Bρ (x) ∂η |ω|=1 ∂ρ ∂ρ |ω|=1

Whence, ∆u = 0 on the whole Ω. Consider then a sequence {un }n∈N of harmonic functions converging
towards a certain u locally uniformly on Ω, i.e. for every x ∈ Ω there exists an open neighborhood U
such that un → u uniformly in Ω ∩ U . We shall see that u is harmonic as well. Indeed,
Z Z
1 1
u(x) − u(y) dy ≤ |u(x) − un (x)| + un (x) − un (y) dy
|Bρ (x)| Bρ (x) |Bρ (x)| Bρ (x)
Z Z
1 1
+ un (y) dy − u(y) dy
|Bρ (x)| Bρ (x) |Bρ (x)| Bρ (x)
Z Z
1 1
= |u(x) − un (x)| + un (y) dy − u(y) dy
|Bρ (x)| Bρ (x) |Bρ (x)| Bρ (x)
Z
1
= |u(x) − un (x)| + |un (y) − u(y)| dy,
|Bρ (x)| Bρ (x)

where the first equality follows because the un ’s are harmonic and thus satisfy the mean value property.
Notice that we use the following version of the mean value property: a continuous function f is harmonic
if and only if for every x there exists r > 0 such that f (x) is equal to the average of f on the sphere
{z; |z − x| = ρ}, for every ρ ∈ (0, r). Alternatively, the average can be taken on balls of the same
radii. The term |u(x) − un (x)| goes to zero by local uniform convergence, whereas
Z
1
|un (y) − u(y)| dy
|Bρ (x)| Bρ (x)

goes to zero by local uniform convergence after choosing ρ small enough so that Bρ (x) ⊂ Ux , where
Ux is the open neighborhood of x on which we have uniform convergence. Therefore, the quantity
Z
1
u(x) − u(y) dy
|Bρ (x)| Bρ (x)

3
goes to zero as n → ∞, for all ρ > 0 small enough. In other words u satisfies the mean value property
so that by the above it is harmonic.

Alternatively, we can prove that u satisfying the mean value property is harmonic in a different
way. For any open ball B with B ⊂ Ω, let v ∈ C 2 (B) ∩ C 0 (B) be the solution to the Dirichlet problem
(
∆v = 0 in B,
v=u on ∂B.

The function v is harmonic and satisfies the mean value property. Assume now for the sake of contra-
diction that v − u attains its supremum M 6= 0 in B. Indeed, we assume that u and v are different,
otherwise we would be done. Consider then y ∈ Σ = {x ∈ B; (v − u)(x) = M } and Br (y) ⊂ B. The
mean value property implies
Z
1
M = (v − u)(y) = (v − u)(x) dx ≤ M,
|Br (y)| Br (y)

with equality if and only if v − u is identically equal to M in Br (y). Arbitrarity of y implies that Σ is
open in B. Moreover, it is clearly closed so that connectedness of B implies Σ = B. In other words u
is equal to a harmonic function on B. By arbitrarity of B we have that u is harmonic everywhere.
ExerciseP4 (Hadamard’s Example, 1906). Let B = B1 (0) ⊂ R2 be the open unit ball in R2 and let
∞
u(r, θ) = n=1 n−2 rn! sin(n!θ). Show that the series converges absolutely uniformly in B. Deduce that
u is continuous on the closed unit ball B and is harmonic in B. Show that u does not have finite
Dirichlet energy in B.
Solution 4. Start by writing un (r, θ) := n−2 rn! sin(n!θ), for n ∈ N. We then trivially have the bound
−2 n! −2
|un (r, θ)|
P∞≤ n −2r n! ≤ n for all (r, θ) ∈ B. Thus, by the Weierstrass M-test we conclude that the
series n=1 n r sin(n!θ) converges absolutely uniformly on the closed unit ball B. Moreover, the
un ’s are continuous from which we deduce that u is continuous as well, cf. uniform limit theorem or
simply arguing by continuity of the un ’s combined with the uniform convergence of the partial sums.
We now wish to show that u is harmonic in B. To this sake we need to differentiate the sum defining
it. Consider for instance differentiation in the θ variable. Formally interchanging derivative and sum
we would get
∞   X ∞
∂ X ∂ 1 n! 1 n!
u(r, θ) = 2
r sin(n!θ) = r cos(n!θ)n!.
∂θ n=1
∂θ n n=1
n2
However, provided that the sum on the right-hand side of the above converges uniformly we know
that the first identity is justified. Applying again the Weierstrass M-test we can infer uniform absolute
convergence of the sum under consideration. One argues similarly to show that
∞
∂u X 1
= 2
n!rn!−1 sin(n!θ),
∂r n=1
n

∞
∂2u X 1
= n!(n! − 1)rn!−2 sin(n!θ)
∂r2 n=2
n 2

and
∞
∂2u X 1
2
= − 2
(n!)2 rn! sin(n!θ).
∂θ n=1
n

Recall then that the Laplacian in polar coordinates in R2 is given by

1 ∂2u 1 ∂u ∂ 2 u 1 ∂2u
 
1 ∂ ∂u
∆u = r + 2 2 = + 2 + 2 2.
r ∂r ∂r r ∂θ r ∂r ∂r r ∂θ

4
Therefore, we can now show that u is harmonic
∞ ∞ ∞
X 1 n!−2
X 1 n!−2
X 1
∆u = 2
n!r sin(n!θ) + 2
n!(n! − 1)r sin(n!θ) − 2
(n!)2 rn!−2 sin(n!θ)
n=1
n n=2
n n=1
n
∞ ∞ ∞ ∞
X 1 n!−2
X 1 2 n!−2
X 1 n!−2
X 1
= 2
n!r sin(n!θ) + 2
(n!) r sin(n!θ) − 2
n!r sin(n!θ) − 2
(n!)2 rn!−2 sin(n!θ)
n=1
n n=2
n n=2
n n=1
n
1 1
= sin(θ) − sin(θ) = 0.
r r
Alternatively, one can prove that the un ’s are harmonic, so that the partial sums are harmonic as well
and by the uniform convergence of the partial sums, combined with Q3 allows to conclude that u is
harmonic as well.
We know prove that u does not have finite Dirichlet energy, thus concluding the exercise. Indeed,
for any radius ρ > 0, we can calculate
Z Z 2π Z ρ
|Du|2 dx ≥ |∂r u(r, θ)|2 rdrdθ
Bρ 0 0
∞ 2
Z 2π Z ρ X 1
= 2
n!rn!−1 sin(n!θ) rdrdθ
0 0 n=1
n
Z 2π Z ρ ∞
X 1
= n!m!rn!+m!−2 sin(n!θ) sin(m!θ) rdrdθ
0 0 n,m=1 n2 m2
∞ Z 2π Z ρ
X 1
= 2 m2
sin(n!θ) sin(m!θ) dθ rn!+m!−1 dr
n,m=1
n 0 0
∞ Z 2π
X 1 n!m! n!+m!
= 2 m2 n! + m!
ρ sin(n!θ) sin(m!θ) dθ
n,m=1
n 0
∞
X n! 2n!
=π ρ ,
n=1
2n4

where we used the the following elementary fact

Z 2π (
0 if n 6= m,
sin(n!θ) sin(m!θ) dθ =
0 π if n = m.

Thus, summarising, we have for every ρ > 0 and every m ∈ N the following lower bound
Z m
X n! 2n!
|Du|2 dx ≥ π ρ ,
Bρ n=1
2n4

from which we infer that the Dirichlet energy is infinite as required.

The main point of the above exercise is to show that there is no full equivalence between solving
the Dirichlet problem associated with the Laplacian and the minimization of the associated Dirichlet
integral. However, one can pretty straightforwardly prove the following result.
Proposition 1. Let Ω ⊂ Rn be an open set and let u ∈ C 2 (Ω) satisfy E(u) < ∞. Then the following
holds.
(i) If ∆u = 0 in Ω, then E(u + v) > E(u) for all nontrivial v ∈ Cc∞ (Ω).
(ii) Conversely, E(u + v) ≥ E(u) for all v ∈ Cc∞ (Ω), then ∆u = 0 in Ω.
Exercise 5. let u be a harmonic function on the open unit ball B = B1 (0) ⊂ Rn . If u(x) < |x|2 for
some ∂B1/4 (0), show that u(y) < |y|2 for some y ∈ ∂B1/2 (0).

5
Solution 5. Define f (x) := |x|2 − u(x) and notice that ∆f = 2n ≥ 0. Consider then as domain B1/2 (0)
and assume for the sake of contradiction that u(y) ≥ |y|2 for all y ∈ ∂B1/2 (0), i.e. f ≤ 0 on ∂B1/2 (0).
By hypothesis we know that there exists x ∈ ∂B1/4 (0) for which f (x) > 0. Combining this fact with
the weak maximum principle for subharmonic functions we obtain the desired contradiction:

0 < sup f = sup f ≤ 0.

B1/2 (0) ∂B1/2 (0)

Exercise 6. Let Ω be a bounded domain in Rn , and let u ∈ C 2 (Ω) ∩ C 0 (Ω) be a critical point (i.e. a
solution of the Euler-Lagrange equation) of the Allen-Cahn functional
Z
1
E(u) = |Du|2 + W (u),
Ω 2

where W : R → R is a smooth double well potential, that is a non-negative function with two non-
degenerate local minima at +1 and −1 with values W (1) = W (−1) = 0, one local maximum at at a
point in (−1, 1) and no other critical points. A standard example considered in the literature is the
function W (t) = 41 (t2 − 1). If u(x) ∈ (−1, 1) for all x ∈ ∂Ω, show that u(x) ∈ [−1, 1] for all x ∈ Ω.
Show in fact u(x) ∈ (−1, 1) for all x ∈ Ω.
Solution 6. We start by calculating the Euler-Lagrange equation for the Allen-Cahn functional. To do
so set
d
E(u + tφ) = 0, for all φ ∈ Cc∞ (Ω).
dt t=0
We then get
Z
d d 1
0= E(u + tφ) = |Du + tDφ|2 + W (u + tφ) dx
dt t=0 dt t=0 Ω 2
Z
d 1 1
= |Du|2 + t2 |Dφ|2 + thDu, Dφi + W (u + tφ) dx
dt 2 2
Z t=0 Ω
= hDu, Dφi + W 0 (u)φ dx
ZΩ
= φ(−∆u + W 0 (u)) dx.
Ω

Since the above holds for all φ ∈ Cc∞ (Ω) we obtain −∆u + W 0 (u) = 0. Moreover, u ∈ C 2 (Ω) so that
this equation is satisfied classically. To answer the first question of the exercise assume that u attains
a maximum strictly larger that 1 at a point x0 ∈ Ω. Then, ∆u(x0 ) ≤ 0 by virtue of it being a local
maximum. However, thanks to the Euler-Lagrange equation we have

∆u(x0 ) = W 0 (u(x0 )) > 0,

giving the desired contradiction. The inequality follows by property of the double well potential.
Indeed, it is increasing at points strictly larger than 1. One can similarly show that u ≥ −1.
We shall reason more abstractly to solve the second question. Indeed, consider u1 and u2 solutions of
the Euler-Lagrange equation. Their difference is solution of the linearised equation. More specifically,
Z 1
d 0
∆(u1 ) − ∆(u2 ) − W 0 (u1 ) + W 0 (u2 ) = ∆(u1 − u2 ) − W ((1 − t)u2 + tu1 ) dt
0 dt
Z 1
= ∆(u1 − u2 ) − W 00 ((1 − t)u1 + tu2 )(u1 − u2 ) dt
0
= ∆(u1 − u2 ) + c(u1 − u2 ),

where c is the smooth function defined by

Z 1
c=− W 00 ((1 − t)u1 + tu2 ) dt.
0

6
Thus, v = u1 − u2 solves ∆v + cv = 0. Choosing then u1 = u and u2 = 1 we have v ≤ 0 by the
above. Thus, by the generalization of the comparison principle due to Serrin, see for instance [2,
Theorem 2.10], we have v < 0 or v ≡ 0 in Ω. The latter being impossible, recall that we are looking
for non-trivial solutions, we have that the former is satisfied, i.e. u < 1 on Ω.
Exercise 7. Let F(u) = Ω F (x, Du(x)) dx, where Ω is a domain in Rn and F : Ω × Rn → R is in
R

C 3 (Ω × Rn ) and is locally uniformly convex in the Rn variables.

(i) If F (x, p) is independent of x, show that a C 2 critical point of F cannot attain a local maxi-
mum/minimum at a point in Ω unless it is constant. Show also that the difference of two critical
points of F, for now a general F = F (x, p) as above, cannot attain a local maximum/minimum
in Ω unless it is constant.
(ii) Now suppose that Ω = Ω0 ×R, where Ω0 is a bounded domain in Rn−1 . Let (x, y) ∈ Rn−1 ×R denote
coordinates in Rn , and assume that F ((x, y), p) is independent of y. Let φ : ∂Ω = ∂Ω0 × R → R be
continuous, and Lipschitz with respect to the y variable with uniform Lipschitz constant bounded
by L and independent of the x variable. If u ∈ C 2 (Ω) ∩ C 0 (Ω) is a critical point of F such
that u|∂Ω = φ and if u(x, y) and u(x, −y) converge uniformly on Ω0 as y → ∞, show that
supΩ |Dy u| ≤ L.
Solution 7.
Exercise 8. Let B be the open unit ball in Rn and let u ∈ C 2 (B) ∩ C 1 (B) be a solution to the minimal
surface equation in B. If the boundary of graph(u) := {(x, u(x)); x ∈ B} is contained in an affine
hyperplane P ⊂ Rn+1 , show that graph(u) is contained in P .
Solution 8. The hyperplane P can be written in the form {x ∈ Rn+1 ; ṽ(x) = 0}, where v is the smooth
function given by ṽ(x) = hx, ηi + α, where η ∈ S n and α ∈ R. Alternatively, the hyperplane can be
written as graph of a smooth function v related to ṽ. This function v trivially satisfies the minimal
surface equation, so that we have
! !
∇v ∇u
div p = 0 and div p = 0.
1 + |∇v|2 1 + |∇u|2

Moreover, on ∂B we have equality between u and v. We claim that v = u on B. To this sake, we set
w = u − v and we derive the equation that it satisfies. Let
p
F (p) = p : Rn → Rn ,
1 + |p|2
be the vector function defining the minimal surface equation. Then,
Z 1
d
Fi (Du) − Fi (Dv) = Fi (tDu + (1 − t)Dv) dt
0 dt
Z 1
∂Fi
= (Dv + t(Du − Dv))Dj (u − v) dt
0 ∂pj
= aij Dij w,
where the Einstein summation convention is in force and the coefficients are given by
Z 1
∂Fi ∂Fi δij pi pj
aij (x) = (Dv + t(Du − Dv)) dt and =p − .
0 ∂pj ∂pj 1 + |p|2 (1 + |p|2 )3/2

Therefore, w satisfies Di (aij Dj w) = 0, or equivalently, since w is C 2 , aij Dij w + bi Di w = 0 with the

obvious changes of the coefficients. We now claim that the equation satisfied by w is uniformly elliptic.
Indeed,
hp, ξi2 |ξ|2 |p|2 |ξ|2
   
∂Fi 1 2
ξi ξj = |ξ| − 2
≥ 1 − 2
= ,
∂pj (1 + |p|2 )1/2 1 + |p| (1 + |p|2 )1/2 1 + |p| (1 + |p|2 )3/2

7
for any ξ ∈ Rn . The inequality follows by Cauchy-Schwarz. Furthermore, we have the trivial bound

∂Fi |ξ|2
ξi ξj ≤ , (1)
∂pj (1 + |p|2 )3/2
again for every ξ ∈ Rn . Therefore,
Z 1 Z 1
∂Fi |ξ|2
aij ξi ξj = (Dv + tD(u − v))ξi ξj dt ≥ 2 )3/2
dt ≥ |ξ|2 λ,
0 ∂pj 0 (1 + |Dv + t(Du − Dv)|

where λ > 0 follows from u and v being C 1 (B). Indeed, by compactness of B we can find a minimum
for the integrand function. Boundedness of the coefficients follows from (1). From the strong maximum
principle for divergence form elliptic equations we get that w ≡ 0 on the whole B, thus showing what
u ≡ v. Geometrically, the minimal graph coincides with the hyperplane.

For the more differential geometric inclined reader there is a proof based on the harmonicity of
the coordinate functions. Denote by Σ the submanifold of Rn+1 given by the graph of u and by gΣ
the Riemannian metric induced on Σ. Recall that we endow Rn+1 with the standard flat metric. Let
ei = ∇Rn+1 xi be the i-th coordinate vector field on Rn+1 and consider η ∈ Cc∞ (Σ) and vanishing on
∂Σ. Here i ∈ {1, 2, . . . , n + 1}. The divergence of ηei is then given by

divΣ (ηei ) = h∇Σ η, ei i = h∇Σ η, ∇Rn+1 i,

where ∇Σ is the induced connection on Σ and divΣ is the divergence operator on Σ. Thus,
Z Z Z Z
− hηei , Hi dµgΣ = divΣ (ηei ) dµgΣ = h∇Σ η, ∇Rn+1 xi i dµgΣ = − η∆Σ xi dµgΣ , (2)
Σ Σ Σ Σ

where dµgΣ denotes the volume form on Σ, H is the mean curvature vector and ∆Σ is the Laplacian
on Σ. From (2) we deduce that the restriction of the coordinate functions of Rn+1 to Σ are harmonic
functions. Using this fact we can prove the so called convex hull property of minimal surfaces, from
which we are going to infer the desired claim. More specifically, we are going to prove that Σ ⊂
Conv(∂Σ), where the latter term denotes the convex hull of ∂Σ. As a convex set we have the following
characterization of the convex hull \
Conv(∂Σ) = H,
H∈H

where H is the collection of hyperplanes containing ∂Σ and H is an element of such collection. The
latter can be written as {x ∈ Rn+1 ; hx, ei ≤ a}, for certain e ∈ Sn and a ∈ R. By the above, the function
u(x) = hx, ei is harmonic on Σ so that maximum principle implies that u reaches its maximum on ∂Σ.
Thus, there exists z ∈ ∂Σ such that u(z) ≥ u(x) for all x ∈ Σ, so that hx, ei ≤ hz, ei ≤ a. Consequently,
every x ∈ Σ satisfies x ∈ H for all H ∈ H and then x ∈ Conv(∂Σ), giving the desired inclusion.
Remark 1. More generally, one could show that if the hyperplane is on one side of the minimal graph
then it stays on the same side or it coincides. In mathematical terms this translates to u1 ≤ u2 on ∂Ω,
then either u1 ≡ u2 or u1 < u2 , where u1 and u2 are two arbitrary solutions of the minimal surface
equation.
Remark 2. The second proof works for arbitrary compact minimal submanifold, not necessarily minimal
graphs, in arbitrary codimension. The argument of the proof of the convex hull property can be
rephrased as follows: translating a hyperplane towards a minimal surface, the first point of contact
must be on the boundary.
Exercise 9. Give a direct proof of the strong maximum principle that does not rely on the Hopf
boundary point lemma by completing the following outline.
Let Ω be a bounded domain in Rn , L = aij Dij + bi Di + c a uniformly elliptic operator on Ω with
|b |/λ + |c|/λ bounded, where λ(x) > 0 is the smallest eigenvalue of (aij (x)). Let u ∈ C 2 (Ω) satisfy
i

Lu ≥ 0 in Ω and suppose that M = supΩ u(x) < ∞ and that there is a point x0 ∈ Ω with u(x0 ) = M .
Let Σ = {x ∈ Ω; u(x) = M }. We wish to show that Σ = Ω in all three of the following cases:

8
(A) c ≡ 0;
(B) C ≤ 0 and M ≥ 0;
(C) M = 0 with no assumption on the sign of c.
Suppose Σ 6= Ω.
(i) Show that tehre is a point y ∈ Ω \ Σ and R > 0 such that BR (y) ⊂ Ω \ Σ and Σ ∩ ∂BR (y)
consists precisely of one point y1 . Choose R1 > 0 with R1 < R such that BR1 (y1 ) ⊂ Ω. Let
2 2
v(x) = e−α|x−y| − e−αR anad let w = u + v for some constants α > 0 and  > 0. Consider
first the cases (A) and (B).
(ii) Show that we may choose α large enough to ensure that Lv > 0 on BR1 (y1 ).
(iii) With α chosen as in the previous point, show that we may choose  > 0 small enough to ensure
that w < M on ∂BR1 (y1 ).
(iv) With α and  chosen as in the two previous pointsl, show that w attains its maximum on BR1 (y1 )
at an interior point z ∈ BR1 (y1 ) and that w(z) ≥ M . Show that this leads to a contradiction in
cases (A) and (B).
(v) Complete the proof by showing that case (C) can be reduced to case (B).
Solution 9. We shall follow the outline in the statement of the exercise.
(i)
(ii)
(iii)
(iv)
(v)
Exercise 10. Show that a C 2 domain in Rn satisfies the interior sphere condition at every point of
its boundary.
Solution 10. As ∂Ω is a C 2 -submanifold we know that its curvature is well-defined, hence we can take
as interior sphere the osculating sphere or any sphere that has radius smaller than the inverse of the
curvature.
Alternatively, one can proceed as follows. Consider a point x0 ∈ ∂Ω and assume, after translation,
that x0 coincides with the origin. Because ∂Ω is C 2 we can find a radius R > 0 and a C 2 -function
n−1
F : BR ⊂ Rn−1 → R such that BR (x0 ) ∩ ∂Ω can be expressed as the graph of it. Moreover, we
have that (x0 , xn ) ∈ Ω ∩ BR (0) if and only if F (x0 ) < xn . Without loss of generality assume that
DF (0) = 0. In other words, starting from F consider the function F̃ (x) = F (x) − DF (x) · x − F (0).
We shall denote such function F in the coming paragraphs. By the multivariate Taylor theorem with
integral remainder we have the bound
X Z 1
|F (x0 )| ≤ (x0 )β (1 − t)|Dβ F (tx0 )| dt ≤ C(n)|x0 |2 max |D2 F |
n−1
0 BR
|β|=2

n−1
for all x0 ∈ BR (0) and where D2 F denotes the Hessian of F , and C(n) is a dimensional constant
used to bound the (x0 )β term. Consider now a ball centered at z = (0, r) for a certain r > 0 small
enough so that Br (z) ⊂ BR (0). A point (x0 , xn ) lies in said ball provided (x0 )2 + (xn − r)2 < r2 .
Thus, we are left with showing that, upon taking a smaller radius, elements of this ball are inside our
domain. In order for (x0 , xn ) to be in Ω ∩ BR (0) we need to have |F (x0 )| < xn . Using our bound on
|F (x0 )| we find
|F (x0 )| ≤ C(n)|x0 |2 max |D2 F | ≤ C(n) r2 − (xn − r2 )2 max |D2 F | ≤ 2C(n)xn r max |D2 F |.


n−1 n−1 n−1
BR BR BR

9
Choosing
1
r< ,
2C(n) maxB n−1 |D2 F |
R

and possibly taking r even smaller so that it lies entirely in the ball BR (0) we are done.
Remark 3. One can relax the regularity of the boundary in the previous exercise. More precisely, Ω
satisfies the uniform interior and exterior sphere sphere condition if and only if the boundary ∂Ω is
C 1,1 . Uniform here means that the radius can chosen independently of the boundary point. A simple
example of two-dimensional domain for which the interior sphere condition fails is the domain

Ωα = {(x, y) ∈ R2 ; y > |x|1+α },

for a fixed α ∈ [0, 1). We exclude the endpoint α = 1 because we have just seen that the inte-
rior/exterior sphere property holds there. The domain Ωα is C 1,α -regular but does not satisfies the
interior/exterior sphere property. Indeed, as there is an issue at the origin. As in the proof above,
we would consider a circle centered at (0, r) and it would touch
√ tangentially the y = 0 plane. The
equation of the bottom branch of the circle would be y = r − r2 − x2 = O(x2 ) which lies below the
graph |x|1+α .
Remark 4. In the above proof a bound on the Hessian can be considered as a bound on the principal
curvatures of the surface at the origin.

10
Example Sheet 2
• Typed solution of Q5.

Exercise 1. This exercise will show that Schauder estimates cannot be improved to encompass the
cases α = 0, 1.

(i) Let α be a multi-index with |α| = 3 and let Q be a homogeneus degree 3 harmonic polynomial
with Dα Q 6= 0. Choose η ∈ Cc∞ (Rn ) with η ≡ 1 on B1 (0) and η = 0 on Rn \ B2 (0). Define
∞ ∞
X 1 X 1
u(x) = η(2k x)Q(x) = (ηQ)(2k x),
k k23k
k=0 k=1

so that
∞
X 1
∆u(x) = g(x) = ∆(ηQ)(2k x).
k2k
k=1
1 n 2,1
Show that g ∈ C (R ) but u 6∈ C (U ) for any open neighborhood of the origin U . Why does
existence of this example show that there is no C 2,1 Schauder estimate?
(ii) Let α be a multi-index with |α| = 2 and P be a homogeneous degree 2 harmonic polynomial with
Dα P 6= 0, e.g. P = x1 x2 , D12 P = 1. Choose η as before and define in a similar manner
∞
X 1
f (x) = ∆(ηP )(2k x).
k
k=1

Show that f is continuous but ∆u = f does not have a C 2 solution in any neighborhood of the
origin. Why does this imply that there is no C 2 Schauder estimate, i.e. that we cannot take
α = 0 in the interiori C 2,α Schauder estimate in a ball?
Solution 1. We shall prove only the first point as the second one follows a similar argument.
Start by noticing the following (ηQ)(x) = 0 for |x| > 2, by definition of η. Similarly, (ηQ)(2k x) = 0,
whenever |2k x| > 2, i.e. when |x| > 1/2k−1 . In particular, this is true for any compact set away from
zero, so that the sum is locally finite. Thus, it converges pointwise and uniformly on compact subsets
of Rn not containing the origin. This implies continuity of u everywhere except at the origin. The
same argument holds for derivatives of η, so that we have that u is smooth everywhere except at the
origin. At any x 6= 0 we can therefore differentiate the sum term by term
∞
X 1
∆u(x) = ∆(ηQ)(2k x), (3)
k2k
k=1

expanding the Laplacian we obtain

∞ ∞
X 1 k
X 1
∆u(x) = (∆ηQ + 2hDη, DQi + η∆Q) (2 x) = (∆ηQ + 2hDη, DQi) (2k x),
k2k k2k
k=1 k=1

where the second equality follows by harmonicity of Q. As in the statement of the exercise denote
g := ∆u. At x 6= 0 the function g is continuous. Furthermore, we can differentiate the sum term by
term to infer C 1 -regularity away from the origin.
We now claim that g is continuous at the origin. Note that all derivative of η vanish outside
B2 \ B1 , so that all terms in the sum (3) vanish, except possibly for the one where 1 ≤ |2k x| ≤ 2, i.e.
− ln2 |x| ≤ k ≤ 1 − ln2 |x|. The integer k is uniquely determined by these inequalities and depends on
x, i.e. we have k = k(x). Thus, we can write g as
1
g(x) = (∆ηQ + 2hDη, DQi) (2k x).
k2k

11
Note that k → ∞ if and only if x → 0. To conclude continuity at the origin we have that g can be
writte as P/Q, where P is the continuous function given by ∆ηQ + 2hDη, DQi that admits a minimum
and a maximum as a continuous function on a compact set. On the other hand, Q(x) = k(x)2k(x)
which can be seen to diverge as x → 0 simply by the bounds we have on the function k(x), thus proving
that g(x) → 0.
We claim now that g ∈ C 1 (Rn ). Note that outside the origin we can differentiate term by term to
obtain
∞
X 1
D(∆η)Q + ∆ηDQ + 2D2 ηQ + 2DηD2 Q (2k x)


Dg(x) =
k
k=1

Note that both ∆Q and DQ are homogeneous polynomial of degree 1 and respectively 2. Also
introduce kP k = sup|x|=1 |P (x)|, for all polynomials P . As before, there is at most one k for which the
above sum is not zero. Thus, we can bound the expression for Dg(x), at that particular k, as follows
1
|Dg(x)| ≤ |D(∆η)(2k x)kQk|2k x|3 + |∆η(2k x)|kDQk|2k x|2
k
+ 2k Hess η(2k x)|kDQk|2k x|2 + 2|Dη(2k x)|k Hess Qk|2k x|.

Thus, we can find a constant C, depending on Q, η and their derivatives such that |Dg(x)| ≤ C|2k x|/k,
for all x ∈ Rn \ {0}. As |2k x| < 2 and k → ∞, we obtain Dg(x) → 0, as x → 0. Thus, g ∈
C 0 (Rn ) ∩ C 1 (Rn \ {0}) and Dg(x) → 0, as x → 0. We now claim that g ∈ C 1 (Rn ) and Dg(0) = 0. To
prove it we use the mean value theorem. Assume for the sake of simplicity that we are in dimension
one. For every x ∈ R \ {0}, there exists cx ∈ (0, x) (where we assumed without loss of generality x > 0)
such that
f (x) − f (0)
f 0 (cx ) = .
x
Now, by hypothesis limx→0 f 0 (cx ) exists and by definition of derivative we have differentiability at the
origin. The fact that Dg(0) = 0 follows from the definition of derivative again.
We now claim that u 6∈ C 2,1 (U ) for any neighborhood U of the origin. For x 6= 0, calculate the
following
∞
X 1
Dij u(x) = (Dij ηQ + Di ηDj Q + Dj ηDi Q + ηDij Q) (x2k ).
k2k
k=1

The first three elements, namely (Dij η)Q and Di ηDj Q, together with Dj ηDi Q are dealt with as
before. They converge to zero as x → 0. We are left with the fourth term, the one without derivatives
of η. Denote α = (i, j, k) a multi-index of order three and recall that by hypothesis Dα Q 6= 0. We
claim that
1
|Dij u(tek )| → ∞, t → 0.
t
Note that Dij Q(2m tek ) = Dα Q(0)2m t as Dij Q is a homogeneous polynomial of degree one and Dα Q(0)
is the leading coefficient. Therefore, we compute
∞ ∞ ∞
X 1 m
X 1 α m m α
X 1
m
(ηDij Q) (2 te k ) = m
D Q(0)η(2 te k )2 t = tD Q(0) η(2m tek ).
m=1
m2 m=1
m2 m=1
m

Because of the asymptotics

∞
X 1
η(2m tek ) ∼ | ln2 (−t)|,
m=1
m

we conclude that, as t → 0, the sum diverges. As Dij u(0) = 0, we have that u is not Lipschitz
continuous in any neighborhood of the origin, i.e. u 6∈ C 2,1 (U ).
The function just constructed is an example of a function with Lipschitz continuous Laplacian with
is not C 2,1 itself. This in principle would not contradict an α = 1 version of Schauder estimates as
u would be a priori assumed to the C 2,1 -regular. However, the example we just constructed would

12
violate the corollary of Schauder estimates appearing at the end of this solution. Indeed, if a version
of such corollary would exist for α = 1, we could apply it to the sequence of functions defined by
m
X 1
um (x) = (ηQ)(2k x),
k23k
k=0

to deduce that u ∈ C 2,1 .

Theorem 1. Let α ∈ (0, 1) and {un }n∈N ⊂ C 2,α (B 1 ) such that there exists M ≥ 0 satisfying

|un |0;B1 + |∆un |0,α;B1 ≤ M,

for n ∈ N. Then, there exists a subsequence {unj }j converging locally in C 2 to a function u ∈ C 2,α (B 1 ).
Moreover, for every ρ > 0 there exists Cρ = C(ρ, α, n) such that

|un |2,α;Bρ ≤ Cρ (|∆u|0,α;B1 + |u|0;B1 ) .

Exercise 2. Arzela-Ascoli theorem. Give a proof of the following version of the Arzela-Ascoli theorem.
If Ω is an open subset of Rn , 0 < α ≤ 1 and {uj }∞
j=1 is a sequence of functions in C
k,α
(Ω) satisfying

sup |uj |k,α;Ω0 ≤ C(Ω0 ), for each open set Ω0 b Ω,

j∈N

for some constant C depending on Ω0 , then there is a subsequence {ujl }∞

l=1 and a function u ∈ C
k,α
(Ω)
k 0 0
such that ujl → u in C (Ω ) for each open subset Ω b Ω.
Solution 2. We start by considering Ω bounded. Fix k ∈ N. The bound

sup |uj |k,α;Ω0 ≤ C, for each open set Ω0 b Ω,

j∈N

gives equicontinuity of Dl u and uniform boundedness for every l ≤ k. Thus, the usual Arzela-Ascoli
theorem implies the existence of a convergence subsequence and a limit ul , for every Ω0 b Ω and for
every l ≤ k. Recall that the convergence is uniform. The issue of the derivatives of the limit being
the limit of the derivatives is dealt with in the usual way (it follows by uniform convergence of the
sequence and its derivatives). We now check that the limit u belongs to C k,α (Ω0 ) and to do so we
have to prove that Dk u ∈ C 0,α (Ω0 ). This follows once again from the uniform bound we have on the
sequence. More precisely,

|Dk uj (x) − Dk uj (y)| ≤ |uj |k,α;Ω0 |x − y|α ≤ C(Ω0 )|x − y|α ,

and
|Dk uj (x) − Dk uj (y)|
≤ C(Ω0 ).
|x − y|α

Letting j → ∞, we deduce by uniform convergence

|Dk u(x) − Dk u(y)|

≤ C(Ω0 ).
|x − y|α

Taking the supremum over Ω0 gives the desired bound [Dk u]0,α;Ω0 ≤ C(Ω0 ). In other words u ∈
C k,α (Ω0 ). To conclude the proof in the bounded Ω case, we are left with proving that the limit and the
subsequence are independent of the subdomain Ω0 ⊂ Ω. Start by noting that if Ω0 ⊂ Ω00 , convergence
in Ω00 implies convergence in Ω0 , so that on smaller subsets the same limit and subsequence work. Now
let,
Ω := {x ∈ Ω; dist(x, ∂Ω) > }.

13
These sets are open and compactly contained in Ω for all  > 0. Furthermore, they form an exhaustion
of Ω, whence, every Ω0 b Ω will be eventually contained in some Ω for  > 0 small enough. Using a
diagonal argument we have the existence of u ∈ C k,α (Ω) such that ulm → u in C k (Ωn ) for all n ∈ N.
If we now assume Ω to be unbounded, define

Ω,k = {x ∈ Ω; dist(x, ∂Ω) > , |x| ≤ k},

for  > 0 and k ≥ 0. Consider then kn = n and n = 1/n for n ∈ N. Every Ω0 b Ω will eventually
be contained in a certain Ωn ,kn for which we can apply the previous argument. A diagonal argument
allows again to conclude.
Exercise 3. Establish the following interpolation inequality: given  > 0, a positive integer l and
α ∈ (0, 1], there is a constant C ∈ (0, ∞) depending only on n, l, α and  such that if u ∈ C l,α (BR (x0 ))

Rk |Dk u|0;BR (x0 ) ≤ C|u|0;BR (x0 ) + Rα+l [Dl u]α;BR (x0 ) , (4)

for all 0 ≤ k ≤ l.
Solution 3. First of all notice that we can assume without loss of generality that R = 1 and x0 = 0.
Indeed, consider the function v(x) = u(x0 +Rx). Thus, v ∈ C l,α (B1 (0)). Moreover, Dk v = Rk Du(x0 +
Rx), so that

|Dk v|0;B1 (0) = Rk |Dk u|0;BR (x0 ) and [Dl v]α;B1 (0) = Rl+α [Dl u]α;BR (x0 ) .

The inequality in (4) would then become

|Dk v|0;B1 (0) ≤ C|u|0;B1 (0) + [Dl u]α;B1 (0) .

We now claim that it is enough to prove

|u|l;B1 (0) ≤ C|u|0;B1 (0) + |u|l,α;B1 (0) . (5)

To see why this is the case recall that |u|l,α;B1 (0) = |u|l;B1 (0) + [Dl u]α;B1 (0) so that (5) becomes

(1 − )|u|l;B1 (0) ≤ C|u|0;B1 (0) + [Dl u]α;B1 (0) ,

or
C  
|u|l;B1 (0) ≤ |u|0;B1 (0) + [Dl u]α;B1 (0) = C̃|u|0;B1 (0) + [Dl u]α;B1 (0) .
1− 1− 1−
Notice then that the function ψ(x) = x/(1−x) is a bijection between (0, 1) and (0, ∞), so that /(1−)
can be made equal to any δ by appropriately choosing the value of  ∈ (0, 1). Moreover, we trivially
have |Dk u|0;B1 (0) ≤ |u|l;B1 (0) for all k ∈ {0, 1, . . . , l}. Therefore, it suffices to prove (5) for  ∈ (0, 1) to
establish the desired inequality for all  > 0. In order to finish the exercise we argue by contradiction.
Then for every n ∈ N there exists un ∈ C l,α (B1 (0)) such that

|un |l;B1 (0) > n|un |0;B1 (0) + |un |l,α;B1 (0) .

Normalising the un ’s by writing vn = un /|un |l;B1 (0) we obtain

1 > n|vn |0;B1 (0) + |vn |l,α;B1 (0) , for all n ∈ N,

from which we deduce

1 1
|vn |0;B1 (0) < |vn |l,α;B1 (0) < , (6)
n 
again for all n ∈ N. Applying Arzela-Ascoli we have the existence of a subsequence {vni }∞ i=1 converging
in C l (B1 (0)) to a certain v ∈ C l,α (B1 (0)). However, by the first inequality in (6) we have v ≡ 0 which
gives the desired contradiction once it is combined with |vn |l;B1 (0) = 1 and the C l (B1 (0)) convergence
of the vn ’s towards v.

14
 The above exercise is an instance of a more general general result about interpolation that goes
under the name of Ehrling’s Lemma.
Lemma 1 (Ehrling’s Lemma). Let (X, k · kX ), (Y, k · kY ) and (Z, k · kZ ) be three Banach spaces such
that the first one is compactly embedded in the second one and the second one is continuously embedded
in the third one. Then, for every  > 0, there exists a constant C() such that, for all x ∈ X, the
inequality
kxkY ≤ kxkX + C()kxkZ
holds.
Proof. If the result were not true, we would be able to find 0 and a sequence {xn }n∈N ⊂ X, such that
kxn kY > 0 kxn kX + nkxn kZ .
Normalise the sequence so that kxn kY = 1 for every n ∈ N and the above inequality reduces to
1 > 0 kxn kX + nkxn kZ .
Now, X is compactly embedded in Y , whence the existence of a, not relabelled, subsequence and a
limit x ∈ Y . By continuity of the inclusion Y ,→ Z we have that xn → x in Z as well. In particular,
kxn kZ ≤ 1/n, from which we deduce that kxkZ = 0, i.e. x = 0. This gives the desired contradiction,
as kxn kY = 1.
Remark 5. The problem with the above lemma, and proof, is that it does not give any information on
the constant C() and its size or behaviour. It is the consequence of an abstract proof. However, in
certain specific situations one can find such a constant explicitly without invoking Ehrling’s lemma.
Exercise 4 (Schwarz’s Reflection principle for harmonic functions). Let Ω+ be an open subset of
the upper-half space Rn+ = {x ∈ Rn ; xn > 0}. Let then T = ∂Ω+ ∩ {xn = 0} and Ω− be the
reflection of Ω+ with respect to {xn = 0}, i.e. Ω− = {(x1 , x2 , . . . , −xn ); (x1 , x2 , . . . , xn ) ∈ Ω+ }. Let
v ∈ C 2 (Ω+ ) ∩ C 0 (Ω+ ∪ T ) and let v̄ : Ω+ ∪ T ∪ Ω− → R be the odd reflection of v with respect to
{xn = 0} defined by
(
v(x1 , x2 , . . . , xn ) if (x1 , x2 , . . . , xn ) ∈ Ω+ ∪ T ,
v̄(x1 , x2 , . . . , xn ) =
−v(x1 , x2 , . . . , −xn ) if (x1 , x2 , . . . , xn ) ∈ Ω− .

If v is harmonic in Ω+ and v = 0 on T , show that v̄ is of class C 2 and is harmonic in its domain.

Solution 4. We are going to use the following version of the mean value property: a continuous function
f is harmonic if and only if for every x there exists r > 0 such that f (x) is equal to the average of f
on the sphere {z; |z − x| = ρ}, for every ρ ∈ (0, r).
The function v̄ is continuous on Ω+ ∪ T and reflection is a continuous operation, so that it is also
continuous on Ω− . We are then left with checking continuity at T . This follows from v being 0 on
T . We now check harmonicity in the whole domain, from which we will deduce C 2 -regularity. By
hypothesis we know that v is harmonic in Ω+ , thus C 2 , so that v̄ is harmonic in Ω+ ∪ Ω− . Indeed,
odd reflection of v with respect to {xn = 0} does not affect the Laplacian. Consequently, we are left
to show that v̄ is harmonic at points both in the interior and in T . Let x0 ∈ T ∩ Int(Ω+ ∪ Ω− ∪ T ).
Since Int(Ω+ ∪ Ω− ∪ T ) is open there exists r0 > 0 such that Br0 (x0 ) ⊂ Ω+ ∪ Ω− ∪ T . Thus,
Br0 (x0 ) ∩ {xn > 0} ⊂ Ω+ ∩ Br0 (x0 ).
However, we also have the converse inclusion
Ω+ ∩ Br0 (x0 ) ⊂ Br0 (x0 ) ∩ {xn > 0},
simply because Ω+ ⊂ {xn > 0}. As a result we have Br0 (x0 ) ∩ {xn > 0} = Ω+ ∩ Br0 (x0 ). Now, for
any 0 < r < r0 we have
Z Z Z Z
v̄(y) dy = v̄(y) dy + v̄(y) dy + v̄(y) dy = 0,
Br (x0 ) Br (x0 )∩Ω+ Br (x0 )∩Ω− Br (x0 )∩T

15
where the second equality follows because the second integral is minus the first one, while the third
one is over a set of measure zero. However, v̄|T = 0 so that
Z
1
v̄(x0 ) = v(x) dy,
ωn rn Br (x0 )

for x0 ∈ T ∩ Int(Ω+ ∪ Ω− ∪ T ). In particular, the function v̄ is continuous and satisfies the mean value
property. We deduce that it is harmonic in its domain and consequently C 2 .
Exercise 5 (Boundary Version Absorbing Lemma). Given θ ∈ (0, 1) and µ ∈ R, there exists δ =
+
δ(n, θ, µ) ∈ (0, 1) and C = C(n, θ, µ) > 0 such that, if R > 0, B = {Bρ (y); Bρ (y) ⊂ BR (0)},
+ + n + + +
B = {Bρ (y); y = 0, Bρ (y) ⊂ BR (0)} and S : B ∪ B → R is a subadditive function satisfying
+
ρµ S(Bθρ (y)) ≤ δρµ S(Bρ+ (y)) + γ, if Bρ+ (y) ∈ B + ,
and
ρµ S(Bρθ (y)) ≤ γ, if Bρ (y) ∈ B,
then
+
Rµ S(BθR (0)) ≤ Cγ.
Here Bρ+ (y) = Bρ (y) ∩ Rn+ .
PN
Solution 5. Recall that S : B ∪ B + → R being a subadditive function means S(A) ≤ j=1 S(Aj ),
SN
whenever A, A1 , . . . , AN ∈ B ∪ B + with A ⊂ j=1 Aj .
Exercise 6. Let α ∈ (0, 1) and let u ∈ C 2,α (Rn ) be a solution to Lu = 0 on Rn , where L = aij Dij is
a constant coefficient elliptic operator. If u satisfies the growth condition |u(x)| ≤ γ(1 + |x|k ) for some
constant γ and all x ∈ Rn , show that u must be a polynomial of degree at most k.
Solution 6. Assume to start that k = 2. Fix a radius R > 0 and consider the ball BR (0). From the
scale invariant interior Schauder estimates we have the bound
|u|02,α;BR (0) ≤ C |u|0;BR (0) + R2 |f |0;BR (0) + R2+α [f ]α;BR (0) = C|u|0;BR (0) ,

as our inhomogeneous term is zero, the constant C is independent of R and where

k
X
|u|0k,α;Bρ (x0 ) := ρj |Dj u|0;Bρ (x0 ) + ρk+α [Dk u]α;Bρ (x0 ) .
j=0

In particular,
[D2 u]α,BR (0) ≤ CR−2−α |u|0;BR (0) ≤ γR−2−α Cγ(1 + R2 ) → 0,
as R → +∞, thus proving that D2 u must be constant and consequently u must be a polynomial of
degree at most 2. The same holds for arbitrary k ∈ N. By higher regularity interior Schauder estimates
we know that u is smooth as the coefficients are of L are. We can also apply the corresponding inequality
to find the bound
Rk+α [Dk u]α;BR (0) ≤ C(k)|u|0;BR (0) ,
where C(k) is a constant depending on k, from which we deduce, thanks to |u(x)| ≤ γ(1 + |x|k ), that
Dk u is constant and consequently that u must be a polynomial of degree at most k.
Exercise 7. Let α ∈ (0, 1) and aij , bi , c ∈ C 0,α (B1 (0)) satisfy the usual condition
|aij |0,α;B1 (0) + |bi |0,α;B1 (0) + |c|0,α;B1 (0) ≤ β,
for some constant β and for all i, j ∈ {1, . . . , n} and aij ζi ζj ≥ λ|ζ|2 for some constant λ > 0 and all
x ∈ B1 (0) and ζ ∈ Rn . The interior Schauder estimates proven in lecture are


|u|2,α;B1/2 (0) ≤ C |u|0;B1 (0) + |Lu|0,α;B1 (0) ,

for any function u ∈ C 2,α (B1 (0)), where Lu = aij Dij + bi Di + c and C = C(n, α, β, λ).

16
 (i) Give examples to show that this estimate cannot be improved to either of the following three cases.
(a) |u|2,α;B1/2 (0) ≤ C|Lu|0,α;B1 (0) ;


(b) |u|2,α;B1/2 (0) ≤ C |u|0;B1 (0) + |Lu|0;B1 (0) ;


(c) |u|2,α;B1 (0) ≤ C |u|0;B1 (0) + |Lu|0,α;B1 (0) .
(ii) If u is harmonic on B1 (0), use the mean value property to show directly, in other words without
relying on the above, that there is a constant C1 := C1 (n) such that

|u|2,α;B1/2 (0) ≤ C1 kukL1 (B1 (0)) .

(iii) In case of a general L as before, does there exist C2 = C2 (n, α, β, λ) such that


|u|2,α;B1/2 (0) ≤ C2 kukL1 (B1 (0)) + |Lu|0,α;B1 (0) ?

Solution 7. (i) We shall see that all three estimates are false, thus showing that the theorem proven
in class is sharp.
(a) Consider the polynomial u(x) = x on R for the operator L = ∆. Then Lu = 0, while
|u|2,α;B1/2 (0) ≥ 1 by definition, giving the desired contradiction. Notice that we could have
added first order terms to the operator under consideration.
1
(b) Consider again the Laplacian L = ∆ and u(x) = N3 sin(N x), in dimension one. Simply by
differentiating we obtain the bounds
1 1 1
|u(x)| ≤ , |Du(x)| ≤ and |D2 u(x)| ≤ .
N3 N2 N
However, for the α-Hölder seminorm of second derivatives we have the following

1 1 1
[D2 u]α;B1/2 (0) = sup α N
sin(N x) − sin(N y)
x,y∈B1/2 (0) |x − y| N
x6=y

= sup |x − y|1−α | cos(N ξ)|

x,y∈B1/2 (0)
x6=y
1
≥ sup |x − y|1−α
2 |x|,|y|<π/(12N )
x6=y
1  π 1−α
≥ ,
2 6N
where the second equality follows from the mean value theorem for ξ ∈ [x, y] and the first
inequality follows by properties of the supremum and bounding from below the cosine term.
The last inequality is due again to the properties of the supremum. Were the claimed in-
equality true we would then get from the above
 
1  π 1−α 1 1
≤C + 3 ,
2 6N N N

for all N ∈ N and where C does not depend on N . The above can be written in the form
 
1 α 1−α 1
N (π/6) ≤ C 1 + 2 ≤ 2C,
2 N

from which we get the desired contradiction.

17
 (c) Consider the boundary value problem
(
Luk = 0 in B1 (0),
uk = φk on ∂B1 (0),

with uk ∈ C 2,α (B1 (0)) and with boundary values satisfying |φk |0;B1 (0) < ∞ uniformly, while
|φk |2,α;∂B1 (0) diverges as k → +∞. From the claimed estimate we would obtain


|uk |2,α;B1 (0) = |uk |2,α;B1 (0) ≤ C |uk |0;B1 (0) + |Luk |0,α;B1 (0) = C|uk |0;B1 (0) ,

where the first and last identity follows by the definition of | · |2,α;Ω -norm for a certain domain
Ω and the fact that the uk ’s belong to C 2,α (B1 (0)). Moreover, the left-hand side satisfies

|φk |2,α;∂B1 (0) ≤ |uk |2,α;B1 (0) ,

from which we deduce

|φk |2,α;∂B1 (0) ≤ C|uk |0;B1 (0) .
If we further assume that c ≡ 0, i.e. that we have no zeroth order terms, we can apply
maximum principle to obtain

|φk |2,α;∂B1 (0) ≤ C|φk |0;∂B1 (0) ,

from which we infer the desired contradiction thanks to the hypothesis on the sequence {φk }k .

Alternatively, a more concrete counterexample is given by the function u(x, y) = enx cos(ny),
for n ∈ N, with the operator L = ∆. Notice that u is harmonic, so that the claimed inequality
becomes
|u|2,α;B1 (0) ≤ C|u|0;B1 (0) .
More specifically, the right-hand side of the above is given by en , whereas the left-hand side
is bounded from below by nen . This inequality being valid for all n ∈ N, we infer the desired
contradiction.

(ii) Consider x ∈ B1/2 (0) and calculate using the mean value property

2n
Z Z
1
|u(x)| = u(y) dy ≤ |u(x)| dy = C1 (n)kukL1 (B1 (0)) .
ωn (1/2)n B1/2 (x) ωn B1/2 (x)

Notice then that derivatives of u are harmonic as well, so that, for i ∈ {1, 2, . . . , n} we obtain

2n
Z
|Di u(x)| = Di u(y) dy
ωn B1/2 (x)

2n
Z
= ui (y)νi dHn−1 (y)
ωn ∂B1/2 (x)

2n
Z
≤ |u(y)| dHn−1 (y) ≤ C1 (n)kukL1 (B1 (0)) ,
ωn ∂B1/2 (x)

where νi is the ith component of the out unit normal. Thus,

|u|0;B1/2 (0) + |Du|0;B1/2 (0) ≤ C1 (n)kukL1 (B1 (0)) .

18
 We are now left with second derivatives and the α-Hölder seminorm. Again, Dij u is harmonic
for all i, j ∈ {1, 2, . . . , n} so that for x ∈ B1/2 (0) we obtain
Z
1
|Dij u(x)| = Dij u(y) dy
ωn (1/4)n B1/4 (x)

4n
Z
= Di (Dj u)(y) dy
ωn B1/4 (x)

4n
Z
= Dj uνi dHn−1 (y)
ωn ∂B1/4 (x)
4n
Z
≤ |Dj u| dHn−1 (y)
ωn ∂B1/4 (x)
4n
≤ |∂B1/4 (x)||Dj u|0;∂B1/4 (x) .
ωn

we can then ensure that |Du|0;B3/4 (0) is bounded by C1 (n)kukL1 (B1 (0)) repeating the same proof
as before. Thus,
4n 4n
sup |Dij u(x)| ≤ C(n) sup |Dj u|0;∂B1/4 (x) ≤ C(n)|Du|0;B3/4 (0) ,
x∈B1/2 (0) ωn x∈B1/2 (0) ωn

where the last term is then bounded by C1 (n)kukL1 (B1 (0)) . To conclude we need to bound the
term [D2 u]α;B1/2 (0) . Start by noticing

|x − y| = |x − y|1−α |x − y|α ≤ diam(Ω)1−α |x − y|α = 21−α |x − y|α ,

so that it is enough to bound the term [D2 u]1;B1/2 (0) . Indeed,

|D2 u(x) − D2 u(y)| |D2 u(x) − D2 u(y)|

[D2 u]α;B1/2 (0) = sup ≤ 21−α
sup = 21−α [D2 u]1;B1/2 (0)
x,y∈B1/2 (0) |x − y|α x,y∈B1/2 (0) |x − y|
x6=y x6=y

Applying the mean value theorem we infer

|D2 u(x) − D2 u(y)| ≤ |D3 u|0;B1/2 (0) |x − y|.

Bounding the term |D3 u|0;B1/2 (0) in terms of C1 (n)kukL1 (B1 (0)) exactly as before we get the
required conclusion.
(iii) Recall that by Ehrling’s Lemma, see Lemma 1, we have that for all p ∈ [1, ∞) and for every δ > 0
there exists C = C(δ, p) such that for all u ∈ C 2,α (B1 (0)) we have

|u|0;B1 (0) ≤ δ|u|2,α;B1 (0) + C(δ, p)kukLp (B1 (0)) .

Thus,


|u|2,α;B1/2 (0) ≤ C(α, n, L) |Lu|0,α;B1 (0) + δ|u|2,α;B1 (0) + C(δ, p)kukLp (B1 (0)) .

Choosing δ > 0 small enough we can apply Simon’s absorbing lemma to get the desired result.
In particular the result holds for p = 1.

19
1 Example Sheet 3
• Discussion on uniqueness and Peano/Picard-Lindelof theorems to get uniqueness for associated
ODEs;
• Odd extension for question of stricly elliptic;
• Explain the WHY in Q5;

Exercise 1. Let α ∈ (0, 1) and f ∈ C 0,α (B), where B = B1 (0) is the open unit ball in Rn . Give
an explicit construction of a sequence of functions fk ∈ Cc∞ (Rn ) such that the sequence { fk |B }k∈N is
uniformly bounded in C 0,α (B) and fk → f uniformly in B. Use this, together with the fact that the
Dirichlet problem ∆u = g in B, with boundary values u = φ on ∂B has a solution u ∈ C ∞ (B) for any
g ∈ C ∞ (B) and any φ ∈ C ∞ (B), to show that for any φ ∈ C 0 (B), the Dirichlet problem ∆u = f in
B, with boundary values u = φ on ∂B has a unique solution u ∈ C 2,α (B) ∩ C 0 (B). Show also that this
solution is in C 2,α (B) if the boundary value is φ ∈ C 2,α (B).
Solution 1. Extend f to f˜ ∈ Cc0,α (Rn ), and φ to φ̃ ∈ Cc0 (Rn ). For the latter any continuous extension
theorem, e.g. Tietze extension theorem, does the job. On the other hand, for the former use the
following expression
f (y) = inf (f (y) + C|x − y|α ) ,
y∈B

where C is the Hölder constant of f . The extension f is uniformly continuous. To make it compactly
supported, whence getting f˜, one can then multiply by a smooth cut-off function. Choose η ∈ Cc∞ (Rn )
n
R
satisfying η ≡ 0 on R \ B1 (0), non-negativity everywhere η ≥ 0, and Rn η = 1. Let ησ (x) =
σ −n η(x/σ). Mollify f˜ and φ̃ with respect to such η and choose a sequence σk → 0+ . We denote
Z Z
fk (x) := f˜(y)ησk (x − y) dy = f˜(x − y)ησk (y) dy,
Rn Rn

and Z Z
φk (x) := φ̃(y)ησk (x − y) dy = φ̃(x − y)ησk dy.
Rn Rn
The functions fk and φk are smooth for all k ∈ N, thus, we have existence of a smooth function uk
solving the boudnary value problem ∆uk = fk in B and uk = φk on ∂B. Furthermore, by definition
of convolution, we have fk → f uniformly in B and φk → φ uniformly in B. We can further deduce
Z
|fk (x) − fk (z)| ≤ |f˜(x − y) − f˜(z − y)|ησk dy ≤ [f˜]0;α |x − y|α ,
Rn

which, combined with the usual convolution bound supB |fk | ≤ supB |f˜| for all k ∈ N, implies |fk |0;α ≤
|f˜|0;α . Similarly, sup |φk | ≤ sup |φ̃|. Note now that

∆(uk − ul ) = fk − fl

in B, as well as uk − ul = φk − φl on ∂B. The maximum principle a priori estimate implies

sup |uk − ul | ≤ sup |φk − φl | + C sup |fk − fl |.

B ∂B B

The sequence {uk }k∈N is therefore uniformly Cauchy and we deduce existence of a limit u ∈ C 0 (B)
such that uk → u uniformly in B, as continuity is preserved by uniform limits. Moreover, by taking
limits we have u = φ on ∂B as well. The idea is now to upgrade the regularity of u via interior
Schauder estimates. More precisely, for any B 0 b B we have
     
|uk |2,α;B ≤ C sup |uk | + |fk |0,α;B ≤ C sup |φk | + |f˜|0,α;Rn ≤ C sup |φ| + |f˜|0,α;Rn ,
B ∂B ∂B

20
where C = C(B 0 , B). This bound is again independent of k, so that by Arzela-Ascoli we have the
2
existence of a, not relabelled, subsequence {uk } converging in Cloc to some limit in C 2,α (B). This limit
must be u by uniqueness of limits. Taking again limits in the equation ∆uk = fk we obtain ∆u = f
satisfying the required boundary values. Assuming φ ∈ C 2,α (B), the argument follows the same line,
up to replacing interior Schauder estimates with global ones.
Remark 6. Note that relaxing the regularity hypothesis on f is not possible. One can either argue as
Exercise 1 of Example Sheet 2 or can consider the following counterexample in the n dimensional ball
BR (0), with R < 1:
x21 − x22
 
n+2 1
f (x) := + .
2|x|2 (− log |x|)1/2 2(− log |x|)3/2
For further details see [2, Pag. 65].
Exercise 2 (C 1,α -Schauder Theory). Let α ∈ (0, 1) and let u ∈ C 1,α (B1 (0)) be a weak solution of
the divergence form equation Di (aij Dj u) + bi Di u + cu = f in B1 (0), where aij ∈ C 0,α (B1 (0)) and
bi , c, f ∈ L∞ (B1 (0)) with |aij |0,α;B1 (0) + |bi |0;B1 (0) + |c|0;B1 (0) ≤ β and aij (x)ζ i ζ j ≥ λ|ζ|2 for some
constants β, λ > 0 and all x ∈ B1 (0) and ζ ∈ Rn . By making appropriate modifications to the scaling
argument presented in the lecture, the one that proves the C 2,α Schauder estimates, show that


|u|1,α;B1/2 (0) ≤ C |u|0;B1 (0) + |f |0;B1 (0) ,

for some constant C = C(n, α, β, λ).

Solution 2. Let us simplify the statement of the exercise by assuming Lu = Di (aij Dj u). We proceed
as in the proof of interior C 2,α -Schauder estimates given in class. It suffices to prove that for every
δ > 0 we have existence of a constant C > 0 such that

[Du]α;B1/2 ≤ δ[Du]α;B1 + C (|u|1;B1 + |f |0;B1 .)

We argue by contradiction. Assume that there exists a sequence of operators Lk with coefficients
aij
k uniformly bounded in C
1,α
(B 1 (0)), i.e. satisfying |aij
k |0,α;B1 (0) ≤ β, and a sequence of functions
∞
{fk }k∈N in L (B1 (0)) for which

[Duk ]α;B1/2 ≥ δ[Duk ]α;B1 + k (|uk |1;B1 + |fk |0;B1 ) . (7)

The proof is in two steps, first rescale the uk ’s around points where the Hölder coefficient of Du
diverges, and then show that this rescaled sequence converges to a weak solution of a PDE with
constant coefficients. Derive then a contradiction using a Liouville type argument. We start with a
claim. For pk , qk ∈ B1 (0) so that
1
|Duk (pk ) − Duk (qk )| ≥ |pk − qk |α [Duk ]α;B1 ,
2
we have |pk − qk | → 0, as k → ∞. Assume without loss of generality that [Duk ]α;B1 ≤ M for all k ≥ 0.
We could even do more and normalise the whole sequence by [Duk ]α;B1 to get [Duk ]α;B1 = 1. Rescale
uk to
1
ũk (x) = 1+α uk (pk + ρk x),
ρk M
and substract its derivative at the origin to get

Uk (x) = ũk (x) − ũk (0) − Dũk (0)x,

which is equivalenty defined as

1
Uk (x) = (uk (pk + ρk x) − uk (pk ) − ρk Duk (pk )x) .
ρ1+α
k M
In order to make the notation less convoluted we shall drop reference to k, and replace Uk with U .
The function U satisfies

21
 (i) U is defined at least on B1/2ρ ⊂ Rn ;
(ii) U (0) = 0 and DU (0) = 0;
(iii) [DU ]α;B1/2ρ ≤ 1;
(iv) In the ball B1/2ρ , the function U is a weak solution of the following PDE

Di (aij (p + rx)Dj U (x)) = r1−α f (p + rx) + r−α Di (aij (p + rx) − aij (p))Dj u(p) .

We only prove the last claim, the first ones follow trivially from the definition of U . Start then with
the left-hand side of the equation and test it against a function Φ ∈ Cc∞ (B1/2ρ ) to obtain
Z
aij (p + rx)Dj U (x)Di Φ(x) dx.
B1/2r

Then, note that

1
Dj U (x) = (Dj u(p + rx) − Dj u(p)) ,
ρα M
so that we obtain
Z
aij (p + rx)Dj U (x)Di Φ(x) dx
B1/2ρ
Z Z
1 1
= α
aij (p + rx)Dj u(p + rx)Di Φ(x) dx − α
aij (p + rx)Dj u(p)Di Φ(x) dx.
ρ M B1/2ρ ρ M B1/2ρ

Let now y = p + ρx and consider φ ∈ Cc∞ (B1/2 (p)) so that φ(p + ρx) = Φ(x). Consequently, differen-
tiating Φ we obtain Di Φ(x) = ρDi φ(p + ρx) and the above expression becomes
Z
aij (p + rx)Dj U (x)Di Φ(x) dx
B1/2ρ
Z Z
1 ij 1
= a (y)Dj u(y)Di φ(y) dy − aij (y)Dj u(p)Di φ(y) dy.
ρα−1+n M B1/2 (p) ρα−1+n M B1/2

We now treat the two terms separately. From the weak PDE, we have that the first term is equal to
Z
1
− α−1+n f (y)φ(y) dy.
ρ M B1/2 (p)

As far as the second term is concerned, we can add a constant term to obtain
Z Z
1 ij 1
aij (y) − aij (p) Di u(p)φ(y) dy.


α−1+n
a (y)D j u(p)Di φ(y) dy = α−1+n
ρ M B1/2 (p) ρ M B1/2 (p)

Changing variables back to B1/2ρ (0) we infer

Z Z !
1 1 ij ij


f (p + ρx)Φ(x) dx − a (p + ρx) − a (p) Dj u(p)Di Φ(x) dx
ρα−1 M B1/2ρ (0) ρ B1/2ρ (0)

from which the claim follows. The sequence Uk is locally uniformly bounded in C 2,α (Rn ), and we
may thus diagonally extract a subsequence converging locally strongly in C 2 (Rn ) to a function U ∈
C 2,α (Rn ). The limiting function U satisfies the same Hölder bound as the Uk in the sequence, i.e.
[DUk ]α;Rn ≤ 1. On the other hand, the sequence pk − qk /|pk − qk | is bounded and we may thus extract
again a convergent subsequence to a point x0 , where Du(x0 ) ≥ 1/2. Consequently, the function U is

22
not identically zero. We can derive a contradiction with the help of the following result. We claim
that the function U is a weak solution of the following PDE in Rn :
Di (aij (0)Dj U ) = 0.
We start from the PDE satisfied by the Uk . There are essentially three steps to prove such claim. We
have to show that
Di (aij ij
k (pk + ρk x)Dj Uk (x)) → Di (a (0)Dij U ),
and
1
f (pk + ρk x) → 0,
ρ1−α
k
together with  
ij ij
ρ−α
k Di (ak (pk + ρk x) − ak (pk ))Dj uk (pk ) → 0,

all of them in the weak sense. They all follow from dominated convergence, whence we only explain
the first one. Fix a test function Φ ∈ Cc∞ (Rn ) with support contained in BR . Take k large enough so
that 1/2ρk > R. The integrand in
Z
aij
k (pk + ρk x)Dj Uk (x)Di Φ(x) dx
Rn

is bounded uniformly by
|aij α
k |0;B1 · |Uk |1,BR · |Φ|1;R ≤ βR |Φ|1;R .
n n

Moreover, it converges pointwise to

aij (0)Dj U (x)Di Φ(x),
thus concluding the proof.
Exercise 3. Let Ω be a bounded domain in Rn , α ∈ (0, 1) and L = aij Dij + bi Di + c be strictly elliptic
in Ω with coefficients in C 0,α (Ω) and c ≤ 0. Let f ∈ C 0,α (Ω).
(i) Suppose v ∈ C 2 (Ω). Show that v is a subsolution of Lu = f in Ω if and only if v ≤ vB in B for
every open ball B with B ⊂ Ω, where vB is the unique function in C 2 (B) ∩ C 0 (B) satisfying
(
LvB = f in B,
(8)
vB = v on ∂B.

(ii) If u ∈ C 2 (Ω) ∩ C 0 (Ω) satisfies Lu = f in Ω, show that for every x ∈ Ω,

u(x) = sup{v(x); v ∈ C 2 (Ω) ∩ C 0 (Ω), v is a subsolution to Lu = f in Ω, v ≤ u on ∂Ω}.

Solution 3. Note the sign of the zeroth order term c.

(i) Start by assuming that v is a subsolution of Lu = f , i.e. Lv ≥ f . Then, L(v − vB ) ≥ 0 and
consequently
sup (v − vB ) ≤ sup (v − vB ) = 0,
B ∂B

from which we conclude. Conversely, assume v ≤ vB in B where vB ∈ C 2 (B) ∩ C 0 (B) satisfies

the system in (8). Assume, for the sake of contradiction, that Lv < f in Ω and construct vB in
a given ball. Then, L(vB − v) > (f − f ) = 0 and vB − v = 0, so that by maximum principle we
obtain the desired contradiction with v ≤ vB (as we would obtain vB − v > 0).
(ii) Note that u is itself a subsolution of Lu = f satisfying u ≤ u on ∂Ω. Consequently, by definition
of supremum, we have the inequality
u(x) ≤ sup{v(x); v ∈ C 2 (Ω) ∩ C 0 (Ω), v is a subsolution to Lu = f in Ω, v ≤ u on ∂Ω}.
To establish the converse consider v ∈ C 2 (Ω) ∩ C 0 (Ω) a subsolution Lv ≥ f with v ≤ u on ∂Ω.
Thus, L(v − u) ≥ 0, so that maximum principle implies supΩ (v − u) ≤ sup∂Ω (v − u)+ = 0, i.e.
v ≤ u everywhere. Taking the supremum allows to conclude the proof.

23
Exercise 4. Let Ω be a bounded domain in Rn , α ∈ (0, 1) and let L = aij Dij + bi Di + c be strictly
elliptic in Ω with aij , bi and c in C 0,α (Ω) and c ≤ 0. Let f ∈ C 0,α (Ω).
(i) If v, w ∈ C 0 (Ω) are respectively a subsolution and a supersolution to Lu = f in Ω with v ≤ w on
∂Ω, show that v ≤ w in Ω. Show in fact that either v < w in Ω or v = w in Ω.
(ii) If v1 , v2 ∈ C 0 (Ω) are subsolutions to Lu = f in Ω, show that v = max{v1 , v2 } is again a
subsolution to Lu = f in Ω.
(iii) If v ∈ C 0 (Ω) is a subsolution to Lu = f in Ω and B is an open ball such with B ⊂ Ω and V is
the L-lift of v in B, show that V is a subsolution to Lu = f in Ω.
Solution 4. Note again the sign of the zeroth order term c. To solve this exercise we are going to use
the characterization of subsolutions proven previously.
(i) Compute L(v − w) = Lv − Lw ≥ 0 in Ω with boundary values satisfying v − w ≤ 0. Thus, setting
ψ = v − w we obtain Lψ ≤ 0 in Ω and ψ ≥ 0 on the boundary. Consequently, by maximum
principle, we obtain either φ = 0 in Ω or ψ < 0 in Ω, giving the desired conclusion.
Alternatively, one can proceed via the characterization we have established in the previous ex-
ercise. For the sake of completeness we shall give the proof. Consider B b Ω and let ψB be
the solution of the boundary value problem LψB = 0 in B and ψB = ψ on ∂B. Let then
vB and wB be the corresponding solutions of the boundary value problem for v and w. Thus,
ψB = ψ = v −w = vB −wB on ∂B so that by maximum principle we conclude that ψB = vB −wB
in the whole B. On the other hand, v ≤ vB and w ≥ wB (as w is a supersolution) in B, so that
ψ = v − w ≤ vB − wB = ψB in B, as required.
(ii) Consider an open ball B b Ω and the corresponding solution of
(
LvB = f in B,
(9)
vB = v on ∂B.

In particular, v ≥ v1 on ∂B, likewise for v2 . Note that L(vB − v) ≤ f − f = 0, from which we

deduce, by maximum principle, vB ≥ v1 in B. As the same holds for v2 we obtain that vB ≥ v
in B and the proof is concluded thanks to the characterization we have for subsolutions.
(iii) We start by recalling the definition of L-lift. Let B be a ball compactly contained in Ω. Let
uB ∈ C 2,α (B) ∩ C 0 (B) be the unique solution to
(
LuB = f in B,
uB = u on ∂B.
Define then (
uB on B,
U :=
u on Ω \ B.
Then U is called the L-lift of u. We wish to prove that the L-lift V of a subsolution v is a
subsolution as well. Consider a new ball B 0 b Ω and let VB 0 be the solution LVB 0 = f in B 0
and VB 0 = V on ∂B 0 . Without loss of generality assume that B 0 ∩ B 6= ∅ and B 0 6⊂ B, the two
converse cases being immediate. By construction VB 0 ≥ V on ∂B 0 ∩ B. As v is a subsolution, we
additionally have VB 0 ≥ v in B 0 (by the usual argument L(VB 0 − v) ≤ 0 in B 0 and VB 0 ≥ v on
∂B 0 , so that maximum principle allows to conclude). Since V = v on ∂B ∩ B 0 , we have VB 0 ≥ V
on ∂B ∩ B 0 , and indeed on B 0 \ B. Therefore, VB 0 ≥ V on the boundary of B 0 ∩ B. Maximum
principle implies then that VB 0 ≤ V on B ∩ B 0 , thus concluding the proof.
Exercise 5. Give an example of a bounded domain Ω and a uniformly elliptic operator L = aij Dij +
bi Di + c with coefficients in C 0,α (Ω) for some α ∈ (0, 1) to show that without the condition c ≤ 0, the
problem (
Lu = f in Ω,
u=φ on ∂Ω,

24
need not be solvable in C 2,α (Ω) ∩ C 0 (Ω) for every f ∈ C 0,α (Ω) and φ ∈ C 0 (∂Ω). Give an example in
which c ≤ 0 fails and yet for each f ∈ C 0,α (Ω) and φ ∈ C 0 (∂Ω), the problem
(
Lu = f in Ω,
u=φ on ∂Ω,

has a unique solution u ∈ C 0 (Ω) ∩ C 2,α (Ω).

Solution 5. Let Ω = (0, π) and L = ∆ + 1. Then, c ≡ 1 > 0 and the operator L is uniformly elliptic.
Moreover, the coefficients are in C 0,α (Ω) for any α ∈ (0, 1). Consider then the problem

Lu = 0
 in (0, π),
u(0) = 0,

u(π) = 1.


The solution to this problem is of the form u(x) = A sin(x) + B cos(x), so that the boundary value
u(0) = 0 gives B = 0. However, we would also have

1 = u(π) = A sin(π) = 0,

giving the desired contradiction. Hence, the above boundary value problem has no solution.
For the second part of the exercise let L as above and consider the domain Ω = (0, π/2). It can
then be shown as above that the homogeneous problem

Lu = 0
 in (0, π/2),
u(0) = 0,

u(π/2) = 0


has only the trivial solution u ≡ 0 (WHY). Thus, by the Fredholm alternative we can solve the
associated inhomogeneous problem (
Lu = f in Ω,
u=φ on ∂Ω,
for each f ∈ C 0,α (Ω) and φ ∈ C 0 (∂Ω), finding a solution u ∈ C 2,α (Ω).
Exercise 6. Let α ∈ (0, 1) and Ω ⊂ Rn be a bounded C 2,α -domain. Let L = aij Dij + bi Di + c be
strictly elliptic in Ω with aij , bi , c ∈ C 0,α (Ω). Show that the set K of functions u ∈ C 2,α (Ω) satisfying
Lu = 0 in Ω and u = 0 on ∂Ω is a finite dimensional subspace of C 2,α (Ω).
Solution 6. By the proof of the Fredholm alternative we know that Lu = 0 if and only if u = −σL−1 σ (u)
and also |u|2,α;Ω = |σ||L−1
σ (u)| ≤ C|u| 0,α;Ω . Consequently, considering the unit ball in the C 0,α
-norm
gives a uniform bound on the C 2,α -norm. Arzela-Ascoli gives a convergent subsequence, from which we
deduce that the unit ball in (K, |·|0,α;Ω ) is compact and, consequently, the space K is finite dimensional.
As a finite dimensional space all norms will then be equivalent, in particular when equipped with the
C 2,α -norm.
The above exercise is part of a more general result that goes under the name of Fredholm alternative
and whose statement follows.
Theorem 2 (Fredholm Alternative). Let α ∈ (0, 1) and Ω ⊂ Rn be a bounded C 2,α -domain. Let
L = aij Dij + bi Di + c be strictly elliptic in Ω with aij , bi , c ∈ C 0,α (Ω). Then,
(a) Either, the homogeneous problem Lu = 0 in Ω with Dirichlet boundary condition u = 0 on ∂Ω has
only the trivial solution, in which case inhomogeneuous problem Lu = f in Ω with u = φ on ∂Ω
has a unique solution in C 2,α (Ω) for all f ∈ C 0,α (Ω) and φ ∈ C 2,α (Ω).
(b) Or, the homogeneous problem Lu = 0 in Ω with Dirichlet boundary condition u = 0 on ∂Ω has
non-trivial solutions which form a finite dimensional subspace of C 2,α (Ω).

25
Exercise 7. Consider the rectangle Q = (0, π) × (0, 1) ⊂ R2 and let u ∈ C 0 (Q) ∩ C 2 (Q) satisfy the
equation uxx + y 2 uyy = 0 in Q.
(i) Assume that u(0, y) = u(π, y) = 0 for y ∈ (0, 1). Prove that u must be of the form
X
u(x, y) = An fn (y) sin(nx),
n∈Z
√
(1+ 1+4n2 )/2
for constants An ∈ R, where fn (y) = y .
(ii) Show that there exists φ ∈ C 0 (∂Q) for which the Dirichlet problem
(
uxx + y 2 uyy = 0 in Q,
u=φ on ∂Q,

has no solution u ∈ C 0 (Ω) ∩ C 2 (Q).

(iii) Why does this not contradict the general existence theory established in the lecture?
Solution 7. We prove the three claims separately.
(i) Let u ∈ C 0 (Q) ∩ C 2 (Q) be a solution of the equation uxx + y 2 uyy = 0 in Q. Fix y ∈ (0, 1) and
consider the Fourier expansion of u(·, y) seen only as a function of x. It reads
∞
g0 (y) X
u(x, y) = + (fn (y) cos(2nx) + gn (y) sin(2nx)) ,
2 n=1

where the coeffiecients are given by

2 π π
Z Z
2
fn (y) = u(x, y) cos(2nx) dx and gn (y) = u(x, y) sin(2nx) dx
π 0 π 0

The boundary condition u(0, y) = u(π, y) = 0, for y ∈ (0, 1), implies fn (y) = 0 for all y ∈ (0, 1),
as well as g0 (y) = 0, so that the above expression becomes
∞
X
u(x, y) = gn (y) sin(2nx).
n=1

We now claim that the coefficients gn ’s satisfies the ordinary differential equation y 2 gn00 − 4n2 gn =
0, for all n ∈ N. Indeed, differentiating directly the expression for gn (x) and using the relation
uxx + y 2 uyy = 0 in Q immediately allows to conclude. An inspection of the ordinary differential
equation at hand will give that a solution is given by the following
( √ √ )
αn 1 − 1 + 16n2 1 + 1 + 16n2
gn (y) = y with αn ∈ , ,
2 2

up to a multiplicative constant An ∈ R. As u is assumed to be bounded in Q up to the boundary,

we can exclude αn < 0, so that u takes the required form
∞ √
X 1+ 1+16n2
u(x, y) = An y 2 sin(2nx),
n=1

for all (x, y) ∈ Q. Alternatively, we can use separation of variables and look for a solution of the
form u(x, y) = X(x)Y (y). We would then obtain

Y 00 2 X 00
y + = 0,
Y X

26
 where k ∈ Z and whose solutions satisfy the ordinary differential equations X 00 = −n2 X and
yY 00 = n2 Y . Thus, X is a linear combination of cos(nx) and sin(nx). The boundary condition
ensures that we get only the sin terms as required. One can argue as before to find that Y has
the form of a polynomial y αn with
( √ √ )
αn 1 − 1 + 4n2 1 + 1 + 4n2
Y (y) = y with αn ∈ , .
2 2

The same boundedness argument allows to exclude αn < 1, so that we have

√
X 1+ 1+4n2
u(x, y) = An y 2 sin(nx),
n∈Z

for all (x, y) ∈ Q. There seems to be a discrepancy between the two solutions constructed. More
precisely, the Fourier series gives, apparently, less terms.
(ii) Since the gn (y) all vanish when y = 0, it suffices to prescribe as boundary value a function φ that
does not vanish on {y = 0} ∩ ∂Q to construct an unsolvable Dirichlet problem.

(iii) The operator is not strictly elliptic.

Exercise 8 (Interior Regularity of Solutions). Let Ω ⊂ Rn be open, α ∈ (0, 1) and suppose that
aij , bi , c ∈ C k,α (Ω) for some integer k ≥ 0. Suppose L = aij Dij + bi Di + c is strictly elliptic in Ω and
f ∈ C k,α (Ω) and that u ∈ C 2 (Ω) satisfies Lu = f in Ω. Show that u ∈ C k+2,α (Ω) and that for open
sets Ω0 , Ω1 with Ω0 b Ω1 b Ω, we have that

|u|k+2,α;Ω0 ≤ C (|u|0;Ω1 + |f |k,α;Ω1 ) ,

where C = C(n, k, L, α, Ω0 , Ω1 ).
Solution 8. We start by proving the exercise in the case k = 0, i.e. that a C 2 solution of Lu = f is in
fact in C 2,α (Ω). To this sake define the operator L̃ = aij Dij + bi Di and consider the equation L̃u = f˜,
where f˜ = f − cu ∈ C 0,α (Ω). Fix now a ball B = Bρ (y) b Ω. By solvability of the Dirichlet problem
for balls we obtain that (
L̃v = f˜ in B,
v=u on ∂B,

has a solution v ∈ C 2,α (B) ∩ C 0 (B). However, L̃(u − v) = 0 in B, u = v on ∂B and u − v ∈ C 0 (B).

By the weak maximum, we get that u = v in B, so that u ∈ C 2,α (B). Notice that we can apply it
since the operator L̃ has no zeroth order term. By arbitrarity of B we conclude that u ∈ C 2,α (Ω).
We shall now turn to the proof of the exercise. We work by induction on k, the case k = 0 being
the above, we now focus on k = 1. To this sake assume aij , bi , c, f ∈ C 1,α (Ω) and u ∈ C 2,α (Ω). The
latter can be done because of the k = 0 case presented above. We claim that u ∈ C 3,α (Ω) and that
the estimate
|u|3,α;Ω0 ≤ C (|u|0;Ω1 + |f |1,α;Ω1 )
holds, where Ω0 and Ω1 are open sets such that Ω0 b Ω1 b Ω. Fix a direction el , for a certain
l ∈ {1, 2, . . . , n}, and for 0 < |h| < dist(Ω0 , Ω1 ) define the difference quotient operator:

g(x + hel ) − g(x)

δl,h g(x) := ,
h
where g : Ω1 → R is an arbitrary function. Notice that δl,h g will be defined only on Ω0 . Thus, we

27
calculate
1h i
L(δl,,h u)(x) = L(u(x + hel )) − Lu(x)
h
1 h ij i
= a (x)Dij u(x + hel ) + bi (x)Di u(x + hel ) + c(x)u(x + hel ) − f (x)
h
1 h ij
= (a (x) − aij (x + hel ))Dij u(x + hel ) + (bi (x) − bi (x + hel ))Di u(x + hel )
h i
+ (c(x) − c(x + hel ))u(x + hel ) + f (x + hel ) − f (x)
= −(δl,h aij )(x)Dij u(x + hel ) − (δl,h bi )(x)Di u(x + hel ) − (δl,h c)(x)u(x + hel ) + (δl,h f )(x)
=: Fl,h (x)

where for the third equality we used Lu(x + hel ) = f (x + hel ). Whence, we have an elliptic equation
for the difference quotient: L(δl,h u) = Fl,h . Now, as f ∈ C 1,α (Ω), we have
1 1
f (x + hel ) − f (x)
Z Z
1 d
(δl,h f )(x) = = f (x + thel ) dt = (Dl f )(x + thel ) dt,
h h 0 dt 0

so that
sup |(δl,h f )| ≤ sup |Df |
Ω0 Ω1

and, for x 6= y,
Z 1
(δl,h f )(x) − (δl,h f )(y) = [(Dl f )(x + thel ) − (Dl f )(y + thel )] dt,
0

from which we deduce [δl,h f ]α;Ω0 ≤ [Dl f ]α;Ω1 . In other words, δl,h f ∈ C 0,α (Ω0 ) with the bound

|δl,h f |0,α;Ω0 = |δl,h f |0;Ω0 + [δl,h f ]α;Ω0 ≤ |Df |0;Ω1 + [Dl f ]α;Ω1 ≤ |f |1,α;Ω1 . (10)

We get similar bounds for the other coefficients:

|δl,h aij |0;Ω0 ≤ |Daij |0;Ω1 , |δl,h bi |0;Ω0 ≤ |Dbi |0;Ω1 and |δl,h c|0;Ω0 ≤ |Dc|0;Ω1 ;

and
[δl,h aij ]α;Ω0 ≤ [Dl aij ]α;Ω1 , [δl,h bi ]α;Ω0 ≤ [Dl bi ]α;Ω1 and [δl,h c]α;Ω0 ≤ [Dl c]α;Ω1 .
Therefore, by interior Schauder estimates, we infer

|δl,h u|2,α;Ω0 ≤ C (|Fl,h |0,α;Ω1 + |δl,h u|0;Ω1 ) ≤ C (|u|2,α;Ω1 + |f |1,α;Ω1 ) , (11)

where the second equality follows from the above discussion and the constant C is allowed to depend on
the coefficients aij , bi and c. For instance, arguing as in (10), we have |δl,h u|0;Ω1 ≤ |u|1,α;Ω1 ≤ |u|2,α;Ω1 .
To be even more precise we need to work with another Ω̃ satisfying Ω0 b Ω̃ b Ω1 to get the right
bounds. Note that the right-hand side of (11) does not depend on h. Thus, we have an uniform bound
for |δl,h u|2,α;Ω0 . Applying Arzela-Ascoli we obtain a subsquence (hj ) converging to zero and such that
δl,hj u → w, for a certain w ∈ C 2,α (Ω0 ). The convergence is in C 2 (Ω00 ), where Ω00 is an arbitrary open
set with Ω00 b Ω0 . However, we know that δl,hj u → Dl u, from which we infer Dl u = w ∈ C 2,α (Ω0 ),
i.e. u ∈ C 3,α (Ω0 ). The arbitrarity of Ω0 implies that u ∈ C 3,α (Ω). Combining this last statement with
(11) we get the desired estimate, thus concluding the case k = 1. Indeed, since u ∈ C 3,α (Ω) we have
|δl,h u|2,α;Ω0 → |Dl u|2,α;Ω0 for every l ∈ {1, 2, . . . , n}. Thus,

|Dl u|2,α;Ω0 ≤ C (|u|2,α;Ω1 + |f |1,α;Ω1 ) .

To obtain the desired estimate for |u|3,α;Ω0 one needs to write the norm explicitly and realize that the
above inequality suffices to draw the conclusion. As before, to be fully rigorous one might have to
consider another Ω̃ satisfying Ω0 b Ω̃ b Ω1 and restrict h further, this should give the desired bound.

28
 We now assume the statement to be true for k − 1 and we prove it for k. More precisely, we
assume aij , bi , c, f ∈ C k,α (Ω), together with u ∈ C k+1,α (Ω), and we show that u ∈ C k+2,α (Ω). The
required estimate will follow similarly. Notice that, as before, we assumed u ∈ C k+1,α (Ω) instead of
u ∈ C k+1 (Ω). This is again due to the discussion at the beginning of the proof. Start by considering
γ, a multi-index of length |γ| = k − 1. The product rule of derivatives allows us to write
X γ!
L(Dγ u) = Dγ f − Dγ−β aij Dβ Dij u + Dγ−β bi Dβ Di u + Dγ−β cDβ u ,


β!(γ − β)!
β<γ

where β < γ means βi ≤ γi for all i and βi < γi for some i. The right-hand side has the right regularity,
i.e. it is C 1,α so that we can apply the case k = 1 to conclude, as before, that Dγ u ∈ C 3,α (Ω), whence
u ∈ C k+2,α (Ω). The desired estimate holds similarly.
Remark 7. An immediate application of the above exercise is the proof of smoothness of minimal
graphs. Indeed, consider u ∈ C 2 (Ω), a solution of the minimal surface equation:
!
∇u
div p = 0 in Ω.
1 + |∇u|2

Written in non-divergence form it reads

 
ij Di uDj u
a Dij u := δij − Dij u = 0,
1 + |Du|2
p
where we have multiplied by 1 + |Du|2 to simplify the equation. The solution u being C 2 we have
by the above exercise that it is actually C 2,α . Thus, Di u ∈ C 1,α (Ω), for all i ∈ {1, 2, . . . , n} so that
aij ∈ C 1,α (Ω). By the above we obtain u ∈ C 3,α . Therefore, we infer aij ∈ C 2,α (Ω) and reasoning by
induction we have u ∈ C k,α (Ω) for all k, i.e. u is a smooth solution of the minimal surface equation.
Theorem 3 (Leray-Schauder fixed point theorem). Let X be a Banach space and T : X → X be a
continuous compact mapping, not necessarily linear. Suppose that there is a constant M > 0 such
that kxkX < M whenever x satisfies x = σT (x) for some σ ∈ [0, 1]. Then, there is x ∈ X such that
x = T (x).
Exercise 9. Let α ∈ (0, 1), Ω be a bounded C 2,α domain in Rn , aij , bi ∈ C 0,α (Ω×R×Rn ), φ ∈ C 2,α (Ω)
and suppose that the matrix [aij (x, z, p)]ij is positive definite for all (x, z, p) ∈ Ω × R × Rn . Let Q be
the quasilinear operator defined by

Qu = aij (x, u, Du)Dij u + b(x, u, Du),

and for σ ∈ [0, 1], let

Qσ u = aij (x, u, Du)Dij u + σb(x, u, Du).
Let then X = C 1,β (Ω) equipped with the usual C 1,β -norm, where β ∈ (0, 1). For v ∈ X, let T v be the
solution u to the quasilinear problem
(
aij (x, v, Dv)Dij u + b(x, v, Dv) = 0 in Ω,
u=φ on ∂Ω.

(i) Show that T is a well-defined map from X into X. Show that in fact T v ∈ C 2,αβ (Ω).
(ii) Show that the map T : X → X is continuous and compact.
(iii) Suppose that there is a fixed constant M > 0 - depending only on n, α, β, Ω and φ - such that for
any σ ∈ [0, 1] and any u ∈ C 2,αβ (Ω) solving the Dirichlet problem
(
Qσ u = 0 in Ω,
u = σφ on ∂Ω;

29
 we have that |u|1,β;Ω < M . Show that the Dirichlet problem
(
Qu = 0 in Ω,
u=φ on ∂Ω;

has a solution in C 2,α (Ω).

Remark 8. The above exercise reduces the question of solvability of the Dirichlet problem for Q in
C 2,α (Ω) to establishing for some β ∈ (0, 1), a uniform a priori bound on the C 1,β (Ω) norm of the
C 2,αβ (Ω) solutions to a one parameter family of related quasilinear problems.
Solution 9. We shall denote for the rest of the exercise ãij (x) := aij (x, v(x), Dv(x)).
(i) We start by proving that the coefficients ãij are C 0,α (Ω) for a fixed v ∈ X. Indeed,
|ãij (x) − ãij (y)| = |aij (x, v(x), Dv(x)) − aij (y, v(y), Dv(y))|
≤ |(x, v(x), Dv(x)) − (y, v(y), Dv(y))|α

α/2
= |x − y|2 + |v(x) − v(y)|2 + |Dv(x) − Dv(y)|2

α/2
≤ |x − y|2 + C|x − y|2β + C|x − y|2β
≤ C|x − y|αβ ,

where we used that aij ∈ C 0,α (Ω × R × Rn ) in the first inequality, the fact that v ∈ C 1,β (Ω)
in the second one. As far as the last one is concerned, we exploited Ω being a bounded domain
to obtain |x − y| ≤ diam(Ω)1−β |x − y|β . Thus, we see that ãij ∈ C 0,αβ (Ω) and similarly for
b̃(x) = b(x, v(x), Dv(x)). Now, as 0 < αβ < α we have C 2,α (Ω) ⊂ C 2,αβ (Ω) so that φ ∈ C 2,αβ (Ω).
Consequently, we infer that the problem
(
ãij Dij u + b̃ = 0 in Ω,
u=φ on ∂Ω;

has a solution in C 2,αβ (Ω) by general elliptic theory as long as the matrix (ãij )ij is positive
definite. This follows from (aij )ij being well-defined and v and Dv being continuous functions on
the compact set Ω, so that the domain of the whole vector (x, v(x), Dv(x)) is valued in a compact
set. Therefore, we conclude that T (v) ∈ C 2,αβ (Ω) and in particular T (v) ∈ C 1,β (Ω), so that the
map is well-defined.
(ii) We start by showing that T is compact and to this sake consider a bounded sequence {vm } ⊂
C 1,β (Ω). We claim that {T (vm )}m has a convergent subsequence. Global Schauder estimates
imply
|T (vm )|2,α;Ω ≤ C (|T (vm )|0;Ω + |vm |1,β;Ω + |φ|2,α;Ω ) .
By the weak maximum principle we have |T (vm )|0;Ω ≤ |φ|0;Ω , where this last term can then be
absorbed in |φ|2,α;Ω , thus giving a uniform bound on |T (vm )|2,α;Ω . By Arzela-Ascoli we have the
existence of a convergent subsequence of {T (vm )}m converging in C 2 to some u ∈ C 2,αβ (Ω). As
a consequence we obtain the convergence in C 1,β and that u is in C 1,β (Ω) as well. We have thus
proven that T is compact. We shall now turn our attention to showing that T is continuous.
Consider a sequence {vm }m converging in the C 1,β -sense towards a certain v and recall by the
closed graph theorem that proving continuity of T is equivalent to proving that the graph of it is
closed. By the above we know that {T (vm )}m is precompact in C 2 so that for any subsequence
{T (vmj )}mj we can find a further subsubsequence, denote it {T (vα )}α , converging in the C 2 -sense
to some u ∈ C 2,αβ (Ω). We then obtain
aij (x, v, Dv)Dij u + b(x, v, Dv) = lim aij (x, vα , Dvα )Dij T (vα ) + b(x, vα , Dvα )


α→∞
= lim 0 (12)
α→∞
= 0,

30
 where the first equality follows by C 2 convergence of T (vα ), C 1,β convergence of the vm ’s and
continuity of the coefficients. The second equality is due to T (vα ) being the solution of the
equation so that the term in the limit is constantly equal to 0. In particular, (12) means T (v) =
u, so that we have shown that every subsequence of {T (vm )}m has a further subsubsequence
converging to T (v). Whence, we must have T (vm ) → T (v), showing that T is continuous.

(iii) In order to apply the Leray-Schauder theorem suppose v = σT (v) for some σ ∈ [0, 1] and
v ∈ C 1,β (Ω). Thus, we have that v/σ solves the the boundary value problem
(
aij (x, v, Dv)Dij u + b(x, v, Dv) = 0 in Ω,
u=φ on ∂Ω;

or alternatively (
aij (x, v, Dv)Dij v + σb(x, v, Dv) = 0 in Ω,
u = σφ on ∂Ω;
which is precisely the boundary value problem Qσ u = 0 in Ω with v = σφ on the boundary
∂Ω. Whence, we can apply the Leray-Schauder fixed point theorem to infer the existence of
u ∈ C 1,β (Ω) such that T (u) = u, which means that u solves the boundary value problem
(
aij (x, u, Du)Dij u + σb(x, u, Du) = 0 in Ω,
u=φ on ∂Ω;

As T maps C 1,β (Ω) into C 2,αβ (Ω) we have that the solution u actually belongs to the latter space.
Consequently, Du ∈ C 1,αβ (Ω) so that a reasoning similar to the one in the first point allows us
to conclude that aij (x, u, Du) and b(x, u, Du) belongs to C 0,α (Ω). From general elliptic theory
we deduce u ∈ C 2,α (Ω), obtaining the required regularity on the solution and thus concluding
the exercise.

31
Example Sheet 4
• Correct typed solution of Q1.

Lemma 2 (Morrey decay lemma). If Ω ⊂ Rn is open and u ∈ W 1,2 (Ω) satisfies

Z
ρ2−n |Du|2 ≤ M 2 ρ2α ,
Bρ (y)

for some fixed constants α ∈ (0, 1], M > 0 and all balls Bρ (y) ⊂ Ω. Then, u is a.e. equal to a function
in C 0,α and satisfies, for any ball Bρ (y) ⊂ Ω, the estimate ρα [u]α;Bρ/2 (y) ≤ CM . Furthermore, one
can prove that for any Ω0 b Ω we have the bound

|u|0;Ω0 + ρα [u]α;Ω0 ≤ C M + kukL2 (Ω) ,

where C = C(n, α, Ω0 , Ω). Eventually, one could deduce the same result from the stronger estimate
Z
|u − ux,r |2 ≤ M 2 rn+2α , for any ball Br (x) ⊂ Ω,
Br (x)

where u is assumed to be in L2 (Ω) and

Z
1
ux,r = u.
|B(x, r)| Br (x)

Proof. For a proof we refer to the reader to Theorem 3.1 and Corollary 3.2 of [2].
Remark 9. For other strengthening/versions of this result we refer the reader to [2, Chapter 3].
Exercise 1 (Continuity of weak solutions in dimensions 1 and 2). Let Λ, λ > 0 be constants, and
let aij be bounded, measurable functions on the unit ball B = B1 (0) ⊂ Rn , with kaij kL∞ (B) ≤ Λ
and aij (x)ζ i ζ j ≥ λ|ζ|2 for all ζ ∈ Rn and a.e. x ∈ B. Let u ∈ W 1,2 (B) be a weak solution of
Di (aij Dj u) = 0 in B. Show that for any constant c ∈ R and any ball Bρ (y) ⊂ B, we have
Z Z
C
|Du|2 ≤ 2 |u − c|2 ,
Bρ/2 (y) ρ Bρ (y)\Bρ/2 (y)

where C = C(λ, Λ). Use this and Poincare’s inequality, for the appropriate domain, to deduce that
there exists γ = γ(λ, Λ) ∈ (0, 1) such that
Z Z
|Du|2 ≤ γ |Du|2 ,
Bρ/2 (y) Bρ (y)

whenever Bρ (y) ⊂ B and hence that

Z  2α Z
σ
|Du|2 ≤ C |Du|2 , (13)
Bσ (y) ρ Bρ (y)

for all σ ∈ (0, ρ], whenever Bρ (y) ⊂ B. Without appealing to the De Giorgi-Nash-Moser theory
presented in the lecture show the following.
(a) If n = 1, prove that u ∈ C 1/2 (B) and that [u]1/2;B1/2 ≤ C|u|0;B1 for some fixed constant C =
C(λ, Λ) independent of u.
(b) If n = 2 prove that u ∈ C 0,α (B) and that [u]α;B1/2 ≤ C|u|0;B1 for some α ∈ (0, 1) and C = C(λ, Λ)
independent of u.
Note where this simple argument fails to prove continuity of weak solutions when n ≥ 3.

32
Solution 1. To prove the first inequality use as test function φ2 u in the weak formulation of the
equation, where φ is a C 1 function that is zero on the boundary of Bρ (y). We are going to choose φ
more precisely later on. Calculating we obtain
Z
0= φ2 uDi (aij Dj u)
Bρ (y)
Z
=− Di (φ2 u)aij Dj u
Bρ (y)
Z
Di uφ2 + 2φDi φu aij Dj u


=−
Bρ (y)
Z Z
=− Di uφ2 aij Dj u − 2φuDi φaij Dj u,
Bρ (y) Bρ (y)

so that ellipticity of the aij ’s implies

Z Z
λ |Du|2 φ2 ≤ −2 φuaij Di φDj u
Bρ (y) Bρ (y)
Z Z !1/2 Z !1/2
2 2 2 2
≤ 2Λ |u||φ||Du||Dφ| ≤ 2 φ |Du| |u| |Dφ| ,
Bρ (y) Bρ (y) Bρ (y)

where the second equality follows from the bound on the aij ’s and the last one by Cauchy-Schwarz.
Dividing the left-hand side and the right-hand side by the left-hand side and squaring we get
Z Z
2 2 4
|Du| φ ≤ 2 |u|2 |Dφ|2 .
Bρ (y) λ Bρ (y)

Applying the above result to the function v = u − c for c ∈ R we further deduce

Z Z
2 2 4
|Du| φ ≤ 2 |u − c|2 |Dφ|2 .
Bρ (y) λ Bρ (y)

Choosing now (
1 on Bρ/2 (y)
φ(x) = ρ−|x|
ρ/2 on Bρ (y) \ Bρ/2 (y),
we get as desired Z Z
C2
|Du| ≤ 2 |u − c|2 ,
Bρ/2 (y) ρ Bρ (y)\Bρ/2 (y)

after noting that |Dφ| ≤ C/ρ. To obtain the second result recall Poincare’s inequality on an annulus
Z Z
|u − ux0 ,r |2 ≤ Cr2 |Du|2 ,
Br (x0 )\Br/2 (x0 ) Br (x0 )\Br/2 (x0 )

where ux0 ,r denotes the average of u over the annulus Br (x0 ) \ Br/2 (x0 ). Thus,
Z Z Z
C
|Du|2 ≤ 2 ρ2 |Du|2 = C |Du|2 ,
Bρ/2 (y) ρ Bρ (y)\Bρ/2 (y) Bρ (y)\Bρ/2 (y)

where we chose c = uρ,y . From this we further deduce, by adding on the left-hand side and right-hand
side of the above the integral of the ball of radius ρ/2,
Z Z
2
(C + 1) |Du| ≤ C |Du|2 ,
Bρ/2 (y) Bρ (y)

33
i.e. Z Z Z
C
|Du|2 ≤ |Du|2 = γ |Du|2 ,
Bρ/2 (y) C +1 Bρ (y) Bρ (y)

where γ ∈ (0, 1). We now wish to prove (13). To do so, consider σ ∈ (0, ρ] and let k ∈ N be such that
2−k ρ < σ ≤ 2−k+1 ρ and set 2α = | ln γ|/ ln(2) ∈ (0, 1). Compute,
Z Z Z
2 2 k−1
|Du| ≤ |Du| ≤ γ |Du|2
Bσ (y) B2−k+1 ρ (y) Bρ (y)
Z  2α Z  2α Z

k−1 2σ σ
= 2−2α |Du|2 ≤ |Du|2 = C |Du|2 ,
Bρ (y) ρ Bρ (y) ρ Bρ (y)

where C = 4 and where we used the bound 2−k < σ/ρ. To conclude the exercise we shall prove the
desired continuity of weak solutions in dimension 1 and 2.
(a) For n = 1 we trivially have the following from (13)
Z Z Z
2 2 2α
ρ |Du| ≤ |Du| ≤ ρ |Du|2 = M ρ2α ,
Bρ (0) Bρ (0) B1 (0)

where the first inequality follows because ρ ∈ (0, 1). To conclude we apply Morrey’s decay lemma
Rto all radii ρ2 ≤ ρ0 for some fixed ρ0 < 1. The constant M appearing in the lemma will then be
Bρ (y)
|Du| . An improved version of the reverse Poincaré’s inequality we have proven above reads
0
kDukL2 (Ω) ≤ CkukL2 (Ω0 ) for any Ω b Ω0 . Thus, bounding the L2 -norm of Du by the L2 -norm of
u and this last one by the L∞ -norm on B1 (0) we can conclude. One has to be careful with the
numerology of the radii and with the subdomains of B1 (0).
(b) The argument for n = 2 follows similarly.
This simple argument fails in dimension greater than 3 because the power ρ2−n changes from positive
to negative.
Exercise 2. Let u ∈ W 1,2 (BR (y)) be a non-negative weak solution of Di (aij Dj u) = 0 in BR (y), where
the coefficients aij are measurable and satisfy, for some constants λ, Λ > 0
n
X
|aij |2L∞ (BR (y)) ≤ Λ2 and aij ζ i ζ j ≥ λ|ζ|2 for a.e. x ∈ BR (y) and all ζ ∈ Rn .
i,j=1

Recall that in the lecture it was shown that for each p > 1, there is a constant C = C(n, λ, Λ, p) such
that !1/p
Z
−n/p p
kukL∞ (BR/2 (y)) ≤ CR u .
BR (y)

Deduce from this that for each θ ∈ (0, 1) and p > 1, the following inequality hold
Z !1/p
−n/p −n/p p
kukL∞ (BθR (y)) ≤ C(1 − θ) R u .
BR (y)

Solution 2. Consider z ∈ BθR (y) and apply the result from the lecture to the ball B(1−θ)R (z). The
choice of radii allows us to infer that we are still on the ball BR (y). Thus, we obtain

|u(z)| ≤ sup |u| ≤ C(1 − θ)−n/p R−n/p kukLp (B(1−θ)R (z)) ≤ C(1 − θ)−n/p R−n/p kukLp (BR (y)) .
B(1−θ)R/2 (z)

Passing now to the supremum over BθR (y) we obtain the desired estimate

kukL∞ (BθR (y)) ≤ C(1 − θ)−n/p R−n/p kukLp (BR (y)) .

34
Exercise 3. Prove the following three statements.
1. Show that for each β ∈ (0, 1) and α > 0, there exists a constant C = C(β, α) such that if
0 ≤ τ0 < τ1 and if f : [τ0 , τ1 ] → R is a bounded function satisfying

f (t) ≤ βf (s) + (s − t)−α γ, (14)

for some γ and all s, t ∈ [τ0 , τ1 ] with t < s, then

f (t) ≤ C(s − t)−α γ (15)

for all s, t ∈ [τ0 , τ1 ] with t < s.

2. Let the hypothesis be as in Exercise 2 and let p > 0. Show that for any R1 < R we have
Z !1/p
1 −n/p
kukL∞ (BθR1 (y)) ≤ kukL∞ (BR1 (y)) + C(1 − θ)−n/p R1 p
u , (16)
2 BR1 (y)

where C = C(n, λ, Λ, p).

3. Deduce that the result of Exercise 2 holds for all p > 0.
Solution 3. We shall proceed with the proof of the three statements.
1. To prove the claimed inequality start by fixing τ0 ≤ t < s ≤ τ1 and let τ ∈ (0, 1). Define the
sequence {tn } recursively by setting t0 = t and tn+1 = tn + (1 − τ )τ n (s − t). Notice that tn is an
increasing sequence and that
n
X
lim tn = t + (1 − τ )(s − t) lim τ k = s.
n→+∞ n→+∞
k=0

Applying repeatedly inequality (14) we obtain

γ
f (t) = f (t0 ) ≤ βf (t1 ) +
(t1 − t0 )α
γ
= βf (t1 ) + (s − t)−α
(1 − τ )α
n−1
γ X
≤ β n f (tn ) + α
(s − t)−α β k τ −kα .
(1 − τ )
k=0

Choosing then τ such that βτ −α < 1, we infer from the convergence of the above sum as n → +∞
the desired inequality
γ
f (t) ≤ (s − t)−α C(α, β) = C(s − t)−α γ.
(1 − τ )α

2. To prove (18) start by noticing that for p ∈ (0, 2) the following estimate hold
Z Z
2−p
2
u ≤ kukL∞ (BR (y)) up .
1
BR1 (y) BR1 (y)

From Exercise 2 in the case p = 2 we further infer

−n/2
kukL∞ (BθR1 (y)) ≤ C(1 − θ)−n/2 R1 kukL2 (BR1 (y)) ,

Combining the two inequality we deduce

−n/2 1−p/2 p/2
kukL∞ (BθR1 (y)) ≤ C(1 − θ)−n/2 R1 kukL∞ (BR kukLp (BR . (17)
1 (y)) 1 (y))

35
 Recall now Young’s inequality for products
1 1
ab ≤  aq + −r/q br ,
q r

for a, b,  > 0 and q, r > 1 satisfying the classical conjugate exponent identity q −1 + r−1 = 1.
Apply it to (17) with

1−p/2 2 −n/2 p/2 2 r 2−p

a = kukL∞ (BR , q= , b = C(1−θ)−n/2 R1 kukLp (BR , r= and =
1 (y)) 2−p 1 (y)) p q p
to obtain
2−p p −n/p
kukL∞ (BθR1 (y)) ≤  kukL∞ (BR1 (y)) + −r/q (1 − θ)−n/p R1 kukLp (BR1 (y)) .
p 2
1
Choosing  = 2−p we obtain the desired inequality, namely

1 −n/p
kukL∞ (BθR1 (y)) ≤ kukL∞ (BR1 (y)) + C(1 − θ)−n/p R1 kukLp (BR1 (y)) . (18)
2

3. In order to deduce the same inequality as in Exercise 2 we shall combine the results of the previous
two points. More specifically, introduce the function f (t) = kukL∞ (Bt (y)) , where t ∈ [0, R).
Notice that f (0) = 0 and that the proof of the first point works if the function f is bounded on
an open interval of the form (τ0 , τ1 ). Moreover, f is bounded on [0, R) by (18). To lighten the
notation set r = θR1 and notice that 0 < r < R1 < R, so that (18) now reads

1 C
f (r) ≤ f (R1 ) + kukLp (BR1 (y)) .
2 (R1 − r)n/p

The inequality in (15) allows us to conclude that

−n/p
kukL∞ (BθR1 ) ≤ C(1 − θ)−n/p R1 kukLp (BR1 (y)) ,

where we used r = θR1 .

1,2
Exercise 4. Let B be the open unit ball in Rn , and let u ∈ Wloc (B) be a weak solution of the strictly
elliptic equation Di (aij Dj u) = 0 in B with locally bounded coefficients aij . Use the test function
vk = min{ku+ , 1}φ, where φ ∈ W01,2 (B) with φ ≥ 0 to show that u+ is again a weak solution.

Solution 4. Start by remarking that vk = min{ku+ , 1}φ, where φ ∈ W01,2 (B) with φ ≥ 0, is in
the right function space. Indeed, by Stampacchia’s Theorem 1 the minimum and maximum of two
Sobolev functions is again a Sobolev function. Thus, u+ ∈ W 1,2 (B) and similarly for the minimum.
Furthermore, multiplying by φ we obtain the right boundary condition. Thus, vk is rightfully a test
function and we shall use to prove that u+ is again a weak solution of the strictly elliptic equation
Di (aij Dj u) = 0 in B with bounded coefficients aij . Introduce the sets Ωk = {x ∈ B; 1 ≤ ku+ } for
k ∈ N. Then,
Z Z Z Z
0≥ aij Di uDj vk = aij Di uDj (min{ku+ , 1}φ) = aij Di uDj φ + aij Di uDj (ku+ φ).
B B Ωk Ωck

Develop now the product on the right-hand side of the last integral to obtain
Z Z Z
aij Di uDj φ + k φaij Di uDj u+ + k u+ aij Di uDj φ ≤ 0.
Ωk Ωck Ωck

1 The theorem states that the composition of Lipschitz functions and Sobolev functions is again a Sobolev function.

Here compose u with the maximum and minimum functions, which are trivially 1-Lipschitz.

36
We shall consider separately the two integrals over Ωck . Indeed,
Z Z Z
k φaij Di uDj u+ = k φaij Di u+ Dj u+ ≥ λk |Du+ |2 φ,
Ωck Ωck Ωck

where the inequality follows by strict ellipticity of the aij ’s and the equality follows from the fact that
u+ is zero on the set {u < 0}, so that, where the integral is non-zero, we have u = u+ . Thus, we can
drop this term by positivity to infer
Z Z Z Z Z
aij Di uDj φ + k u+ aij Di uDj φ ≤ aij Di uDj φ + λk |Du+ |2 φ + k u+ aij Di uDj φ ≤ 0,
Ωk Ωck Ωk Ωck Ωck

or, alternatively, Z Z
aij Di uDj φ ≤ −k u+ aij Di uDj φ.
Ωk Ωck

The right-hand side of the above can then be bounded as follows

Z Z Z
−k u+ aij Di uDj φ = −k u+ aij Di u+ Dj φ ≤ Λ |Du+ ||Dφ|,
Ωck Ωck Ωck

where we used that 1 ≤ ku+ on Ωck . Here Λ denotes, as in the rest of the example sheet, the upper
bound of the aij ’s. As far as the right-hand side is concerned, decompose u in positive and negative
part to get Z Z Z
aij Di u+ Dj φ ≤ Λ |Du||Dφ| + aij Di u− Dj φ = I + II.
Ωk Ωck Ωk

To conclude notice that Ωk ⊂ Ωk+1 and that χΩk → χ{u≥0} pointwise a.e., so that, by dominated
convergence, we obtain
Z Z
I=Λ |Du+ ||Dφ| → Λ |Du+ ||Dφ| = 0,
Ωck {u≤0}

whereas Z Z
−
II = ij
a Di u Dj φ → aij Di u− Dj φ = 0,
Ωk {u≥0}

because u− |{u≥0} = 0. Therefore,

Z Z
aij Di u+ Dj φ = aij Di u+ Dj φ ≤ 0,
B {u≥0}

thus giving the desired inequality and proving that u+ is a subsolution as well.
Exercise 5 (Strong maximum principle for divergence forms operators). Let Ω ⊂ Rn be a domain,
and u ∈ W 1,2 (Ω) be a weak subsolution of the strictly elliptic equation Di (aij Di u) = 0 in Ω with
coefficients aij ∈ L∞ (Ω). If for some ball B b Ω we have supB u = supΩ u, show that u must be
constant in Ω.
Solution 5. Recall the weak Harnack inequality for non-negative supersolutions proven in class. As-
suming the usual hypothesis on the coefficients we have that for a weak supersolution u of Lu = 0 in
Ω, for any BR (y) b Ω and for any p ∈ [1, n/(n − 2)) the following bound holds
R−n/p kukLp (BR (y)) ≤ C inf u,
BR/2 (y)

where C = C(n, λ, Λ, p). The exercise is an application of this Harnack inequality. Indeed, denote by
l = supB u = supΩ u and note that v = l − u is a supersolution (given that u is a subsolution) and
apply Harnack’s inquality with p = 1,
 
−n
R kl − ukL (B/2) ≤ C inf (l − u) = C l − sup u = 0.
1
B B

37
Note that in the previous inequality we could have taken
 
−n
R kl − ukL1 (B) ≤ C inf (l − u) = C l − sup u = 0.
2B 2B

At this point a closed and open argument should allow to conclude, due to the connectedness of Ω.
Otherwise, fix a point p ∈ Ω and find a collection of balls Bi , for i ∈ {1, 2, . . . k} such that p ∈ Bk , and
Bi+1 ∩ Bi 6= ∅ and 2Bj ⊂ Ω. In other words, by path connectedness we can find a curve joining any
element of our initial ball B with any point p. Cover this ball by finitely many balls and apply the
previous argument with the Harnack inequality to any such ball, starting from the first one intersecting
B. Thus, at p the function u is constantly equal to l and by arbitrarity of p the proof is concluded.
For the following exercise recall Exercise 8 of the third Example Sheet. Moreover, we shall need the
following result, which is a global version of the interior De Giorgi-Nash-Moser continuity estimate
proven in the lecture.
Theorem 4. Let Ω be a bounded C 2 domain in Rn and aij ∈ L∞ (Ω) with
n
X
kaij kL∞ (Ω) ≤ γ and aij ζ i ζ j ≥ λ|ζ|2 for a.e. x ∈ Ω and all ζ ∈ Rn ,
i,j=1

for some constants λ, γ > 0. Let φ ∈ C 1 (Ω). There exists β = β(n, λ, γ, φ, Ω) ∈ (0, 1) and C =
C(n, λ, γ, φ, Ω) such that if u ∈ W 1,2 (Ω) ∩ C 0 (Ω) is a weak solution to
(
Di (aij Dj u) = 0, in Ω,
u = φ, on ∂Ω;

then u ∈ C 0,β (Ω) and |u|0,β;Ω ≤ C.

We will also need the following.
Theorem 5. Let Ω be a bounded open domain satisfying a uniform exterior cone condition for some
cone C. Let L = Di (aij Dj ) be a uniformly elliptic operator on Ω and let u ∈ W 1,2 (B1 ) ∩ C 0 (B1 ) be a
weak solution of Lu = 0 in Ω. If there exist constants α0 , K > 0 such that for all x ∈ ∂Ω and r > 0
the following hold
oscBr (x)∩∂Ω u ≤ Krα0 ,
then there exists α = α(n, λ, Λ, α0 , C, diam(Ω)) and C = C(n, λ, Λ, α0 , C, diam(Ω)) such that u ∈
C 0,α (Ω) with the estimate
|u|0,α;Ω ≤ C (|u|0;Ω + K) .
Exercise 6. Let α ∈ (0, 1) and let Ω be a bounded C 2,α domain in Rn . Let φ ∈ C 2,α (Ω). Suppose
that there exists a constant K > 0, depending on n, α, Ω and φ, such that for each σ ∈ [0, 1] and each
u ∈ C 2 (Ω) solving   
div √ Du = 0 in Ω,
1+|Du|2
u = σφ on ∂Ω,


we have sup∂Ω |Du| < K. Show that there is a unique u ∈ C 2,α (Ω) solving
  
div √ Du = 0 in Ω,
1+|Du|2
u=φ on ∂Ω.


Solution 6. We start by clarifying the minimal surface operator we are going to work with. The
minimal surface equation in divergence form is the one given in the statement of the exercise:
n
!
X Di u
Mu = Di p = 0.
i=1
1 + |Du|2

38
Calculating we have
n
! n
X δij Di uDj u X
Mu = p − Dij u = aij (x, u, Du)Dij u = 0.
i,j=1
1 + |Du|2 (1 + |Du|2 )3/2 i,j=1

Notice that multiplying the above by (1 + |Du|2 )3/2 we obtain

n
X
δij (1 + |Du|2 ) − Di uDj u Dij u = 0,


Qσ u = Qu =
i,j=1

an equivalent expression for the minimal surface equation. Notice that in fact Qσ does not depend
on σ. Compare it with the statement of the Leray-Schauder theorem and the associated exercise on
the third example sheet. Now, u ∈ C 2 (Ω), so that aij ∈ C 1 (Ω) for all i, j ∈ {1, 2, . . . , n}. By higher
interior regularity we have that u is smooth in interior. For further details we refer the reader to the
remark following Exercise 7 in the third example sheet. In particular, we can differentiate the equation
M u = 0 with respect to any index k ∈ {1, 2, . . . , n} to yield the following, where we denoted w = Dk u,
 
n n
X 1 X Di uDj u
0= Di  p Di w − Dj w
i=1
1 + |Du| 2
j=1
(1 + |Du|2 )3/2
  
n n
X δ ij
X D i uD j u
= Di  p − 2 )3/2
 Dj w
i=1
1 + |Du| 2
j=1
(1 + |Du|
n
X
Di αij (x, u, Du)Dj w .


=
i,j=1

In other words w = Dk u satisfies a divergence form equation for every k ∈ {1, 2, . . . , n}. Moreover,
the αij ’s are strictly elliptic. Indeed, for every ξ ∈ Rn we obtain

|ξ|2 Di uDj u
αij ξi ξj = p − ξi ξj
1 + |Du|2 (1 + |Du|2 )3/2
|ξ|2 hDu, ξi2 |ξ|2 |Du|2 |ξ|2
 
=p − ≥ 1− ≥ ≥ λ0 |ξ|2 ,
1 + |Du|2
p
1 + |Du|2 (1 + |Du|2 )3/2 1 + |Du|2 (1 + |Du|2 )3/2

for a certain λ0 > 0 uniform on the domain. The existence of such a λ0 follows from the fact that
u ∈ C 1 (Ω). The inequality in the above follows from Cauchy-Schwarz. Consider then
n
X n
X
v = |Du|2 = (Dk u)2 = wk2 .
k=1 k=1

We claim that v is a subsolution of Di (αij Dj v) = 0. Indeed, by direct calculation we get

n
X
Di (αij Dj v) = 2Di αij ωk Dj ωk

k=1
Xn n

X
= 2 Di ωk Dj ωk αij + ωk Di (αij Dj ωk ) ≥ 2λ0 |Dwk |2 ≥ 0,
k=1 k=1

which gives, for every φ ∈ W 1,2 (Ω) such that φ ≥ 0, the following inequality
Z Z n Z
X
− αij Di vDj φ = Di (αij Dj w)φ ≥ 2λ|Dwk |2 φ ≥ 0.
Ω Ω k=1 B

39
Thus, by maximum principle for divergence form equations we obtain

sup |Du| ≤ sup |Du| < K = K(n, α, Ω, φ),

Ω ∂Ω

where the second inequality follows by assumption. To conclude we only need to bound oscillations
along the boundary to apply the boundary regularity version of De Giorgi-Nash-Moser. Consequently,
we will infer that |Du|0,β;Ω is bounded. More specifically, u = σφ on ∂Ω and the data φ is assumed to
be in C 2,α (Ω), so that for all x ∈ ∂Ω and all r > 0 we get

oscBr (x)∩∂Ω |Du| ≤ σ[Dφ]α;Ω rα .

Notice that for this it would have been enough to have φ ∈ C 1,α0 (Ω) for a certain α0 > 0. For constants
β = β(n, λ, Λ, α0 , C, diam(Ω)) and C = C(n, λ, Λ, α0 , C, diam(Ω)) we obtain

|Du|0,β;Ω ≤ C (|Du|0;Ω + σ[Dφ]α;Ω ) ≤ C(K + M ),

where the last inequality is independent of σ. Summarising, we proved

|uσ |1,β;Ω ≤ C,

for a certain C = C(n, λ, Λ, α, diam(Ω), φ) and where uσ is a solution of the minimal surface equation
M u = 0 with boundary data u = σφ. The Leray-Schauder fixed point theorem allows to to conclude
that we have existence u ∈ C 2,α (Ω) of the minimal surface equation with boundary values u = φ.
Uniqueness follows as in Exercise 8 of the first example sheet. Namely, argue by contradiction and
suppose that there are two solutions of the minimal surface equation with the same boundary condition.
Their difference satisfies then a linear partial differential equations with zero boundary data. Applying
the weak maximum principle we infer that the solution is identically zero, thus implying the desired
uniqueness.
Remark 10. In order to apply the Leray-Schauder fixed point theorem, it would have been sufficient
to obtain these bounds for solutions uσ which are known a priori to belong to C 2,αβ (Ω). However, this
was not necessary in our argument.
Remark 11. The above result reduces the solvability of the Dirichlet problem for the minimal surface
equation on a bounded C 2,α domain Ω to an a priori boundary gradient estimate for solutions; with
further work, it can be shown that such an estimate holds for arbitrary boundary data φ ∈ C 2,α (Ω)
if and only if the boundary ∂Ω is mean convex, i.e. has non-negative mean curvature with respect to
the inward pointing unit normal. In contrast, for Laplace’s equation or indeed for other more general
linear equations, no such geometric condition on the domain is necessary for solvability.
Remark 12. The remarkable fact of the De Giorgi-Nash-Moser theory is not requiring any regularity of
the coefficients. Whence its essential role in the regularity theory of minimisers of elliptic functionals.
Finally, let us mention that the De Giorgi-Nash-Moser theory does not hold for systems. This is a
significant obstacle for the regularity theory of minimal surfaces in higher codimensions.

Exercise
Pn 7. Let Ω ⊂ Rn be a bounded domain and aij be bounded, measurable functions on Ω with
ij 2 2 ij i j 2 n
i,j=1 ka kL∞ (Ω) ≤ Λ and a (x)ζ ζ ≥ λ|ζ| for all ζ ∈ R and a.e. x ∈ Ω, where λ, Λ > 0 are
1,2 ij
constants. Let u ∈ W (Ω) be a weak solution of Di (a Dj u) = 0 in Ω.
(i) (Weak maximum principle). Prove that

sup u ≤ sup u+ .
Ω ∂Ω

Here for v ∈ W 1,2 (Ω) we define

sup v = inf{k ∈ R; (v − k)+ ∈ W01,2 (Ω)}.

∂Ω

40
 (ii) A standard existence theorem implies that for any given φ ∈ W 1,2 (Ω), there is a unique u ∈
W 1,2 (Ω) in Ω and u − φ ∈ W01,2 (Ω). Give a proof of this, without appealing to a more general
theorem. Use this result to prove that for any φ ∈ C 0 (∂Ω), there exists a unique solution u ∈
1,2
Wloc (Ω) ∩ C 0 (Ω) to the Dirichlet problem Di (aij Dj u) = 0 weakly in Ω and u = φ on ∂Ω. Does
this solution necessarily belong to W 1,2 (Ω)?.
Solution 7. We shall prove the two claims separately.
(i) We start by recalling that u ∈ W 1,2 (Ω) is a weak solution of Di (aij Dj u) = 0 in Ω provided
Z
aij Di vDj u = 0,
Ω

for all v ∈ W01,2 (Ω) with v ≥ 0. Thus, considering v = max(u − l, 0), where l = sup∂Ω u+ we
obtain Z Z
ij
0= a Di uDj (u − l) ≥ λ |Du|2 .
Ω∩supp v supp v∩Ω

Note that we can use such v as a test function in the weak formulation as it is trivially positive
and it vanishes on the boundary. Now, this inequality implies Du = 0 a.e. on supp v. Since
Dv = Du on supp v and Dv = 0 otherwise, we conclude that Dv = 0 a.e. on Ω, i.e. v is constant
a.e. As v ∈ W01,2 (Ω) we conclude that v ≡ 0 a.e. or, in other words, u ≤ sup∂Ω u+ . One could
have concluded that Dv = 0 a.e. in the following way:
Z Z
aij Di vDj v = aij Di vDj u = 0,
Ω Ω

from which, once again by ellipticity, one can deduce the desired result.
(ii) We assume the coefficients aij to be symmetric. This is done so that the following simple argument
using the Riesz representation theorem can be used. Otherwise, one can apply Lax-Milgram’s
theorem.
For u, v ∈ W01,2 (Ω), let hu, vi = Ω aij Di uDj v. This clearly defines an inner product on W01,2 (Ω)
R

One can check linearity, symmetry and non-degeneracy pretty trivially. However, it is worth
recording that by virtue of the inequality hu, ui ≥ γ Ω |Du|2 ≥ and the boundedness of aij , the
R

induced norm kuk = hu, ui1/2 is equivalent to the usual norm kuk2H 1 = Ω |Du|2 . Consequently,
R
0

W01,2 (Ω) equipped with this inner product is a Hilbert space. Consider then the functional
F : W01,2 (Ω) → R defined by Z
F (v) = − aij Di ϕDj v.
Ω

It is clearly linear and, by Hölder’s inequality and the above lower bound on hu, ui, we have
 
|F (v)| ≤ γ −1/2 sup sup |aij | kDϕkL2 (Ω) kvk,
1≤i,j≤n Ω

from which we deduce that F is bounded. Hence, by the Riesz representation theorem, there is a
unique function w ∈ W01,2 (Ω) such that hu, vi = F (v), for every v ∈ W01,2 (Ω). Letting u = w + ϕ,
we have that u ∈ W 1,2 (Ω) and u − ϕ ∈ W01,2 (Ω), as well as
Z
aij Di uDj v = 0,
Ω

for each v ∈ W01,2 (Ω), i.e. u is a weak solution of Di (aij Dj u) = 0 in Ω.

1,2
We are left with proving that for any φ ∈ C 0 (∂Ω), there exists a unique solution u ∈ Wloc (Ω) ∩
C 0 (Ω) of the Dirichlet problem Di (aij Dj u) = 0 weakly in Ω and u = φ on the boundary. Extend
φ to a continuous function with compact support defined on the whole of Rn . We shall not change

41
 the notation in the coming paragraphs, i.e. denote it φ ∈ Cc0 (Rn ). Mollifying the sequence as we
have done many times, we have the existence of a sequence of smooth functions φk converging
uniformly towards φ. Uniformly in Rn and on the boundary ∂Ω. By the above, we have the
existence of a weak solution um ∈ W 1,2 (Ω) to the system
(
Di (aij Dj v) = 0 in Ω,
v = φm on ∂Ω.

Note that the same reasoning gives that v = um − ul is a solution of

(
Di (aij Dj v) = 0 in Ω,
v = φm − φl on ∂Ω.

By part (i) of the exercise we infer the uniform bound

kum − ul kL∞ (Ω) ≤ Ckφm − φl kL∞ (∂Ω) .

Moreover, by boundedness of the domain Ω one can obtain L2 -bounds as well. Consequently, as
the sequence of {φm } is Cauchy, the sequence {um } is Cauchy as well. Thus, we have that it
converges to a certain u uniformly and in L2 as well. In particular, u = φ on ∂Ω. To conclude
we need to examine what happens to the L2 -norm of |Dum | in order to infer that we have a
weak solution of our initial system. The inequality proven in Exercise 1 can be adapted to the
following: kDvkL2 (Ω0 ) ≤ CkvkL2 (Ω̃) , for any Ω0 b Ω̃. Thus, we have that the sequence {um } is
also Cauchy in H 1 (Ω0 ) for any Ω0 b Ω. Consider an exhaustion of such subsets of our domain Ω.
1,2
Using a diagonal argument we have the existence of f ∈ Wloc . Uniqueness of the limit implies
that f has to agree a.e. with the u we found above. Passing to the limit in the weak formulation
of the equation we find that u is a weak solution of Di (aij Dj u) = 0 with the correct boundary
values. We are left with checking that u is in C 0 (Ω). Proving that u is continuous in every
domain of the exhaustion implies u ∈ C 0 (Ω), which, combined with φ ∈ C 0 (∂Ω) implies the
desired result. The global De Giorgi-Nash-Moser regularity theorem allows to conclude. One can
reapply the sequence argument to any domain of the exhaustion and conclude.
We now construct a counterexample proving that the solution find above does not necessarily
belong to W 1,2 (Ω). Recall Exercise 4 from the first Example Sheet:
∞
X 1 n!
u(r, θ) = r sin(n!θ),
n=1
n2

is harmonic in B ⊂ R2 the unit ball, and continuous in B. However, u does not have finite
Dirichlet energy, thus proving that u 6∈ W 1,2 (B).
Remark 13. More interesting is the proof of (i) when the differential operator under consideration has
non-zero first and zero order terms. For a proof of this we refer the reader to [1, Theorem 8.1].

2 Miscellaneous Facts and Results

Definition 1. Let Ω be an open set of Rn and let x0 ∈ Ω. Let u : Ω\{x0 } → R be a harmonic function
in Ω \ {x0 }. Then, x0 is a removable singularity of u if there exists a harmonic function u in Ω such
that u = u in Ω \ {x0 }.
We then have the following result on the extendibility of harmonic functions.
Exercise 1. Let u be a harmonic function in Ω\{x0 } and assume that u(x) = o(Γ(x−x0 )), as x → x0 ,
where Γ is the fundamental solution of the Laplacian with a singularity at the origin, i.e. assume that

lim u(x)|x − x0 |n−2 = 0, if n ≥ 3,

x→x0

42
and
u(x)
lim = 0, if n = 2.
x→x0 log |x − x0 |
Then, x0 is a removable singularity of u.
Remark 14. Note that the above theorem is sharp as x0 is not a removable singularity for Γ(x − x0 ).
Moreover, this theorem is the analogue of Riemann’s one on the removability of singularities for
holomorphic functions.
Proof. Let r > 0 be such that Br (x0 ) ⊂ Ω and denote by u the solution of the Dirichlet problem
(
∆u = 0, in Br (x0 ),
u(x) = u(x), on ∂Br (x0 ).

Then, by maximum principle u is bounded in Br (x0 ) and it is harmonic in Br (x0 ). To conclude the
proof it suffices to show that u = u on Br (x0 ). Let then w = u − u. Such a function w is harmonic in
Br (x0 ) \ {x0 } and satisfies

w(x)
lim = 0, if n ≥ 3,
x→x0 r 2−n − |x − x0 |2−n

as well as
w(x)
lim = 0, if n = 2.
x→x0 log(r) − log |x − x0 |
In other words, for every  > 0, there exists δ > 0 such that for all x with |x − x0 | < δ, we have

|w(x)| ≤  r2−n − |x − x0 |2−n , if n ≥ 3,

and
|w(x)| ≤  (log(r) − log |x − x0 |) , if n ≥ 3.
Observe now that the terms on the right-hand sides of the above equations are harmonic functions in
the respective dimensions and, moreover, they vanish on ∂Br (x0 ). Thus, by weak maximum principle
applied to the annulus δ ≤ |x − x0 | ≤ r we infer

|w(x)| ≤  r2−n − |x − x0 |2−n , if n ≥ 3,

and
|w(x)| ≤  (log(r) − log |x − x0 |) , if n ≥ 3,
for all x such that |x − x0 | ≤ r. Arbitrariness of  > 0 allows to conclude that w = 0, from which we
deduce the desired result.
Assume without loss of generality that x0 is given by the origin. The above theorem is essentially
telling us that if the function u blows up slower than the fundamental solution, then it can be extended
to a harmonic function on the whole domain. Conversely, if a function blows up to quickly at the origin,
then it cannot be extended. For instance, in Rn the function u can decay at most like 1/|x|n−2− , for
 arbitrarily close to zero. In R4 we cannot have (or better extend) functions blowing up like 1/|x|2 .
Note that this implies in particular that a bounded harmonic function on a punctured domain can be
extended.
Exercise 2. Recall that the Bernstein theorem for minimal graphs says that if 1 ≤ n ≤ 7, and if
u ∈ C 2 (Rn ) is a solution of the minimal surface equation
!
Du
div p =0
1 + |Du|2

43
on Rn , then u is an affine function. Use this result to prove that if 1 ≤ n ≤ 7 and u ∈ C 2 (Rn ) solves
!
Du
div p = κ,
1 + |Du|2

for some constant κ, then u is an affine function.

Solution 1. Integrate the constant mean curvature equation over the ball BR (0) to obtain
Z
Du x
p · = κ|BR (0)|,
1 + |Du| 2 R
∂BR (0)

where we applied the divergence theorem to obtain the expression on the left-hand side, while the term
on the right-hand side is simply given by κωn Rn . Note now that the integrand on the left-hand side is
smaller than 1, so that we deduce κωn Rn ≤ |∂BR (0)| = nωn Rn−1 . Dividing this identity by Rn and
letting R → ∞ we infer κ = 0. Apply then the Bernstein theorem for solutions of the minimal surface
equation to conclude that u is affine.
Exercise 3. Let α ∈ (0, 1) and let Ω ⊂ Rn be a bounded, mean-convex, C 2,α -domain. Recall that for
any given φ ∈ C 2,α (Ω), there is a unique function u ∈ C 2,α (Ω) solving the minimal surface equation
M(u) = 0 in Ω, with u = φ on ∂Ω. Use this result, the interior gradient estimate for solutions of
the minimal surface equation, together with an approximation argument to show that for any given
φ ∈ C 0 (∂Ω), there is a unique function u ∈ C 2 (Ω) ∩ C 0 (Ω) such that M(u) = 0 in Ω and u = φ on
∂Ω.
Exercise 4. In contrast to the case of harmonic functions and the associated Dirichlet energy, if
Ω ⊂ Rn is a bounded Lipschitz domain and u ∈ C 2 (Ω) ∩ C 0 (Ω) is a solution of the minimal surface
equation !
Du
div p =0
1 + |Du|2
in Ω, then the area Z p
A(u, Ω) = 1 + |Du|2
Ω

of graph(u) is finite. Prove this by completing the following outline.

(a) Show that if A(u, Ω) < +∞, the u minimises the area functional in the sense that A(u, Ω) ≤
A(v, Ω) for every Lipschitz function v on Ω with u = v on ∂Ω.

(b) For δ > 0, let Ωδ := {x ∈ Ω; dist(x, ∂Ω) > δ}. Show that
Z
n 1
A(v, Ωδ ) ≤ H (Ω) + |Du| + Hn (Ωδ \ Ω2δ ) sup |u|.
Ωδ \Ω2δ δ Ω

Deduce that A(u, Ω2δ ) ≤ 2Hn (Ω) + c supΩ |u|, where c = c(n), and conclude that A(u, Ω) < ∞.
[Hint: For the first estimate, consider the function v(x) = u(x)γδ (d(x)) for x ∈ Ωδ , where d(x) =
dist(x, ∂Ω) and γδ : [δ, ∞) → ∞ is the function γδ (t) := 2 − δ −1 t if δ ≤ t ≤ 2δ and γδ (t) = 0 if
t > δ.]
Exercise 5 (Method of sub/super solutions for semi-linear equations). Let Ω be a bounded C 2,α -
domain in Rn for some α ∈ (0, 1), and let ψ ∈ C 2,α (Ω). Let Q be the differential operator defined
by
Qu := ∆u − V (u),
where V : R → R is a given, smooth, non-increasing function. Suppose that there are functions ϕ± ∈
C 2 (ω) ∩ C 0 (Ω), with Qϕ+ ≤ 0 ≤ Qϕ− in Ω, and ϕ− ≤ ϕ+ in Ω, together with ϕ+ ≥ ψ ≥ ϕ− on ∂Ω

44
(a) Show that there exists a function u1 ∈ C 2,α (Ω) with ϕ− ≤ u1 ≤ ϕ+ in Ω such that ∆u1 = V (ϕ− )
in Ω and u1 = ψ on ∂Ω.
(b) Deduce that there exists a sequence of functions uk ∈ C 2,α (Ω) with ϕ− ≤ u1 ≤ u2 ≤ . . . ≤ ϕ+ in
Ω, such that ∆uk = V (uk−1 ) in Ω, uk = ψ on ∂Ω, for each k = 1, 2, . . ., where u0 = ϕ− .
(c) Show that there exists a constant C > 0 such that the functions uk in the previous step satisfy
1
|uk |2,α;Ω ≤ |uk−1 |2,α;Ω + C,
2
for all k = 2, 3, . . .. [Hint: The following interpolation inequality might be helpful: for each  > 0,
there exists γ() > 0 such that for any v ∈ C 2,α (Ω) there holds

|v|0,α;Ω ≤ |v|2,α;Ω + γ()|v|0;Ω .]

(d) Deduce that there exists u ∈ C 2,α (Ω) with ϕ− ≤ u ≤ ϕ+ in Ω, such that Qu = 0 in Ω and u = ψ
on ∂Ω.

References
[1] David Gilbarg and Neil S. Trudinger, Elliptic partial differential equations of second order, second
ed., Grundlehren der mathematischen Wissenschaften [Fundamental Principles of Mathematical
Sciences], vol. 224, Springer-Verlag, Berlin, 1983. MR 737190 42
[2] Qing Han and Fanghua Lin, Elliptic partial differential equations, Courant Lecture Notes in Math-
ematics, vol. 1, New York University, Courant Institute of Mathematical Sciences, New York;
American Mathematical Society, Providence, RI, 1997. MR 1669352 7, 21, 32

45